text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'In a simple model of propagation of asymmetric Gaussian beams in nonlinear waveguides, described by a reduction to ordinary differential eqautions of generalized nonlinear Schrödinger equations (GNLSEs) with cubic-quintic (CQ) and saturable (SAT) nonlinearities and a graded-index profile, the beam widths exhibit two different types of beating behavior, with transitions between them. We present an analytic model to explain these phenomena, which originate in a $1:1$ resonance in a 2 degree-of-freedom Hamiltonian system. We show how small oscillations near a fixed point close to $1:1$ resonance in such a system can be approximated using an integrable Hamiltonian and, ultimately, by a single first order differential equation. In particular, the beating transitions can be located from coincidences of roots of a pair of quadratic equations, with coefficients determined (in a highly complex manner) by the internal parameters and initial conditions of the original system. The results of the analytic model agree with numerics of the original system over large parameter ranges, and allow new predictions that can be verified directly. In the CQ case we identify a band of beam energies for which there is only a single beating transition (as opposed to $0$ or $2$) as the eccentricity is increased. In the SAT case we explain the sudden (dis)appearance of beating transitions for certain values of the other parameters as the grade-index is changed.'
author:
- |
David Ianetz$^{1,2,3}$ and Jeremy Schiff$^{1,4}$\
$^1$Department of Mathematics, Bar-Ilan University, Ramat-Gan 5290002, Israel\
$^2$Holon Institute of Technology (HIT), Holon 5810201, Israel\
$^3$E-mail: [email protected]\
$^4$E-mail: [email protected]
bibliography:
- 'w.bib'
title: Analytic Methods to Find Beating Transitions of Asymmetric Gaussian Beams in GNLS equations
---
Introduction
============
In the sequence of papers [@Ianetz2010PRA_1; @Ianetz2010PRA_2; @Ianetz2013PRA] a variational approach was taken to investigate the propagation of asymmetric (elliptic) Gaussian beams in nonlinear waveguides, with cubic-quintic and saturable nonlinearities and a parabolic graded-index (GRIN) profile, as described by suitable generalized nonlinear Schrödinger equations (GNLSEs). The beam widths in the two transverse directions to the direction of propagation were found to obey a set of ordinary differential equations which can be identified as the equations of motion of a point particle in certain rather complicated, but tractable, 2d potentials. Numerical analysis of these equations revealed “beating” phenomena: in addition to fast oscillations, the beam widths exhibit a (relatively) slow periodic variation. Furthermore, two types of beating were identified: In type I beating the amplitude of oscillation of the beam width in one direction remains greater than the amplitude of oscillation in the other direction, whereas in type II, there is an interchange between the widths in the two transverse directions. The type of beating depends on the parameters of the system and initial eccentricity of the beam. Remarkably, as the initial eccentricty or other parameters are changed, there can be a transition between types, and this transition is characterized by a singularity in the ratio of the periods of the beating and of the fast oscillatory motion.
The intention of the current paper is to provide a theoretical analysis of the beating phenomena and, in particular, to present an approximate analytic method to find the transitions between types. The relevant tool is the analysis of small oscillations in $2$ degree-of-freedom Hamiltonian systems near a fixed point which is close to $1:1$ resonance. The fact that resonance is the source of “beating” or “energy transfer” phenomena in mechanical systems is well known. A classic example can be found in the paper of Breitenburger and Mueller [@BM1] on the elastic pendulum, which the authors describe as a “paradigm of a conservative, autoparametric system with an internal resonance”. The paper [@BM1] has other features in common with our work (such as the use of action-angle variables and the fact that the analytic approximation used is a single elliptic function equation) but it is in the much simpler context of $1:2$ resonance. For other examples of autoparametric resonance see, for example, [@vpapr; @Hallerbook]. The most widely used tool for analysis of systems near resonance is the mutliple time scale method, see for example [@kc; @ManMan] for thorough presentations and many examples. For a typical modern application see [@VS1; @VS2]. However, averaging techniques present an alternative [@SVM], and in the context of Hamiltonian systems, working in action-angle coordinates has substantial advantages [@gensap]. A typical study of a system near resonance will involve looking at the bifurcations of special solutions. In this context much attention has been paid to the definition and identification of [*nonlinear normal modes*]{} — see [@mikhbar] for a review, and [@RPV] for an example in the context of $1:1$ resonance.
The $2$ degree-of-freedom Hamiltonian systems we study have a discrete symmetry, and are approximated by a family of systems with $1:1$ resonance studied nearly $40$ years ago by Verhulst [@Verhulst1]. Verhulst showed the existence of an approximate second integral and used this to study bifurcations of special solutions and their stability. Our work differs from that of Verhulst and other works on $1:1$ resonance in several regards. The bifurcation question we pose depends not only on the internal parameters of the system, but also on the initial conditions. The question is not only one of identifying different types of solutions of the system, but also seeing how the type of solution changes as both the initial condition and internal system parameters are varied. We have not seen a similar study in the highly complex context of $1:1$ resonance. Our methodology uses action-angle variables and canonical transformations (though in an appendix we show how to apply standard two time scale techniques). Unlike in most existing studies, it is necessary to compute the relevant canonical transformation to [*second*]{} order. However, this does not affect the result that once the correct canonical transformation has been applied, the resulting approximating Hamiltonian depends only on a single combination of the angle variables and is integrable. The equations of motion for the integrable Hamiltonian can be reduced to a single first order differential equation, and the rich bifurcation structure of the systems we study can reduces to understanding the bifurcations of roots of a pair of quadratic polynomials, with coefficients that depend (in a complex, nonexplicit manner) on the internal parameters of the systems and the initial conditions. Comparison with numerical results shows our method gives high-quality results in a significant region of parameter space, and allows a variety of interesting new predictions.
The structure of this paper is as follows. In the next section we review the relevant models from nonlinear optics and the collective variable approximation to obtain equations for the propagation of beam widths, and present the main findings of papers [@Ianetz2010PRA_1; @Ianetz2010PRA_2; @Ianetz2013PRA] and some further numerical results. In section 3, we develop our method of integrable approximation for small oscillations in a $2$ degree-of-freedom Hamiltonian system near a fixed point close to $1:1$ resonance. In section 4 we describe the application of this method to the specific systems relevant to beam propagation, confirming existing numerical results and presenting new predictions. In section 5 we summarize and conclude. Appendix A completes some technical details omitted from the main text, and Appendix B describes an alternate method of approximation of the full equations using a two time expansion. This is a more [*ad hoc*]{} approach than the one explained in section 3, but we include it as it is more commonly used in the literature, and for certain values of parameters it gives better results.
Before closing this introduction we mention a number of points concerning the relevance of the work in this paper to optical solitons. We will describe in next section the manner in which we use ordinary differential equations (ODEs) to study the behavior of solutions of GNLSEs. The use of ODEs to study GNLSEs is widespread, see for example [@Skarka2006PRL; @Skarka2008JOA; @Skarka2012PS; @hemalomed; @Skarka2014PRA; @Aleksic2015PRA] In particular, the last two papers use ODE methods in the study of rotating solitons. Our work extends the catalog of interesting bifurcations that can be observed in the context of GNLSEs; for another example; see the papers [@gmc1; @gmc2] for a case of a saddle-loop bifurcation. Finally, we mention that we neglect dispersive terms in the GNLSEs we study. This is justifiable in the context of new optical materials [@Kim2002OL; @Moon2008JNCS; @JU2011OE; @gold; @noptrev] characterized by Kerr coefficients of the order $10^{-11}$–$10^{-12}$ ${\rm cm}^2/{\rm W}$, making the critical intensity for self-focusing small enough that it can be reached using microsecond pulses and possibly even continuous wave (CW) laser beams.
Models, the collective variable approach and numerical results
==============================================================
We consider beam propagation in a nonlinear, graded-index fiber, as described by one of the following GNLSEs: $$\begin{aligned}
2i\psi_z + \psi_{xx} + \psi_{yy} + \left( |\psi|^2 - Q|\psi|^4 - g(x^2+y^2) \right) \psi &=& 0\ ,
\label{glnse1}\\
2i\psi_z + \psi_{xx} + \psi_{yy} + \left( \frac{|\psi|^2}{1+\alpha^2|\psi|^2} - g(x^2+y^2) \right) \psi &=& 0\ .
\label{glnse2} \end{aligned}$$ Here, modulo suitable normalizations [@Kivshar2003book; @Ianetz2013PRA], $\psi$ is the strength of the electric field, $z$ is the longitdinal coordinate, $x,y$ are transverse coordinates, and $Q,\alpha,g$ are parameters. The first equation is the case of cubic-quintic nonlinearity (CQ), the second is the case of saturable nonlinearity (SAT). In the low intensity limit these models are similar, but for higher intensity they display different physical properties. In both cases, the higher order nonlinearity prevents beam collapse associated with the standard Kerr nonlinearity [@Chen2004PRE; @Kivshar2003book]. The term $-g(x^2+y^2)\psi$ reflects the graded-index nature of the fibre, that the refractive index $n$ falls with distance $r$ from the center of the fibre according to the law $n^2 = n_0^2 - Gr^2$; the physical significance of this is explained in [@Sodha1977book; @Ghatak1978book; @Ianetz2013PRA].
The collective variable approximation (CVA), introduced for the study of self-focusing beams in [@Anderson1979PF; @Anderson1979PF2; @Anderson1983PRA], is a variational technique to approximate solutions of nonlinear Schrödinger-type equations which has been used and validated in many different situations [@Malomed2002PO]. The method replaces partial differential equations such as (\[glnse1\]) and (\[glnse2\]) by a system of ordinary differential equations for the coefficients of an ansatz for the full solution. The GNLSEs (\[glnse1\]) and (\[glnse2\]) are variational equations for action principles based on the Lagrangian densities $$\begin{aligned}
{\cal L}_{\rm CQ} &=& i\left( \psi\psi^*_z-\psi^*\psi_z \right) + \left|\psi_{x}\right|^2 + \left|\psi_{y}\right|^2
- \frac12 |\psi|^4 + \frac{Q}{3}|\psi|^6
+ g(x^2+y^2) |\psi|^2 \ , \label{lag1}\\
{\cal L}_{\rm SAT} &=& i\left( \psi\psi^*_z-\psi^*\psi_z \right) + \left|\psi_{x}\right|^2 + \left|\psi_{y}\right|^2
+ \frac{\ln\left(1 + \alpha^2|\psi|^2\right)-\alpha^2|\psi|^2}{\alpha^4}
+ g(x^2+y^2) |\psi|^2\ . \nonumber \\
&& \label{lag2}\end{aligned}$$ We assume $\psi$ takes the form of the [*trial function*]{} $$\label{trial_function}
\psi_{T}(x,y,z) = A(z)
\exp \left(i\phi(z)
-\frac {x^2} {2a_x^2(z)} + ib_x(z)x^2
-\frac {y^2} {2a_y^2(z)} + ib_y(z)y^2
\right)\ ,$$ where $A,\phi,a_x,a_y,b_x,b_y$ are currently undetermined, real functions of only the longitudinal coordinate $z$. This trial function describes an elliptic Gaussian beam with $a_x,a_y$ representing the widths of the beam in the $x,y$ directions. $b_x,b_y$ describe curvatures of the beam wavefront, $A$ is the normalized amplitude of the electric field, and $\phi$ is a longitudinal phase factor. Our choice of a Gaussian shape for the trial function is appropriate because the Gaussian is an exact solution of the linear Schrödinger equation for GRIN waveguides [@Sodha1977book; @Ghatak1978book]. Substituting the trial function in the Lagrangian densities (\[lag1\]),(\[lag2\]) and computing the integrals over the variables $x,y$ we obtain reduced densities for the functions $A,\phi,a_x,a_y,b_x,b_y$. The corresponding Euler-Lagrange equations in the CQ case are $$\begin{aligned}
\dot{A}&=& -(b_x +b_y)A \ , \\
\dot{a}_{x,y} &=& 2a_{x,y}b_{x,y}\ , \\
\dot{b}_{x,y} &=& \frac{1}{2a^4_{x,y}} - 2b^2_{x,y} - \frac{g}{2}
- \frac{A^2}{a^2_{x,y}}\left( \frac{1}{8}- \frac{QA^2}{9} \right)\ , \\
\dot{\phi} &=& - \frac{1}{2a^2_x} - \frac{1}{2a^2_y} + \left( \frac{3}{8} - \frac{5QA^2}{18} \right) A^2
\ . \end{aligned}$$ Here a dot denotes differentiation with respect to $z$. In the SAT case the equations for $A,a_x,a_y$ remain the same, but those for $b_x,b_y,\phi$ are replaced by $$\begin{aligned}
\dot{b}_{x,y} &=& \frac {1} {2a_{x,y}^4} -2b_{x,y}^2 - \frac {g} {2} +
\frac {\ln(1+ \alpha^2 A^2)+\textrm{Li}_{2}(-\alpha^2 A^2)}{2\alpha^4 A^2a_{x,y}^2}\ , \\
\dot{\phi} &=& -\frac{1}{2a_x^2} - \frac{1}{2a_y^2}
+ \frac{\alpha^2A^2-2\ln(1+\alpha^2A^2)- \textrm{Li}_2(-\alpha^2 A^2)}{2\alpha^4A^2}\ .\end{aligned}$$ Here $\textrm{Li}_2(x)=\sum_{k=1}^{\infty} \frac{x^k}{k^2}$ is the Spence or dilogarithm function [@Abramowitz1970book]. For both CQ and SAT cases we observe that $A^2 a_xa_y$ is conserved [@Ianetz2010PRA_1; @Ianetz2010PRA_2; @Ianetz2013PRA], and we write $A^2 a_x a_y = 4 E$ ($4E$ is the beam energy), and use this to eliminate $A(z)$ . Furthermore, $\phi(z)$ evidently plays no role in determining the other functions and can be computed by a simple quadrature once the other functions have been found. Furthermore, it is clear that we can write $b_x$ ($b_y$) in terms of $a_x$ ($a_y$) and its $z$-derivative. Thus we can reduce the system of $6$ equations to a pair of second order equations for $a_x,a_y$. After some more calculation it emerges that the equations are simply the equations of motion $$\ddot{a}_x = -\frac{\partial V}{\partial a_x} \ , \qquad
\ddot{a}_y = -\frac{\partial V}{\partial a_y}
\label{em}$$ for a particle in a potential $V(a_x,a_y)$, where for CQ $$V = V_{CQ} \equiv \frac12 \left( \frac1{a_x^2} + \frac1{a_y^2} \right)
- \frac{E}{a_x a_y}
+ \frac{16 QE^2}{9a_x^2a_y^2}
+ \frac{g}{2} \left( a_x^{2}+a_y^{2} \right)\ ,
\label{VCQ}$$ and for SAT $$V =
V_{SAT} \equiv \frac12 \left( \frac1{a_x^2} + \frac1{a_y^2} \right)
-\frac{a_xa_y}{4E\alpha^4} {\rm Li}_2 \left( -\frac{4E\alpha^2}{a_xa_y} \right)
+ \frac{g}{2} \left( a_x^{2}+a_y^{2} \right)\ .
\label{VSAT}$$
Thus integration of equations (\[em\]) for potentials (\[VCQ\]) and (\[VSAT\]) provides a first approximation to solutions of the GNLSEs (\[glnse1\]) and (\[glnse2\]). Full numerical solutions of GNLSEs have been given in both the SAT [@Yang2002PRE] and the CQ [@Michinel2002PRE] cases with $g=0$. In [@Michinel2002PRE] it was shown that the breathing frequencies found numerically are similar to those obtained by the CVA technique. In [@Yang2002PRE] it was shown that the shape of the beam obtained numerically for a saturable medium remains similar to Gaussian, even for an asymmetric initial condition. However, use of direct numeric methods to give an overall picture of the behavior of a GNLSE, as a function of all the various parameters, remains a computationally overwhelming task, and having an qualitatively correct analytic or semianalytic model is therefore useful for developing physical insight [@Kivshar2003book; @Agrawal2007book].
Appropriate initial conditions for (\[em\]) are $$a_x(0) = a_0 r \ , \qquad a_y(0) = \frac{a_0}{r} \ , \qquad
\dot{a_x}(0) = \dot{a_y}(0) = 0\ . \label{ic0}$$ The latter two conditions are equivalent to taking $b_x(0)=b_y(0)=0$. Note that both the CQ and the SAT system have a scaling symmetry $$\begin{array}{lllll}
a_x \rightarrow \lambda a_x \ , &
a_y \rightarrow \lambda a_y\ , &
a_0 \rightarrow \lambda a_0 \ , &
r \rightarrow r \ , &
z \rightarrow \lambda^2 z \ ,\\
Q \rightarrow \lambda^2 Q\ , &
\alpha \rightarrow \lambda \alpha\ , &
g \rightarrow \lambda^{-4} g\ , &
E \rightarrow E . &
\end{array}$$ Thus for CQ we do not need to study the dependence of solutions on the $5$ parameters $Q,g,E,a_0,r$, but only on the $4$ scale invariant quantities $Qa_0^{-2}, ga_0^4, E, r$. On occasion we will work with the scale invariant quantity $K_{\rm CQ} = 4QEa_0^{-2}$ instead of the quantity $Qa_0^{-2}$. (For SAT, replace all instances of $Q$ in the previous two sentences with $\alpha^2$, and $K_{\rm SAT}=4\alpha^2Ea_0^{-2}$.) Note that since there is symmetry in both the models between $a_x$ and $a_y$, there is a $r\rightarrow \frac1{r}$ inversion symmetry, and thus we need only study $r\le 1$ or $r\ge 1$.
In the papers [@Ianetz2010PRA_1; @Ianetz2010PRA_2; @Ianetz2013PRA] the ODE systems above were studied numerically. For appropriate choices of the parameters “beating” phenomena were observed: in addition to (relatively) fast “breathing” oscillations, the beam widths exhibit a (relatively) slow periodic variation. Two types of beating were identified: In type I beating, the amplitude of oscillation of the beam width in one direction remains greater than the amplitude of oscillation in the other direction. In type II beating, there is an interchange between the widths in the two transverse directions. This is illustrated in Figure 1, which shows solutions of the CQ system for $E=2.039$, $K_{\rm CQ}=0.71$, $ga_0^4=0.01$, and two choices of $r$: $r=1.14$ gives type I beating, whereas $r=1.16$ gives type II beating.
![\[fig:epsart\] Dependence of the ratio of periods of slow beating and fast oscillatory motion, $L_{\rm beat}/L_{\rm br}$, on the parameter $r^2$ for the CQ model, and for various values of the parameter $ga_0^4$. $E=2.039$ and $K_{\rm CQ}=0.71$ throughout. (1) (dash-dot black) $ga_0^4=0$, (2) (dashed blue) $ga_0^4=0.01$, (3) (solid red) $ga_0^4=0.02$, (4) (dotted green) $ga_0^4=0.03$. Roman numerals indicate the type of beating in case (2), $ga_0^4=0.01$.](Fig_2.pdf){width="70.00000%"}
The type of beating depends on the parameters of the system and, as evident from Figure 1, on the initial eccentricity of the beam. Remarkably, as the initial eccentricty is increased, or as other parmeters are changed, there can be a transition between types. The approach to this transition is characterized by a divergence in the ratio of the periods of the slow beating and of the fast oscillatory motion. In Figure 2 this ratio (determined from numerical simulations) is plotted as a function of $r^2$ for the CQ system, with $E=2.039$, $K_{\rm CQ}=0.71$ and $ga_0^4=0,0.01,0.02,0.03$. (The reason for the choice of the coordinate $r^2$ on the $x$-axis is simply to make the plot clearer.) For $r$ just above $1$ the beating is type II, then there is a transition to type I, and then a second transition back to type II. The dependence on the system parameters of the two critical values of $r$, which we denote collectively by $r_c$, is explored further in Figure 3. In Figure 3a the values of $r_c$ are plotted as a function of $ga_0^4$ for three different values of $K_{\rm CQ}$ and a constant value of $E$; in Figure 3b $r_c$ is plotted as a function of $ga_0^4$ for three different values of $E$ and a constant value of $K_{\rm CQ}$. In general we see that $r_c$ increases as a function of $ga_0^4$ (for fixed $E,K_{\rm CQ}$). From Figure 3b we see that since the (solid) red is above the (dashed) blue is above the (dot-dashed) black, $r_c$ also increases as a function of $E$ (for fixed $ga_0^4,K_{\rm CQ}$). But in Figure 3a we see there is difference between the upper and lower branches of $r_c$. We deduce that the higher value of $r_c$ also increases with $K_{\rm CQ}$ (for fixed $E,ga_0^4$), but the lower value decreases.
We shall see later that for other values of $ga_0^4,K_{\rm CQ},E$ there can be just a single transition or no transitions at all as $r$ is increased from $1$. Transitions between beating types are also observed in the SAT system, again with a complex dependence on the parameters $ga_0^4$, $K_{\rm SAT}$ and $E$. The aim of this paper is to provide an integrable approximation for equations (\[em\]) with potentials (\[VCQ\]) and (\[VSAT\]) which provides a theoretical model to predict where the transitions between types take place.
Small oscillations near $1:1$ resonance
=======================================
In this section we describe a general process of approximation near a $1:1$ resonance for a $2$ degree-of-freedom Hamiltonian system with Hamiltonian $$H = \frac12 \left(p_x^2 + p_y^2 \right) + V(a_x,a_y)\ .
\label{H}$$ Here $a_x,a_y$ are the coordinates, $p_x,p_y$ are the conjugate momenta, and the potential $V$ (which typically will depend on a number of parameters) is symmetric, $V(a_x,a_y)=V(a_y,a_x)$. We assume that for typical values of the parameters the potential has an isolated symmetric minimum (at $a_x=a_y=a_{\rm min}$, say) at which the system is close to $1:1$ resonance. Note that because of the symmetry, $\displaystyle{ \frac{ \partial^2V}{\partial a_x^2} (a_{\rm min}, a_{\rm min})
= \frac{ \partial^2V}{\partial a_y^2} (a_{\rm min},a_{\rm min})}$. Thus the Hessian matrix of the potential at $(a_{\rm min},a_{\rm min})$ has eigenvectors $\left( \begin{array}{c} 1 \\ \pm 1 \end{array} \right)$ with eigenvalues $\displaystyle{\frac{ \partial^2V}{\partial a_x^2} (a_{\rm min},a_{\rm min})
\pm \frac{ \partial^2V}{\partial a_xa_y} (a_{\rm min},a_{\rm min})}$. The condition for being close to $1:1$ resonance (i.e. equal eigenvalues) is therefore simply $ \frac{ \partial^2V}{\partial a_xa_y} (a_{\rm min},a_{\rm min}) \approx 0 $. For these systems we study orbits with initial conditions as given in (\[ic0\]). The process of approximating such a system with an integrable system has $3$ steps.
The [*first step*]{} is to expand in normal coordinates near the fixed point, retaining only terms up to order $4$ in the potential. Thus we write $$a_x = a_{\rm min} + \frac{\zeta_2+\zeta_1}{\sqrt{2}} \ , \qquad a_y = a_{\rm min} + \frac{\zeta_2-\zeta_1}{\sqrt{2}}$$ and expand to fourth order to obtain $$H_1 = \frac12 \left( p_1^2 + p_2^2 + \omega_1^2 \zeta_1^2 + \omega_2^2 \zeta_2^2 \right) + a_1 \zeta_1^2 \zeta_2 + a_2 \zeta_2^3
+ a_3 \zeta_1^4 + a_4 \zeta_1^2 \zeta_2^2 + a_5 \zeta_2^4
\label{H1}$$ where $p_1,p_2$ are the conjugate momenta to the coordinates $\zeta_1,\zeta_2$, and $\omega_1,\omega_2,a_1,a_2,a_3,a_4,a_5$ are constants that depend on the parameters of the original potential $V$. The Hamiltonian $H_1$ has $\zeta_1 \rightarrow -\zeta_1$ symmetry as a consequence of the symmetry of $H$, and is the general Hamiltonian with this symmetry and a quartic potential. Aspects of the behavior of this Hamiltonian at, or close to, $1:1$ resonance have been studied previously, for example, in [@Verhulst1; @montaldi1; @montaldi2; @pm3; @pm1; @pm2]. The initial conditions for this system, corresponding to (\[ic0\]) are $$\zeta_1(0) = \frac{a_0}{\sqrt{2}} \left(r-\frac1{r} \right)
\ , \qquad \zeta_2(0) = \frac1{\sqrt{2}} \left(a_0\left(r+\frac1{r} \right) - 2 a_{\rm min} \right)
\ , \qquad
{p_1}(0) = {p_2}(0) = 0 \ . \label{ic1}$$ Symmetric solutions correspond to the initial condition $\zeta_1(0)=0$. In regarding $H_1$ as an approximation for $H$ we are neglecting terms of fifth order and above.
The [*second step*]{} is to make the canonical transformation to action-angle coordinates associated with the quadratic part of the Hamiltonian $H_1$, i.e. to substitute $$\begin{array}{ll}
\zeta_1 = \sqrt{\frac{2J_1}{\omega_1}} \cos\theta_1\ , & p_1 = - \sqrt{2J_1\omega_1} \sin\theta_1 \ , \\
\zeta_2 = \sqrt{\frac{2J_2}{\omega_2}} \cos\theta_2\ , & p_2 = - \sqrt{2J_2\omega_2} \sin\theta_2 \ .
\end{array}$$ This gives $$\begin{aligned}
H_2 &=& \omega_1 J_1 + \omega_2 J_2
+ \left( \frac{2J_1a_1}{\omega_1} +\frac{3 J_2 a_2}{\omega_2} \right) \sqrt{\frac{J_2}{2\omega_2}} \cos\theta_2 \nonumber\\
&&
+ \left(\frac{J_2}{2\omega_2}\right)^{3/2} 2a_2\cos 3\theta_2
+ a_1 \frac{J_1}{\omega_1} \sqrt{\frac{J_2}{2\omega_2}}
\left( \cos(2\theta_1-\theta_2) + \cos(2\theta_1+\theta_2) \right) \nonumber\\
&& + 3a_3 \frac{J_1^2}{2\omega_1^2}
+ a_4 \frac{J_1J_2}{\omega_1\omega_2}
+ 3a_5 \frac{J_2^2}{2\omega_2^2}
+ a_4 \frac{J_1J_2}{2\omega_1\omega_2}\left( \cos(2\theta_1-2\theta_2) + \cos(2\theta_1+2\theta_2) \right)
\nonumber \\
&& \left( 2a_3 \frac{J_1^2}{\omega_1^2} + a_4 \frac{J_1J_2}{\omega_1\omega_2} \right)\cos(2\theta_1)
+ a_3 \frac{J_1^2}{2\omega_1^2} \cos(4\theta_1)
+\left(a_4\frac{J_1J_2}{\omega_1\omega_2}+ 2a_5\frac{J_2^2}{\omega_2^2} \right)\cos(2\theta_2) \nonumber\\
&& +a_5\frac{J_2^2}{2\omega_2^2}\cos(4\theta_2) \ . \label{H2} \end{aligned}$$ Here $\theta_1,\theta_2$ are the angle variables, and $J_1,J_2$ the conjugate actions. The initial conditions for the action variables are $$J_1(0) = \frac{a_0^2\omega_1}{4} \left(r-\frac1{r} \right)^2 \ , \qquad
J_2(0) = \frac{\omega_2}{4} \left(a_0\left(r+\frac1{r} \right) - 2 a_{\rm min} \right)^2\ .
\label{Jic}$$ The initial conditions for the angle variables depend on the sign of $\zeta_1(0)$ and $\zeta_2(0)$. If $\zeta_1(0)>0$ ($\zeta_2(0)>0$) then, from (\[ic1\]) we should take $\theta_1(0)=0$ ($\theta_2(0)=0$) and otherwise $\theta_1(0)=\pi$ ($\theta_2(0)=\pi$). Due to the $\zeta_1\rightarrow -\zeta_1$ symmetry of $H_1$ the Hamiltonian $H_2$ has period $\pi$ (and not $2\pi$) as a function of $\theta_1$ and thus the choice of the $\theta_1$ initial condition is irrelevant. The choice of the $\theta_2$ initial condition, however, is important. We are introducing a non-physical discontinuity in the approximation procedure when the sign of $\zeta_2(0)$ changes, i.e. when $r+\frac1{r}=\frac{2a_{\rm min}}{a_0}$. We will see the effects of this later, in our results for the SAT potential.
The [*third step*]{} involves a canonical change of coordinates $(\theta_1,\theta_2,J_1,J_2)\rightarrow (\phi_1,\phi_2,K_1,K_2)$ defined by a generating function of the second type $G_2(\theta_1,\theta_2,K_1,K_2)$ [@Goldstein], chosen to eliminate the nonresonant terms from the Hamiltonian (i.e. all the trigonometric terms of order $||J||^{3/2}$ or $||J||^2$ except the one involving $\cos(2\theta_1-2\theta_2)$.) The full change of coordinates is given by $$\begin{array}{ll}
\phi_1 = {\displaystyle{\frac{\partial G_2}{\partial K_1}}} \ , &
J_1 = {\displaystyle{\frac{\partial G_2}{\partial \theta_1}}} \ , \\[\bigskipamount]
\phi_2 = {\displaystyle{\frac{\partial G_2}{\partial K_2}}} \ , &
J_2 = {\displaystyle{\frac{\partial G_2}{\partial \theta_2}}} \ .
\end{array}
\label{ct}$$ The generating function $G_2$ should be taken in the form $$\begin{aligned}
G_2 &=& K_1\theta_1 + K_2\theta_2 + A_1 \sin\theta_2 + A_2\sin 3\theta_2 + A_3 \sin(2\theta_1-\theta_2)
+ A_4 \sin(2\theta_1+\theta_2) \\
&& + A_5 \sin2\theta_1 + A_6 \sin 4\theta_1 + A_7 \sin2\theta_2 + A_8 \sin 4\theta_2
+ A_9 \sin6\theta_2 + A_{10} \sin (2\theta_1 +2\theta_2)\\
&&
+ A_{11} \sin (2\theta_1 +4\theta_2)
+ A_{12} \sin (4\theta_1 +2\theta_2)
+ A_{13} \sin (2\theta_1 -4\theta_2)
+ A_{14} \sin (4\theta_1 -2\theta_2) \end{aligned}$$ where the coefficients $A_1,\ldots,A_{14}$ are functions of $K_1,K_2$, which are chosen to eliminate the nonresonant trigonometric terms in the Hamiltonian to required order. $A_1,A_2,A_3,A_4$ are of order $||K||^{3/2}$ and $A_5,\ldots,A_{14}$ are of order $||K||^2$. The calculations are long, but straightforward with the help of a symbolic manipulator, and the final Hamiltonian is found to be simply $$H_3 = \omega_1 K_1 + \omega_2 K_2
+ b_1 K_1^2 + b_2 K_1K_2 + b_3 K_2^2
+ (b_4 K_1^2 + b_5 K_1K_2) \cos\left( 2(\phi_1-\phi_2) \right)
\label{H3}$$ where $$\begin{aligned}
b_1 &=& \frac{3a_{{3}}}{2\omega_{1}^{2}}
- \frac{a_{{1}}^{2}\left( 8\omega_{{1}}^{2}-3\omega_{{2}}^{2} \right)}
{4\omega_1^{2}\omega_2^{2} \left( 2\omega_1-\omega_2 \right) \left( 2\omega_1 + \omega_2 \right)}
\nonumber\\
b_2 &=& \frac{a_{{4}}}{\omega_{{1}}\omega_{{2}}} - \frac{3 a_1 a_2}{\omega_1\omega_2^3}
- \frac {2a_{{1}}^2}
{\omega_{{1}}\omega_{{2}} \left( 2\omega_{{1}}-\omega_{{2}} \right) \left( 2\omega_{{1}}+\omega_{{2}} \right) }
\nonumber\\
b_3 &=& \frac{3a_{{5}}}{2\omega_{{2}}^{2}} - \frac{15 a_{{2}}^{2}}{4\omega_{{2}}^{4}}
\label{bs}\\
b_4 &=& \frac { \left( \omega_{{2}}-\omega_{{1}} \right) a_{{1}}^{2} }
{2\omega_{{1}}^{2}\omega_{{2}}^{2} \left( 2\omega_{{1}}-\omega_{{2}} \right) }
\nonumber\\
b_5 &=& \frac {a_{{4}}}{2\omega_{{1}}\omega_{{2}}}
- \frac{a_1a_2(4\omega_1^2-3\omega_1\omega_2-4\omega_2^2)}
{2\omega_{{1}}\omega_{{2}}^{3} \left( 2\omega_{{1}}-\omega_{{2}} \right)\left( 2\omega_{{1}}+\omega_{{2}} \right)}
- \frac{a_1^2}{\omega_1^2\omega_2(2\omega_1-\omega_2)}
\ . \nonumber\end{aligned}$$
[*The Hamiltonian $H_3$ given in (\[H3\]) is an integrable approximation of the original Hamiltonian $H$ given in (\[H\]).*]{} $H_3$ is a [*normal form*]{} for the “natural” Hamiltonian $H_1$ at or near $1:1$ resonance. Note that in the case of exact resonance $\omega_1=\omega_2$ the coefficient $b_4$ vanishes. Also close to resonance, the corresponding term in $H_3$ is of lower order than the other terms, and in [@pm3; @pm1; @pm2] it is omitted. However, we choose to retain it to avoid any assumption on the relative orders of magnitude of $|\omega_1 - \omega_2|$ and $||K||$. The integrability of $H_3$ is evident, as it only depends on the modified angle variables $\phi_1,\phi_2$ through the combination $\phi_1-\phi_2$. As a consequence the quantity $K_1+K_2$ is conserved, in addition to the Hamiltonian itself. We denote the value of the Hamiltonian by ${\cal E}$ and the value of $K_1+K_2$ by $P$ (these should be computed from the system parameters and initial conditions). The full equations of motion are $$\begin{aligned}
\dot{\phi}_1 = \frac{\partial H_3}{\partial K_1}
&=& \omega_1 + 2 b_1 K_1 + b_2 K_2 + (2 b_4 K_1 + b_5 K_2) \cos(2(\phi_1-\phi_2)) \ , \label{fe1}\\
\dot{\phi}_2 = \frac{\partial H_3}{\partial K_2}
&=& \omega_2 + b_2 K_1 + 2 b_3 K_2 + b_5 K_1 \cos(2(\phi_1-\phi_2)) \ , \label{fe2}\\
\dot{K}_1 = -\frac{\partial H_3}{\partial \phi_1}
&=& 2 K_1 (b_4 K_1 + b_5 K_2) \sin (2(\phi_1-\phi_2)) \ , \label{fe3}\\
\dot{K}_2 = -\frac{\partial H_3}{\partial \phi_2}
&=& -2 K_1 (b_4 K_1 + b_5 K_2) \sin (2(\phi_1-\phi_2)) \ . \label{fe4}\end{aligned}$$ Using the two conservation laws it is possible to eliminate $K_2$ and $\phi_1-\phi_2$ from the $K_1$ equation of motion to get a single equation for $K_1$: $$\begin{aligned}
\dot{K}_1^2 &=& - 4
((b_1-b_2+b_3+b_4-b_5)K_1^2+((b_2-2b_3+b_5)P-\omega_2+\omega_1)K_1+b_3P^2+\omega_2P-{\cal E}) \nonumber \\
&&
((b_1-b_2+b_3-b_4+b_5)K_1^2+((b_2-2b_3-b_5)P-\omega_2+\omega_1)K_1+b_3P^2+\omega_2P-{\cal E})\ .\nonumber \\
&& \label{K1e}\end{aligned}$$
Equation (\[K1e\]) is a central result of this paper. To solve (\[K1e\]) it is necessary to translate the initial conditions for $J_1,J_2,\theta_1,\theta_2$ into initial conditions for $K_1,K_2$. This step requires details of the canonical tranformation. Due to their length, the full equations determining the initial values of $K_1,K_2$ are given in Appendix A (equations (\[Keq1\])-(\[Keq2\])). Note there are two cases depending on whether $\theta_2(0)$ is $0$ or $\pi$. Note also that [*there is no guarantee that these equations will have a solution with real, positive $K_1,K_2$*]{}. In the case of the SAT system, for a certain range of parameter values we have experienced numerical problems with the solution of (\[Keq1\])-(\[Keq2\]), specifically for initial values of $J_2$ close to zero, close to the jump from $\theta_2=0$ to $\theta_2=\pi$. However, typically there are values of $K_1(0),K_2(0)$ close to the given values of $J_1(0),J_2(0)$.
Once the initial values of $K_1,K_2$ have been computed, the values of the constants ${\cal E}$ and $P$ can be found and equation (\[K1e\]) can be solved. The right hand side of (\[K1e\]) is the product of two quadratic factors in $K_1$, with up to $4$ real roots, and typical solutions will be oscillatory between two roots. When there is a [*double root*]{} then there is the possibility of the period of the oscillation becoming [*infinite*]{}, marking a bifurcation in the solution. There are two ways that a double root can occur, by the vanishing of the discriminant of one of the quadratic factors, or by one of the roots of the first factor coinciding with one of the roots of the second. The discriminants of the quadratic factors are $$\begin{aligned}
\Delta_1 &=&
\left( (b_2+b_5)^2 - 4b_3 (b_1 + b_4) \right) P^2
+ 2 \left( (b_2 - 2 b_3 + b_5) \omega_1 + (-2b_1 + b_2 - 2 b_4 + b_5) \omega_2 \right) P \nonumber \\
&& + 4 (b_1 - b_2 + b_3 + b_4 - b_5){\cal E} + (\omega_1 - \omega_2) ^ 2 \ , \label{Del1} \\
\Delta_2 &=&
\left( (b_2-b_5)^2 -4b_3 (b_1 - b_4) \right) P^2
+ 2 \left( (b_2 - 2 b_3 - b_5) \omega_1 + (-2b_1 + b_2 + 2 b_4 - b_5) \omega_2 \right) P \nonumber \\
&& + 4 (b_1 - b_2 + b_3 - b_4 + b_5){\cal E} + (\omega_1 - \omega_2) ^ 2 \ . \label{Del2}\end{aligned}$$ A simple algebraic manipulation shows that the first factor and second factor have coincident roots if either $\Delta_3=0$ or $\Delta_4=0$, where $$\begin{aligned}
\Delta_3 &=& b_3 P^2 + \omega_2 P - {\cal E}\ , \label{Del3} \\
\Delta_4 &=& \frac{ b_1 b_5^2 - b_2 b_4 b_5 + b_3 b_4^2}{(b_4-b_5)^2} P^2 +
\frac{b_4\omega_2-b_5\omega_1}{b_4-b_5} P - {\cal E} \ . \label{Del4}\end{aligned}$$ From (\[K1e\]), we see that the first case occurs when the repeated root is at $K_1=0$.
It should be emphasized that the occurence of a double root on the RHS of (\[K1e\]) is a [*necessary*]{} condition for a bifurcation of the solution (giving rise to a transition between types) but not a [*sufficient*]{} condition. For example, if the solution is describing an oscillation on the interval between two adjacent roots of the RHS, and the two other roots outside this interval merge, this will have no effect on the solution. We illustrate, in Figure 4, with two concrete examples of equation (\[K1e\]) emerging from the CQ system described in Section 2. In both cases $Qa_0^{-2}=0.077$ and $ga_0^4=0$; in the first case $r=1.01$ and in the second case $r=1.045$. In both cases we plot the roots of the RHS as a function of the single remaining parameter $E$. (The choice to plot the roots for fixed values of $Qa_0^{-2}$, $ga_0^4$ and $r$ and to vary $E$ is just an illustration; we could just as easilly vary any of the other parameters or a combination thereof.) In the first case there are $4$ points $P_1,P_2,P_3,P_4$ at which there are double roots; however, transitions only occur at the two points $P_1,P_4$ (marked in Figure 4 with large dots). In the second case there are $5$ points $P_1,P_2,P_3,P_4,P_5$ at which there are double roots; however, transitions only occur at the two points $P_2,P_4$. In both cases, the first transition is from type II to type I, and the second transition is from type I to type II, as indicated by Roman numerals on the plot.
The theoretical explanation of this is as follows. In the first case, $r=1.01$, there are $4$ values of $E$ for which there is a double root. The points labelled $P_3$ and $P_4$ on the diagram are associated with the vanishing of the discriminant $\Delta_2$; the point labelled $P_1$ is a double root at $0$, associated with the condition $\Delta_3=0$, and the point labelled $P_2$ is associated with the vanishing of the discriminant $\Delta_1$. The motion takes place between the root that is at $K_1\approx 0.00015$ and an adjacent root: for values of $E$ below $P_2$ the adjacent root is below, for values of $E$ above $P_2$ the adjacent root is above. Thus the double root at $P_1$ indicates a value of $E$ for which there is a bifurcation, and the period of oscillation diverges. The double root at $P_2$ is a special solution for which $K_1$ and $K_2$ are constant (looking at (\[fe3\])-(\[fe4\]) it can be seen that there are $3$ kinds of solution of this type, each corresponding to vanishing of one of the three factors on the RHS of this equation; these are related to the [*nonlinear normal modes*]{} of the system [@mikhbar; @pm1; @pm2]). The point $P_2$ does not, however, give rise to a transition in behavior of the CQ system; the beating period diverges there, but the type does not change. The double root at $P_3$ also does not mark a transition. This is precisely the case described above, in which the oscillation is on the interval between 2 roots, and the other two roots outside this interval merge. The point $P_4$, however, does mark a second transition, from type I beating back to type II.
Proceeding to the second example in Figure 4, there are now $5$ cases of a double root. $P_2$ and $P_4$ are associated with the vanishing of the discriminant $\Delta_2$, $P_3$ with the vanishing of the discriminant $\Delta_1$. $P_1$ is the case of a double root at zero associated with the condition $\Delta_3=0$, and $P_5$ is associated with the final possibility, $\Delta_4=0$. There are however only $2$ transitions, associated with the points $P_2$ and $P_4$, for similar reasons to the case described in the previous paragraph.
In this section we have explained how small oscillations of the original Hamiltonian (\[H\]) near its fixed point and near (symmetric) $1:1$ resonance can be approximated using the integrable Hamiltonian (\[H3\]) and the single differential equation (\[K1e\]). We have arrived at a simple analytic approximation for beating transitions, viz. [*a necessary condition for a transition between type I and type II beating is the vanishing of one of the four quantities $\Delta_1,\Delta_2,\Delta_3,\Delta_4$ given in (\[Del1\]),(\[Del2\]),(\[Del3\]),(\[Del4\]).*]{} It should be emphasized that this far from trivializes the original problem. There is substantial complexity hidden in the relationship between parameters and initial conditions of the original Hamiltonian and those of the integrable Hamiltonian. Also determining which of the vanishing conditions gives a physical transition can be subtle. In Section 4 we apply the approximation to the CQ and SAT models from Section 2 and validate its predictions against numerical results.
Application to the models
=========================
The CQ Model
------------
The potential of the CQ model, given by (\[VCQ\]), has an isolated minimum when $$a_x = a_y = a_{\rm min} \equiv \frac{4 E\sqrt{2Q}}{3\sqrt{E-1}} C_0$$ where $C_0>0$ is a solution of the equation $$1 - C_0^2 = \frac{1024E^4Q^2g}{81(E-1)^3} C_0^6 \ .$$ The $1:1$ resonance condition is $E= E_{\rm res}$ where $$% E_{\rm res}C_0^2 = 2(E_{\rm res}-1) \quad {\rm or} \quad E_{\rm res} = 2 - \frac{8192 E_{\rm res}^2Q^2g}{81}\ .
E_{\rm res} = \frac{4}{1 + \sqrt{1 + \frac{65536 Q^2g}{81}}} \label{rescon}$$ In the case of zero grade index, $g=0$, we have $C_0=1$ and the resonance condition is simply $E=2$. The model is valid if the parameters $E,Q,g$ are chosen so that $E\approx E_{\rm res}$ and the initial conditions (see (\[ic0\])) satisfy $a_0\approx a_{\rm min}$ and $r\approx 1$.
The relevant parameters for the quartic Hamiltonian (\[H1\]) are $$\begin{aligned}
\omega_1^2 &=& \frac{81}{512}\ \frac{(E-1)^2((2-E)C_0^2+E-1)}{E^4Q^2C_0^6} \nonumber\\
\omega_2^2 &=& \frac{81}{512}\ \frac{(E-1)^3(3-2C_0^2)}{E^4Q^2C_0^6} \nonumber\\
a_1 &=& \frac{243}{8192}\
\frac{ (E-1) ^{5/2} \left( 2(E-3) C_0^{2}-3(E-1) \right)}{ {Q}^{5/2}C_0^{7}{E}^{5} } \nonumber \\
a_2 &=& \frac{243}{8192}\
\frac { ( E-1) ^{7/2} (2C_0^{2}-5) }{ {Q}^{5/2}C_0^{7}{E}^{5}} \label{findas} \\
a_3 &=& \frac{729}{262144}\
\frac { (E-1) ^{3} ( 2(5-E)C_0^{2}+3(E-1) ) }{ C_0^{8}{E}^{6}{Q}^{3}} \nonumber\\
a_4 &=& \frac {729}{131072}\
\frac{(E-1)^3\left( 10(3-E)C_0^{2}+ 21(E-1) \right) }{C_0^{8}{E}^{6} {Q}^{3} } \nonumber\\
a_5 &=& \frac {3645}{262144}\
\frac{(E-1)^4( 7-2 C_0^{2} )}{C_0^{8}{E}^{6}{Q}^{3}} \nonumber\end{aligned}$$
The detailed recipe for checking whether a given set of parameters and initial conditions $E,Q,g,a_0,r$ might give rise to a transition is as follows:
1. Compute the coefficients $\omega_1^2,\omega_2^2,a_1,a_2,a_3,a_4,a_5$ using (\[findas\]). This is the only stage of the recipe that is model dependent. Compute the coefficients $b_1,b_2,b_3,b_4,b_5$ from (\[bs\]).
2. Compute the initial conditions $J_1(0),J_2(0)$ from (\[Jic\]) and $\theta_1(0),\theta_2(0)$ from the comments following (\[Jic\]). In the case of CQ, all the parameter values which we used gave $\theta_1(0)=0$ (we took $r>1$ throughout) and $\theta_2(0)=\pi$.
3. Compute the initial conditions $K_1(0),K_2(0)$ using (\[Keq1\])-(\[Keq2\]). This is the only stage of the recipe that is not completely explicit, and involves solving two equations in two variables. If no real solution can be found, the method fails. A suitable initial guess for the solution is $K_1(0)\approx J_1(0)$ and $K_2(0)\approx J_2(0)$.
4. Determine the value of ${\cal E}$, the constant value of the Hamiltonian $H_3$ using (\[H3\]), taking $\cos(2(\phi_1-\phi_2))=1$. Determine the value of $P=K_1+K_2$.
5. Compute $\Delta_1,\Delta_2,\Delta_3,\Delta_4$ from (\[Del1\]),(\[Del2\]),(\[Del3\]),(\[Del4\]). Values of $E,Q,g,a_0,r$ for which any of these quantities vanish are candidates for transitions.
Figure 5 displays results. Figure 5a shows numeric values and candidate analytic approximations of $r_c$ as a function of $E$ for $Qa_0^{-2}=0.077$ and $ga_0^4=0.01$. The dots denote numeric values of transitions in the original system. The solid curves show candidate analytic approximations of 3 distinct types: (1) (black) values for which $\Delta_2=0$ (a closed loop with a cusp on the axis at $r=1$), (2) (green) values for which $\Delta_3=0$ (a simple open curve) and (3) (red) values for which $\Delta_4=0$ (two crossing open curves). For the values of $Qa_0^{-2}$ and $ga_0^4$ specified, it seems there are two branches of parameter values for which there are transitions. We denote the lower branch (on the plot) by $r_{c,1}(E)$, which exists for $E$ greater than a certain value which we denote by $E_{c,1}$, and the upper branch by $r_{c,2}(E)$, which exists for $E$ greater than a certain value which we denote by $E_{c,2}$, with $E_{c,2} \approx 1.975 < E_{c,1} \approx 1.977$. On the lower branch, as $E$ increases from $E_{c,1}$, $r_{c,1}(E)$ at first follows the approximation $\Delta_2=0$, until a triple point at which the curves $\Delta_2=0$ and $\Delta_4=0$ intersect. As $E$ increases further, $r_{c,1}(E)$ follows the approximation $\Delta_4=0$. Surprisingly, this approximation stays reasonably accurate for the full range shown on the figure, even though $r_{c,1}(E)$ rises to approximately $1.12$. On the upper branch, as $E$ increases from $E_{c,2}$, $r_{c,2}(E)$ at first follows the approximation $\Delta_3=0$, until a triple point at which the curves $\Delta_2=0$ and $\Delta_3=0$ intersect. As $E$ increases further, $r_{c,2}(E)$ follows the approximation $\Delta_2=0$. However, the quality of this approximation rapidly decreases as $E$ and $r_{c,2}(E)$ increase further, with the discrepancy already visible on the plot for $r_c\approx 1.06$.
Figure 5b shows numeric values and the [*correct*]{} analytic approximation (made up of pieces of the curves $\Delta_2=0$, $\Delta_3=0$ and $\Delta_4=0$) in the cases (1) $ga_0^4 = 0$, (2) $ga_0^4 = 0.01$, (3) $ga_0^4 = 0.02$, all for $Qa_0^{-2}=0.077$. In addition, stars indicate numerical values of transitions obtained for the quartic system with Hamiltonian (\[H1\]). For small values of $r$, the numerical values for transitions for the exact Hamiltonian and the approximate quartic Hamiltonian (\[H1\]) are, as we would expect, very close. However as $r$ increases, we see that the results for the quartic Hamiltonian rapidly diverge from the results for the exact Hamiltonian, while, remarkably, the analytic approximation continues to be a reasonable approximation for the exact Hamiltonian. This may find an explanation in the fact that while the exact Hamitlonian (\[H\]), the quartic approximation (\[H1\]) and the integrable approximation (\[H3\]) all agree close to the fixed point, the global properties of the exact Hamiltonian are expected to be closer to those of the integrable approximation than the quartic approximation.
Note also in Figure 5b that the intercepts of the curves on the $E$ axis, that we have denoted above by $E_{c,1}$ and $E_{c,2}$, are very close to the values of $E$ determined by the resonance condition (\[rescon\]), which are (1) $E=2$, (2) $E \approx 1.977$, (3) $E\approx 1.954$. However, even though the intercepts for the two curves obtained for each set of parameter values are very close, they are not identical. This is something that is difficult to establish [*a priori*]{} by direct numerics for the original systems (as the beating periods, for values of $r$ close to $1$, are very long), but once the analytic approximation is available to give accurate candidate values for the transition locations, it is possible to verify them [*a posteriori*]{}. Thus [*in the small band of values $E_{c,2}<E<E_{c,1}$ there is only a single beating transition as the beam eccentricity is increased.*]{} As $r$ is increased from $1$ there is immediately type I beating, and as $r$ is increased further there is only a single transition to type II (as opposed, for example, to the sitution in Figure 2, where as $r$ is changed from $1$ type II beating is seen, and then there are two transitions). Using the analytic approximation it can be shown (see Appendix A) that the points $E_{c,1}, E_{c,2}$ are determined by the conditions $$\omega_1 - \omega_2 + P(b_2-2b_3 \mp b_5) = 0 \label{Ecconds}$$ (minus for $E_{c,1}$, plus for $E_{c,2}$) for a solution with $r=1$. (The condition $r=1$ implies $J_1(0) = K_1(0)=0$, and then equation (\[Keq2\]) gives a single equation from which to determine $K_2(0)$ from $J_2(0) =
\omega_2 \left(a_0 - a_{\rm min} \right)^2$.) Figure 6 shows the dependence of $E_{c,1}$ and $E_{c,2}$ on $ga_0^4$ for two values of $Qa_0^{-2}$, as computed by the analytic approximation, along with a few numeric values (computed [*a posteriori*]{}). In addition the value of $E_{\rm res}$ from (\[rescon\]) is shown, this being the value of $E$ for which there is exact $1:1$ resonance in the linear approximation. We see that the values of $E_{c,1}$, $E_{c,2}$ and $E_{\rm res}$ all decrease monotonically with $ga_0^4$.
The SAT Model
-------------
The potential of the SAT model, given by (\[VSAT\]), has an isolated minimum when $$a_x = a_y = a_{\rm min} \equiv a_0 \sqrt{\frac{K_{\rm SAT}}{K_0}}$$ where $K_0$, which depends on the parameters $E,K_{\rm SAT}, ga_0^4$, is a solution of the equation $$\frac{K_0^2}{4E} + {\rm Li}_2 ( -K_0) + \ln(1+K_0) - \frac{K_{\rm SAT}^2ga_0^4}{4E}
= 0 \ . \label{Hdef}$$ (Recall that the constant $K_{\rm SAT}$ is defined by $K_{\rm SAT}=4\alpha^2Ea_0^{-2}$.) The resonance condition can be written $K_0 = K_{\rm res}$ where $K_{\rm res}$ is the solution of $${\rm Li}_2 ( -K_{\rm res}) + 2\ln(1+K_{\rm res}) - \frac{K_{\rm res}}{1+K_{\rm res}} = 0 \ .$$ $K_{\rm res}$ has numerical value approximately $5.017$. We recall that for our analytic model to be most effective we need to be near resonance, and the initial conditions should be close to the minimum, i.e. $a_0\approx a_{\rm min}$, or $K_{\rm SAT}\approx K_0$, and $r\approx 1$. These conditions give $K_{\rm SAT} \approx K_{\rm res} = 5.017$ and $E \approx 6.550( 1 - ga_0^4)$. In practice we will look at a large range of values of $E$ and $K_{\rm SAT}$, but focus on this region. We also recall that in our model the sign of $\zeta_2(0)$ (as given in (\[ic1\]) plays a critical role. From (\[Hdef\]) we have $\zeta_2(0)=0$ (or equivalently $K_{\rm SAT}=K_0$) when [@Ianetz2013PRA] $$E = \frac{ -K_{\rm SAT}^2(1-ga_0^4)} { 4( {\rm Li}_2(-K_{\rm SAT})+\ln(1+K_{\rm SAT})) } \ . \label{z20sign}$$
The relevant parameters for the quartic Hamiltonian (\[H1\]) in the SAT case are $$\begin{aligned}
\omega_1^2 &=& \frac{2}{a_0^4}\left( \frac{K_0^2}{K_{\rm SAT}^2} + ga_0^4 \right)
\nonumber\\
\omega_2^2 &=& \frac{4}{a_0^4K_{\rm SAT}^2}\left( -2E\ln(1+K_0) + K_0^2 + \frac{2 E K_0}{1+K_0} \right) \nonumber\\
a_1 &=& \frac{\sqrt{2K_0}}{K_{\rm SAT}^{5/2}a_0^5} \left( 2E\ln(1+K_0) - 3K_0^2-\frac{2EK_0}{1+K_0} \right) \nonumber\\
a_2 &=& \frac{\sqrt{2K_0}}{3K_{\rm SAT}^{5/2}a_0^5} \left(-2E\ln(1+K_0) -3K_0^2 + \frac{2EK_0(1+3K_0)}{(1+K_0)^2} \right)
\label{asforSAT} \\
a_3 &=& \frac{K_0}{4K_{\rm SAT}^3a_0^6} \left( -2E\ln(1+K_0) + 5K_0^2 + \frac{2EK_0}{1+K_0} \right)
\nonumber\\
a_4 &=& \frac{K_0}{2K_{\rm SAT}^3a_0^6} \left( -2E\ln(1+K_0) + 15K_0^2 + \frac{2EK_0(1-K_0)}{(1+K_0)^2} \right)
\nonumber\\
a_5 &=& \frac{K_0}{12K_{\rm SAT}^3a_0^6} \left( 2E\ln(1+K_0) + 15K_0^2 - \frac{2EK_0(1+10K_0 + K_0^2)}{(1+K_0)^3} \right) \ .
\nonumber \end{aligned}$$
The method is identical to that given for CQ in the previous subsection, so we can immediately present results. For fixed values of $E$ and $ga_0^4$ we look for values of $r$ giving beating transitions as a function of $K_{\rm SAT}$. Both numeric and analytic results suggest there is a qualitative difference in behavior for $E$ above and below a critical threshold, and our results are consitent with the value of this threshold being approximately $6.550( 1 - ga_0^4)$, as found above. Figure 7 displays results for $ga_0^4=0$ and $E=6.3$ (below the threshold, left) and $E=6.7$ (above the threshold, right). The numeric results show that below the threshold, there are two ranges of $K_{\rm SAT}$ for which there is a single beating transition, from type I (for $r$ below $r_c$) to type II (for $r$ above $r_c$). For values of $K_{\rm SAT}$ below or above these two ranges, there is only type I beating, and for values between the two ranges there is only type II beating. The analytic approximation reproduces these results well. In this region of parameter space there are values for which $\Delta_2=0$ (indicated in black in the figure) and $\Delta_3=0$ (indicated in green). It is the latter that are physically relevant, and the values of $r_c$ predicted by the analytic model are accurate for a good range. Moving “above the threshold”, numerics show there is a range of values of $K_{\rm SAT}$ for which, as $r$ is increased from $1$, the beating is initially type II, then there is a transition to type I. For some of these values there is then a further transition back to type II for quite high values of $r$. It should be mentioned that these latter transitions were initially discovered using the analytic approximation, and confirmed numerically [*a posteriori*]{}. The analytic approximation reproduces the first transition very well, using pieces of the $\Delta_2=0$ and $\Delta_4=0$ degeneracy curves. The upper transition is not reproduced well, which is not surprising bearing in mind the values of $r$ involved. Pieces of the $\Delta_2=0$ and $\Delta_3=0$ degeneracy curves are close to some of the results, but for a small range of values of $K_{\rm SAT}$ and $r$ the model fails as there is no solution of equations (\[Keq1\])-(\[Keq2\]). Two branches of the $\Delta_3=0$ degeneracy curve come to an abrupt end (in the plot we have connected the ends with a dashed line, which is not associated with any degeneracy). The values of parameters involved are precisely those for which $\zeta_2(0)\approx 0$.
Figure 8 enlarges upon these results for different values of $E$ and $ga_0^4$. In the 4 panels here, the upper panels (a and b) show results for values of $E$ above the threshold, and the lower panels (c and d) show results for values of $E$ below the threshold. In the left panels (a and c), $ga_0^4=0$, in the right panels (b and d), $ga_0^4=0.02$. For $ga_0^4=0$, the values $E=6.3,6.4,6.5,6.7,7.0$ are shown, the first three of which are below the threshold (in panel c), and the last two above the threshhold (in panel a). For $ga_0^4=0.02$, the values $E=6.3,6.4,6.5,6.7$ are shown, the first two of which are below the threshold (in panel d), and the last two above the threshold (in panel b). Note specifically that for $ga_0^4=0$ the case $E=6.5$ is below the threshold (approximately $6.55$), while for $ga_0^4=0.02$ it is above (as the threshold drops to approximately $6.42$). Thus (for example) for $E=6.5$, $K_{\rm SAT}=5$ and $ga_0^4=0$, no beating transitions are observed as the beam eccentricity is increased; but if the grade index is changed to $ga_0^4=0.02$, there are two beating transitions. The analytic theory fully explains this phenomenon. Indeed, for all the cases shown in Figure 8, the analytic theory is in excellent quantitative agreement with numerics for lower values of $r$, and gives reasonable qualitative predictions for higher values of $r$.
Another conclusion from Figure 8 is that for values of $E$ below the threshold, we can find two values of $K_{\rm SAT}$ that give rise to a given value of $r_c$, but for $E$ above the threshold this need not be the case; furthermore the gap in $r_c$ values increases with the given value of $E$. In Figure 9 we illustrate this phenomenon more clearly. For the case $ga_0^4=0$, we show contours in the $K_{\rm SAT},E$ plane that give rises to the values $r_c = \frac1{0.95}\approx 1.053$ (black), $r_c = \frac1{0.92}\approx 1.087$ (blue) and $r_c = \frac1{0.895}\approx 1.117$ (red). It is clear that the “gap” between the two branches of each contour increases with $r$. Note that in the upper branch of each contour there is a small section denoted by a dashed line where the analytic method fails (the dashed line is a straight line between the last two points on each side for which the method works). As expected, the regions where the method fails straddle the curve (\[z20sign\]), incidicated by a dashed turquoise curve. Note further that in many cases the analytic method works well far beyond the region in which this is expected, but there are some exceptions.
Conclusions and discussion
==========================
In this paper we have described the beating phenomena observed in the equations of motion for the beam widths obtained in a collective variable approximation to solution of the GNLSEs relevant for beams in nonlinear waveguides with cubic-quintic (CQ) and saturable (SAT) nonlinearities and a graded-index profile. We have described the different types of beating, and the transitions between them. Arguing that the origin of these phenomena is in a $1:1$ Hamiltonian resonance, we have developed an approximation scheme for small oscillations in a class of 2 degree-of-freedom Hamiltonian systems with an isolated fixed point close to $1:1$ resonance. We have shown that such oscillations can be described by an integrable Hamiltonian, or, alternatively, a single first order differential equation (\[K1e\]). Understanding the bifurcations of the system, which include the beating transitions, can be reduced to looking at the bifurcations of the roots of a pair of quadratic equations. Applying our general methodology to the specific cases of the CQ and SAT models we managed to reproduce numerical results for beating transitions over a large range of parameter values. The theory allows us to map out the regions (of parameter space and beam eccentricities) where beating transitions do and do not exist. Amongst other things, in the CQ case we identified a band of beam energies for which there is only a single beating transition (as opposed to $0$ or $2$) as the beam eccentricity is increased, and in the SAT case we explained the appearance and disappearance of transitions with changes of the grade-index.
We expect our methods to have applications to related problems in nonlinear optics, for nonlinearities other than the ones studied here, for different beams, such as super-Gaussian beams [@supergauss], and for optical bullets [@Kivshar2003book; @bullets]. We are encouraged by the fact that there is some recent experimental evidence [@skarkarobust] of breathing in optical solitons, albeit in a dissipative setting. We also hope the general theory of $1:1$ resonances that we have developed will find application in the settings of nonlinear mechanics and astronomy, as well as suitable extensions for $1:1:1$ resonances in higher dimensional systems (see for example the recent papers [@JS1; @JS2]).
Further Technical Details
=========================
As explained in section 3, the Hamiltonian (\[H3\]) is in an integrable approximation to the Hamiltonian (\[H2\]), and is obtained from (\[H2\]) via a canonical transformation and neglecting higher order terms. The only need for explicit details of the canonical transformation is to compute the initial conditions of the variables $K_1,K_2$ from the initial conditions of $J_1,J_2$ given in (\[Jic\]). The equations to be solved are $$\begin{aligned}
J_1 &=& K_{{1}}
\mp {\frac {4K_{{1}}a_{{1}}}{4\omega_{{1}}^{2}- \omega_{{2}}^{2}}\sqrt {{\frac {2K_{{2}}}{\omega_{{2}}}}}}
+ \left(
\frac{a_1^2 \left( 48\omega_1^{4}-8\omega_1^{3}\omega_2-40\omega_2^{2}\omega_1^{2}+2\omega_1\omega_2^{3}+5\omega_2^{4} \right)}
{4\omega_{{2}}^{2}\omega_{{1}}^{3} \left( 2\omega_{{1}}+ \omega_{ {2}} \right)^{2} \left( 2\omega_{{1}}-\omega_{{2}}\right) ^{2}}
-{\frac {5a_{{3}}}{{2\omega_{{1}}}^{3}}}
\right) K_1^2 \nonumber \\
&& + \left( \frac { \left( 40\omega_{{1}}^{3}+28\omega_{{1}}^{2}\omega_{{2}}-6\omega_{{1}}\omega_{{2}}^{2}
-3\omega_{{2}}^{3} \right) a_{{1}}^{2} }
{\omega_{{1}}^{2}\omega_{{2}} \left( 2\omega_{{1}}-\omega_{{2}} \right) ^{2} \left( 2\omega_{{1}}+\omega_{{2}} \right)^{2}
\left( \omega_{{1}}+\omega_{{2}} \right) }
\right. \nonumber \\
&& \left.
+\frac { \left( 12\omega_{{1}}^{3}+11\omega_{{1}}^{2}\omega_{{2}}-10\omega_{{1}}\omega_{{2}}^{2}
-6\omega_{{2}}^{3} \right) a_ {{1}}a_{{2}}}
{2\omega_{{1}}^{2}\omega_{{2}}^{3} \left( \omega_{{1}}+\omega_{{2}} \right) \left( 4\omega_{{1}}^{2}-\omega_{{2}}^{2}\right) }
- \frac {\left( 3\omega_{{1}}+2\omega_{{2}} \right) a_{{4}}}
{2\omega_{{1}}^{2}\omega_{{2}} \left( \omega_{{1}}+\omega_{{2}} \right) } \right) K_1K_2 \ , \label{Keq1}\\
J_2 &=& K_{{2}}
\mp 2 \left(
\frac { \left( 2\omega_{{1}}^{2}-\omega_{{2}}^{2} \right)K_{{1}}a_{{1}}}
{\omega_{{1}}\omega_{{2}} \left( 4\omega_{{ 1}}^{2}-\omega_{{2}}^{2} \right) }
+ \frac{K_{{2}}a_{{2}}}{\omega_{{2}}^{2}}
\right) \sqrt{\frac {2K_{{2}}}{\omega_{{2}}}}
+ \left( \frac{33a_2^2}{4\omega_2^5} -\frac {5a_5}{2\omega_{{2}}^{3}} \right) K_2^2 \nonumber \\
&&
+ \frac{\left( 16\omega_{{1}}^{4}+8\omega_{{1}}^{3}\omega_{{2}} - 12\omega_{{1}}^{2}\omega_{{2}}^{2}
-2\omega_{{1}}\omega_{{2}}^{3}+3\omega_{{2}}^{4} \right) {a_{{1}}}^{2}{K_{{1}}}^{2}}
{2{\omega_{{1}}}^{2}\omega_{{2}}^{3} \left( 2\omega_{{1}}-\omega_{{2}} \right)^{2}\left(2\omega_{{1}}+\omega_{{2}}\right)^{2} }
\nonumber\\
&&
+ \left(
\frac {\left( 8\omega_{{1}}^{4}+16\omega_{{1}}^{3}\omega_{{2}}-10\omega_{{1}}^{2}\omega_{{2}}^{2}
-8\omega_{{1}}\omega_{{2}}^{3}+\omega_{{2}}^{4} \right) {a_{{1}}}^{2}}
{\omega_{{1}}^{2}\omega_{{2}}^{2} \left( 2\omega_{{1}}-\omega_{{2}} \right)^{2}
\left( 2\omega_{{1}}+2\omega_{{2}} \right) ^{2} \left( \omega_{{1}}+ \omega_{{2}} \right) }
\right. \nonumber\\
&& \left.
+\frac { \left( 40\omega_{{1}}^{3}+44\omega_{{1}}^{2}\omega_{{2}}-9\omega_{{1}}\omega_{{2}}^{2}-16\omega_{{2}}^{3} \right) a_{{1}}a_{{2}}}
{2\omega_{{1}}\omega_{{2}}^{4} \left( \omega_{{1}}+\omega_{{2}} \right)
\left( 2\omega_{{1}}-\omega_{{2}} \right)\left( 2\omega_{{1}}+\omega_{{2}} \right)
}
-\frac { \left( 2\omega_{{1}}+3\omega_{{2}} \right) a_{{4}}}
{2\omega_{{1}}\omega_{{2}}^{2} \left( \omega_{{1}}+\omega_{{2}} \right) }
\right) K_1K_2 \ . \label{Keq2}\end{aligned}$$ Here the upper signs should be taken in the square roots terms in the case $\theta_2(0)=0$ and the lower signs in the case $\theta_2(0)=\pi$.
In Section 4.1, in the study of the CQ system, we stated the conditions (\[Ecconds\]) for the value $r_c$ giving a beating transition to tend to $1$. We briefly describe the origin of these conditions. The symmetric solutions with $a_x=a_y$ of (\[em\]), arising from the initial condition $r=1$, correspond to solutions with $K_1\equiv 0$ of (\[fe1\])-(\[fe4\]). From (\[K1e\]), the values of $P$ and ${\cal E}$ for such a solution must evidently satisfy $b_3P^2 + \omega_2 P -{\cal E}=0$, which is just the condition $\Delta_3=0$, see (\[Del3\]). As explained in Section 3, a necessary condition for a beating transition is the vanishing of one of the quantites $\Delta_1,\Delta_2,\Delta_3,\Delta_4$. To determine $E_{c,1}$ in Section 4.1 we want $r_c\rightarrow 1$ for a solution of $\Delta_{2}=0$. Clearly this requires $\Delta_2=\Delta_3=0$, and some simple algebra then gives the condition $\omega_1-\omega_2 + P(b_2-2b_3-b_5)=0 $. To determine $E_{c,2}$, however, is not so straightforward, as for this we want we want $r_c\rightarrow 1$ for a solution of $\Delta_{3}=0$, and apparently we do not have two equations. The resolution of this conundrum is as follows: Although we stated above that the symmetric solutions of (\[em\]) correspond to solutions with $K_1\equiv 0$ of (\[fe1\])-(\[fe4\]), the latter in fact provide a [*blow up*]{} of the former — there is a $3$ parameter family of the latter and only a $2$ parameter family of the former. Solving (\[fe1\])-(\[fe4\]) in the case $K_1\equiv 0$, we obtain $K_2=P$ (constant), $\phi_2 = \phi_{2}(0) + (\omega_2 + 2 b_3P)z $, and that $\phi_1$ must satisfy the ODE $$\dot{\phi}_1 = \omega_1 + b_2 P + b_5 P \cos \left( 2\left( \phi_{2}(0) + (\omega_2 + 2 b_3P)z - \phi_1 \right)\right) \ .$$ This latter equation can be solved explicitly, and for a general choice of the constant of integration will give a complicated function $\phi_1(z)$. However, for a beating transition we seek a solution that is characterized by a single frequency, i.e. we need $$\phi_1(z) = \phi_1(0) + (\omega_2 + 2 b_3 P) z$$ Substituting this in the differential equation, we obtain $$\omega_2 + 2 b_3 P = \omega_1 + b_2 P + b_5 P \cos \left( 2\left( \phi_{2}(0) - \phi_1(0) \right)\right) \ .$$ Since the initial conditions $\phi_1(0),\phi_2(0)$ take the values $0$ or $\pi$, we deduce that $\omega_1 - \omega_2 + P(b_2 - 2b_3 + b_5) = 0$, as required.
A two time expansion approach
=============================
In this appendix we outline a two time expansion approach [@kc; @ManMan] which is an alternative to the procedure based on canonical transformations given in Section 3.
We wish to look at solutions of the Hamiltonian system with Hamiltonian (\[H\]) and initial conditions (\[ic0\]). We assume the system has an isolated symmetric minimum at which the system is close to $1:1$-resonance. To apply a two time technique we need to introduce a small parameter $\epsilon$ explicitly into the equations. Our systems involve a number of system parameters, for example in the CQ case, the parameters $E,Q,g$, for which the resonance condition is (\[rescon\]). We introduce a small parameter by selecting one system parameter and writing this as its value at resonance plus a small perturbation. However, for reasons described in [@Verhulst1], the “small perturbation” here should be [*quadratic*]{} in the small parameter. Thus, for example in CQ, we have to consider two possibilities, $ E = E_{\rm res} \pm \epsilon^2 $ where $E_{\rm res}$ (which depends on the other system parameters $Q,g$) is the value of $E$ at resonance. The two resulting expansions will differ just in signs. This is the counterpart in the two time method of the need to choose $\theta_2(0)$ to be $0$ or $\pi$ in Section 3 and the resulting choice of signs in equations (\[Keq1\])-(\[Keq2\]). However, we emphasize that it is not the same, so the resulting method is different, in particular, the “choice” in Section 3 involves the initial conditions as well as the system parameters.
Taking, as before, the minimum of the potential $V$ to be at $a_x=a_y=a_{\rm min}$ we now write $$a_x = a_{\rm min} + \epsilon \tilde{a}_x \ , \qquad a_y = a_{\rm min} + \epsilon \tilde{a}_y$$ and expand to 4th order in $\epsilon$. The order $0$ terms are irrelevant and can be discarded. The order $1$ terms vanish by definition of $a_{\rm min}$. In the other terms there is dependence on all the system parameters. However by making the assignment of the form $E = E_{\rm res} \pm \epsilon^2$, discarding all terms of order higher than $4$ and a suitable rescaling, we obtain an approximate potential of the form $$\begin{aligned}
\tilde{V} &=& \frac12 C_1 (\tilde{a}_x^2 + \tilde{a}_y^2)
+ \epsilon \left( C_2( \tilde{a}_x^3 + \tilde{a}_y^3) + C_3\tilde{a}_x\tilde{a}_y (\tilde{a}_x + \tilde{a}_y) \right) \\
&& + \epsilon^2 \left( C_4( \tilde{a}_x^4 + \tilde{a}_y^4) + C_5\tilde{a}_x\tilde{a}_y (\tilde{a}_x^2 + \tilde{a}_y^2)
+ C_6 \tilde{a}_x^2\tilde{a}_y^2 + C_7(\tilde{a}_x^2 + \tilde{a}_y^2) + C_8 \tilde{a}_x\tilde{a}_y \right) \ . \end{aligned}$$ Here $C_1,\ldots,C_8$ are all functions of the system parameters excluding the parameter replaced by $\epsilon$. Note that as a result of the dependence of the system parameters on $\epsilon$ there are now quadratic terms in $\tilde{a}_x,\tilde{a}_y$ in the $O(\epsilon^2)$ terms.
Following the usual two time formalism, we seek solutions of the system with potential $\tilde{V}$ in the form $$\begin{aligned}
\tilde{a}_x &=& \Lambda_1(\epsilon^2 z) \cos (\sqrt{C_1}z) + \Lambda_2(\epsilon^2 z) \sin (\sqrt{C_1}z) +
\epsilon \tilde{a}_{x,1} (z,\epsilon^2 z) + \epsilon^2 \tilde{a}_{x,2} (z,\epsilon^2 z) + \ldots \\
\tilde{a}_y &=& \Lambda_3(\epsilon^2 z) \cos (\sqrt{C_1}z) + \Lambda_4(\epsilon^2 z) \sin (\sqrt{C_1}z) +
\epsilon \tilde{a}_{y,1} (z,\epsilon^2 z) + \epsilon^2 \tilde{a}_{y,2} (z,\epsilon^2 z) + \ldots\end{aligned}$$ Here $\Lambda_1(\epsilon^2 z), \Lambda_2(\epsilon^2 z), \Lambda_3(\epsilon^2 z), \Lambda_4(\epsilon^2 z)$ are functions of the slow variable $\epsilon^2 z$. Substituting in the equations of motion and equating order-by-order, the first order terms $\tilde{a}_{x,1}, \tilde{a}_{y,1}$ can be determined, and a system of first order equations is obtained that $\Lambda_1,\Lambda_2,\Lambda_3,\Lambda_4$ must satisfy to guarantee the absence of secular terms in $\tilde{a}_{x,2}, \tilde{a}_{y,2}$. Writing $$\begin{aligned}
R_1 &=& \Lambda_1^2 + \Lambda_2^2 + \Lambda_3^2 + \Lambda_4^2 \\
R_2 &=& \Lambda_1^2 + \Lambda_2^2 - \Lambda_3^2 - \Lambda_4^2 \\
R_3 &=& \Lambda_1 \Lambda_3 + \Lambda_2 \Lambda_4 \\
R_4 &=& \Lambda_1 \Lambda_4 - \Lambda_2 \Lambda_3\end{aligned}$$ (c.f. [@RPV; @pm2; @hh]) we obtain the system $$\begin{aligned}
R_1' &=& 0 \nonumber \\
R_2' &=& 4 R_4( \gamma_1 R_1 + \gamma_2 + (\gamma_3+\gamma_4)R_3) \label{Reqs} \\
R_3' &=& -\gamma_3 R_2 R_4 \nonumber \\
R_4' &=& - R_2( \gamma_1 R_1 + \gamma_2 + \gamma_4 R_3) \nonumber\end{aligned}$$ where the constants $\gamma_1,\gamma_2,\gamma_3,\gamma_4$ are certain combinations of the constants $C_1,\ldots,C_8$. (Note $R_3^2 + R_4^2 = \frac14 ( R_1^2 - R_2^2 )$.) Thus $R_1$ is an invariant, as are the quantities $$Q_2 = R_2^2 + 4\left(1 + \frac{\gamma_4}{\gamma_3} \right)
\left( R_3 + \frac{\gamma_1 R_1 + \gamma_2}{\gamma_3+\gamma_4} \right)^2
\ ,
\qquad Q_3 = R_4^2 - \frac{\gamma_4}{\gamma_3}
\left( R_3 + \frac{\gamma_1 R_1+ \gamma_2}{\gamma_4} \right)^2 \ .$$ Note that $R_1,Q_2,Q_3$ are related by $$Q_2 + 4Q_3 = R_1^2 - \frac{4(\gamma_1R_1+\gamma_2)^2}{\gamma_4(\gamma_3+\gamma_4)}\ .$$ Using the invariants it is possible to write a single differential equation for the quantity $R_3$: $$(R_3')^2 = -4 \gamma_4 \left(\gamma_3 + \gamma_4 \right)
\left( \left( R_3 + \frac{\gamma_1 R_1 + \gamma_2}{\gamma_3+\gamma_4} \right)^2
-\frac{\gamma_3}{4(\gamma_3 + \gamma_4 )} Q_2 \right)
\left( \left( R_3 + \frac{\gamma_1 R_1+ \gamma_2}{\gamma_4} \right)^2
+ \frac{\gamma_3 Q_3}{\gamma_4} \right)\ .
\label{R3eq}$$ This has the same form as (\[K1e\]) — the right hand side is a product of two quadratic factors in $R_3$ — and similar techniques can be used to discuss bifurcations of its solutions. Specifically, there can be a double root if the discriminant of one of the factors vanishes (i.e. if $Q_2$ or $Q_3$ vanish), or if the factors have a common root. The latter happens in the two cases $$\left( (2\gamma_1\pm\gamma_4)R_1+2 \gamma_2\right)^2+ 4Q_3 \gamma_3\gamma_4 = 0\ .
\label{zzzz}$$ As in Section 3, detecting beating transitions requires translating the initial conditions to the constants of motion $R_1,Q_2,Q_3$ and checking up to $4$ conditions.
We have implemented this method for the CQ and SAT systems and found some satisfactory results which we do not report here; in certain cases the results were better than those found using the method based on canonical transformations. However there are numerous reasons to prefer the method based on canonical transformations. The two time method requires deciding how to explicitly introduce a small parameter and different ways of doing this give different results. It also requires advance knowledge of the correct relative order of magnitude of the oscillations around the fixed point and the deviation of the system parameters from their resonance values. In general, the algebraic manipulations required to implement the two time method, most of which we have omitted in our account here, are substantially more complicated than those required for the method based on canonical transformations; in particular the reduction of the system (\[Reqs\]) to a single differential equation (\[R3eq\]) is a surprise, that emerges from [*ad hoc*]{} manipulations, whereas the parallel steps in the canonical formalism are standard, based on the integrability of the Hamiltonian (\[H3\]). Finally, from our numerical experiments it emerges that while the results based on the vanishing of the discriminant of one of the factors of the right hand side of (\[R3eq\]) are good, the results based on conditions (\[zzzz\]) are poor.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the dyadic model with viscosity and additive Gaussian noise as a simplified version of the stochastic Navier–Stokes equations, with the purpose of studying uniqueness and emergence of singularities. We prove path–wise uniqueness and absence of blow–up in the intermediate intensity of the non–linearity, morally corresponding to the 3D case, and blow–up for stronger intensity. Moreover, blow–up happens with probability one for regular initial data.'
address: 'Dipartimento di Matematica, Università di Pisa, Largo Bruno Pontecorvo 5, I–56127 Pisa, Italia'
author:
- Marco Romito
date: 'October 29, 2011'
title: 'Uniqueness and blow–up for the noisy viscous dyadic model'
---
Introduction
============
Motivations {#motivations .unnumbered}
-----------
Uniqueness is a problem with many facet for PDEs and different problems may require different approaches. When turning to stochastic PDEs, the problem acquires new levels of complexity, as uniqueness for stochastic processes can be understood in different ways (path–wise, in law, etc.). There are several recent result on this topic (see for instance [@Fla11] for a review).
A prototypical example of PDE without uniqueness are the Navier–Stokes equations, where the issue of uniqueness is mixed with the issue of regularity and emergence of singularities (see [@Fef06]). The analysis of the stochastic version of the equations has also received a lot of attention, and in recent years, by means of a clever way to solve the Kolmogorov equation, Da Prato and Debussche [@DapDeb03] (see also [@DebOda06]) have shown existence of Markov families of solutions to the problem, proving that such Markov families admit, under suitable assumptions on the noise, a unique invariant measure, with exponential convergence rate [@Oda07]. In [@FlaRom06; @FlaRom08] similar results have been obtained with a completely different method, based on the Krylov selection method [@Kry73]. Related results can be found in [@FlaRom07; @DapDeb08; @Fla08; @Rom08; @Rom08a; @Rom08b; @RomXu09; @Rom10a; @AlbDebXu10].
The two methods apply equally well in more general situations (as done for instance in [@BloFlaRom09]). The purpose of this paper is to analyse, with a view on uniqueness and emergence of blow–up, a much simpler infinite dimensional stochastic equation which anyway retains most of the characteristics of the original problem and which makes the methods of [@DapDeb03] and [@FlaRom08] applicable. To understand which system could provide a good candidate, the essential point is to identify the non–linearity. If we look at the Navier–Stokes non–linearity on the torus with periodic boundary conditions, we have in Fourier variables, $$(u\cdot\nabla)u
= {\mathrm{i}}\sum_{\mathbf{k}\in{\mathbf{Z}}^3,\mathbf{k}\neq0}
\Bigl(
\sum_{\mathbf{l}+\mathbf{m}=\mathbf{k},\ \mathbf{l},\mathbf{m}\neq0} (u_\mathbf{l}\cdot\mathbf{m})u_\mathbf{m}
\Bigr){\operatorname{e}}^{{\mathrm{i}}\mathbf{k}\cdot x},$$ that is, the $\mathbf{k}^\text{th}$ mode interacts with almost every other mode. The most reasonable simplification is to reduce the complexity of the interaction to a finite number of modes, while keeping the orthogonality property (which gives the conservation of energy), and the simplest is the nearest neighbour interaction. This is the dyadic model.
The dyadic model {#the-dyadic-model .unnumbered}
----------------
The dyadic model has been introduced in [@FriPav04a; @KatPav05] as a model of the interaction of the energy of an inviscid fluid among different packets of wave–modes (shells). It has been lately studied in [@KisZla05; @Wal06; @CheFriPav07; @BarFlaMor08; @BarFlaMor09b] and in the inviscid and stochastically forced case in [@BarFlaMor09a; @BrzFlaNekZeg10; @BarFlaMor10]).
The viscous version has been studied in [@FriPav04b; @Che08; @CheFri09], and in particular [@Che08] proves blow–up of *positive* solutions for the problem with non–linearity of strong intensity, and later in [@BarMorRom10] the authors prove well–posedness and convergence to the inviscid limit, again for positive solutions, with non–linearity of intensity of “Navier–Stokes” type.
In this paper we study the dyadic model with additive noise, $$\label{e:dyadic}
dX_n = \bigl( - \nu\lambda_n^2 X_n
+ \lambda_{n-1}^\beta X_{n-1}^2
- \lambda_n^\beta X_n X_{n+1}\bigr)\,dt
+ \sigma_n\,dW_n,
\qquad n\geq1,$$ where $\lambda = 2$, $\lambda_n = \lambda^n$ and $X_0\equiv0$. The dispersion coefficients satisfy suitable assumptions (see Assumption \[a:noise\]) and the parameter $\beta$ measures the relative intensity of the non–linearity with respect to the linear term.
The non–linear term *formally* satisfies the following property, $$\sum_{n=1}^\infty X_n\bigl(\lambda_{n-1}X_{n-1}^2 - \lambda_n^\beta X_nX_{n+1}\bigr)
= 0,$$ providing an a–priori bound of $X = (X_n)_{n\geq1}$ in $\ell^2({\mathbf{R}})$ which is independent of $\beta$. On the other hand the linear term and the non–linear term are roughly of the same size if $$\lambda_n^2 X_n\approx\lambda_n^\beta X_n^2
\qquad\leadsto\qquad
\lambda_n^{\beta-2}X_n\approx O(1),$$ which is the essential reason why local strong solutions exist when the initial condition decays at least as $\lambda_n^{-(\beta-2)}$ (see Theorem \[t:strong\]). If $\beta\leq 2$ this is always true due to the $\ell^2$ bound explained above, and in fact [@Che08] proves that the non–rndom problem has a unique global solution. The same method also works in the noisy case and it is straightforward to prove path–wise uniqueness in the case $\beta\leq2$.
By a scaling argument (see for instance [@Che08]), one can “morally” identify the dyadic model with the Navier–Stokes equations when $\beta\approx\tfrac52$. In [@BarMorRom10] well–posedness is proved in a range which includes the value $\tfrac52$, but only for *positive* solutions (positivity is preserved by the unforced dynamics). It is clear that, as is, positivity is broken by the random perturbation.
Main results {#main-results .unnumbered}
------------
This paper contains a thorough analysis of the case $\beta>2$, which can be roughly summarised in the table below.
$\beta\leq 2$ $2<\beta\leq 3$ $\beta>3$
------------ --------------- ----------------- -----------
blow–up NO NO${}^\star$ YES
uniqueness YES YES ?
We prove path–wise uniqueness (Theorem \[t:unique\]) in the range $\beta\in(2,3]$ by decomposing the solution in a quasi–positive component and a residual term. Quasi–positivity (see Section \[s:negative\]) is preserved by the system as long as the random perturbation is *not too strong* (the meaning of this will be explained below). Under the same conditions the residual term is small and provides a lower bound for the solution.
The quasi–positivity, together with the invariant area argument of [@BarMorRom10], implies smoothness of the solution (Theorem \[t:smooth\]), where by smoothness we mean that $(\lambda_n^\gamma X_n)_{n\geq1}$ is bounded for every $\gamma$. This result holds for $\beta\in(2,\beta_c)$, where $\beta_c\in(2,3]$ is the value idenfied in [@BarMorRom10].
When $\beta>3$ we use an idea of [@Che08], which only works for positive solutions, together with quasi–positivity, to identify a set of initial conditions which leads to blow–up with positive probability (Theorem \[t:blowup\]). While there are already cases where emergence of blow–up is proved, as for instance [@DebDeb02; @DebDeb05] for the Schrödinger equation, [@MueSow93; @Mue00; @DozLop09] for the nonlinear heat equation (there is also [@BonGro09], but their result is essentially one dimensional and no ideas for infinite dimensional systems are involved), such results essentially prove only that blow–up occurs with positive probability.
Our main result on blow–up for the dyadic model, Theorem \[t:noway\], states that blow–up occurs with full probability, as long as the initial condition satisfies $\lambda_n^\alpha X_n(0) = O(1)$ for some $\alpha>\beta-2$. This is optimal since it corresponds to the condition which ensures the existence of a local smooth solution, as remarked before. Essentially we prove that the $\lambda_n^{-(\beta-2)}$–decay is transient.
It remains open to understand uniqueness for $\beta>3$, since blow–up rules out the use of smooth solutions, making path–wise uniqueness an harder problem. Uniqueness in law may still be achievable.
Methods {#methods .unnumbered}
-------
Our results are essentially based on the four following ideas, introduced in this paper to analyse the model.
### Quasi–positivity {#quasipositivity .unnumbered}
As already mentioned, the deterministic dynamics preserves positivity and an external forcing in principle destroys this. We prove that a negative lower bound can still be proved under the condition that the random perturbation is not too big (in a sense clarified in the next idea) and the negative part of the initial condition is small. This cannot be true for general initial states, and in fact we require this to be true only for the modes $X_n(0)$ of the initial condition corresponding to $n$ large enough.
### Irrelevance of the perturbation {#irrelevance-of-the-perturbation .unnumbered}
The deterministic results of [@Che08; @BarMorRom10] are not directly applicable due to the presence of the random term, responsible of the loss of positivity. Quasi–positivity works with a small perturbation and in general the random perturbation does not stay small.
We recover smallness in two ways, equally effective: on the one hand we do not need to have a small effect of the perturbation for every component, but only for large enough modes (in the duality frequency/wavelength this would correspond to small scales). On the other hand, due to viscosity, the effect of the randomness can be quantified in terms of regularity ([i. e.]{} decay in terms of powers of $\lambda_n$) and, roughly speaking, if the effect is finite with respect to some decay, then it is small with respect to any weaker decay.
### Contraction of the negative components {#contraction-of-the-negative-components .unnumbered}
In order to prove that blow–up happens with full probability, the above ideas are not sufficient, as they only insure that it happens with positive probability. We give a stronger form of quasi–positivity (Lemma \[l:contraction\]), namely that the negative parts of the solution become smaller in a finite time, depending only on the size of the initial condition in $H$ and on the size of the random perturbation.
### Recurrence {#recurrence .unnumbered}
We use the contraction of negative components to identify an event, spanning a time interval, thus an event for trajectories, that leads to a set where blow–up occurs. An argument of recurrence for these sets finally shows that blow–up has full probability. We remark that recurrence is not at all obvious, since for $\beta>3$ the energy estimate is not strong enough to provide existence of a stationary solution with standard methods for dissipative stochastic PDEs.
Structure of the paper {#structure-of-the-paper .unnumbered}
----------------------
Section \[s:preliminary\] contains the basic definitions and assumptions, as well as the different notions of solution that will be used, together with some existence results. In Section \[s:negative\] we show the control on negative components, which the basic result on which the paper is structured. Uniqueness and well–posedness are then proved in Section \[s:unique\] for the intermediate range $2<\beta\leq3$. Section \[s:bu\_abstract\] contains some preliminary and general considerations on the blow–up time. These considerations are used in Section \[s:blowup\] to first identify events that lead to blow–up, then to prove that such sets are, conditional to the absence of blow–up, recurrent and hence that blow–up occurs with full probability.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The author wish to thank the *Institut Élie Cartan* of the *Université Henri Poincaré, Nancy I*, where part of this work has been done, for the kind hospitality and the exciting working environment. Warmful thanks to David Barbato, Franco Flandoli and Johnathan Mattingly for useful conversations and insights on the subject. This work is dedicated to Martina, she has arrived whilst the paper was being completed.
Preliminary results and definitions {#s:preliminary}
===================================
We start by stating the assumptions on the strength of the noise we shall consider.
\[a:noise\] The sequence $(\sigma_n)_{n\geq1}$ of non–negative real numbers satisfies the following assumption of regularity: there is $\alpha_0\in{\mathbf{R}}$ such that $$\label{e:noise}
\sup_{n\geq1}\bigl(\lambda_n^{\alpha_0}\sigma_n\bigr)
< \infty.$$
We shall see that, in order to ensure existence of solutions for problem , we need to impose a restriction on the possible values of $\alpha_0$, depending on the value of the parameter $\beta$.
\[a:noise2\] If $\beta$ is the parameter of problem , then the number $\alpha_0$ of Assumption \[a:noise\] above satisfies $$\label{e:noise2}
\alpha_0>\max\bigl\{\tfrac12(\beta-3), \beta-3\bigr\}$$
Notations
---------
Set $\lambda = 2$ and $\lambda_n = \lambda^n$. For every $\alpha\in{\mathbf{R}}$ denote by $V_\alpha$ the (Hilbert) space $$V_\alpha
= \{(x_n)_{n\geq1}: \sum_{n=1}^\infty (\lambda_n^\alpha x_n)^2<\infty\},$$ with scalar product $\langle\cdot,\cdot\rangle_\alpha$ and norm $\|\cdot\|_\alpha$ given by $$\langle x,y\rangle_\alpha
= \sum_{n=1}^\infty \lambda_n^{2\alpha} x_n y_n,
\qquad
\|x\|_\alpha
= \Bigl(\sum_{n=1}^\infty (\lambda_n^\alpha x_n)^2\Bigr).$$ Set in particular $H = V_0$ and $V = V_1$.
Let $\Omega_\beta = C([0,\infty);V_{-\beta})$ and define on $\Omega_\beta$ the canonical process $(\xi_t)_{t\geq0}$ defined as $\xi_t(\omega)=\omega(t)$, $t\geq0$. Define on $\Omega_\beta$ the canonical filtration $({\mathscr{B}}_t)_{t\geq0}$ where ${\mathscr{B}}_t$ is the Borel $\sigma$–field of $C([0,t];V_{-\beta})$.
Definitions of solution
-----------------------
We turn to the definition of solution. We shall consider first strong solutions, which are unique, regular but defined on a (possibly) random interval. Then we will consider weak solutions (in three different flavours).
### Strong solutions
We first state the definition of local strong solution.
\[d:strong\] Let $\mathcal{W}$ be an Hilbert sub–space of $H$. Given a probability space $(\Omega, {\mathscr{F}}, {\mathbb{P}})$ and a cylindrical Wiener process $(W_t, {\mathscr{F}}_t)_{t\geq0}$ on $H$, a *strong solution in $\mathcal{W}$* with initial condition $x\in\mathcal{W}$ is a pair $(X(\cdot;x),\tau_x^\mathcal{W})$ such that
- $\tau_x^\mathcal{W}$ is a stopping time with ${\mathbb{P}}[\tau_x^\mathcal{W}>0]=1$,
- $X(\cdot;x)$ is a process defined on $[0,\tau_x^\mathcal{W})$ with $\widetilde{\mathbb{P}}[X(0,x)=x]=1$,
- $X(\cdot;x)$ is continuous with values in $\mathcal{W}$ for $t<\tau_x^\mathcal{W}$,
- $\|X(t;x)\|_{\mathcal{W}}\to\infty$ as $t\uparrow\tau_x^\mathcal{W}$, $\widetilde{\mathbb{P}}$–[a. s.]{},
- $X(\cdot;x)$ is solution of on $[0,\tau_x^\mathcal{W})$.
The strong solution turns out to be a Markov process (and even a strong Markov process, but we do not need this fact here) in the following sense (see [@IkeWat81] for further details). Set $\mathcal{W}' = \mathcal{W}\cup\{{\lower-1.8pt\hbox{\footnotesize\danger}\xspace}\}$, where the terminal state [-1.8pt]{}is an isolated point. Define the set $\overline W(\mathcal{W}')$ of all paths $\omega:[0,\infty)\to\mathcal{W}'$ such that there exists a time $\zeta(\omega)\in[0,\infty]$ with $\omega$ continuous with values in $\mathcal{W}$ on $[0,\zeta(\omega))$ and $\omega(t)={\lower-1.8pt\hbox{\footnotesize\danger}\xspace}$ for $t\geq\zeta(\omega)$. The strong solution defined above can be extended as a process in $[0,\infty)$ with values in $\mathcal{W}'$ in a canonical way, achieving value [-1.8pt]{}for $t\geq\tau_x^{\mathcal{W}}$. We say that the strong solution is Markov when the process on the extended state space $\mathcal{W}'$ is a Markov process.
\[t:strong\] Let $\beta>2$ and assume , . Let $\alpha\in(\beta-2,\alpha_0+1)$, then for every $x\in V_\alpha$ there exists a strong solution $(X(\cdot;x), \tau_x^\alpha)$ with initial condition $x$.
Moreover, the solution is unique in the sense that if $(X(\cdot;x),\tau_x)$ and $(X'(\cdot;x),\tau_x')$ are two solutions, then ${\mathbb{P}}[\tau_x=\tau_x']=1$ and $X(\cdot;x)=X'(\cdot;x)$ for $t<\tau_x$.
Finally, the process $(X(\cdot;x))_{x\in V_\alpha}$ is Markov, in the sense given above.
Existence and uniqueness are essentially based on the same ideas of [@Rom10a Theorem 5.1], but with simpler estimates, we give a sketch of their proof because we will use some of the definitions we introduce later. Let $\chi\in C^\infty([0,\infty))$ be decreasing and such that $\chi(u)=1$ for $u\leq 1$ and $\chi(u)=0$ for $u\geq2$, and consider the problem $$\label{e:cutoff}
dX_n^R = - \nu\lambda_n^2 X_n^R\,dt
+ \chi_R(\|X^R\|_\alpha)\bigl(\lambda_{n-1}^\beta (X_{n-1}^R)^2 - \lambda_n^\beta X_n^R X_{n+1}^R\bigr)\,dt
+ \sigma_n\,dW_n.$$ The problem has a (path–wise) unique global solution for every $x\in V_\alpha$, which is continuous in time with values in $V_\alpha$. Given $x\in V_\alpha$, define $\tau_x^{\alpha,R}$ as the first time $t$ when $\|X^R(t)\|_\alpha=R$, then $\tau_x^\alpha = \sup_{R>0}\tau_x^{\alpha,R}$ and the strong solution $X(t;x)$ coincides with $X^R(t;x)$ for $t\leq\tau_x^{\alpha,R}$. By uniqueness the definition makes sense. Markovianity follows by the Markovianity of each $X^R$.
We notice that by path–wise uniqueness, if $x\in V_\alpha$, then $\tau_x^\alpha=\tau_x^{\alpha'}$ for every $\alpha'\in(\beta-2,\alpha)$.
### Weak martingale solutions
The fact that the *blow–up time* $\tau_x^\alpha$ associated to a strong solution may (or may not) be infinite is the main topic of discussion of the paper. In order to have global solutions in the general case we state a weaker definition of solution.
\[d:weak\] Given $x\in H$, a probability measure ${\mathbb{P}}$ on $\Omega_\beta$ is a *weak martingale solution* of problem with initial condition $x$ if
1. for every $n\geq1$, the process $$M^n_t
= \xi_n(t) - \xi_n(0)
+ \int_0^t \bigl( \nu\lambda_n^2 \xi_n(s) - \lambda_{n-1}^\beta \xi_{n-1}^2 + \lambda_n^\beta \xi_n \xi_{n+1}\bigr)\,ds,$$ defined through the canonical process $\xi$, is a continuous square integrable martingale with quadratic variation $\sigma_n^2$ (hence in particular a Brownian motion if $\sigma_n\neq0$),
2. ${\mathbb{P}}[\xi(0) = x] = 1$.
It can be easily seen that this is equivalent to the standard definition of martingale solution (a proof is given for instance in [@Fla08] for the Navier–Stokes equations, but can be easily adapted to this problem).
Clearly both definition, of strong and weak solution, extend straightforwardly respectively to initial ${\mathscr{F}}_0$–measurable random variables and to initial distributions on $H$.
It turns out (see Theorem \[t:markov\] below) that the strong solution represents the unique local solution, thus all weak solutions with the same initial condition coincide with it up to the *blow–up* time $\tau_x^\alpha$, if we require that weak solutions satisfy an additional property. Before stating the property, we need a few definitions for preparation.
Let ${\mathbb{P}}$ be a weak martingale solution and define for each $n$ such that $\sigma_n\neq0$ the process $W_n = \tfrac1{\sigma_n}M_n$ (and choose a $W_n$ independent from all the other if $\sigma_n=0$). Under ${\mathbb{P}}$ the $(W_n)_{n\geq1}$ are independent standard one–dimensional Brownian motions and the sequence defined as $$Z_n(t) = \sigma_n\Bigl(W_n(t)
- \nu\lambda_n^2\int_0^t{\operatorname{e}}^{-\nu\lambda_n^2(t-s)}W_n(s)\,ds\Bigr)$$ solves the following system, $$\label{e:stokes}
\begin{cases}
dZ_n + \nu\lambda_n^2 Z_n\,dt = \sigma_n\,dW_n,\\
Z_n(0) = 0,
\end{cases}
\qquad
n\geq1.$$ Moreover, if $Y = \xi - Z$, where $\xi$ is the canonical process on $\Omega_\beta$, thus a solution (in the sense of Definition \[d:weak\]) of under ${\mathbb{P}}$, then ${\mathbb{P}}$–[a. s.]{}, $$\label{e:v}
\dot Y_n + \nu\lambda_n^2 Y_n
= \lambda_{n-1}^\beta(Y_{n-1}+Z_{n-1})^2 - \lambda_n^\beta(Y_n+Z_n)(Y_{n+1}+Z_{n+1}).$$ Define ${\mathcal{G}}_t$ as $$\begin{gathered}
{\mathcal{G}}_t(y,z)
= \|y(t)\|_H^2
+ 2\nu\int_0^t\|y(s)\|_V^2\,ds - {}\\
- 2\int_0^t\Bigl(\sum_{n=1}^\infty\lambda_n^\beta\bigl(y_n+z_n)(y_{n+1}z_n - y_n z_{n+1}\bigr)\Bigr)\,ds,\end{gathered}$$ it is easy to see (using the lemma below) that ${\mathcal{G}}_t(Y,Z)$ is finite if $Y\in L^\infty_{{\textrm{\tiny loc}}}(0,\infty;H)\cap L^2_{{\textrm{\tiny loc}}}(0,\infty;V)$ and Assumptions \[a:noise\], \[a:noise2\] hold, and is jointly measurable in the variables $(t,y,z)$ (see [@BloFlaRom09; @Rom08b] for a related problem). We first give a regularity result for $Z$, which is standard (see [@DapZab92])
\[l:zreg\] Assume with $\alpha_0\in{\mathbf{R}}$. Given $\alpha<\alpha_0+1$, then almost surely $Z\in C([0,T];V_\alpha)$ for every $T>0$. Moreover, for every $\epsilon\in(0,1]$, with $\epsilon<\alpha_0+1-\alpha$, there are ${\ensuremath{c_{\text{\ref{[}}}}}-1.$$]{l:zreg}>0$ and ${\ensuremath{c_{\text{\ref{[}}}}}-2.$$]{l:zreg}>0$, such that for every $T>0$, $${\mathbb{E}}\Bigl[\exp\Bigl(\frac{{\ensuremath{c_{\text{\ref{l:zreg}-2.$\epsilon$}}}}}{T^\epsilon}
\sup_{[0,T]}\|Z(t)\|_\alpha^2\Bigr)\Bigr]
\leq {\ensuremath{c_{\text{\ref{l:zreg}-1.$\epsilon$}}}}.$$
\[d:energy\] A weak martingale solution ${\mathbb{P}}$ with initial condition $x$ is an *energy* martingale solution if
1. ${\mathbb{P}}[Y\in L^\infty_{{\textrm{\tiny loc}}}(0,\infty;H)\cap L^2_{{\textrm{\tiny loc}}}([0,\infty);V)] = 1$, where $Y = \xi - Z$ and $Z$ is the solution to associated to ${\mathbb{P}}$,
2. there is a set $T_{\mathbb{P}}\subset(0,\infty)$ of null Lebesgue measure such that for every $s\not\in T_{\mathbb{P}}$ and every $t>s$, $${\mathbb{P}}[{\mathcal{G}}_t(Y,Z)\leq{\mathcal{G}}_s(Y,Z)] = 1.$$
Assume with $\alpha_0>0$. As in [@FlaRom06; @FlaRom07] one can consider an alternative definition where the almost sure energy inequality is replaced by the supermartingale condition given as follows:
- ${\mathbb{P}}[L^\infty(0,\infty;H)\cap L^2(0,\infty;V)] = 1$,
- for every $m\geq1$ the process $${\mathcal{E}}_t^m
= \|\xi(t)\|_H^{2m}
+ 2m\nu\int_0^t \|\xi(s)\|_H^{2m-2}\|\xi(s)\|_1^2\,ds
- m(2m-1)\sigma^2\int_0^t \|\xi(s)\|_H^{2m-2}\,ds$$ is an *almost sure supermartingale*, that is there is a set $T_{\mathbb{P}}\subset(0,\infty)$ of null Lebesgue measure such that for every $s\not\in T_{\mathbb{P}}$ and every $t\geq s$, $${\mathbb{E}}^{\mathbb{P}}[{\mathcal{E}}^m_t|{\mathscr{B}}_s]
\leq {\mathcal{E}}_s^m,$$
where $\sigma^2 = \sum_n \sigma_n^2$.
Given $\alpha>\beta-2$ and $R>0$, define the following random times on $\Omega_\beta$, $$\label{e:tauinfty}
\tau_\infty^\alpha
= \inf\{t\geq0: \|\omega(t)\|_\alpha=\infty\},
\qquad
\tau_\infty^{\alpha,R}
= \inf\{t\geq0: \|\omega(t)\|_\alpha>R\},$$ with the understanding that each random time is $\infty$ if the set is empty.
\[t:markov\] Let $\beta>2$ and assume , .
- For every $x\in H$ there exists at least one energy martingale solution ${\mathbb{P}}_x$.
- If $\alpha\in(\beta-2,1+\alpha_0)$, $x\in V_\alpha$ and ${\mathbb{P}}_x$ is an energy martingale solution with initial condition $x$, then $\tau_x^\alpha = \tau_\infty^\alpha$ under ${\mathbb{P}}_x$ and for every $t>0$, $$\xi_s = X(s;x),
\quad s\leq t,\quad
\qquad{\mathbb{P}}_x-a.s.\text{ on }\{\tau_x^\alpha>t\},$$ where $(X(\cdot;x),\tau_x^\alpha)$ is the strong solution with initial condition $x$ defined on $\Omega_\beta$.
- There exists at least one family $({\mathbb{P}}_x)_{x\in H}$ of energy martingale solutions satisfying the almost sure Markov property, namely for every $x\in H$ and every bounded measurable $\phi:H\to{\mathbf{R}}$, $${\mathbb{E}}^{{\mathbb{P}}_x}\bigl[\phi(\xi_t)|{\mathscr{B}}_s]
= {\mathbb{E}}^{{\mathbb{P}}_{\omega(s)}}[\phi(\xi_{t-s})],
\qquad {\mathbb{P}}_x-a. s.,$$ for almost every $s\geq0$ (including $0$) and for all $t\geq s$.
The proof of the first fact can be done as in [@BarBarBesFla06]. The proofs of the other two facts are entirely similar to those of Theorem 2.1 of [@Rom08b] and Theorem 3.6 of [@Rom10a] and we refer to these reference for further details.
A natural way to prove existence of weak solution (and in fact this is the way it is done in [@BarBarBesFla06]) is to use finite dimensional approximations, namely, consider for each $N\geq1$ the solution $(X_n^{(N)})_{1\leq n\leq N}$ to the following finite dimensional system, $$\label{e:galerkinx}
\begin{cases}
\dot X_1^{(N)}
= -\nu\lambda_1^2X_1^{(N)}
- \lambda_1^\beta X_1^{(N)} X_2^{(N)}
+ \sigma_1\,dW_1,\\
\dots,\\
\dot X_n^{(N)}
= - \nu\lambda_n^2X_n^{(N)}
+ \lambda_{n-1}^\beta(X_{n-1}^{(N)})^2
- \lambda_n^\beta X_n^{(N)} X_{n+1}^{(N)}
+ \sigma_n\,dW_n,\\
\dots,\\
\dot X_N^{(N)}
= - \nu\lambda_N^2X_N^{(N)}
+ \lambda_{N-1}^\beta(X_{N-1}^{(N)})^2
+ \sigma_N\,dW_N,
\end{cases}$$ Given $x\in H$, let ${\mathbb{P}}^{(N)}_x$ be the probability distribution on $\Omega_\beta$ of the solution of the above system with initial condition $x^{(N)}=(x_1,x_2,\dots,x_N)$.
\[d:galerkin\] Given $x\in H$, a *Galerkin* martingale solution is any limit point in $\Omega_\beta$ of the sequence $({\mathbb{P}}^{(N)}_x)_{N\geq1}$.
It is easy to verify (it is indeed the proof of existence in Theorem \[t:markov\] and we refer to [@BarBarBesFla06] for details in a similar problem) that *Galerkin* martingale solutions are *energy* martingale solutions.
The results stated in this section are essentially independent of the structure of the non–linear part and are only based on the fact that that the non–linearity is polynomial and at each mode involves only a finite number of other modes.
The results of the rest of the paper are on the other hand strongly based on the structure of the non–linear part. Besides the physical motivations of [@FriPav04a], from the point of view of studying a simplified version of Navier–Stokes, there are several possibilities for *nearest–neighbour* interaction among modes. One can write [@KisZla05] any nearest–neighbour non–linearity as $$\dot X_n
= -\nu\lambda_n^2 X_n + (a_1 B_n^1(X) + a_2 B_n^2(X)),$$ where $B_n^1$ is the one corresponding to the dyadic model and $B_n^2(x) = \lambda_{n+1}^\beta x_{n+1}^2 - \lambda_n^\beta x_{n-1} x_n$. In [@KisZla05] the authors notice that the inviscid problem with non–linearity $B_n^2$ is well–posed. Hence the dyadic model is the difficult, hence the most meaningful, part of any nearest–neighbour non–linearity.
Control of the negative components {#s:negative}
==================================
Given $\beta>2$, $\alpha\in{\mathbf{R}}$ and $c_0>0$, consider the solution $Z$ of and define the following process, $$\label{e:N}
N_{\alpha,c_0}(t)
= min\bigl\{m\geq1: |Z_n(s)|\leq c_0\nu\lambda_{n-1}^{-\alpha}
\text{ for }s\in[0,t]\text{ and }n\geq m\bigr\},$$ with $N_{\alpha,c_0}(t) = \infty$ if the set is empty.
\[l:moments\_N\] Given $\beta>2$, assume and let $\alpha<\alpha_0+1$. Then for every $\gamma\in(0,\alpha_0+1-\alpha)$ and $\epsilon\in(0,1]$, with $\epsilon<\alpha_0+1-\alpha-\gamma$, there are two numbers ${\ensuremath{c_{\text{\ref{l:moments_N}-1}}}}>0$ and ${\ensuremath{c_{\text{\ref{l:moments_N}-2}}}}>0$, depending only on $\epsilon$, $\gamma$ and $\alpha_0$, such that for every $t>0$ and $n\geq1$, $${\mathbb{P}}[N_{\alpha,c_0}(t) > n]
\leq {\ensuremath{c_{\text{\ref{l:moments_N}-1}}}}
{\operatorname{e}}^{-{\ensuremath{c_{\text{\ref{l:moments_N}-2}}}}\frac{c_0\nu}{t^\epsilon}\lambda_n^\gamma}.$$ In particular, $${\mathbb{E}}\bigl[{\operatorname{e}}^{\lambda_{N_{\alpha,c_0}(t)}^\gamma}\bigr]
<\infty,$$ and ${\mathbb{P}}[N_{\alpha,c_0}(t) = n] > 0$ for every $n\geq1$.
For $n\geq 1$, $$\{N_{\alpha,c_0}(t)\leq n\}
= \Bigl\{\sup_{k\geq n}\sup_{[0,t]}\lambda_{k-1}^\alpha|Z_k(s)|\leq c_0\nu\Bigr\},$$ hence if $\gamma<\alpha_0+1-\alpha$ and $k\geq n$, $$\sup_{[0,t]}\lambda_{k-1}^\alpha|Z_k(s)|
\leq \lambda_{n-1}^{-\gamma}\sup_{[0,t]}\|Z(s)\|_{\alpha+\gamma},$$ and therefore by Chebychev’s inequality and Lemma \[l:zreg\], $${\mathbb{P}}[N_{\alpha,c_0}(t)>n]
\leq {\mathbb{P}}\Bigl[\sup_{[0,t]}\|Z(s)\|_{\alpha+\gamma} > c_0\nu\lambda_{n-1}^\gamma\Bigr]
\leq {\ensuremath{c_{\text{\ref{l:zreg}-1.$\epsilon$}}}}{\operatorname{e}}^{-{\ensuremath{c_{\text{\ref{l:zreg}-2.$\epsilon$}}}}\frac{c_0\nu}{t^\epsilon}\lambda_{n-1}^\gamma},$$ for every $\epsilon\in(0,1]$ with $\epsilon<\alpha_0+1-\alpha-\gamma$. The double–exponential moment follows easily from this estimate.
We finally prove that ${\mathbb{P}}[N_{\alpha,c_0}(t)=n]>0$. We prove it for $n=1$, all other cases follow similarly. By independence, $${\mathbb{P}}[N_{\alpha,c_0}(t)=1]
= \exp\Bigl(- \sum_{n=1}^\infty-\log{\mathbb{P}}\bigl[\sup_{[0,t]}\lambda_{k-1}^\alpha|Z_k(s)|\leq c_0\nu\bigr]\Bigr)$$ and it is sufficient to show that the series above is convergent. By , $${\mathbb{P}}\bigl[\sup_{[0,t]}\lambda_{k-1}^\alpha|Z_k(s)|\leq c_0\nu\bigr]
\geq {\mathbb{P}}\bigl[\sup_{[0,t]}|\zeta(\lambda_n^2s)|\leq 2^\alpha c_0\nu\lambda_n^{\alpha_0+1-\alpha}\bigr],$$ where $\zeta$ is the solution of the one–dimensional SDE $d\zeta + \nu\zeta\,dt = dW$, with $\zeta(0) = 0$, and the conclusion follows by the fact that $\alpha<1+\alpha_0$ and standard tail estimates on the one–dimensional Ornstein–Uhlenbeck process (see for instance [@Dir75]).
The lemma below is the crucial result of the paper. To formulate its statement, we introduce the finite dimensional approximations of the problem. Consider problem and, for an integer $N\geq1$, the finite dimensional approximations of , $$\label{e:galerkin}
\begin{cases}
\dot Y_1^{(N)} = -\nu\lambda_1^2Y_1^{(N)} - \lambda_1^\beta X_1^{(N)} X_2^{(N)},\\
\dots,\\
\dot Y_n^{(N)} = -\nu\lambda_n^2Y_n^{(N)} +\lambda_{n-1}^\beta(X_{n-1}^{(N)})^2 - \lambda_n^\beta X_n^{(N)} X_{n+1}^{(N)},\\
\dots,\\
\dot Y_N^{(N)} = -\nu\lambda_N^2Y_N^{(N)} +\lambda_{N-1}^\beta(X_{N-1}^{(N)})^2,
\end{cases}$$ where for $n=1,\dots,N$ we have set $X_n^{(N)} = Y_n^{(N)} + Z_n$. It is easy to verify that the above SDE admits a unique global solution.
\[l:crucial\] Let $\beta>2$ and assume , . Let $\alpha\in[\beta-2,1+\alpha_0)$ and consider $C_0>0$, $a_0>0$ and $n_0\geq1$ such that $$\label{e:crucial}
c_0 \leq a_0
\qquad\text{and}\qquad
c_0 < \sqrt{a_0}\bigl(\lambda_{n_0}^{\frac12(\alpha+2-\beta)}-\sqrt{a_0}\bigr).$$ Given $T>0$, let $N\geq1$ and assume that $\lambda_{n-1}^\alpha X_n^{(N)}(0)\geq - a_0\nu$ for all $n=n_0,\dots,N$.
If $N > N_{\alpha,c_0}(T)$, then $Y_n^{(N)}(t)\geq -a_0\nu\lambda_{n-1}^{-\alpha}$ for all $t\in[0,T]$ and all $n\geq n_0\vee N_{\alpha,c_0}(T)$.
For simplicity we drop in this proof the superscript ${}^{(N)}$. We can first assume that $\lambda_{n-1}^\alpha Y_n(0)>-\nu a_0$ for $n\geq n_0,\dots,N$ (the equality will be included by continuity), then the same is true in a neighbourhood of $t=0$. Let $t_0>0$ be the first time when at least for one $n$, $\lambda_{n-1}^\alpha Y_n(t_0) = -\nu a_0$. Let $n\geq n_0\vee N_{\alpha,c_0}(T)$ be one of such indices, then $$\begin{aligned}
\dot Y_n(t_0)
&\geq -\nu\lambda_n^2 Y_n(t_0)
- \lambda_n^\beta(Y_n(t_0)+Z_n(t_0))(Y_{n+1}(t_0)+Z_{n+1}(t_0))\\
&\geq a_0\nu^2\lambda^2\lambda_{n-1}^{2-\alpha}
+ \lambda_n^\beta\bigl(a_0\nu\lambda_{n-1}^{-\alpha}-Z_n(t_0)\bigr)(Y_{n+1}(t_0)+Z_{n+1}(t_0))\\
&\geq a_0\nu^2\lambda^2\lambda_{n-1}^{2-\alpha}
- \lambda_n^\beta\bigl(a_0\nu\lambda_{n-1}^{-\alpha}-Z_n(t_0)\bigr)(Y_{n+1}(t_0)+Z_{n+1}(t_0))_-
\end{aligned}$$ since $\nu a_0\lambda_{n-1}^{-\alpha}-Z_n(t_0)\geq0$ for $n\geq N_{\alpha,c_0}(T)$, and where $x_- = \max(-x,0)$. We also know that $Y_{n+1}(t_0)\geq -a_0\nu\lambda_n^{-\alpha}$, hence $Y_{n+1}(t_0)+Z_{n+1}(t_0)\geq -\nu(a_0+c_0)\lambda_n^{-\alpha}$ and so $(Y_{n+1}(t_0)+Z_{n+1}(t_0))_-\leq \nu(a_0+c_0)\lambda_n^{-\alpha}$. We also have $a_0\nu\lambda_{n-1}^{-\alpha}-Z_n(t_0)\leq \nu(a_0+c_0)\lambda_{n-1}^{-\alpha}$, so in conclusion $$\dot Y_n(t_0)
\geq \nu^2\lambda^2\lambda_{n-1}^{2-\alpha}
\bigl(a_0 - \lambda_n^{\beta-2-\alpha}(a_0+c_0)^2\bigr)
>0$$ and the lemma is proved.
The next theorem is based on Lemma \[l:crucial\] and shows that the process can explode *only* in the positive area.
Given $\beta>2$, assume and . Let $\alpha\in(\beta-2,\alpha_0+1)$ and $x\in V_\alpha$, and let $(X(\cdot;x),\tau_x^\alpha)$ be the strong solution in $V_\alpha$ with initial condition $x$. Then for every $T>0$ and $p\geq1$, $${\mathbb{E}}\Bigl[\sup_{n\geq1}\sup_{t\in[0,T\wedge\tau_x^\alpha]}
\bigl(\lambda_{n-1}^\alpha\bigl(X_n(t)\bigr)_-\bigr)^p\Bigr]
<\infty.$$ In particular, $$\inf_{n\geq1}\inf_{t\in [0,\tau_x^\alpha\wedge T]}\lambda_{n-1}^\alpha X_n
> -\infty,
\qquad {\mathbb{P}}\text{--a.~s.}$$
Fix $x\in V_\alpha$ and $T>0$, and set $a_0=\tfrac14$ and $c_0=\tfrac16$, so that condition holds for any $n_0$. Next, choose $n_0\geq1$ as the smallest integer such that $\lambda_{n-1}^\alpha x_n\geq -\frac14\nu$ for all $n\geq n_0$. With the choice $c_0=\tfrac16$, set $\mathcal{Z}_{\alpha,T} = \{N_{\alpha,\frac16}(T)<\infty\}$, which by Lemma \[l:moments\_N\] is an event of probability one.
Lemma \[l:crucial\] implies that on $\{\tau_x^\alpha>T\}$, $$Y_n(t)\geq -\frac14\nu\lambda_{n-1}^{-\alpha},
\qquad\text{for }n\geq n_0\vee N_{\alpha,\frac16}(T).$$ Indeed, we can set $x^{(N)}=(x_1,\dots,x_N)$ and notice that on the event $\{\tau_x^\alpha>T\}$, problem has a unique solution, hence for every $N$ the solution of with initial condition $x^{(N)}$ converges to the solution of with initial condition $x$ (where the convergence is component–wise uniform in time on $[0,T]$).
Let $N_1 = n_0\vee N_{\alpha,\frac16}(T)$, it is clear that $N_1$ has the same finite moments of $N_{\alpha,\frac16}(T)$, moreover on $\{\tau_x^\alpha>T\}$, $$\lambda_{n-1}^\alpha X_n(t)
\geq
\begin{cases}
-\lambda_{N_1-1}^\alpha \sup_{t\in[0,T]}\|X(t)\|_H,
&\qquad n<N_1,\\
-\frac5{12}\nu
&\qquad n\geq N_1,
\end{cases}$$ for every $n\geq1$, and so $$\sup_{n\geq1}\sup_{t\in[0,T]}\lambda_{n-1}^\alpha\bigl(X_n(t)\bigr)_-
\leq \frac5{12}\nu + \lambda_{N_1-1}^\alpha\sup_{t\in[0,T]}\|X(t)\|_H$$ From Lemma \[l:moments\_N\] and the fact that ${\mathbb{E}}[\sup_{[0,T]}\|X(t)\|_H^p]$ is finite for every $p\geq1$, the estimate in the statement of the theorem readily follows.
Given an initial condition $x\in V_\alpha$, if we set $$\tau_{x,\pm}^\alpha
= \sup\{t: sup_{n\geq1}\lambda_n^\alpha(X_n)_\pm<\infty\},$$ then $\tau_x^\alpha = \min(\tau_{x,+}^\alpha,\tau_{x,-}^\alpha)$, and the previous theorem essentially states that $\tau_x^\alpha=\tau_{x,+}^\alpha$.
Given $\beta>2$, assume and . Given $\alpha\in(\beta-2,\alpha_0+1)$ and $x\in V_\alpha$, assume additionally either that problem , with initial condition $x$, admits a unique solution, for almost every possible value assumed by $Z$, or that we are dealing with a *Galerkin* solution starting in $x$. Then for every $T>0$ and $p\geq1$, $${\mathbb{E}}\Bigl[\sup_{n\geq1}\sup_{t\in[0,T]}
\bigl(\lambda_{n-1}^\alpha\bigl(X_n(t)\bigr)_-\bigr)^p\Bigr]
<\infty.$$ In particular, $$\inf_{n\geq1}\inf_{t\in [0,T]}\lambda_{n-1}^\alpha X_n
> -\infty,
\qquad {\mathbb{P}}\text{--a.~s.}$$
We simply notice that in the proof of the theorem above we have used the piece of information $\tau_x^\alpha>T$ only to ensure that problem admits a unique solution.
On the other hand, if we are dealing with a Galerkin solution, then up to a sub–sequence we still have component–wise uniform convergence in time.
Uniqueness and regularity for 2<beta<=5/2 {#s:unique}
===============================================
In this section we prove two results, the first concerning path–wise uniqueness, the second concerning regularity (absence of blow–up), which are essentially extensions of the corresponding results for the case without noise. The extension is made possible by means of the control of negative components shown in Section \[s:negative\]. The path–wise uniqueness result below is restricted to *Galerkin* solutions (see Definition \[d:galerkin\]).
\[t:unique\] Let $\beta\in(2,3]$ and assume that , hold. Let $X(0)\in V_{\beta-2}$, then there exists a (path–wise) unique solution of with initial condition $X(0)$, in the class of *Galerkin* martingale solutions.
We do not know if there is uniqueness in some larger class (energy martingale solutions or weak martingale solutions), neither we know if a Galerkin martingale solution may develop blow–up. By slightly restricting the range of values of $\beta$, we do actually have an improvement over the previous result.
\[t:smooth\] There exists $\beta_c\in (\tfrac52,3]$ such that the following statement holds true. Assume that , hold and let $\beta\in(2,\beta_c)$ and $\alpha\in(\beta-2,1+\alpha_0)$, then for every $x\in V_\alpha$, $\tau_x^\alpha=\infty$. In particular, path–wise uniqueness holds in the class of *energy* martingale solutions.
The proof of Theorem \[t:unique\]
---------------------------------
The proof is based on the idea in [@BarMorRom10 Proposition 3.2] which builds up on a result in [@BarFlaMor09b]. Both results are proved for positive solutions (and no noise).
It is sufficient to show uniqueness on any finite time interval, so we fix an arbitrary $T>0$ and show that there is only one solution on $[0,T]$. We will use Lemma \[l:crucial\] with $c_0=\tfrac16$ and $a_0=\tfrac14$ (so that holds for any $n_0$). In fact, since a Galerkin martingale solution is the component–wise limit of finite dimensional approximations, the bounds of the lemma remain stable in the limit to the infinite dimensional system.
Since $X(0)\in V_{\beta-2}$, we know that there is $n_0\geq1$ such that $\lambda_n^{\beta-2}X_n(0)\geq -\tfrac14\nu$. We set $N_0 = 1 + n_0\vee N_{\beta-2,\frac16}(T)$.
Let $X^1$, $X^2$ two solutions with the same given initial condition $X^1(0) = X^2(0) = X(0)$. By Lemma \[l:crucial\] we know that $X_n^i(t)\geq Z_n(t) - \tfrac14\nu\lambda_{n-1}^{2-\beta}$ for $n\geq N_0$, $t\in[0,T]$ and $i=1,2$. Set now $$A_n = X_n^1 - X_n^2,
\qquad
B_n = \frac12\nu\lambda_{n-1}^{2-\beta} - 2 Z_n$$ and $$\psi_\ell(t) = \sum_{n=1}^{N_0-1}\frac{A_n^2}{\lambda_n},
\qquad
\psi_{h,N}(t) = \sum_{N_0}^N\frac{A_n^2}{\lambda_n},
\qquad
\psi_N(t) = \psi_\ell(t) + \psi_{h,N}(t),$$ and notice that $X_n^1 + X_n^1 + B_n\geq0$ if $t\in[0,T]$ and $n\geq N_0$. A simple computation shows that $$\begin{gathered}
\dot A_n
= - \nu\lambda_n^2 A_n
+ \lambda_{n-1}^\beta(X_{n-1}^1+X_{n-1}^2)A_{n-1} + {}\\
- \frac12\lambda_n^\beta\bigl[(X_n^1+X_n^2)A_{n+1}
+ (X_{n+1}^1+X_{n+1}^2)A_n\bigr],
\end{gathered}$$ hence for $N > N_0$, $$\begin{gathered}
\frac{d}{dt}\psi_{h,N}
+ 2\nu\sum_{n=N_0}^N \lambda_n A_n^2
= - \sum_{n=N_0}^N \lambda_n^{\beta-1}(X_{n+1}^1 + X_{n+1}^2) A_n^2 + {}\\
- \lambda_N^{\beta-1}(X_N^1+X_N^2)A_NA_{N+1}
+ \lambda_{N_0-1}^{\beta-1}(X_{N_0-1}^1+X_{N_0-1}^2)A_{N_0-1}A_{N_0} =\\
= { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{1}}\kern-2pt}
}
\xspace
} + { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{2}}\kern-2pt}
}
\xspace
}_N + { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{3}}\kern-2pt}
}
\xspace
}_{N_0}.
\end{gathered}$$ For the first term we notice that $B_{n+1}\leq \tfrac56\nu\lambda_n^{2-\beta}$, hence $${ \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{1}}\kern-2pt}
}
\xspace
}
\leq \sum_{n=N_0}^N\lambda_n^{\beta-1}B_{n+1}A_n^2
\leq \nu\sum_{n=N_0}^N\lambda_n A_n^2.$$ For the second term, $$\sum_{N=1}^\infty \int_0^T { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{2}}\kern-2pt}
}
\xspace
}_N\,dt
\leq \sup_{[0,T]}\|X^1 + X^2\|_H\int_0^T\|A\|_{\frac12(\beta-1)}^2\,ds,$$ and the quantity on the right–hand side is [a. s.]{} finite since by $Z$ is [a. s.]{} in $C([0,T];V_\gamma)$ for every $\gamma<1+\alpha_0$ (hence for $\gamma=\tfrac12(\beta-1)$), and since by Definition \[d:energy\], $V\in L^2([0,T];V)$ (and $\tfrac12(\beta-1)\leq1$). This implies that [a. s.]{} $\int_0^T{ \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{2}}\kern-2pt}
}
\xspace
}_N\,dt\to0$ as $N\to\infty$.
Likewise (recall that $X_0^1 = X_0^2 = 0$), $$\frac{d}{dt}\psi_\ell
\leq - \sum_{n=1}^{N_0-1} \lambda_n^{\beta-1}(X_{n+1}^1 + X_{n+1}^2) A_n^2
- { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{3}}\kern-2pt}
}
\xspace
}_{N_0}
\leq \lambda_{N_0-1}^\beta\bigl(\sup_{[0,T]}\|X^1 + X^2\|_H\bigr)\psi_\ell
- { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{3}}\kern-2pt}
}
\xspace
}_{N_0},$$ and in conclusion $$\frac{d}{dt}\psi_N
\leq \lambda_{N_0-1}^\beta\bigl(\sup_{[0,T]}\|X^1 + X^2\|_H\bigr)\psi_\ell
+ { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{2}}\kern-2pt}
}
\xspace
}_N.$$ Integrate in time (recall that $\psi_N(0) = 0$) and take the limit as $N\uparrow\infty$, $$\psi(t)
\leq \lambda_{N_0-1}^\beta\bigl(\sup_{[0,T]}\|X^1 + X^2\|_H\bigr) \int_0^t\psi(s)\,ds$$ where $\psi_N\uparrow\psi$ and $\psi(t) = \|A(t)\|_{-\frac12}^2$. Since the term $\lambda_{N_0-1}^\beta\bigl(\sup_{[0,T]}\|X^1 + X^2\|_H\bigr)$ is [a. s.]{} finite, by Gronwall’s lemma it follows that [a. s.]{} $A(t)=0$ for all $t\in[0,T]$.
The proof of Theorem \[t:smooth\]
---------------------------------
We start by giving a minimal requirement for smoothness of solutions to which is analogous to the criterion developed in the noise–less case (see [@BarMorRom10]).
Given $T>0$ define the subspace $K_T$ of $\Omega_\beta$ as $$K_T
= \bigl\{\omega\in\Omega_\beta: \lim_n\bigl(\max_{t\in[0,T]}\lambda_n^{\beta-2}|\omega_n(t)|\bigr) = 0\bigr\}.$$ Clearly $L^\infty(0,T;V_\alpha)\subset K_T$ for all $\alpha>\beta-2$.
Under the assumptions of Theorem \[t:smooth\], let $x\in V_\alpha$. If ${\mathbb{P}}_x$ is an energy martingale solution (Definition \[d:energy\]) with initial condition $x$ and $\tau_\infty^\alpha$ is the random time defined in , then for every $T>0$, $$\{\tau_\infty^\alpha>T\}
= K_T,
\qquad\text{under }{\mathbb{P}}_x.$$
Fix $\alpha\in(\beta-2,\alpha_0+1)$, $x\in V_\alpha$ and an energy martingale solution ${\mathbb{P}}_x$ starting at $x$, and let $\tau_\infty^\alpha$ be the random time defined in . Assume $\tau_\infty^\alpha(\omega)>T$, then there is $R_0>\|x\|_\alpha$ such that $\tau_\infty^{\alpha,R_0}(\omega)>T$, where $\tau_\infty^{\alpha,R_0}$ is also defined in . In particular $\|\xi_t(\omega)\|_\alpha\leq R_0$ for $t\in[0,T]$. Hence $$\begin{gathered}
\lambda_n^{\beta-2}\max_{[0,T]}|\xi_n(t,\omega)|
\leq \lambda_n^{\beta-2-\alpha}\max_{[0,T]}|\lambda_n^\alpha\xi_n(t,\omega)|\leq\\
\leq \lambda_n^{\beta-2-\alpha}\sup_{[0,T]}\|\xi_t(\omega)\|_\alpha
\leq R_0\lambda_n^{\beta-2-\alpha}
\end{gathered}$$ and $\omega\in K_T$.
Vice versa, assume that $\omega\in K_T$ and choose a decreasing sequence $(M_n)_{n\geq1}$ such that $M_n\downarrow0$ and $\lambda_n^{\beta-2}\max_{[0,T]}|\xi_n(t)|\leq M_n$. Set $u_n = \lambda_n^\alpha\xi_n$ and $m_n = \max_{[0,T]}|u_n(t)|$, then $$\begin{aligned}
|u_n(t)|
&\leq |u_n(0)|
+ \lambda_n^\alpha|Z_n(t)|
+ M_{n-1}\lambda_n^2\int_0^t{\operatorname{e}}^{-\nu\lambda_n^2(t-s)}\bigl(\lambda^{\alpha-2}|u_{n-1}| + \lambda^{2-\beta}|u_n|\bigr)\,ds\\
&\leq |u_n(0)|
+ \bigl(\sup_{[0,T]}\lambda_n^\alpha|Z_n(t)|\bigr)
+ \frac1\nu\lambda^{\alpha-2}M_{n-1}m_{n-1}
+ \frac1\nu\lambda^{2-\beta}M_{n-1}m_n
\end{aligned}$$ and in conclusion $$\Bigl(1 - \frac1\nu\lambda^{2-\beta}M_{n-1}\Bigr)m_n
\leq |u_n(0)|
+ \bigl(\sup_{[0,T]}\lambda_n^\alpha|Z_n(t)|\bigr)
+ \frac1\nu\lambda^{\alpha-2}M_{n-1}m_{n-1}.$$ For $n$ large enough we have that $$M_{n-1}\leq\frac{\nu}{2\lambda^{2-\beta}+2\lambda^{\alpha-2}}
\qquad\text{and\qquad}
\frac{\lambda^{\alpha-2}M_{n-1}}{\nu - \lambda^{2-\beta}M_{n-1}}\leq \frac12,$$ and in particular $\bigl(1 - \tfrac1\nu\lambda^{2-\beta}M_{n-1}\bigr)\geq\tfrac12$, hence $$m_n\leq A_n + \frac12 m_{n-1}$$ where $A_n = 2\lambda_n^\alpha|x_n| + 2 \bigl(\sup_{[0,T]}\lambda_n^\alpha|Z_n(t)|\bigr)$. By Lemma \[l:zreg\] (applied to an $\alpha'>\alpha$) we have $\sum_n A_n^2<\infty$ with probability $1$, since $\sum A_n^2\leq c(\|x\|_\alpha^2 + \sup_{[0,T]}\|Z\|_{\alpha'}^2)$. Moreover $$m_n^2
\leq A_n^2 + \frac14 m_{n-1}^2 + A_n m_{n-1}
\leq \frac32 A_n^2 + \frac34 m_{n-1}^2,$$ and by solving the recursive inequality we get $\sum_n m_n^2<\infty$, which in particular implies that $\tau_\infty^\alpha(\omega)>T$.
In order to prove Theorem \[t:smooth\], we essentially prove that, given a smooth $x$ and $T>0$, there is a solution ${\mathbb{P}}_x$ starting at $x$ that satisfies ${\mathbb{P}}_x[K_T]=1$, and hence is the unique solution.
Fix $\alpha\in(\beta-2,\alpha_0+1)$, $x\in V_\alpha$, $T>0$ and an energy martingale solution ${\mathbb{P}}_x$ starting at $x$, and let $\tau_\infty^\alpha$ be the random time defined in . Without loss of generality we can assume that ${\mathbb{P}}_x$ is a Galerkin solution (see Definition \[d:galerkin\]). Indeed, by Theorem \[t:markov\] the random time $\tau_\infty^\alpha$ coincides [a. s.]{} with the lifespan $\tau_x^\alpha$ of the strong solution with the same initial condition $x$, and the probability of the event $\{\tau_\infty^x\}$ does not depend on ${\mathbb{P}}_x$, but only on the strong solution.
Since ${\mathbb{P}}_x$ is a Galerkin solution, there are $(x^{(N_k)},{\mathbb{P}}^{(N_k)})_{k\in{\mathbf{N}}}$ such that $x^{(N_k)}\to x$ in $H$ and ${\mathbb{P}}^{(N_k)}\rightharpoonup{\mathbb{P}}_x$ in $\Omega_\beta$, and for each $k$, ${\mathbb{P}}^{(N_k)}$ is the solution of , with dimension $N_k$ and initial condition $x^{(N_k)}$. By definition we also have that $x_n^{(N_k)}=x_n$ for $n\leq N_k$.
By a standard argument (Skorokhod’s theorem) there are a common probability space $(\widetilde\Omega,\widetilde{\mathscr{F}},\widetilde{\mathbb{P}})$ and random variables $X^{(N_k)}$, $X$ on $\widetilde\Omega$ with distributions ${\mathbb{P}}^{(N_k)}$, ${\mathbb{P}}_x$ respectively, such that, in particular, $X_n^{(N_k)}\to X_n$, $\widetilde{\mathbb{P}}$–[a. s.]{}, uniformly on $[0,T]$ for all $n\geq1$ and all $T>0$.
Let $\epsilon>0$ be small enough such that $\alpha>\beta-2+2\epsilon$ and $6-2\beta-3\epsilon>0$. We will use Lemma \[l:crucial\] with $a_0<\tfrac12$ (to be chosen later in the proof) and $c_0=\tfrac13 a_0$. Let $\overline n$ be the smallest integer such that $\lambda_{n-1}^\alpha |x_n|\leq a_0\nu$ for all $n\geq\overline n$, and set $N_0 = \overline n\vee N_{\beta-2+2\epsilon,c_0}(T)$.
For each integer $n_0\geq1$ and real $M>0$ define the sets $$A_M(n_0)
= \Bigl\{\sup_{[0,T]}\bigl(|X_{n_0-1}^{(N_k)}| + |X_{n_0}^{(N_k)}|\bigr)
\leq M\text { for all $k$ such that }N_k\geq n_0\Bigr\},\\$$ Notice that $\widetilde{\mathbb{P}}\bigl[\bigcup_{M>0}A_M(n_0)\bigr] = 1$ since $X_n^{(N_k)}\to X_n$ uniformly, $\widetilde{\mathbb{P}}$–[a. s.]{}, for all $n\geq1$, hence $$\widetilde{\mathbb{P}}\Bigl[\bigcup_{n_0\geq 1,\ M>0} \bigl(\{N_0 = n_0\}\cap A_M(n_0)\bigr)\Bigr]
= 1.$$ Fix $n_0\geq1$, $M>0$ and $N\geq n_0$, then everything boils down to prove that $K_T$ happens on $\{N_0 = n_0\}\cap A_M(n_0)$ for $(X_n^{(N_k)})_{n_0\leq n\leq N_k}$ uniformly in $k$. In the following we work path–wise for $\omega\in\{N_0 = n_0\}\cap A_M(n_0)$ and we aim to use the method in [@BarMorRom10], which is based on proving that the area shown in Figure \[f:trapping\] is invariant for the dynamics of the (suitably rescaled) solution.
(0,0) – (0.1,0) – (0.1,0.1) – (0,0.1); (0,0) – (1.5,0) node \[right\] [$U_n$]{}; (0,0) – (0,1.1) node \[above\] [$U_{n+1}$]{}; (0,0) – (0, 0.6) node\[left\] [$\theta$]{} – (0.53,1) – (1,1) – (1,0.66) node\[right\] [$c$]{}; plot (,[.82\*(-0.1)\*(-0.1)]{}); (0.1,0) node\[below\] [$\delta$]{} – (0,0); (0,1) node\[left\] [$1$]{} – (0.53,1) (1,0) node\[below\] [$1$]{} – (1,0.66); (0.5,0.5) node [$A$]{}; (0,0.3) – (0.1,0.3) node\[right\] [$\vec{n}_1$]{}; (0.26,0.8) – (0.33,0.72) node\[right\] [$\vec{n}_2$]{}; (0.77,1) – (0.77,0.9) node\[left\] [$\vec{n}_3$]{}; (1,0.84) – (0.9,0.84) node\[below, text centered\] [$\vec{n}_4$]{}; (.67,0.27) – (0.6,0.33) node\[left\] [$\vec{n}_5$]{}; (0.05,0) – (0.05,0.1) node\[right\] [$\vec{n}_6$]{}; (-.5,.8) node \[left\] [$g(x)$]{} to \[out=-20,in=135\] (.2,.8); (.7,.27) to \[in=160,out=-45\] (1.5,.27) node \[right\] [$h(x)$ (or $h_\eta(x)$)]{};
In [@BarMorRom10] the area $A$ in the picture is defined for $\beta\leq\tfrac52$ with the values $c=\lambda^{-(6-2\beta-3\epsilon)}$, $g(x)=\min(mx+\theta,1)$, $h(x) = c(\max(x-\delta,0)/(1-\delta))^{\lambda^2}$, $\delta=\tfrac1{10}$, $\theta=\tfrac35$ and $m=\tfrac34$. To deal with the random perturbation we shall need to slightly modify the set $A$, in particular we will replace the border delimited by $h(x)$ with a new function $h_\eta$ (see ).
First we define the re–scaling. Let $\epsilon_n = \nu\lambda_{n-1}^{-2\epsilon}$ and $$\begin{cases}
u_n = \lambda_n^\epsilon(\lambda_n^{\beta-2}Y_n^{(N_k)} + a_0\epsilon_n),\\
v_n = \lambda_n^\epsilon(a_0\epsilon_n - \lambda_n^{\beta-2}Z_n),
\end{cases}$$ then we know by Lemma \[l:crucial\] that $$\label{e:uvbounds}
\begin{cases}
u_n\geq0,\\
\frac23a_0\lambda_n^\epsilon\epsilon_n
\leq v_n
\leq \frac53a_0\lambda_n^\epsilon\epsilon_n
\end{cases}
\qquad\text{for all }n_0\leq n\leq N_k.$$ We have that $$\begin{gathered}
\dot u_n
= \lambda_n^{2-\epsilon}\Bigl[
- \nu\lambda_n^\epsilon u_n
+ a_0\nu\lambda_n^{2\epsilon}\epsilon_n
+ \lambda^{\beta-4+2\epsilon}(u_{n-1} - v_{n-1})^2 + {}\\
- \lambda^{2-\beta-\epsilon}(u_n - v_n)(u_{n+1} - v_{n+1})
\Bigr].
\end{gathered}$$ Define $\delta_0>0$ as $$\frac1\delta_0
= \max\Bigl\{\frac1\delta,
\lambda_{n_0}^{\beta-2+\epsilon}M + 2a_0\lambda_{n_0}^\epsilon\epsilon_{n_0-1},
\sup_{n\geq n_0}\lambda_n^\epsilon(\lambda_n^{\beta-2}x_n+2a_0\epsilon_n)\Bigr\}$$ and replace $u_n$ with $U_n(t) = \delta_0^2 u_n(\delta_0 t)$. It turns out that $U_n(0)\leq\delta_0\leq\delta$ for all $n\geq n_0$ and $\max_{[0,T]} U_{n_0-1}\leq\delta_0\leq\delta$ and $\max_{[0,T]} U_{n_0}\leq\delta_0\leq\delta$, since $$|U_{n_0}|
\leq \delta_0^2\lambda_n^\epsilon\bigl(\lambda_{n_0}^{\beta-2}|X_{n_0}^{(N_k)}| + a_0\epsilon_{n_0} + \lambda_{n_0}^{\beta-2}|Z_{n_0}|\bigr)
\leq \delta_0^2\bigl(\lambda_{n_0}^{\beta-2+\epsilon}M + 2a_0\lambda_{n_0}^\epsilon\epsilon_{n_0}\bigr)
\leq \delta$$ (and similarly for $U_{n_0-1}$). Moreover the $U_n$ satisfy $$\dot U_n
= \lambda_n^{2-\epsilon}
\Bigl[
\delta_0^3 P_n^0(U,V)
+ \delta_0 P_n^1(U,V)
+ \frac1{\delta_0}\lambda^{\beta-4+2\epsilon} P_n^2(U,V)
\Bigr],$$ where $$\begin{aligned}
P_n^0
&= a_0\nu\lambda_n^{2\epsilon}\epsilon_n
+ \lambda^{\beta-4+2\epsilon} V_{n-1}^2
- \lambda^{2-\beta-\epsilon} V_n V_{n+1},\\
P_n^1
&= - \nu\lambda_n^\epsilon U_n
- 2\lambda^{\beta-4+2\epsilon} V_{n-1} U_{n-1}
+ \lambda^{2-\beta-\epsilon} (V_{n+1} U_n + V_n U_{n+1}),\\
P_n^2
&= U_{n-1}^2 - \lambda^{6-2\beta-3\epsilon} U_n U_{n+1},
\end{aligned}$$ and $V_n(t) = v_n(\delta_0 t)$.
We proceed now as in [@BarMorRom10] and consider for each $n\geq n_0$ the coupled systems in $(U_n,U_{n+1})$, $$\label{e:field}
\frac{d}{dt}
\begin{pmatrix}
U_n\\
U_{n+1}
\end{pmatrix}
= \lambda_n^{2-\epsilon}\Bigl(
\delta_0^3{\mathfrak{B}}_n^0
+ \delta_0{\mathfrak{B}}_n^1
+ \frac1{\delta_0}\lambda^{\beta-4+2\epsilon}{\mathfrak{B}}_n^2\Bigr)$$ where $${\mathfrak{B}}_n^i
= \begin{pmatrix}
P_n^i\\
\lambda^{2-\epsilon}P_{n+1}^i
\end{pmatrix},
\qquad i=0,1,2,$$ The goal is to prove that $(U_n(t))_{n\geq n_0}$ is uniformly bounded in $n$ and $t$. Indeed, we will show that $0\leq U_n(t)\leq 1$ for all $n\in\{n_0,\dots,N_k\}$, which in turns implies that $$-\lambda_n^\epsilon\epsilon_n \leq \lambda_n^{\beta-2+\epsilon}Y_n^{(N_k)} \leq \frac1{\delta_0^2}.$$ for all $n\in\{n_0,\dots,N_k\}$ and all $t\in[0,T]$. Since $Y_n^{(N_k)}\to Y_n$ uniformly on $[0,T]$ for each $n$, the same bound holds for the limit $Y$ and, due to Lemma \[l:zreg\], $X\in K_T$.
It remains to show that $U_n\leq 1$, and to this end we show that each pair $(U_n(t), U_{n+1}(t))$, for $n\geq n_0$, remains in $A$ (as it is already in the interior of $A$ at time $t=0$ by the choice of $\delta_0$). This will be again a dynamical system proof and it is enough to prove that the vector field in which gives $(\dot U_n, \dot U_{n+1})$ points inwards on the boundary of $A$, or equivalently that the scalar product of the vector fields with the normal inwards is positive.
We make two preliminary remarks: first there is nothing to prove for the two pieces of boundary corresponding to the normal vectors $\vec{n}_1$ and $\vec{n}_6$, as this is essentially already ensured by Lemma \[l:crucial\], and second that since $A$ is convex it is also possible to check that the product of each of the three vector fields ${\mathfrak{B}}_n^0$, ${\mathfrak{B}}_n^1$ and ${\mathfrak{B}}_n^2$ separately is positive.
*The vector field ${\mathfrak{B}}^1$*. We will use , that $U\leq 1$ in $A$ and that $\epsilon_n$ is non–increasing. We will obtain lower bounds that are positive if the number $a_0$ is chosen small enough (with size depending only on the constants $m$, $\beta$ and $\epsilon$, but not on $M$, $n_0$ or $\delta_0$).
On the border with normal $\vec{n}_2 = (g',1) = (m,-1)$ we have $U_n\in[0,\tfrac{1-\theta}{m}]$ and $U_{n+1}=mU_n+\theta$, hence $\lambda^2 U_{n+1} - m U_n\geq\lambda^2\theta$ and $$\label{e:b12}
\begin{aligned}
{\mathfrak{B}}_n^1\cdot \vec{n}_2
& = m P_n^1 - \lambda^{2-\epsilon} P_{n+1}^1\\
&\geq \nu\lambda_n^\epsilon(\lambda^2 U_{n+1} - m U_n) - 4a_0(m\lambda^{\beta-4+\epsilon}+\lambda^{4-\beta})\lambda_n^\epsilon\epsilon_{n-1}\\
&\geq \lambda_n^\epsilon\bigl(\nu\lambda^2\theta - a_0 c(m,\beta,\epsilon)\epsilon_{n-1}\bigr)
\end{aligned}$$ On the border with normal $\vec{n}_3 = (0, -1)$ we have $U_{n+1} = 1$, hence $$\label{e:b13}
{\mathfrak{B}}_n^1\cdot \vec{n}_3
= - \lambda^{2-\epsilon} P_{n+1}^1
\geq \lambda^{2-\epsilon} \lambda_{n+1}^\epsilon (\nu - 4a_0\lambda^{2-\beta}\epsilon_{n+1})$$ Similarly, on the border with normal $\vec{n}_4 = (-1, 0)$ we have $U_n = 1$, hence $$\label{e:b14}
{\mathfrak{B}}_n^1\cdot \vec{n}_4
= - P_n^1
\geq \lambda_n^\epsilon (\nu - 4a_0\lambda^{2-\beta}\epsilon_n).$$ Finally we consider the border with normal $\vec{n}_5$. Prior to this we give the precise definition of this piece of the boundary. For $\eta\in(0,1)$ define $\varphi_\eta(x) = \bigl((x-\eta)/(1-\eta)\bigr)^{\lambda^2}$, $x\in[\eta,1]$, and notice that $\varphi_\eta$ is positive, increasing, convex and $$x\varphi_\eta' - \lambda^2\varphi_\eta
\geq \frac{\lambda^2\eta}{1-\eta}\Bigl(\frac{\delta-\eta}{1-\eta}\Bigr)^{\lambda^2-1}
>0.$$ In [@BarMorRom10] it was used $h(x) = c\varphi_\delta(x)$. Here we consider $$\label{e:hdef}
h_\eta(x)
= \frac{c}{1-\varphi_\eta(\delta)}\bigl(\varphi_\eta(x) - \varphi_\eta(\delta)\bigr),
\qquad x\in[\delta,1],$$ with $\eta<\delta$ and we notice that $h_\eta$ is positive, increasing, convex, $h_\eta(\delta)=0$, $h_\eta(1) = c$ and $h_\eta\to h$ in $C^1([\delta,1])$ as $\eta\uparrow\delta$. Moreover, $$x h_\eta' - \lambda^2 h_\eta
\geq \frac{c\lambda^2\delta}{1-\varphi_\eta(\delta)}\,\frac{(\delta-\eta)^{\lambda^2-1}}{(1-\eta)^{\lambda^2}}
= c_{\delta-\eta}
>0.$$ With this inequality in hands, we proceed with the estimate of ${\mathfrak{B}}_n^1\cdot\vec{n}_5$. On the border with normal $\vec{n}_5 = (-h_\eta'(U_n), 1)$ we have $U_n\in[\delta,1]$ and $U_{n+1} = h_\eta(U_n)$, hence $$\label{e:b15}
\begin{aligned}
{\mathfrak{B}}_n^1\cdot\vec{n}_5
& = \lambda^{2-\epsilon} P_{n+1}^1 - h_\eta'(U_n)P_n^1\\
&\geq \lambda_n^\epsilon\bigl[\nu\bigl(h_\eta'(U_n)U_n - \lambda^2 h_\eta(U_n)\bigr)
- 4a_0\lambda^{\beta-2+\epsilon}\epsilon_n - 4a_0\lambda^{2-\beta-\epsilon}h'(U_n)\epsilon_n\bigr]\\
&\geq \lambda_n^\epsilon\bigl(\nu c_{\delta-\eta} - a_0 c(\beta,\epsilon,\delta)\epsilon_n\bigr),
\end{aligned}$$ since $h_\eta'\leq c\lambda^2 /(1-2\delta)$.
*The vector field ${\mathfrak{B}}^2$*. The fact that ${\mathfrak{B}}^2_n\cdot\vec{n}_3$ and ${\mathfrak{B}}^2_n\cdot\vec{n}_4$ are positive follows immediately from [@BarMorRom10], since in those computations the boundary function $h$ is not involved. As it regards ${\mathfrak{B}}^2_n\cdot\vec{n}_2$ and ${\mathfrak{B}}^2_n\cdot \vec{n}_5$, it is easy to see that they are both continuous functions of $h$ and $h'$ and with the values of the parameters chosen in [@BarMorRom10] (as well as here), both scalar products have positive minima. Hence the same is true if we replace $h$ with $h_\eta$, for $\eta$ small enough, since $h_\eta\to h$ in $C^1([\delta,1])$.
*The vector field ${\mathfrak{B}}^0$*. Using we have that $$\begin{aligned}
|P_n^0|
&\leq a_0\nu\lambda_n^{2\epsilon}\epsilon_n + \lambda^{\beta-4+2\epsilon} V_{n-1}^2 + \lambda^{2-\beta-\epsilon} V_n V_{n+1}\\
&\leq \lambda_n^\epsilon\bigl[a_0\nu\lambda_n^\epsilon\epsilon_n + 4a_0^2(\lambda^{\beta-4} + \lambda^{2-\beta})\lambda_n^\epsilon\epsilon_{n-1}^2\bigr]
\end{aligned}$$ and the quantity above can be made a small fraction of $\lambda_n^\epsilon$ if $a_0$ is small enough. This ensures that together ${\mathfrak{B}}_n^1$ and ${\mathfrak{B}}_n^0$ point inward, since the lower bounds for ${\mathfrak{B}}_n^1$ in formulae , , , are strictly positive and independent of $n$, for a suitable choice of $a_0$ (up to the common multiplicative factor $\lambda_n^\epsilon$).
The proof we have given (due to the choice of the numbers $m$, $\theta$, $\delta$) works for $\beta\leq\tfrac52$, hence we can consider $\beta_c$ slightly larger than $\tfrac52$. A larger value of $\beta_c$ may be considered (see [@BarMorRom10 Remark 2.2]).
The blow–up time {#s:bu_abstract}
================
In this section we analyse with more details the blow–up time introduced in Definition \[d:strong\]. We give some general results which hold beyond the dyadic model under examination and which will be used in the next section to prove that blow–up happens with probability $1$. Example \[x:onedim\] shows that the fact that blow–up happens almost surely is a property that in general depends on the structure of the drift and hence strongly motivates our analysis.
Thus, consider the local strong solution $(X(\cdot;x), \tau_x)_{x\in\mathcal{W}}$ of a stochastic equation (in finite or infinite dimension) on the state space $\mathcal{W}$, where $\mathcal{W}$ is a separable Hilbert space. By comparing with our case, we assume that
- ${\mathbb{P}}[\tau_x>0]=1$ for all $x\in\mathcal{W}$,
- $X(\cdot;x)$ is continuous for $t<\tau_x$ with values in $\mathcal{W}$,
- $X(\cdot;x)$ is the maximal local solution, namely either $\tau_x=\infty$ or $\|X(t;x)\|_{\mathcal{W}}\to\infty$ as $t\uparrow\infty$, ${\mathbb{P}}$–[a. s.]{},
- $(X(\cdot;x), \tau_x)_{x\in\mathcal{W}}$ is Markov (in the sense given in Theorem \[t:strong\]),
- all martingale solutions coincide with the strong solution up to the blow–up time $\tau_x$.
The last statement plainly implies that the occurrence of blow–up is an intrinsic property of the local strong solution and does not depend in an essential way from weak solutions.
Define for $x\in\mathcal{W}$ and $t\geq0$, $$\flat(t,x)
= {\mathbb{P}}[\tau_x>t],$$ then we know that $\flat(0,x)=1$ and $\flat(\cdot,x)$ is non–increasing. Set $$\flat(x)
= \inf_{t\geq0} \flat(t,x) = {\mathbb{P}}[\tau_x=\infty].$$ We prove a dichotomy for the supremum of $\flat$, namely this quantity is either equal to $0$ or to $1$.
\[l:solotre\] Consider the family of processes $(X, \tau)$ on $\mathcal{W}$ as above. If there is $x_0\in\mathcal{W}$ such that ${\mathbb{P}}[\tau_{x_0}=\infty]>0$, then $$\sup_{x\in\mathcal{W}}{\mathbb{P}}[\tau_x=\infty]
= 1$$
By the Markov property, $$\flat(t+s,x)
= {\mathbb{P}}[\tau_x>t+s]
= {\mathbb{E}}[{\mathbbm{1}}_{\{\tau_x>t\}}\flat(s,X(t;x))]$$ and in the limit as $s\uparrow\infty$, by monotone convergence, $$\label{e:solotre}
\flat(x)
= {\mathbb{E}}[{\mathbbm{1}}_{\{\tau_x>t\}}\flat(X(t;x))].$$ Set $c=\sup\flat(x)$, then by the above formula, $$\flat(x_0)
= {\mathbb{E}}[{\mathbbm{1}}_{\{\tau_{x_0}>t\}}\flat(X(t;x_0))]
\leq c{\mathbb{E}}[{\mathbbm{1}}_{\{\tau_{x_0}>t\}}]
= c \flat(t,x_0),$$ and, as $t\uparrow\infty$, we get $\flat(x_0)\leq c \flat(x_0)$, that is $c\geq1$, hence $c=1$.
If we have the additional information that there is $x_0\in\mathcal{W}$ such that $\flat(x_0) = 1$, then something more can be said. Indeed, ${\mathbbm{1}}_{\{\tau_{x_0}>t\}} = 1$ almost surely, and, using again formula , $${\mathbb{E}}[\flat(X(t;x_0))]
= {\mathbb{E}}[{\mathbbm{1}}_{\{\tau_{x_0}>t\}}\flat(X(t;x_0))]
= \flat(x_0)
= 1,$$ hence $\flat(X(t;x_0)) = 1$, almost surely for every $t>0$. This is very close to prove that $\flat\equiv0$, and in fact [@FlaRom08 Theorem 6.8] proves, although with a completely different approach, that under the assumptions of
- (suitable) strong Feller regularity,
- *conditional irreducibility*, namely for every $x\in\mathcal{W}$, $t>0$ and every open set $A$ in $\mathcal{W}$, ${\mathbb{P}}[X(t;x)\in A,\tau_x>t]>0$,
$\flat(x_0)=1$ implies that $\flat\equiv1$ on $\mathcal{W}$.
\[p:recurrent\] Consider the family $(X,\tau)$ of processes as above. Assume that, given $x\in\mathcal{W}$, there are a closed set $B_\infty\subset\mathcal{W}$ with non–empty interior and three numbers $p_0\in(0,1)$, $T_0>0$ and $T_1>0$ such that
- ${\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}=\infty,\tau_x=\infty]=0$,
- ${\mathbb{P}}[\tau_y\leq T_0]\geq p_0$ for every $y\in B_\infty$,
where the (discrete) hitting time $\sigma_{B_\infty}^{x,T_1}$ of $B_\infty$, starting from $x$, is defined as $$\sigma_{B_\infty}^{x,T_1}
= \min\{k\geq0: X(kT_1;x)\in B_\infty\},$$ and $\sigma_{B_\infty}^{x,T_1} = \infty$ if the set is empty. Then $${\mathbb{P}}[\tau_x<\infty]\geq\frac{p_0}{1+p_0}.$$
\[r:recurrent\] The first condition in the above proposition can be interpreted as a conditional recurrence: knowing that the solution does not explode, it will visit $B_\infty$ in a finite time with probability $1$.
The first assumption says that ${\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}>n,\tau_x>n]\downarrow0$ as $n\to\infty$. If ${\mathbb{P}}[\tau_x=\infty]=0$, then ${\mathbb{P}}[\tau_x<\infty]=1\geq\tfrac{p_0}{1+p_0}$ and there is nothing to prove. If on the other hand ${\mathbb{P}}[\tau_x=\infty]>0$, then ${\mathbb{P}}[\tau_x>n]>0$ for all $n\geq1$ and, since ${\mathbb{P}}[\tau_x>n]\downarrow{\mathbb{P}}[\tau_x=\infty]$ as $n\to\infty$, we have that $${\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}\leq n|\tau_x>n]
= 1 - \frac{{\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}>n,\tau_x>n]}{{\mathbb{P}}[\tau_x>n]}
\longrightarrow1,
\qquad
n\to\infty.$$ For $n\geq1$, $${\mathbb{P}}[\tau_x>n+T_0]
\leq {\mathbb{P}}[\tau_x>n+T_0,\ \sigma_{B_\infty}^{x,T_1}\leq n]
+ {\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}>n].$$ The strong solution is Markov, hence $$\begin{aligned}
{\mathbb{P}}[\tau_x>n+T_0,\ \sigma_{B_\infty}^{x,T_1}\leq n]
&= \sum_{k=0}^n {\mathbb{P}}[\tau_x>n+T_0,\ \sigma_{B_\infty}^{x,T_1} = k]\\
&\leq (1-p_0){\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}\leq n].
\end{aligned}$$ In conclusion $$\begin{aligned}
{\mathbb{P}}[\tau_x>n+T_0]
&\leq (1-p_0){\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}\leq n] + {\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}>n]\\
&= 1 - p_0{\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}\leq n]\\
&\leq 1 - p_0{\mathbb{P}}[\sigma_{B_\infty}^{x,T_1}\leq n|\tau_x>n]{\mathbb{P}}[\tau_x>n]
\end{aligned}$$ and in the limit as $n\to\infty$, $${\mathbb{P}}[\tau_x=\infty]
\leq 1 - p_0{\mathbb{P}}[\tau_x=\infty],$$ that is ${\mathbb{P}}_x[\tau_x=\infty]\leq\tfrac1{1+p_0}$.
\[c:recurrent\] Assume that there are $p_0\in(0,1)$, $T_0>0$ and $B_\infty\subset\mathcal{W}$ such that the assumptions of the previous proposition hold for every $x\in\mathcal{W}$ (the time $T_1$ may depend on $x$). Then for every $x\in\mathcal{W}$, $${\mathbb{P}}[\tau_x<\infty] = 1.$$
The previous proposition yields $$\sup_{x\in\mathcal{W}}{\mathbb{P}}[\tau_x=\infty]
\leq\frac{1}{1+p_0}
<1,$$ hence the dichotomy of Lemma \[l:solotre\] immediately implies that ${\mathbb{P}}[\tau_x<\infty] = 1$ for every $x\in\mathcal{W}$.
\[x:onedim\] Here we analyse a simple one dimensional example that shows that the fact that blow–up occurs with full probability does depend on the structure of the drift. We remark that the one dimensional case is fully understood (see for instance [@KarShr91]), our proofs below are elementary and mimic the proofs of the next section.
Consider the one–dimensional SDEs $$dX = f_i(X)\,dt + dW,
\qquad i=1,2,$$ with initial condition $X(0) = x\in{\mathbf{R}}$, where $$f_1(x) =
\begin{cases}
x^2, &\quad x\geq0,\\
x, &\quad x<0,
\end{cases}
\qquad\quad
f_2(x) =
\begin{cases}
x^2, &\quad x\geq0,\\
-x, &\quad x<0.
\end{cases}$$ It is easy to see (using the Feller test, see for instance [@KarShr91 Proposition 5.22]) that with the drift $f_1$ the blow–up function $\flat$ defined in the proof of Lemma \[l:solotre\] verifies $0<\flat(x)<1$, while with the drift $f_2$ the blow–up function is $\flat(x)\equiv0$.
In view of the results proved above and the analysis of the next section (see Theorem \[t:blowup\]), we observe that
- with both drifts $f_1,f_2$ and with $B_\infty = \{x\geq1\}$, there are $p_0\in(0,1)$ and $T_0>0$ such that ${\mathbb{P}}[\tau_x\leq T_0]\geq p_0$ for all $x\in B_\infty$ (in other words, the second assumption of Proposition \[p:recurrent\] holds),
- the first assumption of Proposition \[p:recurrent\] holds for $f_2$ but *not* for $f_1$,
- in both cases ${\mathbb{E}}\bigl[\sup_{[0,T]}(X_n)_-^p\bigr]<\infty$ for all $T>0$ and $p\geq1$.
We give an elementary proof of the first statement, which is modeled on the proof which will be used in the next section for the blow–up of the infinite–dimensional problem . Given an initial condition $x\in [1,\infty)$, we prove that $${\mathbb{P}}\Bigl[\Bigl\{\sup_{t\in [0,2]}|W_t|\leq\frac14\Bigr\}\cap\{\tau_x>2\}\Bigr]=0.$$ Indeed, set $Y_t = X_t - W_t$, so that $Y_0=x$ and $$dY
= dX - dW
= (Y + W)^2,$$ in particular $Y_t\geq1$. On the event $\{\sup_{t\in [0,2]}|W_t|\leq\tfrac14\}$, $$\dot Y
\geq Y^2 - 2|W|Y
\geq Y(Y-\frac12)
\geq \frac12Y^2,$$ hence by comparison $Y_t$ (and hence $X_t$) explodes before time $\tfrac{2}{x}\leq 2$.
Blow–up for beta>3 {#s:blowup}
=====================
In this section we first prove that there are sets in the state space which lead to a blow–up with positive probability. The underlying idea is to use again Lemma \[l:crucial\] to adapt the estimates of [@Che08], which work only for positive solutions.
In the second part of the section we show that such sets are achieved with full probability when the blow–up time is conditioned to be infinite. The general result of the previous section immediately implies that blow–up occurs with full probability
Blow–up with positive probability
---------------------------------
Given $\alpha>\beta-2$, $p\in(0,\beta-3)$, $a_0>0$ and $M_0>0$, define the set $$\label{e:bu_set}
B_\infty(\alpha,p,a_0,M_0)
= \{x\in V_\alpha: \|x\|_p\geq M_0\text{{} and {}}
\inf_{n\geq1}\bigl(\lambda_{n-1}^{\beta-2}x_n\bigr)\geq -\nu a_0\}.$$ We will show that for suitable values of $a_0$ and $M_0$ there is a positive probability that the process solution of with initial condition in the above set blows up in finite time.
\[t:blowup\] Let $\beta>3$ and assume and . Given $\alpha\in(\beta-2,\alpha_0+1)$, $p\in(0,\beta-3)$, and $a_0\in(0,\tfrac14]$, there exist $p_0>0$, $T_0>0$ and $M_0>0$ such that for each $x\in B_\infty(\alpha,p,a_0,M_0)$ and for every energy martingale weak solution ${\mathbb{P}}_x$ with initial condition $x$, $${\mathbb{P}}_x[\tau_\infty^\alpha\leq T_0]
\geq p_0.$$
Choose $c_0>0$ such that $c_0\leq a_0$ and $c_0\leq\sqrt{a_0}(1-\sqrt{a_0})$, and consider the random integer $N_{\alpha,c_0}(T_0)$ defined in , where the value of $T_0$ will be specified at the end of the proof. Set $$p_0
= {\mathbb{P}}_x[N_{\alpha,c_0}(T_0)=1],$$ we remark that $p_0>0$ by Lemma \[l:moments\_N\] and its value depends only on the distribution of the solution of , thus is independent from the martingale solution ${\mathbb{P}}_x$. The theorem will be proved if we show that $$\label{e:bu_claim}
{\mathbb{P}}_x[\tau_\infty^\alpha>T_0,\ N_{\alpha,c_0}(T_0)=1]
= 0,$$ indeed $${\mathbb{P}}_x[\tau_\infty^\alpha\leq T_0]
= 1 - {\mathbb{P}}_x[\tau_\infty^\alpha\leq T_0,\ N_{\alpha,c_0}(T_0)>1]
\geq 1 - {\mathbb{P}}_x[N_{\alpha,c_0}(T_0)>1]
= p_0.$$ We proceed with the proof of and we work path–wise on the event $$\Omega(\alpha,T_0)
= \{\tau_\infty^\alpha>T_0\}\cap\{N_{\alpha,c_0}(T_0)=1\}.$$ Let $Z$ be the solution of and $Y = X - Z$ the solution of . On the event $\{\tau_\infty^\alpha>T_0\}$ has a unique solution on $[0,T_0]$, and on the event $\{N_{\alpha,c_0}(T_0)=1\}$ we have that $\lambda_{n-1}^{\beta-2}|Z_n(t)|\leq c_0$ for every $t\in[0,T_0]$ and every $n\geq1$. Set $$\eta_n = X_n - Z_n + a_0\nu\lambda_{n-1}^{2-\beta},$$ then by this position $\eta = (\eta_n)_{n\geq1}$ satisfies the system $$\begin{cases}
\dot\eta_n
= - \nu\lambda_n^2\eta_n
+ a_0\nu^2\lambda^2 \lambda_{n-1}^{4-\beta}
+ \lambda_{n-1}^\beta X_{n-1}^2
- \lambda_n^\beta X_n X_{n+1},\\
\eta_n(0) = X_n(0) + a_0\nu\lambda_{n-1}^{2-\beta}.
\end{cases}
\qquad
n\geq1$$ Moreover, by Lemma \[l:crucial\] (applied with the values of $a_0$ and $c_0$ we have fixed), it turns out that $\eta_n(t;\omega)\geq0$ for each $t\in[0,T_0]$, $n\geq1$ and $\omega\in\Omega(\alpha,T_0)$.
Fix a number $b>0$, which will be specified later in formula , then $$\begin{aligned}
\frac{d}{dt}\bigl(\eta_n^2 + b\eta_n\eta_{n+1}\bigr)
& = - 2\nu\lambda_n^2\eta_n^2 - b\nu(1+\lambda^2)\lambda_n^2\eta_n\eta_{n+1}\\
&\quad + a_0\lambda^2\nu^2(2 + b\lambda^{4-\beta})\lambda_{n-1}^{4-\beta}\eta_n + a_0 b\lambda^2\nu^2\lambda_{n-1}^{4-\beta}\eta_{n+1}\\
&\quad + 2\lambda_{n-1}^\beta X_{n-1}^2\eta_n + b\lambda_{n-1}^\beta X_{n-1}^2 \eta_{n+1} + b\lambda_n^\beta X_n^2\eta_n\\
&\quad - 2\lambda_n^\beta X_n X_{n+1}\eta_n - b\lambda_n^\beta X_n X_{n+1}\eta_{n+1} - b\lambda_{n+1}^\beta\eta_n X_{n+1}X_{n+2}\\
& = { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{1}}\kern-2pt}
}
\xspace
} + { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{2}}\kern-2pt}
}
\xspace
} + { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{3}}\kern-2pt}
}
\xspace
} + { \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{4}}\kern-2pt}
}
\xspace
}.
\end{aligned}$$ By the choice of $a_0$ and $c_0$, we have that $(a_0 + c_0)^2\leq a_0$, hence $${ \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{2}}\kern-2pt}
}
\xspace
}
\geq \lambda^2\nu^2(a_0+c_0)^2(2 + b\lambda^{4-\beta})\lambda_{n-1}^{4-\beta}\eta_n
+ b\lambda^2\nu^2(a_0 + c_0)^2\lambda_{n-1}^{4-\beta}\eta_{n+1}.$$ The third term on the right–hand side of the formula above can be estimated from below using Young’s inequality as $$\begin{aligned}
{ \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{3}}\kern-2pt}
}
\xspace
}
&\geq 2\lambda_{n-1}^\beta X_{n-1}^2\eta_n
+ b\lambda_n^\beta X_n^2\eta_n\\
&\geq \bigl[(1+\lambda^{-2p})\lambda_{n-1}^\beta\eta_{n-1}^2\eta_n - 4\nu^2(a_0+c_0)^2\frac{\lambda^{2\beta+2p-4}}{\lambda^{2p}-1}\lambda_{n-1}^{4-\beta}\eta_n\bigr]\\
&\quad + b\bigl[\lambda_n^\beta\eta_n^3
- 2\lambda^{\beta-2}\nu (a_0+c_0)\lambda_n^2\eta_n^2\bigr]
\end{aligned}$$ since $X_n = \eta_n - (a_0\nu\lambda_{n-1}^{2-\beta} - Z_n)$ and $0\leq (a_0\nu\lambda_{n-1}^{2-\beta} - Z_n)\leq \nu(a_0+c_0)\lambda_{n-1}^{2-\beta}$. Likewise, the fourth term can be estimated from below as $$\begin{aligned}
{ \ensuremath{
\framebox{\tiny\textbf{\kern-2pt\textsf{4}}\kern-2pt}
}
\xspace
}
&\geq - 2\lambda_n^\beta\eta_n^2\eta_{n+1}
- b\lambda_n^\beta\eta_n\eta_{n+1}^2
- b\lambda_{n+1}^\beta\eta_n\eta_{n+1}\eta_{n+2}\\
&\quad - \nu^2(a_0+c_0)^2(2\lambda^2 + b\lambda^{6-\beta})\lambda_{n-1}^{4-\beta}\eta_n
- b\lambda^2\nu^2(a_0+c_0)^2\lambda_{n-1}^{4-\beta}\eta_{n+1},
\end{aligned}$$ where we have dropped the positive terms, using again the fact that $(a_0\nu\lambda_{n-1}^{2-\beta} - Z_n)\geq0$. The three above estimates together yield $$\frac{d}{dt}\bigl(\eta_n^2 + b\eta_n\eta_{n+1}\bigr)
+ 2\nu\lambda_n^2\eta_n^2 + b\nu(1+\lambda^2)\lambda_n^2\eta_n\eta_{n+1}
\geq A_n + B_n + C_n,$$ where $$\begin{aligned}
A_n &= b\lambda_n^\beta\eta_n^3
+ (1+\lambda^{-2p})\lambda_{n-1}^\beta\eta_{n-1}^2\eta_n
- 2\lambda_n^\beta\eta_n^2\eta_{n+1}
- b\lambda_n^\beta\eta_n\eta_{n+1}^2\\
&\quad - b\lambda_{n+1}^\beta\eta_n\eta_{n+1}\eta_{n+2},\\
B_n &= -2b\lambda^{\beta-2}\nu(a_0 + c_0)\lambda_n^2\eta_n^2,\\
C_n &= - 4\nu^2(a_0+c_0)^2\frac{\lambda^{2\beta+2p-4}}{\lambda^{2p}-1}\lambda_{n-1}^{4-\beta}\eta_n.
\end{aligned}$$ By using Young’s inequality in the same way as [@Che08], $$\label{e:Anineq}
\begin{aligned}
A_n
&\geq b\lambda_n^\beta\eta_n^3
+ (1+\lambda^{-2p}) \lambda_{n-1}^\beta\eta_{n-1}^2\eta_n
- 2\lambda_n^\beta\eta_n^2\eta_{n+1}
- \frac12 b\lambda_n^\beta \bigl(\eta_n^2\eta_{n+1} + \eta_{n+1}^3\bigr)\\
&\quad- \frac14 b\lambda_{n+1}^\beta\bigl(2\eta_n^2\eta_{n+1} + \eta_{n+1}^2\eta_{n+2} + \eta_{n+2}^3\bigr)\\
& = b\lambda_n^\beta\eta_n^3
- \frac12 b\lambda_n^\beta \eta_{n+1}^3
- \frac14 b\lambda_{n+1}^\beta\eta_{n+2}^3
+ (1+\lambda^{-2p}) \lambda_{n-1}^\beta\eta_{n-1}^2\eta_n\\
&\quad -\bigl(2+\frac12b+\frac12\lambda^\beta b\bigr)\lambda_n^\beta\eta_n^2\eta_{n+1}
- \frac14 b\lambda_{n+1}^\beta \eta_{n+1}^2\eta_{n+2}
\end{aligned}$$ Summing up the above estimates yields $$\begin{gathered}
\frac{d}{dt}\Bigl[\sum_{n=1}^\infty\lambda_n^{2p}\bigl(\eta_n^2 + b\eta_n\eta_{n+1}\bigr)\Bigr]
+ 2\nu\|\eta\|_{1+p}^2 + {}\\
+ b\nu(1+\lambda^2)\sum_{n=1}^\infty\lambda_n^{2+2p}\eta_n\eta_{n+1}
\geq \sum_{n=1}^\infty \lambda_n^{2p} (A_n + B_n + C_n).
\end{gathered}$$ By using and shifting the summation index we have $$\begin{aligned}
\sum_{n=1}^\infty \lambda_n^{2p} A_n
&\geq b\Bigl(1 - \frac12\lambda^{-(\beta+2p)} - \frac14\lambda^{-(\beta+4p)}\Bigr)\sum_{n=1}^\infty \lambda_n^{\beta+2p}\eta_n^3\\
&\quad + \Bigl[\lambda^{2p} - 1 - b\Bigl(\frac12+\frac12\lambda^\beta+\frac14\lambda^{-2p}\Bigr)\Bigr]\sum_{n=1}^\infty\lambda_n^{\beta+2p}\eta_n^2\eta_{n+1}\\
& = k_1 \sum_{n=1}^\infty \lambda_n^{\beta+2p}\eta_n^3,
\end{aligned}$$ where we have chosen $b$ so that $$\label{e:choosea}
\lambda^{2p} - 1 - b\Bigl(\frac12+\frac12\lambda^\beta+\frac14\lambda^{-2p}\Bigr)
= 0.$$ The other two terms are simple, indeed $$\sum_{n=1}^\infty \lambda_n^{2p} B_n
= -2b\lambda^{\beta-2}\nu(a_0+c_0)\sum_{n=1}^\infty\lambda_n^{2+2p}\eta_n^2
= - k_2 \|\eta\|_{1+p}^2,$$ and, by the Cauchy–Schwarz inequality, $$\begin{aligned}
\sum_{n=1}^\infty \lambda_n^{2p} C_n
&= -4\nu^2(a_0+c_0)^2\frac{\lambda^{3\beta+2p-8}}{\lambda^{2p}-1}\sum_{n=1}^\infty\lambda_n^{3-\beta+p}(\lambda_n^{1+p}\eta_n)\\
&\geq -4\nu^2(a_0+c_0)^2\frac{\lambda^{3\beta+2p-8}}{\lambda^{2p}-1}
\bigl(\lambda^{2(3-\beta+p)} - 1\bigr)^{-\frac12}\|\eta\|_{1+p}\\
& = - k_3 \|\eta\|_{1+p},
\end{aligned}$$ we have used the fact that $p<\beta-3$. Since on the other hand $$\begin{aligned}
2\nu\|\eta\|_{1+p}^2
+ b\nu(1+\lambda^2)\sum_{n=1}^\infty\lambda_n^{2+2p}\eta_n\eta_{n+1}
&\leq \nu\bigl(2 + b(1+\lambda^2)\lambda^{-1-p}\bigr)\|\eta\|_{1+p}^2\\
&= k_4\|\eta\|_{1+p}^2,
\end{aligned}$$ and $$\|\eta\|_{1+p}^2
= \sum_{n=1}^\infty\bigl(\lambda_n^{\frac13(\beta+2p)}\eta_n\bigr)^2\lambda_n^{-\frac23(\beta-3-p)}
\leq \Bigl(\frac1{k_5}\sum_{n=1}^\infty \lambda_n^{\beta+2p}\eta_n^3\Bigr)^{\frac23},$$ all the estimates obtained so far together yield $$\dot H + k_4\psi \geq k_1k_5\psi^{\frac32} - k_2\psi - k_3\sqrt\psi,$$ where $$H(t)
= \sum_{n=1}^\infty\lambda_n^{2p}\bigl(\eta_n^2 + b\eta_n\eta_{n+1}\bigr),
\qquad
\psi(t)
= \|\eta\|_{1+p}^2.$$ Finally, $H\leq (1+b\lambda^{-p})\psi = k_6 \psi$, and it is easy to show by a simple argument (following for instance the one in [@Che08]) that if $$H(0)
> M_0^2
:= \frac{k_6}{k_1 k_5}\bigl(k_4+k_2+\sqrt{(k_4+k_2)^2 + 2k_1k_3k_5}\bigr)
\quad\text{and}\quad
T_0
> \frac{4k_6^{\frac32}}{k_1k_5\sqrt{H(0)}},$$ the solution of the differential inequality given above for $H$ becomes infinite before time $T_0$.
Ineluctable occurrence of the blow–up
-------------------------------------
The theorem in the previous section has shown that if the initial condition is *not too negative* and the noise is *not too strong*, so that the the process is not kicked away from the quasi–positive area, then the deterministic dynamics dominates and the process explodes. In this section we show that the sets that lead to blow–up are conditionally recurrent, in the sense of Proposition \[p:recurrent\] (see Remark \[r:recurrent\]).
\[t:noway\] Let $\beta>3$ and assume and . Assume moreover that the set $\{n\geq1: \sigma_n\neq0\}$ is non–empty. Given $\alpha\in(\beta-2,1+\alpha_0)$, for every $x\in V_\alpha$ and every energy martingale solution ${\mathbb{P}}_x$ with initial condition $x$, $${\mathbb{P}}_x[\tau_\infty^\alpha<\infty] = 1.$$
Before proving the theorem, we need a few preliminary results. Our strategy is to use Corollary \[c:recurrent\] to prove that blow–up happens with full probability. Hence we need to prove that sets satisfy the assumptions of the corollary. The idea is to find conditions on the size of the solution in $H$ and on the structure of the perturbation, so that the system is led to a set . Lemma \[l:contraction\] shows that the negative part of the solution becomes small, while Lemma \[l:expansion\] shows that the size of the solution becomes large. Finally, Lemma \[l:recurrent\] shows that the conditions found happen with full probability, conditional to the absence of blow–up.
We need a slight improvement of Lemma \[l:crucial\]. That lemma showed that if the noise is *small* from mode $N_{\alpha,c_0}(T)$ on, and if the initial condition is *not too negative* from mode $n_0$ on, then the system is *not too negative* from mode $\max(n_0,N_{\alpha,c_0}(T))$ on. In the next lemma we show that after a short time, which depends only on the size of the initial condition *in $H$*, modes becomes *not too negative* and the above result holds from mode $N_{\alpha,c_0}(T)$ on, regardless of the value of $n_0$.
We prove the result on the event $\{N_{\alpha,c_0}(T)=1\}$, because the control of the $H$ norm, given in the next lemma, is simpler.
\[l:Hcontrol\] Let $\beta>3$ and assume and . There exists ${\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}>0$ such that for $\alpha\in(\beta-2,1+\alpha_0)$, for every $x\in V_\alpha$, every energy martingale solution ${\mathbb{P}}_x$ starting at $x$, every $T>0$ and every $c_0>0$ with $4 c_0(1+\lambda^{\beta-3})\leq 1$, $$\sup_{[0,T]} \|X(t)\|_H
\leq \|x\|_H + {\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}\nu,$$ ${\mathbb{P}}_x$–[a. s.]{} on the event $\{\tau_\infty^\alpha> T\}\cap\{N_{\beta-2,c_0}(T)=1\}$.
As already remarked, since we work on $\{\tau_\infty^\alpha> T\}$, problem has a unique solution, so we can work directly on $Y$ (the rigorous proof proceeds through Galerkin approximations). We know that $\lambda_n^{\beta-2}|Z_n(t)|\leq c_0\nu$ for $t\in[0,T]$ and $n\geq1$, hence $$\begin{aligned}
\frac{d}{dt}\|Y\|_H^2 + 2\nu\|Y\|_1^2
&\leq 2\sum_{n=1}^\infty \lambda_n^\beta(Y_nY_{n+1}Z_n + Y_{n+1}Z_n^2 - Y_n^2Z_{n+1} - Y_nZ_nZ_{n+1})\\
&\leq 2c_0\nu(1+\lambda^{\beta-3})\|Y\|_1^2
+ 2c_0^2\nu^2\frac{\lambda^{\beta-2}}{\lambda^{\beta-3}-1}\|Y\|_1\\
&\leq 4c_0\nu(1+\lambda^{\beta-3})\|Y\|_1^2
+ \frac{\lambda^{2\beta-4}}{2(\lambda^{\beta-3}-1)}c_0^3\nu^3.
\end{aligned}$$ Using the assumption on $c_0$ and the fact that $\|Y\|_1\geq\lambda\|Y\|_H$, we finally obtain $$\frac{d}{dt}\|Y\|_H^2 + \nu\lambda^2\|Y\|_H^2
\leq k_0 c_0^3\nu^3,$$ where the value of $k_0$ follows from the inequality above and depends only on $\beta$. By integrating the differential inequality, we get $$\|Y(t)\|_H^2
\leq \|x\|_H^2 + \frac{k_0}{\lambda^2}c_0^3\nu^2,
\quad t\in[0,T],$$ and in conclusion for $t\in[0,T]$, $$\|X(t)\|_H
\leq \|Y(t)\|_H + \|Z(t)\|_H
\leq \|x\|_H + \bigl(\sqrt{k_0}\lambda^{-1} + (1-\lambda^{4-2\beta})^{-\frac12}\bigr)\nu,$$ where we have used the fact that $c_0\leq 1$.
\[l:contraction\] Let $\beta>3$ and assume and . For every $M>0$, $a_0\in(0,\frac14]$ and $c_0<a_0$, with $4 c_0(1+\lambda^{\beta-3})\leq 1$, there exists $T_M>0$ such that for every $x\in V_\alpha$, with $\alpha\in(\beta-2,1+\alpha_0)$ and $\|x\|_H\leq M$, and every energy martingale solution ${\mathbb{P}}_x$, $$\inf_{n\geq 1} \bigl(\lambda_{n-1}^{\beta-2} X_n(T_M)\bigr)
\geq -(a_0 + c_0)\nu,$$ ${\mathbb{P}}_x$–[a. s.]{} on the event $\{\tau_\infty^\alpha> T_M\}\cap\{N_{\beta-2,c_0}(T_M)=1\}$.
Let $n_0$ be the first integer such that $\inf_{n\geq n_0}\lambda_n^{\beta-2}x_n\geq -a_0\nu$. By the choice of $c_0$ and $a_0$, we know that $c_0<\sqrt{a_0}(1-\sqrt{a_0})$, hence Lemma \[l:crucial\] implies that $\lambda_n^{\beta-2}Y_n(t)\geq -a_0\nu$ holds for every $t\in[0,T_M]$ and every $n\geq n_0$. If $n_0=1$ there is nothing to prove, so we consider the case $n_0>1$.
The idea here is to show that the negative part of the mode $n_0-1$ becomes closer to $0$ within a time $T_{n_0-1}$. After this time, again by Lemma \[l:crucial\], the negative part of the mode $n_0-1$ stays small, and we can apply the same contraction idea on the negative part of the component $n_0-2$, which then will become small as well within a time $T_{n_0-2}$, and so on. The sequence of time intervals turns out to be summable and dependent only on the size of the initial condition in $H$.
By these considerations it is sufficient to prove the following statement:
> given $n>1$, if we know that for a time $t_0>0$, $$\sup_{k\geq 1}\sup_{[t_0,T]}\lambda_{k-1}^{\beta-2}|Z_k|\leq c_0\nu
> \quad\text{and}\quad
> \sup_{[t_0,T]}\lambda_n^{\beta-2}(Y_{n+1})_-\leq a_0\nu$$ then at time $t_0 + T_n$, where $$T_n(\|x\|_H,c_0,a_0)
> = \frac2\nu(\beta-2)\log\lambda\frac{n-1}{\lambda_n^2}
> + \frac2\nu\lambda_n^{-2}\log\Bigl(1\vee\frac{\|x\|_H+{\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}\nu}{(a_0-c_0)\nu}\Bigr),$$ we have that $$Y_n(t_0 + T_n)\geq -a_0\nu\lambda_{n-1}^{2-\beta}.$$
Before proving the claim, we notice that $\sum_n T_n<\infty$, hence we can choose $$T_M
= \sum_{n=1}^\infty T_n(M,c_0,a_0).$$ We conclude the proof of the lemma by showing the above claim. Set $$\eta_n = Y_n + c_0\nu\lambda_{n-1}^{2-\beta},$$ then $X_n = \eta_n - (c_0\nu\lambda_{n-1}^{2-\beta} - Z_n)$ and $$\begin{aligned}
\dot\eta_n
&= -\nu\lambda_n^2\eta_n
+ c_0\nu^2\lambda^2\lambda_{n-1}^{4-\beta}
+ \lambda_{n-1}^\beta X_{n-1}^2
- \lambda_n^\beta X_n X_{n+1}\\
&\geq -(\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n
+ c_0\nu^2\lambda^2\lambda_{n-1}^{4-\beta}
+ \lambda_n^\beta (c_0\nu\lambda_{n-1}^{2-\beta} - Z_n)X_{n+1}\\
&\geq -(\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n
+ c_0\nu^2\lambda^2\lambda_{n-1}^{4-\beta}
- \lambda_n^\beta (c_0\nu\lambda_{n-1}^{2-\beta} - Z_n)(X_{n+1})_-.
\end{aligned}$$ By the assumption of the claim, $(X_{n+1})_-\leq(a_0+c_0)\nu\lambda_n^{2-\beta}$ and $(c_0\nu\lambda_{n-1}^{2-\beta} - Z_n)\leq 2c_0\nu\lambda_{n-1}^{2-\beta}$, hence $$\begin{aligned}
\dot\eta_n
&\geq - (\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n
+ c_0\lambda^2\nu^2\bigl(1 - 2(a_0+c_0)\bigr)\lambda_{n-1}^{4-\beta}\\
&\geq - (\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n,
\end{aligned}$$ which implies that for $t\geq t_0$, $$\begin{gathered}
\eta_n(t)
\geq \eta_n(t_0)\exp\Bigl(-\int_{t_0}^t(\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\,ds\Bigr)\geq\\
\geq -(\eta_n(t_0))_-\exp\Bigl(-\int_{t_0}^t(\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\,ds\Bigr)
\geq -(\eta_n(t_0))_-{\operatorname{e}}^{-\frac12\nu\lambda_n^2(t-t_0)},
\end{gathered}$$ since, using the fact that $a_0+c_0\leq\tfrac12$, $$\nu\lambda_n^2 + \lambda_n^\beta X_{n+1}
\geq \nu\lambda_n^2 - \nu(a_0+c_0)\lambda_n^2
= \nu\lambda_n^2\bigl(1 - (a_0+c_0)\bigr)
\geq \frac12\nu\lambda_n^2.$$ Finally, by Lemma \[l:Hcontrol\], $$(\eta_n(t_0))_-
\leq (Y_n(t_0))_-
\leq |Y_n(t_0)|
\leq \|Y(t_0)\|_H
\leq \|x\|_H + {\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}\nu,$$ and it is elementary now to check that at time $t_0+T_n$, $$\begin{gathered}
Y_n(t_0+T_n)
= \eta_n(t_0+T_n) - c_0\nu\lambda_{n-1}^{2-\beta}\geq\\
\geq -(\|x\|_H + {\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}){\operatorname{e}}^{-\frac12\nu\lambda_n^2 T_n} - c_0\nu\lambda_{n-1}^{2-\beta}
\geq -a_0\nu\lambda_n^{2-\beta},
\end{gathered}$$ which concludes the proof of the claim.
In order to show that the hitting time for sets like , where there is a uniform estimate on the probability of blow–up, is finite, we also need to ensure that the size of the solution is large enough, while being *not too negative*.
At this stage the noise is crucial, although one randomly perturbed component is enough for our purposes. The underlying ideas of the following lemma come from control theory, although we do not need sophisticated results like [@Shi06; @Rom04], because a quick and strong impulse turns out to be sufficient.
\[l:expansion\] Under the assumptions of Theorem \[t:noway\], let $m$ be the smallest element of the set $\{n\geq1:\sigma_n\neq0\}$. Given $M_1>0$, $M_2>0$, and $a_0,a_0',c_0>0$ such that $c_0<a_0<a_0'<\tfrac14$ and $c_0+a_0<a_0'$, for every $X(0)\in V_\alpha$, with $\|X(0)\|_H\leq M_1$ and $\inf_{n\geq 1}\lambda_{n-1}^{\beta-2} X_n(0)\geq -a_0\nu$, there exists $T>0$, whose value depends on $M_1$, $M_2$, $c_0$, $a_0$, $a_0'$ and $m$, such that
- $\lambda_{n-1}^{\beta-2} X_n(t) \geq -(a_0' + c_0)\nu$ for every $n\geq1$ and $t\in[0,T]$,
- $\|X(T)\|_H\geq M_2$,
on the event $$\{\tau_\infty^\alpha>T_M\}
\cap\{\sup_{[0,T]}\lambda_{n-1}^{\beta-2}|Z_n(t)|\leq c_0\nu\text{ for }n\neq m\}
\cap\{\sup_{[0,T]}\lambda_{m-1}^{\beta-2}|Z_m(t) - \psi(t)|\leq c_0\nu\},$$ where $\psi:[0,T]\to{\mathbf{R}}$ is an non–decreasing continuous function such that $\psi(0)=0$ and $\psi(T)$ large enough depending on the above given data (its value is given in the proof).
We work on the event given in the statement of the proof.
*Step 1: estimate in $H$.* We first give an estimate of $X$ in $H$. Set $\widetilde\psi=\sup_{[0,T]}\|Z_m\|_H\leq\psi(T) + c_0\nu$, then $$\begin{aligned}
\frac{d}{dt}\|Y\|_H^2 + 2\nu\|Y\|_1^2
&\leq 2\sum_{n=1}^\infty \lambda_n^\beta X_n(Z_n Y_{n+1} - Y_nZ_{n+1})\\
&\leq 2\sum_{n=1}^\infty \lambda_n^\beta
\bigl(|Z_n Y_n Y_{n+1}| + Z_n^2 |Y_{n+1}| + |Z_{n+1}|Y_n^2 + |Z_n Z_{n+1}Y_n|\bigr)\\
&\quad + 2\lambda_{m-1}^\beta \bigl(|Z_m|Y_{m-1}^2 + |Z_{m-1}Z_mY_{m-1}|\bigr)\\
&\quad + 2\lambda_m^\beta \bigl(||Z_mY_mY_{m+1} + Z_m^2 |Y_{m+1}| + |Z_mZ_{m+1}Y_m|\bigr)\\
&\leq 4c_0\nu(1+\lambda^{\beta-3})\|Y\|_1^2 + k_0c_0^3\nu^3\\
&\quad + \lambda_m^\beta(2\widetilde\psi^2 + 4\widetilde\psi+ 4c_0\nu\lambda^{\beta-3}\widetilde\psi)(1+\|Y\|_H^2)\\
&\leq \nu\|Y\|_1^2 + k_0\nu^3 + 16\lambda_m^\beta(1+\nu)(1+\widetilde\psi^2)(1+\|Y\|_H^2),
\end{aligned}$$ where we have estimated the terms without $Z_m$ as in Lemma \[l:Hcontrol\] (the constant $k_0$ is the same of that lemma) and the terms with $Z_m$ bounding the components of $Y$ with $\|Y\|_H$. If $k_1=k_0\nu^3$, $k_2=16\lambda_m^\beta(1+\nu)$ and $M_3(T,\psi(T))^2 = (M_1^2 + k_1/k_2)\exp\bigl(k_2T(1+\widetilde\psi^2)\bigr)$, it follows from Gronwall’s lemma that $\sup_{[0,T]}\|Y(t)\|_H^2\leq M_3(T,\psi(T))^2$. Since on the given event we have that $\|Z(t)\|_H\leq \lambda^{2-\beta}(\lambda^{2-\beta}-1)^{-1} + c_0\nu + \psi(T)$ for every $t\in[0,T]$, we finally have that $$\begin{aligned}
\sup_{[0,T]}\|X(t)\|_H
&\leq c_0\nu
+ \frac{\lambda^{2-\beta}}{\lambda^{2-\beta}-1}
+ \psi(T)
+ \bigl(M_1 + \sqrt{\tfrac{k_1}{k_2}}\bigr){\operatorname{e}}^{k_2 T(1+(\psi(T)+c_0\nu)^2)}\\
&=: M_4(T,\psi(T))
\end{aligned}$$
*Step 2: Large size at time $T$.* Using the previous estimate we have $$\begin{aligned}
X_m(t)
&= {\operatorname{e}}^{-\nu\lambda_m^2t}X_m(0)
+ Z_m(t)
+ \int_0^t {\operatorname{e}}^{-\nu\lambda_m^2(t-s)}\bigl(
\lambda_{m-1}^\beta X_{m-1}^2 - \lambda_m^\beta X_m X_{m+1}\bigr)\,ds\\
&\geq - a_0\nu\lambda_{m-1}^{2-\beta}
+ \bigl(\psi(t) - c_0\nu\lambda_{m-1}^{2-\beta}\bigr)
- \lambda_m^\beta t \sup_{[0,T]}\|X\|_H^2,
\end{aligned}$$ hence for $t=T$ we have $$X_m(T)
\geq \psi(T) - \nu - \lambda_m^\beta M_4(T,\psi(T)) T.$$ We can choose $\psi(T) = M_2 + 2\nu$ and notice that $M_4(T,M_2+2\nu) T\to 0$ as $T\downarrow0$, hence we can choose $T$ small enough so that $\lambda_m^\beta M_4(T,\psi(T)) T\leq\nu$, to obtain $X_m(T)\geq M_2$ and in conclusion $\|X(T)\|_H\geq M_2$.
*Step 3: Bound from below for $n=m$.* The choice of $\psi(T)$ and the computation in the above step yield $$\lambda_{m-1}^{\beta-2} X_m(t)
\geq - (a_0 + c_0)\nu
- \lambda^{2-\beta}M_4(T, M_2+2\nu)\lambda_m^{2\beta-2} T,$$ since $\psi$ is non–negative. By assumption we have that $a_0+c_0<a_0'$, hence, possibly fixing a smaller value of $T$ than the one chosen in the previous step, we can ensure that $X_m$ has $-a_0'\nu\lambda_{m-1}^{2-\beta}$ as lower bound on $[0,T]$.
*Step 4: Bound from below for $n\neq m$.* If $n>m$, the proof proceeds as in Lemma \[l:crucial\], since $X_m$ appears in the system of equations for $(Y_n)_{n>m}$ only through the *positive* term $\lambda_m^\beta X_m^2$ in the equation for the $(m+1)^\text{th}$ component (and since $c_0<\sqrt{a_0'}(1-\sqrt{a_0'})$ by the choice of $c_0$, $a_0'$).
If $n<m$, the proof follows by finite induction. Indeed, the lower bound is true for $n=m$ by the previous step. Next we prove that if $\lambda_n^{2-\beta} Y_{n+1}(t)\geq -a_0'\nu$ for all $t\in[0,T]$, then $\lambda_{n-1}^{2-\beta} Y_n\geq -a_0'\nu$. We prove this similarly to the proof of Lemma \[l:contraction\]. Indeed, set $\eta_n = Y_n + a_0'\nu\lambda_{n-1}^{2-\beta}$, then using only that $\lambda_{n-1}^{\beta-2}|Z_n(t)|\leq c_0\nu$ and $(X_{n+1}(t))_-\leq (a_0'+c_0)\nu\lambda_n^{2-\beta}$ for $t\in[0,T]$, $$\begin{aligned}
\dot\eta_n
&\geq - (\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n
+ a_0'\nu^2\lambda^{\beta-2}\lambda_n^{4-\beta}
- \lambda_n^\beta(a_0'\nu\lambda_{n-1}^{2-\beta} - Z_n)(X_{n+1})_-\\
&\geq - (\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n
+ \nu\lambda^{\beta-2}\bigl(a_0' - (a_0'+c_0)^2\bigr)\lambda_n^{4-\beta}\\
&\geq - (\nu\lambda_n^2 + \lambda_n^\beta X_{n+1})\eta_n,
\end{aligned}$$ since $c_0<a_0'<\tfrac14$, hence $c_0<\sqrt{a_0'}(1-\sqrt{a_0'})$. The fact that $\eta_n(0)\geq0$ implies that $\eta_n(t)\geq0$.
Lemma \[l:contraction\] and Lemma \[l:expansion\] suggest how to have a *well-behaved* noise, namely a random perturbation that leads the system from a ball in $H$ to a pre–blow–up set . Let $c_0>0$, $t_0>0$, $T_c>0$, $T_e>0$ and $\psi:[0,T_e]\to{\mathbf{R}}$ be a non–negative non–increasing function, and define the event $$\mathcal{N}(t_0;c_0,T_c,T_e,\psi)
= \mathcal{N}_c(c_0,t_0,T_c)\cap \mathcal{N}_e(c_0,t_0+T_c,T_e,\psi)$$ where $$\mathcal{N}_c(c_0,t_0,T_c)
= \bigl\{
\lambda_{n-1}^{\beta-2}|Z_n^c(t)|\leq c_0\nu\text{ for all }n\geq1\text{ and }t\in[t_0,t_0+T_c]
\bigr\}\\$$ and $$\begin{gathered}
\mathcal{N}_e(c_0,t_0+T_c,T_e,\psi)
= \Bigl\{
\sup_{[t_0+T_c,t_0+T_c+T_e]}\lambda_{m-1}^{\beta-2}|Z_m^e(t) - \psi_{t_0+T_c}(t)|\leq c_0\nu,\\
\lambda_{n-1}^{\beta-2}|Z_n^e(t)|\leq c_0\nu\text{ for all }n\neq m\text{ and }t\in[t_0+T_c,t_0+T_c+T_e]
\Bigr\},\end{gathered}$$ $\psi_s:[s,s+T_e]\to{\mathbf{R}}$ is defined for $s\geq0$ as $\psi_s(t) = \psi(t-s)$, for $t\in[s,s+T_e]$, $m$ is the smallest integer of the set $\{n:\sigma_n\neq0\}$, and for every $n\geq1$, $$\begin{gathered}
Z^c_n(t)
= \sigma_n\int_{t_0}^t {\operatorname{e}}^{-\nu\lambda_n^2(t-t_0)}\,dW_n,
\qquad t\in[t_0,t_0+T_c],\\
Z^e_n(t)
= \sigma_n\int_{t_0+T_c}^t {\operatorname{e}}^{-\nu\lambda_n^2(t-t_0-T_c)}\,dW_n,
\qquad t\in[t_0+T_c,t_0+T_c+T_e].
\end{gathered}$$ Given a weak martingale solution ${\mathbb{P}}_x$ with initial condition $x$, the following facts are straightforward:
- $\mathcal{N}_c(c_0,t_0,T_c)$ and $\mathcal{N}_e(c_0,t_0+T_c,T_e,\psi)$ are independent,
- they both have positive probability (Lemma \[l:moments\_N\] ensures that $\mathcal{N}_c$ has positive probability and the same methods can be used for $\mathcal{N}_e$),
- the value of their probability is independent of $t_0$,
- if $t_0,T_c,T_e$ and $t_0'$ are given such that $t_0+T_c+T_e\leq t_0'$, then $\mathcal{N}(t_0;c_0,T_c,T_e,\psi)$ and $\mathcal{N}(t_0';c_0,T_c,T_e,\psi)$ are independent.
The basic idea for the proof of Theorem \[t:noway\] is to show that a pre–blow–up set is recurrent for the system, knowing that there is no blow–up. Since a *well-behaved* randomness leads the system from a ball in $H$ to a pre–blow–up set, it is sufficient to show that the fact that the system hits a ball in $H$, of some radius, and *afterwards* the randomness is *well-behaved* happens with probability one.
\[l:recurrent\] Let $\beta>3$, assume and and consider $\alpha\in(\beta-2,\alpha_0+1)$. There exists ${\ensuremath{c_{\text{\ref{l:recurrent}}}}}>0$ such that if $M>0$, $T_c>0$, $T_e>0$, $c_0>0$, and $\psi:[0,T_e]\to{\mathbf{R}}$ is a non–decreasing and non–negative function and $$\frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2} + {\operatorname{e}}^{-\nu\lambda^2(T_c+T_e)}
< 1,$$ then for every $x\in V_\alpha$ and every energy martingale solution ${\mathbb{P}}_x$ starting at $x$, $${\mathbb{P}}_x\Bigl[\{\tau_\infty^\alpha=\infty\}
\cap\bigcap_{k\geq1}\Bigr(
\{\|X(kT)\|_H\leq M\}\cap\mathcal{N}(k T;c_0,T_c,T_e,\psi)\Bigl)^c\Bigr]
= 0,$$ where $T = T_c + T_e$.
We first obtain a quantitative estimate on the return time in balls of $H$ of the Markov process $X^R(\cdot;x)$, solution of problem with initial condition $x\in V_\alpha$, which has been introduced in the proof of Theorem \[t:strong\], and then get the same estimate for the strong solution. This will allow us to prove the lemma.
*Step 1.* It is straightforward to prove by Itô’s formula (applied to $\|X^R(t;x)\|_H^2$) and Gronwall’s lemma that $$\label{e:expo}
{\mathbb{E}}[\|X^R(t;x)\|_H^2]
\leq \|x\|_H^2{\operatorname{e}}^{-2\nu\lambda^2t} + {\ensuremath{c_{\text{\ref{l:recurrent}}}}},$$ where ${\ensuremath{c_{\text{\ref{l:recurrent}}}}}=(2\nu\lambda^2)^{-1}\sum_{n=1}^\infty\sigma_n^2$, and where the last series is convergent due to and to the assumption $\alpha_0>\beta-3$.
*Step 2.* We use the previous estimate to show that $$\label{e:decay}
{\mathbb{P}}\bigl[\|X^R(kT;x)\|_H\geq M\text{ for }k=1,\dots,n\bigr]
\leq \Bigl({\operatorname{e}}^{-\nu\lambda^2T} + \frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2}\Bigr)^{n-1}.$$ We proceed for instance as in [@Deb11 Lemma III.2.4], we detail the proof for the sake of completeness. Define, for $k$ integer, $C_k = \{\|X^R(kT;x)\|_H\geq M\}$ and $B_k = \bigcap_{j=0}^k C_j$, and set $\alpha_k = {\mathbb{E}}[{\mathbbm{1}}_{B_k}\|X^R(kT;x)\|_H^2]$ and $p_k = {\mathbb{P}}[B_k]$. By the Markov property, Chebychev’s inequality and , $${\mathbb{P}}[C_{k+1}|{\mathscr{F}}_{kT}]
\leq \frac1{M^2}{\operatorname{e}}^{-\nu\lambda^2T}\|X^R(kT;x)\|_H^2
+ \frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2},$$ hence $$p_{k+1}
= {\mathbb{E}}\bigl[{\mathbbm{1}}_{B_k}{\mathbb{P}}[C_{k+1}|{\mathscr{F}}_{kT}]\bigr]
\leq \frac1{M^2}{\operatorname{e}}^{-\nu\lambda^2T}\alpha_k
+ \frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2}p_k.$$ On the other hand, by integrating on $B_k$, we get $$\alpha_{k+1}
\leq {\mathbb{E}}[{\mathbbm{1}}_{B_k}\|X^R((k+1)T;x)\|_H^2]
\leq {\operatorname{e}}^{-\nu\lambda^2T}\alpha_k + {\ensuremath{c_{\text{\ref{l:recurrent}}}}}p_k$$ If $(\widetilde\alpha_k)_{k\in{\mathbf{N}}}$ and $(\widetilde p_k)_{k\in{\mathbf{N}}}$ are the solutions to the recurrence system $$\begin{cases}
\widetilde\alpha_{k+1}
= {\operatorname{e}}^{-\nu\lambda^2T}\widetilde\alpha_k + {\ensuremath{c_{\text{\ref{l:recurrent}}}}}\widetilde p_k,\\
\widetilde p_{k+1}
= \frac1{M^2}{\operatorname{e}}^{-\nu\lambda^2T}\widetilde\alpha_k + \frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2}\widetilde p_k,
\end{cases}
\qquad k\geq1,$$ with $\widetilde p_1 = p_1$ and $\widetilde\alpha_1 = \alpha_1$, then $\alpha_k\leq\widetilde\alpha_k$ and $p_k\leq\widetilde p_k$ for all $k\geq1$ and easily follows, since $\widetilde{\alpha}_k = M^2\widetilde p_k$ for $k\geq2$.
*Step 3.* By the previous step, $$\begin{aligned}
\lefteqn{{\mathbb{P}}[\|X(kT;x)\|_H\geq M, k=1,\dots,n,\tau_x^\alpha=\infty]}\qquad\\
&\leq\lim_{R\uparrow\infty}{\mathbb{P}}[\|X(kT;x)\|_H\geq M, k=1,\dots,n,\tau_x^{\alpha,R}>nT]\\
&= \lim_{R\uparrow\infty}{\mathbb{P}}[\|X^R(kT;x)\|_H\geq M, k=1,\dots,n,\tau_x^{\alpha,R}>nT]\\
&\leq\limsup_{R\uparrow\infty}{\mathbb{P}}[\|X^R(kT;x)\|_H\geq M, k=1,\dots,n]\\
&\leq\Bigl({\operatorname{e}}^{-\nu\lambda^2T} + \frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2}\Bigr)^{n-1},
\end{aligned}$$ since $\tau_x^\alpha = \sup_{R>0}\tau_x^{\alpha,R}$. Hence, if we define the (discrete) hitting time $K_1 = \min\{k\geq0: \|X(kT;x)||_H\leq M\}$ of the ball $B_M(0)$ in $H$ (and $K_1=\infty$ if the set is empty), then $K_1<\infty$ on the event $\{\tau_x^\alpha=\infty\}$. Likewise, define the sequence of (discrete) return times $K_2 = \min\{k>K_1: \|X(kT;x)||_H\leq M\}$, …, $K_j = \min\{k>K_{j-1}: \|X(kT;x)||_H\leq M\}$, …, with the understanding that a return time is infinite if the corresponding set is empty (this in particular happens if some previous return time is already infinite). The previous step immediately implies that $K_j<\infty$ for each $j\geq1$ on the event $\{\tau_x^\alpha=\infty\}$.
*Step 4.* Consider for $k\geq 1$ the events $\mathcal{N}_k = \mathcal{N}(kT;c_0,T_c,T_e,\psi)$. By the considerations we have stated above, we know that ${\mathbb{P}}[\mathcal{N}_k]$ is constant in $k$, and we set $p={\mathbb{P}}[\mathcal{N}_k]$. Moreover, by the choice of $T$, it turns out that $\mathcal{N}_1, \mathcal{N}_2, \dots, \mathcal{N}_k, \dots$ are all independent. Set $\mathcal{N}_\infty = \emptyset$ and define the time $$L_0
= \min\{j\geq1: {\mathbbm{1}}_{\mathcal{N}_{K_j}} = 1\},$$ and $L_0 = \infty$ if the set is empty. Notice that if $L_0$ is finite, then $\|X(K_{L_0}T;x)\|_H\leq M$ and the random perturbation is *well-behaved* in the time interval $[K_{L_0}T, K_{L_0}T + T_c + T_e]$. Hence the lemma is proved if we show that $$\label{e:claim}
{\mathbb{P}}[L_0 = \infty, \tau_x^\alpha=\infty]
= 0.$$
*Step 5.* Given an integer $\ell\geq1$, we have that $$\begin{aligned}
{\mathbb{P}}[L_0>\ell, \tau_x^\alpha=\infty]
&= {\mathbb{P}}[\mathcal{N}_{K_1}^c\cap\dots\mathcal{N}_{K_\ell}^c\cap\{\tau_x^\alpha=\infty\}]\\
&= \sum_{k_1=1}^\infty\dots\sum_{k_\ell=k_{\ell-1}+1}^\infty {\mathbb{P}}[S_\ell(k_1,\dots,k_l)\cap\{\tau_x^\alpha=\infty\}],
\end{aligned}$$ where we have set $$S_\ell(k_1,\dots,k_\ell)
= \mathcal{N}_{K_1}^c\cap\dots\mathcal{N}_{K_\ell}^c
\cap\{K_1=k_1,\dots,K_\ell=k_\ell\}.$$ Notice that $S_{\ell}(k_1,\dots,k_\ell)\in{\mathscr{F}}_{(k_\ell+1)T}$, hence by the Markov property, $$\begin{aligned}
\lefteqn{{\mathbb{P}}[S_\ell(k_1,\dots,k_l)\cap\{\tau_x^\alpha>(k_\ell+1)T\}]}\qquad\\
&= {\mathbb{E}}\bigl[{\mathbbm{1}}_{S_{\ell-1}(k_1,\dots,k_{\ell-1})}
{\mathbbm{1}}_{\{\tau_x^\alpha>(k_{\ell-1}+1)T\}} {\mathbbm{1}}_{\{K_\ell=k_\ell\}}
{\mathbb{P}}[\mathcal{N}_{k_\ell}^c\cap\{\tau_{X(k_\ell T;x)}^\alpha>T\}|{\mathscr{F}}_{k_\ell T}]\bigr]\\
&\leq(1-p){\mathbb{P}}[S_{\ell-1}(k_1,\dots,k_{\ell-1})\cap
\{\tau_x^\alpha>(k_{\ell-1}+1)T\}\cap\{K_\ell=k_\ell\}].
\end{aligned}$$ By summing up over $k_\ell$, we have $$\begin{gathered}
\sum_{k_\ell=k_{\ell-1}+1}^\infty {\mathbb{P}}[S_\ell(k_1,\dots,k_\ell)\cap\{\tau_x^\alpha>(k_\ell+1)T\}]\leq\\
\leq (1-p) {\mathbb{P}}[S_{\ell-1}(k_1,\dots,k_{\ell-1})\cap\{\tau_x^\alpha>(k_{\ell-1}+1)T\}]
\end{gathered}$$ and the iteration of this argument finally yields $${\mathbb{P}}[L_0>\ell, \tau_x^\alpha=\infty]
\leq (1-p)^\ell,$$ hence the claim .
We have all the ingredients to complete the proof of the main theorem of this section.
Fix $\alpha\in(\beta-2,1+\alpha_0)$, $\overline p\in(0,\beta-3)$ and $\overline{a_0}\in(0,\tfrac14]$, and consider the values $\overline{p_0}>0$, and $\overline{M_0}>0$ given by Theorem \[t:blowup\]. In view of Corollary \[c:recurrent\], it is enough to prove that the (sampled) arrival time to $B_\infty(\alpha,\overline p, \overline{a_0}, \overline{M_0})$ is finite on the event $\{\tau_x^\alpha=\infty\}$, for all $x\in V_\alpha$. To prove this, by virtue of Lemma \[l:recurrent\], it is sufficient to prove that there are $M,T_c,T_e,c_0>0$ and a function $\psi$ as in the statement of the lemma such that ${\operatorname{e}}^{-\nu\lambda^2T_c}
+\frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2}<1$ and $$\label{e:claim2}
\left.
\begin{array}{c}
\|X(t_0;x)\|_H\leq M\\
\mathcal{N}(t_0;c_0,T_c,T_e,\psi)
\end{array}
\right\}
\quad\Rightarrow\quad
X(t_0+T_c+T_e;x)\in B_\infty(\alpha,\overline p, \overline{a_0}, \overline{M_0}).$$ Indeed, the left–hand side of the above implication happens almost surely on $\{\tau_x^\alpha<\infty\}$ for some integer $k$ such that $t_0 = k(T_c+T_e)$, hence the right–hand side happens with the same probability and ${\mathbb{P}}[\sigma_{B_\infty}^{x,T_c+T_e}=\infty,\tau_x^\infty=\infty] = 0$.
We conclude with the proof of . We first notice that in Lemma \[l:contraction\], the larger we choose $M$, the larger is the time $T_c$ we get, hence we apply Lemma \[l:contraction\] with $a_0 = \frac18\overline{a_0}$, $c_0<\min\{\frac18\overline{a_0}, (4(1+\lambda^{\beta-3}))^{-1}\}$ and $M>0$ large enough so that the time $T_c$ whose existence is ensured by the lemma satisfies ${\operatorname{e}}^{-\nu\lambda^2T_c} +\frac{{\ensuremath{c_{\text{\ref{l:recurrent}}}}}}{M^2}<1$. Moreover we know that
- $\inf_{n\geq 1}\bigl(\lambda_{n-1}^{\beta-2}X_n(t_0+T_c)\bigr)
\geq -(a_0 + c_0)\nu\geq -\frac14 \overline{a_0}\nu$ on the event $\{\tau_x^\alpha=\infty\}\cap\mathcal{N}_c(c_0,t_0,T_c)$,
- $\|X(t_0+T_c)\|_H\leq \|X(t_0)\|_H + {\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}\nu\leq M + {\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}\nu$,
where the second statement follows from Lemma \[l:Hcontrol\]. We apply then Lemma \[l:expansion\] with $M_1=M + {\ensuremath{c_{\text{\ref{l:Hcontrol}}}}}\nu$, $M_2=\overline{M_0}$, $a_0 = \frac14\overline{a_0}$, $a_0' = 2a_0$ and $c_0$ as above, hence there is $T_e>0$ such that
- $\inf_{n\geq 1}\bigl(\lambda_{n-1}^{\beta-2}X_n(t_0+T_c+T_e)\bigr)
\geq -(a_0' + c_0)\nu\geq - \overline{a_0}\nu$ on the event $\{\tau_x^\alpha=\infty\}\cap\mathcal{N}_e(c_0,t_0+T_c,T_e,\psi)$,
- $\|X(t_0+T_c+T_e)\|_{\overline p}\geq \lambda^{\overline p}\|X(t_0+T_c+T_e)\|_H\geq \overline{M_0}$,
that is $X(t_0+T_c+T_e)\in B_\infty(\alpha,\overline p,\overline{a_0},
\overline{M_0})$.
[10]{}
Sergio Albeverio, Arnaud Debussche, and Lihu Xu, *Exponential mixing of the 3[D]{} stochastic [N]{}avier-[S]{}tokes equations driven by mildly degenerate noises*, 2009.
David Barbato, Michele Barsanti, Hakima Bessaih, and Franco Flandoli, *Some rigorous results on a stochastic [GOY]{} model*, J. Stat. Phys. **125** (2006), no. 3, 673–712.
David Barbato, Franco Flandoli, and Francesco Morandin, *Anomalous dissipation in a stochastic inviscid dyadic model*, 2010, [[arXiv:1007.1004](http://arxiv.org/abs/1007.1004)]{}, to appear on Ann. Appl. Probab.
, *A theorem of uniqueness for an inviscid dyadic model*, C. R. Math. Acad. Sci. Paris **348** (2010), no. 9-10, 525–528.
, *Uniqueness for a stochastic inviscid dyadic model*, Proc. Amer. Math. Soc. **138** (2010), no. 7, 2607–2617.
, *Energy dissipation and self-similar solutions for an unforced inviscid dyadic model*, Trans. Amer. Math. Soc. **363** (2011), 1925–1946.
David Barbato, Francesco Morandin, and Marco Romito, *Smooth solutions for the dyadic model*, Nonlinearity **24** (2011), no. 11, 3083–3097.
Dirk Bl[ö]{}mker, Franco Flandoli, and Marco Romito, *Markovianity and ergodicity for a surface growth [PDE]{}*, Ann. Probab. **37** (2009), no. 1, 275–313. [ [``](http://www.ams.org/mathscinet-getitem?mr=2489166) ]{}
Zdzislaw Brze[ź]{}niak, Franco Flandoli, Misha Neklyudov, and Boguslaw Zegarli[ń]{}ski, *Conservative interacting particles system with anomalous rate of ergodicity*, 2010, [[arXiv:1012.0582](http://arxiv.org/abs/1012.0582)]{}.
Alexey Cheskidov, *Blow-up in finite time for the dyadic model of the [N]{}avier-[S]{}tokes equations*, Trans. Amer. Math. Soc. **360** (2008), no. 10, 5101–5120.
Alexey Cheskidov and Susan Friedlander, *The vanishing viscosity limit for a dyadic model*, Phys. D **238** (2009), no. 8, 783–787.
Alexey Cheskidov, Susan Friedlander, and Nata[š]{}a Pavlovi[ć]{}, *Inviscid dyadic model of turbulence: the fixed point and [O]{}nsager’s conjecture*, J. Math. Phys. **48** (2007), no. 6, 065503, 16.
Giuseppe Da Prato and Arnaud Debussche, *Ergodicity for the 3[D]{} stochastic [N]{}avier-[S]{}tokes equations*, J. Math. Pures Appl. (9) **82** (2003), no. 8, 877–947.
, *On the martingale problem associated to the 2[D]{} and 3[D]{} stochastic [N]{}avier-[S]{}tokes equations*, Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. **19** (2008), no. 3, 247–264.
Giuseppe Da Prato and Jerzy Zabczyk, *Stochastic equations in infinite dimensions*, Encyclopedia of Mathematics and its Applications, vol. 44, Cambridge University Press, Cambridge, 1992.
A. de Bouard and A. Debussche, *On the effect of a noise on the solutions of the focusing supercritical nonlinear [S]{}chrödinger equation*, Probab. Theory Related Fields **123** (2002), no. 1, 76–96. [ [``](http://www.ams.org/mathscinet-getitem?mr=1906438) ]{}
Anne de Bouard and Arnaud Debussche, *Blow-up for the stochastic nonlinear [S]{}chrödinger equation with multiplicative noise*, Ann. Probab. **33** (2005), no. 3, 1078–1110. [ [``](http://www.ams.org/mathscinet-getitem?mr=2135313) ]{}
Arnaud Debussche, *Ergodicity results for the stochastic [N]{}avier–[S]{}tokes equations: an introduction*, Topics in mathematical fluid–mechanics, Lecture Notes in Math., Springer, Berlin, to appear, Lectures given at the C.I.M.E.–E.M.S. Summer School in applied mathematics held in Cetraro, September 6–11, 2010, Edited by Franco Flandoli and Hugo Beirao da Veiga.
Arnaud Debussche and Cyril Odasso, *Markov solutions for the 3[D]{} stochastic [N]{}avier-[S]{}tokes equations with state dependent noise*, J. Evol. Equ. **6** (2006), no. 2, 305–324.
John P. Dirkse, *An absorption probability for the [O]{}rnstein-[U]{}hlenbeck process*, J. Appl. Probability **12** (1975), no. 3, 595–599.
M. Dozzi and J. A. Lopez, *Finite-time blowup and existence of global positive solutions of a semi-linear [SPDE]{}*, 2009, [[arXiv:0908.3364](http://arxiv.org/abs/0908.3364)]{}.
C. L. Fefferman, *Existence and smoothness of the navier-stokes equation*, The millennium prize problems, Clay Math. Inst., Cambridge, MA, 2006, pp. 57–67.
Julian Fern[á]{}ndez Bonder and Pablo Groisman, *Time-space white noise eliminates global solutions in reaction-diffusion equations*, Phys. D **238** (2009), no. 2, 209–215. [ [``](http://www.ams.org/mathscinet-getitem?mr=2516340) ]{}
Franco Flandoli, *An introduction to 3[D]{} stochastic fluid dynamics*, S[PDE]{} in hydrodynamic: recent progress and prospects, Lecture Notes in Math., vol. 1942, Springer, Berlin, 2008, Lectures given at the C.I.M.E. Summer School held in Cetraro, August 29–September 3, 2005, Edited by Giuseppe Da Prato and Michael R[ö]{}ckner, pp. 51–150.
, *Random perturbation of [PDE]{}s and fluid dynamic model*, Lecture [N]{}otes in [M]{}athematics, vol. 2015, Springer, Berlin, 2011, [É]{}cole d’[É]{}t[é]{} de Probabilit[é]{}s de Saint-Flour XL.
Franco Flandoli and Marco Romito, *Markov selections and their regularity for the three-dimensional stochastic [N]{}avier-[S]{}tokes equations*, C. R. Math. Acad. Sci. Paris **343** (2006), no. 1, 47–50. [ [``](http://www.ams.org/mathscinet-getitem?mr=2241958) ]{}
, *Regularity of transition semigroups associated to a 3[D]{} stochastic [N]{}avier-[S]{}tokes equation*, Stochastic differential equations: theory and applications (Peter H. Baxendale and Sergey V. Lototski, eds.), Interdiscip. Math. Sci., vol. 2, World Sci. Publ., Hackensack, NJ, 2007, pp. 263–280. [ [``](http://www.ams.org/mathscinet-getitem?mr=2393580) ]{}
, *Markov selections for the 3[D]{} stochastic [N]{}avier-[S]{}tokes equations*, Probab. Theory Related Fields **140** (2008), no. 3-4, 407–458. [ [``](http://www.ams.org/mathscinet-getitem?mr=2365480) ]{}
Susan Friedlander and Nata[š]{}a Pavlovi[ć]{}, *Blowup in a three-dimensional vector model for the [E]{}uler equations*, Comm. Pure Appl. Math. **57** (2004), no. 6, 705–725.
, *Remarks concerning modified [N]{}avier-[S]{}tokes equations*, Discrete Contin. Dyn. Syst. **10** (2004), no. 1-2, 269–288, Partial differential equations and applications.
Nobuyuki Ikeda and Shinzo Watanabe, *Stochastic differential equations and diffusion processes*, North-Holland Publishing Co., Amsterdam, Oxford, New York, 1981.
Ioannis Karatzas and Steven E. Shreve, *Brownian motion and stochastic calculus*, second ed., Graduate Texts in Mathematics, vol. 113, Springer-Verlag, New York, 1991.
Nets Hawk Katz and Nata[š]{}a Pavlovi[ć]{}, *Finite time blow-up for a dyadic model of the [E]{}uler equations*, Trans. Amer. Math. Soc. **357** (2005), no. 2, 695–708 (electronic).
Alexander Kiselev and Andrej Zlato[š]{}, *On discrete models of the [E]{}uler equation*, Int. Math. Res. Not. (2005), no. 38, 2315–2339.
Nicolai V. Krylov, *The selection of a [M]{}arkov process from a [M]{}arkov system of processes, and the construction of quasidiffusion processes*, Izv. Akad. Nauk SSSR Ser. Mat. **37** (1973), 691–708.
Carl Mueller, *The critical parameter for the heat equation with a noise term to blow up in finite time*, Ann. Probab. **28** (2000), no. 4, 1735–1746. [ [``](http://www.ams.org/mathscinet-getitem?mr=1813841) ]{}
Carl Mueller and Richard Sowers, *Blowup for the heat equation with a noise term*, Probab. Theory Related Fields **97** (1993), no. 3, 287–320. [ [``](http://www.ams.org/mathscinet-getitem?mr=1245247) ]{}
Cyril Odasso, *Exponential mixing for the 3[D]{} stochastic [N]{}avier-[S]{}tokes equations*, Comm. Math. Phys. **270** (2007), no. 1, 109–139.
Marco Romito, *Ergodicity of the finite dimensional approximation of the 3[D]{} [N]{}avier-[S]{}tokes equations forced by a degenerate noise*, J. Statist. Phys. **114** (2004), no. 1-2, 155–177. [ [``](http://www.ams.org/mathscinet-getitem?mr=2032128) ]{}
, *Analysis of equilibrium states of [M]{}arkov solutions to the 3[D]{} [N]{}avier-[S]{}tokes equations driven by additive noise*, J. Stat. Phys. **131** (2008), no. 3, 415–444. [ [``](http://www.ams.org/mathscinet-getitem?mr=2386571) ]{}
, *An almost sure energy inequality for [M]{}arkov solutions to the 3[D]{} [N]{}avier-[S]{}tokes equations*, 2009, [[arXiv:0902.1407](http://arxiv.org/abs/0902.1407)]{}, to appear on Quaderni di Matematica.
, *Critical strong [F]{}eller regularity for [M]{}arkov solutions to the [N]{}avier–[S]{}tokes equations*, J. Math. Anal. Appl. **384** (2010), 115–129, [[arXiv:1003.4623](http://arxiv.org/abs/1003.4623)]{}.
, *The martingale problem for [M]{}arkov solutions to the [N]{}avier-[S]{}tokes equations*, Seminar on [S]{}tochastic [A]{}nalysis, [R]{}andom [F]{}ields and [A]{}pplications [VI]{}, Progr. Probab., vol. 63, Birkhäuser, Basel, 2011, pp. 227–244.
Marco Romito and Lihu Xu, *Ergodicity of the 3[D]{} stochastic [N]{}avier-[S]{}tokes equations driven by mildly degenerate noise*, Stoch. Proc. Appl. **121** (2011), no. 4, 673–700.
Armen Shirikyan, *Approximate controllability of three-dimensional [N]{}avier-[S]{}tokes equations*, Comm. Math. Phys. **266** (2006), no. 1, 123–151.
Fabian Waleffe, *On some dyadic models of the [E]{}uler equations*, Proc. Amer. Math. Soc. **134** (2006), no. 10, 2913–2922 (electronic).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Elimination of quantifiers is shown to fail dramatically for a group of well-known mathematical theories (classically enjoying the property) against a wide range of relevant logical backgrounds. Furthermore, it is suggested that only by moving to more extensional underlying logics can we get the property back.
*Keywords:* quantifier elimination, Routley-Meyer semantics, relevant logic, model theory, paraconsistent mathematics.
*Math subject classification:* Primary, 03C10, 03B47 ; Secondary, 03B53.
author:
- 'Guillermo Badia, Andrew Tedder'
bibliography:
- 'BDLO\_QE.bib'
nocite: '[@*]'
title: 'ON ELIMINATION OF QUANTIFIERS IN SOME NON-CLASSICAL MATHEMATICAL THEORIES'
---
Introduction
============
The property of *quantifier elimination* for a theory has the following classical uses ([@Hodges93], pp. 68-69): (1) classification of structures up to elementary equivalence, (2) completeness and decidability proofs for theories, (3) description of elementary embeddings, and (4) description of definable relations in a structure. In this paper, we chose to focus on the latter and, as in [@Marker02], take it to be the fundamental import of quantifier elimination. Our aim is to explore elimination of quantifiers in the non-classical setting of quantificational logics which are sound with respect to the Routley-Meyer semantic framework from relevant logic [@Routley82].[^1] Such quantificational logics include many relevant and paraconsistent logics as well as, say, classical logic.
We should alert the reader that the concern of the present pages will be with *mathematical theories* (theories describing some mathematical structure), so we will disregard things such as the pure theory of equality given that we will assume equality to be a logical notion with a fixed interpretation.[^2] Some examples of mathematical theories are the theory of *dense linear orderings without endpoints*, the theory of *real closed fields*, the theory of *ordered divisible abelian groups* and *Presburger arithmetic*. All these theories are known to have quantifier elimination in the classical setting [@Marker02]. Here we investigate the fate of this property in the Routley-Meyer semantic framework.
We also look at the minimal logical principles that can be added to our formal systems to recover quantifier elimination. This is in line with some recent investigations into the non-classical models of mathematical theories.[^3] Our toy theory will be the theory of dense linear orderings without endpoints (henceforth, DLO), which is the simplest around.[^4]
This paper is (fundamentally) a contribution to the study of paraconsistent mathematical theories (i.e., theories where from a sentence and its negation not everything follows). Our work is motivated by the need of understanding just how different the behaviour of these theories is from that of their classical counterparts. We would like to stress that, surprisingly, most of the research done in this field with logics having a detachable conditional has used algebraic models. In that sense, [@Restall09] seems to be the first (and likely only) place where Routley-Meyer relational models have been employed in the study of substructural mathematical theories as we will do below.
In §\[sec:log\], we review the details of the Routley-Meyer semantics as well as introduce the fundamental definitions we will be working with. In §\[math\], we present the theories we will be studying. In §\[sec:fqe\], we show that quantifier elimination fails for well-known mathematical theories when the underlying logic is any of the most famous systems of relevant logic. In §\[sec:qe\], we discuss how to get quantifier elimination back. Finally, in §\[sec:con\], we summarize our results.
Logical Preliminaries {#sec:log}
=====================
All the languages we will be considering will have connectives $\neg,\land,\lor,\rightarrow,\forall,\exists,\bot,\top$ (with the usual arities) and a denumerable list of individual variables $x_0,x_1,\dots$. As usual, by the *signature* of a language we will mean its relation, function and constant symbols.
We will be presenting relevant logics semantically rather than syntactically. The reason for this is that Fine [@Fine89] has shown that the natural proof-theoretic formulation of the well-know system **R** of relevant logic plus first order quantifiers is incomplete with respect to the Routley-Meyer semantic framework.[^5] Hence, if one desires to find a complete axiomatization for the systems we will call **B**, **R** and **RM** they would need to go beyond what is usually understand proof-theoretically by these calculi. It is not difficult to show, however, that our semantically defined systems are indeed recursively axiomatizable using Craig’s method.
Models {#models .unnumbered}
------
We employ the standard ternary relation semantics, also routinely called the “Routley-Meyer” semantics. A **B**-model is a structure $M=\langle W,R,D,^*,s,V\rangle$ where $W$ is a non-empty set (of worlds), $R\subseteq W^3$, $D$ is a set of objects (the domain), $^*: W \longrightarrow W$, $s \in W$, $V$ is a valuation such that if $w \in W$ and $P^n$ is an $n$-ary predicate, $V(w, P^n) \subseteq D^n$, and moreover:
- For any $\alpha\in W$, $Rs\alpha\alpha$
- for any $\alpha,\beta,\gamma,\delta\in W$, $Rs\alpha\beta$ and $R\beta\gamma\delta$ implies that $R\alpha\gamma\delta$
- For any $\alpha\in W$, $\alpha^{**}=\alpha$
- For any $\alpha,\beta\in W$, if $Rs\alpha\beta$ then $Rs\beta^*\alpha^*$
- If $R0\alpha\beta$, then for any $P^n$, $V(\alpha, P^n) \subseteq V(\beta, P^n)$.
The special predicate = will always be interpreted as the metatheoretic classical identity unless stated otherwise.
The additional ternary relation restrictions necessary for the logic **R**-mingle, or **RM**, are as follows (the axioms enforced by the principles are included):[^6]
- For any $\alpha,\beta, \gamma \in W$, if $R\alpha\beta \gamma$ then $R\alpha \gamma^*\beta^*$ $(A\rightarrow B)\rightarrow (\neg B \rightarrow \neg A)$
- For any $\alpha,\beta,\gamma,\delta\in W$, if $R\alpha\beta\gamma$ then $R\alpha\gamma\beta$.
- For any $\alpha\in W$, $R\alpha\alpha\alpha$.
- If there is an $x\in W$ s.t. $R\alpha\beta x$ and $Rx\gamma\delta$ then there is a $y\in W$ s.t. $R\alpha y\delta$ and $R\beta\gamma y$.
- If $R\alpha\beta\gamma$ then either $Rs\alpha\gamma$ or $Rs\beta\gamma$ $A\rightarrow(A\rightarrow A)$
The correspondence between combinatory logic and propositional logic support the following pairs of combinators and valid formulae:
- $(A\rightarrow B)\rightarrow ((C\rightarrow A)\rightarrow (C\rightarrow B))$
- $((A\rightarrow B)\land A)\rightarrow B$
- $A\rightarrow ((A\rightarrow B)\rightarrow B)$
Given a model $M$, we will use $\overline{a}$ to denote a sequence of individuals of the domain of $M$. Truth of complex formulae at a world is defined as follows:
- $M,w\vDash \bot$ never
- $M,w\vDash A\land B[\overline{a}]$ iff $M,w\vDash A[\overline{a}]$ and $M,w\vDash B[\overline{a}]$
- $M,w\vDash (A\lor B)[\overline{a}]$ iff $M,w\vDash A[\overline{a}]$ or $M,w\vDash B[\overline{a}]$
- $M,w\vDash\neg A [\overline{a}]$ iff $M,w^*\nvDash A[\overline{a}]$
- $M,w\vDash (A\rightarrow B)[\overline{a}]$ iff if $Rww_1w_2$ and $M,w_1\vDash A[\overline{a}]$ then $M,w_2\vDash B[\overline{a}]$
- $M,w\vDash\forall xA[\overline{a}]$ iff for any $y\in D$, $M,w\vDash A[y/x, \overline{a}]$
- $M,w\vDash\exists xA[\overline{a}]$ iff for some $y\in D$, $M,w\vDash A[y/x, \overline{a}]$
For any logic **L** discussed here, set of formulas $\Gamma$ and formula $A$, we will say that $A$ *is deducible from* $\Gamma$ in [**L**]{} or, in symbols, $\Gamma \vDash_{\bf L} A$ when for every [**L**]{}-model $M$, we have that $M,s\vDash \Gamma$ only if $M,s \vDash A$. So deducibility only cares about what happens at the distinguished world $s$. In particular if $\Gamma = \emptyset$ we get a definition of theoremhood in [**L**]{}. In what follows, we will drop the subscript in $\vDash_{\bf L}$ when the context makes clear about what system we are talking.
A theory $T$ *admits quantifier elimination* if, for any formula $A$, there is a quantifier free formula $B$ such that $T\vDash A\leftrightarrow B$.
Next we make a couple of important observations which will be used (often without explicit mention) throughout the paper.
*(Hereditary condition)* Let M be a **B**-model. If $Rs\alpha \beta$ and $A(x)$ is any formula, $M, \alpha \vDash A[a]$ only if $M, \beta \vDash A[a]$.
Let M be a **B**-model. Then $M, s \vDash A \rightarrow B [a]$ iff for any $\alpha$, $M, \alpha \vDash A [a]$ only if $M, \alpha \vDash B [a]$.
Given some equivalence relation $\equiv$ between models, where $N\equiv M$ implies that all sentences true in one of the models are also true in the other, a theory $T$ is *$\omega$-categorical with respect to $\equiv$* if for any countable models $N,M\vDash T$, $N \equiv M$.
A theory $T$ is *negation-complete* if for any formula $A$, either $T\vDash A$ or $T\vDash\neg A$.
A theory $T$ is *complete* if for any models $M, N$ of $T$, $M$ satisfies exactly the same sentences as $N$.
Quantifier elimination is a model-theoretic property in the sense that it considerably simplifies the theory of the definable sets of the models of a given theory: these amount to the things one can define using quantifier free formulas.
Some mathematical theories {#math}
==========================
In this section, we will review the axiomatizations of a our target mathematical theories. The theories as presented are, of course, not the classical theories which go by these names, as the underlying logic differs from classical logic. However, we retain (roughly) the same axioms as the classical theories, though stated in non-classical vocabulary. We shall be focused on properties had by classical formal theories which are lost when the underlying logic is relevant, and so we shall use the usual names for the classical theories to make explicit the comparison.
Dense Linear Orderings {#dense-linear-orderings .unnumbered}
----------------------
In a signature with a binary relation symbol $<$, DLO without endpoints is given by the following axioms:[^7]
1. $\forall x(x<x\rightarrow\bot)$
2. $\forall x,y(x=y\lor x<y\lor y<x)$
3. $\forall x,y,z((x<y\land y<z)\rightarrow x<z)$
4. $\forall x,y(x<y\rightarrow\exists z(x<z\land z<y))$
5. $\forall x\exists y(x<y)$
6. $\forall x\exists y(y<x)$
If DLO is formulated with $\forall x \neg (x<x)$ instead of A1 it is a simple matter to build a one element model for it, so the following result (though easy) is a necessary preliminary.
DLO in any logic **L** extending **B** has no finite models.
This is easily seen essentially as in the classical case by constructing an infinitely ascending (descending) sequence (in the ordering holding at world $s$) of distinct elements of the domain of objects of any **B**-model of DLO. In the reasoning one needs to use the property $Rsss$ of **B**-models in an essential way but only axioms A1, A5 (or alternatively A6) and A3.
We have to insist that the failure of QE for DLO (which will be established below in §\[sec:fqe\]) holds, *a fortiori*, for DLO with $\forall x \neg (x<x)$ replacing A1. We do not have a direct proof of the failure of QE for such alternative formulation of DLO. A similar remark applies to all the theories we will examine in this paper, hence whenever $\neg A$ occurs in a standard classical axiomatisation of some theory, we shall replace this with $A\rightarrow\bot$. For the QE result when the underlying logic is strengthened, DLO in its current form looks like what is required.
Real Closed Fields {#real-closed-fields .unnumbered}
------------------
The signature of the real closed fields has a binary predicate symbol $<$, binary function symbols $+, \times$, unary function symbols $-, ^{-1}$, and constants $0, 1$. The theory RCF has the following axioms:[^8]
1. $\forall x(x<x\rightarrow\bot)$
2. $\forall x,y(x=y\lor x<y\lor y<x)$
3. $\forall x,y,z((x<y\land y<z)\rightarrow x<z)$
4. $\forall x,y((0<x \land 0 < y) \rightarrow 0 < xy)$
5. $\forall x,y, z(x< y \rightarrow x+z<y+z)$
6. $\forall x \exists y (0 < x \rightarrow x = yy)$
7. $\forall x_1, \dots, x_n \exists y (y^n+x_1y^{n-1}+ \dots + x_n =0)$ for each odd $n > 0$.
8. $\forall x, y, z ((x+y)+z=x+(y+z))$
9. $\forall x (x+0=x)$
10. $\forall x (x+(-x) =0)$
11. $\forall x, y (x+y=y+x)$
12. $\forall x, y ((xy)z = x(yz))$
13. $\forall x (x1 = x)$
14. $\forall x (\neg(x = 0) \rightarrow xx^{-1} = 1)$
15. $\forall x, y(xy = yx)$
16. $\forall x, y, z (x(y+z) = (xy)+(xz))$
17. $0 = 1 \rightarrow \bot$
Presburger arithmetic {#presburger-arithmetic .unnumbered}
---------------------
The signature of Presburger arithmetic has a binary predicate symbol $<$, unary predicates $P_n$ ($n > 1$), a binary function symbol $+$, and constants $0, 1$. The axioms are as follows:
1. $\forall x(x<x\rightarrow\bot)$
2. $\forall x,y(x=y\lor x<y\lor y<x)$
3. $\forall x,y,z((x<y\land y<z)\rightarrow x<z)$
4. $\forall x,y, z(x< y \rightarrow x+z<y+z)$
5. $\forall x, y, z ((x+y)+z=x+(y+z))$
6. $\forall x (x+0=x)$
7. $\forall x, y (x+y=y+x)$
8. $0 < 1 $
9. $\forall x((x=0 \lor x<0)\lor (1<x \lor x =1))$
10. $\forall x (P_n(x) \leftrightarrow \exists y (x = \underbrace{y + \dots + y}_{n-\emph{times}})) $ for $n > 1$
11. $\forall x (\bigvee_{i=0}^{n-1} (P_n(x + \underbrace{1 + \dots + 1}_{i-\emph{times}}) \land \bigwedge_{i \neq j} (P_n(x + \underbrace{1 + \dots + 1}_{j-\emph{times}}) \rightarrow \bot))) $ for $n > 1$
Divisible Ordered Abelian Groups {#divisible-ordered-abelian-groups .unnumbered}
--------------------------------
The signature of the theory of divisible ordered abelian groups (DOAG) has a binary predicate symbol $<$, a unary function symbol $-$, a binary function symbol $+$, and the constants $0$. Here is the list of axioms:
1. $\forall x(x<x\rightarrow\bot)$
2. $\forall x,y(x=y\lor x<y\lor y<x)$
3. $\forall x,y,z((x<y\land y<z)\rightarrow x<z)$
4. $\forall x,y, z(x< y \rightarrow x+z<y+z)$
5. $\forall y \exists x (\underbrace{x+x \dots x}_{n-times} = y)$ for $n > 1$.
6. $\forall x, y, z ((x+y)+z=x+(y+z))$
7. $\forall x (x+0=x)$
8. $\forall x (x+(-x) =0)$
9. $\forall x, y (x+y=y+x)$
Algebraically Closed Fields {#sec:acf .unnumbered}
---------------------------
The signature of the algebraically closed fields has binary function symbols $+, \times$, unary function symbols $-, ^{-1}$, and constants $0, 1$. The theory ACF has the following axioms:
1. $\forall x, y, z ((x+y)+z=x+(y+z))$
2. $\forall x (x+0=x)$
3. $\forall x (x+(-x) =0)$
4. $\forall x, y (x+y=y+x)$
5. $\forall x, y ((xy)z = x(yz))$
6. $\forall x (x1 = x)$
7. $\forall x (\neg(x = 0) \rightarrow xx^{-1} = 1)$
8. $\forall x, y(xy = yx)$
9. $\forall x, y, z (x(y+z) = (xy)+(xz))$
10. $0 = 1 \rightarrow \bot$
11. $\forall x_1, \dots, x_n \exists y (y^n+x_1y^{n-1}+ \dots + x_n =0)$ for each $n > 0$.
Failure of Quantifier Elimination {#sec:fqe}
=================================
In this section we study the failure of QE for most of the theories from §\[math\]. Note that we employ the following seriality condition: $\forall x\exists y,z (Rxyz)$. This is need for the proof, but is not motivated by any considerations other than that it is so needed. We have not discovered a means of obtaining this important result without the seriality condition.
\[1\] Let **L** be any logic extending **B** such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$ (a seriality requirement). Then if $A(x)$ is a quantifier free formula with one free variable in DLO without endpoints, either $A(x)\leftrightarrow x=x$ or $A(x)\leftrightarrow x<x$.
We proceed by structural induction on $A(x)$:
Base: $A(x)$ is atomic; hence $A(x)$ is either $x=x$ or $x<x$.
Induction step: We proceed by cases on the main connective of $A(x)$, where the induction hypothesis guarantees that any component formula $B(x)$ of $A(x)$ are such that either $B(x)\leftrightarrow x=x$ or $B(x)\leftrightarrow x<x$.
Case 1: $A(x)$ is $B(x)\land C(x)$. By induction hypothesis, we know that any of the following four may be true of $B(x)$, $C(x)$.
- $B(x)\leftrightarrow x=x$,
- $B(x)\leftrightarrow x<x$,
- $C(x)\leftrightarrow x=x$,
- $C(x)\leftrightarrow x<x$.
So, we have four possibilities to consider:
1. \(i) and (iii)
2. \(i) and (iv)
3. \(ii) and (iii)
4. \(ii) and (iv)
Suppose (i) and (iii) hold. Then $(B(x)\land C(x))\leftrightarrow (x=x\land x=x)\leftrightarrow x=x$. The result for (ii) and (iv) is similar.
Suppose (i) and (iv) hold. Then $(x=x\land x<x)\leftrightarrow (B(x)\land C(x))$. So $(B(x)\land C(x))\leftrightarrow x<x$ given that $(x=x\land x<x)\leftrightarrow x<x$ (the right to left implication follows using A1). The result for (ii) and (iii) is similar.
Case 2: $A(x)$ is $B(x)\lor C(x)$. By inductive hypothesis we have the same set of possibilities as above.
Suppose that (i), (iii) hold. As in Case 1, $(x=x\lor x=x)\leftrightarrow x=x$. For (ii), (iv), $(x<x\lor x<x) \leftrightarrow x<x$.
Suppose that (i), (iv) hold. That is, $(B(x)\lor C(x))\leftrightarrow (x=x\lor x<x)$. Well, $x=x\rightarrow(x=x\lor x<x)$. Obviously, $x=x\rightarrow x=x$; we need merely note that by A1 and the transitivity of $\rightarrow$, $x<x\rightarrow x=x$. Given this, we have $(x=x\rightarrow x=x)\land(x<x\rightarrow x=x)$, and we have as an instance of an axiom of **B** that $((x=x\rightarrow x<x)\land(x<x\rightarrow x=x))\rightarrow((x=x\lor x<x)\rightarrow x=x)$, therefore $(x=x\lor x<x) \rightarrow x=x$. So $(B(x)\lor C(x))\leftrightarrow x=x$.
Case 3: Suppose that $A(x)$ is $\neg B(x)$. By inductive hypothesis, either $B(x)\leftrightarrow x=x$ or $B(x)\leftrightarrow x<x$. Suppose that $B(x)\leftrightarrow x=x$. Then $A(x)\leftrightarrow\neg x=x$. It is easy to check that $\neg x=x\rightarrow\bot$ is a theorem (recall that the interpretation of $=$ has been fixed at every world) and $\bot\rightarrow x<x$, hence $A(x)\leftrightarrow x<x$ using A1. Suppose that $B(x)\leftrightarrow x<x$. Then $A(x)\leftrightarrow\neg x<x$. By A1, contraposition, the fixed interpretation of $=$ which makes $x=x \leftrightarrow \top$ hold, and transitivity of $\rightarrow$, it follows that $x=x\leftrightarrow\neg x<x$, hence $A(x)\leftrightarrow x=x$.
Case 4: Suppose that $A(x)$ is $B(x)\rightarrow C(x)$. Again, we have the possibilities (1)–(4) as in Case 1.
Suppose that (i), (iii) hold. Then $(B(x)\rightarrow C(x))\leftrightarrow(x=x\rightarrow x=x)$. Since $x=x\leftrightarrow\top$, we know that $(\top\leftrightarrow\top)\leftrightarrow(x=x\leftrightarrow x=x)$. Similarly, $(\top\leftrightarrow\top)\leftrightarrow\top$, and so $(x=x\leftrightarrow x=x)\leftrightarrow x=x$, and $A(x)\leftrightarrow x=x$.
Suppose that (ii), (iv) hold. Then $(B(x)\rightarrow C(x))\leftrightarrow(x<x\rightarrow x<x)\leftrightarrow(\bot\leftrightarrow\bot)\leftrightarrow\top\leftrightarrow x=x$. So $A(x)\leftrightarrow x=x$.
Suppose that (i), (iv) hold. Then $(B(x)\rightarrow C(x))\leftrightarrow(x=x\rightarrow x<x)$. We know that $(x=x\rightarrow x<x)\leftrightarrow (\top\rightarrow\bot)$. Note that using the seriality condition on our frames, $(\top\rightarrow\bot)\leftrightarrow\bot$ must hold. It suffices to note that never $\top\rightarrow\bot$ holds at any world $\alpha$ of any model, for otherwise, using seriality, $\bot$ would have to hold at some world, which is impossible. Hence, $(x=x\rightarrow x<x)\leftrightarrow\bot$, and the definition of $\bot$ gives us that $\bot\leftrightarrow x<x$, so $A(x)\leftrightarrow x<x$.
Suppose that (ii), (iii) hold. Then $(B(x)\rightarrow C(x))\leftrightarrow(x<x\rightarrow x=x)$ and we know that $(x<x\rightarrow x=x)\rightarrow x=x$. Similarly, we know that $x=x\rightarrow(x<x\rightarrow x=x)$ given that $x<x\rightarrow x=x$ holds at every world in every model, and so $A(x)\leftrightarrow x=x$.
Let **L** be any relevant logic extending **B** up to **RM**. Then QE fails for DLO in **L**.
Let $N=\langle W,D,R,^*,s,V\rangle$ the model where:
- $W=\{s,t\}$
- $D=\mathbb{Q}$
- $R=\{\langle s,s,s\rangle,\langle s,t,t\rangle,\langle t,s,s\rangle,\langle t,s,t\rangle,\langle t,t,s\rangle,\langle t,t,t\rangle, \langle s,t,s\rangle\}$
- $V$ assigns formulae containing $<$ the values consonant with the usual ordering on $Q$ at world $s$. At world $t$, $V$ assigns the usual ordering on $<$ for rationals in the interval $[2,3]$ and assigns false to any atomic formula involving rationals outside this interval. At every world, $=$ is interpreted as real identity.
- $s$ is the designated world.
- $^*=\{\langle s,t\rangle,\langle t,s\rangle\}$
The reader can easily convince themselves that DLO is validated at $s$. On the other hand, it is a tedious, but mechanical process to check that $N$ satisfies the frame conditions to validate the logic **RM**.[^9]
The definition of $\{2\}$ at world $s$ is the formula
$\neg\exists x(x<y)\land(\forall x (x=x)\rightarrow\exists x(y<x))$.
This is due to the fact that for any $a$,
$N,s\vDash \neg\exists x(x<y)\land(\forall x (x=x)\rightarrow\exists x(y<x))[a]$ iff $a=2$.
First, it is easy to see that $N,s\vDash \neg\exists x(x<y)\land(\forall x (x=x)\rightarrow\exists x(y<x))[2]$ since $N,t\nvDash \exists x(x<y)[2]$ by definition of $N$ and $N,s\vDash\forall x (x=x)\rightarrow\exists x(y<x)[2]$ given that $N,w\vDash \exists x(y<x)[2]$ for any $w\in W$. Finally, note that by definition of $N$, 2 is the only element $a \in D$ for which $N,t\nvDash \exists x(x<y)[a]$ and $N,t\vDash \exists x(y<x)[a]$.
These results show that $\neg\exists x(x<y)\land(\forall x (x=x)\rightarrow\exists x(y<x))$ is not equivalent to $\top$ or $\bot$, (that is, is not equivalent to $x<x$ or $x=x$). Now, given that the models for [**RM**]{} satisfy the seriality condition, by Lemma \[1\], $\neg\exists x(x<y)\land(\forall x (x=x)\rightarrow\exists x(y<x))$ is not equivalent to a quantifier free formula.[^10]
Failure of $\omega$-categoricity with respect to an equivalence relation {#failure-of-omega-categoricity-with-respect-to-an-equivalence-relation .unnumbered}
------------------------------------------------------------------------
There exist countable models $N,M\vDash$ DLO s.t. not $N\equiv M$.
Let $M=\langle W,D,R,^*,s,V\rangle$ where:
- $W=\{s\}$, $R=W^3$, $^*=W^2$
- $D=\mathbb{Q}$
- $s$ is the designated world
- $V$ assigns the usual $<$-ordering on $\mathbb{Q}$.
- $V$ assigns = real identity.
Given the definition of $V$, $M\vDash$ DLO, and since $D=\mathbb{Q}$, $M$ is countable. $M,s\vDash\forall x (x = x)\rightarrow$ A3 and $M,s\nvDash$ A3 $\land\neg$ A3.
Let $N=\langle W,D,R,^*,s,V\rangle$ where:
- $W=\{s,t\}$
- $D=\mathbb{Q}$
- $R=\{\langle s,s,s\rangle,\langle s,t,t\rangle\}$
- $^*=\{\langle s,t\rangle,\langle t,s\rangle\}$
- $s$ is the designated world
- $V$ assigns the usual $<$-ordering on $\mathbb{Q}$ at $s$, and assigns $<$ to $\varnothing$ at $t$.
- $V$ assigns = real identity at every world.
Given the definition of $V$ for $<$ at $s$, $N\vDash$ DLO. Again, $N$ is countable. We need only note that $N,s\nvDash\forall x (x = x)\rightarrow$ A3 and $N,s\vDash$ A3 $\land\neg$ A3. Hence, we have formulae $A,B$ s.t. $M\vDash A$ and $N\nvDash A$, and $M\nvDash B$ and $N\vDash B$, and so not $M\equiv N$.
Failure of Negation Completeness {#failure-of-negation-completeness .unnumbered}
--------------------------------
There exists a formula $A$ such that DLO in RM proves neither $A$ nor $\neg A$.
Consider the RM model $M$ where:
- $W=\{s\}$, $R=W^3$, $^*=W^2$
- $D=\mathbb{Q}$
- $s$ is the designated world
- $V$ assigns the usual $<$-ordering on $\mathbb{Q}$
On this model, the ‘classical’ model for RM, we have that $M,s\nvDash\neg(\forall x\neg x<x\rightarrow\forall x\exists yy<x)$. In the model $N$ defined in Theorem 3, we have that $N,t\nvDash\forall x\neg x<x\rightarrow\forall x\exists yy<x$ because $N,t\vDash\forall x\neg x<x$ and $N,t\nvDash\forall x\exists yy<x$. Hence, it is the case that DLO$\nvDash \forall x\neg x<x\rightarrow\forall x\exists yy<x$ and DLO$\nvDash\neg(\forall x\neg x<x\rightarrow\forall x\exists yy<x)$.
So, for DLO, we have the failure of negation completeness, quantifier elimination, $\omega$-categoricity, and it’s fairly obvious that we also have the failure of model-completeness. Essentially, in any relevant logic, DLO fails to have any of the nice model-theoretic properties it enjoys in the classical context. At base, one can chalk this up to the fact that in the relevant setting, we have much more power to construct countermodels.
\[t\] Let **L** be any relevant logic between **B** and **RM**. Then QE fails for RCF in **L**.
Consider the first order structure $\langle \mathbb{R}, <_{\mathbb{R}}, +_{\mathbb{R}}, -_{\mathbb{R}}, ^{-1_{\mathbb{R}}}, \times_{\mathbb{R}}, 0_{\mathbb{R}}, 1_{\mathbb{R}} \rangle$. By the upward Löwenheim-Skolem theorem, there is an elementary (in the sense of classical first order logic) extension $\langle \mathbb{R}^{\prime}, <_{\mathbb{R}^{\prime}}, +_{\mathbb{R}^{\prime}}, -_{\mathbb{R}^{\prime}}, ^{-1_\mathbb{R}^{\prime}}, \times_{\mathbb{R}^{\prime}}, 0_{\mathbb{R}^{\prime}}, 1_{\mathbb{R}^{\prime}} \rangle$ of $\langle \mathbb{R}, <_{\mathbb{R}}, +_{\mathbb{R}}, -_{\mathbb{R}}, ^{-1_{\mathbb{R}}}, \times_{\mathbb{R}}, 0_{\mathbb{R}}, 1_{\mathbb{R}} \rangle$ such that $|\mathbb{R}^{\prime}| = 2^{2^{\omega}}$. So in particular, $\langle \mathbb{R}^{\prime}, <_{\mathbb{R}^{\prime}}, +_{\mathbb{R}^{\prime}}, -_{\mathbb{R}^{\prime}}, ^{-1_\mathbb{R}^{\prime}}, \times_{\mathbb{R}^{\prime}}, 0_{\mathbb{R}^{\prime}}, 1_{\mathbb{R}^{\prime}} \rangle$ is a model of the classical theory of real closed fields. Furthermore, it can be guaranteed that $|\mathbb{R}| < |\{x \in \mathbb{R}^{\prime} : 0_{\mathbb{R}^{\prime}} <_{\mathbb{R}^{\prime}} x\}|$.
Define the ordering $<^{\prime}$ on $\mathbb{R}^{\prime}$ as:
$x<'y$ iff $0 <_{\mathbb{R}^{\prime}}y-x$ and $y-x \in \mathbb{R}$,
where the expression $y-x$ abbreviates $y+(-x)$ as usual. First, we see that for $x \in \mathbb{R}^{\prime}$, it cannot be the case that $x<'x$ since $0 <_{\mathbb{R}^{\prime}}x-x = 0$ is false. So $<'$ is irreflexive. If $x, y, z \in \mathbb{R}^{\prime}$, $x<'y$ and $y<'z$, i. e., $0 <_{\mathbb{R}^{\prime}}y-x \in \mathbb{R}$ and $0 <_{\mathbb{R}^{\prime}}z-y \in \mathbb{R}$, then $0 <_{\mathbb{R}^{\prime}}z-x=(y-x)+(z-y) \in \mathbb{R}$ (for $\mathbb{R}$ is certainly closed under $+$), which means that $x <' z$. So $<'$ is transitive. Now if $0<'x$ and $0<'y$, i.e., $0 <_{\mathbb{R}^{\prime}}x-0=x \in \mathbb{R}$ and $0 <_{\mathbb{R}^{\prime}}y-0=y \in \mathbb{R}$ then surely $0 <_{\mathbb{R}^{\prime}}xy-0 = xy \in \mathbb{R}$ (for $\mathbb{R}$ is also closed under $\times$), i.e., $0<'xy$. Finally, if $x<'y$, i. e., $0 <_{\mathbb{R}^{\prime}}y-x \in \mathbb{R}$ then since $z-z=0$, it follows that $0 <_{\mathbb{R}^{\prime}} (y+z) - (x+z) = (y-x) + (z-z) \in \mathbb{R}$, which means that $x+z<'y+z$.
Now consider a Routley-Meyer model $N^{\prime}=\langle W,D,R,^*,s,V\rangle$ where
- $W=\{s, t\}$
- $D=\mathbb{R}^{\prime}$
- All constants and function symbols of the language of RCF are interpreted on $\mathbb{R}^{\prime}$ as in $\langle \mathbb{R}^{\prime}, <_{\mathbb{R}^{\prime}}, +_{\mathbb{R}^{\prime}}, -_{\mathbb{R}^{\prime}}, ^{-1_\mathbb{R}^{\prime}}, \times_{\mathbb{R}^{\prime}}, 0_{\mathbb{R}^{\prime}}, 1_{\mathbb{R}^{\prime}} \rangle$.
- $R=\{\langle s,s,s\rangle,\langle s,t,t\rangle,\langle t,s,s\rangle,\langle t,s,t\rangle,\langle t,t,s\rangle,\langle t,t,t\rangle, \langle s,t,s\rangle\}$
- $V$ assigns the symbol $<$ the (strict) ordering $<_{ \mathbb{R}^{\prime}}$ on $\mathbb{R}^{\prime}$ at world $s$. At world $t$, $V$ assigns the ordering $<'$ as the interpretation of the symbol $<$. Note that since $<' \subset <_{\mathbb{R}^{\prime}}$, $V$ is an admissible valuation in the Routley-Meyer semantics. At every world, $=$ is interpreted as real identity.
- $s$ is the designated world.
- $^*=\{\langle s,t\rangle,\langle t,s\rangle\}$.
This structure is a model of RCF. Next take any $r \in \{x \in \mathbb{R}^{\prime} : 0_{\mathbb{R}^{\prime}} <_{\mathbb{R}^{\prime}} x\}\setminus \mathbb{R}$ (recall that we made sure that such an $r$ exists). Now, $(r+_{\mathbb{R}^{\prime}} 1_{\mathbb{R}^{\prime}}) - 1_{\mathbb{R}^{\prime}} = r$. But then it cannot be that $1_{\mathbb{R}^{\prime}} <^{\prime}(r+_{\mathbb{R}^{\prime}} 1_{\mathbb{R}^{\prime}})$. Hence,
$N^{\prime}, s \nvDash \forall y, z ((0 < y < z \land \neg(y<z)) \rightarrow \neg \forall x (x=x) )$.
Observe that in this model $\neg \forall x (x=x) )$ is essentially just $\bot$.
Now take the model $N^{\prime \prime}$ of RCF which is just $N^{\prime}$ but with $D = \mathbb{R}$. With this modification, the ordering $<^{\prime}$ basically collapses to $<_{\mathbb{R}}$. It is easy to see then that
$N^{\prime \prime}, s \vDash \forall y, z ((0 < y < z \land \neg(y<z)) \rightarrow \neg \forall x (x=x) )$.
Finally, one can show by induction on formula complexity that if $A(x)$ is a quantifier free relevant formula in the full language of RCF and $\overline{a}$ a sequence of objects from the domain $\mathbb{R}$ of $N^{\prime \prime}$, $w \in W$, then
$N^{\prime \prime}, w \vDash A[\overline{a}]$ iff $N^{\prime }, w \vDash A[\overline{a}]$.
But with the above results this implies that the formula in the free variable $x$,
$\forall y, z ((x < y < z \land \neg(y<z)) \rightarrow \neg \forall x (x=x) )$
cannot be equivalent to any quantifierless formula of the language of RCF[^11], so the theory lacks QE.
Let **L** be any relevant logic between **B** and **RM**. Then QE fails for Presburger arithmetic in **L**.
This follows as for RCF except that we work with $\mathbb{Z}$ rather than $\mathbb{R}$, and a smaller signature. The reader can fill in the details.
Let **L** be any relevant logic between **B** and **RM**. Then QE fails for DOAG in **L**.
The proof is left to the reader and it follows as for the case of RCF again, by using $\mathbb{Q}$ rather than $\mathbb{R}$.
To end this section we wish to remark that the failure of QE presented here cannot be fixed by taking a weaker notion of QE where we merely demand interdeducibility with a quantifier free formula as opposed to equivalence in the sense of a relevant biconditional. The reason is that our arguments show that the formulas witnessing the failure of QE cannot be interdeducible with any quantifierless formula either. To see this note that, for instance, in the proof of Theorem \[t\], we actually showed that there cannot be any quantifierless formula $A(x)$ such that:
$N^{\prime \prime}, s \vDash \forall y, z ((0 < y < z \land \neg(y<z)) \rightarrow \neg \forall x (x=x) )$ iff $N^{\prime }, s \vDash A[0]$,
which means that no such quantifierless formula can be interdeducible with $\forall y, z ((x < y < z \land \neg(y<z)) \rightarrow \neg \forall x (x=x) )$.
Theories with Quantifier Elimination {#sec:qe}
====================================
In this section we study under which circumstances can we get QE back. Some proofs are simply arguments from [@Marker02], given in the Routley-Meyer semantic framework.
\[2\] Let **L** be any logic extending **B** + such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$, $M=\langle W,R,D,^*, s, V\rangle$ an **L**-model for DLO, $A(\overline{x})$ a formula of DLO, and $\overline{a}$ a sequence of elements of D. Then for any $w \in W$, $M, w \vDash A [\overline{a}]$ iff $M, s \vDash A [\overline{a}]$.
The frame condition for (the principle $A\rightarrow(B\rightarrow A))$ is:
\(1) $R\alpha \beta \gamma$ only if $Rs\alpha \gamma$.
Now we proceed by induction of the complexity of $A(\overline{x})$. Suppose first that $A(\overline{x})$ is an atomic formula of the form $x_i < y_j$. Then if $M, s \vDash A [\overline{a}]$, $M, w \vDash A [\overline{a}]$ follows by the Hereditary condition and the fact that $Rssw$ in the presence of (1) since $Rsww$. For the converse suppose that $M, w \vDash A [\overline{a}]$, which means that $M, w \vDash x_i < y_j [a_ia_j]$. Now if $M, s \nvDash x_i < y_j [a_ia_j]$, since $M$ is a model for $DLO$, either $M, s \vDash x_i = y_j [a_ia_j]$ or $M, s \vDash y_j < x_i [a_ia_j]$. If the first, it must be that $M, w \vDash x_i = y_j [a_ia_j]$ and since $(x_i = y_j \land x_i < y_j) \rightarrow x_i < x_i $ is a theorem of **L**, it follows that $M, s \vDash (x_i = y_j \land x_i < y_j) \rightarrow x_i < x_i [a_ia_j]$, so since $Rsww$, it must be that $M, w \vDash x_i < x_i [a_ia_j]$, which is impossible. If, on the other hand, $M, s \vDash y_j < x_i [a_ia_j]$, then $M, w \vDash y_j < x_i [a_ia_j]$ and given that $Rsww$ it must be the case that $M, w \vDash x_i < x_i [a_ia_j]$, i.e., $M, w \vDash x_i < x_i [a_i]$, which is impossible again. Hence, $M, s \vDash x_i < y_j [a_ia_j]$, as desired.
Suppose next that $A(\overline{x})$ is of the form $\neg B(\overline{x})$. If $M, s \vDash \neg B [\overline{a}]$ then $M, w \vDash \neg B [\overline{a}]$ by the Hereditary condition. Now if $M, s \nvDash \neg B [\overline{a}]$ then $M, s^* \vDash B [\overline{a}]$ and, by a double application of the inductive hypothesis, $M, s \vDash B [\overline{a}]$ and $M, w^* \vDash B [\overline{a}]$, which in turn implies that $M, w\nvDash \neg B [\overline{a}]$, as wanted.
Let $A(\overline{x})$ be of the form $B \rightarrow C$. One direction of the result follows by the Hereditary condition again. For the converse, suppose that $M, s \nvDash B \rightarrow C [\overline{a}]$, so there are $y, z \in W$ such that $M, y \vDash B [\overline{a}]$ and $M, z \nvDash C [\overline{a}]$. By inductive hypothesis, it follows that $M, s \vDash B [\overline{a}]$ and $M, s \nvDash C [\overline{a}]$. By assumption there are $u, v \in W$ such that $Rwuv$ (this is the only point in the proof where we need this assumption). Consequently, by inductive hypothesis again that $M, u \vDash B [\overline{a}]$ and $M, v \nvDash C [\overline{a}]$. But then $M, w \nvDash B \rightarrow C [\overline{a}]$, as desired.
The remaining cases are similarly straightforward.
\[c\] *(Cantor’s Theorem)* Let **L** be any logic extending **B** + such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$, $M=\langle W,R,D,^*, s, V\rangle, N= \langle W^{\prime},R^{\prime},D^{\prime},^{*^{\prime}}, s^{\prime}, V^{\prime}\rangle$ two countable **L**-models for DLO, and $a_1, \dots , a_n \in D$, $b_1, \dots , b_n \in D^{\prime}$ sequences such that $a_1 < \dots < a_n$ and $b_1 < \dots < b_n$ hold at $s$ and $s^{\prime}$ respectively. Then there is a bijective mapping $f: D \longrightarrow D^{\prime}$ such that $f(a_i)=b_i$ and for any formula $A(\overline{x})$ of DLO and sequence of elements $\overline{a}$ of D,
$M, s \vDash A [\overline{a}]$ iff $N, s^{\prime} \vDash A [f(\overline{a})]$.
The function $f$ is built by a back-and-forth argument.[^12] We make sure by construction that $M, s \vDash x<y [\overline{a}]$ iff $N, s^{\prime} \vDash x< y [f(\overline{a})]$.
Let $c_0, c_1, c_2 \dots $ and $d_0, d_1, d_2 \dots $ be enumerations of $D\setminus \{a_1, \dots, a_n\}$ and $D^{\prime}\setminus \{b_1, \dots, b_n\}$ respectively. Next we build a sequence $f_0 \subseteq f_1 \subseteq \dots$ of bijections $f_i: D_i \longrightarrow D^{\prime}_i $ such that $ D_i \subset D$ and $ D^{\prime}_i \subset D^{\prime}$ are finite and $M, s \vDash x<y [a, b]$ iff $N, s^{\prime} \vDash x< y [f_i(a), f_i(b)]$ for any $a, b \in D_i$. The idea is to build the sequences such that $\bigcup D_i =D$ and $\bigcup D^{\prime}_i =D^{\prime}$, so $\bigcup f_i$ would be the $f$ required by the theorem.
[Stage]{} 0: Just put $D_0 = \{a_1, \dots, a_n\}$, $D^{\prime}_0 = \{b_1, \dots, b_n\}$ and let $f_0$ be the mapping $a_i \mapsto b_i$.
[Stage]{} n+1=2m+1: When $c_m \in D_n$ then just put $D_{n+1}=D_n$, $D^{\prime}_{n+1}$ and $f_{n+1}=f_n$. On the other hand suppose that $c_m \notin D_n$. We want $D_{n+1}$ to be $D_n \cup \{c_m\}$ in this case. This means that we need to choose $d \in D^{\prime}\setminus D^{\prime}_{n}$ carefully for $f_{n+1}=f_n \cup \{\langle c_m, d\rangle\}$ and $D^{\prime}_{n+1}= D^{\prime}_{n}\cup\{d\}$ be as required by the construction.
Keep in mind that we are working with models of DLO. Hence, there are three mutually exclusive possibilities:
- $M, s \vDash x<y [c, c_m]$ for each $c \in D_n$
- $M, s \vDash x<y [c_m, c]$ for each $c \in D_n$
- there are $a, b \in D_n$ such that $M, s \vDash x<y [a, b]$, $M, s \vDash x<y < z [a, c_m, b]$ and for all $c\in D_n$ either $M, s \vDash x<y \vee x=y [c, a]$ or $M, s \vDash x<y \vee x=y [b, c]$.
If (1) holds, then since $D_n^{\prime}$ is finite, we can find $d \in D^{\prime}\setminus D^{\prime}_{n}$ such that $N, s^{\prime} \vDash x<y [c, d]$ for each $c \in D^{\prime}_{n}$ (this comes essentially from the same argument showing that $DLO$ has no models with a finite domain). A similar thing is the case when (2) holds. In case (3), given that $N, s^{\prime} \vDash x<y [f_n(a), f_n(b)]$ by construction, since $N$ is a model of $DLO$ just take $d \in D^{\prime}\setminus D^{\prime}_{n}$ such that $N, s^{\prime} \vDash x< z<y [f_n(a), d, f_n(b)]$.
[Stage]{} n+1=2m+2: When $d_m \in D^{\prime}_n$, we simply let $D_{n+1}=D_n$, $D^{\prime}_{n+1}= D^{\prime}_{n}$ and $f_{n+1}=f_n$. Otherwise, we proceed in an analogous way to the previous case making sure that $D_{n+1}$, $D^{\prime}_{n+1}$ and $f_{n+1}$ are such that $d_m \in D^{\prime}_{n+1}$.
After having constructed $f$, we establish the equivalence in the proposition by induction on the complexity of $A(\overline{x})$. If $A(\overline{x})$ is atomic, then the result follows by construction of $f$.
Suppose next that $A(\overline{x})$ is of the form $\neg B(\overline{x})$. If $M, s \vDash \neg B [\overline{a}]$, so $M, s^* \nvDash B [\overline{a}]$, hence $M, s \nvDash B [\overline{a}]$ by Proposition \[2\]. So $N, s^{\prime} \nvDash B [f(\overline{a})]$ by inductive hypothesis, and $N, s^{\prime *^{\prime}} \nvDash B [f(\overline{a})]$ by the previous theorem again, so $N, s^{\prime *^{\prime}} \vDash \neg B [f(\overline{a})]$ as desired. The converse is symmetric.
Now let $A(\overline{x})$ be of the form $B \rightarrow C$. Suppose that $M, s \vDash B \rightarrow C [\overline{a}]$. Assume further that $N, s^{\prime} \vDash B [f(\overline{a})]$, so by inductive hypothesis, it must be that $M, s \vDash B [\overline{a}]$. Since $Rsss$, $M, s \vDash C [\overline{a}]$, which, by inductive hypothesis, means that $N, s^{\prime} \vDash C [f(\overline{a})]$. Then, by Proposition \[2\], it must be that $N, w \vDash B [f(\overline{a})]$ only if $N, w \vDash C [f(\overline{a})]$ for any $w \in W^{\prime}$, so, in particular, $N, s^{\prime} \vDash B \rightarrow C [f(\overline{a})]$ holds. The converse follows by symmetry.
The remaining cases are straightforward and the quantificational case uses the fact that $f$ is surjective.
\[3\] Let **L** be any logic extending **B** + such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$. Then DLO in **L** is complete.
Suppose not, that is, there are two **L**-models $M=\langle W,R,D,^*, s, V\rangle$ and $ N= \langle W^{\prime},R^{\prime},D^{\prime},^{*^{\prime}}, s^{\prime}, V^{\prime}\rangle$ for $DLO$ such that there is a sentence $A$ of $DLO$ for which $M, s \vDash A$ but $N, s^{\prime} \nvDash A$. The cardinality of both $D$ and $D^{\prime}$ has to be infinite. Since the language of quantificational relevant logic can be embedded in a satisfaction preserving way into the language of classical first order logic, by the classical downward Löwenheim-Skolem theorem, $M$ and $N$ can both be assumed to be countable. By Proposition \[c\], they have to make exactly the same sentences hold at $s$ and $s^{\prime}$, which contradicts our assumption.
Let **L** be any logic extending **B** + such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$ and **L** is strongly complete with respect to the Routley-Meyer semantics. Then DLO in **L** is negation complete.
Take any sentence $A$ of $DLO$. Consider an arbitrary **L**-model $M$ for $DLO$. Then either $M, s \vDash A$ or $M, s \nvDash A$. The second implies $M, s^* \nvDash A$ by Proposition \[2\], with $s^*$ substituted for $w$. This is equivalent to $M, s \vDash \neg A$. Since, according to Proposition \[3\], all **L**-models of $DLO$ make exactly the same sentences true, it must be the case that either every **L**-model for $DLO$ make $A$ the case or that every **L**-model for $DLO$ make $\neg A$ hold. Hence, either $DLO \vDash A$ or $DLO \vDash \neg A$.
Let **L** be any logic extending **B** + such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$ and **L** is strongly complete with respect to the Routley-Meyer semantics. Then DLO in **L** has QE.
Take any countable model $M=\langle W,R,D,^*, s, V\rangle$ of $DLO$ in **L**. First, if $A$ is a sentence and $M, s \vDash A$, using Proposition \[2\], it must be that $M, s \vDash A \leftrightarrow x =x$, which means that $DLO \vDash A \leftrightarrow x =x$ by the completeness of $DLO$. On the other hand if $M, s \nvDash A$, we similarly obtain that $DLO \vDash A \leftrightarrow x < x$.
Suppose now that $A(\overline{x})$ is a formula in the free variables $x_1, \dots, x_n$. Let $h: \{\langle i, j\rangle : 1 \leq i < j \leq n\} \longrightarrow \{0, 1, 2\}$ and consider the formula $B_h(x_1, \dots, x_n)$ defined as follows:
$ \bigwedge_{h(i, j)=0} x_i=x_j \land \bigwedge_{h(i, j)=1} x_i<x_j \land \bigwedge_{h(i, j)=2} x_j<x_i$.
Write $\Lambda_A$ for the set of all $B_h(x_1, \dots, x_n)$ with $h: \{\langle i, j\rangle : 1 \leq i < j \leq n\} \longrightarrow \{0, 1, 2\}$ such that there is a sequence of elements $\overline{a}$ of $D$ for which $M, 0 \vDash A(x_1, \dots, x_n) \land B_h(x_1, \dots, x_n) [\overline{a}]$. $\Lambda_A$ will, of course, always be finite.
If $\Lambda_A = \emptyset$, then for all sequences $\overline{a}$ of $D$ it must be that $M, s \nvDash A[\overline{a}]$ for note that every $\overline{a}$ would satisfy some $B_h(x_1, \dots, x_n)$. This means that $M, s \vDash \forall x_1, \dots, x_n (A \leftrightarrow x_{n+1} < x_{n+1})$, which implies that $DLO \vDash \forall x_1, \dots, x_n (A \leftrightarrow x_{n+1} < x_{n+1})$.
On the other hand, if $\Lambda_A \neq \emptyset$, we claim that $M, s \vDash \forall x_1, \dots, x_n (A \leftrightarrow \bigvee \Lambda_A)$, which implies that $DLO \vDash \forall x_1, \dots, x_n (A \leftrightarrow \bigvee \Lambda_A)$. The half $M, s \vDash \forall x_1, \dots, x_n (A \rightarrow \bigvee \Lambda_A)$ is clear since, as we said, every sequence of objects of $D$ satisfies some $B_h(x_1, \dots, x_n)$. For the other half, let $a_1, \dots a_n \in D$ and $M, s \vDash B_h[a_1, \dots a_n]$ for some $B_h(x_1, \dots, x_n) \in \Lambda_A$. By definition of $\Lambda_A$ there is a sequence of objects $b_1, \dots b_n \in D$ such that $M, s \vDash A(x_1, \dots, x_n) \land B_h(x_1, \dots, x_n) [b_1, \dots b_n ]$. Now take the mapping $f: D \longrightarrow D$ such that $f(a_i)=b_i$ given by our version of Cantor’s theorem. Since $M, s \vDash A [b_1, \dots b_n ]$, it must be that $M, s \vDash A [a_1, \dots a_n ]$. This shows that $M, s \vDash \forall x_1, \dots, x_n ( \bigvee \Lambda_A \rightarrow A)$, as desired.
An example of a logic in which $DLO$ will have all the nice features described above is **B** + + (since all frames for this logic satisfy $\forall x(Rxsx)$).
Let **L** be any logic extending **B** + . Then for any model $M=\langle W,R,D,^*, s, V\rangle$ of $DLO$ in **L** and formula $A(\overline{x})$ there is a quantifier free formula $B(\overline{x})$ such that for every sequence of objects $\overline{a}$ in $D$, $M, s\vDash A[\overline{a}]$ iff $ M, s \vDash B[\overline{a}]$.
Using the downward Löwenheim-Skolem theorem and the fact that Routley-Meyer models are first order structures, it is not difficult to see that it suffices to establish the theorem for models with a countable domain of objects. By examining the proof of the previous theorem, it is easy to see how it can be adapted *mutatis mutandis* up to the case that $\Lambda_A \neq \emptyset$. At this point what is required is an instance of Cantor’s theorem saying that where $M=\langle W,R,D,^*, s, V\rangle$ is a countable **L**-model for DLO, and $a_1, \dots , a_n \in D$, $b_1, \dots , b_n \in D$ sequences such that $a_1 < \dots < a_n$ and $b_1 < \dots < b_n$ hold at $s$, then there is a bijective mapping $f: D \longrightarrow D$ such that $f(a_i)=b_i$ and for any formula $A(\overline{x})$ of DLO, $w \in W$ and sequence of elements $\overline{a}$ of D,
$M, w \vDash A [\overline{a}]$ iff $M, w \vDash A [f(\overline{a})]$.
A simple modification of the proof of Cantor’s theorem suffices to establish it. The construction of $f$ is basically as before. To prove the above equivalence we need to note first that for any sequence of objects $\overline{a}$, $M, w \vDash A [\overline{a}]$ iff $M, s \vDash A [\overline{a}]$ if $A$ is atomic. This follows by the same argument used for the basis of the inductive proof of Proposition \[2\]. Now a simple induction on formula complexity takes care of the rest.
So if the interest in QE was to reduce the complexity of the definable sets of any given model of $DLO$, **B** + is a logic extending **B** which has this effect, and we see no way to regain this result without employing , or some similar principle which has the effect of eliminating all the worlds in the Routley-Meyer model except the designated world. It is important to note that this theorem implies that one cannot use the same strategy we used to refute QE for $DLO$ in **B** in order to refute QE for $DLO$ in **B** + . For there is no model $M$ for $DLO$ in **B** + and formula $A(\overline{x})$ such that $A(\overline{x})$ does not define a set already definable by a quantifierless formula at $s$ in $M$.
Let **L** be any logic extending **B** such that its frames satisfy the condition $\forall x \exists y, z (Rxyz)$. Then ACF in **L** has QE.
We start by showing that if $M=\langle W,R,D,^*, s, V\rangle$ is an **L**-model for ACF, $A(\overline{x})$ a formula of ACF, and $\overline{a}$ a sequence of elements of D. Then for any $w \in W$, $M, w \vDash A [\overline{a}]$ iff $M, s \vDash A [\overline{a}]$. This follows by an easy induction on formula complexity as before. So, in these models, a formula of the form $A \rightarrow B$ comes down semantically to the same as a formula of the form $\neg A \vee B$, which means that all relevant formulas in these models are equivalent to $\rightarrow$-free formulas. Since ACF based on classical logic has QE, it suffices to show that every theorem of classical ACF is also a theorem of ACF in **L**. But the contrapositive of this is easily established, for if $M=\langle W,R,D,^*, s, V\rangle$ is an **L**-model for ACF and $M, s \nvDash A $ for some relevant formula $A$ (which without loss of generality can be taken to be $\rightarrow$-free), then a classical model refuting $A$ is easily extracted by interpreting function symbols in $D$ as they are interpreted at world $s$ in the Routley-Meyer model.[^13]
The above proof works, in fact, for any theory whose classical counterpart has QE and that does not contain non-logical symbols other than function and constants.
This positive result (as the one for DLO) is, of course, sensitive to a change of our background assumptions on equality. In this paper we have taken the predicate to mean real identity under all circumstances. This position is certainly rejected by most relevant logicians, though. This is because it produces validities like $A \rightarrow \forall x (x=x)$ for any arbitrary $A$. Relevant logicians would normally require just something like principles (R), (T) and (L) in the theorem below to hold. In such case, QE can be shown to fail by arguments similar to the ones in §\[sec:fqe\]. This indicates that there is very little hope for positive results on QE when $=$ is allowed all the freedom most relevant logicians would grant it. In other words, for the dedicated relevant logician, QE fails for quite simple reasons in most circumstances.[^14]
Suppose that all we require semantically from the predicate $=$ is that it satisfies the principles of reflexivity, transitivity, and a version of Leibniz’s law
- $\forall x x=x$,
- $\forall x, y, z (x=y \wedge y=z \rightarrow x=z)$,
- $\forall x, y (x=y \wedge A(x) \rightarrow A(y))$ for all relevant formulas $A$.
Let **L** be any logic between **B** and **RM**. Then QE fails for ACF in **L**.
We argue in a rather similar way as in the failure of QE for RCF. However, this time we use the classical structure $\langle \mathbb{A}, +_{\mathbb{A}}, -_{\mathbb{A}}, ^{-1_\mathbb{A}}, \times_{\mathbb{A}}, 0_{\mathbb{A}}, 1_{\mathbb{A}} \rangle$ of algebraic numbers and some uncountable elementary extension (again in the classical sense) $\langle \mathbb{A}^{\prime}, +_{\mathbb{A}^{\prime}}, -_{\mathbb{A}^{\prime}}, ^{-1_\mathbb{A}^{\prime}}, \times_{\mathbb{A}^{\prime}}, 0_{\mathbb{A}^{\prime}}, 1_{\mathbb{A}^{\prime}} \rangle$.
Now consider a Routley-Meyer model $N^{\prime}=\langle W,D,R,^*,s,V\rangle$ where
- $W=\{s, t\}$
- $D=\mathbb{A}^{\prime}$
- All constants and function symbols of the language of ACF are interpreted on $\mathbb{A}^{\prime}$ as in $\langle \mathbb{A}^{\prime}, +_{\mathbb{A}^{\prime}}, -_{\mathbb{A}^{\prime}}, ^{-1_\mathbb{A}^{\prime}}, \times_{\mathbb{A}^{\prime}}, 0_{\mathbb{A}^{\prime}}, 1_{\mathbb{A}^{\prime}} \rangle$.
- $R=\{\langle s,s,s\rangle,\langle s,t,t\rangle,\langle t,s,s\rangle,\langle t,s,t\rangle,\langle t,t,s\rangle,\langle t,t,t\rangle, \langle s,t,s\rangle\}$
- $V$ assigns to the symbol $=$ real equality on $\mathbb{A}^{\prime}$ at world $s$. At world $t$, $V$ assigns equality restricted to $\mathbb{A}$ as the interpretation of the symbol $=$. Note that since $V(=, t) \subset V(=, s)$, $V$ is a valuation in the Routley-Meyer semantics.
- $s$ is the designated world.
- $^*=\{\langle s,t\rangle,\langle t,s\rangle\}$.
This structure is a model of ACF. Next take any $r \in \mathbb{A}^{\prime}\setminus \mathbb{A}$ (recall that we made sure that such an $r$ exists since $\mathbb{A}^{\prime}$ is uncountable while $\mathbb{A}$ is countable). Now,
$N^{\prime}, s \vDash \neg(r=0) \wedge \neg (r=r)$,
so,
$N^{\prime}, s \vDash \exists x( \neg(x=0) \wedge \neg (x=x))$.
Now take the model $N^{\prime \prime}$ of ACF which is just $N^{\prime}$ but with $D = \mathbb{A}$. With this modification, the predicate $=$ becomes real equality at every world in the model, so
$N^{\prime \prime}, s \nvDash \exists x( \neg(x=0) \wedge \neg (x=x))$.
One can show by induction on formula complexity that if $A(x)$ is a quantifier free relevant formula in the language of ACF and $\overline{a}$ a sequence of objects from the domain $\mathbb{A}$ of $N^{\prime \prime}$, $w \in W$, then
$N^{\prime \prime}, w \vDash A[\overline{a}]$ iff $N^{\prime }, w \vDash A[\overline{a}]$.
But with the above results this implies that the formula in the free variable $y$,
$\exists x( \neg(x=y) \wedge \neg (x=x))$
cannot be equivalent to any quantifierless formula of the language of ACF, so the theory does not have QE.
Conclusion
==========
One of the main morals is that the property of quantifier elimination for a number of well-known mathematical theories (involving non-logical *predicate* – and not just *function* – symbols) is easily lost in many relevant logics. Naturally, this is because logics that bring the property back are nearly classical, in that they enforce strong extensional equivalences which are rejected in a relevant setting. So when working in a more fine-tuned context, with fewer principles available, a property like QE which marks a theory’s simplicity should no longer be expected to hold. As anyone who has tried to do non-classical mathematics can testify, there is a substantial jump in complexity from classical to non-classical mathematics.
It is of definite importance to the study of relevant logic and mathematics that plays a key role in capturing classical results for DLO. Thinning does play a key role, and if the relevance properties of the logic involve rejecting , then logics which respect relevance provide a view of standard mathematical theories far different from the orthodoxy. This should not be particularly surprising in itself, but what is surprising is that in a logic as weak as **B**, the addition of just and a plausible seriality condition on Routley-Meyer frames goes a long way to capturing classical results. That is, insofar as one may want QE and categoricity considerations are useful for purposes (1)–(4) in §1, at least is required (and indeed, in addition to nothing beyond **B** is required). The dedicated relevantist, classical logician, or non-partisan observer can, of course, take these results negatively, as showing that irrelevance really plays a core role in standard mathematical model-theoretic reasoning. However, there is a more interesting response available, to which we have gestured. This is that even in the case of simple mathematical theories like those considered here, a logical approach maintaining relevance opens a view which, as Meyer and Mortensen claimed, “cannot impoverish insight into the nature of mathematical structures, but rather can only enrich it” [@MeyerMortensen84].
\[sec:con\]
Acknowledgements {#acknowledgements .unnumbered}
================
There are many people deserving of thanks for input on various stages of this project. We would like to thank the audiences of the *Frontiers of Non-Classicality: Logic, Mathematics, Philosophy* workshop held in Auckland in Jan. of 2016 and the *Third Workshop* held in Edmonton in May of 2016 for their comments, and the organisers of these events. More particular thanks is due to Zach Weber and an anonymous referee for valuable commentary on drafts of this paper.
Tedder: I would like to thank the Department of Philosophy at the University of Connecticut for travel funding for the Auckland and Edmonton workshops.
Badia: I acknowledge the support by the Austrian Science Fund (FWF): project I 1923-N25 (*New perspectives on residuated posets*).
[^1]: Quantifier elimination has been recently studied in [@Ellison08] for the also non-classical context of intuitionistic logic.
[^2]: Indeed, as can be extracted from §\[sec:qe\], if $=$ is taken as a non-logical symbol then the pure theory of equality will certainly fail to have QE.
[^3]: Papers with results which are directly relevant to our work are [@Restall09; @Mortensen88] and [@Ferguson12]. Other papers in relevant arithmetic are [@Meyer76], [@MeyerMortensen84], and [@MeyerFriedman92]. For the broader class of paraconsistent arithmetics, see [@Priest97], [@Priest00], [@ParisPathmanathan06], and [@ParisSirokofskich08].
[^4]: Details on the *classical* model theoretic properties of DLO, and other mathematical theories considered here can be found in [@Marker02], or any standard model theory textbook. Our axioms for RCF in §3.2 are found in [@Sacks10].
[^5]: A semantics which brings completeness back is given in [@Mares06].
[^6]: Sans-serif uppercase letters give the Combinatory names of implication axioms, and are used throughout, in line with standard usage in the relevant literature. A good introduction to Combinatory logic can be found in [@Bimbo11], and details regarding the correspondence between combinators and implicational formulae can be found in [@AndersonBelnapDunn92]\[§71\] or in [@DunnRestall02].
[^7]: DLO with a paraconsistent underlying logic has been studied in [@Ferguson12] as formulated in the $\rightarrow$-free fragment of our language (where $A \rightarrow B$ is replaced by $\neg A \vee B$).
[^8]: For a different (essentially second order) approach to $\mathbb{R}$ in a paraconsistent setting, see [@WeberJordens12].
[^9]: It has been verified using PROVER9, a computer proof assistant McCune [@Prover9]
[^10]: The formula $\forall y, z (\neg (x<y<z) \vee x<y<z)$ is another example of something that has no quantifierless equivalent in DLO. So, in fact, $\rightarrow$ is not essential to construct such a formula.
[^11]: In fact, the presence of $\rightarrow$ is not essential here, the formula $\exists y, z ((x < y)\land(y < z) \land \neg(y<z))$ would also do the trick.
[^12]: See [@Marker02] for examples of such arguments in a classical setting.
[^13]: This kind of argument would have also established QE for DLO for logics extending **B** + and with frames satisfying seriality. However, our previous argument for DLO was much more informative.
[^14]: Note, our results do not directly rule out building theories with QE by means of Skolemization, so we cannot make this claim quite unequivocally.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We consider the complexity of the <span style="font-variant:small-caps;">Independent Set Reconfiguration</span> problem under the Token Sliding rule. In this problem we are given two independent sets of a graph and are asked if we can transform one to the other by repeatedly exchanging a vertex that is currently in the set with one of its neighbors, while maintaining the set independent. Our main result is to show that this problem is PSPACE-complete on split graphs (and hence also on chordal graphs), thus resolving an open problem in this area.
We then go on to consider the $c$-<span style="font-variant:small-caps;">Colorable Reconfiguration</span> problem under the same rule, where the constraint is now to maintain the set $c$-colorable at all times. As one may expect, a simple modification of our reduction shows that this more general problem is PSPACE-complete for all fixed $c\ge 1$ on chordal graphs. Somewhat surprisingly, we show that the same cannot be said for split graphs: we give a polynomial time ($n^{O(c)}$) algorithm for all fixed values of $c$, except $c=1$, for which the problem is PSPACE-complete. We complement our algorithm with a lower bound showing that $c$-<span style="font-variant:small-caps;">Colorable Reconfiguration</span> is W\[2\]-hard on split graphs parameterized by $c$ and the length of the solution, as well as a tight ETH-based lower bound for both parameters.
author:
- Rémy Belmonte
- Eun Jung Kim
- Michael Lampis
- Valia Mitsou
- Yota Otachi
- Florian Sikora
bibliography:
- 'reconfiguration.bib'
title: 'Token Sliding on Split Graphs[^1]'
---
[^1]: Supported by JSPS and MAEDI under the Japan-France Integrated Action Program (SAKURA) Project GRAPA 38593YJ, by FMJH program PGMO and EDF via project 2016-1760H/C16/1507 “Stability versus Optimality in Dynamic Environment Algorithmics” and project “ESIGMA” (ANR-17-CE23-0010), and by JSPS KAKENHI Grant Numbers JP18K11157, JP18K11168, JP18K11169, JP18H04091.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we study the possibility of probing the low-energy part of the spectrum of non-thermal electrons in galaxy clusters by detecting their non-thermal bremsstrahlung (NTB) emission in the soft gamma ray band, using instruments like e-ASTROGAM. Using the Coma cluster as a reference case, we find that, for very low values of the minimum energy of the electrons, in principle the NTB is detectable, but this situation is possible only for conditions that can be maintained only for a short time compared to the cluster lifetime. The possibility of constraining the low energy spectrum of non-thermal electrons through NTB is therefore hard to achieve in next years.'
author:
- 'P. Marchegiani'
title: 'Probing the low-energy spectrum of non-thermal electrons in galaxy clusters with soft gamma ray observations'
---
Introduction
============
Our knowledge of the properties of non-thermal electrons in galaxy clusters is based on the observation of diffuse radio emission like radio halos and relics, that are produced by synchrotron emission of relativistic electrons having energy of the order of GeV interacting with magnetic fields with intensity of the order of $\mu$G (e.g. Feretti et al., 2012). The same electrons are also expected to produce Inverse Compton Scattering (ICS) emission when interacting with the Cosmic Microwave Background (CMB) photons (Perola & Reinhardt, 1972); however, this emission has not yet been firmly detected even in nearby and rich clusters like Coma (e.g. Gastaldello et al., 2015).
While there are hints of a steepening of the electrons spectrum at high energies giving origin to an analogous steepening in the integrated radio halo spectrum (see Thierbach et al., 2003 for the Coma case), the lower energy part of the electrons spectrum remains basically unconstrained, with the possible exception of the detection in a few clusters of an excess in the Extreme Ultraviolet band (EUV; see e.g. Lieu et al., 1996, Bowyer et al., 2004), that can be due to ICS of electrons with energy of the order of 100–200 MeV (Sarazin & Lieu, 1998), but also to thermal bremsstrahlung from gas with low temperature (Lieu et al., 1996).
Due to the steep spectra usually found for non-thermal electrons in galaxy clusters (see e.g. Feretti et al., 2012), the low-energy part of the electrons spectrum is expected to provide the largest contribution to the total number and the total energy content of these electrons. Therefore it would be important to identify a way to probe the electrons spectrum at low energies, in order to estimate the ratio between the energy content stored in non-thermal electrons and the other components of the Intra Cluster Medium (ICM), like thermal gas, magnetic field, and non-thermal protons (see, e.g., Colafrancesco & Marchegiani, 2011 for an analogous discussion in radio galaxies).
Electrons with energies lower than $\sim100$ MeV are expected to have a short lifetime compared with the cluster one because of the Coulomb interactions with the thermal electrons, with the lifetime being shorter for lower electrons energies (e.g. Sarazin, 1999). As a results, non-thermal electrons spectra that are injected with an initial power-law shape are expected to flatten at low energies (e.g. Lieu et al., 1999); this fact can be phenomenologically parametrized through the value of the minimum Lorentz factor $\gamma_{min}$ for which the electrons spectrum can be described by a power law for $\gamma>\gamma_{min}$ (e.g. Colafrancesco & Marchegiani, 2011). The energy content stored in the non-thermal electrons is strongly depending on this value.
A possible way to probe the low-energy part of non-thermal electrons spectrum in galaxy clusters is given by the study of the non-thermal Sunyaev-Zel’dovich effect (SZE), i.e. the distortion of the CMB spectrum by ICS produced in ionized media present in cosmic structures (e.g. Colafrancesco, Marchegiani & Palladino, 2003). A non-thermal contribution to the total SZE has been possibly detected in the Bullet cluster (Colafrancesco, Marchegiani & Buonanno, 2011; Marchegiani & Colafrancesco, 2015), suggesting the possibility of a very small value of the minimum energy of the electrons ($\gamma_{min} \sim 1.4$). Also, a statistical study of the SZE in 23 radio halo clusters (Colafrancesco et al., 2014) suggested an average value of the ratio between non-thermal and thermal electrons pressures of the order of $X\equiv P_{non-th}/P_{th}\approx 0.55$. Moreover, an interpretation of a stacked analysis of Planck HFI data in galaxy clusters (Hurier, 2016) in terms of a non-thermal contribution to the SZE (Marchegiani & Colafrancesco, 2017) has indicated a low value of $\gamma_{min}$ in this sample, that corresponds to a possible pressure ratio of the order of $X\sim20-30\%$. These results are in contradiction with the values of the upper limits on the pressure ratio between non-thermal protons and thermal gas derived from Fermi-LAT gamma ray upper limits in galaxy clusters, that are of the order of $2-6\%$ (Huber et al., 2013, Prokhorov & Churazov, 2014); therefore, it would be useful to identify another possible way to detect low-energy non-thermal electrons, in order to check the validity of these results.
Non-thermal electrons with $E<100$ MeV should emit by non-thermal bremsstrahlung (NTB) when interacting with the thermal gas nuclei (e.g., Enßlin, Lieu & Biermann, 1999); this radiation should be emitted in the region of the electromagnetic spectrum under 100 MeV, that is presently almost unexplored, but can become accessibile in the future thanks to proposed instruments like e-ASTROGAM[^1] and AMEGO[^2]. Therefore in this paper we study the possibility to use NTB as a probe of low-energy non-thermal electrons at the light of the properties of these instruments, using the Coma cluster as a case of study. Coma is a rich and nearby cluster where a bright radio halo has been observed (e.g. Deiss et al., 1997), meaning that a high level of diffuse non-thermal activity is present; for these reasons, it looks as one of the candidates where in principle the observation of non-thermal emissions is more favored compared to other clusters.
We adopt in this paper a simple phenomenological model, i.e. a power-law spectrum for non-thermal electrons for $\gamma>\gamma_{min}$, and explore the consequences on the produced NTB by varying the values of $\gamma_{min}$. Since the angular resolution of the proposed instruments (de Angelis et al., 2018) is larger than the Coma cluster angular size, we consider the emission integrated on the whole cluster as a point-like source.
The outline of the paper is the following: in Section 2 we describe the theoretical methods used in this paper. In Section 3 we present the results, and summarize and discuss our results in Section 4. Throughout the paper, we use a flat, vacuum–dominated cosmological model following the results of Planck, with $\Omega_m = 0.308$, $\Omega_{\Lambda} = 0.692$ and $H_0 =67.8$ km s$^{-1}$ Mpc$^{-1}$ (Ade et al., 2016).
Methods
=======
NTB is produced by interactions of non-thermal electrons with thermal nuclei. For an electrons spectrum giving the particle density per unit of energy and volume $N_e(E,r)$, the NTB emissivity is given by: $$j_{B}(E_\gamma,r)=\int d E N_e(E,r) P_{B}(E_\gamma,E,r)$$ (e.g. Schlickeiser, 2002), where $E_\gamma$ is the energy of the emitted photon, and $$P_B(E_\gamma,E,r)=\beta c E_\gamma n_{th}(r) \sigma(E_\gamma,E),
\label{eq.pb.brem}$$ where $n_{th}(r)$ is the thermal gas density, and where the cross section for relativistic electrons is given by: $$\begin{aligned}
\sigma(E_\gamma,E) & = & \frac{1}{E_\gamma} \frac{3}{8\pi} \alpha \sigma_T \left[ \left[ 1+
\left(1-\frac{E_\gamma}{E} \right)^2 \right] \phi_1 + \right. \nonumber \\
& & - \left. \frac{2}{3}
\left(1-\frac{E_\gamma}{E} \right) \phi_2 \right],\end{aligned}$$ being $\alpha=1/137.036$ the fine-structure constant and $\sigma_T=6.652\times10^{-25}$ cm$^{2}$ the Thomson cross section. For unshielded target nuclei (like in the case of the ionized intra cluster medium) we can assume $\phi_1=\phi_2=Z^2\phi_u$, with $Z=1$ and $$\phi_u=4 \left[ \ln \left[ \frac{2E}{m_e c^2} \left( \frac{E-E_\gamma}{E_\gamma}
\right) \right] - \frac{1}{2} \right].$$ In the following we assume for the non-thermal electrons spectrum a simplified phenomenological model, assuming that they have a radial profile proportional to the thermal gas one $g_{th}(r)$, and a power-law spectrum with a low-energy cut-off: $$N_e(\gamma,r)=k_0 \gamma^{-s_e}g_{th}(r) \;\; \mbox{for } \gamma \geq \gamma_{min} ,
\label{spettroele}$$ with a typical value of the spectral index in galaxy clusters $s_e=3.7$, and where the parameter $k_0$ provides the normalization of the non-thermal electrons density.
The thermal gas density profile in the Coma cluster is derived from X-ray measures: $$n_{th}(r)=n_0 g_{th}(r) = n_0 \left[1+\left(\frac{r}{r_c}\right)^2\right]^{-q_{th}}$$ with parameters $n_0=3.4 \times 10^{-3} h_{70}^{1/2}$ cm$^{-3}$, $r_c= 300 h_{70}^{-1}$ kpc and $q_{th}=1.125$ taken from Briel, Henry & Böhringer (1992).
As effect of NTB, it is expected that an electron with energy $E$ has the maximum of its emission at approximately $E_\gamma\sim E/2$, and that, for an electrons spectrum as in eq.(\[spettroele\]), the NTB has a flux spectrum $F_B\propto E_\gamma^{-s_e+1}$ (e.g. Longair, 1994). As a consequence, we expect the NTB has a steep photon spectrum, with a sudden low-energy break at an energy $E_{\gamma}\sim \gamma_{min} m_e c^2/2$, under which the photons spectrum should strongly decrease.
Non-thermal electrons emit also by ICS with the CMB photons in a frequency range going from EUV to gamma rays, if the electrons spectrum does not steepen at high energies, with a spectral index $\alpha_{ICS}=-(s_e-1)/2$, i.e. flatter than the NTB one. The combination of NTB and ICS therefore has a flatter component, given by ICS, with a peak produced by NTB at an energy around $E_{\gamma}\sim \gamma_{min} m_e c^2/2$; the detection of such a peak can in principle allow to derive the value of $\gamma_{min}$.
Non-thermal electrons emit also by synchrotron in the radio band when interacting with the intra-cluster magnetic field. The intensity and the spatial profile of the magnetic field can be derived from Faraday Rotation measures, that in the case of Coma provide as a best fit a central value of $B_0=4.7$ $\mu$G and a radial profile proportional to $g_{th}^{1/2}(r)$ (Bonafede et al., 2010). In the following we normalize the electrons density $k_0$ by requiring that the radio flux produced at 1.4 GHz for these properties of the magnetic field is equal to the observed value of 640 mJy (Deiss et al., 1997); we choose this frequency because it is the highest frequency where the spectrum of the Coma radio halo has a power-law shape (Thierbach et al., 2003). We also note that this simple model does not take into account the possible effect of re-acceleration of electrons that can be provided by turbulences, that can boost the electrons spectrum only for electrons that emit in the radio band, i.e. with $\gamma\sim10^3-10^4$ (e.g. Brunetti et al., 2001). As a consequence, in presence of this boosting produced by turbulences, the ratio between the number of electrons emitting in radio and the ones with lower energies would be higher, and the flux produced by NTB would be reduced compared to the radio one. For this reason, in our simplified model the resulting NTB should be considered as an upper limit.
The low-part of the electrons spectrum is basically unconstrained; at low energies the electrons loose their energy mainly because of the Coulomb interactions with the thermal gas, and for low energies the electrons lifetime can be smaller than $10^8$ yrs (e.g. Sarazin, 1999); low energy electrons therefore should be present only for a short time after being accelerated. Electrons with $\gamma\sim100-200$ should instead have a longer lifetime, even comparable with the cluster one (e.g. Enßlin et al., 1999, Brunetti et al., 2001), so it is reasonable to think that a low-energy cutoff in the electrons spectrum should not be higher than these values.
In the following, we normalize the density of non-thermal electrons to the radio flux at 1.4 GHz as described, and for several values of $\gamma_{min}$ we calculate the ICS and NTB emissions in the soft gamma ray band, comparing the results with the expected sensitivity of e-ASTROGAM; we also calculate the energy content stored in non-thermal electrons compared to the thermal one, and the heating rate provided by non-thermal electrons compared with the cooling rate due to thermal bremsstrahlung, in order to check if the physical configurations associated to these assumptions are problematic for the stability of the ICM.
The energy content stored in the non-thermal electrons is given by: $$U_e = \int_{\gamma_{min}}^\infty d\gamma N_e(\gamma) (\gamma-1) m_e c^2 ,$$ that can be compared with the energy stored in the thermal gas, $U_{th}=3n_{th}k_B T_e$, that is $\sim1.3\times10^{-10}$ erg cm$^{-3}$ at the center of the Coma cluster.
Non-thermal electrons can also heat the ICM because of Coulomb interactions. The heating rate induced by the electrons having the spectrum assumed in Eq.(\[spettroele\]) is given by: $$\dot{ \epsilon}_h \equiv {d \epsilon \over dt} \Biggr|_{_{\rm heat}} = \int_{E_{min}}^{\infty } N_{e}(E)
\bigg({dE \over dt}\bigg) dE \; .
\label{eq.heating}$$ For a relativistic electron with velocity $v=\beta c$ and Lorentz factor $\gamma$, the heating rate on thermal gas due to Coulomb interactions is: $${dE\over dt}\approx K \, Z^2\,{1\over \beta}\,
\left [\ln {2\, m_{\rm e}\, c^2\, \beta^2 \gamma^2\over I_{\rm p}}
-\beta^2\right],
\label{dedx}$$ where $Z^2$ is the (suitably averaged) squared charge of the plasma’s nuclei, $K\!=\!4\,
\pi\, n_{\rm th}\, r_{\rm e}^2\, m_{\rm e}\, c^3$ and $r_{\rm e}\!=\!e^2/m_{\rm e}\,
c^2\!\simeq\!2.82$ fm. Here $I_{\rm p}\!=\!\hbar\, \omega_{\rm p}$, with $\omega_{\rm p}\!=\! [4\pi\,
n_{\rm th}\,e^2/m_{\rm e}]^{1/2}$ being the plasma frequency (e.g. Colafrancesco & Marchegiani, 2008). The heating rate can be compared with the cooling rate of the thermal gas, given by $$\dot{ \epsilon}_c \equiv {d\epsilon\over dt}\Biggr|_{_{\rm cool}}= \sqrt{2^{11}\pi^3\over 3^3}\; {e^6\sqrt{m_{\rm e} }\over h\,m_{\rm e}^2\, c^3}\,
{\bar G}\, {\bar z} \, n_{\rm th}^2\, \sqrt{kT}$$ (e.g. Longair, 1994), where $\bar z$ is an average charge of the IC plasma and ${\bar G}$ is the Gaunt factor; the cooling rate at the center of the Coma cluster is $\dot{ \epsilon}_c\sim1.6\times10^{-28}$ erg cm$^{-3}$ s$^{-1}$. The values of the heating rate are dependent on the value of the normalization of the electrons spectrum and the electrons minimum energy (see eq.\[eq.heating\]), and for the models we are using are reported in Table \[tab1\].
Results
=======
The normalization of the electrons spectrum required to not produce a radio emission in excess compared to the observed one is $k_0=1.6\times10^{-4}$ cm$^{-3}$; since the radio emission is not sensible to the low-energy spectrum of electrons, this value does not depend on $\gamma_{min}$. We consider several value of $\gamma_{min}$ between 3 and 200, and calculate the energy content stored in the relativistic electrons and the heating rate provided by them. The results are in Table \[tab1\].
[\*[4]{}[c]{}]{} $\gamma_{min}$ & $U_e/U_{th}$ & $\dot{ \epsilon}_h$ & $\dot{ \epsilon}_h/\dot{ \epsilon}_c$\
& & erg cm$^{-3}$ s$^{-1}$ &\
3 & $7.7\%$ & $1.1\times10^{-26}$ & 72\
5 & $3.3\%$ & $2.9\times10^{-27}$ & 18\
10 & $1.2\%$ & $4.5\times10^{-28}$ & 2.8\
30 & $0.20\%$ & $2.4\times10^{-29}$ & 0.15\
100 & $0.026\%$ & $1.0\times10^{-30}$ & $6.3\times10^{-3}$\
200 & $0.0073\%$ & $1.6\times10^{-31}$ & $9.9\times10^{-4}$\
The high-energy emission provided by relativistic electrons by NTB and ICS for different values of $\gamma_{min}$ is shown in Fig.\[spectrum\], where it is compared with the expected sensitivity of e-ASTROGAM for an effective exposure of one year (de Angelis et al., 2018). From this figure, it is possible to see that the expected soft gamma ray emission from diffuse relativistic electrons is below the e-ASTROGAM sensitivity for $\gamma_{min}>5$.
![High energy emission in the Coma cluster produced via ICS and NTB by non-thermal electrons with spectrum as in Eq.(\[spettroele\]) with $s_e=3.7$ and different values of the minimum energy of the electrons: $\gamma_{min}=200$ (solid line), 100 (dashed), 30 (dot-dashed), 10 (three dots-dashed), 5 (long dashed), and 3 (dotted). Plotted are also the upper limits in the hard X-ray band from NuSTAR (Gastaldello et al., 2015) and in the gamma ray from Fermi-LAT (Ackermann et al., 2010), the thermal emission from the Intra Cluster Medium (blue line), the Astro-H (HXI and SGD) sensitivity for 100 ks of time integration (from http://astro-h.isas.jaxa.jp/researchers/sim/sensitivity.html; black thick line), and the sensitivity of e-ASTROGAM for an effective exposure of one year (de Angelis et al., 2018; red solid thick line). For reference also the e-ASTROGAM sensitivity divided by a factor of 10 (red dashed thick line) and 100 (red dot-dashed thick line) is plotted.[]{data-label="spectrum"}](Marchegiani_fig1.eps){width="\columnwidth"}
For $\gamma_{min} \le 5$ the soft gamma ray emission is in principle observable with e-ASTROGAM, but the physical conditions in this case are problematic: first, the resulting energy content stored in the relativistic electrons is bigger than 3%, i.e. of the order of the upper limits on the cosmic ray protons derived from Fermi-LAT upper limits (Huber et al., 2013; Prokhorov & Churazov, 2014), while it is expected that non-thermal electrons should contain only a fraction of the energy stored in cosmic ray protons of the order of $10^{-2}$ or smaller (e.g. Vazza & Br[ü]{}ggen, 2014). Second, electrons with such a low value of minimum energy should have a very small lifetime, of the order of $10^7-10^8$ yrs because of the Coulomb losses (e.g. Sarazin, 1999), and therefore they should be present only for a small fraction of the cluster lifetime, if no steady sources of acceleration are present. Finally, as it is possible to see from Table \[tab1\], where we also show the values of the ratio between the heating and the cooling rates, the heating rate provided by electrons with low $\gamma_{min}$ is much higher than the cooling rate of the gas, and this can be a problem for the stability of the ICM. However, this last problem is not too serious if the electrons have a short lifetime: for example, in the case $\gamma_{min}=3$ the heating rate at the center of the Coma cluster is $\dot{ \epsilon}_h\sim1.1\times10^{-26}$ erg cm$^{-3}$ s$^{-1}$. Assuming the lifetime of these electrons to be of the order of $10^8$ yrs (really it is probably smaller, see Sarazin, 1999), the energy density provided at the center of the cluster during this lifetime is $\sim3.6\times10^{-11}$ erg cm$^{-3}$. This number is smaller than the energy density of the thermal gas at the center of the cluster, $\sim1.3\times10^{-10}$ erg cm$^{-3}$, therefore the heating provided by non-thermal electrons during their lifetime is not sufficient to overheat the ICM. Obviously, this fact excludes the possibility of a steady source for acceleration of electrons, and puts a limit on the time for which NTB from low-energy electrons is in principle observable.
In Fig.\[spectrum\], we also show for reference the sensitivity of hypothetical instruments with sensitivities 10 and 100 times better than e-ASTROGAM. In the first case, NTB emission would be detectable for values up to $\gamma_{min}\simgt 10$, and in the second case it would be detectable also for $\gamma_{min}\sim200$, for which the lifetime of non-thermal electrons is expected to be higher than the one of the Universe (Sarazin, 1999). However, such sensitivity levels are much better than the e-ASTROGAM one, and therefore it is hard to believe they can be reached in next years.
From Fig.\[spectrum\] it is also possible to see that in the Hard X-ray band the NTB can provide a significant contribution compared to the ICS one for small values of $\gamma_{min}$, that in principle can be observed with an instrument like Astro-H, even if in this band this emission is lower than the thermal one, so it would be necessary to disentangle thermal and non-thermal emissions.
We also calculated the flux in the EUV and Soft X-ray (SXR) bands produced by the non-thermal electrons by ICS of the CMB photons, and compared the results with the flux measured with EUVE in the 0.13–0.18 keV band, $F_{EUV}\sim4.1\times10^{-12}$ erg cm$^{-2}$ s$^{-1}$ (Bowyer et al., 2004), and with the flux measured by ROSAT in the 0.2–0.4 keV band, $F_{SXR}\sim1.1\times10^{-11}$ erg cm$^{-2}$ s$^{-1}$ (Bonamente et al., 2002). The ICS flux calculated in our model, that does not depend on the value of $\gamma_{min}$ because it is produced by electrons with $\gamma>250$, is $F_{ICS}\sim8.0\times10^{-14}$ erg cm$^{-2}$ s$^{-1}$ in the EUV band, and $F_{ICS}\sim1.4\times10^{-13}$ erg cm$^{-2}$ s$^{-1}$ in the SXR band. Both these values are much lower than the observed ones, suggesting that the observed emissions in these bands should have a different origin, like a thermal origin from a low-temperature component of the ICM (e.g. Lieu et al., 1996).
Finally, NTB does not appear to be observable in the Fermi-LAT band, where instead the emission of hadronic origin, not considered in this paper, should be dominant (e.g. Colafrancesco & Blasi, 1998).
Discussion and conclusions
==========================
The study of the low-energy part of the spectrum of relativistic electrons in galaxy clusters by detecting their NTB appears to be a hard task with the next generation of instruments operating in the soft gamma ray band.
As pointed out by Petrosian (2001), NTB is inefficient compared with the Coulomb losses rate: the consequence is that the electrons with low energy have a very short lifetime compared to the cluster one, and that the heating rate provided by these electrons is very high, while the NTB emission is not very strong.
We have found that in principle there is a narrow region of the parameters space (with $\gamma_{min}\le 5$) where the NTB in the Coma cluster can be observed with e-ASTROGAM. However, these electrons can be present only for a short fraction of the cluster lifetime because of their short lifetime, and because otherwise they would overheat the cluster gas. Also, the energy content of these electrons should be of the order of the upper limits on cosmic ray protons energy content in galaxy clusters: this can be a problem because, on the basis of our present understanding of mechanisms of particles acceleration in galaxy clusters, one should expect that the energy content of the electrons should be lower than the one of protons (e.g. Vazza & Brüggen, 2014), even if some attempt to identify some acceleration mechanism efficient on electrons but not on protons in galaxy clusters has been done (e.g. Wittor, Vazza & Brüggen, 2016).
On the other hand, since several studies of non-thermal SZE in galaxy clusters (Colafrancesco et al., 2011, 2014; Marchegiani & Colafrancesco, 2017) indicate a possible low value of $\gamma_{min}$ and a high energy content stored in non-thermal electrons, it still appears useful to try to detect galaxy clusters with e-ASTROGAM in order to check these results.
It is also necessary to point out that, even if a cluster would be in principle observable, other issues could be present in order to establish if the observed emission would be really produced by diffuse electrons. In fact, the expected angular resolution of e-ASTROGAM, although improved compared to previous gamma ray instruments, in the energy range 30–100 MeV is of the order of 1.5 degrees (de Angelis et al., 2018), therefore basically all cluster core regions (including Coma) should appear as point-like sources; therefore, the presence of other sources with a soft gamma ray emission (e.g. AGNs) in the field of view of the instrument would be difficult to remove from the cluster one.
Another possible problem can be the gamma ray emission produced by decay of neutral pions produced in hadronic interactions between cosmic ray protons and thermal gas nuclei (e.g. Colafrancesco & Blasi, 1998). Although still not firmly detected in galaxy clusters (e.g. Ackermann et al., 2016), this emission should be dominant at energies of the order of 70 MeV or more, and this is another factor that limits the possibility to detect the NTB in galaxy clusters only to the case of low values of $\gamma_{min}$.
In order to improve the situation, it would be necessary an instrument with a sensitivity at least 10 times (or preferably more) better than e-ASTROGAM. However, considering that this instrument adopts very advanced technological solutions for the present day, it is difficult to think that these improved sensitivity levels can be reached in next years. We also note that, since most of other rich clusters are more distant than Coma, the situation is not expected to be easier in other clusters, especially in clusters that do not host a radio halo, suggesting that the non-thermal activity in these clusters can be lower than in the Coma case.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is based on the research supported by the South African Research Chairs Initiative of the Department of Science and Technology and National Research Foundation of South Africa (Grant No 77948). PM acknowledges support from the Department of Science and Technology/National Research Foundation (DST/NRF) Square Kilometre Array (SKA) post-graduate bursary initiative under the same Grant. I thank the Referee for useful comments and suggestions.
Ackermann, M., et al., 2010, ApJL, 717, L71
Ackermann, M., et al., 2016, ApJ, 819, 149
Ade, P. A. R., et al., 2016, A&A, 594, A13
Bonafede, A., Feretti, L., Murgia, M., et al., 2010, A&A, 513, A30
Bonamente, M., Lieu, R., Joy, M. K., & Nevalainen, J. H., 2002, ApJ, 576, 688
Bowyer, S., Korpela, E. J., Lampton, M., & Jones, T. W., 2004, ApJ, 605, 168
Briel, U. G., Henry, J. P., & Böhringer, H., 1992, A&A, 259, L31
Brunetti, G., Setti, G., Feretti, L., & Giovannini, G., 2001, MNRAS, 320, 365
Colafrancesco, S., & Blasi, P., 1998, Astroparticle Physics, 9, 227
Colafrancesco, S., & Marchegiani, P., 2008, A&A, 484, 51
Colafrancesco, S., & Marchegiani, P., 2011, A&A, 535, A108
Colafrancesco, S., Marchegiani, P., & Palladino, E., 2003, A&A, 397, 27
Colafrancesco, S., Marchegiani, P., & Buonanno, R., 2011, A&A, 527, L1
Colafrancesco, S., Emritte, M. S., Mhlahlo, N., & Marchegiani, P., 2014, A&A, 566, A42
de Angelis, A., et al., 2018, Journal of High Energy Astrophysics, 19, 1
Deiss, B. M., Reich, W., Lesch, H., & Wielebinski, R., 1997, A&A, 321, 55
Enßlin, T. A., Lieu, R., & Biermann, P. L., 1999, A&A, 344, 409
Feretti, L., Giovannini, G., Govoni, F., & Murgia, M., 2012, A&ARv, 20, 54
Gastaldello, F., et al., 2015, ApJ, 800, 139
Huber, B., Tchernin, C., Eckert, et al., 2013, A&A, 560, A64
Hurier, G., 2016, A&A, 596, A61
Lieu, R., Mittaz, J., Bowyer, S., et al., 1996, Science, 274, 1335
Lieu, R., Ip, W.-H., Axford, W. I., & Bonamente, M., 1999, ApJ, 510, L25
Longair, M., 1994, *High Energy Astrophysics*, Cambridge University Press
Marchegiani, P., & Colafrancesco, S., 2015, MNRAS, 452, 1328
Marchegiani, P., & Colafrancesco, S., 2017, MNRAS, 469, 4644
Perola, G. C., & Reinhardt, M., 1972, A&A, 17, 432
Petrosian, V., 2001, ApJ, 557, 560
Prokhorov, D. A., & Churazov, E. M., 2014, A&A, 567, A93
Sarazin, C. L., 1999, ApJ, 520, 529
Sarazin, C. L., & Lieu, R., 1998, ApJ, 494, L177
Schlickeiser, R., 2002, *Cosmic Ray Astrophysics*, Springer-Verlag, Berlin
Thierbach, M., Klein, U., & Wielebinski, R. 2003, A&A, 397, 53
Vazza, F., & Br[ü]{}ggen, M., 2014, MNRAS, 437, 2291
Wittor, D., Vazza, F., & Brüggen, M. 2016, Galaxies, 4, 71
[^1]: http://eastrogam.iaps.inaf.it/
[^2]: https://asd.gsfc.nasa.gov/amego/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We have shown earlier that hyperfine spectroscopy in a vapor cell using co-propagating pump-probe beams has many advantages over the usual technique of saturated-absorption spectroscopy using counter-propagating beams. The main advantages are the absence of crossover resonances, the appearance of the signal on a flat (Doppler-free) background, and the higher signal-to-noise ratio of the primary peaks. Interaction with non-zero-velocity atoms causes additional peaks, but only one of them appears within the primary spectrum. We first illustrate the advantages of this technique for high-resolution spectroscopy by studying the $D_2$ line of Rb. We then use an acousto-optic modulator (AOM) for frequency calibration to make precise hyperfine-interval measurements in the first excited $P_{3/2}$ state of $^{85,87}$Rb and $^{133}$Cs.'
author:
- 'Alok K. Singh, Sapam Ranjita Chanu, Dipankar Kaundilya and Vasant Natarajan'
title: 'Hyperfine spectroscopy using co-propagating pump-probe beams'
---
Introduction
============
High-resolution laser spectroscopy has been revolutionized in the last two decades with the advent of low-cost tunable diode lasers [@WIH91]. These diodes, when placed in an external cavity with optical feedback, have frequency uncertainty of about 1 MHz, which is small enough for hyperfine transitions in atoms to be resolved [@MSW92; @BRW01]. Hyperfine spectroscopy, particularly in the low-lying electronic states of alkali-metal atoms, plays an important role in fine-tuning atomic wavefunctions used in theoretical calculations. This is because comparison between theoretical and experimental determinations of hyperfine structure provides a stringent test of atomic calculations in the vicinity of the nucleus [@SJD99]. In addition, hyperfine structure in these multielectron atoms is sensitive to core polarization and core correlation effects [@AIV77].
Many of the alkali-metal atoms have transitions to the first-excited state (so-called $D$ lines) which are in the near infrared, and therefore accessible with diode lasers. They also have a high-enough vapor pressure near room temperature that the spectroscopy can be done in a vapor cell. The thermal motion of the atoms inside the cell causes [*Doppler broadening*]{}, which is typically 100 times larger than the natural width of the hyperfine transitions. The standard technique to overcome the first-order Doppler effect is to use a [*counter-propagating*]{} pump beam to saturate the transition for zero-velocity atoms, in what is called saturated-absorption spectroscopy (SAS) [@DEM82].
Most atoms also have several closely-spaced hyperfine levels within the Doppler profile. In these cases, it is well known that SAS also produces spurious crossover resonances in between each pair of hyperfine transitions. They occur because, for some non-zero-velocity group, the pump drives one transition while the probe drives the other. In earlier work [@BAN03], we have shown that the use of co-propagating pump and probe beams overcomes the problem of crossover resonances. Closely-spaced levels that are not resolved in SAS can be resolved by this technique. Probe transmission in such multilevel atoms is caused by the phenomenon of electromagnetically induced transparency (EIT) [@HAR97; @DAN05], and population depletion due to optical pumping. Furthermore, by scanning only the pump beam, the signal appears on a flat background without the underlying Doppler profile seen in SAS. This is advantageous for applications such as laser locking or laser frequency measurement.
In this work, we present a complete study of the spectra taken with the co-propagating technique in the $D_2$ line of the two isotopes of Rb. In particular, we show that the effect of non-zero velocity groups is to cause additional peaks. However, in contrast to SAS, only one of these spurious peaks appears within the spectrum, and the real peaks are unaffected. Interestingly, the spurious peak within the spectrum is almost negligible for transitions starting from the upper ground hyperfine level, but highly prominent for transitions starting from the lower level. Such differences between the two levels have also been seen in EIT, arising from the fact that the closed transition is $F \rightarrow (F+1)$ for the upper hyperfine set and $F
\rightarrow (F-1)$ for the lower hyperfine set.
For a second set of experiments, we have used a single laser along with an acousto-optic modulator (AOM) to produce the pump-probe beams with a precisely-controlled frequency offset. We then obtain the entire spectrum by scanning the frequency of the AOM. The scan axis is guaranteed to be linear because it is determined by the frequency of the rf oscillator driving the AOM. A curve fit to the observed spectrum yields the hyperfine interval. The measurements have an accuracy of 20 kHz, which is comparable to the accuracy of other techniques.
In recent work from our laboratory, we have reported high-accuracy values for the hyperfine constants in the $D$ lines of all alkali atoms [@DAN08]. The hyperfine intervals in that work were obtained (with an accuracy of 6 kHz) by [*locking*]{} the AOM to the neighboring transition. One of the uncertainties when locking the AOM is whether the lock point is exactly at the center of the peak, since any shift would cause a systematic error in the measurement. Though we had done experiments to verify that this error was less than 2 kHz, we wanted to repeat the measurements with another technique that was not at all susceptible to errors arising from lock-point uncertainty. As we will see below, the co-propagating technique achieves precisely this. In addition, by measuring the entire spectrum and looking at the symmetry of the line shape, we can be sure that other sources of error are not significant. The current set of measurements in $^{85}$Rb and $^{133}$Cs, although having slightly smaller precision, are consistent with our earlier work.
Spectroscopy on the $D_2$ line of rubidium
==========================================
The schematic for the first set of experiments is shown in Fig. \[schema\]. The pump and probe beams are derived from two home-built frequency-stabilized diode laser systems [@BRW01] operating on the 780 nm $D_2$ line of Rb. The linewidth of the lasers after stabilization is of the order of 1 MHz. The output beams are elliptical with $1/e^2$ size of $2 \times 4$ mm and powers of around 10 $\mu$W each. Part of the probe laser is sent into a Rb SAS cell and the laser is locked to a hyperfine transition using fm modulation spectroscopy. The pump laser is scanned around the same set of transitions. The beams are mixed in a polarizing beamsplitter cube (PBS) and [*copropagate*]{} through a room-temperature vapor cell (5 cm long) with orthogonal linear polarizations. Halfwave retardation plates in the path of each beam allow precise control of their powers. The probe beam is separated using a second PBS, and its transmitted signal is detected with a photodiode. The PBS’s have extinction ratios of $1000:1$, ensuring good purity of the detected signal.
Spectrum in $^{87}$Rb, $F=2 \rightarrow F'=1,2,3$
-------------------------------------------------
As mentioned before, the main advantage of the co-propagating configuration is the absence of spurious crossover resonances. This difference is seen clearly in Fig. \[satabs\]. In (a), we show the usual saturated-absorption spectrum for the $F=2 \rightarrow
F'=1,2,3$ transitions in $^{87}$Rb. The spectrum is Doppler corrected, which is necessary because probe absorption through a vapor cell will show a broad Doppler profile when the probe addresses a velocity group different from that resonant with the pump. When the pump and probe are resonant with the same velocity class, we get transmission peaks. As expected, there are three hyperfine peaks and three crossovers. The crossovers are more prominent than the actual peaks because two velocity classes contribute to each crossover resonance, compared to one (zero-velocity) class for each hyperfine peak. Probe transparency is primarily caused by two effects: (i) saturation of absorption caused by the strong pump beam, and (ii) optical pumping into the $F=1$ ground hyperfine level for open transitions (i.e. those involving the $F'=1$ and 2 excited levels). In addition, there will be population redistribution among the magnetic sublevels, which can cause increased absorption or transparency depending on the $F$ values of the levels. The linewidth of the peaks in the figure is about 12 MHz, compared to the natural linewidth of 6 MHz. This increase is typical in SAS and arises due to a misalignment angle between the beams and power broadening.
(a)\
(b)\
(c)
Now let us consider the spectrum shown in (b) taken with the co-propagating configuration. The probe is locked to the $F=2 \rightarrow 3$ transition and the pump is scanned across the set of $F=2 \rightarrow F'=1,2,3$ transitions. Since the probe is locked, its transmitted signal primarily corresponds to absorption by zero-velocity atoms (i.e.atoms moving perpendicular to the laser beam) making transitions to the $F'=3$ level. The signal remains flat (or Doppler free) until the pump also comes into resonance with a transition for the same zero-velocity atoms. Thus there are three transmission peaks at the locations of the hyperfine transitions, with no crossover resonances in between. The hyperfine peaks are located at $-423.600$ MHz, $-266.657$ MHz, and 0 [@DAN08], all measured with respect to the frequency of the locked probe laser. The linewidth of the peaks is about 19 MHz, which is only 50% larger than the linewidth obtained in the saturated-absorption spectrum. The primary cause for the transparency peaks is the phenomenon of EIT in this V-type system. The pump laser causes an [*AC Stark shift of the ground level*]{} (creation of dressed states [@COR77]) and hence reduces probe absorption at line center. In addition, there are effects of saturation and optical pumping, but these are less important than the EIT effects.
Since the experiments are done in a vapor cell with the full Maxwell-Boltzmann distribution of velocities, we have to consider that there will be two additional velocity classes that absorb from the locked probe: both moving in the same direction as the probe but with velocities such that one drives transitions to the $F'=2$ level (266.657 MHz lower) and the second to the $F'=1$ level (423.600 MHz lower). Each of these will cause three additional transparency peaks from the mechanisms discussed above. The first velocity class moves at 208 m/s and will cause peaks at $-156.943$ MHz, 0, and $+266.657$ MHz, i.e., a set of peaks shifted up by 266.657 MHz. The second velocity class moves at 330 m/s and will cause peaks at 0, $156.943$ MHz, and $+423.600$ MHz, i.e., a set of peaks shifted up by 423.600 MHz. Thus there will be 7 peaks in all, with 3 real peaks and 4 spurious ones. However, only the peak at $-156.943$ MHz will appear within the spectrum, caused by the probe driving the $F=2 \rightarrow F'=2$ transition and the pump driving the $F=2 \rightarrow F'=1$ transition. The other three spurious peaks will lie outside the spectrum to the right. This is indeed what is observed in Fig.\[satabs\](b): there is a small peak at $-157$ MHz within the spectrum.
The above explanation is borne out by the calculated spectrum shown in Fig. \[satabs\](c). Using a density-matrix formulation, we can calculate the absorption of a probe laser in a V-type system [@DAN05]. The calculation is done for multiple hyperfine levels with full thermal averaging. The only adjustable parameters are the relative amplitudes of the three EIT resonances. As seen from the figure, the calculation reproduces the locations of the peaks in the measured spectrum. The observed linewidth is slightly larger than the calculated one, but this could be because of a small misalignment angle between the beams, which is known to broaden the EIT resonance [@CAT04]. The calculation shows that there is only one spurious peak within the spectrum. However, if we extend the calculation up to $+500$ MHz, we see all the seven peaks mentioned in the previous paragraph. The extended calculation is shown in Fig. \[theory\].
The advantages of this scheme are quite clear from Fig.\[satabs\](b). The spectrum appears on a flat background, obviating the need for Doppler subtraction as in the case of SAS. There are no crossover resonances, which often swamp the true peaks. And there is only one additional peak within the spectrum due to absorption by non-zero velocity atoms. In Fig. ref[coprop]{}, we show the effect of pump power on the peaks. As the power is varied from $0.33$ to $1.66$ times the probe power, the three main peaks remain quite prominent with good signal-to-noise ratio. The additional peak at $-157$ MHz increases in height, but not significantly. By comparison, a good saturated-absorption spectrum requires the pump-probe power ratio to be accurately controlled to a value of $3:1$, with loss in signal at lower pump powers and power broadening at higher powers. In Fig. \[coprop\](b), we show a multipeak Lorentzian fit to the spectrum measured with a pump power of 15 $\mu$W. The residuals show that the line shape of all the peaks is Lorentzian.
(a)\
(b)
Spectrum in $^{87}$Rb, $F=1 \rightarrow F'=0,1,2$
-------------------------------------------------
The same advantages are seen in the spectrum of transitions starting from the lower ground hyperfine level ($F=1
\rightarrow F'=0,1,2$) shown in Fig. \[rb10\]. The Doppler-subtracted saturated-absorption spectrum on top has 6 peaks including the 3 crossover resonances. The spectrum with the co-propagating beams shown below is taken with the probe locked to the $F=1 \rightarrow 0$ transition. The beam powers are 9 $\mu$W (probe) and 15 $\mu$W (pump). It appears on a flat background and shows the 3 hyperfine peaks without any crossovers in between. The 3 hyperfine peaks are located at 0, $+72.223$ MHz, and $+229.166$ MHz. Two of the additional peaks are seen, one at $-72.223$ MHz (outside the spectrum) and the other at $+156.943$ MHz (within the spectrum). The additional peak within the spectrum is due to atoms moving with a velocity of 52 m/s such that the probe drives the $F=1 \rightarrow F'=1$ transition and the pump drives the $F=1 \rightarrow F'=2$ transition. This spurious peak is more prominent compared to transitions starting from the upper ground level \[see Fig. \[satabs\](b)\] (though the real peaks still have high signal-to-noise ratio). The difference arises due to the fact that the closed transition for this set is the $F=1 \rightarrow F'=0$ transition, which has fewer magnetic sublevels in the excited state compared to the ground state. This leads to population trapping in the $m_F = \pm
1$ sublevels, and the relative importance of EIT effects in causing probe transparency (which is the same for all peaks) increases.
Spectrum in $^{85}$Rb, $F=3 \rightarrow F'=2,3,4$
-------------------------------------------------
The improvement with this technique is much more dramatic in the spectra of the other isotope, $^{85}$Rb. For transitions starting from the upper ground level ($F=3
\rightarrow F'=2,3,4$) shown in Fig. \[rb34\], the hyperfine peaks corresponding to $F'=2$ and 4 in the saturated-absorption spectrum are barely visible. In the co-propagating spectrum shown below, the peaks become prominent. The hyperfine intervals [@DAN08] are such that the real peaks are at $-184.390$ MHz, $-120.966$ MHz, and 0, while the additional peaks are at $-63.424$ MHz, $+63.424$ MHz, $+120.966$ MHz, and $+184.390$ MHz. The additional peak within the spectrum (at $-63.424$ MHz) is almost negligible, as was observed for upper-level transitions in $^{87}$Rb. The beam powers are 9 $\mu$W (probe) and 15 $\mu$W (pump).
Spectrum in $^{85}$Rb, $F=2 \rightarrow F'=1,2,3$
-------------------------------------------------
For transitions starting from the lower ground level ($F=2
\rightarrow F'=1,2,3$) shown in Fig. \[rb21\], the hyperfine interval between $F'=1$ and 2 is so small that the crossover resonance in the saturated-absorption spectrum completely swamps the $F'=1$ peak. However, the spectrum with the co-propagating technique shows the peak well resolved. The real peaks are located at 0, $+29.268$ MHz, and $+92.692$ MHz, while the additional peaks are at $-92.692$ MHz, $-63.424$ MHz, $-29.268$ MHz, and $+63.424$ MHz [@DAN08]. The additional peak within the spectrum (at $+63.424$ MHz) is quite prominent as in the case of transitions starting from the lower hyperfine level in $^{87}$Rb, again because the closed $F=2 \rightarrow F'=1$ transition has population trapping in the $m_F = \pm 2$ sublevels. There are two additional peaks appearing outside the spectrum to the left, which are closer because of the smaller hyperfine intervals. The beam powers are 9 $\mu$W (probe) and 15 $\mu$W (pump).
Hyperfine measurements using an acousto-optic modulator
=======================================================
The above experiments were done using separate pump and probe lasers. However, it is possible to do the experiment with just one laser by using an AOM to produce the scanning pump beam. The scan range of an AOM is limited to about 20 MHz, but this is large enough to scan across a hyperfine peak. The main advantage of using an AOM is that the frequency-scan axis (with respect to the probe beam) is both linear and calibrated by the rf frequency of the driver powering the AOM, thus allowing the hyperfine interval to be measured accurately. By measuring the entire peak, potential systematic errors due to locking of the pump laser to a peak are avoided. In addition, if there is a systematic shift in the lock point of the probe laser, this will not cause an error in the interval because the frequency which brings the pump into resonance will also be shifted by the same amount, and hence the AOM offset (with respect to the probe) for the spectrum will remain the same.
We have therefore used a single laser and a scanning AOM to measure hyperfine intervals in the $D_2$ lines of Rb and Cs. The measurements are motivated by the fact that there are several experimental values reported in the literature that are somewhat discrepant from each other. In many cases, we feel that a potential source of error is the uncertainty in locking to a peak. In our current technique, measuring the entire spectrum avoids such errors, as discussed before.
The experimental schematic for this second set of experiments is shown in Fig. \[schema2\]. As before, the primary laser is a frequency-stabilized diode laser. The probe beam is derived after locking the laser to a hyperfine transition using SAS. The scanning pump beam is frequency offset from the probe using a double-passed AOM. The frequency is adjusted so that the pump is resonant with a nearby hyperfine transition whose interval has to be measured. The pump intensity is stabilized to better than 1% in a servo-loop by controlling the rf power driving the AOM. The two beams co-propagate through a vapor cell kept inside a magnetic shield. The residual field (measured with a three-axis fluxgate magnetometer) is below 5 mG. The beams have orthogonal linear polarizations and are mixed and separated using PBS’s. The beam powers are about 15 $\mu$W each and adjusted to get good signal-to-noise ratio in the spectrum.
The experiment proceeds as follows. An rf frequency generator whose timebase is referenced to an ovenized quartz clock (uncertainty less than $10^{-8}$) is used to drive the AOM. A computer program is used to set the rf frequency, and the probe signal is measured and recorded. The frequency is changed in steps of 0.1 MHz over a range of 15 MHz to obtain the complete spectrum. A curve fit to the spectrum yields the AOM frequency at the peak center, which is the hyperfine interval.
Error analysis
--------------
Systematic errors can arise due to one of the following reasons.
1. [*Radiation-pressure effects.*]{} Radiation pressure causes velocity redistribution of the atoms in the vapor cell. In the SAS technique, the opposite Doppler shifts for the counter-propagating beams can result in asymmetry of the observed lineshape. However, with co-propagating beams, the effects are less important because the Doppler shift will be the same for both beams and will not affect the hyperfine interval, similar to how the interval is insensitive to any detuning of the probe from resonance.
2. [*Effect of stray magnetic fields.*]{} The primary effect of a magnetic field is to split the Zeeman sublevels and broaden the line without affecting the line center. However, line shifts can occur if there is asymmetric optical pumping into Zeeman sublevels. For a transition $\ket{F,m_F}
\rightarrow \ket{F',m_{F'}}$, the systematic shift of the line center is $\mu_B(g_{F'}m_{F'} -
g_Fm_F)B$, where $\mu_B=1.4$ MHz/G is the Bohr magneton, $g$’s denote the Landé $g$ factors of the two levels, and $B$ is the magnetic field. The selection rule for dipole transitions is $\Delta m
= 0,\pm 1$, depending on the direction of the magnetic field and the polarization of the light. Thus, if the beams are linearly polarized, there will be no asymmetric driving and the line center will not be shifted. We therefore minimize this error in two ways. First, we use polarizing cubes to ensure that the beams have near-perfect linear polarization. Second, we use a magnetic shield around the cell to minimize the field.
The experiment is repeated by reversing the scan direction to check for errors that might depend on which direction the rf generator is scanned. Another source of error is whether the intensity stabilization servo-loop stays locked. But if this loses lock, it shows up in the spectrum as an asymmetry of the line shape. Indeed, both the sources of error discussed above also show up as asymmetry of the line shape. Thus a symmetric line shape is a good indication that the measurement proceeded correctly. From the residual asymmetry, we estimate the systematic errors to be 20 kHz.
Measurements in the $5P_{3/2}$ state of $^{87}$Rb
-------------------------------------------------
The first set of measurements were done in $^{87}$Rb. The different values in the literature are consistent with each other, and have an accuracy of 10 kHz. Therefore, our main motivation was to see if the scanning-AOM technique worked well and our error budget was proper.
A typical spectrum with the probe locked to the $F=2
\rightarrow F'=3$ transition and pump scanning across the $F=2 \rightarrow F'=2$ transition is shown in Fig.\[rb22\]. We saw earlier that the line shape was well described by a Lorentzian. We therefore fit a Lorentzian curve to the spectrum and extract the peak center. Note the symmetry of the spectrum and the high signal-to-noise ratio. The $\{ 3-2 \}$ interval is twice the center frequency (because the AOM is double passed). For technical reasons, there is an additional AOM with a fixed frequency in the path of the probe, and this offset has to be added to obtain the interval.
The average value from 14 individual measurements is listed in Table \[t1\]. The standard deviation of the set is 32 kHz, which means the expected error in the mean is $32/\sqrt{14} = 8.6$ kHz, less than our estimated error of 20 kHz. This value is compared to other values reported in the literature. The two most accurate measurements [@DAN08; @YSJ96] have uncertainties below 10 kHz and have overlapping error bars. The more recent measurement [@DAN08] is also from our laboratory and used an AOM to measure the interval, but the AOM was locked to the peak. The current measurement obtained by measuring the entire spectrum is consistent with this value, thus giving confidence in the current technique. The only slightly discrepant measurement is from the work in Ref.[@BGR91], which is $1.7 \sigma$ away.
$\{ 3-2 \} $ Interval Reference
----------------------- -----------
266.653(20) This work
266.657(8) [@DAN08]
266.650(9) [@YSJ96]
266.503(84) [@BGR91]
: Comparison of measurements of hyperfine intervals in the $5P_{3/2}$ state of $^{87}$Rb to previous results. The last row is calculated from the $A$ and $B$ coefficients reported therein. All values in MHz.
\[t1\]
Measurements in the $5P_{3/2}$ state of $^{85}$Rb
-------------------------------------------------
With the reliability of the technique established with measurements in $^{87}$Rb, we turned to the other isotope, namely $^{85}$Rb. The probe was locked to the $F=3
\rightarrow F'=4$ transition, and the pump was scanned either across the $F=3 \rightarrow F'=3$ transition or the $F=3 \rightarrow F'=2$ transition. Typical spectra for the two cases are shown in Fig. \[rb33\]. As before, Lorentzian fits to the measured spectra were used to determine the peak center, and thus the $\{ 4-3 \}$ and $\{
3-2 \}$ intervals.
\(a) (b)
The average values for the two intervals are listed in Table \[t2\]. For the $\{ 4-3 \}$ interval, the standard deviation from a set of 14 measurements is 35 kHz. For the $\{ 3-2 \}$ interval, the standard deviation from 11 measurements is 34 kHz. These values are also compared to other values in the literature. There are two non-overlapping sets for the $\{ 4-3 \}$ interval. The value from Ref. [@BGR91] is $3.7 \sigma$ away from the other two values, which are both from our laboratory and both of which relied on AOM locking. Ref. [@BGR91] is also the work in which the value in $^{87}$Rb was discrepant by $1.7 \sigma$. The current measurement is consistent with our previous ones. All the values for the $\{ 3-2 \}$ interval are consistent with each other.
$\{ 4-3 \} $ Interval $\{ 3-2 \} $ Interval Reference
----------------------- ----------------------- -----------
120.958(20) 63.436(20) This work
120.966(8) 63.424(6) [@DAN08]
120.960(20) 63.420(31) [@RKN03]
120.506(124) 63.402(93) [@BGR91]
: Comparison of measurements of hyperfine intervals in the $5P_{3/2}$ state of $^{85}$Rb to previous results. The last row is calculated from the $A$ and $B$ coefficients reported therein. All values in MHz.
\[t2\]
Measurements in the $6P_{3/2}$ state of $^{133}$Cs
--------------------------------------------------
The next set of measurements was done on the $D_2$ line in $^{133}$Cs at 852 nm. For this, the probe was locked to the $F=4 \rightarrow F'=5$ transition and the pump was scanned either across the $F=4 \rightarrow F'=4$ transition or the $F=4 \rightarrow F'=3$ transition. Representative spectra for the two cases are shown in Fig. \[cs44\]. Lorentzian fits to the spectra yielded the line center and hence the $\{ 5-4 \}$ and $\{ 4-3 \}$ intervals.
\(a) (b)
The average value for the $\{ 5-4 \}$ interval is 251.031(20) MHz, as listed in Table \[t3\]. This was obtained from a set of 27 independent measurements with a standard deviation of 21 kHz. We concentrated on this interval because two of the values reported in the literature, 251.092(2) MHz from Ref. [@GDT03] and 251.000(20) from Ref. [@TAW88], differ by $4.5
\sigma$. The more recent measurement of the two [@GDT03] was done using a frequency comb. An earlier measurement from our laboratory using the AOM locking technique [@DAN05] yielded a result of 251.037(6) MHz, which was consistent with the earlier value at the $1.5
\sigma$ level, but totally inconsistent with the frequency-comb result (difference of $9\sigma$). Our current value vindicates our earlier result since it is consistent with the work in Ref. [@TAW88] but not with the frequency-comb result.
$\{ 5-4 \} $ Interval $\{ 4-3 \} $ Interval Reference
----------------------- ----------------------- -----------
251.031(20) 201.260(20) This work
251.037(6) 201.266(6) [@DAN05]
251.092(2) 201.287(1) [@GDT03]
251.000(20) 201.240(20) [@TAW88]
: Comparison of measurements of hyperfine intervals in the $6P_{3/2}$ state of $^{133}$Cs to previous results. All values in MHz.
\[t3\]
For the $\{ 4-3 \}$ interval, we obtain an average value of 201.260(20) MHz from a set of 10 measurements with a standard deviation of 33 kHz. The value from the work in Ref. [@TAW88] was 201.240(20) MHz, while the more recent frequency-comb work in Ref. [@GDT03] reported a value of 201.287(1) MHz. The inconsistency of $2.4 \sigma$ is smaller but still quite significant. Again, our previous result of 201.266(6) MHz obtained with AOM locking [@DAN05] overlapped with the earlier value but was inconsistent with the frequency-comb result. Our new value, though with a larger error bar, gives confidence in the previous measurement.
Conclusions
===========
In summary, we have shown that hyperfine spectroscopy with co-propagating beams in a vapor cell has several advantages over conventional saturated-absorption spectroscopy. In addition to the usual mechanisms responsible for probe transparency, there are EIT effects that enhance the peaks, which is supported by density-matrix calculations. As a result, the primary peaks are more prominent and appear with good signal-to-noise ratio. The transmitted signal appears on a flat background (Doppler-free) and does not have the problem of crossover resonances in between hyperfine transitions (which are stronger and often swamp the true peaks). Absorption by non-zero velocity groups causes additional peaks, but only one of them appears within the spectrum. These observations are again supported by density-matrix calculations taking the thermal velocity distribution into account. An important difference between transitions starting from the upper ground level and transitions starting from the lower ground level, is that the additional peak is almost negligible in the first case and quite prominent in the second case. This difference arises because of the difference in number of magnetic sublevels for the closed transition in each set.
We have adapted this technique to make measurements of hyperfine intervals by using one laser along with an AOM to produce the scanning pump beam. We measure intervals in the $D_2$ lines of Rb and Cs with 20 kHz precision. By measuring the entire spectrum and looking at the symmetry of the line shape, we avoid several potential sources of systematic error. The measurements are consistent with earlier results from our laboratory obtained by locking the AOM to the frequency difference, and show that our earlier error budget was reasonable.
This work was supported by the Department of Science and Technology, India. V.N. acknowledges support from the Homi Bhabha Fellowship Council and A.K.S. from the Council of Scientific and Industrial Research, India.
[10]{}
C. E. Wieman and L. Hollberg, “Using diode lasers for atomic physics,” Rev. Sci. Instrum. **62**, 1–20 (1991).
K. B. MacAdam, A. Steinbach, and C. Wieman, “A narrow-band tunable diode laser system with grating feedback, and a saturated absorption spectrometer for [C]{}s and [R]{}b,” Am. J. Phys. **60**, 1098–1111 (1992).
A. Banerjee, U. D. Rapol, A. Wasan, and V. Natarajan, “High-accuracy wavemeter based on a stabilized diode laser,” Appl. Phys. Lett. **79**, 2139–2141 (2001).
M. S. Safronova, W. R. Johnson, and A. Derevianko, “Relativistic many-body calculations of energy levels, hyperfine constants, electric-dipole matrix elements, and static polarizabilities for alkali-metal atoms,” Phys. Rev. A **60**, 4476–4487 (1999).
E. Arimondo, M. Inguscio, and P. Violino, “Experimental determinations of the hyperfine structures in the alkali atoms,” Rev. Mod. Phys. **49**, 31–75 (1977).
W. Demtröder, *Laser Spectroscopy* (Springer-Verlag, Berlin, 2003), 3rd ed.
A. Banerjee and V. Natarajan, “Saturated-absorption spectroscopy: eliminating crossover resonances by use of copropagating beams,” Opt. Lett. **28**, 1912–14 (2003).
S. E. Harris, “Electromagnetically induced transparency,” Phys. Today **50**, 36–39 (1997).
D. Das and V. Natarajan, “Hyperfine spectroscopy on the [$6P_{3/2}$ state of $^{133}$Cs]{} using coherent control,” Europhys. Lett. **72**, 740–746 (2005).
D. Das and V. Natarajan, “High-precision measurement of hyperfine structure in the [D]{} lines of alkali atoms,” J. Phys. B **41**, 035001 (12pp) (2008).
C. Cohen-Tannoudji and S. Reynaud, “Modification of resonance [R]{}aman scattering in very intense laser fields,” J. Phys. B. **10**, 365–383 (1977).
P. R. S. Carvalho, L. E. E. de Araujo, and J. W. R. Tabosa, “Angular dependence of an electromagnetically induced transparency resonance in a [D]{}oppler-broadened atomic vapor,” Phys. Rev. A **70**, 063818 (2004).
J. Ye, S. Swartz, P. Jungner, and J. L. Hall, “Hyperfine structure and absolute frequency of the [$^{87}$Rb 5P$_{3/2}$]{} state,” Opt. Lett. **21**, 1280–1282 (1996).
G. P. Barwood, P. Gill, and W. R. C. Rowley, “Frequency measurements on optically narrowed [Rb-stabilised]{} laser diodes at 780 nm and 795 nm,” Appl. Phys. B **53**, 142–147 (1991).
U. D. Rapol, A. Krishna, and V. Natarajan, “Precise measurement of the hyperfine structure in the 5[P]{}$_{3/2}$ state of $^{85}$[R]{}b,” Eur. Phys. J. D **23**, 185–188 (2003).
V. Gerginov, A. Derevianko, and C. E. Tanner, “Observation of the nuclear magnetic octupole moment of $^{133}$[C]{}s,” Phys. Rev. Lett. **91**, 072501 (2003).
C. E. Tanner and C. Wieman, “Precision measurement of the hyperfine structure of the $^{133}$[Cs]{} $6{P}_{3/2}$ state,” Phys. Rev. A **38**, 1616–1617 (1988).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We achieved ohmic contacts down to 5 K on standard n-doped Ge samples by creating a strongly doped thin Ge layer between the metallic contacts and the Ge substrate. Thanks to the laser doping technique used, Gas Immersion Laser Doping, we could attain extremely large doping levels above the solubility limit, and thus reduce the metal/doped Ge contact resistance. We tested independently the influence of the doping concentration and doped layer thickness, and showed that the ohmic contact improves when increasing the doping level and is not affected when changing the doped thickness. Furthermore, we characterised the doped Ge/Ge contact, showing that at high doping its contact resistance is the dominant contribution to the total contact resistance.'
author:
- 'F. Chiodi'
- 'A. D. Chepelianskii'
- 'C. Gardes'
- 'G. Hallais'
- 'D. Bouchier'
- 'D. Débarre'
title: 'Laser doping for ohmic contacts in n-type Ge'
---
*Introduction –* Cryogenic detectors are currently used in space imaging to achieve better energy resolution and lower noise levels. However, to further improve the quality of the detection, a close by cryogenic electronics is needed. Silicon JFET, despite their very good performances down to 40 K, are not suitable at lower temperatures due to the carriers freeze-out, so that new materials need to be explored. Germanium JFET are an interesting possibility, having high impedance, low leakage currents and low 1/f noise [@Todi; @Das]. However, their development is subordinate to the resolution of some technical issues, among which the difficulty in creating ohmic contacts over n-type Ge regions. Indeed, two main problems arise when contacting n-Ge with a metal: on one side, a strong Fermi-level pinning results in a high effective electron Schottky barrier height, independent on the chosen metal; on the other, low solubility and fast dopant diffusion make it difficult to create an intermediate thin layer of strongly doped semiconductor, as is often done to achieve ohmic contacts. Several approaches have been validated to realise ohmic, low resistance contacts on n-Ge, such as the suppression of the Fermi level pinning by the introduction of ultrathin SiN [@Kobayashi] and Ge$_3$N$_4$ [@Lieten] layers or by sulfur passivation [@Thathachary]. Alternatively, Sb $\delta$-doping [@Sawano], dopant segregation during Ni germanidation [@Firrincieli2] and laser annealing [@Wang; @Firrincieli] were successfully used to obtain high doping concentrations and ohmic contacts. Because of the higher dopant activation, RTA (rapid thermal annealing) [@Chui; @Thareja] and more recently laser annealing [@Wang] produced the higher concentrations and the lower contact resistances. Indeed, as the laser pulse duration is extremely short ($\sim 20\,$ns), the dopants don’t have the time to diffuse to their equilibrium concentration, and it is thus possible to attain a meta-stable state where the active concentration is higher than the solubility limit. Between the laser annealing techniques, Gas Immersion Laser Doping (GILD) has a particular interest, as it doesn’t require an implantation step prior the laser annealing. The samples are thus free of all deep defects introduced by the ion implantation, which cannot be completely erased even by fast laser annealing. Moreover, GILD gives the possibility to dope a specific area of a device and, without breaking the vacuum, to add a thin ultra-doped layer on top to insure the ohmic contact, eliminating the need for further processes.\
In this paper, we demonstrate ohmic contacts on standard n-doped Ge samples of concentration $10^{15}\,$ cm$^{-3}$, by creating between the metallic contact (M) and the Ge substrate (nGe) a strongly Phosphorus doped thin Ge layer (n$^{++}$Ge) of resistivity $\rho_{++}$ down to $5\times 10^{-4} \,\Omega \, cm$. We tested the influence of the doped layer thickness $t_{++}$ and doping concentration $n_{++}$ on the M/n$^{++}$Ge contact resistance $\rho_{c,M}$. We then characterised the nGe/n$^{++}$Ge contact, showing that its contact resistance $\rho_{c,n}$ is independent on doping thickness and concentration. The distinction between the different contributions of the contact resistance has been often overlooked in the literature, but shouldn’t be neglected: indeed, we showed that $\rho_{c,n}$ dominates the total contact resistance as soon as $\rho_{c,M}$ is better than $2 \times 10^{-3} \, \Omega \,$cm$^2$.\
In view of the utilisation of Ge-based devices such as Ge JFET at low temperature, we tested the I-V curves down to 5 K, and confirmed that the I-V characteristics remain ohmic even at low temperatures. The total contact resistance $\rho_{c,t}$ was also extracted, showing that the contact is better than at room temperature down to 13 K, with a minimum resistance at 40 K where $\rho_{c,t} (T)/\rho_{c,t}(300K)=13 \,\%$.
*Laser doping and sample fabrication–* The quality of the realised interfaces rests on the many advantages of the adopted doping technique, Gas Immersion Laser Doping (GILD). The doping takes place in a ultra high vacuum chamber, at a base pressure of $10^{-9}\,$mbar which insures a very low impurity level. A puff of the precursor gas, PCl$_3$, is injected using a pulse valve onto the Ge sample surface, where it saturates the chemisorption sites. The gas is continuously pumped and, after a small delay, the substrate is melted by a pulsed excimer XeCl laser, of pulse duration 25 ns. The phosphorous can thus diffuse into the liquid germanium phase and is incorporated in the lattice as the liquid/solid interface moves back to the surface at the end of the irradiation. A Ge:P crystal is thus created by liquid phase epitaxy above the underlying Ge substrate (Fig. \[VI\]-a).\
Because of the short pulse duration and the high recrystallization speed, high phosphorous concentrations, larger than the solubility limit ($\sim$ 1$\times 10^{19}$cm$^{-3}$) [@Harame], can be obtained by multiple process repetitions [@Bhaduri; @Cammilleri]. As the dose of active dopants is determined exclusively by the number of laser shots, the doping depth can be independently tuned by controlling the laser energy. Above the Ge melting threshold ($\sim 250\,$mJ/cm$^2$), we can thus linearly increase the melted depth up to a few hundreds of nanometers. To monitor the doped depth, we measure the in-situ time resolved transient reflectivity at 675 nm for each laser shot. Since the reflectivity changes in liquid Ge, we can directly measure the melting duration, which is proportional to the doping depth [@Kerrien2003]. To improve the uniformity of the $2\,$mm$\times2\,$mm laser spots, the spatial inhomogeneity of the laser energy density was reduced to less than $1\%$ by a careful optical treatment of the laser beam (using, in particular, a fly eye homogenizer). This results in a constant P concentration in the doped volume and a sharp 2D Ge:P/Ge interface, as demonstrated by SIMS measurements [@Cammilleri].
All the laser-doped samples were fabricated on a single n-type 101 Ge wafer of doping concentration $10^{15}\,$ cm$^{-3}$. Two different thicknesses were studied, $t_{++}$=85 and 175 nm (laser energy density of 450 and 640$\,$mJ/cm$^2$ respectively), and for each a varying number of laser shots (1-5-10-50) determined the doping level. This resulted in the concentration range $n_{++}=5\,\times 10^{18}- 1\,\times 10^{20}\,$cm$^{-3}$. Three identical set of samples were realised for each condition. After a surface desoxidization, Ti (15 nm)/Al (200 nm) rectangular contacts were deposited in a Transmission Line Method (TLM) configuration by e-beam evaporation. Ohmic M/nGe contacts were also demonstrated for Ni/Au and Ti/Au contacts, stressing the robustness of the obtained results.
*Results –* Resistance measurements were performed on the M/n$^{++}$Ge/nGe samples with a set of 9 rectangular TLM. Ohmic $I-V$ characteristics were consistently obtained (Fig. \[VI\]-b), proof of the good quality of the contact realised. The resistance, dominated by the doped layer resistance, decreased when increasing the doping level $n_{++}$ or the doped thickness $t_{++}$. Only in the particular case of one laser shot, rectifying I-V characteristics were found for both doping thicknesses (Fig. \[VI\]-b, inset). This could be attributed to the imperfect dopant diffusion in the melted layer induced by a single laser shot, as was already shown in Si:B laser doped systems [@Bhaduri], confirming the necessity of an efficient doping process.
![(color online) a) GILD process: chemisorption of the precurson gas over the sample surface; melting of the substrate by the laser pulse and P introduction in the liquid Ge; fast cooling and epitaxy of a Ge:P layer on top of the Ge substrate. b) Representative $I(V)$ curves for as-made samples with different doping concentrations (blue and green: 10 and 50 laser shots, $t_{++}=$85 nm; red and black: 5 and 10 laser shots, $t_{++}=$175 nm) Inset: Non linear $|I|(V)$ curves for $t_{++}=$175 nm and 1 laser shot, after the $n_{++}$ layer etch between the contacts; units are the same as in the main frame. c) Ohmic $I(V)$ characteristics at 300K (orange), 200 K (red), 100 K (green), 40 K (blue), 10 K (pink) for an etched sample with $t_{++}=$175 nm and 5 laser shots ($n_{++}=2.7 \times 10^{19}$cm$^{-3}$). The voltage of the curve at 10 K has been rescaled by a factor of 10. []{data-label="VI"}](Fig1.jpg){width="0.8\columnwidth"}
![(color online) a) $R(L)$ curves for $n_{++}=$1.1$\times 10^{20}$cm$^{-3}$ and $t_{++}=$85 nm as realised (red triangles), after the trench etch around the contacts (blue dots) and after the deep etch through the doped layer (black square). b) Simulations and dimensionless $w R(L)/\rho_{sub}$ data for pristine samples after the trench etch and c) after the deep etch through the doped layer. Symbols from top to bottom on pannel b): black: $n_{++}=5.6\times 10^{19}\,$cm$^{-3}$, $t_{++}$=85 nm; red: $n_{++}=1.1\times 10^{20}\,$cm$^{-3}$, $t_{++}$=85 nm; green: $n_{++}=2.7\times 10^{19}\,$cm$^{-3}$, $t_{++}$=175 nm; blue: $n_{++}=5.4\times 10^{19}\,$cm$^{-3}$, $t_{++}$=175 nm; light blue: $n_{++}=2.7\times 10^{20}\,$cm$^{-3}$, $t_{++}$=175 nm. Lines represent our best-fitting numerical simulations. The same notations apply to c). []{data-label="RL"}](Fig2.jpg){width="\columnwidth"}
The contact resistance consists of two independent contributions given by the M/n$^{++}$Ge and n$^{++}$Ge/nGe contact resistances $\rho_{c,M}$ and $\rho_{c,n}$. To gain access to $\rho_{c,n}$, we removed the doped Ge layer between the contacts by anisotropic reactive ion etching. The Al contacts were unaffected by the CHF$_3$ based etch and acted as an etching mask, preserving the doped layer intact under the contacts. In the exposed areas between the contacts the etch was deep enough to completely remove the n$^{++}$Ge, penetrating up to $1.6\,\mu$m into the nGe substrate. The resulting samples also had linear $I-V$ characteristics, proving that ohmic contacts are obtained both at the M/n$^{++}$Ge and at the n$^{++}$Ge/nGe interfaces. We measured the temperature dependence of the $I-V$ characteristics of the etched samples (Fig. \[VI\]-c) and found that, independently of the doping concentration, the contacts remain ohmic down to 5 K. To separate the contribution of the metal/doped Ge interface from the doped Ge/Ge interface, we compared the resistance $R$ of pristine and etched samples for different separations $L$ between the TLM contacts. The room temperature dependence $R(L)$ measured for the two types of samples is shown on Fig. \[RL\]-a. We attribute the non-linearity of the $R(L)$ dependence to the finite ratio between the substrate thickness $t_{sub}$=$340\, \mu$m and the TLM distance $L=5-300 \,\mu$m. Indeed the $R(L)$ dependence of the etched samples shows a more pronounced non-linearity, which can be easily understood by considering that in the unetched bilayers most of the current is transmitted through the thin doped layer, satisfying in part the thin film approximation leading to a linear $R(L)$. To confirm that current paths at the end of the contacts do not contribute significantly to the $R(L)$ dependence, we prepared a deep trench surrounding the TLM contacts. As shown on Fig. \[RL\]-a this procedure left the $R(L)$ dependence unchanged except at the highest $L$ values. This observation is consistent with the small ratio between $L$ and the width of the contacts $w = 800\;{\rm \mu m}$ in our samples. In our future discussion, we will indicate as pristine samples the samples surrounded by a trench, with the doped film intact between the contacts. We performed a systematic investigation of the $R(L)$ dependence for pristine and etched samples prepared with different doping concentrations of the n$^{++}$Ge layer (see Figs. \[RL\]-b,c). For the etched samples the doping level can only change the offset at $L=0$ in the $R(L)$ dependence. Our measurements demonstrate that the offset is actually almost independent of the doping concentration. Also, the offset increases significantly after etching for all the samples. These two observations suggest that the n$^{++}$Ge/nGe contact resistance is the dominant contribution to the total contact resistance and that it is almost independent of the doping in the n$^{++}$Ge layer. To further confirm this qualitative analysis and to extract quantitative parameters from our measurements we simulated the shape of the non-linear $R(L)$ curves by solving numerically the Laplace equations on the electrostatic potential inside the bilayer using a finite elements method. The simulations were performed in a two dimensional geometry where the effects of the finite contact width ($w = 800\;{\rm \mu}$m) were neglected. This model allowed us to take into account the geometrical effects due to the finite thickness of the Ge sample, and to describe the current exchange between the n$^{++}$Ge and nGe layers in presence of a finite contact resistance between the layers.
As demonstrated by the quality of the fits in Fig. \[RL\]-b and \[RL\]-c we managed to obtain a very good quantitative description of the measured $R(L)$ dependences. The $R(L)$ curves of the etched samples were fitted with a single adjustable parameter corresponding to a substrate resistivity $\rho_{sub} = 0.95 \,\Omega cm$, in good agreement with the manufacturer value $\rho_{sub} = 1-1.3 \,\Omega cm$. The total contact resistance $\rho_{c,t}$ could then be deduced from the resistance offset at $L=0$ in the $R(L)$ dependences. The analysis of the $R(L)$ curves for the pristine samples involved two fitting parameters, the doped layer square resistance $R_{sq} = \rho_{++}/t_{++}$ and $\rho_{c,n}$. The contact resistance $\rho_{c,M}$ was deduced as previously from the offset at $L=0$. We checked the validity of our fitting procedure by measuring in a Van der Pauw configuration the doped layer square resistance $R_{sq}$ of a second set of identical laser spots with no metallic contacts, finding a relative error with the fitting results within $\sim$10$\%$. Moreover, the sum of $\rho_{c,M}$ and $\rho_{c,n}$ extracted from the resistance of the pristine samples reproduced the values of the total contact resistance $\rho_{c,t}$ measured on the etched samples to within $5\%$. We are thus confident that the combination of our experimental results and numerical simulations allows us to measure the contact resistances at both the n$^{++}$Ge/nGe and M/n$^{++}$Ge interfaces.
![(color online) Contact resistances $\rho_{c,n}$ and $\rho_{c,M}$ as a function of the doped layer resistivity $\rho_{++}$. Top panel: $\rho_{c,n}$, broadly independent of the resistivity. Bottom panel: $\rho_{c,M}$, decreasing linearly with the decrease of the doped layer resistivity (increasing concentration) down to a minimum of $5 \times 10^{-6} \, \Omega cm^2$. Two different series of samples are shown: the blue symbols (squares for $t_{++}$=85 nm and dots for $t_{++}$=175 nm) and the black symbols (triangle for $t_{++}$=85 nm and star for $t_{++}$=175 nm); originally identical, the latter had a more through desoxidation. Lines are guides to the eye. Inset: total contact resistance $\rho_{c,t}$ normalised to its value at 300 K, as a function of temperature. []{data-label="Rc"}](Fig3.jpg "fig:"){width="0.9\columnwidth"}\[t\]
In Fig. \[Rc\] we plot the extracted contact resistances $\rho_{c,M}$ and $\rho_{c,n}$ against the resistivity of the laser doped layer $\rho_{++}$. Two sets of samples are presented; a more thorough deoxydation of one set allowed us to reduce $\rho_{c,M}$ down to around $5\,\times 10^{-6} \,\Omega \,$cm$^2$, a value above but comparable with state of the art results ($\sim 10^{-6} - 5\times 10^{-7}\, \Omega \,$cm$^2$) [@Firrincieli2; @Gallacher]. We observe that $\rho_{c,M}$ decreases linearly with $\rho_{++}$ producing a better ohmic M/n++ interface at higher doping concentration (corresponding to lower $\rho_{++}$ values). On the other hand, the contact resistance at the n$^{++}$Ge/nGe interface is found to be independent on the doping level in the n++ layer. Thus at high dopings, $\rho_{c,n}$ gives the dominant contribution to the total contact resistance. This highlights the importance of the n$^{++}$Ge/nGe interface whose contribution has so far been neglected, as to our knowledge previous experiments measured only the contact resistance at the M/n$^{++}$Ge interface. In view of the cryogenic electronics applications of our method, we measured the temperature dependence of the resistance between two different TLMs ($L$=200 $\mu$m and $L$=300 $\mu$m). The characteristic extracted total contact resistance $\rho_{c,t}$ decreases monotonically with the temperature down to 40 K, where its value is only 13 $\%$ of its room temperature value (Fig. \[Rc\], inset). A sharp increase is measured at lower temperatures due to the diminished carrier density activation; however, at 16 K, the contact is still only half of its value at room temperature. A behaviour similar to the one we observe above 50 K was reported for the contact resistance decrease of strongly doped Si:P/metal layers [@Swirhun].
*Conclusions –* In conclusion, we have shown that ohmic contacts down to 5 K can be consistently realised on a n-doped Ge substrate by strong laser doping of a thin surface layer. We have measured the contact resistance between the Ge substrate and the metallic electrodes, separating the contributions of the M/n$^{++}$Ge and of the n$^{++}$Ge/nGe interfaces. While $\rho_{c,n}$ was shown to be independent on the thickness and concentration of the laser doped layer, we could strongly suppress $\rho_{c,M}$ by increasing the doping level. The GILD laser doping technique was fundamental in creating the high doping of the thin epitaxied layers ensuring a good contact resistance. Since GILD also allows to realise the doping in the active-device regions this technique may be crucial for cryogenic temperatures Ge-based electronics.
We are grateful to J.-L. Perrossier for technical help and discussions. This work was supported in part by the Réseau RENATECH and European FP7 grant No. 263455.
[99]{}
R. M. Todi, S. Sonde, E. Simoen, and C. Claeysd; K. B. Sundaram, Appl. Phys. Lett. [**90**]{}, 043501 (2007)
N.C. Das, C. Monroy, and M. Jhabvala, Solid State Electronics [**44**]{}, 937 (2000)
M. Kobayashi, A. Kinoshita, K. Saraswat, H.-S. P. Wong, and Y. Nishi, J. Appl. Phys. [**105**]{}, 023702 (2009)
R. R. Lieten et al., Journal of The Electrochemical Society [**158**]{}, H358 (2011)
A. V. Thathachary, K. N. Bhat, N. Bhat, and M. S. Hegde, Appl. Phys. Lett. [**96**]{}, 152108 (2010)
K. Sawano, Y. Hoshi, K. Kasahara, K. Yamane, K. Hamaya, M. Miyao, and Y. Shiraki Appl. Phys. Lett. [**97**]{}, 162108 (2010)
A. Firrincieli, et al., Appl. Phys. Lett. [**99**]{}, 242104 (2011)
A. Firrincieli, K. Martens, E. Simoen, C. Claeys, and J.A. Kittl, Microelectronic Engineering [**106**]{}, 129 (2013)
Chen Wang, et al., Applied Physics Express [**6**]{}, 106501 (2013)
C. O. Chui, L. Kulig, J. Moran, W. Tsai, and K. C. Saraswat, Appl. Phys. Lett. [**87**]{}, 091909 (2005)
G. Thareja, S.-L. Cheng, T. Kamins, K. Saraswat, and Y. Nishi, IEEE Electron Device Lett. [**32**]{}, 608 (2011)
D. Harame, et al., SiGe, Ge, and Related Compounds 4: Materials, Processing, and Devices [**6**]{} (2010)
A. Bhaduri et al., [*[Appl. Surf. Sc.]{}*]{} **258**, 9228 (2012)
D. Cammilleri et al., Thin Solid Films [**517**]{}, 75 (2008)
G. Kerrien et al., [*[Appl. Surf. Sci.]{}*]{} **208**, 277 (2003)
K. Gallacher, P. Velha, D. J. Paul, I. MacLaren, M. Myronov, and D. R. Leadley, Appl. Phys. Lett. [**100**]{}, 022113 (2012)
S. E. Swirhun and R. M. Swanson, IEEE Electron Device Lett. [**EDL-7**]{}, 3 (1986)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Consider a randomly oriented graph $G=(V,E)$ and let $a$, $s$ and $b$ be three distinct vertices in $V$. We study the correlation between the events $\{a\to s\}$ and $\{s\to b\}$. We show that, when $G$ is the complete graph $K_n$, the correlation is negative for $n=3$, zero for $n=4$, and that, counter-intuitively, it is positive for $n\ge 5$. We also show that the correlation is always negative when $G$ is a cycle, $C_n$, and negative or zero when $G$ is a tree (or a forest).'
address:
- 'Department of Mathematics, Uppsala University, P.O. Box 480, SE-751 06, Uppsala, Sweden.'
- 'Department of Mathematics, KTH-Royal Institute of Technology, SE-100 44, Stockholm, Sweden.'
author:
- Sven Erick Alm
- Svante Linusson
title: 'A counter-intuitive correlation in a random tournament'
---
[^1]
[^2]
Introduction {#S:Intro}
============
Given a graph $G=(V,E)$ we orient each edge with equal probability for the two possible directions and independent of all other edges. This model has been studied previously in for instance [@G00; @SL2; @CM]. Let $a$, $s$ and $b$ be three distinct vertices in $V$. The object of this paper is to study the correlation of the two events $\{a\to s\}$, that there exists a directed path from $a$ to $s$, and $\{s\to b\}$. One might intuitively guess that they are always negatively correlated, i.e. that $P(a\to s, s\to b)< P(a\to s)\cdot P(s\to b)$. This is however not true for all graphs. In fact, the smallest counterexample is the graph on four vertices with all edges except $\{a,b\}$ present, see Section \[S:CE\].
In section \[S:Kn\] we prove that for the complete graph, $K_n$, the events are negatively correlated for $n=3$, independent for $n=4$ and positively correlated for $n\ge 5$. The complementary events, $A:=\{a \nto s\}$, that there does not exist a directed path from $a$ to $s$, and $B:=\{s\nto b\}$ have the same covariance, and we show that their relative covariance, $(P(A\cap B)-P(A)\cdot P(B))/P(A\cap B)$ converges to $1/3$ as $n\to\infty$.
In Section \[S:Rec\] we give exact recursions for the probabilities $P(A)$ and $P(A\cap B)$, and compute these for $n\le15$. For completeness, in Section \[S:CT\] we show that the events are negatively correlated when $G$ is a cycle and negatively correlated or independent when $G$ is a tree (or a forest). We end with stating a number of conjectures and open problems.
In a coming paper, [@AL2], we will study this problem when $G$ is the random graph $G(n,p)$.
The question studied here was posed in [@SL2]. There it was proved that under this model for any vertices $a,b,s,t\in V$ the events $\{s\to a\}$ and $\{s\to b\}$ are never negatively correlated. This was shown to be true also if we first conditioned on $\{s\nto t\}$, i.e. As a sort of converse it was also proved that $P(s\to a, b\to t | s\nto t )\le P(s\to a | s\nto t)\cdot P(b\to t | s\nto t)$. The proofs in [@SL2] relied heavily on the results in [@vdBK] and [@vdBHK], where similar statements were proved for edge percolation on a given graph and a result from [@CM] that relates the random orientation with edge percolation. This cluster of questions on correlation of paths have been inspired by an interesting conjecture due to Kasteleyn, named the Bunkbed conjecture by Häggström [@OH2], see also [@SL1] and Remark 5 in [@vdBK].
[**Acknowledgment:**]{} We thank Svante Janson, Stanislav Volkov and Johan Wästlund for fruitful discussions.
A counter-intuitive example {#S:CE}
===========================
Let $G$ be a graph with four vertices and all edges except the one between $a$ and $b$ present, see Figure \[F:pos\], and let $C:=\{a\to s\}$ and $D:=\{s\to b\}$.
![A counterexample with positive correlation[]{data-label="F:pos"}](k4-1){height="3cm"}
Then, $P(C)=P(D)=\frac12+\frac18+\frac1{32}=\frac{21}{32}$ and $P(C\cap D)=\frac14+\frac1{16}+\frac1{16}+\frac1{16}=\frac7{16}$, so that $$P(C\cap D)-P(C)\cdot P(D)=\frac7{16}-\left(\frac{21}{32}\right)^2=\frac7{1024}>0.$$ Note that if we relabel the vertices in this graph as in Figure \[F:neg\], we still get $P(C)=P(D)=\frac{21}{32}$, but $P(C\cap D)=\frac14+\frac18+\frac1{32}=\frac{13}{32}$, so that $$P(C\cap D)-P(C)\cdot P(D)=\frac{13}{32}-\left(\frac{21}{32}\right)^2=-\frac{25}{1024}<0.$$ In fact, the labeling in Figure \[F:pos\] is the only graph with four vertices that gives a positive correlation.
![A slightly modified example with negative correlation[]{data-label="F:neg"}](k4-2){height="30mm"}
The complete graph $K_n$ {#S:Kn}
========================
Let $G$ be the complete graph, $K_n$, and orient the edges either way independently with probability $\frac{1}{2}$. For three different vertices, $a$, $s$ and $b$, of $K_n$ we want to know if the event $\{a \to s\}$ and the event $\{s\to b\}$ are positively or negatively correlated. It turns out to be easier to study the correlation of the complementary events, i.e. $A:=\{a \nto s\}$, that there does not exist a directed path from $a$ to $s$, and the event $B:=\{s\nto b\}$, which have the same correlation.
Think of the vertices of $K_n$ as $[n]:=\{1,\dots,n\}$. To estimate $P(A)$, the following lemma will be useful.
\[L:a\] For all $n\ge0$, $$a(n):=\sum_{k=1}^{n-1} \binom{n}{k}\sum_{m=1}^{n-k} \binom{n-k}{m}\Big(\frac12\Big)^{km}\le5.6\cdot\Big(\frac74\Big)^n.
$$
As $a(0)=a(1)=0$, we will assume that $n\ge2$. We will use that $(\frac12)^{km}=(\frac14)^m\cdot(\frac12)^{(k-2)m}\le(\frac14)^m\cdot(\frac12)^{k-2}=4\cdot(\frac12)^k\cdot(\frac14)^m$, if $m\ge1$ and $k\ge2$, and split the sum into two parts, $k=1$ and $k\ge 2$. $$\begin{aligned}
a(n)&=n\cdot\sum_{m=1}^{n-1} \binom{n-1}{m}\Big(\frac12\Big)^m +\sum_{k=2}^{n-1} \binom{n}{k}\sum_{m=1}^{n-k} \binom{n-k}{m}\Big(\frac12\Big)^{km}\\
&\le n\cdot\Big(\frac32\Big)^{n-1}
+4\cdot\sum_{k=2}^{n-1} \binom{n}{k}\Big(\frac12\Big)^k\sum_{m=1}^{n-k} \binom{n-k}{m}\Big(\frac14\Big)^{m}\\
&\le n\cdot\Big(\frac32\Big)^{n-1}+4\cdot\sum_{k=2}^{n-1} \binom{n}{k}\Big(\frac12\Big)^k\Big(\frac54\Big)^{n-k}\\
&\le n\cdot\Big(\frac32\Big)^{n-1}+4\cdot\Big(\frac74\Big)^n.\end{aligned}$$ The lemma follows by showing that $n\cdot(\frac32)^{n-1}\le1.6\cdot(\frac74)^n$ holds for all $n\ge2$.
As $k=1$ contributes $n\cdot(\frac32)^{n-1}$, this gives a lower bound for $a(n)$. This is in fact the dominating term, so that $a(n)$ asymptotically is of order $p(n)\cdot(\frac32)^n$, where $p(n)$ is a polynomial in $n$. By splitting the sum into three parts, we can, for example, show that $a(n)\le13.6\cdot(\frac{13}8)^n$ for all $n\ge0$.
\[L:A\] For all $n\ge2$, $$\Big(\frac12\Big)^{n-2}\left(1-\Big(\frac12\Big)^{n-1}\right)\le P(A)\le\Big(\frac12\Big)^{n-2}\left(1+3.2\cdot\Big(\frac78\Big)^{n-1}\right).
$$
A necessary condition for $A$ is that the edge between $a$ and $s$ is directed from $s$ to $a$. Let $E$, with $P(E)=1/2$, denote this event.
Let $O_a$ and $O_s$ denote the sets of points in $[n]\setminus\{a,s\}$ that can be reached from $a$ and $s$ respectively in one step. Similarly, let $I_a$ and $I_s$ denote the sets of points in $[n]\setminus\{a,s\}$ that can reach $a$ and $s$ respectively in one step.
For the lower bound, note that $E\cap(O_a=\emptyset)\Rightarrow A$ and $E\cap(I_s=\emptyset)\Rightarrow A$, so that $$\begin{aligned}
P(A)&\ge P((O_a=\emptyset)\cup(I_s=\emptyset))/2=\left(\Big(\frac12\Big)^{n-2}+\Big(\frac12\Big)^{n-2}-\Big(\frac12\Big)^{2n-4}\right)/2\\
&=\Big(\frac12\Big)^{n-2}\left(1-\Big(\frac12\Big)^{n-1}\right).\end{aligned}$$ For the upper bound, note that $A\Rightarrow E\cap(O_a\subset O_s)\cap F$, where $F$ is the event that the points in $O_a$ have no directed edges to points in $I_s$. Note that, if $k=|O_a|$ and $m=|O_s|$, then $k\le m$ is necessary, and there are $k(n-2-m)$ edges in $F$.
$$\begin{aligned}
P(A)&\le P((O_a\subset O_s)\cap F)/2\\
&=\frac12\cdot\sum_{k=0}^{n-2}\binom{n-2}k\Big(\frac12\Big)^{n-2}\sum_{m=k}^{n-2}\binom{n-2-k}{m-k}\Big(\frac12\Big)^{n-2}\cdot\Big(\frac12\Big)^{k(n-2-m)}\\
&=\Big(\frac12\Big)^{2n-3}\sum_{k=0}^{n-2}\binom{n-2}k\sum_{m=0}^{n-2-k}\binom{n-2-k}m\Big(\frac12\Big)^{k\cdot m}\\
&=\Big(\frac12\Big)^{2n-3}\sum_{m=0}^{n-2}\binom{n-2}m+\Big(\frac12\Big)^{2n-3}\sum_{k=1}^{n-2}\binom{n-2}k\\
&\phantom{aa}+\Big(\frac12\Big)^{2n-3}\sum_{k=1}^{n-2}\binom{n-2}k\sum_{m=1}^{n-2-k}\binom{n-2-k}m\Big(\frac12\Big)^{k\cdot m}\\
&=\Big(\frac12\Big)^{2n-3}\cdot 2^{n-2}+\Big(\frac12\Big)^{2n-3}\cdot(2^{n-2}-1)\\
&\phantom{aa}+\Big(\frac12\Big)^{2n-3}\sum_{k=1}^{n-3}\binom{n-2}k\sum_{m=1}^{n-2-k}\binom{n-2-k}m\Big(\frac12\Big)^{k\cdot m}\\
&\le\Big(\frac12\Big)^{n-2}+\Big(\frac12\Big)^{2n-3}\cdot a(n-2)
\le\Big(\frac12\Big)^{n-2}+5.6\cdot\Big(\frac12\Big)^{2n-3}\Big(\frac74\Big)^{n-2}\\
&\le\Big(\frac12\Big)^{n-2}\left(1+5.6\cdot\frac47\cdot\Big(\frac78\Big)^{n-1}\right)
=\Big(\frac12\Big)^{n-2}\left(1+3.2\cdot\Big(\frac78\Big)^{n-1}\right),\end{aligned}$$
where the function $a$ was defined in Lemma \[L:a\].
\[T:A\] $$\lim_{n\to\infty} 2^{n-2}\cdot P(A)=1.$$
Follows immediately from Theorem \[L:A\].
To estimate $P(A\cap B)$, the following lemma, in combination with Lemma \[L:a\], is useful.
\[L:b\] For all $n\ge0$, $$b(n):=\sum_{k=1}^{n-1}\binom nk\sum_{i=1}^{n-1-k}\binom{n-k}i\Big(\frac12\Big)^{ki}\sum_{m=1}^k\binom km\Big(\frac12\Big)^{m(n-k-i)}
\le4\cdot\Big(\frac74\Big)^n.$$
Note first that $b(0)=b(1)=b(2)=0$, so that we may assume that $n\ge3$. Note also that for $k,i\ge1$, $ki=(k-1)(i-1)+k+i-1\ge k+i-1$, so that $(\frac12)^{ki}\le2(\frac12)^k(\frac12)^i$. This gives $$\begin{aligned}
b(n)&\le4\cdot\sum_{k=1}^{n-1}\binom nk\cdot\Big(\frac12\Big)^{k}\sum_{i=1}^{n-1-k}\binom{n-k}i\cdot\Big(\frac12\Big)^{i}\cdot\Big(\frac12\Big)^{n-k-i}
\sum_{m=1}^k\binom km\cdot\Big(\frac12\Big)^{m}\\
&\le4\cdot\sum_{k=1}^{n-1}\binom nk\cdot\Big(\frac12\Big)^{k}\cdot1\cdot\Big(\frac32\Big)^{k}
=4\cdot\sum_{k=1}^{n-1}\binom nk\cdot\Big(\frac34\Big)^{k}\le4\cdot\Big(\frac74\Big)^{n}.\end{aligned}$$
\[L:AB\]For all $n\ge3$, $$\Big(\frac12\Big)^{2n-3}\left(3-2\Big(\frac12\Big)^{n-3}\right)\le P(A\cap B)
\le\Big(\frac12\Big)^{2n-3}\left(3+20.8\cdot\Big(\frac78\Big)^{n-3}\right).$$
A necessary condition for $A\cap B$ is that the edge between $a$ and $s$ is directed from $s$ to $a$, that the edge between $s$ and $b$ is directed from $b$ to $s$ and that the edge between $a$ and $b$ is directed from $b$ to $a$. Let $E$, with $P(E)=1/8$, denote this event.
Let $O_a$, $O_s$ and $O_b$ denote the sets of points in $[n]\setminus\{a,s,b\}$ that can be reached from $a$, $s$ and $b$ respectively in one step. Similarly, let $I_a$, $I_s$ and $I_b$ denote the sets of points in $[n]\setminus\{a,s,b\}$ that can reach $a$, $s$ and $b$ respectively in one step. For the lower bound, note that $E\cap(O_a=O_s=\emptyset)\Rightarrow A\cap B$, $E\cap(O_a=I_b=\emptyset)\Rightarrow A\cap B$ and $E\cap(I_s=I_b=\emptyset)\Rightarrow A\cap B$, so that $$\begin{aligned}
P(A\cap B)&\ge P((O_a=O_s=\emptyset)\cup(O_a=I_b=\emptyset)\cup(I_s=I_b=\emptyset))/8\\
&=\Big(\frac12\Big)^3\left(3\Big(\frac12\Big)^{2n-6}-2\Big(\frac12\Big)^{3n-9}\right)\\
&=\Big(\frac12\Big)^{2n-3}\left(3-2\Big(\frac12\Big)^{n-3}\right),\end{aligned}$$ as $(O_a=O_s=\emptyset)\cap(I_s=I_b=\emptyset)=\emptyset$.
For the upper bound, we note that $A\cap B\Rightarrow E\cap(O_a\subset O_s\subset O_b)\cap F$, where $F$ denotes the event that the points in $O_a$ have no directed edges to points in $I_s$ and the points in $O_s$ have no directed edges to points in $I_b$, so that $P(A\cap B)\le P((O_a\subset O_s\subset O_b)\cap F)/8$. Let $k=|O_s|$, $i=|I_b|$ and $m=|O_a|$. Then $0\le k\le n-3$, $0\le i\le n-3-k$ and $0\le m\le k$ and the direction of all $3(n-3)+3=3n-6$ edges connected to $a$, $s$ and $b$ are determined. Further, the event $F$ determines the direction of $ki+m(n-3-k-i)$ edges, so that $$\begin{aligned}
P(A\cap B)
&\le\Big(\frac12\Big)^{3n-6}\sum_{k=0}^{n-3}\binom{n-3}k\sum_{i=0}^{n-3-k}\binom{n-3-k}i\Big(\frac12\Big)^{ki}
\sum_{m=0}^k\binom km\Big(\frac12\Big)^{m(n-3-k-i)}\\
&=\Big(\frac12\Big)^{3n-6}\cdot(S_1+S_2+S_3+S_4+S_5+S_6+S_7),\end{aligned}$$ where the triple sum is split into seven parts:\
1: $k=0\Rightarrow m=0$, 2: $k=n-3\Rightarrow i=0$, 3: $i=m=0,1\le k\le n-4$,\
4: $i=0,m\ge1,1\le k\le n-4$, 5: $m=0,i\ge1,1\le k\le n-4$,\
6: $i=n-3-k,m\ge1,1\le k\le n-4$,\
7: $1\le k\le n-4,1\le i\le n-4-k, 1\le m\le k$.\
The first three cases correspond to the three cases of the lower bound, $$\begin{aligned}
S_1&=\sum_{i=0}^{n-3}\binom{n-3}i=2^{n-3},\\
S_2&=\sum_{m=0}^{n-3}\binom{n-3}m=2^{n-3},\\
S_3&=\sum_{k=1}^{n-4}\binom{n-3}k=2^{n-3}-2\le2^{n-3}.\end{aligned}$$ The next three can be expressed by the function $a$ of Lemma \[L:a\], $$\begin{aligned}
S_4&=\sum_{k=1}^{n-4}\binom{n-3}k\sum_{m=1}^k\binom km\cdot\Big(\frac12\Big)^{m(n-3-k)}
=\sum_{j=1}^{n-4}\binom{n-3}j\sum_{m=1}^{n-3-j}\binom{n-3-j}m\cdot\Big(\frac12\Big)^{mj}\\
&=a(n-3)\le5.6\cdot\Big(\frac74\Big)^{n-3},\\
S_5&=\sum_{k=1}^{n-4}\binom{n-3}k\sum_{i=0}^{n-3-k}\binom{n-3-k}i\cdot\Big(\frac12\Big)^{ki}=a(n-3)\le5.6\cdot\Big(\frac74\Big)^{n-3},\\
S_6&=\sum_{k=1}^{n-4}\binom{n-3}k\cdot\Big(\frac12\Big)^{k(n-3-k)}\sum_{m=1}^k\binom km
\le\sum_{k=1}^{n-4}\binom{n-3}k\sum_{m=1}^k\binom km\cdot\Big(\frac12\Big)^{m(n-3-k)}\\
&=\sum_{i=1}^{n-4}\binom{n-3}i\sum_{m=1}^{n-3-i}\binom{n-3-i}m\cdot\Big(\frac12\Big)^{im}
=a(n-3)\le5.6\cdot\Big(\frac74\Big)^{n-3},\end{aligned}$$ and the last by the function $b$ of Lemma \[L:b\], $$\begin{aligned}
S_7&=\sum_{k=1}^{n-4}\binom{n-3}k\sum_{i=1}^{n-4-k}\binom{n-3-k}i\cdot\Big(\frac12\Big)^{ki}\sum_{m=1}^k\binom km\cdot\Big(\frac12\Big)^{m(n-3-k-i)}\\
&=b(n-3)\le4\cdot\Big(\frac74\Big)^{n-3}.\end{aligned}$$ Collecting the estimates gives the lemma.
\[T:AB\] $$\lim_{n\to\infty} 2^{2n-3}\cdot P(A\cap B)=3.$$
Follows immediately from Theorem \[L:AB\].
\[T:relkorr\] $$\lim_{n\to\infty} \frac{P(A\cap B)-P(A)\cdot P(B)}{P(A\cap B)}=\frac13.$$
Follows immediately from Theorems \[T:A\] and \[T:AB\] as $P(B)=P(A)$.
\[R:korr\] Note that $$\frac{P(A\cap B)-P(A)\cdot P(B)}{P(A\cap B)}=1-\frac{P(A)\cdot P(B)}{P(A)\cdot P(B\,|\,A)}=1-\frac{P(B)}{P(B\,|\,A)},$$ so that Theorem \[T:relkorr\] can be formulated as $$\lim_{n\to\infty}\frac{P(B)}{P(B\,|\,A)}=\frac23,
\text{ or equivalently }
\lim_{n\to\infty}\frac{P(B\,|\,A)}{P(B)}=\frac32.$$
Theorem \[T:relkorr\] shows that the events $A=\{a\nto s\}$ and $B=\{s\nto b\}$ are positively correlated for sufficiently large $n$. From this follows that the complementary events $C=\{a\to s\}$ and $D=\{s\to b\}$ also are positively correlated for sufficiently large $n$. It is in fact true for all $n\ge5$ as the next theorem shows.
\[T:korr\] The events $A=\{a\nto s\}$ and $B=\{s\nto b\}$ are negatively correlated for $n=3$, independent for $n=4$ and positively correlated for $n\ge5$.
From Lemmas \[L:A\] and \[L:AB\] we get $$\begin{aligned}
P(A\cap B)-&P(A)\cdot P(B)=P(A\cap B)-(P(A))^2\\
&\ge\Big(\frac12\Big)^{2n-3}\cdot\left(3-2\Big(\frac12\Big)^{n-3}\right)-\left\{\Big(\frac12\Big)^{n-2}\cdot\left(1+3.2\cdot\Big(\frac78\Big)^{n-1}\right)\right\}^2\\
&\ge\Big(\frac12\Big)^{2n-4}\cdot\left(6-4\cdot\Big(\frac12\Big)^{n-3}-1-6.4\cdot\Big(\frac78\Big)^{n-1}-10.24\cdot\Big(\frac{49}{64}\Big)^{n-1}\right)\\
&=\Big(\frac12\Big)^{2n-4}\cdot\left(5-\Big(\frac12\Big)^{n-5}-6.4\cdot\Big(\frac78\Big)^{n-1}-10.24\cdot\Big(\frac{49}{64}\Big)^{n-1}\right)\\
&=\Big(\frac12\Big)^{2n-4}\cdot(5-c(n)),\end{aligned}$$ where $c(n)$ is a decreasing function of $n$, with $c(8)<5$, so that the theorem holds for $n\ge8$. The remaining cases, $3\le n\le7$, are proved using the recursion formulas in Lemmas \[L:f\] and \[L:g\] in the next section.
Exact recursions {#S:Rec}
================
In this section we will derive recursions for $P(A)$ and $P(A\cap B)$. For $n\ge 2$, $s\in [n]$ and $K\subset [n]\setminus \{s\}$ let $\{K\nto s\}$ denote the event $\{a\nto s$ for every $a\in K\}$.
With $|K|=k$ define $$f(n,k):=P_n(K\nto s),$$ where in particular $f(n,0)=1$. Also set $f(1,0)=1$ for convenience.
For $n\ge 3$ and $s,b\in [n], K\subset [n]\setminus \{s,b\}$, $s\neq b$ and $|K|=k$ define: $$g(n,k):=P_n(K\nto s,s\nto b),$$ where in particular $g(n,0)=f(n,1)$. Also let $g(2,0):=f(2,1)=1/2$.
\[L:f\] For $n\ge k+1\ge 2$ we have $$f(n,k)=\sum_{i=0}^{n-k-1} \binom{n-k-1}{i}\frac{(2^{k}-1)^i}{2^{k(n-k)}} f(n-k,i).$$
We first deduce the following recursion: $$P_n(K\nto s)=\sum_{L\subset [n]\setminus (K\cup \{s\})} \frac{1}{2^{k(n-k-|L|)}}\Big(1-\frac{1}{2^k}\Big)^{|L|} P_{n-k}(L\nto s).$$
This can be seen as follows. We think of the set $L$ as the vertices that we can reach from vertices in $K$ in one step. Clearly, we must have $s\notin L$. Every edge from $[n]\setminus (K\cup L)$ to $K$ must be oriented in that direction, which gives the first power of $1/2$. To reach a vertex $c\in L$ in one step it must not be the case that all edges $\{a,c\}, a\in K$ are directed away from $c$, which gives the second factor. The edges within $K$ make no difference and we have considered all edges going between $K$ and $[n]\setminus K$. We are left with a situation where we must make sure that there is no directed path from any vertex in $L$ to $s$ passing over vertices in $[n]\setminus K$, which gives the last term. The recursion in the lemma is easily deduced from this, by summing over the size, $i=|L|$, of $L$.
\[L:g\] For $n\ge k+2\ge 3$ we have $$g(n,k)=\sum_{i=0}^{n-k-2} \binom{n-k-2}{i}\frac{(2^k-1)^i}{2^{k(n-k)}} g(n-k,i).$$
First note that $\{K\nto s\}$ and $\{s\nto b\}$ implies $\{K\nto b\}$. Reasoning as in the previous proof, we obtain $$P_n(K\nto s,s\nto b)=\sum_{L\subset [n]\setminus (K\cup \{s\}\cup \{b\})} \frac{1}{2^{k(n-k-|L|)}}\Big(1-\frac{1}{2^k}\Big)^{|L|}
P_n(L\nto s, s\nto b).$$
Using the lemmas, we can recursively compute the desired probabilities $P(A)=f(n,1)$ and $P(A\cap B)=g(n,1)$. Exact and numerical values of $P(A)$ and $P(A\cap B)$ were computed and are given in Table \[Tab:corr\] together with numerical values of the relative covariance $(P(A\cap B)-P(A)\cdot P(B))/P(A\cap B)$ for $3\le n\le13$.
----------------- ---------------------------------- ------------ ------------------------------------------ ------------------ ------------------------------------------
$n$ [$P(A)\cdot 2^{\binom{n}{2}}$]{} [$P(A)$]{} [$P(A\cap B)\cdot 2^{{\binom{n}{2}}}$]{} [$P(A\cap B)$]{} $\frac{P(A\cap B)-P(A)P(B)}{P(A\cap B)}$
\[-10pt\]
\[-10pt\] [2]{} [1]{} [0.5000]{}
[3]{} [3]{} [0.3750]{} [1]{} [0.1250000]{} [$-$0.125000]{}
[4]{} [16]{} [0.2500]{} [4]{} [0.0625000]{} [0.000000]{}
[5]{} [150]{} [0.1465]{} [26]{} [0.0253906]{} [0.154898]{}
[6]{} [2504]{} [0.0764]{} [272]{} [0.0083008]{} [0.296523]{}
[7]{} [77472]{} [0.0369]{} [4672]{} [0.0022278]{} [0.387428]{}
[8]{} [4677904]{} [0.0174]{} [139696]{} [0.0005204]{} [0.416449]{}
[9]{} [571023120]{} [0.0083]{} [7928624]{} [0.0001154]{} [0.401547]{}
[10]{} [142058571776]{} [0.0040]{} [917140928]{} [0.0000261]{} [0.374613]{}
[11]{} [71626948215168]{} [0.0020]{} [220836999808]{} [0.0000061]{} [0.355191]{}
[12]{} [72752562631695616]{} [0.0010]{} [109473061398784]{} [0.0000015]{} [0.344746]{}
[13]{} [148346259329909191680]{} [0.0005]{} [110228037783934976]{} [0.0000004]{} [0.339426]{}
----------------- ---------------------------------- ------------ ------------------------------------------ ------------------ ------------------------------------------
: Probabilities and relative covariances[]{data-label="Tab:corr"}
Cycles and trees {#S:CT}
================
Let $G=C_n$, the cycle with $n$ vertices, and let $c$, $d$ and $n-c-d$ denote the distances (number of edges) between $a$ and $s$, $s$ and $b$, and $b$ and $a$, respectively. We assume the three vertices to be distinct, so that $c,d,n-c-d\ge1$. Further, let $C:=\{a\to s\}$ and $D:=\{s\to b\}$. Then, $$\begin{aligned}
P(C)&=\left(\frac12\right)^c+\left(\frac12\right)^{n-c}-\left(\frac12\right)^n,\\
P(D)&=\left(\frac12\right)^d+\left(\frac12\right)^{n-d}-\left(\frac12\right)^n,\\
P(C\cap D)&=\left(\frac12\right)^c\cdot\left(\frac12\right)^{d}+\left(\frac12\right)^n,\end{aligned}$$ where the last term corresponds to the case when there is a directed path $s\to a\to b\to s$. This gives $$\begin{aligned}
P(C\cap &D)-P(C)\cdot P(D)\\
&=\left(\frac12\right)^{c+d}+\left(\frac12\right)^n-\left(\Big(\frac12\Big)^c+\Big(\frac12\Big)^{n-c}-\Big(\frac12\Big)^n\right)
\cdot\left(\Big(\frac12\Big)^d+\Big(\frac12\Big)^{n-d}-\Big(\frac12\Big)^n\right)\\
&=\left(\frac12\right)^{2n}\cdot\left(2^n-2^{n-c+d}-2^{n+c-d}-2^{c+d}+2^{n-c}+2^{n-d}+2^c+2^d-1\right)\\
&=\left(\frac12\right)^{2n}\cdot\left(2^{n-c+d}\big(2\cdot2^{c-d}-1-2^{2(c-d)}\big)\right.\\
&\phantom{\left(\frac12\right)^{2n}\cdot\qquad}\left.-2^n\big(1-2^{-c}-2^{-d}\big)-\big(2^{c+d}-2^c-2^d+1\big)\right)\\
&\le-\left(\frac12\right)^{2n}\cdot\left(2^{n-c+d}\cdot\big(2^{c-d}-1)^2+\big(2^c-1\big)\cdot\big(2^d-1\big)\right)\\
&\le-\left(\frac12\right)^{2n},\end{aligned}$$ with equality if and only if $c=d=1$.
We have proved the following theorem.
\[T:cycles\] When $G$ is the cyclic graph with $n$ nodes, $C_n$, the covariance between the events $\{a\to s\}$ and $\{s\to b\}$ is at most $-\left(\frac12\right)^{2n}$, with equality if and only if the vertices $a$ and $b$ are adjacent to $s$.
For trees the situation is even simpler, as there are no cycles so that there is a unique path between any two vertices. Two cases can occur. If the path between $a$ and $b$ passes $s$, the paths between $a$ and $s$ and between $s$ and $b$ have no edges in common, so that the events $\{a\to s\}$ and $\{s\to b\}$ are independent. On the other hand, if the path between $a$ and $b$ does not pass $s$, then the events are disjoint, so that $P(\{a\to s\}\cap\{s\to b\})=0$, and the covariance is strictly negative. For a forest, if not all three vertices are in the same tree, then trivially $P(\{a\to s\}\cap\{s\to b\})=0$ and at least one of $P(a\to s)$ and $P(s\to b)$ is zero, so that the events are independent.
When $G$ is a tree (or a forest), the events $\{a\to s\}$ and $\{s\to b\}$ are either independent or mutually exclusive.
Open problems and conjectures {#S:PC}
=============================
From Theorem \[T:cycles\] and the observations in Section \[S:CE\], we make the following conjecture.
For any connected graph $G=(V,E)$ and three distinct vertices $a$, $s$ and $b$ in $V$; if $s$ has degree at most two, then the events $\{a\to s\}$ and $\{s\to b\}$ are independent or negatively correlated.
Any connected simple graph $G=(V,E), |V|\ge 3$ belongs to (at least) one of the following classes.
- For any three distinct vertices $a,b,s\in V(G)$, the events $\{a\to s\}$ and $\{s\to b\}$ are non-positively correlated.
- There exist three distinct vertices $a,b,s\in V(G)$, such that the events $\{a\to s\}$ and $\{s\to b\}$ are negatively correlated and there exist three distinct vertices $a',b',s'\in V(G)$, such that the events $\{a'\to s'\}$ and $\{s'\to b'\}$ are positively correlated. Or there exist three distinct vertices $a,b,s\in V(G)$, such that the events $\{a\to s\}$ and $\{s\to b\}$ are independent.
- For any three distinct vertices $a,b,s\in V(G)$, the events $\{a\to s\}$ and $\{s\to b\}$ are non-negatively correlated.
We have shown that trees and cycles belong to Class I, $K_n,n\ge 5$ belongs to Class III and $K_4$ minus one edge belongs to Class II. Note that when we have independent events there may be some overlap between the classes, in particular $K_4$ belongs to all three classes.
For large $n$ most graphs will belong to Class II.
In fact we guess that for $n$ large enough, the graphs in Class I are joins (in some vague sense) of cycles and trees. It would be interesting if it was possible to characterize the graphs in Class I. We formulate the following more specific questions. Recall that outerplanar graphs are the graphs that do not have $K_4$ or $K_{2,3}$ as minors.
Are all graphs in Class I (with $|V(G)|\ge 5$) outerplanar?
Similarly one could ask for a characterization of the graphs in Class III. The following subproblem would also be interesting if it could be solved.
For a given $n$, what is the smallest number $k$ such that there exist $k$ edges whose removal from $K_n$ gives a graph not in Class III?
Finally we ask if Class I and Class III are monotone.
Is it true that if $G$ belongs to Class I, but not to Class II, then so does any connected subgraph obtained by removing one edge? Similarly, is it true that if $G$ belongs to Class III, but not to Class II, then so does $G$ plus any new edge?
The results in [@AL2] seem to suggest that Class III is larger than Class I. Is this true?
[ABX]{}
Sven Erick Alm and Svante Linusson, Correlations for paths in random orientations of $G(n,p)$, [*Preprint 2009*]{}[****]{}. arXiv:0906.0720.
Jacob van den Berg and Jeff Kahn, A correlation inequality for connection events in percolation, [*Annals of Probability *]{}[**29**]{} No. 1 (2001), 123–126.
Jacob van den Berg, Olle Häggström and Jeff Kahn, Some conditional correlation inequalities for percolation and related processes, [*Rand. Structures Algorithms*]{} [**29**]{} (2006), 417–435.
Geoffrey R. Grimmett, Infinite Paths in Randomly Oriented Lattices, [*Random Structures and Algorithms*]{} [**18**]{}, Issue 3, (2001), 257 – 266.
Geoffrey R. Grimmett, [*Percolation*]{}, Springer-Verlag, Berlin, (1999).
Olle H[ä]{}ggstr[ö]{}m, Probability on Bunkbed Graphs, [*Proceedings of FPSAC’03, Formal Power Series and Algebraic Combinatorics*]{} Linköping, Sweden 2003. Available at\
http://www.fpsac.org/FPSAC03/ARTICLES/42.pdf
Svante Linusson, On percolation and the bunkbed conjecture, [*Preprint 2008*]{}[****]{}. arXiv:0811.0949.
Svante Linusson, A note on correlations in randomly oriented graphs, [*Preprint 2009*]{}.[****]{} arXiv:0905.2881.
Colin McDiarmid, General percolation and random graphs, [*Adv. in Appl. Probab.*]{} [**13**]{}, 40–60 (1981).
[^1]: Svante Linusson is a Royal Swedish Academy of Sciences Research Fellow supported by a grant from the Knut and Alice Wallenberg Foundation.
[^2]: This research was conducted when both authors visited the Institut Mittag-Leffler (Djursholm, Sweden).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we continue the investigation reported by [@RMAA] concerning the morphology of binary configurations obtained via the collapse of rotating parent gas structures with total masses in the range of M$_T$= 1 to 5 M$_{\odot}$. Here we extend the mass range and consider the collapse of two uniform gas clumps of M$_T$= 50 and 400 M$_{\odot}$, so that they also rotates rigidly in such a way that its approximate virial parameter takes the values of 0.5, 1.5, and 2.5 and their collapse is induced initially by implementing an azimuthal mass perturbation. To assess the effects of the total mass of the parent gas structure on the nature of the resulting binary configurations, we also consider the collapse of two cores of M$_T$= 1 and 5 M$_{\odot}$. We calculate the collapse of all these parent gas structures using three values of the ratio of thermal energy to potential energy and for two values of the mass perturbation amplitude. We next calculate the binary separations, masses and integral properties of the binary fragments and present them in terms of the total mass of the parent structure. For most of our models, we finally calculate the $\beta$ extreme value, so that a model with a slightly higher $\beta$ value would no longer collapse.'
author:
- 'Guillermo Arreaga-García'
title: Comparing binary systems from rotating parent gas structures with different total masses
---
Introduction {#intro}
============
Stars are formed in gravitationally collapsed clouds made of molecular hydrogen gas. These gas clouds appear to be composed of many well-defined gas substructures, so that these assembled clouds can be very large and massive. Thus, all gas structures have been characterized according to [@bergin] by means of their size and mass. For instance, a gas clump is a gas structure with mass and size in the range of 50-500 M$_{\odot}$ and 0.3-3 pc, while for a gas core these values are in the range of 0.5-5 M$_{\odot}$ and 0.03-0.2 pc.
Numerical simulations of the collapse of a rotating, self-gravitating, isolated cloud cores, began to be performed many years ago; see [@bodentohlineblack], [@boden]. One of the classic models of binary formation is based on the collapse of a rotating spherical core of 1 M$_{\odot}$, in which an azimuthal symmetric mass perturbation was initially implemented such that a binary system was formed by prompt fragmentation; see the so-called “standard isothermal test case” calculated by [@boss1979], [@boss1991], [@truelove98], [@klein99], [@boss2000], and [@kitsionas], among others.
It has always been very important to determine whether or not a gas structure is stable against gravitational collapse. As early as two decades ago, theorists proposed collapse and fragmentation criteria by constructing configuration diagrams, the axes of which are usually the ratio of thermal energy to potential energy, denoted by $\alpha$, versus the ratio of rotational energy to gravitational energy, denoted by $\beta$; see for instance [@miyama], [@hachisu1], [@hachisu2], and [@tsuribe1].
It must be emphasized that numerical simulations are still necessary to conclusively show the final configuration obtained from the collapse of a particular model of gas structure, as all these $\alpha$ versus $\beta$ configuration diagrams mentioned are, above all, mostly indicative, due to the fact that the parameter space that determines the final simulation outcome is very large.
We mention that the role played by the total mass of the parent core on the collapse results has been considered for a long time ago. For instance, [@Tscharnuter] chose initial conditions in order to study the collapse of a cloud of M=60 M$_\odot$ and compared it to the collapse of three models of 1 M$_\odot$. [@Rozyczka] investigated the collapse and fragmentation in four cloud models, of which the first three are for 1 M$_\odot$ and the last one is a model of a very massive uniform cloud of 5000 M$_\odot$. Furthermore, [@boss1986] studied in depth the collapse of rotating, uniform non-isothermal clouds by considering a four dimensional parameter space composed by $\alpha$, $\beta$, $T$ and the total mass M. The upper limit of M was 2 M$_{\odot}$ while the range of $\alpha$ and $\beta$ varied within 0.01-0.50 and 0-0.33, respectively. He found four different types of configurations and noted that as the mass $M$ decreased, cloud fragmentation is less favored.
More recently, [@RMAA] reported numerical simulations focusing on only one type of binary configuration, consisting of a well-defined pair of mass condensations that approach each other, achieve rotational speed, swing past each other, and finally separate to form the desired binary system, in which the mass condensations orbit around one another at the end of the simulation. A schematic diagram was also presented in which this particular binary configuration can be located among others. Similar configurations were also obtained by [@hennebelle] by increasing the external pressure on a rotating core. The configuration that interested [@RMAA] corresponds to the disk-bar type fragmentation studied by [@matsumoto] and [@tsuribe2].
The schematic diagram reported by [@RMAA] was constructed with the total mass of the parent core M$_T$ as the vertical axis in the range of 0.75 to 5 M$_{\odot}$, versus the dimensionless ratio of rotational energy to gravitational energy $\beta$, as the horizontal axis in the range of 0.1 to 0.21. The dimensionless ratio of thermal energy to potential energy, $\alpha$, was kept fixed for all simulations. It was also noted that the increase in the total mass of the parent core diminished the formation of the desired binary configuration and instead favored the formation of a single central mass condensation surrounded by a disk as the final result.
Observationally, much effort has been put forth for many years in order to determine some physical property that could be useful in deciding whether or not a particular gas structure will collapse. Particularly, the virial parameter, denoted here by $\alpha_{vir}$, has recently been measured in order to characterize the dynamical state of gas structures. For instance, [@kauffmann] recently compiled a catalog of 1325 molecular gas clouds of very different sizes, including estimates of their virial parameters.
There exists a critical virial parameter, denoted here by $\alpha_{vir}^{crit}$, which comes from a stability study in which perturbations both in pressure and density gradients were considered; it has been mathematically determined that $\alpha_{vir}^{crit} \approx 2$. The common belief in the recent past has been that most gas structures have $\alpha_{vir} > \alpha_{vir}^{crit}$, so that either they do not collapse or will even be diffused into the interstellar medium. It is important to note that [@kauffmann] recently observed low values of $\alpha_{vir}$ for regions of high-mass star formation; that is, these gas structures have $\alpha_{vir} < \alpha_{vir}^{crit}$, so they are expected to collapse.
The virial parameter is defined as the ratio of virial mass M$_{vir}$ to the total mass M$_T$ of a gas structure, so that in the case of a spherical model of radius $R$, it is given by $\alpha_{vir}=\frac{5 \, \sigma_v \, R}{G\, M_T}$, where $\sigma_v$ is the velocity dispersion and $G$ is the Newton’s gravitational constant. It should be noted that the velocity dispersion entering in the calculation of $\alpha_{vir}$ includes both the thermal and the non-thermal velocity components.
When the velocity distribution of a particular model of a gas structure is considered to be composed only of a thermal component, it is possible to relate the $\alpha_{vir}$ to the $\beta$ ratio by means of $\alpha_{vir}= f \, 2 \beta$, where $f$ is a form factor, which is empirically included to take into account modifications for non-homogeneous and non-spherical gas distributions.
As it seems that the dimensionless critical virial parameter, which separates the gas clouds that collapse from those that do not, is close to 2, we here prepare initial conditions of the SPH particles to have a rigidly rotating gas structure, such that its $\beta$ is given by 0.1, 0.3, and 0.48. A possible approximation would be that of using the average velocity $<v>$ of the particle distribution instead of the $\sigma_v$. Thus, for a numerical simulation, we would have a virial parameter approximated by $\alpha_{vir}=\frac{5 \, <v>^2 \, R}{G\, M_T}$. So that the values of $\beta$ given above will correspond to simulations with their approximated $\alpha_{vir}$ values given by 0.5, 1.5, and 2.5, respectively.
Thus, the question raised by the virial parameter in the observational field, of finding where is the transition between the cores that collapse to those that do not, in this paper we translated this question to the numerical simulations field, so that we here locate those models that are near to the non-collapsing regime, such that we call them “the last collapsing configuration”. As expected, the occurrence of this last collapsing configuration depends on initial values of $\alpha$ and $\beta$ and on the total mass of the parent structure.
With regard to the theoretical aspect, we mention that self-similar solutions for the collapse of isothermal spheres were first obtained by [@larson] and [@penston], and later by [@shu]. [@hunter] performed a detailed mathematical analysis of the collapse problem and re-discovered previous solutions and also obtained new solutions. Other self-similar solutions were also found for the collapse of adiabatic spheres, in which a polytropic equation of state was assumed. All these solutions provide an accurate description of the early collapse stage, which mostly proceeds isothermally, as the gas is optically thin to its own radiation. A radial density profile was obtained from these self-similar solutions. In the case of a rotating collapse with pressure, [@tohline2] was able to get also the mathematical expression of a radial density profile by using an approach based on the virial theorem. The physical parameters of the parent core are included in the density profile, so that it can be simply scaled for a different set of parameters of a second parent core.
Motivated by the work of [@RMAA] mentioned above, in this paper we want to study the formation of binary systems via the collapse of a more massive gas structure; thus, we carry out a fully three-dimensional set of numerical hydrodynamical simulations aimed to model the gravitational collapse of a M$_T \approx $ 50 and 400 $\,$ M$_{\odot} $ clumps, by using the SPH particle technique, such that a finite number of SPH particles is used to sample the entire gas structure. The formation of two antipode mass seeds during the early collapse stage of the clumps is here enforced by implementing a mass perturbation with the same mathematical structure of the density perturbation successfully used in the classic collapse calculation of [@boss2000] for an isothermal one solar mass core.
When the peak density of the collapsing core is high enough, the central region of the core become opaque to its own radiation; thus the isothermal regime breaks down so that an adiabatic regime begins. In our case, this transition is taken into account by using the barotropic equation of state, first proposed by [@boss2000], which is characterized by a critical density.
Now, the value of this critical density depends on how much mass surrounds and obscures the central region of the core. In general, it is expected that the more massive the parent cores are, the critical density at which the obscuration occurs may be lower. Strictly speaking, this would imply to change the value of the critical density of the barotropic equation of state in our suite of simulations.
In this paper, we were able to follow the collapse of the clumps up to three orders of magnitude in density within this adiabatic regime, where there are no scaling solutions known, out of which new solutions can be simply scaled from the initial mass of the parent core, as can be done in the isothermal collapse. Because of this, the treatment of this advanced collapse stage must be investigated by means of numerical simulations. As expected, for the first collapse stage, no significance differences in the collapse of a higher mass structure are seen. However, as we will see later, the influence of a higher mass parent structure on the collapse get manifested only in the non-isothermal collapse stage; see [@sterzik].
In this paper we re-simulate two core models of 1 and 5 M$_{\odot}$ with $\beta$ also given by 0.1, 0.3, and 0.48, respectively. These models are similar but not identical to those already reported by [@RMAA]. We consider here all those collapse models with values of $\alpha$ given by 0.1, 0.2, and 0.3, in order to make a comparison of our results with those of [@riaz], who studied the thermal sensitivity of binary formation via the collapse of cores. To complement the qualitative study of binary configurations that we carry out in the first part of this paper, we next present a quantitative analysis of some of the resulting binary configurations, including the calculation of the binary separation and masses of the fragments of some particular binary configurations, and we present them in terms of the total mass of the parent structure.
It should be mentioned that we have not considered here the collapse of a more massive gas structure, such as a gas cloud with mass and size of the order of 10$^3$-10$^4 \,$ M$_{\odot}$ and 2-15 pc. This is due to the fact that it has been suggested by several authors that the observed masses and core separations in regions of high-mass star formation cannot be explained by invoking only fragmentation via gravitational instability; see [@liu] among others. Thus, the formation process of very massive dense cores in high-mass star formation regions needs other mechanisms to be taken into account other than that of thermal fragmentation (also called Jeans fragmentation), for instance, turbulent fragmentation or magneto-hydrodynamical induced fragmentation.
However, for both the cores and the intermediate mass clumps, such as those considered in this paper, it is expected that only thermal fragmentation is relevant to explain their fragmentation properties; see [@palau14], [@palau15] and [@busquet].
We finally mention that the collapse of cores to model isolated or binary star formation is an active field of research, for example, the evolution of rotating cloud cores including turbulent velocity initial distributions has been investigated by many authors; see for instance: [@bate2002], [@bate2003], [@delgadonante2004a], [@delgadonante2004b], [@batebonnell2005] and [@goodwin]. Collapse calculations of an isothermal cloud core rotating in a uniform magnetic field have been presented, among others, by [@price], [@hennnebelleteyssier], [@machida2] and [@bosskeiser].
The physical state of the parent gas structures {#sec:parent}
===============================================
All the parent gas structures considered in this paper are uniform spheres, which are rigidly rotating around the $z$ axis with an angular velocity $\Omega$, so the initial velocity of the $i-th$ SPH particle is given by $(-\Omega\, y_i,\Omega\, x_i,0)\;$.
The time needed for a test particle to reach the center of a gas sphere, when gravity is the only force acting on it, is defined as the free fall time $t_{ff}$, and it is given by $t_{ff} \approx \sqrt{ \frac{3\, \pi}{32 \, G \, \rho_0}}$, where $\rho_0$ is the average density. We will use $\rho_0$ and $t_{ff}$ to normalize the figures that appear in the coming sections.
Following [@bodeniv], the dynamical state of a general gas structure is usually characterized by the values of the dimensionless ratios $\alpha$ and $\beta$, which are given by
$$\begin{array}{l}
\alpha \equiv \frac{E_{therm}}{\left|E_{grav}\right|} \vspace{0.25 cm}\\
\beta \equiv \frac{E_{rot}}{\left|E_{grav}\right|}
\end{array}
\label{defalphabeta}$$
For a spherical model of a gas structure of total mass $M_T$, the total gravitational potential energy is approximated by $<E_{grav}> \approx - \frac{3}{5} \; \frac{G\, M_T^2}{R}$, where $R$ is the sphere radius. The average total thermal energy $<E_{therm}>$ (kinetic plus potential interaction terms of the molecules) is $ <E_{therm}> \approx \frac{3}{2} {\cal N} \, k \, T = \frac{3}{2} M_T\, c_0^2$, where $k$ is the Boltzmann constant, $T$ is the equilibrium temperature, ${\cal N}$ is the total number of molecules in the gas, and $c_0$ is the speed of sound. The rotational energy of the clump is approximately given by $E_{rot}=\frac{1}{2} \; I\,\Omega^2=\frac{1}{2}\,\frac{J^2}{I}
\approx \frac{1}{5}\, M_T\, R^2 \, \Omega^2$, where $I\approx \frac{2}{5} \, M_T\,R^2$ is the moment of inertia and $J=I\,\Omega_0$ is the total angular momentum.
If we consider that the velocity dispersion $\sigma_v$ of a gas structure is directly related only to the thermal velocity component, then $\sigma_v = \sqrt{\frac{k\, T_{kin} }{17 \, m_H} }$ where $m_H$ is the molecular mass of the main gas component and $T_{kin}$ is the kinetic temperature. In this case, an approximation to be used in this paper would be using the average velocity $<v>$ of the simulation particles instead of the $\sigma_v$. Thus, we would have a virial parameter approximated by $\alpha_{vir}=\frac{5 \, <v>^2 \, R}{G\, M_T}$.
In this paper, we have carefully selected the values of $c_0$ and $\Omega$, so the energy ratios obtained by considering all the simulation particles of an initial configuration (the snapshot zero of a simulation) are initially given by a pairs of values taken from the following sets:
$$\begin{array}{l}
\alpha \equiv 0.3, 0.2, 0.1 \vspace{0.25 cm}\\
\beta \equiv 0.1, 0.3 , 0.48
\end{array}
\label{valuesalphaybeta}$$
so that the corresponding values of the virial parameter for all our models are approximately given, respectively, by $$\alpha_{vir} \equiv 0.5,1.5, 2.5
\label{valuesalphavir}$$
The cores {#subsec:core}
---------
We will consider the gravitational collapse of two cores: the first is a variant of the so-called “standard isothermal test case,” which was first calculated by [@boss1979] and later calculated by [@burkertboden93] and [@bateburkert97]; the main outcome of this classic model was a protostellar binary system. In this paper, the core radius is R=4.99 $\times 10^{16} \,$ cm $\equiv$ 0.016 pc and its mass is M$_T$=1 M$_{\odot}$. Thus, the average density and the corresponding free fall time of this core are $\rho_0=$3.8 $\times 10^{-18}\, $ g cm$^{-3}$ and $t_{ff} \approx $1.0 $\times 10^{12} \,$ s $\equiv $ 34077 yr, respectively.
The radius and mass of the second core are R=6.16 $\times 10^{17} \,$ cm $\equiv $0.2 pc and M$_T$=5 M$_{\odot}$, respectively. This core has an average density of $\rho_0=$1.0 $\times 10^{-20}\, $ g cm$^{-3}$ and a $t_{ff} \approx $2.0 $\times 10^{13} \,$ s $\equiv 660997 \,$ yr, respectively.
Despite the size and mass differences of these gas structures, both of them are still cores, as defined statistically by [@bergin].
The clumps {#subsec:clump}
----------
The first clump structure considered in this paper has a radius and mass given by R=0.3 pc $\equiv$ 9.24 $\times 10^{17}$ cm and M$_T$=50 M$_{\odot} $, respectively. Its average density is given by $\varrho_1=$3.0 $\times 10^{-20}$ g cm$^{-3}$.
The radius and mass of the second clump are given by R=1 pc $\equiv$ 3.08 $\times 10^{18}\,$cm and M$_T$=400 M$_{\odot} $, respectively. Its average density is $\varrho_2=$6.5 $\times 10^{-21}\, $ g cm$^{-3}$ and its corresponding free fall time is $t_{ff} \approx 2.6 \times 10^{13} \,$ s $\equiv $826 247 yr.
We emphasize that the radius and mass of the clumps have been changed with respect to those of the cores. In spite of this, it is still possible to make a comparison of the collapse results, as the focus of this paper is to study the effects of changing the values of $\alpha$ and $\beta$ on the collapse of different parent gas structures. Besides, if we had kept unchanged some physical property of the clump structures, so that we had used the core radius, while the total parent mass is different to that of the cores, then the resulting clump structure would be un-physical, in the statistical sense defined by [@bergin].
The mass perturbation {#subs:initialmasspert}
---------------------
For every SPH particle i of a simulation with total number of particles N$_p$, that is, i=1..N$_p$, the particle mass is given by $m_0=\frac{M_T}{N_p}$; then we implement a mass perturbation of the type
$$m_i=m_0+m_0*a \cos\left(m\, \phi_i \right)
\label{masspert}$$
where the perturbation amplitude is set to the values of a=0.1 and a=0.25; the mode is fixed to m=2 and $\phi$ is the azimuthal spherical coordinate.
This mass perturbation scheme was successfully applied in many papers on collapse; see for instance [@NuestroApJ], [@NuestroRMAA], [@NuestroPlummer], and notably [@springel], when the Gadget2 code was proven, among other tests, with the calculation of the isothermal collapse, where the unequal mass particles method was validated.
Resolution {#subs:resol}
----------
[@truelove] demonstrated the need of fulfill appropriate spatial resolution requirements in order to avoid the occurrence of artificial fragmentation in a collapse simulation.
Following @bateburkert97, the smallest mass that a $SPH$ calculation can resolve, $m_r$, is given by $m_r \approx 2 \, N_{neigh} \, M_J$ where $N_{neigh}$ is the number of neighboring particles included in the $SPH$ kernel and $M_J$ is the spherical Jeans mass $M_J$, which is defined by
$$M_J \equiv \frac{4}{3}\pi \; \rho \left(\frac{ \lambda_J}{2} \right)^3
= \frac{ \pi^\frac{5}{2} }{6} \frac{c^3}{ \sqrt{G^3 \, \rho_m} } \;.
\label{mjeans}$$
where $\lambda_J$ is Jeans wavelength and $\rho_m$ is the peak density reached in a simulation. Therefore, the smallest mass particle $m_p$ in our simulations must at least be such that $\frac{m_p}{m_r}<1$.
Let us assume that $N_{neigh}=40$, $\rho_m$=1.0 $\times \, 10^{-11}$ g cm$^{-3}$ and that the sound speed $c$ varies within the values 10 369 and 17 868 cm s$^{-1}$, which correspond to runs with $\alpha$=0.1 and 0.3 of the 1 M$_{\odot}$ model, respectively. In this case, the $m_r$ is within 3.0 $\times 10^{27}$ and 1.5 $\times 10^{28}$ g. In the simulations of a 1 M$_{\odot}$ core, we used N$_p$=2 000 000 of SPH particles, so that $m_p$ is $m_p=9.9 \times 10^{26}$ g, and then we have that the ratio of masses is $m_p/m_r$ 0.06 and 0.3. Therefore The number of particles is high enough to fulfill the resolution requirements described by [@truelove].
For the 5 M$_{\odot}$ core, the particle mass is $m_p$=4.9 $\times 10^{27}$ g while the sound speed varies within 9500.0 and 12400.0 cm s$^{-1}$, so that by using again N$_p$=2 000 000 of SPH particles, the ratios $m_p/m_r$ are appropriate up to peak densities smaller than $\rho_m$=5.0 $\times \, 10^{-12}$.
For the 50 M$_{\odot}$ clump, we used 12 000 000 of SPH particles, so the particle mass is $m_p=8.2 \times 10^{27}$ g. In this case, the sound speed varies within 17 497 and 29 576 cm s$^{-1}$, and for a peak density of 1.0 $\times 10^{11}$, $m_r$ are 0.5 and 0.1, respectively.
For the 400 M$_{\odot}$ clump, a very large number of particles is needed for the simulations to fulfill the resolution requirements. Because computational limitations, we could not achieve the desired resolution and despite of this, we evolved these simulations using only 2 million of SPH particles; for this lack of reliability, we consider them useful only for comparison with the lower mass simulations.
The barotropic equation of state {#subs:thermo}
--------------------------------
In order to ensure the change in the thermodynamic regime from isothermal to adiabatic, we here implement the barotropic equation of state proposed by [@boss2000]:
$$p= c_0^2 \rho \left[ 1 +
\left( \frac{\rho}{\rho_{crit}}\right)^{\gamma -1 }
\right]
\label{beos}$$
where $\gamma$=5/3 and c$_0$ is the sound speed. The critical density $\rho_{crit}$ has only been given here the value $\rho_{crit}=5.0 \times 10^{-14} \,$ g cm$^{-3}$. As we will show in the following sections, in the simulations considered in this paper, the average peak density increases up to 3 orders of magnitude within the adiabatic regime.
Evolution Code {#subs:code}
--------------
To follow the gravitational collapse of our models, in this paper we use the fully parallelized particle-based code Gadget2, which is based on the tree-PM method for computing the gravitational forces and on the standard SPH method for solving the Euler equations of hydrodynamics; see [@springel] and also [@gadgets]. The Gadget2 code has implemented a Monaghan-Balsara form for the artificial viscosity; see [@mona1983] and [@balsara1995]. The strength of the viscosity is regulated by setting the parameter $\alpha_{\nu} = 0.75$ and $\beta_{\nu}=\frac{1}{2}\, \times \alpha_v$; see Equations 11 and 14 in [@springel]. We have fixed the Courant factor to $0.1$.
The initial setup {#subs:setup}
-----------------
Herein we used a radial mesh with spherical geometry, such that a set of concentric shells was created and populated with SPH particles by means of a Monte Carlo scheme to set the initial particle configuration.
Thus, all the SPH particles of each simulation, were located randomly in all the available surfaces of each spherical shell. The total mass in each shell is kept constant, so that the global density of the gas structure is also constant. In order to have a constant density distribution in a local sense, we next applied a radial perturbation to all the particles of a given shell such that any particle could be randomly displaced radially outward or inward, but preventing a perturbed particle from reaching another shell; see section 2.1 of [@RMAA].
The collapse models {#sec:modelos}
-------------------
In Table \[tab:modelos\] we summarize the models considered in this paper and the main configurations obtained. Column 1 shows the number of the model; column 3 shows the total mass of the parent structure; column 4 shows the value of the mass perturbation amplitude, defined by Eq.\[masspert\]; column 5 and 6 show the values of the dimensionless ratios $\alpha$ and $\beta$, respectively; column 7 shows the type of configuration obtained and column 8 shows the number of the figure and panel, in which the solution configuration is shown[^1].
We now emphasize five types of solution configurations that appear frequently, which may be termed as follows: (i) a face-to-face configuration, characterized by the presence of two fragments located one in front of the other, so that the connecting filament has been entirely accreted; (ii) a connecting filament configuration, in which the filament still plays a relevant role as a gas bridge between the two fragments; (iii) a central primary configuration, characterized by a rotationally supported solution surrounded by spiral arms; (iv) a filament configuration, characterized by a thin and dense longitudinal collapsed structure without fragments present; (v) a binary configuration, characterized by two fragments, so that they are orbiting one about the other.
All of these configurations have already been observed before in the literature; in the next section we describe how these configurations are transformed one into another because the change in parameters of the parent gas structure.
Results {#sec:resultados}
=======
The main results of this paper are shown in color iso-density plots for a slice of particles around the equatorial plane of the spherical gas structure.
As we want to compare several collapse models in this paper, we use only one panel per model, which usually will correspond to the last available snapshot, and we use three panels to compose a mosaic in order to illustrate the resulting configurations obtained from a parent gas structure of total mass M$_T$.
It should be mentioned that the vertical and horizontal axes of the iso-density plots indicate length in terms of the sphere radius, which varies from zero to one in all the plots for the initial snapshot, irrespective of the model; therefore the Cartesian axes X and Y vary initially from -1 to 1. In order to facilitate the comparison of the results, in each mosaic at least, we use the same length scale for all the panels.
A comparison between models is still possible even at slightly different output times because most of the configurations have already entered a stable stage. Otherwise, the time evolution should be done with the sink technique introduced by [@batebonnellprice95], so that the system can be evolved further in time.
The panels of a mosaic have the same value of the virial parameter $\alpha_{vir}$ (equivalently, with the same $\beta$ values of 0.1,0.3 and 0.48) for a given value of $\alpha$. We observed that all the models with the highest $\alpha$ and $\beta$, given by 0.3 and 0.48, respectively, do not collapse but get dispersed. As we want to include only collapsing configurations, these models have been skipped and for this reason some mosaics are formed only by two panels. The last collapsing configurations have been grouped in other figures, because they are more likely to make better sense in mathematical terms, as their $\beta$ values may be too high to keep physical significance in the theory of star formation.
Finally, for each model with M$_T$, $\alpha$ and $\beta$ given, we build two mosaics, each corresponding to a perturbation amplitude $a$ with values of $a=$0.1 and $a=$0.25, respectively; see Eq. \[masspert\]; see Table \[tab:modelos\].
The models with M$_T$= 1 M$_{\odot}$ {#subsec:rescore1}
------------------------------------
The figures \[CPrueba60p1\], \[CPrueba60p2\], and \[CPrueba60p3\] show the resulting systems obtained from the gravitational collapse of a M$_T$= 1 M$_{\odot}$ core, similar to “the standard test case”. In Fig.\[CPrueba60p1\] all the panels show a well-defined pair of mass condensations connected by a filament. When $\beta$ increases, the filament disappears so that the last collapsing configuration shown in panel 1 of Fig.\[lastcolconfM1\] is a well defined face-to-face configuration. By observing the axes of the panels it can be seen that the separation of the resulting fragments has significantly been increased for this last collapsing configuration.
When $\alpha$ is increased from 0.1 to 0.2, the pair of fragments can move one towards the other, so that an orbiting binary system is formed for lowest value of $\beta$, and for higher values of $\beta$, the configurations end with a central mass condensation surrounded by short spiral arms, as can be seen in the left and right panels of Fig.\[CPrueba60p2\] and also in the panel 2 of Fig.\[lastcolconfM1\].
In Fig.\[CPrueba60p3\], one can see that even for the lowest value of $\beta$, the resulting configuration is a central primary, which is rotationally supported, such that the gravitational collapse slows down.
It is interesting to observe the change in the outcome of models with the same $\beta$, but in which the $\alpha$ has been increased: from a face-to-face configuration type to a central primary type. This behavior was also observed by [@riaz].
When the mass perturbation is increased from $a=$0.1 to $a=$0.25, we can see in Fig.\[CPrueba70p1\] and in panel 4 of Fig.\[lastcolconfM1\] that a similar behavior is obtained to that which was seen in Fig. \[CPrueba60p1\], but the binary separations are quite larger than that seen in the corresponding panels of Fig.\[CPrueba60p1\].
We also noted that the number of orbiting binary systems is greatly increased, as can be seen in Figs.\[CPrueba70p1\] and \[CPrueba70p2\]; only the last collapsing configuration finishes as a central primary, as can be seen in panel 6 of Fig.\[lastcolconfM1\].
The models with M$_T$= 5 M$_{\odot}$ {#subsec:rescore2}
------------------------------------
For all the values of $\beta$, we see that a pair of mass condensations connected by a very smooth filament are formed, as can be seen in Fig.\[CPrueba30p1\]. When $\alpha$ is increased from 0.1 to 0.2, the mass condensations are very weak or they even disappear, while the filaments are thinner and more pronounced; see Fig. \[CPrueba30p2\]. For the highest value of $\alpha=$0.3 considered here, we see in Fig.\[CPrueba30p3\] that the corresponding filaments become thinner and narrower and again a rotationally supported central primary surrounded by long spiral arms is seen as the last collapsing configuration for the highest value of $\beta$.
When the perturbation scale takes the new value of $a=0.25$, the filaments almost disappear and instead a pair of mass condensations are seen located face-to-face and more pronounced than before, as one can see in Fig.\[CPrueba40p1\]. Later, when $\alpha$ increases to 0.2, the mass condensations are weaker but still face-to-face, while no dense filaments are seen in Fig.\[CPrueba40p2\]. We observe again that the binary separation is significantly increased in these runs, shown in Fig.\[CPrueba40p1\] and Fig.\[CPrueba40p2\], as compared with those of Fig.\[CPrueba30p1\] and Fig.\[CPrueba30p2\].
For the highest $\alpha=$0.3 value considered here, we see in Fig.\[CPrueba40p3\] the appearance of weaker filaments and again the transition to a central primary.
In any run of the models with M$_T$= 5 M$_{\odot}$ we have not directly observed the formation of an orbiting binary system.
The clump models {#subsec:resclump}
----------------
We first notice that there is almost no difference between the configurations obtained for the clump of M$_T$= 50 M$_{\odot}$ with those obtained for the M$_T$= 400 M$_{\odot}$, as can be seen by comparing Figs.\[CPrueba80p1\],\[CPrueba80p2\],\[CPrueba80p3\] and \[CPrueba90p1\],\[CPrueba90p2\], \[CPrueba90p3\] with the corresponding figures of the largest mass clump. We take advantage of this similarity by using only the last collapsing configurations of the M$_T$= 400 M$_{\odot}$ clump to compare with the lower mass models of previous sections.
In Fig. \[CPrueba80p1\] one can see that configurations of the face-to-face type connected by a weak filament are mainly formed for the two lower values of $\beta$. These filaments are fainter than those observed in the M$_T$= 1 M$_{\odot}$ model shown in Fig.\[CPrueba60p1\] while they are very similar to those obtained for the M$_T$= 5 M$_{\odot}$ model shown in Fig.\[CPrueba30p1\].
In the right panel of Fig. \[CPrueba0p1\] and panel 1 of Fig.\[lastcolconfM400\], which corresponds to a $\beta=$0.3 and $\beta=$0.7, respectively, still show the face-to-face configuration, as was the case for the two models of smaller total mass considered; see the panels 1 of Fig.\[lastcolconfM1\] and Fig.\[lastcolconfM5\].
When $\alpha$ takes the value of 0.2, the filament formed for a large $\beta$ is fragmenting as can be seen in the right panel of Fig. \[CPrueba80p2\], while the opposite case was seen for the M$_T$= 5 M$_{\odot}$ model in Fig.\[CPrueba30p2\]; that is, the filament becomes denser and thinner while in the M$_T$= 5 M$_{\odot}$ model shown in Fig.\[CPrueba60p2\] there is no filament for comparison.
When $\beta$ takes its highest possible value, still in the collapsing regime, the filament of the M$_T$= 400 M$_{\odot}$ model is replaced by a central primary configuration surrounded by spiral arms; see the panel 2 and 3 of Fig. \[lastcolconfM400\] and the bottom left panel of Fig. \[CPrueba0p3\]. This behavior was also observed in the M$_T$= 5 M$_{\odot}$ models in Fig.\[CPrueba30p2\] and Fig.\[CPrueba30p3\] but was not observed in the M$_T$= 1 M$_{\odot}$ models with $\alpha=$0.2 in Fig.\[CPrueba60p2\] nor with $\alpha=$0.1 in Fig.\[CPrueba60p1\], although it occurred for this model with $\alpha=$0.3, as can be seen in Fig.\[CPrueba60p3\].
When the mass perturbation is increased to $a=$0.25, more configurations of the face-to-face type appear connected by a tenuous bridge of gas; see Fig. \[CPrueba50p1\]. This behavior was also observed in the M$_T$= 5 M$_{\odot}$ model shown in Fig.\[CPrueba40p1\], but it must be mentioned that the binary separation is much larger in the former model than in the latter. Something similar was also observed for the M$_T$= 1 M$_{\odot}$ model, but there the filaments are more pronounced, as can be seen in Fig.\[CPrueba70p1\].
The runs with $\alpha=$0.3 are similar for the M$_T$= 50 M$_{\odot}$ models and M$_T$= 5 M$_{\odot}$ shown in Fig. \[CPrueba50p2\] and Fig.\[CPrueba40p1\], respectively, in which we see configurations of the face-to-face type. It should be noted that in the M$_T$=1 M$_{\odot}$ model, the configurations formed are not at all similar to the ones seen for the corresponding two models with larger parent mass, as an orbiting binary configuration is only seen in Fig.\[CPrueba70p2\].
For the $M_T=$ 400 $M_{\odot}$ models with $\alpha=$0.3, we see a well-defined filament that is transformed to a central primary configuration when $\beta$ takes its highest value; see Fig. \[CPrueba50p3\]; this was also observed in Fig.\[CPrueba40p3\] for the M$_T$= 5 M$_{\odot}$ model. But again, there is no possible comparison with the configuration obtained for the M$_T$= 1 M$_{\odot}$ model shown in Fig.\[CPrueba70p3\], in which orbiting binaries were indeed observed even for this higher value of $\alpha$.
For all the models of M$_T$= 50 M$_{\odot}$, we did not observe the formation of any orbiting binary system.
The last collapsing configuration {#subsec:lastcoll}
---------------------------------
We measure the uppermost $\beta$ values for all models except for the clump of M$_T$= 50 M$_{\odot}$, so that the parent structure is still within the collapsing regime to the left of these curves, while to the right of each curve, there is no collapse. The configurations found are shown in Figs. \[lastcolconfM1\], \[lastcolconfM5\] and \[lastcolconfM400\].
In Fig. \[lcurve\] we observe that the curve for the models with a=0.25 are always located to the right of the corresponding curve for a=0.1; this fact is easy to understand, as the mass perturbation weighs more, then the resulting fragments are more massive, so they need more rotational energy to reach the level in which the equilibrium between the centrifugal force and the gravitational force is to be overcome. When this happens, the fragments are separated indefinitely and the gas structure does not collapse.
A quantitative comparison of the resulting configurations {#subsec:qc}
---------------------------------------------------------
We measure the most important integral properties of the resulting collapse configurations: for binaries, we consider the fragment masses and the binary separations; for the central primaries, we consider the central mass and radius; for all the resulting fragments or primaries, we also calculate the values of the energy ratios $\alpha_f$ and $\beta_f$.
In order to calculate these properties, we proceeded as follows: we locate the highest density particle in the fragment’s region; this particle is considered the center of the fragment. The binary separation is simply calculated as the distance between the centers associated with each fragment. We next find all the particles which have density above or equal to some minimum density value, given in advance as $\log_{10} \left( \rho_{min}/\rho_0 \right)=0.0$ for all the models. We finally check that these particles are also located within a given maximum radius r$_{max}$ from the fragment’s center. This set of particles allowed us to calculate the fragment integral properties, as the mass, including the $\alpha_f$ and $\beta_f$; see [@SegRMAA].
The results obtained for the face-to-face configurations are shown in Figs.\[SepFrags\], \[MassFrags\] and \[AlphavsBetaFrags\]. For the central primary configurations, we report the mass of the central condensations results and their integral properties, which are shown in Figs.\[MassCenP\] and \[AlphavsBetaCenP\].
All the fragment properties of the face-to-face and of central primary configurations have also been reported in Tables \[tab:propfrags\] and \[tab:propfrags2\], respectively. The entries of these Tables are as follows. The first column shows the number of the model; the total parent mass M$_T$ is shown in the second column; column third shows r$_{max}$ in terms of the initial sphere radius R$_0$; columns 4, 5 and 6 of Table \[tab:propfrags\] show the mass of the fragments given in terms of M$_T$ and the binary separation in astronomical units; in the case of Table \[tab:propfrags2\], the last column shows the mass of the central primary region in terms of M$_T$. The last columns give the values of the $\alpha_f$ and $\beta_f$.
Discussion {#sec:dis}
==========
The mass seeds, implemented by means of the mass perturbation of Eq. \[masspert\], accrete mass slowly during the early evolution stage of all the simulations until they become a well-defined pair of mass condensations that move through a gas of particles. Shortly thereafter, when the mass condensations are massive enough, their translational motion can be slowed down or even stopped as a consequence of the dynamical friction of the surrounding gas. Then, the mass condensations appear to be face-to-face. At this time, a competition begins between their gravitational force, which favors their approach, and the centrifugal force due to the parent structure rotation, which favors their separation. Also, the mass condensations may or may not be connected by a filament, and as we have seen, these connecting filaments can be of very different types depending on the total mass of the parent structure and on the values of $\alpha$ and $\beta$.
For the M$_T$= 1 M$_{\odot}$ core models, a good combination of these forces occurs such that the assembled mass condensations are sufficiently massive yet the centrifugal force is strong enough for their gravitational attraction to make them approach each other, swing past each other, and finally separate to form an orbiting binary system. We have observed that the occurrence of these events is significantly increased when the mass perturbations weigh more, as the mass condensations are naturally more massive to overcome their centrifugal force more easily, while the $\beta$ must still be high enough to make them swing past one another.
By contrast, when the mass condensations are too massive, they approach each other, make contact, and finally merge to form a single central primary condensation. In this case, the gravitational force between the mass condensations has too easily overcome the rotational centrifugal force.
For the M$_T$= 5 M$_{\odot}$ and M$_T$= 50 M$_{\odot}$ models, we observed that the mass condensations remain face-to-face in most of the runs, such that if $\beta$ increases further, then the separation of the face-to-face configuration gets larger up to the point that the parent structure no longer collapses. However, for higher values of $\beta$, these face-to-face configurations give rise to a central primary configuration, such that for a slightly higher $\beta$, the parent structure does not collapse but instead expands outwards but with spherical symmetry.
We see in Fig.\[SepFrags\] that the binary separation is correlated with the mass of the parent gas structure. It seems that no correlation exists in the case of the fragment mass, as can be see in Fig.\[MassFrags\]. This missing correlation is expected on physical grounds, in fact, for the Taurus dark cloud, @myers reported a correlation between the mass of the newly formed stars and the mass of the associated dense proto-stellar cores. More recently, large proto-stellar masses were observed by [@tobin].
A potentially interesting issue is the fact that some binary systems have a barycentre that is not at the Cartesian coordinate origin; see for instance the panels of Fig.\[CPrueba30p1\]. A possible explanation for this fact is that the masses of the fragments of some binary systems are different; see Fig.\[MassFrags\] and Table\[tab:propfrags\]. While it is true that the perturbation mass implemented in Eq.\[masspert\] has axial symmetry, it should be remembered that each spherical radial shell was randomly populated by means of a Monte Carlo method; see \[subs:setup\]. We think that the origin of the unequal masses of the fragments and the displacement of the barycentre can be explained by the development of an asymmetrical random mass seed, that grow over time and are manifested mainly later, in the more evolved binary systems. The fact that some binary systems show upward displacement while others downward, gives support to this explanation.
We notice that the mass fraction of the central primary configurations is similar for all the models reported in Table \[tab:propfrags2\]; this is so because the number of SPH particles entering into the central region is similar for all the models; therefore, the mass of the central primary configuration does not scale with the mass of the parent structure;see Fig.\[MassCenP\].
It is clear that the values calculated for the energy ratios $\alpha_f$ and $\beta_f$, unfortunately depend on the values chosen for $\rho_{min}$ and $r_{max}$, so that there inevitably is certain ambiguity in defining the fragment’s boundaries. Despite of this, when we calculated the $\alpha_f$ and $\beta_f$ values we observe that some binary fragments do show a clear tendency to virialize, as it can be appreciated in Fig.\[AlphavsBetaFrags\]. We emphasize that a similar conclusion can be drawn from the calculation of @NuestroRMAA, where plots of the $\alpha_f$ and $\beta_f$ time evolution were also presented. On the contrary, for the central primaries, we do not observe any trend to approach the virial line; see Fig.\[AlphavsBetaCenP\].
Concluding Remarks {#sec:conclu}
==================
In this paper, we have considered the gravitational collapse of several rotating spherical gas structures of very different size and mass, such that the collapse was triggered initially by means of the same azimuthal mass perturbation, whose amplitude has been allowed to take the values $a=$0.1 and $a=$0.25. For the sake of comparison between our models, we have used the same set of values of the $\alpha$ and $\beta$ ratios.
The range of $\alpha$ considered here, 0.1-0.3, was similar to that of [@riaz]; this range reflects the observational fact that the temperatures of the gas structures are mostly in the range of 9-12 K; see [@bergin]. The range of $\beta$ considered here, 0.1,0.3 and 0.48, was initially motivated by recent observations of the virial parameters of clouds, so that the critical virial parameter is within the chosen range, as we mentioned in Sect.\[intro\].
However, we have extended this range of $\beta$ in order to look for the last collapsing configurations, for which purpose the new range has been increased to values of 0.9 for the lowest $\alpha$ and around 0.4 for the highest $\alpha$. This range of $\beta$ values is wider than those considered by other authors; see for instance, [@tsuribe1], who used a maximum $\beta$ of 0.3.
We have thus revisited many binary configurations already seen in other works, but here the configurations change depending on the mass of the parent structure. Based on these results, we present the following assessment of the effect of parent mass on the collapse.
The resulting configurations for panels with $\alpha=$0.1 have a well-defined pair of mass condensations connected by a dense filament for the M$_T$= 1 M$_{\odot}$ model. The connecting filaments become weaker and almost disappear when the total mass increases to M$_T$= 5 M$_{\odot}$ and M$_T$= 50 M$_{\odot}$, respectively. The last collapsing configurations are of the same kind; that is, of the face-to-face type; see Sect.\[sec:resultados\].
When $\alpha=$0.2, we have observed the formation of many orbiting binary configurations for the M$_T$= 1 M$_{\odot}$ model; this tendency becomes more pronounced when the mass perturbation increases from $a=$0.1 to $a=0.25$, so that the more massive the mass condensations, the greater the probability of an orbiting binary configuration.
For the M$_T$= 5 M$_{\odot}$ model, what has been observed primarily are configurations where thin and dense filaments connect two very small mass condensations. When the mass of the parent gas structure is M$_T$= 50 M$_{\odot}$, mainly face-to-face configurations are formed, in which the mass condensations are predominant and the filaments become thicker and less dense than those seen in the previous M$_T$=5 M$_{\odot}$ model. The last collapsing configurations are of three kinds; i.e., some runs finish with a central primary (see Figs.\[CPrueba60p1\] and \[CPrueba0p1\]); other runs finish with a binary (see Figs.\[CPrueba70p2\], Figs.\[CPrueba40p2\] and \[CPrueba50p2\]), and finally, only one run finishes with a thin and dense filament (see Fig.\[CPrueba30p2\]).
When $\alpha=$0.3, we have seen the formation of most of the types of binary configurations discussed in this paper; i.e., some runs show a central primary, and other runs show a well-defined pair of mass condensations connected by a filament that becomes weaker as the total parent mass increases. However, the last collapsing configurations are of the type of a central primary for all the models.
It must be emphasized that we have seen more differences between the outcomes of the M$_T$=1 M$_{\odot}$ model and those obtained from the M$_T$=5 M$_{\odot}$, M$_T$=50 M$_{\odot}$ and M$_T$=400 M$_{\odot}$ models. In addition, we have seen many similarities in the outcomes of these last two models. These observations indicate that there is a critical total mass not much larger than one solar mass, for which the mass condensations to be formed will be too massive and the centrifugal force will not be able to make them start to orbit around each other.
As we saw in Sect.\[sec:dis\], the binary systems formed from massive gas structures can be the origin of wide binaries, in which a very large separation between the mass components can be reached for large values of $\beta$. It is interesting to mention that a high frequency of wide binaries has recently been observed in regions having very different physical properties; see [@duch].
Lastly, we mentioned that the mass perturbation shown in Eq. \[masspert\] was invented to favor fragmentation of the so-called “standard isothermal test case” of M$_T$=1 M$_{\odot}$. For the results outlined in this paper and those reported in other papers, we conclude that this mass perturbation scheme works quite well for the low-mass gas structure, but needs some improvement for the higher mass case.
In order to calculate the integral properties of the some configurations, we took particles by applying selection criteria based on two parameters, a minimum density and a maximum radius, whose values were fixed in advance. Thus, one would expect slight differences in the reported results as they are definition-dependent. Nevertheless, we find that there is a clear correlation between total mass of the parent gas structure and the separation of binary fragments; we do not observe a similar correlation for the mass of the fragments.
It must be emphasized that the initial grid of spherical concentric shells, where SPH particles were randomly located on all the available radial surface of each shell, appears to be a well-suited setup for numerical simulations of binary formation as we have obtained in a natural way, that the mass of the fragments in binary configurations are not always equal one to the other.
The author thankfully acknowledge the computer resources, technical expertise and support provided by the Laboratorio Nacional de Supercómputo del Sureste de M' exico through the grant number O-2016/047; I thank ACARUS-UNISON for the use of their computing facilities in the development of this manuscript.
Arreaga-Garcia, G., Klapp, J., Sigalotti, L.G. and Gabbasov, R. 2007, ApJ, [**666**]{}, pp.290-308.
Arreaga-Garcia, G., Saucedo, J., Duarte, R. and Carmona, J. 2008, Rev. Mex. Astron. Astrophys, [**44**]{}, pp.259-284.
Arreaga-Garcia, G., and Klapp, J. 2010, Astron. Astrophys, [**509**]{}, A96.
Arreaga-Garcia, G. and Saucedo, J., 2012, Rev. Mex. Astron. Astrophys, [**48**]{}, Num. 1,pp.61-84.
Arreaga-Garcia, G., 2016, Rev. Mex. Astron. Astrophys, [**52**]{}, Num. 1,pp.1-15.
Balsara, D. 1995, J. Comput. Phys., [**121**]{}, 357.
Bate, M.R., Bonnell, I.A. and Price, N.M., 1995, MNRAS, [**277**]{}, pp. 362-376.
Bate, M.R. and Burkert, A., 1997, MNRAS, [**288**]{}, 1060.
Bate, M. R., Bonnell, I. A. and Bromm, V. 2002, MNRAS, [**332**]{},pp. L65-L68.
Bate, M. R., Bonnell, I. A. and Bromm, V. 2003, MNRAS, [**339**]{}, pp.577-599.
Bate, M. R. and Bonnell, I. A. 2005, MNRAS, [**356**]{}, pp. 1201-1221.
Bergin, E. and Tafalla, M. 2007, Annu. Rev. Astro. Astrophys., [**45**]{}, 339
Bodenheimer, P., [*Principles of star formation*]{}, Springer-Verlag, 2011.
Bodenheimer, P., Burkert, A., Klein, R.I. and Boss, A.P., in [*Protostars and Planets IV*]{}, Eds. V.G. Mannings, A.P. Boss and S.S. Russell, University of Arizona Press, Tucson, AZ, USA.
Bodenheimer, P., Tohline, J. E. and Black, D. C., 1980, International Astronomical Union, Colloquium on Stellar Hydrodynamics, 58th, Los Alamos, N. Mex., Aug. 12-15, 1980. Space Science Reviews, vol. 27, Nov.-Dec. 1980, p. 247-252; Discussion, p. 253.
Boss, A.P., 1986, ApJ, [**62**]{},pp. 519-552.
Boss, A.P., 1991, Nature, [**351**]{}, 298.
Boss, A.P., Fisher, R.T., Klein, R. and McKee, C.F. 2000, ApJ, [**528**]{}, 325.
Boss, A.P. and Keiser, S.A., 2013, ApJ, [**764**]{}, pp.1-13.
Boss, A. P. and Bodenheimer, P. 1979, ApJ, [**234**]{}, 289
Burkert, A. and Bodenheimer, P. 1993, MNRAS, [**264**]{}, 798.
Busquet, G., Estalella, R., Palau, A., Liu, H.B., Zhang, Q., Girart, J.M., Gregorio-Monsalvo, I., Pillai, T., Anglada, G. and Ho, P.T.P, 2016, ApJ, [**819**]{}, Issue 2, arXiv:1602.02500.
Delgado-Donate, E. J., Clarke, C. J. and Bate, M. R. 2004, MNRAS, [**347**]{}, pp.759.
Delgado-Donate, E. J., Clarke, C. J., Bate, M. R., and Hodgkin, S. T. 2004, MNRAS, [**351**]{}, pp. 617.
Duchêne, G., Bouvier, J., Bontemps, S., André, P. and Motte, F. 2004, Astron. Astrophys, [**427**]{}, 651.
Goodwin, S. P., Kroupa, P., Goodman, A. and Burkert, A. 2007, in Protostars and Planets V, ed. B. Reipurth, D. Jewitt, and K. Keil (Tucson: Univ. Arizona Press), 133.
Hachisu, I. and Heriguchi, Y., 1984, Astron. Astrophys,[**140**]{},259.
Hachisu, I. and Heriguchi, Y., 1985, Astron. Astrophys, [**143**]{},355.
Hennebelle, P., Whitworth, A. P., Cha, S.-H. and Goodwin, S. P. 2004, MNRAS, [**348**]{}, 687.
Hennebelle, P. and Teyssier, R., 2008, A&A, [**477**]{}, pp.25-34.
Hunter, C., 1977, ApJ, [**218**]{}, 834.
Kauffmann, J., Pillai, T. and Goldsmith, P.F.,2013, ApJ, [**779**]{}, Issue 2,pp.14.
Kitsionas, S. and Whitworth, A. P. 2002, MNRAS, [**330**]{}, 129.
Klein, R. I., Fisher, R. T., McKee, C. F. and Truelove, J. K. 1999, in Numerical Astrophysics 1998, ed. S. Miyama, K. Tomisaka and T. Hanawa (Dordrecht: Kluwer), 131.
Larson, R.B.,1969, MNRAS, [**145**]{}, 271.
Liu, H.B., Galvan-Madrid, R., Jimenez-Serra, I., Román-Zuñiga,C., Zhang, Q., Li, Z. and Chen, H-R., 2015, ApJ, [**804**]{}, Issue 1, Article id 37.arXiv:1505.04255.
Machida, N.M., 2008, ApJ, [**682**]{}, pp.L1-L4.
Machida, N.M., Tomisaka, K., Matsumoto, T. and Inutsuka, S., 2008, ApJ, [**677**]{}, pp.327-343.
Matsumoto, T. and Hanawa, T. 2003, ApJ, [**595**]{}, 913.
Myers, P. C., [*Fragmentation of Molecular Cores and Star Formation*]{}, ed. E. Falgarone, F. Boulanger and G. Duvert (Dordrecht, Kluwer), 1983, pp.221.
Miyama, S.M., Hayashi, C. and Narita, S., 1984, ApJ, [**279**]{}, 621.
Monaghan, J.J. and Gingold, R.A. 1983, J. Comput. Phys., [**52**]{}, 374
Palau, A., Estallela, R., Girart, J.M., 2014, ApJ, [**785**]{}, 42.
Palau, A., Ballesteros-Paredes, J., Vazquez-Semanedi, E., Sanchez-Monge, A., Estallela, R., Fall, S.M., Zapata, L.A., Camacho, V. Gomez, L. Naranjo-Romero, R., Busquet, G. and Fontani, F., 2015, MNRAS, [**453**]{}, Issue 4, pp.3785-3797.
Penston, M.V.,1969, MNRAS, [**144**]{}, 425.
Price, D.J. and Bate, M.R., 2007, ASTROPHYS SPACE SCI., [**311**]{}, pp.75-80.
Riaz, R., Farooqui, S.Z., and Vanaverbeke, S., 2014, MNRAS, [**444**]{}, Issue 2, pp.1189-1204.
Rozyczka, M., Tscharnuter, M. and Yorke, H.W., 1980, Astronomy and Astrophysics, [**81**]{}, pp.207.
Shu, F.H.,1977, ApJ, [**214**]{}, 488.
Springel, V., Yoshida N. and White S. D. M., 2001, New Astronomy, 6, 79.
Springel, V. 2005, MNRAS, [**364**]{}, 1105.
Stahler, S.W. and Palla, F., [*The formation of stars*]{}, Wiley-Vch, 2004.
Sterzik, M.F., Durisen, R.H. and Zinnecker, H., Astron. Astrophys, [**411**]{}, Num. 2, 2003, pp. 91-97.
Tobin, J.J, Chandler, C., Wilner, D.J., Looney, L.W.,Loinard, L., Chiang, H-F., Hartmann, L., Calvet, N., D’Alessio, P., Bourke, T.L. and Kwon, W., 2013, ApJ, [**779**]{}, Issue 2.
Tohline, J.E., 1981, ApJ, [**248**]{}, pp.717.
Tohline, J.E., 2002, Annu. Rev. Astro. Astrophys, [**40**]{}, pp.349-385.
Truelove, J.K., Klein, R.I., McKee, C.F., Holliman, J.H., Howell, L.H. and Greenough, J.A., 1997, ApJ, [**489**]{}, L179.
Tscharnuter, W., 1975, Astronomy and Astrophysics, [**39**]{}, pp. 207.
Truelove, J. K., Klein, R. I., McKee, C. F., Holliman, J. H., Howell, L. H., Greenough, J. A. and Woods, D. T. 1998, ApJ, [**495**]{}, 821.
Tsuribe, T. and Inutsuka, S-I., 1999, ApJ, [**523**]{}, pp.L155-L158. Tsuribe, T. and Inutsuka, S-I., 1999, ApJ, [**526**]{}, pp.307-313.
Tsuribe, T., 2002, Progress of theoretical physics supplement, [**147**]{}, pp.155-180.
[lcccccc]{} 1 & 1 & 0.1 & 0.1 & 0.1 & connecting filament & \[CPrueba60p1\] l\
2 & 1 & 0.1 & 0.1 & 0.3 & connecting filament & \[CPrueba60p1\] m\
3 & 1 & 0.1 & 0.1 & 0.48 & connecting filament & \[CPrueba60p1\] r\
4 & 1 & 0.1 & 0.1 & 0.758 & face-to-face & \[lastcolconfM1\] 1\
5 & 1 & 0.1 & 0.2 & 0.1 & binary & \[CPrueba60p2\] l\
6 & 1 & 0.1 & 0.2 & 0.3 & connecting filament & \[CPrueba60p2\] m\
7 & 1 & 0.1 & 0.2 & 0.48 & central primary with arms & \[CPrueba60p2\] r\
8 & 1 & 0.1 & 0.2 & 0.5344 & central primary & \[lastcolconfM1\] 2\
9 & 1 & 0.1 & 0.3 & 0.1 & central primary with disk & \[CPrueba60p3\] l\
10 & 1 & 0.1 & 0.3 & 0.3 & central primary with arms & \[CPrueba60p3\] m\
11 & 1 & 0.1 & 0.3 & 0.3983 & central primary & \[lastcolconfM1\] 3\
12 & 1 & 0.25 & 0.1 & 0.1 & connecting thin filament & \[CPrueba70p1\] l\
13 & 1 & 0.25 & 0.1 & 0.3 & connecting filament & \[CPrueba70p1\] m\
14 & 1 & 0.25 & 0.1 & 0.48 & connecting filament & \[CPrueba70p1\] r\
15 & 1 & 0.25 & 0.1 & 0.8871& face-to-face & \[lastcolconfM1\] 4\
16 & 1 & 0.25 & 0.2 & 0.1 & binary & \[CPrueba70p2\] l\
17 & 1 & 0.25 & 0.2 & 0.3 & binary & \[CPrueba70p2\] m\
18 & 1 & 0.25 & 0.2 & 0.48 & binary & \[CPrueba70p2\] r\
19 & 1 & 0.25 & 0.2 & 0.5411& binary & \[lastcolconfM1\] 5\
20 & 1 & 0.25 & 0.3 & 0.1 & binary & \[CPrueba70p3\] l\
21 & 1 & 0.25 & 0.3 & 0.3 & binary & \[CPrueba70p3\] m\
22 & 1 & 0.25 & 0.3 & 0.3983 & central primary& \[lastcolconfM1\] 6\
23 & 5 & 0.1 & 0.1 & 0.1 & connecting filament with disk & \[CPrueba30p1\] l\
24 & 5 & 0.1 & 0.1 & 0.3 & connecting filament & \[CPrueba30p1\] m\
25 & 5 & 0.1 & 0.1 & 0.48 & face-to-face & \[CPrueba30p1\] r\
26 & 5 & 0.1 & 0.1 & 0.6118 & face-to-face & \[lastcolconfM5\] 1\
27 & 5 & 0.1 & 0.2 & 0.1 & thin filament & \[CPrueba30p2\] l\
28 & 5 & 0.1 & 0.2 & 0.3 & thin filament & \[CPrueba30p2\] m\
29 & 5 & 0.1 & 0.2 & 0.48 & thin filament & \[CPrueba30p2\] r\
30 & 5 & 0.1 & 0.3 & 0.1 & connecting filament & \[CPrueba30p3\] l\
31 & 5 & 0.1 & 0.3 & 0.3 & thin filament & \[CPrueba30p3\] m\
32 & 5 & 0.1 & 0.3 & 0.3594 & central primary with arms & \[lastcolconfM5\] 3\
33 & 5 & 0.25 & 0.1 & 0.1 & face-to-face & \[CPrueba40p1\] l\
34 & 5 & 0.25 & 0.1 & 0.3 & face-to-face & \[CPrueba40p1\] m\
35 & 5 & 0.25 & 0.1 & 0.48 & face-to-face & \[CPrueba40p1\] r\
36 & 5 & 0.25 & 0.1 & 0.6989 & face-to-face & \[lastcolconfM5\] 4\
37 & 5 & 0.25 & 0.2 & 0.1 & connecting filament & \[CPrueba40p2\] l\
38 & 5 & 0.25 & 0.2 & 0.3 & face-to-face & \[CPrueba40p2\] m\
39 & 5 & 0.25 & 0.2 & 0.48 & face-to-face & \[CPrueba40p2\] r\
40 & 5 & 0.25 & 0.2 & 0.7467 & face-to-face & \[lastcolconfM5\] 5\
41 & 5 & 0.25 & 0.3 & 0.1 & connecting thin filament & \[CPrueba40p3\] l\
42 & 5 & 0.25 & 0.3 & 0.3 & connecting filament & \[CPrueba40p3\] m\
43 & 5 & 0.25 & 0.3 & 0.3918 & central primary with arms & \[lastcolconfM5\] 6\
44 & 50 & 0.1 & 0.1 & 0.1 & connecting filament & \[CPrueba80p1\] l\
45 & 50 & 0.1 & 0.1 & 0.3 & face-to-face & \[CPrueba80p1\] m\
46 & 50 & 0.1 & 0.1 & 0.48 & face-to-face & \[CPrueba80p1\] r\
47 & 50 & 0.1 & 0.2 & 0.1 & connecting thin filament & \[CPrueba80p2\] l\
48 & 50 & 0.1 & 0.2 & 0.3 & connecting filament & \[CPrueba80p2\] m\
49 & 50 & 0.1 & 0.2 & 0.48 & thin filament & \[CPrueba80p2\] r\
50 & 50 & 0.1 & 0.3 & 0.1 & connecting filament & \[CPrueba80p3\] l\
51 & 50 & 0.1 & 0.3 & 0.3 & thin filament & \[CPrueba80p3\] m\
52 & 50 & 0.25 & 0.1 & 0.1 & face-to-face & \[CPrueba90p1\] l\
53 & 50 & 0.25 & 0.1 & 0.3 & face-to-face & \[CPrueba90p1\] m\
54 & 50 & 0.25 & 0.1 & 0.48 & face-to-face & \[CPrueba90p1\]r\
55 & 50 & 0.25 & 0.2 & 0.1 & connecting filament & \[CPrueba90p2\] l\
56 & 50 & 0.25 & 0.2 & 0.3 & face-to-face & \[CPrueba90p2\] m\
57 & 50 & 0.25 & 0.2 & 0.48 & face-to-face & \[CPrueba90p2\] r\
58 & 50 & 0.25 & 0.3 & 0.1 & connecting filament & \[CPrueba90p3\] l\
59 & 50 & 0.25 & 0.3 & 0.3 & face-to-face & \[CPrueba90p3\] m\
60 & 400 & 0.1 & 0.1 & 0.1 & connecting filament with disk & \[CPrueba0p1\] l\
61 & 400 & 0.1 & 0.1 & 0.3 & connecting filament & \[CPrueba0p1\] m\
62 & 400 & 0.1 & 0.1 & 0.48 & connecting filament & \[CPrueba0p1\] r\
63 & 400 & 0.1 & 0.1 & 0.7 & face-to-face & \[lastcolconfM400\] 1\
64 & 400 & 0.1 & 0.2 & 0.1 & filament & \[CPrueba0p2\] l\
65 & 400 & 0.1 & 0.2 & 0.3 & connecting filament & \[CPrueba0p2\] m\
66 & 400 & 0.1 & 0.2 & 0.48 & filament & \[CPrueba0p2\] r\
67 & 400 & 0.1 & 0.2 & 0.55 & central primary with arms & \[lastcolconfM400\] 2\
68 & 400 & 0.1 & 0.3 & 0.1 & connecting thin filament & \[CPrueba0p3\] l\
69 & 400 & 0.1 & 0.3 & 0.3 & filament & \[CPrueba0p3\] m\
70 & 400 & 0.1 & 0.3 & 0.38 & central primary with arms & \[lastcolconfM400\] 3\
71 & 400 & 0.25 & 0.1 & 0.1 & connecting filament & \[CPrueba50p1\] l\
72 & 400 & 0.25 & 0.1 & 0.3 & connecting filament with disk & \[CPrueba50p1\] m\
73 & 400 & 0.25 & 0.1 & 0.48 & face-to-face & \[CPrueba50p1\] r\
74 & 400 & 0.25 & 0.1 & 0.81 & face-to-face & \[lastcolconfM400\] 4\
75 & 400 & 0.25 & 0.2 & 0.1 & connecting filament with disk & \[CPrueba50p2\] l\
76 & 400 & 0.25 & 0.2 & 0.3 & connecting weak filament & \[CPrueba50p2\] m\
77 & 400 & 0.25 & 0.2 & 0.48 & face-to-face & \[CPrueba50p2\] r\
78 & 400 & 0.25 & 0.2 & 0.55 & face-to-face & \[lastcolconfM400\] 5\
79 & 400 & 0.25 & 0.3 & 0.1 & connecting filament with disk & \[CPrueba50p3\] l\
80 & 400 & 0.25 & 0.3 & 0.3 & connecting weak filament & \[CPrueba50p3\] m\
81 & 400 & 0.25 & 0.3 & 0.385 & central primary with arms & \[lastcolconfM400\] 6\
[lccccccccc]{} 4 & 1 & 0.225 & 5.5e-02 & 4.9e-02 & 6877.1 & 0.11 & 0.28 & 0.12 & 0.22\
5 & 1 & 0.02 & 1.4e-01 & 1.32e-01 & 133.5 & 0.15 & 0.32 & 0.15 & 0.34\
6 & 1 & 0.04 & 8.3e-02 & 1.46e-02 & 1062.3 & 0.13 & 0.30 & 0.07 & 0.72\
15 & 1 & 1.0 & 3.8e-02 & 3.4e-02 & 19004.2 & 0.15 & 0.24 & 0.19 & 0.13\
17 & 1 & 0.04 & 1.3e-01 & 5.6e-03 & 1856.3 & 0.16 & 0.31 & 0.09 & 0.86\
18 & 1 & 0.1 & 1.0e-01 & 1.0e-01 & 2720.3 & 0.12 & 0.31 & 0.12 & 0.31\
19 & 1 & 0.02 & 5.4e-02 & 5.1e-02 & 324.5 & 0.17 & 0.30 & 0.16 & 0.30\
20 & 1 & 0.015 & 1.3e-01 & 1.2e-01 & 333.6 & 0.19 & 0.26 & 0.19 & 0.26\
21 & 1 & 0.025 & 9.8e-02 & 9.8e-02 & 333.8 & 0.17 & 0.29 & 0.17 & 0.29\
25 & 5 & 0.1 & 5.0e-02 & 4.99e-02 & 3206.4 & 7.7 & 0.23 & 6.40 & 0.20\
26 & 5 & 0.15 & 5.3e-02 & 5.25e-02 & 55284.9 & 0.28 & 0.14 & 0.27 & 0.14\
33 & 5 & 0.15 & 2.0e-01 & 1.6e-01 & 2002.2 & 2.24 & 0.50 & 0.50 & 0.58\
34 & 5 & 0.15 & 1.2e-01 & 1.2e-01 & 2848.5 & 5.18 & 0.30 & 5.10 & 0.31\
35 & 5 & 0.15 & 8.4e-02 & 2.2e-02 & 5005.9 & 5.90 & 0.26 & 0.08 & 0.58\
36 & 5 & 0.15 & 5.36e-02 & 5.29e-02 & 86903.3 & 0.14 & 0.23 & 0.07 & 0.22\
38 & 5 & 0.1 & 8.3e-02 & 4.96e-02 & 2502.3 & 5.6 & 0.24 & 0.12 & 0.47\
39 & 5 & 0.1 & 5.7e-02 & 1.04e-02 & 5005.6 & 14.4 & 0.24 & 0.16 & 0.6\
40 & 5 & 0.15 & 4.8e-02 & 4.8e-02 & 99934.4 & 0.40 & 0.20 & 0.17 & 0.22\
45 & 50 & 0.062 & 4.77e-02 & 5.09e-02 & 37642.66 & 0.05 & 0.29 & 0.06 & 0.36\
46 & 50 & 0.125 & 6.74e-02 & 5.35e-02 & 55592.61 & 0.10 & 0.21 & 0.06 & 0.30\
52 & 50 & 0.0625 & 8.19e-02 & 7.87e-02 & 33227.5 & 0.03 & 0.45 & 0.02 & 0.45\
53 & 50 & 0.1 & 7.69e-02 & 7.41e-02 & 53690.4 & 0.05 & 0.31 & 0.04 & 0.28\
54 & 50 & 0.125 & 6.91e-02 & 6.66e-02 & 77490.5 & 0.08 & 0.28 & 0.05 & 0.23\
56 & 50 & 0.05 & 4.63e-02 & 4.38e-02 & 40819.7 & 0.17 & 0.26 & 0.07 & 0.26\
57 & 50 & 0.070 & 4.35e-02 & 4.05e-02 & 61478.6 & 0.14 & 0.28 & 0.11 & 0.21\
59 & 50 & 0.065 & 6.46e-02 & 6.30e-02 & 21099.1 & 0.14 & 0.26 & 0.12 & 0.24\
63 & 400 & 0.25 & 1.28e-02 & 7.17e-05 & 308823.5 & 0.13 & 0.04 & 0.15 & 0.35\
73 & 400 & 0.4 & 2.05e-01 & 2.06e-01 & 259714.3 & 0.2 & 0.2 & 0.15 & 0.23\
74 & 400 & 0.5 & 6.8e-02 & 6.8e-02 & 627298.3 & 0.45 & 0.068 & 0.29 & 0.18\
77 & 400 & 0.1 & 8.78e-02 & 8.65e-02 & 205888.0 & 0.15 & 0.23 & 0.14 & 0.22\
78 & 400 & 0.085 & 1.2e-02 & 4.1e-03 & 124480.4 & 0.17 & 0.03 & 0.23 & 0.23\
[lccccc]{} 8 & 1 & 0.03 & 4.6e-02 & 0.18 & 0.20\
9 & 1 & 0.01 & 1.0e-01 & 0.22 & 0.18\
10 & 1 & 0.02 & 7.8e-02 & 0.17 & 0.22\
11 & 1 & 0.035 & 5.37e-02 & 0.18 & 0.21\
22 & 1 & 0.015 & 1.14e-01 & 0.16 & 0.31\
32 & 5 & 0.015 & 2.15e-02 & 0.15 & 0.26\
43 & 5 & 0.02 & 2.13e-02 & 0.19 & 0.24\
67 & 400 & 0.008 & 9.77e-03 & 0.13 & 0.27\
70 & 400 & 0.005 & 1.20e-02 & 0.18 & 0.23\
81 & 400 & 0.01 & 1.82e-02 & 0.18 & 0.27\
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba60p1\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC6PRadr3p2c_110Ink "fig:"){width="2.2in"} ![\[CPrueba60p1\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC6PRadr3p1c_182Ink "fig:"){width="2.2in"} ![\[CPrueba60p1\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC6PRadr3p3c_400Ink "fig:"){width="2.2in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba60p2\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC6PRadr4p2_043Ink "fig:"){width="2.2in"} ![\[CPrueba60p2\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC6PRadr4p1_105Ink "fig:"){width="2.2in"} ![\[CPrueba60p2\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC6PRadr4p3_109Ink "fig:"){width="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba60p3\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC6PRadr5p2_219Ink "fig:"){width="2.2in"} ![\[CPrueba60p3\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC6PRadr5p1_105Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba70p1\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC7PRadr3p2_043Ink "fig:"){width="2.2in"} ![\[CPrueba70p1\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC7PRadr3p1_050Ink "fig:"){width="2.0in" height="1.75in"} ![\[CPrueba70p1\] Core models with M$_T$= 1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC7PRadr3p3c_082Ink "fig:"){width="2.2in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba70p2\] Core models with M$_T$=1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC7PRadr4p2_050Ink "fig:"){width="2.2in"} ![\[CPrueba70p2\] Core models with M$_T$=1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC7PRadr4p1_080Ink "fig:"){width="2.2in"} ![\[CPrueba70p2\] Core models with M$_T$=1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC7PRadr4p3_100Ink "fig:"){width="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba70p3\] Core models with M$_T$=1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC7PRadr5p2_053Ink "fig:"){width="2.2in"} ![\[CPrueba70p3\] Core models with M$_T$=1 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC7PRadr5p1_094Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba30p1\] Core models with M$_T$=5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC3PRadr3p2_090Ink "fig:"){width="2.2in"} ![\[CPrueba30p1\] Core models with M$_T$=5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC3PRadr3p1_107Ink "fig:"){width="2.2in"} ![\[CPrueba30p1\] Core models with M$_T$=5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC3PRadr3p3_111Ink "fig:"){width="2.2in"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba30p2\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48. This is the last collapsing configuration; therefore, there is a missing panel in Fig.\[lastcolconfM5\].](UC3PRadr4p2_102Ink "fig:"){width="2.2in"} ![\[CPrueba30p2\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48. This is the last collapsing configuration; therefore, there is a missing panel in Fig.\[lastcolconfM5\].](UC3PRadr4p1_103Ink "fig:"){width="2.2in"} ![\[CPrueba30p2\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48. This is the last collapsing configuration; therefore, there is a missing panel in Fig.\[lastcolconfM5\].](UC3PRadr4p3_228Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba30p3\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC3PRadr5p2_070Ink "fig:"){width="2.2in"} ![\[CPrueba30p3\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC3PRadr5p1_113Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba40p1\] Core models with M$_T$=5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC4PRadr3p2_075Ink "fig:"){width="2.2in"} ![\[CPrueba40p1\] Core models with M$_T$=5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC4PRadr3p1_074Ink "fig:"){width="2.2in"} ![\[CPrueba40p1\] Core models with M$_T$=5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC4PRadr3p3_094Ink "fig:"){width="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba40p2\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC4PRadr4p2c_076Ink "fig:"){width="2.2in"} ![\[CPrueba40p2\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC4PRadr4p1c_098Ink "fig:"){width="2.2in"} ![\[CPrueba40p2\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC4PRadr4p3c_095Ink "fig:"){width="2.2in"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba40p3\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (top left) 0.1 (top right) 0.3.](UC4PRadr5p2_072Ink "fig:"){width="2.2in"} ![\[CPrueba40p3\] Core models with M$_T$= 5 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (top left) 0.1 (top right) 0.3.](UC4PRadr5p1_099Ink "fig:"){width="2.2in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba80p1\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC8PRadr3p2_025Ink "fig:"){width="2.2in"} ![\[CPrueba80p1\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC8PRadr3p1_050Ink "fig:"){width="2.2in"} ![\[CPrueba80p1\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC8PRadr3p3_057Ink "fig:"){width="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba80p2\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC8PRadr4p2_009Ink "fig:"){width="2.2in"} ![\[CPrueba80p2\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC8PRadr4p1_052Ink "fig:"){width="2.2in"} ![\[CPrueba80p2\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC8PRadr4p3_262Ink "fig:"){width="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba80p3\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (middle) 0.3.](UC8PRadr5p2p3v2_000Ink "fig:"){width="2.2in"} ![\[CPrueba80p3\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (middle) 0.3.](UC8PRadr5p1v2_070Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba90p1\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC9PRadr3p2p2_006Ink "fig:"){width="2.2in"} ![\[CPrueba90p1\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC9PRadr3p1p2_006Ink "fig:"){width="2.2in"} ![\[CPrueba90p1\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC9PRadr3p3p2_039Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba90p2\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC9PRadr4p2p2_008Ink "fig:"){width="2.2in"} ![\[CPrueba90p2\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC9PRadr4p1p2_003Ink "fig:"){width="2.2in"} ![\[CPrueba90p2\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC9PRadr4p3p2_011Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba90p3\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (middle) 0.3.](UC9PRadr5p2_052Ink "fig:"){width="2.2in"} ![\[CPrueba90p3\] Core models with M$_T$=50 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (middle) 0.3.](UC9PRadr5p1_165Ink "fig:"){width="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba0p1\] Clump models with M$_T$ 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UCPRadr3p2c_019Ink "fig:"){width="2.2in"} ![\[CPrueba0p1\] Clump models with M$_T$ 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UCPRadr3p1c_021Ink "fig:"){width="2.2in"} ![\[CPrueba0p1\] Clump models with M$_T$ 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UCPRadr3p3c_033Ink "fig:"){width="2.2in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba0p2\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UCPRadr4p2_025Ink "fig:"){width="2.2in"} ![\[CPrueba0p2\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UCPRadr4p1_024Ink "fig:"){width="2.2in"} ![\[CPrueba0p2\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UCPRadr4p3_040Ink "fig:"){width="2.2in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba0p3\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3. ](UCPRadr5p2_019Ink "fig:"){width="2.1in" height="1.75in"} ![\[CPrueba0p3\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.1 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3. ](UCPRadr5p1_025Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba50p1\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC5PRadr3p2c_038Ink "fig:"){width="2.2in"} ![\[CPrueba50p1\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC5PRadr3p1c_068Ink "fig:"){width="2.2in"} ![\[CPrueba50p1\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.1; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC5PRadr3p3c_205Ink "fig:"){width="2.2in"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba50p2\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC5PRadr4p2_030Ink "fig:"){width="2.2in"} ![\[CPrueba50p2\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC5PRadr4p1_039Ink "fig:"){width="2.2in"} ![\[CPrueba50p2\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.2; the corresponding $\beta$ are (left) 0.1 (middle) 0.3 (right) 0.48.](UC5PRadr4p3_036Ink "fig:"){width="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[CPrueba50p3\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC5PRadr5p2_016Ink "fig:"){width="2.2in"} ![\[CPrueba50p3\] Clump models with M$_T$= 400 M$_{\odot}$, $a=$0.25 and $\alpha=$0.3; the corresponding $\beta$ are (left) 0.1 (right) 0.3.](UC5PRadr5p1_069Ink "fig:"){width="2.2in"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[lastcolconfM1\] (1) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.758; (2) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.2, $\beta=$0.5344; (3) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3983; (4) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.8871; (5) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.5411; (6) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3983; ](UC6PRadr3p5w_086Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM1\] (1) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.758; (2) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.2, $\beta=$0.5344; (3) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3983; (4) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.8871; (5) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.5411; (6) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3983; ](UC6PRadr4p5w_261Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM1\] (1) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.758; (2) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.2, $\beta=$0.5344; (3) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3983; (4) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.8871; (5) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.5411; (6) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3983; ](UC6PRadr5p1py_243Inkn "fig:"){width="2.2in" height="2.2in"}
![\[lastcolconfM1\] (1) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.758; (2) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.2, $\beta=$0.5344; (3) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3983; (4) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.8871; (5) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.5411; (6) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3983; ](UC7PRadr3p5r_190Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM1\] (1) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.758; (2) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.2, $\beta=$0.5344; (3) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3983; (4) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.8871; (5) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.5411; (6) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3983; ](UC7PRadr4p5r_098Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM1\] (1) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.758; (2) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.2, $\beta=$0.5344; (3) M$_T$=1 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3983; (4) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.8871; (5) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.5411; (6) M$_T$=1 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3983; ](UC7PRadr5p1py_196Inkn "fig:"){width="2.2in" height="2.2in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[lastcolconfM5\] (1) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.6118; (3) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3594; (4) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.6989; (5) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.7467; (6) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3918; ](UC3PRadr3p5t_083Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM5\] (1) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.6118; (3) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3594; (4) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.6989; (5) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.7467; (6) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3918; ](UC3PRadr5p1t_064Inkn "fig:"){width="2.2in" height="2.2in"}
![\[lastcolconfM5\] (1) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.6118; (3) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3594; (4) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.6989; (5) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.7467; (6) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3918; ](UC4PRadr3p5t_070Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM5\] (1) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.6118; (3) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3594; (4) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.6989; (5) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.7467; (6) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3918; ](UC4PRadr4p5w_080Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM5\] (1) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.1, $\beta=$0.6118; (3) M$_T$=5 M$_{\odot}$, $a=$0.1, $\alpha=$0.3, $\beta=$0.3594; (4) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.1, $\beta=$0.6989; (5) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.2, $\beta=$0.7467; (6) M$_T$=5 M$_{\odot}$, $a=$0.25,$\alpha=$0.3, $\beta=$0.3918; ](UC4PRadr5p1w_078Inkn "fig:"){width="2.2in" height="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[lastcolconfM400\] (1) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.1, $\beta=$0.7; (2) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.2, $\beta=$0.55; (3) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.3, $\beta=$0.38; (4) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.1,$\beta=$0.81; (5) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.2,$\beta=$0.55; (6) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.3,$\beta=$0.385; ](UCPRadr3p5q2_057Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM400\] (1) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.1, $\beta=$0.7; (2) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.2, $\beta=$0.55; (3) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.3, $\beta=$0.38; (4) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.1,$\beta=$0.81; (5) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.2,$\beta=$0.55; (6) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.3,$\beta=$0.385; ](UCPRadr4p5q_068Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM400\] (1) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.1, $\beta=$0.7; (2) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.2, $\beta=$0.55; (3) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.3, $\beta=$0.38; (4) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.1,$\beta=$0.81; (5) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.2,$\beta=$0.55; (6) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.3,$\beta=$0.385; ](UCPRadr5p1p_044Inkn "fig:"){width="2.2in" height="2.2in"}
![\[lastcolconfM400\] (1) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.1, $\beta=$0.7; (2) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.2, $\beta=$0.55; (3) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.3, $\beta=$0.38; (4) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.1,$\beta=$0.81; (5) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.2,$\beta=$0.55; (6) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.3,$\beta=$0.385; ](UC5PRadr3p5_082Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM400\] (1) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.1, $\beta=$0.7; (2) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.2, $\beta=$0.55; (3) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.3, $\beta=$0.38; (4) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.1,$\beta=$0.81; (5) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.2,$\beta=$0.55; (6) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.3,$\beta=$0.385; ](UC5PRadr4p5q_041Inkn "fig:"){width="2.2in" height="2.2in"} ![\[lastcolconfM400\] (1) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.1, $\beta=$0.7; (2) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.2, $\beta=$0.55; (3) M$_T$=400 M$_{\odot}$,$a=$0.1,$\alpha=$0.3, $\beta=$0.38; (4) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.1,$\beta=$0.81; (5) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.2,$\beta=$0.55; (6) M$_T$=400 M$_{\odot}$,$a=$0.25,$\alpha=$0.3,$\beta=$0.385; ](UC5PRadr5p1p_046Inkn "fig:"){width="2.2in" height="2.2in"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[lcurve\] Limiting curves for all the models, so that in the region to the right of each curve, the corresponding parent gas structure does not collapse anymore.](AlphavsBetaVer2UpSPunLin){width="3.0in" height="2.5in"}
![\[SepFrags\] Binary separation $s$ in terms of the run number; see Table \[tab:modelos\].](NSepFrags){width="3.0in" height="2.5in"}
![\[MassFrags\] Fragment mass in terms of the run number; see Table \[tab:modelos\].](NMassFrags){width="3.0in" height="2.5in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[AlphavsBetaFrags\] (left) Integral properties for binary fragments; (right) A zoom in of the left panel.](AlphavsBetaFrags "fig:"){width="3.0in" height="2.5in"} ![\[AlphavsBetaFrags\] (left) Integral properties for binary fragments; (right) A zoom in of the left panel.](AlphavsBetaFragsZ "fig:"){width="3.0in" height="2.5in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[MassCenP\] Central primary mass in terms of the run number; see Table \[tab:modelos\].](NMassCenP){width="3.0in" height="2.5in"}
![\[AlphavsBetaCenP\] Integral properties for central primaries.](AlphavsBetaCenP){width="3.0in" height="2.5in"}
[^1]: The letters on the right side of the number of the figure indicate the panel of the mosaic where the configuration is located
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A novel decomposition technique is used to extract the centrality dependence of di-jet properties and yields from azimuthal correlation functions obtained in Au+Au collisions at $\sqrt{s_{_{\rm NN}}}$=200 GeV. The width of the near-side jet shows very little dependence on centrality. In contrast, the away-side jet indicates substantial broadening as well as hints for for a local minimum at $\Delta \phi=\pi$ for central and mid-central events. The yield of jet-pairs (per trigger particle) slowly increases with centrality for both the near- and away-side jets. These observed features are compatible with several recent theoretical predictions of possible modifications of di-jet fragmentation by a strongly interacting medium. Several new experimental approaches, including the study of flavor permutation and higher order multi-particle correlations, that might help to distinguish between different theoretical scenarios are discussed.'
address: 'Department of Chemistry, State University of New York at Stony Brook, Stony Brook, NY 11794-3400, USA'
author:
- 'W. G. Holzmann, N. N. Ajitanand, J. M. Alexander, P. Chung, M. Issah, R. A. Lacey, A. Taranenko, A. Shevel'
title: 'Multiparticle angular correlations: a probe for the sQGP at RHIC'
---
.
Introduction
============
QCD calculations on the lattice indicate a transition from hadronic matter to a deconfined phase of quarks and gluons (QGP) at extremely high temperature and energy density [@Karsch:2002; @Fodor:2001pe]. Heavy ion collisions can deposit large amounts of energy in a small collision volume. Thus, they provide excellent pathways to energy densities and temperatures pertinent to such a phase transition. Indeed, several recent results from Au+Au collisions at the Relativistic Heavy Ion Collider (RHIC) indicate the creation of matter with energy density believed to be sufficiently high for a state of matter with non-hadronic degrees of freedom to be formed [@Adcox:2004mh]. The dynamical evolution of this matter appears to be hydrodynamically driven, implying that the produced medium is strongly interacting [@Lacey:2001va; @Shuryak:2004cy; @Arkadij_ND; @Gyulassy_qgp; @Heinz; @Heinz2]. This matter has been observed to strongly suppress the number of high transverse momentum particles [@ppg003; @ppg014; @star_supp; @ppg023] and to quench the away-side jet in central Au+Au collisions [@Adler:2002ct]. The suppression is commonly ascribed to radiative energy loss of hard scattered partons traversing the high energy density matter prior to the formation of hadrons [@bjorken; @appel1986; @blaizot_mclerran86; @wang2; @gyulassy; @wang].
The influence of parton-medium interactions on jet properties such as the number of jet-associated partner particles per trigger hadron, and jet-topologies, remains an open question of great current interest. Consequently, much recent effort is targeted at understanding how jets couple to the strongly interacting medium produced in energetic Au+Au collision at RHIC [@Ko; @Hwa; @Fries; @Salgado; @stoecker; @stoecker2; @shuryak; @colorwake; @Majumder:2004pt]. At the center of this discussion is the possible creation of jet induced “conical flow" or “mach shocks" analogous to a sonic boom in a fluid [@stoecker; @stoecker2; @shuryak; @colorwake]. If observed, such a phenomena could provide the means for reliable estimates of the speed of sound in the nuclear collision medium produced at RHIC.
In this contribution, we utilize azimuthal angular correlation functions for charged hadrons to examine the influence of medium-effects on jet characteristics and yields. The use of correlation functions provides a novel and effective probe for such effects because they provide simultaneous access to both the (di-) jet signals and harmonic contributions arising from hydrodynamic evolution of the system.
Data Analysis
=============
The analysis presented here uses Au+Au data ($\sqrt{s_{NN}}$=200 GeV) provided by RHIC in the second running period (2001) and recorded by the PHENIX collaboration. The full PHENIX detector setup is described elsewhere [@nim_1]. Charged tracks for this analysis were reconstructed in the two central arms of PHENIX, each of which covers 90 degrees in azimuth. Tracking was performed via a drift chamber and two layers of multi-wire proportional chambers with pad readout (PC1,PC3)[@nim_1]. A combinatorial Hough transform in the track bend plane was used for pattern recognition[@nim_2]. Most conversions, albedo and decays were rejected by requiring a confirmation hit within a 2 $\sigma$ matching window in the PC3. Collision centrality was determined using the beam-beam counters (BBC) and zero degree calorimeters (ZDC) by cutting in the space of BBC versus ZDC analog response [@nim_3].
Following the approach commonly exploited in HBT analyses’, we define the azimuthal angular correlation function as the ratio of a foreground distribution $N_{cor}$, constructed with correlated particle pairs from the same event, and a background distribution $N_{mix}$, obtained by randomly pairing particles from different events within the same multiplicity and vertex classes. $$C\left(\Delta\phi\right) \propto \frac{N_{cor}\left( \Delta\phi
\right)} {N_{mix}\left( \Delta\phi \right)}.$$ Here, $\Delta \phi=(\phi_{1}-\phi_{2})$ is the azimuthal angle difference between particle pairs formed with one hadron from a high-$p_{T}$ “trigger" bin (hereafter labeled A) and another hadron from a lower $p_{T}$ selection (hereafter labeled B). A particular asset of constructing a correlation function is that azimuthal acceptance and detector efficiencies automatically cancel in the ratio of foreground to background distribution. The area normalized correlation function then gives the probability distribution for detecting correlated particle pairs per event within the PHENIX pseudorapidity acceptance ($|\eta|<0.35$).
Figure \[Fi:fig01\] shows the centrality dependence of di-hadron correlation functions obtained from Au+Au collisions with $p_T$ selections of $2.5 < p_T < 4.0$ GeV/c and $1.0 < p_T < 2.5$ GeV/c for trigger and associated hadrons respectively. The correlation function for the most peripheral event sample (cf. Fig. \[Fi:fig01\]f) shows two distinct peaks; a narrow peak at $\Delta \phi=0$ and a somewhat broader peak at $\Delta \phi=\pi$. This asymmetic pattern is compatible with the very well known azimuthal patterns expected for di-hadron correlations resulting from di-jet fragmentation. In contrast to the peripheral events, the correlation functions for more central collisions show less prominent jet-driven asymmetries and appear to be dominated by large anisotropies which can be linked to harmonic correlations resulting from elliptic flow. It is noteworthy that the correlation functions for central - mid-central collisions show a minimum below that expected for harmonic contributions ($\Delta \phi = \pi/2$). Such a shift strongly suggests an away-side jet which is significantly broader than in peripheral collisions.
A particularly important feature of the correlation functions shown in Figure \[Fi:fig01\] is the observation that they can be fully accounted for if one assumes that only two sources of correlations contribute to the measured azimuthal correlation function. That is, a (di-)jet correlations and and harmonic correlations due to elliptic flow [@Adler:2002ct; @Ajitanand:2002qd; @Chiu:2002ma; @Adler:2002tq].
![\[Fi:fig01\]Correlation functions, $C(\Delta\phi)$, for trigger particles in 2.5 GeV/c$<p_{T}^{A}<$4.0 GeV/c and associated particles in 1.0 GeV/c$<p_{T}^{B}<$2.5 GeV/c. The bands show the harmonic contribution within the systematic uncertainty. ](PPG032_Fig1){width="25pc"}
Decomposition of jet and harmonic contributions
===============================================
It follows from the above discussion that a thorough investigation of possible modifications to the (di-) jet pair distribution in Au+Au collisions requires the separation of the jet-signal from the underlying harmonic “background". Reliable procedures for such a decomposition are detailed in Refs. [@Stankus_RP; @Ajit_Methods]. Consequently, only the main steps of the procedure are outlined below.
It can be shown [@Stankus_RP] that the pair correlations from the combination of flow and jet sources are given by $$%
C^{AB} (\Delta \phi ) = a_{o} [C^{AB}_H(\Delta \phi )] + J (\Delta \phi ),
\label{eq1}
%$$ where $C^{AB}_H(\Delta \phi )$ is a harmonic function of effective amplitude v$_{2}$, $$%
C_{H}^{AB} (\Delta \phi )_{ } = [1 + 2 v_{2}
cos 2(\Delta \phi _{ }) ]; \ v_{2}=(v_2^A \times v_2^B).
\label{eq2}$$ and J($\Delta \phi )$ is the (di-)jet function. By rearranging Eq. \[eq1\] one gets $$%
J(\Delta \phi ) = C^{AB}(\Delta \phi ) - a_{o}C^{AB}_{H} (\Delta \phi ).
\label{eq3}
%$$ J($\Delta \phi$) can be extracted if the normalization (a$_{o}$) can be constrained and the harmonic amplitude of the background (v$_{2}$) can be reliably determined. To fix the value of a$_{o}$ we assume that the (di-)jet function has zero yield at the minimum (ZYAM) $\Delta \phi _{min}$. This is equivalent to demanding $a_0C_{H}^{AB}(\Delta{\phi}_{min}) = C^{AB}(\Delta \phi_{min})$. However, no explicit or implicit assumption for the functional form of J($\Delta \phi$) is required. For the results presented here, the minimum is determined by requiring the background to coincide with a functional fit to the correlation function. We obtain $v_2$ by measuring the single particle distributions relative to the reaction plane $\psi_R$. The reaction plane is reconstructed in the beam-beam counters and corrected for dispersion following the procedures outlined in Ref. [@Poskanzer:1998yz]. An important aspect in this regard is the large (pseudo)rapidity separation ($\Delta \left|\eta \right|> 2.75$) between each BBC and the PHENIX central arms. This relatively large $\eta$ separation is believed to substantially reduce possible jet contributions to the determination of $v_2$ [@Adler:2003kt].
Before applying the ZYAM decomposition technique to the analysis of data, we have tested its reliability via a detailed set of Monte Carlo simulations that accounted for the $\phi$ and the $\eta$ acceptance of PHENIX [@Ajit_Methods].
After applying the ZYAM condition to the PHENIX data and utilizing the measured values of $v_2$ ($v_{2}=(v_2^A \times v_2^B$)), the harmonic contribution to the correlation function is given by the solid bands in Fig. \[Fi:fig01\]. The jet-pair distribution is then obtained by building the difference between the correlation function and this harmonic contribution. It can be shown, that the integral of this distribution is proportional to the fraction of jet-correlated particles per event [@Stankus_RP; @Ajit_Methods]. This fraction can be extracted by building the ratio of the sum of $J(\Delta \phi)$ and the sum of $C(\Delta \phi)$ (over all bins in $\Delta \phi$), $$PF = \frac{\sum_i{J(\Delta\phi_i)}}{\sum_i{C(\Delta\phi_i)}}
\label{Eq:JPF_decomp}$$ One readily obtains the efficiency corrected pairs per trigger or conditional yield $CY$ from this fraction via multiplication by the ratio of the average number of detected particle pairs per event $\langle N^{AB}_{d} \rangle$, to the product of the detected singles rates $\langle N^{A}_{d}\rangle$, $\langle N^{B}_{d}\rangle $, followed by a final multiplication with the efficiency corrected singles rate $\langle N^{B}_{eff}\rangle$, for the lower $p_T$ bin B [@Stankus_RP; @Ajit_Methods]. $$%
CY = PF \times \frac{\langle N^{AB}_{d} \rangle}{ \langle N^{A}_{d}\rangle \times \langle N^{B}_{d}\rangle}
\times \langle N^{B}_{eff}\rangle .
\label{Eq:CY_decomp}
%$$
Results
=======
Figure \[Fi:fig02\] shows the centrality dependence of the extracted jet-pair distributions normalized to the number of trigger particles. The most peripheral bin exhibits an azimuthal pattern compatible with regular vacuum fragmentation of (di-)jets. That is, a near-side peak at $\Delta \phi=0$ and an away-side peak at $\Delta \phi=\pi$.
![\[Fi:fig02\]ZYAM-subtracted jet-pair distributions $1/N_{trig}dN/d(\Delta\phi)$. The dashed(solid) histograms show the distributions that would result from subtracting the harmonic contribution after increasing(decreasing) the product $v_2 = v_{2}^{A} \times v_{2}^{B}$ by one interval of the systematic error. The hatched area indicates the systematic error on the background determination (see text).](PPG032_Fig2){width="25pc"}
The jet-pair distributions obtained at other centralities reveal additional striking features. In order to quantify these features, we separate the jet-pair distributions into two regions and assign $\Delta \phi=0 - \Delta \phi_{min}$ and $\Delta \phi_{min} - \Delta \phi=\pi$ to the near- and away-side jets respectively. Both pieces of the jet-pair distribution can then be characterized by their widths (RMS) and their respective per-trigger yields. The resulting values so obtained are plotted in Fig. \[Fi:fig03\] as a function of the number of participants. Similar results for the 0-20% most central d+Au collisions are also included in Fig. \[Fi:fig03\] to facilitate a comparison between the Au+Au and d+Au systems.
Let us first follow the centrality evolution of the near-side jet. The shape of this jet is essentially unchanged across centralities (cf. Fig. \[Fi:fig02\]). However it’s per-trigger yield rises with increasing collision centrality suggesting some degree of modification \[by the medium\] which does not have a strong influence on the near-side jet shape. In contrast to the near-side jet, the shape of the away-side jet-pair distribution broadens significantly as one moves from peripheral events to mid-central events (cf. Fig. \[Fi:fig02\]). In addition, the mid-central distributions (20-40% and 10-20%) indicate an away-side jet distribution that exhibits a local maximum at $\Delta\phi=2\pi/3$ and an apparent minimum at $\Delta\phi=\pi$. These observations are consistent with a picture that involves little or no quenching of jets in peripheral events and strong centrality dependent medium modifications of the away-side jet shape in the mid-central and central events. It is interesting to note that both the near- and away-side per trigger yields (cf. Fig. \[Fi:fig03\]) show a mild rise with increasing centrality, suggesting a possible medium induced modification of both near- and away-side jets. The prominent “hump-backed" shape of the away-side jet is consistent with recent conjectures of a strong coupling between such jets and the high energy density matter that they traverse [@stoecker; @stoecker2; @shuryak; @Salgado].
It is instructive to study the influence of systematic variations in $v_2$ on the jet-pair distribution. The solid (dashed) lines in Fig.\[Fi:fig02\] indicate the conditional yield distributions that is extracted after subtraction of a $v_2$ product that is lowered (raised) by one interval of the systematic error respectively. The systematic error on $v_2$ is dominated by the uncertainty on the reaction plane dispersion. The dotted line shows the jet-pair distribution resulting from a subtraction of a $v_2$ product lowered by twice the systematic uncertainty. In the latter case, the local minimum at $\Delta \phi=\pi$ is no longer significant. However, the away-side peaks for all centralities smaller than $60\%$, remain significantly broader than in the most peripheral event sample.
![(a) Per trigger yields for near- and away-side peaks in the jet pair distribution; and (b) Widths (RMS) of the peaks. The triangles denote results for 0-20% most central d+Au events from a recent PHENIX analysis.\[Fi:fig03\]](PPG032_Fig3){width="25pc"}
Summary of charged hadron results
=================================
To summarize, we have used a novel decomposition technique to extract absolutely normalized jet pair distributions from correlation functions. Our results indicate jet pair distributions for peripheral Au+Au collisions which suggest very little, if any, medium induced modification of jet properties. By contrast, the away-side jet measured in central and mid-central collisions show strong modifications resulting from interactions with the medium. At this juncture, it is very interesting that the observed jet modifications are in qualitative agreement with several theoretical conjectures of how jet characteristics can be modified by a strongly interacting QCD liquid [@Salgado; @stoecker; @stoecker2; @shuryak; @colorwake].
Towards the Future
===================
Future detailed studies of the modification of jet characteristic by the strongly interacting QCD medium will require new measurements as well as enhancements in our analysis techniques. In what follows, we give a few examples which show that such developments are already underway.
Extraction of the jet function via harmonic extinction
------------------------------------------------------
For mid-central collisions, Fig. \[Fi:fig01\] shows pair correlations with strong harmonic components C$_{H}^{AB}(\Delta \phi )$ and relatively weak jet components J($\Delta \phi )$. Only after subtraction of C$_{H}^{AB}(\Delta \phi)$ is the true shape revealed for J($\Delta \phi )$ (cf. Fig.\[Fi:fig02\] ). It would be desirable to select a data set that contained only those correlations due to the jets. In recent work we have developed a new technique which allows data selection such that the harmonic correlations are extinguished ((v$_{2})$ = 0) [@Ajit_Methods]. This technique utilizes $\Delta \phi$ distributions where the trigger particles are selected in a window perpendicular to the reaction plane [@Voloshin:2004]. For these distributions $v_{2}$ can change phase and it is possible to select a cut angle $\phi _{c}$ so that $v_2$ = 0. In Fig. \[fig5\] we demonstrate the extinguishing of harmonic correlations. The filled circles show the inclusive $\Delta \phi$ distribution obtained for a simulation with an input $v_2 \sim 0.16$. The filled squares show the out-of-plane $\Delta \phi$ distribution after the cut angle $\phi _{c}$, is set to the extinction value, ie. $\phi _{c} = \phi _{xt}$ [@Ajit_Methods]. The latter distribution is flat and demonstrates that this technique gives a good method to determine the jet correlations directly from a data set, without the blurring effect from harmonic correlations mediated by the reaction plane. The latter technique provides clear advantages for future jet studies at RHIC.
![\[fig5\]Simulated correlation function for unconstrained particles (filled circles) and for a trigger particle constrained within the cut angle $\phi _{c} = \phi _{xt}$ perpendicular to the reaction plane (filled squares), see text. The results are for a pure harmonic simulation with $v_2 \sim 0.16$.](fig5){width="25pc"}
Flavor permutations
-------------------
The study of jet fragmentation and the influence of possible interactions of the jet with the medium in Au+Au collisions will undoubtedly require investigations which seek to map out the flavor dependence (ie. the dependence on particle species) of these effects. Correlation functions measured with fully identified particles at intermediate and high $p_T$ should serve as a powerful tool to unravel different aspects of the interaction between jets and the medium. Recently, we have begun to measure such correlation functions at RHIC. Figure \[Fi:fig05\] shows preliminary flavor permutation correlation functions for the centrality range 20-40%. The trigger particle $p_T$ bin spans $1.5<p_{T}<2.5$ GeV/c and the assoicated particle has been selected from the range $0.8<p_{T}<1.5$ GeV/c. One can observe strong anisotropies and asymmetries \[in these correlation functions\] reminiscent of those discussed above for charged hadrons. More importantly, one can clearly see that these anisotropies and asymmetries strongly depend on the flavor permutation (ie. the the identity of each particle which is used in the pair correlation measurement). For example, the baryon-baryon correlation function shown in Fig. \[Fi:fig05\] (a) does not appear to exhibit any asymmetries and shows a clear minimum at $\Delta \phi=\pi/2$, ie. the correlation function is essentially harmonic. This suggests little, if any, contributions from (di-) jet sources to the di-baryon correlation functions at these transverse momenta. On the other hand, the meson-meson correlation function shown in Fig. \[Fi:fig05\] (c) is relatively asymmetric, suggesting that jet fragmentation products are predominantly mesons.
![\[Fi:fig05\] Correlation functions for identified trigger particles ($1.5<p_{T}<2.5 GeV/c$) and identified partner particles ($0.8<p_{T}<1.5 GeV/c$). The centrality selection is 20-40%.](permutation){width="25pc"}
These data suggest that a detailed study of correlation functions as a function of flavor permutation holds much promise for providing detailed insights into jet fragmentation and its modification in a strongly interacting medium.
Higher order multiparticle correlations
---------------------------------------
In recent work we have also been pursuing analyses’ which involve higher order multi-particle correlations. Such correlations have been shown to carry additional topological information compared to a two particle correlation analysis [@LauretMulti]. Additional shape information from higher order correlations could potentially provide a powerful tool for distinguishing between different scenarios for jet-modification by the nuclear medium. For example, “mach cone” like structures would potentially manifest themselves in a rather distinctive pattern when correlating more than two particles. Such analyses for identified and unidentified particle multiplets are currently being pursued with vigor.
References {#references .unnumbered}
==========
[<28>]{} F. Karsch, in: [*Lecture Notes in Physics, vol.*]{} [**583**]{} (2002) p.209 Z. Fodor and S.D. Katz, [*JHEP*]{} [**03**]{} (2002) 014 K. Adcox et al., [*Nucl. Phys.*]{} [**A757**]{} (2005) 184 R.A. Lacey, [*Nucl. Phys.*]{} [**A698**]{} (2002) 559 E.V. Shuryak,(2002) [*Preprint*]{} hep-ph/0405066 A. Taranenko, [*Proc. 21st Winter Workshop on Nuclear Dynamics*]{} [*Preprint*]{} nucl-ex/0506019 M. Gyulassy and L. McLerran, [*Nucl. Phys.*]{} [**A750**]{} (2005) 30. U.W. Heinz, (2005) [*Preprint*]{} nucl-th/0504011 U.W. Heinz, (2005) [*Preprint*]{} nucl-th/0407067 K. Adcox [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**88**]{}, 022301 (2002) S.S. Adler [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**91**]{}, 072301 (2003) J. Adams [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**91**]{}, 172302 (2003) S.S. Adler [*et al.*]{}, [*Phys. Rev.*]{} [**C69**]{}, 034910 (2004) C. Adler et al., [*Phys. Rev. Lett.*]{} [**90**]{} 082302 (2003) J.D. Bjorken, (1982) FERMILAB-PUB-82-059-THY D. A. Appel, [*Phys. Rev.*]{} [**D33**]{}, 717 (1986) J. P. Blaizot, [*et al.*]{} [*Phys. Rev.*]{} [**D34**]{}, 2739 (1986) X.N. Wang and M. Gyulassy, [*Phys. Rev. Lett.*]{} [**68**]{}, 1480 (1992); X.N. Wang, [*Phys. Rev.*]{} [**C58**]{}, 2321 (1998) M. Gyulassy et al.,(2003) [*Preprint*]{} nucl-th/0302077. X.-N. Wang,(2004) [*Preprint*]{} nucl-th/0412051
V. Greco [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**90**]{}, 202302 (2003) and [*Phys. Rev.*]{} [**C68**]{}, 034904 (2003) R. C. Hwa and C.B. Yang, [*Phys. Rev.*]{} [**C70**]{} 024905 (2004), and [*J. Phys.*]{} [**G30**]{} (2004) S1117-S1120 R. Fries [*et al.*]{}, [*Preprint*]{} nucl-th/0407102 (2004) N. Armesto et al., (2004)[*Preprint*]{} hep-ph/0411341 H Stöcker, [*Nucl. Phys.*]{} [**A750**]{} (2005) 121 H Stöcker, (2005) [*Preprint*]{} nucl-th/0412022 J. Casalderrey-Solana et al., (2004) [*Preprint*]{} hep-ph/0411315 J. Ruppert and B. Müller, [*Phys. Lett.*]{} [**B618**]{} (2005) 123-130 ([*Preprint*]{} hep-ph/0503158) A. Majumder, [*et al.*]{} [*Preprint*]{} nucl-th/0412061 K. Adcox et al., [*Nucl. Instrum. Meth.*]{} [**A499**]{} (2003) 469 J. Mitchell et al., [*Nucl. Instrum. Meth.*]{} [**A482**]{} (2002) 491 K. Adcox et al., [*Phys. Rev.*]{} [**C69**]{} (2004) 024904
N.N. Ajitanand et al., [*Nucl. Phys.*]{} [**A715**]{} (2003) 765 M. Chiu et al., [*Nucl. Phys.*]{} [**A715**]{} (2003) 761 C. Adler et al., [*Phys. Rev. Lett.*]{} [**90**]{} 032301 (2003) P. Stankus, [*PHENIX Technical Note*]{} [**412**]{} (2005) N.N. Ajitanand et al., (2005) [*Preprint*]{} nucl-ex/0501025, submitted for publication. S.S. Adler et al., [*Phys. Rev. Lett.*]{} [**91**]{} 182301 (2003) S.S. Adler et al., (2004) [*Preprint*]{} nucl-th/0411040 A.M. Poskanzer et al., [*Phys. Rev.*]{} [**C58**]{} (1998) 1671 J. Bielcikova, S. Esumi, K. Filimonov, S. Voloshin and J. P. Wurm, [*Phys. Rev.*]{} [**C69**]{} 021901 (2004) J. Lauret, R. A. Lacey [*et al.*]{}, [*Phys. Lett.*]{} [**B339**]{}, 22 (1994)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The predictability of the atmosphere at short and long time scales, associated with the coupling to the ocean, is explored in a new version of the Modular Arbitrary-Order Ocean-Atmosphere Model (MAOOAM), based on a 2-layer quasi-geostrophic atmosphere and a 1-layer reduced-gravity quasi-geostrophic ocean. This version features a new ocean basin geometry with periodic boundary conditions in the zonal direction. The analysis presented in this paper considers a low-order version of the model with 40 dynamical variables.
First the increase of surface friction (and the associated heat flux) with the ocean can either induce chaos when the aspect ratio between the meridional and zonal directions of the domain of integration is small, or suppress chaos when it is large. This reflects the potentially counter-intuitive role that the ocean can play in the coupled dynamics.
Second, and perhaps more importantly, the emergence of long-term predictability within the atmosphere for specific values of the friction coefficient occurs through intermittent excursions in the vicinity of a (long-period) unstable periodic solution. Once close to this solution the system is predictable for long times, i.e. a few years. The intermittent transition close to this orbit is, however, erratic and probably hard to predict.
This new route to long-term predictability contrasts with the one found in the closed ocean-basin low-order version of MAOOAM, in which the chaotic solution is permanently wandering in the vicinity of an unstable periodic orbit for specific values of the friction coefficient. The model solution is thus at any time influenced by the unstable periodic orbit and inherits from its long-term predictability.
author:
- 'Stéphane Vannitsem[^1]'
- 'Roman Solé-Pomies'
- Lesley De Cruz
bibliography:
- 'maosoam.bib'
title: 'Routes to long-term atmospheric predictability in reduced-order coupled ocean-atmosphere systems – Impact of the ocean basin boundary conditions'
---
Introduction
============
Nowadays important efforts are devoted to the development of forecasting systems for long lead times from seasons to decades. Just like numerical weather prediction (NWP) systems, these forecasts are subject to the property of sensitivity to initial conditions by which a small initial error will rapidly amplify until it reaches a level at which no useful information can be extracted anymore. This level is usually fixed at the point where the mean square error score reaches twice the variance of the climatology of the observable [@Lorenz1982; @Dalcher1987; @Chen1989; @Boer1994; @Savijarvi1995]. This specific level marks the time at which, on average, no correlation exists between the forecast and the observations anymore. So the way the mean square error reaches this level as a function of time is a key feature revealing the potential of extended-range forecasts.
This feature was already recognized by @Lorenz1982 who estimated the limit of predictability of weather systems at mid-latitudes to 10-15 days using a state-of-the-art weather forecasting system (at that time). Similar estimates have been obtained later in more sophisticated NWP models which only describe the evolution of the atmosphere [@Dalcher1987; @Chen1989; @Kalnay2003]. Nowadays, the tendency is to develop Earth System Dynamic (ESD) models encompassing different components of the climate system. The hope is that the mutual interaction between these different components will allow to extend the predictability limit of the atmosphere beyond the 10-15 days barrier.
Interactions between the different components of the climate system are essential to its behaviour on multiple time scales. A particular example is the dynamics of the El Niño–Southern Oscillation (ENSO) phenomenon for which the ocean interacts with the atmosphere in the tropical Pacific, leading to the development of a coupled dynamics with a low-frequency variability developing in the atmosphere on time scales of a few years, e.g. [@Philander1990; @Stuecker2013]. At mid-latitudes, such ocean-atmosphere interactions are less pronounced and the origin of the low-frequency variability is still a matter of debate [@Robertson2000; @Marshall2001]. Some coupled modes have however been found in coupled ocean-atmosphere models or in observational data [@Czaja2002; @VanderAvoird2002; @Kravtsov2007; @Feliks2007; @Minobe2008; @Brachet2012; @Lheveder2015; @Vannitsem2017a]. This suggests that low-frequency variability developing in the atmosphere can be associated with the coupling to the ocean, and could be a source of long-term predictability.
Recently a simple reduced-order multi-scale system composed of a reduced-gravity quasi/geostrophic ocean layer and a two-layer baroclinic atmosphere has been built to sketch the dynamics of the coupled ocean-atmosphere system over mid-latitude ocean basins. The fields described by these equations were developed in Fourier series and severely truncated to a low order, yielding a system of 36 ordinary differential equations describing the dominant modes of the dynamics [@Vannitsem2015a]. This system displays a low-frequency variability (LFV) in the atmosphere provided the friction coefficient between the ocean and the atmosphere is sufficiently high [@Vannitsem2015a]. The emergence of this low-frequency variability is at the origin of the long-term predictability of some observable within the atmosphere as illustrated in [@Vannitsem2017b]. It is however important to notice here that in the context of this coupled model, the LFV does not develop when the friction coupling (which also controls the heat fluxes between the two components) is small. This means that coupling two components, slow and fast, does not imply that the fast system will inherit the LFV associated with the presence of the slow component. In the model version of @Vannitsem2015a, called VDDG in the following, the development of the LFV and the long-term predictability in the atmosphere is the outcome of a bifurcation leading to a drastic qualitative change of dynamics. The solution of the system experiences a catastrophic change when the friction coefficient is increased, leading to an attractor developing close to an unstable periodic orbit. We will refer to this transition as a [*chaos-to-chaos transition*]{}. This new solution wanders chaotically around the unstable periodic orbit, a type of dynamics that has already been isolated in much simpler systems by e.g. @Tel2006 [@Wernecke2017]. This chaos-to-chaos transition has been found in autonomous and non-autonomous versions of the model [@Vannitsem2015b], and in all the experiments that were done with this model configuration [@Vannitsem2015a; @Vannitsem2015b; @Vannitsem2017b; @Demaeyer2018].
One can now wonder whether the presence of this transition is sensitive to changes in the model, such as an increase in the model resolution or a change in the geometry of the system. This has motivated the development of a flexible, generalised version of the model called the Modular Arbitrary-Order Ocean-Atmosphere Model (MAOOAM) [@DeCruz2016]. Using MAOOAM, some preliminary work has been done in answering the first question, by looking at the variability of the solutions when increasing the number of modes. It was shown that the LFV can weaken at intermediate resolutions, but recovers as the number of modes is further increased.
This work aims to address the second question by modifying the basin boundary conditions of the ocean. Instead of using a fully closed ocean basin, periodic boundary conditions are imposed in the zonal direction of the ocean component. In this situation, both the flow in the ocean and in the atmosphere develop in a channel.
This new model configuration can be viewed as a very crude representation of the Southern Ocean encircling Antarctica. However, care should be taken when using the results presented in the following sections to interpret the dynamical behaviour in the Southern Ocean. Indeed, the model does not feature a meridional gradient of density in the ocean component, which is composed of a single homogeneous layer. This gradient is known to be essential in the development of the Antarctic Circumpolar Current (ACC), [@Vallis2006]. It is nonetheless interesting to investigate the robustness of the findings obtained in the context of a closed ocean basin in this new model version. This analysis can provide new ideas in the analysis of the dynamics of a more realistic Southern Ocean configuration.
In Section 2, the new model version is described. The Lyapunov properties characterizing the short-term evolution and the long-term saturation of the error within this model are then discussed in Section 3. The key conclusions are provided in Section 4.
The new version of the model
============================
Dynamical equations {#sec:dyn}
-------------------
The dynamical equations of the model have already been described in @Vannitsem2015b [@DeCruz2016]. These are briefly repeated here for completeness.
### Equations of motion for the atmosphere and the ocean {#ssec:atmos}
The atmospheric model is based on the vorticity equations of a two-layer, quasi-geostrophic flow defined on a $\beta$-plane. The equations in pressure coordinates are $$\begin{aligned}
\frac{\partial}{\partial t} \left( \nabla^2 \psi^1_a \right) + J(\psi^1_a, \nabla^2 \psi^1_a) + \beta \frac{\partial \psi^1_a}{\partial x}
& = & -k'_d \nabla^2 (\psi^1_a-\psi^3_a) + \frac{f_0}{\Delta p} \omega, \nonumber \\
\frac{\partial}{\partial t} \left( \nabla^2 \psi^3_a \right) + J(\psi^3_a, \nabla^2 \psi^3_a) + \beta \frac{\partial \psi^3_a}{\partial x}
& = & +k'_d \nabla^2 (\psi^1_a-\psi^3_a) - \frac{f_0}{\Delta p} \omega \nonumber \\
& & - k_d \nabla^2 (\psi^3_a-\psi_o);
\label{eq:atmos}\end{aligned}$$ here $\psi^1_a$ and $\psi^3_a$ are the streamfunction fields at 250 and 750 hPa, respectively, $\omega = dp/dt$ is the vertical velocity, $f_0$ is the Coriolis parameter at latitude $\phi_0$, and $\beta = df/dy$ the meridional gradient of $f$ at latitude $\phi_0$.
The coefficients $k_d$ and $k'_d$ multiply the surface friction term and the internal friction between the layers, respectively, while $\Delta p = 500$ hPa is the pressure difference between the two atmospheric layers. An additional term has been introduced in this system in order to account for the presence of a surface boundary velocity of the oceanic flow defined by $\psi_o$ whose evolution is based on the reduced-gravity, quasi-geostrophic shallow-water model on a $\beta$-plane : $$\label{eq:QGSW}
\frac{\partial}{\partial t} \left( \nabla^2 \psi_o - \frac{\psi_o}{L_R^2} \right) + J(\psi_o, \nabla^2 \psi_o) + \beta \frac{\partial \psi_o}{\partial x}
= -r \nabla^2 \psi_o + \frac{{\mathrm{curl}}_z \tau}{\rho h}.$$ describing the dynamics within the model ocean’s upper, active layer, where $\rho$ is the density of water of the upper layer, $h$ the depth of this layer, $L_R$ the reduced Rossby deformation radius, $r$ a friction coefficient at the bottom of the active layer, and ${\mathrm{curl}}_z \tau$ is the vertical component of the curl of the wind stress that will be defined later.
### Ocean temperature equation {#ssec:ocean_temps}
We assume that temperature is a passive scalar transported by the ocean currents, but the oceanic temperature field displays strong interactions with the atmospheric temperature through radiative and heat exchanges. Under these assumptions, the evolution equation for the ocean temperature is
$$\label{eq:heat_oc}
\gamma_o ( \frac{\partial T_o}{\partial t} + J(\psi_o, T_o)) = -\lambda (T_o-T_a) + E_R$$
with $$\label{eq:fluxes_oc}
E_R = -\sigma_B T_o^4 + \epsilon_a \sigma_B T_a^4 + R_o.$$
In Equations and above, $E_R$ is the net radiative flux at the ocean surface (positive into the ocean), $R_o$ is the shortwave radiation entering the ocean, $\epsilon_a$ the emissivity of the atmosphere, $\sigma_B$ the Stefan-Boltzman constant, $\gamma_o$ the heat capacity of the ocean, and $\lambda$ is the heat transfer coefficient between the ocean and the atmosphere that combines both the latent and sensible heat fluxes. It is assumed that this combined heat transfer is proportional to the temperature difference between the atmosphere and the ocean.
### Atmospheric temperature equation {#ssec:atmos_temps}
The thermodynamic equation for the atmosphere is written as,
$$\label{eq:heat_atm}
\gamma_a ( \frac{\partial T_a}{\partial t} + J(\psi_a, T_a) -\sigma \omega \frac{p}{R}) = -\lambda (T_a-T_o) + E_{a,R}$$
with $$\label{eq:fluxes_atm}
E_{a,R} = \epsilon_a \sigma_B T_o^4 - 2 \epsilon_a \sigma_B T_a^4 + R_a.$$
In these two equations, $R$ is the gas constant, and $$\sigma = - \frac{R}{p} \Big(\frac{\partial T_a}{\partial p}- \frac{1}{\rho_a c_p}\Big)$$ is the static stability, with $p$ the pressure, $\rho_a$ the air density, and $c_p$ the specific heat at constant pressure; here $\sigma$ is taken to be constant. $R_a(t)$ represents the portion of the short-wave radiative input from the sun directly captured by the atmosphere.
Note also that, thanks to the hydrostatic relation in pressure coordinates and to the ideal gas relation $p=\rho_a R T_a$, the atmospheric temperature $T_a$ can be expressed as $T_a = - (p/R) f_0 (\partial \psi_a/\partial p)$. This expression for $T_a$ can then be used to combine Equations and , as done when deducing the quasi-geostrophic potential vorticity equation [e.g. @Vallis2006].
Domain of integration and parameter values {#sec:maosoamdescription}
------------------------------------------
In the original model version of MAOOAM, the domain of integration is rectangular, with a closed ocean basin and a channel flow for the atmosphere [@Vannitsem2015a; @DeCruz2016]. Free-slip boundary conditions were chosen along the meridional and longitudinal boundaries of the ocean basin, while these are free-slip in the meridional direction and periodic in the zonal direction for the atmosphere. In the new version of the model, the main modification is that we now impose periodic boundary conditions in the zonal direction in the ocean. This model now represents a channel flow for both the ocean and the atmosphere.
These boundary conditions are satisfied by using the basis functions that were previously used for the atmosphere to expand both the atmosphere and the ocean fields. In other words, with the proper normalization, the basis functions for all fields must be of the following form, following the nomenclature of @Cehelsky1987: $$\begin{aligned}
&F^A_{P} (x', y') = \sqrt{2}\, \cos(P y')\label{eqn:FA}\\
&F^K_{M,P} (x', y') = 2\cos(M nx')\, \sin(P y')\label{eqn:FK}\\
&F^L_{H,P} (x', y') = 2\sin(H nx')\, \sin(P y'),\label{eqn:FL}\end{aligned}$$ with $(P, M, H) \in \mathbb{N}^3$. The set of modes that will be used in the present study are such that $P=1, 2$, $M=1, 2$ and $H=1, 2$, implying that 10 modes will be used for the four dynamical fields, $\psi_a=\sum_{i=1}^{10} \psi_{a,i} F_i$, $\theta_a=\sum_{i=1}^{10} \theta_{a,i} F_i$, $\psi_o=\sum_{i=1}^{10} \psi_{o,i} F_i$ and $T_o=\sum_{i=1}^{10} T_{o,i} F_i$, where $F_i$ are simplified notations for the set of modes used. Note that the first mode is $ F_1=\sqrt{2}\, \cos(y')$ whose coefficients are denoted as $\psi_{a,1}$, $\psi_{o,1}$ $T_{o,1}$. These variables will be analyzed in detail in the next section.
Let us now estimate the parameters to some realistic values [@Vannitsem2015b]. Assuming that the wind stress follows the linear relation, $(\tau_x, \tau_y)=C (u-U,v-V)$ – where $(u = -\partial \psi_a/\partial y, v = \partial \psi_a/\partial x)$ are the horizontal components of the geostrophic wind, respectively, and $(U, V)$ the corresponding components of the geostrophic currents in the ocean – one gets, $$\frac{\mathrm{curl}_z \tau}{\rho h} = \frac{C}{\rho h} \nabla^2 (\psi^3_a-\psi_o).
\label{eq:stress}$$ where $C$=$\rho_a C_D |\vec{v}-\vec{V}|$ with $C_D$ the drag coefficient and $|\vec{v}|$, the norm of the velocity. The coefficient $d = C/(\rho_o h)$ characterizes the strength of the mechanical coupling between the ocean and the atmosphere.
Similarly, one can use the Ekman layer theory to relate the coefficient $k_d$ in Equation \[eq:atmos\] to the friction coefficient $C$ in pressure coordinates, $$k_d=\frac{g C}{\Delta p} \, [s^{-1}]$$ and one assumes that $k'_d$=$k_d$ as in [@Charney1980].
The parameter $\lambda$ in Equation \[eq:heat\_oc\] can also be related to the surface friction coefficient as discussed in @Houghton1986, $$\lambda= 1004 \, C \, [W m^{-2} K^{-1}]
\label{lambda}$$
The radiative input is decomposed in two different terms, $R_o=R_{o,0}+\delta R_{o}$, the first one being a constant value in space and the second one a meridionally dependent term $\delta R_{o} = C_o F_1$.
The parameter values used in the present analysis are given in Table \[tab:par\].
Parameter (unit) Value Parameter (unit) Value
---------------------------- ------------------------ --------------------------------------- ------------------------
$L_y = \pi L$ (km) $5.0 \times 10^3$ $\gamma_\text{o}$ (Jm$^{-2}$K$^{-1}$) $4 \times 10^6 \, h$
$f_0$ (s$^{-1}$) $1.195 \times 10^{-4}$ $C_\text{o}$ (Wm$^{-2}$) Variable
$n = 2 L_y / L_x$ Variable $T_\text{o}^-2$ (K) $285$
$R_\text{E}$ (km) $6370$ $\gamma_\text{a}$ (Jm$^{-2}$K$^{-1}$) $1.0 \times 10^7$
$\phi_0$ $-0.3056 \pi$ $C_\text{a}$ (Wm$^{-2}$) $C_\text{o}/3$
$g^\prime$ $3.1 \times 10^{-2}$ $\epsilon_\text{a}$ $0.76$
$r$ (s$^{-1}$) $1.0 \times 10^{-8}$ $\beta$ (m$^{-1}$s$^{-1}$) $1.62 \times 10^{-11}$
$h$ (m) Variable $T_\text{a}^0$ (K) $270$
$d$ (s$^{-1}$) $C/(\rho_o h)$ $\lambda$ (Wm$^{-2}$K$^{-1}$) $1004 \, C$
$k_d$ (s$^{-1}$) $\frac{g C}{\Delta p}$ $R$ (Jkg$^{-1}$K$^{-1}$) $287$
$k_d^\prime $ (s$^{-1}$) $\frac{g C}{\Delta p}$ $\sigma$ (J kg$^{-1}$ Pa$^{-2}$) $2.16 \times 10^{-6}$
$C$ (kg m$^{-2}$ s$^{-1}$) Variable
: List of parameters of the model\[tab:par\]
Four important parameters will be modified in the current investigation, $C$, $n$, $h$ and $C_o$.
Typical solutions of the model {#solution}
------------------------------
Let us first briefly qualitatively analyze the solutions generated by the coupled model. Time series of $\theta_{a,1}$ and $T_{o,1}$ for $C=0.016$ kg m$^{-2}$ s$^{-1}$, $n=1.7$, $C_o=350$ W m$^{-2}$ and $h=1000$ m, are displayed in Fig. \[Figseries\] for about 100,000 days. A first remarkable result is the difference of typical time scale of variability. For the first mode of temperature in the ocean, a low-frequency variability is visible on time scales of the order of 10,000 days; while for the first temperature atmospheric mode, $\theta_{a,1}$, a high frequency variability on a time scale of days is present together with a low-frequency variability coherent with the ocean temperature evolution. This evolution suggests that the dynamics within the atmosphere is influenced by the ocean, and should therefore lead to modifications of the predictability properties of the atmosphere.
![Time series of $\theta_{a,1}$ and $T_{o,1}$ for $C=0.016$ kg m$^{-2}$ s$^{-1}$, $n=1.7$, $C_o=350$ W m$^{-2}$ and $h=1000$ m. Nondimensional units are used for the variables.[]{data-label="Figseries"}](Vannitsem_Fig0){width="90.00000%"}
In order to further clarify the behaviour of the solutions of the system in this range of values, several videos have been prepared and are provided as supplementary material (videos S1-S7). These videos display the solutions generated by the model for $h=1000$ m and two values of $n=1.5$ and $1.7$, and several values of $C$. Six different panels are displayed in each video, the geopotential height difference between two locations in the spatial domain ($x'=0$, $y'=3\pi/4$ and $y'=\pi/4$), a three-dimensional projection of the attractor of the system for the variables ($\psi_{a,1}$, $\psi_{o,1}$ $T_{o,1}$), and four panels representing the solution in the spatial domain. The spatial view is centered at the South pole.
Let us focus for the moment on the solutions obtained with $n=1.7$. For $C=0.01$ kg m$^{-2}$ s$^{-1}$ (video S1), the temporal evolution of the geopotential height displays an erratic behaviour without any visible low-frequency variability. This erratic behaviour is also visible in the three-dimensional plot. The solutions in space display a rapid erratic dynamics for the atmospheric temperature (proportional to the baroclinic streamfunction) and barotropic atmospheric streamfunction, while a much slower evolution is seen for the ocean fields. Interestingly a slow counterclockwise motion is found for the ocean streamfunction, with a westward Rossby wave-type of motion. Overall the dynamics of the atmosphere seems to be well independent of the dynamics of the ocean.
Let us now increase the friction coefficient to $C=0.016$ kg m$^{-2}$ s$^{-1}$ (video S2). In this case, the dynamics displays an intermittent behaviour with periods of strong variability and periods during which the dynamics seems frozen. It is particularly spectacular for the atmospheric streamfunction which seems to be frozen for a while from time to time. In the three-dimensional projection of the attractor this dynamics is marked by a trajectory (the red dot) exploring high values of $\psi_{o,1}$ and small values of $\theta_{o,1}$. The trajectory seems to converge to a domain of the state space associated with a purely periodic motion which is visible in the counterclockwise rotation of the spatial ocean streamfunction field. For even larger values of $C$, a similar picture is found as illustrated in the videos S3 and S4 for $C=0.02$ and $C=0.027$ kg m$^{-2}$ s$^{-1}$, respectively. In the latter case, the almost periodic motion is clearly visible alternating with more erratic behaviour, in particular in the geopotential height. This almost periodic motion has quite a long period of about 10 years for $C=0.027$ kg m$^{-2}$ s$^{-1}$.
These videos illustrate the important role of the ocean in the development of low-frequency variability within the atmosphere. A first interesting aspect is the fact that the atmosphere does not “feel” the ocean if the friction is not large enough (also related to the heat fluxes between the two components through Equation (\[lambda\])). When a sufficiently high value of the friction coefficient $C$ is reached, the solution experiences intermittent excursions toward a situation for which the dynamics of the atmosphere and the ocean are coherent with each other. These different types of behaviour for small and large values of $C$ should play a role in the predictability of the atmosphere. This question will now be addressed by investigating two quantities, the Lyapunov exponents and the error dynamics.
Predictability properties of the model
======================================
As illustrated in section \[solution\], the solutions of the model can display erratic types of behaviour reminiscent of the dynamics of chaotic deterministic dynamical systems. The predictability of such systems is limited in time as a small error in the initial conditions will grow in time until it reaches a level at which the forecast cannot be distinguished from a random draw of a solution on the attractor of the system (a draw from the long-term climatology of the system). This property of sensitivity to initial conditions can be evaluated by computing the Lyapunov exponents and/or by computing the evolution of the error during the entire period of the forecasts. The first approach assumes infinitesimal initial errors and is commonly used to characterize the nature of the dynamics such as stationary, periodic or chaotic solutions. The second approach does not assume that the initial errors are infinitesimal, which is of course more appropriate when dealing with realistic forecasting problems. Both will be used here, first to characterize the nature of the solutions, and second to evaluate the long-term predictability of the flow.
Lyapunov instability of the model
---------------------------------
Let us first briefly define these quantities and then look at the exponents obtained in the present model. More details can be found in e.g. @Kuptsov2012.
### Lyapunov exponents
The evolution laws of a dynamical system like the ones presented in Section \[sec:dyn\] can be written in the compact form $$\frac{d\vec{x}}{dt} = {\vec f}(\vec x, \lambda)
\label{equat}$$ where $\vec x$ is a vector containing the set of relevant variables $\vec x$ = $( x_1, ..., x_n)$. This system of equations is then integrated in time starting from an initial state, $\vec x (t_0)=\vec x_0$. As the real state $\vec x_0$ is never known with infinite precision in practice, a small error, $\delta \vec x_0$, will affect the future forecast. This perturbed initial state generates a new trajectory in phase space. The time-dependent error vector, that is the displacement vector between the reference trajectory and the perturbed one at a given time, is denoted $\delta \vec x (t)$. Provided the initial error is small, its evolution is described by a linearized system of equations, $$\frac{d\delta \vec x}{dt} = \frac{\partial \vec f}{\partial \vec x}_{\vert \vec x(t)}
\delta \vec x
\label {linear}$$ whose formal solution is, $$\delta \vec x (t) = {\bf M}(t,\vec{x}(t_0)) \delta \vec x (t_0)$$ where the matrix $\bf{M}$ is referred as the resolvent matrix. In the ergodic theory of deterministic dynamical systems, the double limit of infinitely small initial errors and infinitely long times, is usually taken, e.g. @Eckmann1985. In this limit the divergence of initially close states is determined by the logarithm of the eigenvalues of the matrix $(\bf{M}^T \bf{M})^{2(t-t_0)}$ that are referred to as the Lyapunov exponents. The full set of Lyapunov exponents of a system is called the Lyapunov spectrum, which is usually represented in decreasing order. Positive Lyapunov exponents indicate the presence of a chaotic dynamics, hence displaying sensitivity to initial conditions.
Several numerical techniques have been developed to evaluate these Lyapunov exponents [@Parker1989; @Kuptsov2012]. One of the most popular methods consists in following the evolution of a set of orthonormal vectors chosen initially at random in the tangent space of the trajectory $\vec x(t)$. This basis is regularly orthonormalized using the standard Gram-Schmidt method to avoid the alignment of all the vectors along the unstable direction associated to the largest Lyapunov exponent, and the amplification along these vectors can then be computed. The logarithm of these amplifications are then computed and rescaled by the time $t-t_0$, which provides the set of Lyapunov exponents. This is the approach adopted here.
### Results for MAOOAM with a channel ocean
Figure \[Figlyap1\] displays the dependence of the first Lyapunov exponent as a function of the friction coefficient $C$ for $C_o=350$ and $C_o=250$ W m$^{-2}$ with $h=1000$ m. Different values of the aspect ratio, $n$, are considered in the different panels. Let us start with small values of the aspect ratio in panel (a) for $C_o=350$ W m$^{-2}$. For very small values of $n$, the solutions found are periodic, leading to a first Lyapunov exponent equal to 0. When $n$ is increased, the solutions become chaotic for an intermediate set of values of the friction coefficient, $C$. This indicates that the coupling between the ocean and the atmosphere through surface friction can induce a chaotic dynamics, a regime referred to as [*ocean-induced chaotic dynamics*]{}. For even larger values of $n$ (panel b), the picture is different, with a decrease of the value of the dominant Lyapunov exponents as a function of the friction coefficient, suggesting a stabilization of the flow through friction. This feature is similar to the dependence found for the original VDDG model with $n=1.5$ [@Vannitsem2017b]. This regime can be referred to as a [*ocean-tempered chaotic dynamics*]{}. A similar picture is found when the solar input is reduced to $C_o=250$ W m$^{-2}$ as illustrated in panels (c) and (d).
When the depth of the ocean layer is reduced to $h=100$ m, shown in Figure \[Figlyap2\], a qualitatively similar picture is found, but with a more complicated dependency structure of the friction coefficient $C$. For instance with $n=1.5$, alternating windows of chaotic and periodic solutions are present, which was not visible for $h=1000$ m at panels (a) and (c) in Figure \[Figlyap1\].
In summary, the contrasting behaviour as a function of $n$ reveals a complicated dependence of the dynamics of the coupled system with respect to the size of the domain (measured by $n$) that can lead to ocean-induced chaos or on the contrary to ocean-tempered chaos when friction is increased. This remarkable result reveals the important potential role played by the ocean, and the nontrivial dependence of this role on the resolved dynamics in the ocean. Indeed, increasing the aspect ratio can also be interpreted as filtering out the large-scale zonal waves in favour of an enhanced description of smaller-scale phenomena, shown to be important for the ocean [@DeCruz2018]. It is however not clear yet whether this result is robust when the number of modes is explicitly increased. This question will be addressed in the future.
![First Lyapunov exponent as a function of the friction coefficient $C$ \[kg m$^{-2}$ s$^{-1}$\] for (a) different values of the aspect ratio $n$ and for $C_o=350$ W m$^{-2}$ and $h=1000$ m; (b) as in (a) but with larger values of $n$; (c) as in (a) but for $C_o=250$ W m$^{-2}$; and (d) as in (c) but for $C_o=250$ W m$^{-2}$.[]{data-label="Figlyap1"}](Vannitsem_Fig1){width="100.00000%"}
![First Lyapunov exponent as a function of the friction coefficient $C$ \[kg m$^{-2}$ s$^{-1}$\] for (a) different values of the aspect ratio $n$ and for $C_o=350$ W m$^{-2}$ and $h=100$ m; (b) as in (a) but with larger values of $n$; (c) as in (a) but for $C_o=310$ W m$^{-2}$; and (d) as in (c) but for $C_o=310$ W m$^{-2}$.[]{data-label="Figlyap2"}](Vannitsem_Fig2){width="100.00000%"}
Error dynamics
--------------
### Definition of the error
In the previous section, the emphasis was placed on the analysis of the Lyapunov exponents characterizing the short term evolution of small initial errors. This analysis, although very useful when looking at the emergence of chaos, should however be complemented by the analysis of the error dynamics when interested in the long-term predictability of the system.
Let us consider a solution of the system in phase space, $\vec{x}(t_0)=\vec{x}_0$, at time $t_0$. Observations of this system are affected by finite-amplitude initial errors that can be for simplicity be considered as a Gaussian white noise, $\vec{\epsilon}$, and the observed state is then, $\vec{x}'(t_0)=\vec{x}_0+\vec{\epsilon}$. One can now measure the error evolution starting from these two initial conditions as $$\vec{E}(t)=\vec{x}'(t)-\vec{x}(t)$$ where $\vec{x}'(t)$ and $\vec{x}(t)$ are the two trajectories starting from the two initial conditions $\vec{x}'(t_0)$ and $\vec{x}(t_0)$ of the perturbed and unperturbed trajectories. Since the amplification of this error fluctuates on the inhomogeneous attractor of the system, an average over the attractor is necessary in order to obtain properties that are independent of the initial state. We do not use the classical norm ($L^2$ norm) but rather the logarithmic norm [@Nicolis1995], $$\langle \ln E_t^2 \rangle = \int \mathrm{d}\vec{\epsilon}_0 \rho_{\epsilon} (\vec{\epsilon}_0) \int \mathrm{d}\vec{x}_0 \rho_{S} (\vec{x}_0) \ln \left[ (\vec{x}'(t) - \vec{x}(t)) \cdot (\vec{x}'(t)) - \vec{x}(t)) \right]
\label{expnorm0}$$ where $\rho_{\epsilon}(\vec{\epsilon}_0) \mathrm{d}\vec{\epsilon}_0$ and $\mathrm{d}\vec{x}_0 \rho_{S} (\vec{x}_0)$, are the probability measure of the initial errors and of the initial conditions on the attractor of the system. This specific norm is chosen on the one hand because it is closely related to the definition of the Lyapunov exponents, and on the other hand because it considerably reduces the fluctuations associated with the finite number of realizations used to compute the mean error. The amplitude of the perturbations $\vec{\epsilon}$ is taken sufficiently small in order to get information of the full error growth evolution, even in the short term regime for which the error evolution follows a linearized system of equations, used to define the Lyapunov exponents. The exponential error dynamics expected in this linearized evolution regime translates, after a short transient evolution, to a linear error amplification in the logarithmic norm defined above. As the error (\[expnorm0\]) reaches a substantial amplitude, the effect of the nonlinear terms on the dynamics cannot be neglected anymore and the rate of amplification of the logarithm of the error starts to decrease, and for long lead times, saturates due to the finite size of the system’s attractor. This evolution is discussed in detail in [@Nicolis1995; @Vannitsem1994; @Vannitsem2016].
As we are interested in the long-term predictability of the atmosphere in the coupled system, the focus is placed on how the error defined in (\[expnorm0\]) saturates for a long lead time. Once this saturation is reached, no predictability is left in the system anymore, see e.g. [@Vannitsem2017b]. In other words, any random state taken on the attractor of the system will provide a skill score comparable to the one obtained with a set of initialized forecasts for long times (when reaching the saturation level).
In [@Vannitsem2016; @Vannitsem2017b] the dynamics of the error is analyzed for the VDDG model version and it is shown that provided the friction is sufficiently high, the error will continue to grow for very long times (up to 100 years) for certain atmospheric variables. As discussed in the latter references, this remarkable result is associated with the development of the low-frequency variability (LFV) into the coupled ocean-atmosphere system through a complex sequence of bifurcations when $C_o$ is increased, for a sufficiently large value of $C$. Beyond this sequence of bifurcations the attractor of the system develops around an unstable periodic orbit with a very long period. This orbit constitutes the backbone of the attractor of the system and controls the long term evolution of the error. Let us now investigate how the error behaves in the new model version discussed here.
### Error evolution in the version of MAOOAM with a channel ocean
Equation \[expnorm0\] is now computed for the version of the model with 40 variables presented in section \[sec:maosoamdescription\], with 1000 realizations starting from different initial states on the attractor of the system. Figure \[Figerr1\] displays the error evolution for the four fields of the model, namely the barotropic and baroclinic streamfunctions of the atmosphere, the ocean temperature and the ocean streamfunction, for the parameters $h=1000$ m, $C_o=350$ W m$^{-2}$, $n=1.7$. The different curves in each panel correspond to different values of the friction coefficient $C$.
For a value of the friction coefficient, $C=0.005$ kg m$^{-2}$ s$^{-1}$, the error amplifies rapidly for the barotropic and baroclinic streamfunctions in the atmosphere, with a saturation level reached before 0.2 years (Figure \[Figerr1\], panels (a) and (b)). For the ocean fields, the picture is different with a very stable error until 0.1 year, after which it increases considerably concomitantly to the error amplification in the atmosphere. This phase is then followed by a slow increases for long lead times up to more than 20 years (Figure \[Figerr1\], panels (c) and (d)). This slow increase of the error in the ocean starts when the error in the atmosphere has already reached its saturation level. A similar picture has been found in the VDDG model version [@Vannitsem2016; @Vannitsem2017b] for small friction coefficients, with a limited predictability to about a month in the atmosphere and a long term predictability in the ocean.
When the friction coefficient is increased, the initial error amplification is slower due to the smaller value of the dominant Lyapunov exponent (Figure \[Figlyap1\]), explaining the slight shift of the error growth curve to the right. Besides this shift, the error behaviour in the atmosphere and the ocean is similar as for $C=0.005$ kg m$^{-2}$ s$^{-1}$. The saturation of the error is however substantially modified when $C$ is further increased. The saturation levels for the atmospheric fields are reached after about 2 years for $C$=0.015 kg m$^{-2}$ s$^{-1}$ and after about 20 years for $C$=0.02 kg m$^{-2}$ s$^{-1}$. In the ocean the error is still increasing after 25 years. This result reveals that the friction coefficient $C$ plays an important role in the long term predictability of the atmospheric fields.
![Evolution of the averaged error measured using the logarithmic norm for (a) the atmospheric barotropic streamfunction; (b) the atmospheric baroclinic streamfunction; (c) the ocean streamfunction; and (d) the ocean temperature. The different curves correspond to different values of the friction coefficient $C$. The parameter values used here are $h=1000$ m, $C_o=350$ W m$^{-2}$, $n=1.7$.[]{data-label="Figerr1"}](Vannitsem_Fig3){width="100.00000%"}
Several other values of $C$ have been explored as illustrated in Figure \[Figerr2\]. In this figure the error for the barotropic streamfunction fields saturates at larger and larger lead times as a function of $C$, for values above $C$=0.01 kg m$^{-2}$ s$^{-1}$. For $C$=0.02 kg m$^{-2}$ s$^{-1}$, the error has not reach its saturation level after about 20 years yet. This feature is reminiscent of the change of long term predictability experienced in the VDDG model [@Vannitsem2015a; @Vannitsem2016; @Vannitsem2017b] for large values of friction, with a long term predictability associated with the presence of an unstable periodic orbit around which the attractor is organized. The dynamics around this unstable periodic orbit however only appears after a chaos-to-chaos transition. This modification is visible in the abrupt change of the dominant Lyapunov exponents as illustrated in Figure 10 of [@Vannitsem2017b]. One can wonder whether a similar change is experienced in the present model version. A first remark is that there is no such drastic change of the dominant Lyapunov exponent in the present model version as shown by the green filled circles in panel (b) of Figure \[Figlyap1\] for $n=1.7$, although the variations of the dominant Lyapunov exponent seem larger when $C$ is increased from 0.01 to 0.02 kg m$^{-2}$ s$^{-1}$.
![Evolution of the averaged error measured using the logarithmic norm for the atmospheric barotropic streamfunction for different values of $C$. The parameter values used here are $h=1000$ m, $C_o=350$ W m$^{-2}$, $n=1.7$.[]{data-label="Figerr2"}](Vannitsem_Fig4){width="100.00000%"}
In order to understand this property, it is necessary to go back to the dynamics described in Section \[solution\]. For large values of the friction coefficient $C$, the solution of the model from time to time makes an excursion in the vicinity of an unstable periodic orbit, for which the atmosphere is almost frozen and closely connected with the motion of the ocean. In such a situation the solution is predictable for quite a long time. In the statistics of the error evolution, this leads to a slow saturation of the error for long times. This intermittent behaviour lies at the origin of the long-term predictability of the atmosphere illustrated in Figure (\[Figerr2\]). For small values of $C$, there is no excursion close to such a periodic solution, leaving the atmosphere behaving with a high level of chaoticity.
In order to further elucidate how excursions close to the periodic solution is inducing long-term predictability, the behaviour of the solutions for several realizations of the error evolution have been investigated in more details. Figure \[Figerr2-real\]a displays several realizations of the error evolution. The first one shows a rapid increase of the error which saturates after about 0.2 years, but the others display saturation only after 0.5 years, 1.5 years, and 2 years. This clearly suggests that in some circumstances long term predictability is present in the coupled system. If one looks at the two extreme cases, realizations 1 and 3, at panels (b) and (c) of Figure \[Figerr2-real\] where the control and perturbed dynamics of the variable $\psi_{a,1}$ are displayed, one can realize that a very different dynamics is found. For Realization 1, an oscillating behaviour (but still chaotic) is present suggesting that the dynamics is partly driven by the existence of an (unstable) periodic orbit. This evolution is accompanied with a small sensitivity to initial states as reflected by the superposition of the control and the perturbed solutions for about up to 2 years. Note also that both solutions are still close to each other along the oscillating pattern up to 5 years, indicating that there is still a good potential for prediction up to that lead time (this phase is less clear in the error evolution since we used the logarithm of the error). For realization 3, the picture is very different with much less apparent oscillations, and much less concordance between the two trajectories. In the latter case, the long term predictability is very low.
Finally it is interesting to see on which part of the attractor these two realizations are located. Figure \[Figerr2-real\]d shows the 3-dimensional projection of the solutions on which the two realisations 1 and 3 of the error evolution are based. These are well separated in state space, with realisation 3 spanning the lower right part of the projected attractor. This lower part has been identified in the video S2 as the region where the periodic orbit should lie.
So it clearly appears that the long term predictability of the solutions is associated with the intermittent excursions of the trajectories in the vicinity of an unstable periodic orbit. As illustrated in video S2, these excursions occur apparently irregularly suggesting intermittent transitions in the vicinity of the unstable periodic orbit. This contrasts with the much more regular wandering behaviour around the unstable periodic orbit found in the context of the VDDG model.
![(a) Four realisations of the error evolution for $C=0.016$ kg m$^{-2}$ s$^{-1}$, $h=1000$ m, $C_o=350$ W m$^{-2}$, $n=1.7$; (b) control and perturbed trajectories for realisation 1; (c) control and perturbed trajectories for realisation 3; and (d) control trajectories of realisations 1 and 3 in a three-dimensional projection of the state space.[]{data-label="Figerr2-real"}](Vannitsem_Fig5){width="100.00000%"}
![Evolution of the averaged error measured using the logarithmic norm for (a) the atmospheric barotropic streamfunction; (b) the atmospheric baroclinic streamfunction; (c) the ocean streamfunction; and (d) the ocean temperature. The different curves correspond to different values of the friction coefficient $C$. The parameter values used here are $h=1000$ m, $C_o=350$ W m$^{-2}$, $n=1.5$.[]{data-label="Figerr3"}](Vannitsem_Fig6){width="100.00000%"}
Let us now turn to the ocean-friction induced chaos as found with $n=1.5$. Figure \[Figerr3\] displays the error evolution for different values of $C$. For $C$ in the chaotic regime, a similar evolution as for $n=1.7$ is found, with a rapid initial amplification and then a saturation phase. For most of the cases explored, the saturation level is only reach at very long lead times, the only case with a short predictability period (still beyond a year) is the one for $C=0.012$ kg m$^{-2}$ s$^{-1}$. To interpret this long term predictability, let us figure out what kind of dynamics is taking place by looking at the different videos made with $n=1.5$ (videos S5-S7). As for $n=1.7$, the solutions behave chaotically with intermittent excursions close to a periodic solution with a long period, leading to a slow saturation of the error for long lead times.
![Evolution of the averaged error measured using the logarithmic norm for (a) the atmospheric barotropic streamfunction; (b) the atmospheric baroclinic streamfunction; (c) the ocean streamfunction; and (d) the ocean temperature. The different curves correspond to different values of the friction coefficient $C$. The parameter values used here are $h=100$ m, $C_o=350$ W m$^{-2}$, $n=1.7$.[]{data-label="Figerr4"}](Vannitsem_Fig7){width="100.00000%"}
A similar analysis of the error evolution is conducted for another depth of the ocean layer, $h=100$ m (Figures \[Figerr4\]-\[Figerr5\]). In this situation, the evolution of the error in the atmosphere does not display any long term predictability, except for $n=1.5$ and $C=0.005$ kg m$^{-2}$ s$^{-1}$. The intermittent behaviour found for $h=1000$ m does not emerge from the dynamics anymore. This contrasting behaviour suggests that the conditions for getting long term predictability are not always met and even when a strong coupling (large friction and large heat transfer between the components of the system) exists, the predictability is still limited to the typical time scale associated with the inverse of the dominant Lyapunov exponent.
![Evolution of the averaged error measured using the logarithmic norm for (a) the atmospheric barotropic streamfunction; (b) the atmospheric baroclinic streamfunction; (c) the ocean streamfunction; and (d) the ocean temperature. The different curves correspond to different values of the friction coefficient $C$. The parameter values used here are $h=100$ m, $C_o=350$ W m$^{-2}$, $n=1.5$.[]{data-label="Figerr5"}](Vannitsem_Fig8){width="100.00000%"}
In summary, the emergence of long term predictability in the present model version considerably differs from the behaviour found in the closed basin version of the model [@Vannitsem2015a; @Vannitsem2015b; @Vannitsem2017b]. In the latter work as already mentioned, a chaos-to-chaos bifurcation was identified leading to the development of a chaotic attractor around an unstable periodic orbit at the origin of a low-frequency variability. The solutions are then wandering around the unstable periodic orbit in a similar way as in simpler systems [@Tel2006; @Wernecke2017]. In the present system, an unstable periodic orbit is also involved, but the solution of the system is only experiencing intermittent excursions in the vicinity of this orbit. These excursions also last longer when the value of $C$ is increased.
Conclusions
===========
A new geometry for the integration of the coupled ocean-atmosphere model, MAOOAM, has been implemented with a channel flow for both the atmosphere and the ocean (periodic boundary conditions in the zonal direction for both model components). This new model version mimics the conditions that could be present around Antarctica. The predictability properties that may arise from the interaction between the ocean and the atmosphere are explored based on both the computation of the Lyapunov exponents and the long-term convergence of the mean square error toward a plateau, signature of the loss of predictability.
A first important result is that the interaction with the ocean can either induce chaos when the aspect ratio between the meridional and zonal length scales is small, or suppress chaos when the aspect ratio is large. This feature has been found to be robust to modifications of the depth of the ocean and the radiative input into the system. The ocean-induced chaotic regime may however be simply an artifact of the truncation of the spatial fields to their low-order versions, here limited to 10 Fourier modes.
A second remarkable result of the analysis is that long-term predictability as measured by the mean square error evolution is not a robust feature emerging from the coupling, as it was the case in the coupled ocean-atmosphere model with closed boundaries for the ocean [@Vannitsem2015a; @Vannitsem2015b; @Vannitsem2017b]. To have long term predictability (in the mean), some specific parameter values (the depth of the ocean, the friction coefficient) should be set in such a way that the solution of the model operates intermittent excursions in the vicinity of an unstable periodic orbit. This clearly demonstrates that emergence of low-frequency variability and the associated long-term predictability is not straightforwardly linked to the coupling between the different components of the climate system. For the interaction with the ocean as discussed here, this emergence clearly depends on the basin geometry in reduced-order models like MAOOAM.
Higher-resolution model versions should be explored in order to get confidence in these results. A preliminary analysis with model versions of higher dimensions with up to 312 variables, half for the atmosphere and half for the ocean, has been performed. Figure 10 illustrates the results for $h=1000$ m, $C_o$= 350 W m$^{-2}$, n=1.7, and different values of $C$ arbitrarily taken. When $C$ is small, a highly erratic behaviour is found, while for larger values of $C$ the behaviour displays strong similarities with the dynamics at lower resolutions shown in Figure 1, with an erratic decrease of the dominant ocean temperature mode close to 0 (see panel (c)). This suggests that the dynamics found for the reduced-order model is a generic dynamics that can be found at higher resolutions. The mechanism behind this dynamics is therefore worth investigating in more realistic models like the one developed by @Hogg2003 and also in realistic climate models.
![Time series of $\theta_{a,1}$ and $T_{o,1}$ with a higher resolution model version with 312 variables for (a) $C=0.01$ kg m$^{-2}$ s$^{-1}$; (b) $C=0.02$ kg m$^{-2}$ s$^{-1}$; (c) $C=0.03$ kg m$^{-2}$ s$^{-1}$ and (d) $C=0.04$ kg m$^{-2}$ s$^{-1}$. The other parameters are $n=1.7$, $C_o=350$ W m$^{-2}$ and $h=1000$ m. Nondimensional units are used for the variables.[]{data-label="Fighires"}](Vannitsem_Fig9){width="100.00000%"}
The present open-channel ocean model has a horizontal geometry with some similarities with the open-channel structure of the Southern Ocean. However the homogeneous vertical and horizontal structure of the model is far from being appropriate to describe the Southern ocean since strong horizontal variations of the density are present inducing the development of the Antarctic Circumpolar Current (ACC), [@Vallis2006]. The model is thus not able to describe the ACC appropriately. One can therefore wonder whether our analysis is relevant for the actual Southern Ocean. This can be addressed by modifying the ocean dynamics allowing for horizontal gradient of densities and the development of convective instabilities. This analysis is planned in the future along the lines of the ocean modelling approach of @Hogg2003 for instance.
Finally, in the VDDG and the current model versions, the long-term predictability is associated with the presence of an unstable periodic orbit whose origin can be traced back to the interaction between the ocean and the atmosphere. This long-term predictability mechanism was already discovered in much simpler “single scale” systems [@Tel2006; @Gros2015; @Wernecke2017] and the dynamics developing in this context is referred to as [*partially predictable chaos*]{}. Our results are reminiscent of this mechanism but here in a multiscale system with very long periods of motion along the unstable periodic orbits. The current results, however, point toward the possibility of having different routes to long-term predictability. A first one has been obtained in [@Vannitsem2015a; @Vannitsem2015b; @Vannitsem2017b] for the closed ocean basin model with a wandering along the unstable periodic orbit, and second as in the present model configuration, an intermittent transition from a highly chaotic attracting set to the vicinity of an unstable periodic orbit. In the latter case, only some forecasts inherit of the long-term predictability properties and the difficulty is to know when this transition will occur. Finding precursors of this transition is essential in order to assess the potential predictability of such forecasts, like for instance when searching for precursors of atmospheric blocking [@Vautard1990; @Matsueda2018] or the precursors of specific ENSO events [@Duan2004].
Supporting information
======================
Seven videos were prepared to visualize the solutions of the model for different values of the parameters, and are provided as supplementary material. They are listed from video S1 to video S7. Table \[tab:vid\] lists the parameter values and the digital object identifiers (DOIs) for each video. The other parameters are fixed at $C_o=350$ W m$^{-2}$ and $h=1000$ m for these simulations.
------- ----- -------------------------- ---------------------------------
Video n C (kg m$^{-2}$ s$^{-1}$) DOI
S1 1.7 0.010 <https://doi.org/10.5446/39179>
S2 1.7 0.016 <https://doi.org/10.5446/39180>
S3 1.7 0.020 <https://doi.org/10.5446/39181>
S4 1.7 0.027 <https://doi.org/10.5446/39182>
S5 1.5 0.010 <https://doi.org/10.5446/39183>
S6 1.5 0.012 <https://doi.org/10.5446/39184>
S7 1.5 0.016 <https://doi.org/10.5446/39185>
------- ----- -------------------------- ---------------------------------
: List of videos\[tab:vid\]
[^1]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The transport of the energy contained in electrons, both thermal and suprathermal, in solar flares plays a key role in our understanding of many aspects of the flare phenomenon, from the spatial distribution of hard X-ray emission to global energetics. Motivated by recent [*RHESSI*]{} observations that point to the existence of a mechanism that confines electrons to the coronal parts of flare loops more effectively than Coulomb collisions, we here consider the impact of pitch-angle scattering off turbulent magnetic fluctuations on the parallel transport of electrons in flaring coronal loops. It is shown that the presence of such a scattering mechanism in addition to Coulomb collisional scattering can significantly reduce the parallel thermal and electrical conductivities relative to their collisional values. We provide illustrative expressions for the resulting thermoelectric coefficients that relate the thermal flux and electrical current density to the temperature gradient and the applied electric field. We then evaluate the effect of these modified transport coefficients on the flare coronal temperature that can be attained, on the post-impulsive-phase cooling of heated coronal plasma, and on the importance of the beam-neutralizing return current on both ambient heating and the energy loss rate of accelerated electrons. We also discuss the possible ways in which anomalous transport processes have an impact on the required overall energy associated with accelerated electrons in solar flares.'
author:
- 'Nicolas H. Bian, Eduard P. Kontar, and A. Gordon Emslie'
bibliography:
- 'bian\_et\_al\_turbulent\_transport.bib'
title: 'SUPPRESSION OF PARALLEL TRANSPORT IN TURBULENT MAGNETIZED PLASMAS AND ITS IMPACT ON NON-THERMAL AND THERMAL ASPECTS OF SOLAR FLARES'
---
Introduction
============
A solar flare involves a complex set of energy release and transport mechanisms, with both non-thermal and thermal elements [see, e.g., @1988psf..book.....T; @2011SSRv..159..107H for reviews]. In particular, it is generally accepted [e.g., @2011SSRv..159..301K] that a significant fraction of the energy released is in the form of deka-keV electrons. These electrons *gain* energy through as-yet-not-fully-understood process(es) associated with the reconnection of stressed, current-carrying magnetic fields [see, e.g., @2011SSRv..159..357Z for a review]. They *lose* energy principally through Coulomb collisions [e.g., @1972SoPh...26..441B; @1978ApJ...224..241E], although ohmic losses associated with driving the beam-neutralizing return current through the finite-resistivity ambient medium , may also be significant.
Early impulsive phase models [e.g., @1971SoPh...18..489B] assumed, largely for simplicity, that the electrons are accelerated out of a “point source” at or near the apex of a coronal loop. However, estimates [e.g., @2003ApJ...595L..97H; @2011SSRv..159..107H] of the number of accelerated electrons required to produce a strong hard X-ray burst, combined with even generous estimates of the number density of electrons in the acceleration region, show that an acceleration region extending over a substantial portion of the flare volume is required. Additional evidence for an extended acceleration region includes the following:
- While many observations [e.g. @2002ApJ...569..459P; @2003ApJ...595L.107E] point to an acceleration site in the corona and the production of hard X-ray footpoints by electrons precipitating into the chromosphere, the appearance of coronal hard X-ray sources , in particular extended nonthermal coronal sources shows that the accelerated electrons are, at least in some events, fully confined to the extended coronal region where the acceleration occurs;
- [*RHESSI*]{} [@2002SoPh..210....3L] observations reveal that the accelerated electron distribution is nearly isotropic [e.g., @2006ApJ...653L.149K]. This favors stochastic acceleration mechanisms operating throughout an extended region [e.g., @1992ApJ...398..350H; @1994ApJS...90..623M; @2008ApJ...687L.111B; @2009ApJ...692L..45B; @2012SSRv..173..535P; @2012ApJ...754..103B];
have shown that the number of electrons trapped in extended coronal sources significantly exceeds the number consistent with purely collisional transport of these electrons to the chromospheric footpoints. They thus argue that some form of non-collisional scattering mechanism confines electrons to the coronal part of the loop (and hence to the acceleration region), even in coronal-source-plus-footpoint events. @2014ApJ...787...86J have noted that observed variation of hard X-ray source length with photon energy is not consistent with a purely collisional transport model but rather with one in which parallel transport proceeds through a process involving both collisions and noncollisional scattering. @2014ApJ...780..176K further showed, through analysis of similar events, that the mean free path associated with the noncollisional process is of order $10^8$ cm, an order of magnitude or so less than the collisional mean free path.
Local fluctuations in the magnetic field are already well known to be responsible for angular scattering and isotropization of the particle distribution, leading to spatial diffusion along the guiding magnetic field [@1966ApJ...146..480J; @1989ApJ...336..243S; @2009ASSL..362.....S]. [*Cross-field*]{} diffusion of particles in turbulent magnetic fields has also long been considered as a mechanism for cosmic-ray transport , for the transport of solar energetic particles [@2003ApJ...590L..53M; @2013ApJ...773L..29L], and for transport of thermal electrons in coronal loops . In this context, @2011ApJ...730L..22K and have shown that the width (perpendicular to the guiding magnetic field) of extended coronal hard X-ray sources increases slowly with energy, consistent with transport of energetic electrons across the guiding field lines through collisionless scattering off magnetic inhomogeneities.
In summary, hard X-ray observations of solar flares strongly suggest that electrons are accelerated in an extended region, within which a combination of Coulomb collisions and collisionless pitch-angle scattering operate. In this paper we explore the implications that the addition of such a non-collisional scattering process has for the transport of electrons of [*all*]{} energies, both thermal and non-thermal, in the flaring corona.
Soft X-ray observations [e.g., @2011SSRv..159..107H as a recent review] show that the overall spectrum (and hence the distribution of emitting electrons in the flaring corona) is often near-Maxwellian, with a temperature of a few $\times 10^7$ K. Such fits to flare soft X-ray spectra also provide estimates of the emission measures $EM \equiv \int n^2 \, dV$ of the soft-X-ray-emitting volume; in large flares this can be $\simeq 10^{49}$ cm$^{-3}$ or higher. Since the emitting volume is of order $10^{27}$ cm$^3$, this requires a density of order $10^{11}$ cm$^{-3}$. At such densities and temperatures, the collisional mean free path is of order $10^8$ cm, significantly less than the characteristic source length $L \sim V^{1/3} \simeq 10^9$ cm. Thus, unlike in other situations (e.g., in the solar wind), collisional processes are important in determining the ambient conditions in the plasma, which accounts for the very good fit of Maxwellian forms to observed soft X-ray spectra.
In a collisional environment, transport of quantities such as heat and electric charge are principally determined [e.g. @1962pfig.book.....S] by local gradients in temperature and density. The presence of additional non-collisional scattering processes does not change this essentially local [*nature*]{} of plasma transport phenomena. However, collisionless scattering processes affect the transport coefficients and hence the values of heat flux and current that arise from prescribed values of the local temperature gradient and large-scale electric field. In this paper, we evaluate this effect quantitatively, and we also discuss how substantial deviations from the classical [@1962pfig.book.....S] values of the thermoelectric transport coefficients can have very significant implications for several areas of importance to the solar flare problem. Such implications include:
- [ Suppression of thermal conduction will result in a higher coronal temperature for a prescribed heating rate (e.g., by nonthermal electrons) and so possibly account for very hot sources observed to be confined in the corona ;]{}
- [ A reduction in the thermal conductivity coefficient $\kappa_\parallel$ will lengthen the conductive cooling time $\tau_{\rm cool} \simeq 3nkT/(\kappa_\parallel T^{7/2}/L^2$) from its classical value [@1980sfsl.work..341M] and so offers a possible resolution to the long-standing conundrum of the apparent need for continued energy input to coronal plasma after the impulsive phase.]{}
- [ Nonthermal electrons lose energy in driving a beam-neutralizing return current through the finite resistivity of the ambient atmosphere. For a given beam current density $j_\parallel$, a reduction in the electrical conductivity $\sigma_\parallel$ increases the role of return current losses relative to Coulomb collisions, and this can have significant implications for the spatial distribution of hard X-ray emission and for the energy deposition profile throughout the ambient atmosphere, and hence the hydrodynamic response of the solar atmosphere [e.g., @1984ApJ...279..896N; @1989ApJ...341.1067M; @2005ApJ...630..573A]. Confinement of nonthermal electrons through enhanced resistivity may also offer an alternative explanation for loop-top coronal sources . And, because the hard X-ray production of a population of accelerated electrons depends inversely on the energy loss rates of the accelerated electrons, the results will change the required total energy content in these accelerated electrons, a quantity of considerable importance in the overall energetics of solar eruptive events [@2012ApJ...759...71E].]{}
In Section \[pitch-angle\] we consider the combined effects of collisional and turbulent scattering on the effective mean free path used to compute quantities such as the thermal conductivity $\kappa_\parallel$ and the electrical conductivity $\sigma_\parallel$. In Section \[modified-spitzer\] we study this problem more formally, using a Chapman-Enskog expansion of the kinetic equation for the electron phase-space distribution function $f(z,v,\mu)$, and we derive expressions for the various thermoelectric coefficients that link transport quantities (thermal flux, electrical current density) to the local environment (temperature gradient, electric field). We do this first for a model of isotropic scattering where the explicit dependence of the diffusion coefficient $D_{\mu\mu}$ on pitch angle $\mu$ is, for simplicity, discarded. Scattering by magnetostatic fluctuations, which is generally not isotropic, is addressed in Section \[dependent\]. In Section \[application\] we discuss the application of these results to several aspects of solar flares, including the heating of coronal plasma by non-thermal electrons and the subsequent cooling of this hot plasma by thermal conduction to the chromosphere (Section \[coronal-heating-cooling\]), and the role of return currents in ohmic heating of the corona and in the dynamics of the nonthermal electron population (Section \[return-current\]). In Section \[summary\] we summarize the results and present our conclusions.
Effects of Pitch-Angle Scattering on Particle Transport {#pitch-angle}
=======================================================
The basis of turbulent scattering theory in plasmas was developed some time ago [see @1969npt..book.....S for a review]. Here we adopt the philosophy that turbulence can be thought of as playing a role similar to collisions, resulting in a “rescaling” of the transport coefficients with respect to their collisional values. We also focus on elastic angular scattering, which is the predominant effect for scattering of particles by low-frequency turbulence [@1966JETP...23..145R].
Modeling the scattering frequency and mean free path {#scattering-parameters}
----------------------------------------------------
The pitch-angle diffusion coefficient $D_{\mu\mu}$ of charged particles in a magnetized plasma is related to the angular scattering frequency $\nu $ (s$^{-1}$) by
$$\label{dmumudef}
D_{\mu\mu}=\nu \, \frac{(1-\mu^2)}{2} \,\,\,,$$
where $\mu$ is the cosine of the pitch-angle relative to the guiding magnetic field. In the case where angular scattering is a superposition of two additive processes, collisional and turbulent, the scattering frequency can be written as
$$\label{dmumu-total}
\nu (v) = \nu_{C}(v) + \nu_{T}(v) \,\,\, .$$
Here the collisional contribution is given by
$$\label{dmumu-c}
\nu_{C}(v) = \frac{4\pi n_e \, e^4 \, \ln \Lambda}{m_e^2} \, \frac{1}{v^3} \equiv \frac{v}{\lambda_{\rm C}(v)} \,\,\, ,$$
where we have introduced the collisional mean free path
$$\label{lambdac-def}
\lambda_{\rm C} (v) = \frac{m_e^2}{4\pi n_e \, e^4 \, \ln \Lambda} \, {v^4}\equiv \lambda_{\rm ei} \, \left( \frac{v}{v_{\rm te}} \right)^{4} \,\,\, .$$
Here $e$ and $m_e$ are the electronic charge (esu) and mass (g), respectively, $n_e$ (cm$^{-3}$) is the ambient electron density, $\ln \Lambda$ is the Coulomb logarithm, and the thermal mean free path
$$\label{lambda-ei}
\lambda_{\rm ei} = \frac{(2 k_B)^2}{2 \pi e^4 \, \ln \Lambda} \, \frac{T_e^2}{n_e}\simeq \frac{10^{4} \, T_{e}^{2}}{n_{e}} \,\,\, ,$$
which, by definition, is the collisional mean free path of electrons with thermal speed $v = v_{\rm te} \equiv \sqrt{2 k_B T_e/m_e}$, where $k_B$ is Boltzmann’s constant.
Similarly, the turbulent contribution to the angular scattering frequency may be written
$$\label{dmumu-t}
\nu_{T}(v) = \frac{v}{\lambda_{\rm T}(v)} \,\,\, ,$$
which involves a (generally velocity-dependent) quantity $\lambda_{\rm T}$, identifiable as the “turbulent mean free path.” The physical origin of the turbulence is, for the moment, left unspecified. However, in order to exploit the analogy with collisional scattering we assume a velocity dependence for the turbulent mean free path of the form
$$\label{alpha-def}
\lambda_{T}(v)=\lambda_{0} \left( \frac{v}{v_{\rm te}} \right )^{\alpha},$$
corresponding to a turbulent scattering frequency
$$\label{nut-v}
\nu_{T}(v)=\frac{v} {\lambda_{0}} \, \left ( \frac{v}{v_{\rm te}} \right )^{-\alpha} \,\,\, .$$
Using this approach, we can express the total scattering frequency resulting from a combination of collisions and turbulent scattering in the form
$$\label{lambda-eff}
\nu (v) = \frac{v}{\lambda(v)} \,\,\, ,$$
where the “effective,” or simply *the*, mean free path $\lambda (v)$ is given by
$$\label{lambda-eff-c-t}
\frac{1}{\lambda(v)} = \frac{1}{\lambda_{\rm C}(v)} + \frac{1}{\lambda_{\rm T}(v)} = \frac{1}{\lambda_{\rm ei}} \left ( \frac{v_{\rm te}}{v} \right )^4 + \frac{1}{\lambda_{\rm T}(v)}
= \frac{1}{\lambda_{\rm ei}} \left ( \frac{v_{\rm te}}{v} \right )^4 + \frac{1}{\lambda_0} \left ( \frac{v_{\rm te}}{v} \right) ^{\alpha} \,\,\, .$$
We will find it convenient to introduce the dimensionless ratio
$$\label{r}
R = \frac{\lambda_{\rm ei}}{\lambda_{0}} \,\,\, .$$
By equations (\[dmumu-c\]), (\[lambdac-def\]), (\[nut-v\]), and (\[r\]), the total scattering frequency can be written in the form
$$\label{nu-v}
\nu(v) = \frac{v_{\rm te}}{\lambda_{\rm ei}} \, \frac{1+R (v/v_{\rm te})^{4-\alpha}}{(v/v_{\rm te})^3} \,\,\, ,$$
with the corresponding expression
$$\label{lambda-v}
\lambda(v) = \frac{\lambda_{\rm ei} \, (v/v_{\rm te})^4}{1+R (v/v_{\rm te})^{4-\alpha}}$$
for the mean free path.
The parameter $R$ is the ratio of two length scales, collisional and turbulent, and here plays the role of a transport reduction factor. In the presence of turbulence, the scattering frequency for thermal electrons with $v\sim v_{\rm te}$ is $\nu=\nu_{\rm ei}(1+R)$, an increase by a factor $(1+R)$ relative to the collisional value $\nu_{\rm ei}=v_{\rm te}/\lambda_{\rm ei}$. When $R$ is large, this increased scattering frequency corresponds to a decrease of the the mean free path by a factor $\simeq$$R$.
Pitch-angle scattering due to turbulence is in general *not* isotropic; the scattering frequency may also depend on pitch angle, say as
$$\label{nu_t_anisotropic}
\nu_{T}(v,\mu) = \frac{v} {\lambda_{0}} \, \left ( \frac{v}{v_{\rm te}} \right )^{-\alpha}|\mu|^{\beta} \,\,\, .$$
In such cases, the mean free path is related to the scattering frequency by the relation
$$\label{oo}
\lambda_{T} (v,\mu) =
\frac{3v}{4}\int _{-1}^{+1} \, \frac{(1-\mu^{2})}{\nu(v,\mu)} \, d\mu \,\,\, .$$
We caution that in the limit of very large $R$, one cannot compute the mean free path simply by first taking this limit in Equation (\[nu-v\]) and then substituting this into Equation (\[oo\]). For example, it is easily checked that, for pure turbulent scattering ($\nu=\nu_{T}$) and certain values of $\beta$ (e.g., $\beta=1$), the expression (\[oo\]) for the purely turbulent mean free path diverges. Physically, this is because for such values of $\beta$ turbulent scattering alone is incapable of scattering particles through the $90^\circ$ ($\mu=0$) “barrier.” It is therefore essential, in general, to include collisional effects even in the limit of large $R=\lambda_{\rm ei}/\lambda_{0}$, so that the mean free path given by Equation (\[oo\]) remains finite. The corresponding scattering frequency is given (cf. Equation (\[nu-v\])) by
$$\label{nu-v-3}
\nu(v,\mu) = \frac{v_{\rm te}}{\lambda_{\rm ei}} \, \frac{1+R \, |\mu|^{\beta} \, (v/v_{\rm te})^{4-\alpha}}{(v/v_{\rm te})^3} \,\,\, .$$
The existence of a turbulent spectrum of magnetic fluctuations transverse to the guiding magnetic field is well-documented [see, e.g., @2009ASSL..362.....S for a review]. Such fluctuations can be considered as quasi-static provided the characteristic particle velocity (a few $v_{\rm te}$ for the electrons that carry the bulk of the thermal flux) is larger than $\lambda _{B}/\tau_{B}$, where $\lambda_{B}$ and $\tau_{B}$ are respectively the correlation length and time associated with the fluctuations. Typically, the turbulent mean free path parameter $\lambda_{0}$ is proportional to the inverse square of the fractional level of the magnetic fluctuations, i.e.,
$$\label{lambda0-lambdaB}
\lambda_{0} \simeq \lambda_{B} \, \left ( \frac{\delta B_{\perp}}{B_{0}} \right )^{-2},$$
so that the value of the ratio $R$ is
$$R = \frac{\lambda_{\rm ei}}{\lambda_{0}} \simeq \left ( \frac{\lambda_{\rm ei}}{\lambda_B} \right ) \,
\left ( \frac{\delta B_\perp}{B_0} \right )^2 = \frac{ 10^4 \, T_e^2 \, (\delta B_\perp/B_0)^2}{n_e \, \lambda_B} \,\,\, .$$
Effect on transport coefficients
--------------------------------
We now consider in a semi-quantitative way the impact of adding turbulent scattering on various transport coefficients. The “collisionality” of a plasma is inversely proportional to the ratio of the collisional mean free path $\lambda_C$ to the temperature gradient length scale $L_T$; low values the collisional Knudsen number ${\rm Kn}_S \equiv \lambda_C/L_T$ imply a high degree of collisionality and vice versa. (Here the subscript $S$ stands for @1962pfig.book.....S.) The collisionality of the solar electron population ranges from quite high (${\rm Kn}_{S}\sim 10^{-2}$) close to the Sun to very low (${\rm Kn}_{S}\sim 1$) in the solar wind at 1 AU. For coronal loops, $L_{\rm T}$ is of the order of the loop length $L$, therefore the collisional Knudsen number ${\rm Kn}_{S} \simeq 10^{4}T_{e}^{2}/n_{e}L$. Taking $n_{e} = 10^{10}$ cm$^{-3}$ and $L=10^{9}$ cm, the collisional Kundsen number ${\rm Kn}_{S}=10^{-15}T_{e}^{2}$, meaning that the range of temperatures $T_{e}=10^{6}-10^{7}$ K corresponds to collisional Knudsen numbers ${\rm Kn}_{S}=10^{-3}-10^{-1}$.
Now let us consider the heat flux density carried by electrons along the field line, which for free-streaming electrons may be straightforwardly written as
$$\label{qparallel}
q_{\parallel} = n_{e} m_e v_{\rm te}^3 = n_e m_e \left ( \frac{2 k_B T_e}{m_e} \right )^{3/2} \,\,\, .$$
However, we know from Fick’s law that when the collisional Knudsen number $\lambda/L_T$ is small, the thermal flux is driven by the local temperature gradient. Hence we write $T^{3/2}_{e} = - \lambda_C \, T_{e}^{1/2}dT_e/dz$, resulting in an approximate expression for the parallel heat flux in a collisional environment:
$$\label{kappa-def-coll}
q_{\parallel} = -\frac{2 n_e \, k_B \, (2 k_B T_e)^{1/2}}{m_e^{1/2}} \, \lambda_C \, \frac{dT_e}{dz} \equiv - \, \kappa_{\parallel,S} \, \frac{dT_e}{dz} \,\,\, .$$
Adding collisionless scattering effects gives
$$\label{kappa-def}
q_{\parallel} = -\frac{2 n_e \, k_B \, (2 k_B T_e)^{1/2}}{m_e^{1/2}} \, \lambda \, \frac{dT_e}{dz} \equiv - \, \kappa_\parallel \, \frac{dT_e}{dz} \,\,\, ,$$
where
$$\kappa_{\parallel}=\frac{2 n_e \, k_B \, (2 k_B T_e)^{1/2}}{m_e^{1/2}} \, \lambda$$
is the thermal conductivity, which, through the form of $\lambda$ (Equation (\[lambda-v\])), includes the effects of both collisional and non-collisional scattering.
When the turbulent transport reduction factor $R = \lambda_{\rm ei}/\lambda_{0}$ is small, the mean free path takes on its collisional value $\lambda \simeq \lambda_{\rm ei}$ and the parallel heat conductivity correspondingly assumes the standard [e.g., @1962pfig.book.....S] collisional form $\kappa_{\parallel S} = 2 n_e k_B \, (2 k_B T_e)^{1/2} \, \lambda_{\rm ei}/m_e^{1/2} = k_B (2 k_B T_e)^{5/2}/(\pi m_e^{1/2} e^4 \ln \Lambda)$. (Note that this is density independent because $\lambda_{ei} \propto T_e^2/n_e$.) On the other hand, when the turbulent reduction factor $R$ is large, the heat conductivity becomes significantly suppressed relative to the collisional value and obeys the (generally density dependent) scaling $\kappa_\parallel = 2 n_e k_B (2 k_B T_e)^{1/2} \lambda_{0}/m_e^{1/2} = R^{-1} \kappa_{\parallel S}$.
Next we consider the [*electrical*]{} conductivity in a turbulent plasma. As usual, this can be obtained by writing the balance between the electric force and the friction force acting on an electron:
$$\label{eqn-motion}
-eE_\parallel - \nu \, m_e v_\parallel = 0 \,\,\, ,$$
where $E_\parallel$ (statvolt cm$^{-1}$) is the component of the electric field in the direction of the background guiding magnetic field. From this we find the parallel current density (defined by $j_\parallel = - e n_e v_\parallel$) to be $j_\parallel = (n_e e^2/\nu \, m_e) \, E_\parallel = (n_e e^2 \lambda/m_e v_{\rm te}) \, E_\parallel$, leading to the Ohm’s law
$$\label{sigma-def}
j_\parallel = \frac{n_{e} e^2 \lambda}{m^{1/2}_e (2 k_B T_e)^{1/2}} \, E_\parallel \equiv \sigma_\parallel \, E_\parallel \,\,\, ,$$
where the electrical conductivity
$$\sigma_\parallel=\frac{n_{e} e^2 \lambda}{m^{1/2}_e (2 k_B T_e)^{1/2}} \,\,\, .$$
Again, when $R\ll 1$, the mean free path takes on its collisional value and the parallel electric conductivity obeys the collisional scaling $\sigma_{\parallel S}=n_e e^2 \lambda_{\rm ei}/m_e v_{\rm Te} = (2 k_B T_e)^{3/2}/(2 \pi m_e^{1/2} e^2 \ln \Lambda)$. On the other hand, when the turbulent reduction factor $R\gg 1$, the electric conductivity becomes significantly suppressed relative to the collisional value and obeys the scaling $\sigma_\parallel = n_e e^2 \lambda_{0}/m_e v_{\rm te} = R^{-1} \, \sigma_{\parallel S}$.
We may summarize the above discussion into one simple and rather obvious expression, valid for $v \simeq v_{\rm te}$,
$$\label{dparallel-ratio}
\frac{\nu_{\rm ei}}{\nu}\sim \frac{\lambda}{\lambda_{\rm ei}}\sim
\frac{{\rm Kn}}{{\rm Kn}_{S}}\sim \frac{D_{\parallel}}{D_{\parallel S}}\sim\frac{\kappa_{\parallel}}{\kappa_{\parallel S}}\sim\frac{\sigma_{\parallel}}{\sigma_{\parallel S}}\sim \frac{1}{1+R} \,\,\, .$$
An enhanced rate of angular scattering yields reduced conductivities compared with the collisional (Spitzer) values.
The modified Spitzer problem {#modified-spitzer}
============================
In this section we will improve upon the dimensional estimates above by using a a more rigorous Chapman-Enskog expansion of the electron kinetic equation, i.e., by solving a standard [@1962pfig.book.....S] collisional transport problem that includes an additional non-collisional source of pitch-angle scattering. The model will involve two important non-dimensional parameters: the Knudsen number ${\rm Kn}$, which measures the degree of collisionality vs. free-streaming and is assumed to be smaller than unity, and the turbulent reduction factor $R$ (as defined above), which measures the additional role of collisionless scattering. Note that these non-dimensional numbers are independent; thus the large-$R$ asymptotic solutions should not be confused with the collisionless limit ${\rm Kn} \rightarrow \infty$: collisions are always essential (even when $R\gg 1$) to drive the electrons toward the Maxwellian distribution and to keep the overall mean free path finite.
Isotropic scattering
--------------------
We first consider the impact of an additional source of *isotropic* scattering on the transport coefficients. (This will set up the framework for later addressing the problem of angular scattering by a spectrum of transverse magnetostatic fluctuations, which is generally not isotropic.) Let us consider, then, the one-dimensional kinetic equation for a gyrotropic ($\partial/\partial \phi = 0$) electron distribution function $f(z,\theta, v, t)$ under the action of an electric field $E_\parallel$ parallel to the ambient magnetic field ${\bf B}$,
$$\label{kinetic}
\frac{\partial f}{\partial t} + v_\parallel \, \mathbf{b}.\nabla f - \frac{eE_\parallel}{m_e} \, \mathbf{b}.\nabla_{\mathbf{v}}f = St^{v}(f) + St^{\theta}(f) \,\,\, ,$$
where $z$ (cm) is the position of the particle gyrocenter along the magnetic field with direction ${\mathbf b} = {\mathbf B_0}/B_0$, $\theta$ is the pitch angle ($\cos \theta = \mathbf {v}.\mathbf{B}_0/vB_0 = v_\parallel/v$) and $v = \sqrt{v_\parallel^2 + v_\perp^2}$ is the electron speed. Transforming to the variables $(z,\mu,v,t)$, with $\mu= \cos \theta$ being the pitch-angle cosine, this Fokker-Planck equation may be rewritten as
$$\label{fokker}
\frac{\partial f}{\partial t} + \mu \, v \, \frac{\partial f}{\partial z} - \frac{e E_\parallel}{m_e} \, \mu\, \frac{\partial f}{\partial v} - \frac{e E_\parallel}{m_e} \, \frac{(1-\mu^2)}{v}
\, \frac{\partial f}{\partial \mu} = St^{v}(f) + St^{\mu}(f) \,\,\, .$$
The velocity scattering operator
$$\label{stv}
St^{v}(f) = \frac{1}{v^{2}} \, \frac{\partial}{\partial v} \left [ v^{2} D(v) \left ( \frac{\partial f}{\partial v} + \frac{m_e}{k_B T_e} \, fv \right ) \right ]$$
describes collisional diffusion in velocity space and collisional drag. The diffusion coefficient in velocity space is given by
$$\label{diff-v}
D(v) = \frac{4 \pi e^{4} \ln \Lambda \, n_e k_B T_e}{m_e^3} \, \frac{1}{v^3} \,\,\, ,$$
whereas the pitch-angle scattering operator takes the form
$$\label{stmu}
St^{\mu}(f) = \frac{\partial}{\partial \mu} \left [ D_{\mu\mu}\, \frac{\partial f}{\partial \mu} \right ] =\frac{\partial}{\partial \mu} \left [ \frac{\nu (v)}{2} \, (1-\mu^{2}) \, \frac{\partial f}{\partial \mu} \right ] \,\,\, .$$
While the velocity-space scattering operator (\[stv\]) is responsible for kinetic energy change, the pitch-angle scattering operator (\[stmu\]) is responsible for momentum change at constant kinetic energy. We recall that the scattering frequency $\nu(v)$, and hence (Equation (\[dmumudef\])) the pitch-angle diffusion coefficient $D_{\mu\mu}$, are sums of two components – collisional and turbulent pitch-angle diffusion; thus $D_{\mu\mu} = D^C_{\mu\mu} + D^T_{\mu\mu}$.
The right hand side of Equation (\[fokker\]) vanishes identically for a Maxwellian distribution, i.e.,
$$\label{st-maxwellian}
St^{v}(f_{0}) + St^{\mu}(f_{0}) =0 \,\,\, ,$$
where
$$\label{maxwellian}
f_{0}(v) = n_e \left ( \frac{m_e}{2\pi k_B T_e} \right )^{3/2} \, e^{-m_e v^2/2k_B T_e} \,\,\, .$$
However, if $T$ and/or $n$ are not uniform – $T_e=T_e(z)$ and/or $n_e=n_e(z)$ – then the (local) Maxwellian does [*not*]{} cancel the spatial transport term on the left hand side of the Fokker-Planck equation (\[fokker\]). This deviation from the homogeneous equilibrium state is responsible for the spatial flux of particles in the plasma and of what they carry (e.g., kinetic energy, electric charge). In this situation, we use a standard Chapman-Enskog expansion of the electron kinetic equation and write the distribution function as the sum of a zeroth order isotropic Maxwellian distribution plus a small flux-carrying anisotropic correction:
$$\label{perturb}
f = f_0(z,v) + \epsilon f_1(z,\mu,v) \,\,\, ,$$
where the expansion parameter $\epsilon$ is of the order of the (small) Knudsen number in the plasma, i.e., $\epsilon \sim$ Kn. All operators being linear, we obtain at order $\epsilon$ an equation for $f_{1}$:
$$\label{first-order}
St^{v}(f_{1}) + \frac{\partial}{\partial \mu} \left [ D_{\mu\mu}(v) \, \, \frac{\partial f_1}{\partial \mu} \right ] =
\mu \, v \, \frac{\partial f_0(z, v)}{\partial z} - \mu \, \left ( \frac{e E_\parallel}{m_e} \right ) \,
\frac{\partial f_0(z, v)}{\partial v} \,\,\, ,$$
where $f_0$ is the Maxwellian distribution (\[maxwellian\]). By solving this equation for the lead anisotropic component $f_1$, we can obtain expressions for both the heat flux and the electric current. This constitutes a standard @1962pfig.book.....S problem which becomes easily tractable in the Lorentz plasma approximation[^1], i.e., if one assumes that the scattering is angular only, i.e., $St^v(f_1)=0$. For such a case, Equation (\[first-order\]) can be immediately integrated once over $\mu$ to give
$$\label{df1dmu-c}
\frac{\partial f_1}{\partial \mu} = \frac{\mu^2}{\nu \, (1-\mu ^2)} \, \left [ v \, \frac{\partial f_0(z, v)}{\partial z}-
\frac{e E_\parallel}{m_e} \, \frac{\partial f_0(z, v)}{\partial v} \right ] + \frac{C}{\nu \, (1-\mu ^2)} \,\,\, .$$
The choice of the integration constant
$$C = - \left [ v \, \frac{\partial f_{0}(z, v)}{\partial z} - \frac{e E_\parallel}{m_e} \, \frac{\partial f_0(z, v)}{\partial v} \right ]$$ yields
$$\label{df1dmu}
\frac{\partial f_1}{\partial \mu} = -\frac{1}{\nu } \left [ v \, \frac{\partial f_0(z, v)}{\partial z}- \frac{e E_\parallel}{m_e} \, \frac{\partial f_0(z, v)}{\partial v} \right ] \,\,\, ,$$
which ensures the regularity of $f_1$ at $\mu=\pm 1$.
Since we are, for now, adopting an isotropic scattering model in which the scattering frequency $\nu$ is independent of the pitch angle cosine $\mu$, one can integrate Equation (\[df1dmu\]) once more to obtain
$$\label{f1-general}
f_1(z, v, \mu) = - \frac{\mu}{\nu} \left [ v \, \frac{\partial f_0(z, v)}{\partial z}- \frac{e E_\parallel}{m_e} \, \frac{\partial f_0(z, v)}{\partial v} \right ] \,\,\, .$$
Evaluating the pertinent space and velocity derivatives of $f_0$, and assuming a constant pressure along $z$: $n_e(z) k_B T_e(z) = P_e = {\rm constant}$, we obtain
$$\label{f1-basic}
f_1 (z, v, \mu) = - \frac{\mu v}{\nu} \left [ \left ( \frac{m_ev^2}{2 k_B T_e} - \frac{5}{2} \right )
\frac{1}{T_e} \, \frac{d T_e}{dz} + \frac{eE_\parallel}{k_B T_e} \right ] \, f_0 \,\,\, ,$$
with
$$\label{dmumu-v}
\nu (v) = \frac{v}{\lambda_{\rm c}(v)} + \frac{v}{\lambda_{\rm T}(v)} = \frac{v_{\rm te}}{\lambda_{\rm ei}} \, \frac{1+R(v/v_{\rm te})^{4-\alpha}}{(v/v_{\rm te})^3} \,\,\, .$$
It is convenient to introduce the normalization
$$\label{x-def}
x = \frac{v}{v_{\rm te}} \,\,\, ,$$
so that the zero-order Maxwellian distribution (Equation (\[maxwellian\])) at a given location (i.e., given values of $T_e$ and $v_{\rm te}$) can be written in the form
$$\label{f0}
f_0(x) = \pi^{-3/2} n_{e} \, v_{\rm te}^{-3} \, e^{-x^{2}} \,\,\, .$$
Further, Equation (\[dmumu-v\]) becomes
$$\label{dmumux3}
\nu = \frac{v_{\rm te}}{\lambda_{\rm ei}} \, \frac{Rx^{4-\alpha}+1}{x^{3}} \,\,\, ,$$
and thus the expression (\[f1-basic\]) for $f_1$ becomes
$$\label{f1-x-mu}
f_1 = - \mu \lambda_{\rm ei} \, \frac{x^4}{Rx^{4-\alpha}+1} \, \left [ \left ( x^2-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz} + \frac{eE_\parallel}{k_B T_e} \right ] \, f_0 \,\,\, .$$
We can now compute the heat flux from
$$\label{heat-flux-def}
q_\parallel(z) = 2 \pi \, \int_0^\infty dv \, v^2 \int_{-1}^1 d\mu \, \mu \, \left ( \frac{m_e v^2}{2} \right ) \, v \, f_1 (z, v, \mu) \,\,\, ,$$
which yields
$$\begin{aligned}
\label{qparallel-mu-v-f1}
q_\parallel(z)
&=& 2\pi k_B T_e \, v_{\rm te}^4\int_0^\infty dx \, x^5 \int_{-1}^1 d\mu \, \mu \, f_1(z,x,\mu)
\cr
&=& - \frac{4}{3 \sqrt{\pi}} \, n_{e} \, k_B T_e \, v_{\rm te} \, \lambda_{\rm ei}
\int_0^\infty dx \, \frac{x^9}{Rx^{4-\alpha}+1} \, \left [ \left ( x^2-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz}+\frac{eE_\parallel}{k_B T_e} \right ] \, e^{-x^2} \,\,\, .\end{aligned}$$
Writing this in the form
$$\label{qparallel-onsager-2}
q_\parallel = - \kappa_\parallel \, \frac{dT_e}{dz} - \alpha_\parallel \, E_\parallel \,\,\, ,$$
gives the thermoelectric coefficients
$$\label{kappa-parallel}
\kappa_{\parallel} = \frac{4}{3\sqrt{\pi}} \, n_e \, k_B \, v_{\rm te} \, \lambda_{\rm ei} \int_0^\infty \frac{x^9}{Rx^{4-\alpha}+1} \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx$$
and
$$\label{alpha-parallel}
\alpha_{\parallel} = \frac{4}{3\sqrt{\pi}} \, n_e \, e \, v_{\rm te} \, \lambda_{\rm ei} \int_0^\infty \frac{x^9}{Rx^{4-\alpha}+1} \, e^{-x^2} \, dx \,\,\, ,$$
respectively.
Similarly, substituting for $f_1$ from Equation (\[f1-basic\]) in the expression for the parallel current density
$$\label{current-density-def}
j_\parallel(z) = - 2 \pi \, e \int_0^\infty dv \, v^2 \int_{-1}^1 d\mu \, \mu \, v \, f_1(z, v, \mu)$$
gives
$$\begin{aligned}
\label{jparallel-mu-v-f1}
j_\parallel(z)
&=& - 2 \pi e \, v_{\rm te}^4 \int dx \, x^3 \int_{-1}^1 d\mu \, \mu \, f_1(z,x,\mu)
\cr
&=& \frac{4}{3 \sqrt{\pi}} \, n_e \, e \, v_{\rm te} \, \lambda_{\rm ei} \, \int_0^\infty dx \, \frac{x^7}{Rx^{4-\alpha}+1} \, \left [ \left ( x^2-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz} + \frac{eE_\parallel}{k_B T_e} \right ] \, e^{-x^2} \,\,\, ,\end{aligned}$$
and writing this in the form
$$\label{jparallel-onsager-2}
j_\parallel = \beta_\parallel \, \frac{dT_e}{dz} + \sigma_\parallel \, E_\parallel$$
gives the identifications
$$\label{beta-parallel}
\beta_{\parallel} = \frac{4}{3\sqrt{\pi}} \, \frac{n_e \, e \, v_{\rm te} \, \lambda_{\rm ei}}{T_e} \int_{0}^{\infty} \frac{x^7}{Rx^{4-\alpha}+1} \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx$$
and
$$\label{sigma-parallel}
\sigma_{\parallel} = \frac{4}{3\sqrt{\pi}} \, \frac{n_e \, e^2 \, v_{\rm te} \, \lambda_{\rm ei}}{k_B T_e} \int_0^\infty \frac{x^7}{Rx^{4-\alpha}+1} \, e^{-x^2} \, dx \,\,\, .$$
Note that enforcing the current neutrality condition $j_\parallel=0$ in Equation (\[jparallel-onsager-2\]) shows that $E_\parallel = (-\beta_\parallel/\sigma_\parallel) \, dT_e/dz$. Substituting this condition into Equation (\[qparallel-onsager-2\]) implies that the thermal conductivity coefficient in a current neutralized environment is given by $\kappa^* = \kappa_\parallel - \alpha_\parallel \beta_\parallel/\sigma_\parallel$.
We now provide explicit expressions for these coefficients in the $R \ll 1$ (collision-dominated) and $R \gg 1$ (turbulence-dominated) limiting regimes.
Limit of small R
----------------
As expected, in the limit $R\ll 1$ we can replace the term $(Rx^{4-\alpha}+1)$ in the denominators of the integrands in expressions (\[kappa-parallel\]), (\[alpha-parallel\]), (\[beta-parallel\]), and (\[sigma-parallel\]) with unity, thus recovering the collisional [@1962pfig.book.....S] values. Noting that the integral
$$\int_0^\infty dx \, x^n \, e^{-x^2} = \frac{1}{2} \, \Gamma \left ( \frac{n + 1}{2} \right ) = \frac{1}{2} \, \left (\frac{n-1}{2} \right ) !$$
gives the purely collisional results
$$\label{kappa-sn}
\kappa_{\parallel} = \frac{2\Gamma(6)- 5\Gamma(5)}{3 \sqrt{\pi}} \, n_e k_B \, v_{\rm te} \, \lambda_{\rm ei} = \frac{40}{\sqrt{\pi}} \, n_e k_B \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, ;$$
$$\label{alpha-sn}
\alpha_{\parallel} = \frac{2 \, \Gamma(5)}{3 \sqrt{\pi}} \, n_e \, e \, v_{\rm te} \, \lambda_{\rm ei} = \frac{16}{\sqrt{\pi}} \, n_e \, e \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, ;$$
$$\label{beta-sn}
\beta_{\parallel}=\frac{2\Gamma (5)-5\Gamma(4)}{3 \sqrt{\pi}} \, \frac{n_e \, e \, v_{\rm te} \, \lambda_{\rm ei}}{T_e}=\frac{6}{\sqrt{\pi}} \, \frac{n_e \, e \, v_{\rm te} \, \lambda_{\rm ei}} {T_e} \,\,\, ;$$
and
$$\label{sigma-sn}
\sigma_{\parallel} =\frac{2 \, \Gamma(4)}{3 \sqrt {\pi}} \frac{n_e e^2 \, v_{\rm te} \, \lambda_{\rm ei}}{k_B T_e} = \frac{4}{\sqrt {\pi}} \frac{n_e e^2 \, v_{\rm te} \, \lambda_{\rm ei}}{k_B T_e} \,\,\, .$$
Finally, in the zero-current regime, the effective thermal conductivity coefficient is
$$\label{kappa-star-sn}
\kappa^{*} \equiv \kappa_\parallel - \frac{\alpha_\parallel \beta_\parallel}{\sigma_\parallel} = \frac{16}{\sqrt{\pi}} \, n_e k_B \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, .$$
Limit of large R
----------------
On the other hand, when $R\gg 1$ we replace the term $(Rx^{4-\alpha}+1)$ in the denominators of the integrands in expressions (\[kappa-parallel\]), (\[alpha-parallel\]), (\[beta-parallel\]), and (\[sigma-parallel\]) with $Rx^{4-\alpha}$. This gives the following expressions:
$$\label{kappa-sn-rgg1}
\kappa_{\parallel} = \frac{2\Gamma(\frac{8+\alpha}{2})-5\Gamma(\frac{6+\alpha}{2})}{3 \sqrt{\pi}} \,\, \frac{n_e k_B \, v_{\rm te} \, \lambda_{\rm ei}}{R}\,\,\, ,$$
$$\label{alpha-sn-rgg1}
\alpha_{\parallel} = \frac{2 \, \Gamma(\frac{6+\alpha}{2})}{3 \sqrt{\pi} \, R} \, \, n_e \, e \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, ,$$
$$\label{beta-sn-rgg1}
\beta_{\parallel} = \frac{2\Gamma(\frac{6+\alpha}{2})-5\Gamma(\frac{4+\alpha}{2})}{3 \sqrt{\pi} \, R} \,\, \frac{n_e \, e \, v_{\rm te} \, \lambda_{\rm ei}}{T_e}\,\,\, ;$$
and
$$\label{sigma-sn-rgg1}
\sigma_{\parallel} = \frac{2 \, \Gamma(\frac{4+\alpha}{2})}{3 \sqrt {\pi} \, R} \,\, \frac{n_e e^2 \, v_{\rm te} \, \lambda_{\rm ei}}{k_B T_e}\,\,\, ;$$
The case $\alpha=0$ is of special interest as it corresponds (Equation (\[alpha-def\])) to a velocity-independent turbulent mean free path $\lambda_T$. In this case we have
$$\label{kappa-sn-rgg1alpha0}
\kappa_{\parallel}=\frac{2}{3\sqrt{\pi} \, R} \, n_e k_B \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, ,$$
$$\alpha_{\parallel}= \frac{4}{3 \sqrt{\pi} \, R} \, n_e \, e \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, ,$$
$$\label{beta-sn-rgg1alpha0}
\beta_{\parallel} = - \frac{1}{3 \sqrt{\pi} \, R} \, \frac{n_e \, e \, v_{\rm te} \, \lambda_{\rm ei}}{T_e} \,\,\, ;$$
and
$$\label{sigma-sn-rgg1alpha0}
\sigma_{\parallel} = \frac{2}{3 \sqrt {\pi} \, R} \, \frac{n_e e^2 \, v_{\rm te} \, \lambda_{\rm ei}}{k_B T_e} \,\,\, .$$
In a zero-current scenario, the effective thermal conductivity coefficient is
$$\label{kappa-star-sn-rgg1alpha0}
\kappa^{*} \equiv \kappa_\parallel - \frac{\alpha_\parallel \beta_\parallel}{\sigma_\parallel} = \frac{4}{3 \sqrt{\pi} \, R} \, n_e k_B \, v_{\rm te} \, \lambda_{\rm ei} \,\,\, .$$
As expected from the simple dimensional arguments of the previous Section, in the turbulence-dominated limit of large $R$ the transport coefficients are suppressed by a factor of $R$ with respect to the Spitzer values.
Figure \[fig:coefficients\] shows the full dependence on $R$ of all four thermoelectric coefficients for the case $\alpha=0$. We note from this Figure that in the turbulence-dominated high-$R$ limit, one cross-transport coefficient ($\beta_\parallel$) has not only changed in magnitude, but also in [*sign*]{}, compared to its collisional value (compare Equations (\[beta-sn\]) and (\[beta-sn-rgg1alpha0\])). Physically, this sign reversal is a consequence of the form of the Maxwellian distribution. In a uniform pressure environment,
$$\label{f0-T-dependence}
f_0(z,v) \sim \frac{n_e(z)}{[T_e(z)]^{3/2}} e^{-mv^2/2 k_B T_e(z)} = \left ( \frac {P_e}{k_B} \right ) \,
[T_e(z)]^{-5/2} \, e^{-mv^2/2 k_B T_e(z)} \,\,\, ,$$
from which we see that there are two contributions to the temperature gradient $dT_e/dz$, one involving the derivative of the normalization coefficient $T_e^{-5/2}$ and the other the derivative of the width of the Maxwellian velocity profile. The relative signs of these terms depend on the value of the speed $v$: increasing the temperature flattens the Gaussian velocity profile, leading to a general increase in $f_0$ at a high speeds but a general decrease at low speeds. As a consequence, the spatial derivative
$$\label{df0bdz}
\frac{\partial f_0(z,v)}{\partial z} = \frac{\partial f_0(z,v)}{\partial T_e} \, \frac{dT_e}{dz} \,\,\, ,$$
that appears in the expression (\[f1-general\]) for the lead anisotropic component $f_1(z,v,\mu)$ changes sign at some value of $v$. The overall sign of a particular transport coefficient (e.g., $\kappa_\parallel$, $\beta_\parallel$) depends on the relative contributions of the particular moments of $f_1(v,z,\mu)$ that appear[^2] in the expression for that coefficient. Higher-order moments of the velocity distribution are dominated by larger values of $v$, so that an increase in temperature generally leads to an increase in that moment, while lower-order moments place a greater weight on lower values of $v$, leading to a reduced increase (or even a decrease) in that moment. And, since (1) the moments of the velocity distribution involved in the calculation of $\kappa_\parallel$ and $\beta_\parallel$ depend on the value of $R$; and (2) the moments involved in the calculation of $\kappa_\parallel$ are generally higher than those involved in the calculation of $\beta_\parallel$, it follows that the signs of the various thermoelectric coefficients are not necessarily invariant with respect to the value of $R$.
Consider for definiteness the case of a negative temperature gradient, i.e., a temperature that decreases in the positive $z$ direction (the case of a positive temperature gradient is entirely similar). Also consider first the collisional (low-$R$) limit, for which the fourth-power dependence of the collision frequency $\nu$ causes all the moments involved to be quite high, specifically, the ninth and eleventh moments for $\kappa_\parallel$ (Equation (\[kappa-parallel\])) and the seventh and ninth moments for $\beta_\parallel$ (Equation (\[beta-parallel\])). These high velocity moments are all dominated by large values of $v$; thus, as we move toward increasing values of $z$ (i.e., toward regions of lower temperature), the predominant effect is a reduction in $f_0(z,v)$ with $z$: $\partial f_0(z,v)/\partial z < 0$. Both the heat flux and electron flow are predominantly carried by high-velocity electrons that move in the positive-$z$ direction to fill this relative void in $f_0$, and hence they are both oriented in the positive-$z$ direction, i.e., along the direction of decreasing temperature. Indeed, we see from Equation (\[f1-general\]) that the lead anisotropic term is aligned with the positive-$z$ direction: $f_1(z,v,\mu) \propto \mu$. The coefficients $\kappa_\parallel$ and $\beta_\parallel$ are therefore both positive (the former since $q_\parallel \sim -\kappa_\parallel \, dT_e/dz$ and the latter since the conventional current $j_\parallel \sim \beta_\parallel \, dT_e/dz$ flows in the opposite direction to the electron flow, i.e., in the negative-$z$ direction).
However, in the turbulence-dominated (high-$R$) limit, the velocity-distribution moments involved are all (four powers) lower because of the weaker velocity dependence of the collision frequency $\nu = \nu_T$ (compare Equation (\[alpha-def\]) \[with $\alpha = 0$\] and Equation (\[lambdac-def\])). Since the expression (\[kappa-parallel\]) for $\kappa_\parallel$ still involves relatively-high-order velocity moments (fifth and ninth), this reduction of all the moment orders is not sufficient to reverse the direction of the heat flux; it remains parallel to the direction of the negative temperature gradient. However, the considerably lower-order velocity moments (third and fifth) present in the expression (\[beta-parallel\]) for $\beta_\parallel$ are such that the value of the pertinent moments are now dominated by lower-velocity electrons. At these velocities a decrease in temperature causes an [*increase*]{} in $f_0(z,v)$. At the pertinent velocities, the spatial derivative of the isotropic zero-order term is therefore now positive: $\partial f_0(z,v)/\partial z > 0$, and so (Equation (\[f1-general\])) the lead anisotropic term now aligns in the negative-$z$ direction ($f_1(z,v,\mu) \propto - \mu$). The net result is that the predominant (low-velocity) electron flow is now in the negative-$z$ direction, i.e., in the direction of positive temperature gradient, antiparallel to the heat flux. The conventional current and the heat flux are now both in the positive $z$ direction, and so $\kappa_\parallel > 0$ and $\beta_\parallel < 0$.
![Transport coefficients values, relative to their collisional (Spitzer; $R=0$) limit as a function of $R$ for the case $\alpha = 0$ (turbulent mean free path $\lambda_T$ independent of velocity). The blue dashed line indicates negative values, which appear only for $\beta_\parallel$ in the high-$R$ limit; see explanation in the text.[]{data-label="fig:coefficients"}](f1.eps){width="0.56\linewidth"}
We remind the reader that all the above results have been derived under the simplifying assumption that the turbulent collision frequency $\nu_T$ is independent of pitch angle $\mu$, corresponding to isotropic scattering. While this has provided illustrative results on the impact of turbulent scattering on the various transport coefficients, in actual physical situations the scattering rate $\nu$ may depend on the pitch angle $\mu$ (and/or on the velocity $v$). This is indeed the case when angular scattering is produced by the presence of a spectrum of magnetic fluctuations in the plasma, next to be considered.
Transport reduction from pitch-angle scattering by transverse magnetic fluctuations {#dependent}
===================================================================================
Magnetostatic fluctuations $\delta \mathbf{B}_\perp$ perpendicular to the background magnetic field $B_0 \hat{\mathbf{z}}$ are a source of pitch-angle scattering in magnetized plasmas. For slab turbulence the pitch-angle diffusion coefficient takes the form
$$\label{stu-w}
D_{\mu\mu}^{T} = \frac{\pi}{2} (1-\mu^2) \, \Omega_{\rm ce} \,
\left . \frac{k_\parallel \, W(k_\parallel)}{B_0^2} \, \right \vert _{k_\parallel = \Omega_{\rm ce}/v_\parallel} \, \,\,\, ,$$
where $W(k_\parallel)$ is the spectral energy density of the magnetic fluctuations in wavenumber, i.e., $\int W(k_\parallel) \, dk_\parallel = (\delta B_\perp)^2$. The significance of the condition
$$\label{omegace}
\Omega_{\rm ce} = k_\parallel v_\parallel \equiv k_{\parallel} \, |\mu| \, v$$
in the expression (\[stu-w\]) is that scattering of electrons occurs as a result of gyro-resonance with zero-frequency magnetic modes during their cyclotronic orbits.
We shall consider the following form of the wavenumber spectrum:
$$W(k_{\parallel})=C(q) \, (\delta B_{\perp})^{2} \, \lambda_{B} \, \left [ 1+(k_\parallel \lambda_{B})^{2} \right ]^{-q} \,\,\, ,$$
where $C(q)$ is a normalization constant. Substituting this in Equation (\[stu-w\]) and using Equation (\[dmumudef\]) gives the corresponding turbulent scattering frequency:
$$\nu_{T} = \frac{2D^T_{\mu\mu}}{1-\mu^2} = \pi \, C(q) \, \Omega_{\rm ce} \,
\left(\frac{\delta B_{\perp}}{B_{0}}\right)^{2} \, \left(\frac{\Omega_{ce} \, \lambda_{B}}{|\mu| v}\right) \left[1+\left(\frac{\Omega_{ce}\lambda_{B}}{|\mu| v}\right)^{2}\right]^{-q} \,\,\, .$$
Notice the explicit dependence on $\mu$ which originates from the resonance condition (\[omegace\]). Introducing the normalized gyroradius (or rigidity parameter)
$$r_{\rm te} = \frac{v_{\rm te}}{\Omega_{\rm ce}\lambda_{B}} \,\,\, ,$$
we obtain, in the regime $r\ll 1$,
$$\nu_{T} = \pi \, C(q) \, \Omega_{\rm ce} \left( \frac{\delta B_{\perp}}{B_{0}} \right)^{2}
r_{\rm te}^{2q-1} \left (\frac{v}{v_{\rm te}} \right )^{2q-1} \, |\mu|^{2q-1} \,\,\, ,$$
which has the form (cf. Equation (\[nu\_t\_anisotropic\]))
$$\label{dmumu-t-mu}
\nu_{T} = \left (\frac{v_{\rm te}}{\lambda_{0}} \right ) \left (\frac{v}{v_{\rm te}} \right )^{1-\alpha} \, |\mu|^{\beta} \,\,\, ,$$
with
$$\label{alpha-beta-q}
1-\alpha=\beta=2q-1$$
and
$$\label{lamb}
\lambda_{0} =\frac{v_{\rm te}}{\pi \, C(q) \, \Omega_{\rm ce} \left( \frac{\delta B_{\perp}}{B_{0}} \right)^{2}
r_{\rm te}^{2q-1}}=\frac{1}{\pi \, C(q)} \, \lambda_{B} \, \left ( \frac{B_{0}}{\delta B_\perp} \right )^{2} \, r_{\rm te}^{2-2q}\,\,\, .$$
Note that the parameter $\lambda_{0}$ is *not* the turbulent mean free path. Rather, the turbulent mean free path is defined in terms of the parallel diffusion coefficient, which is itself a functional of the pitch-angle diffusion coefficient (or the scattering frequency):
$$\lambda_{T}\equiv \frac{3}{v} \, D_{\parallel}=\frac{3v}{8}\int _{-1}^{+1}d\mu \, \frac{(1-\mu^{2})^2}{D_{\mu\mu}^{T}}=
\frac{3v}{2}\int _{0}^{1}d\mu \, \frac{(1-\mu^{2})}{\nu_{T}}\,\,\, .$$
Substituting for $\nu_{T}$ from Equation (\[dmumu-t-mu\]), we obtain an expression for the turbulent mean free path:
$$\lambda_T = \lambda_{T0} \left ( \frac{v}{v_{\rm te} } \right)^{\alpha} = \lambda_{T0} \left ( \frac{v}{v_{\rm te} } \right)^{2 - 2q} \,\,\, ,$$
where $\lambda_{T0}$ is the turbulent mean free path at $v=v_{\rm te}$:
$$\begin{aligned}
\label{lambda_T}
\lambda_{T0} &=& \frac{3}{2 \pi \, C(q)} \, \lambda_{B} \, \left(\frac{B_{0}}{\delta B_{\perp}}\right)^{2} r_{\rm te}^{2-2q}\int _{0}^{1}d\mu \, (1-\mu^{2}) \, \mu^{1-2q} \cr
&=&\frac{3}{4 \pi \, C(q) (1-q)(2-q)} \,
\lambda_{B} \left(\frac{B_{0}}{\delta B_{\perp}}\right)^{2} r_{\rm te}^{2-2q} \,\,\, .\end{aligned}$$
For a typical spectrum, e.g., $2q=5/3$, $\lambda_{\rm T}\sim v^{2-2q}\sim v^{1/3}$, so that $\lambda_{\rm T}$ is only weakly dependent on velocity and hence on temperature ($\lambda_T \propto T_{e}^{1/6}$). Therefore, the case of a velocity-independent (or temperature-independent) mean free path (which corresponds precisely to the case $q=1$) deserves particular attention. However Equation (\[lambda\_T\]) shows that the turbulent mean free path formally diverges as $q \rightarrow 1$, which seems to suggest (Equation \[lambda-eff-c-t\])) that there is no reduction of transport below the collisional value in this case. In practice this divergence in $\lambda_T$ is avoided by considering the additional influence of collisions, with the result that the turbulent mean free path becomes finite and weakly (logarithmically) dependent on velocity [see below and the appendix in @2014ApJ...780..176K].
Returning to the result (\[df1dmu\]), which for an isobaric plasma can be written (cf. Equation (\[f1-basic\]))
$$\label{df1bdmu}
\frac{\partial f_1}{\partial \mu} = - \frac {v}{\nu (v)} \left [ \left ( \frac{m_e v^2}{2k_B T_e}-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz}+\frac{eE_\parallel}{k_B T_e} \right ] \, f_0 \,\,\, ,$$
we now have
$$\label{dmumu-r-mu-x}
\nu(v) = \frac{v_{\rm te}}{\lambda_{\rm ei}} \, \frac{1+R|\mu|^{2q-1}x^{2q+2}}{x^3} \,\,\, ,$$
where we have used the same normalization $x=v/v_{\rm te}$ and definition of the ratio $R = \lambda_{\rm ei}/\lambda_{0}$ as before. Substituting Equation (\[dmumu-r-mu-x\]) into Equation (\[df1bdmu\]), we obtain
$$\label{df1bdu-r-mu-x}
\frac{\partial f_{1}}{\partial \mu} = - \lambda_{\rm ei} \, \frac{x^4}{1+R|\mu|^{2q-1}x^{2q+2}} \, \left [ \left ( x^2-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz} + \frac{eE_\parallel}{k_B T_e} \right ] \, f_0 \,\,\, .$$
This is the generalization of Equation (\[f1-x-mu\]) to the case where $D^{T}_{\mu\mu}$ is proportional to $|\mu|^{2q-1}$ (Equations (\[dmumu-t-mu\]) and (\[alpha-beta-q\])).
Since, in the absence of collisions, the turbulent mean free path diverges for $q \geq 1$, the transport coefficients will also diverge for $q\geq1$ in the absence of collisions. We therefore focus on the physically-relevant $q < 1$ ($\alpha > 0$) cases. The singular case $q=1$ ($\alpha =0$) will be dealt with separately.
The case $q < 1$ {#smallq}
----------------
Except for the additional $\mu$ integral which produces a factor that depends on the value of the dimensionless spectral index $q$, there will be no difference in the scalings for large $R$ compared to those already given above for the case of isotropic scattering. Hence, while the necessary integrals can be evaluated in the large $R$ limit, it is simpler to replace the anisotropic scattering problem by an equivalent isotropic one, i.e., set
$$\label{nu-isotropic}
\nu(v) = \frac{v_{\rm te}}{\lambda_{\rm ei}} \, \frac{1+R x^{2q+2}}{x^3} \,\,\, ,$$
but with
$$R=\frac{\lambda_{\rm ei}}{\lambda_{T0}}$$
instead of $R=\lambda_{\rm ei}/{\lambda_{0}}$. This will lead to the previously given results of Section \[modified-spitzer\] with a somewhat different definition of $R$; these results are valid provided $q<1$, i.e., $\alpha>0$.
The case $q=1$ {#q1}
--------------
In the Lorentzian case $q=1$, corresponding to
$$\label{W-spectrum}
W(k_\parallel) = \frac{(\delta B_\perp)^2}{\pi} \, \frac{(1/\lambda_B)}{(1/\lambda_B)^2+k_\parallel^2} \,\,\, ,$$
we have
$$\label{nuT-Lorentzian}
\nu_T = |\mu| \left ( \frac{\delta B_\perp}{B_0} \right )^2 \frac{v}{\lambda_B} = \frac{|\mu| v}{\lambda_0} \,\,\, ,$$
so that
$$\label{Rm}
R = \frac{\lambda_{\rm ei}}{\lambda_{0}} = \left ( \frac{\lambda_{\rm ei}}{\lambda_B} \right ) \,
\left ( \frac{\delta B_\perp}{B_0} \right )^2 \simeq \frac{ 10^4 \, T_e^2 \, (\delta B_\perp/B_0)^2}{n_e \, \lambda_B} \,\,\, .$$
Equation (\[df1bdu-r-mu-x\]) can be analytically integrated over $\mu$ to give
$$\label{f1-r-mu-x}
f_1 = \begin{cases}
- \!\!\!\!\! &\frac{\lambda_{\rm ei}}{R} \, \ln (1+Rx^4\mu) \, \left [ \left ( x^2-\frac{5}{2} \right ) \frac{1}{T_e} \, \frac{dT_e}{dz} +
\frac{eE_\parallel}{k_B T_e} \right ] \, f_0 \quad ; \quad \mu > 0 \cr
& \frac{\lambda_{\rm ei}}{R} \, \ln (1- Rx^4 \mu)\, \left [ \left ( x^2-\frac{5}{2} \right ) \frac{1}{T_e} \frac{dT_e}{dz} +
\frac{eE_\parallel}{k_B T_e} \right ] \, f_0 \quad ; \quad \mu < 0 \,\,\, .
\end{cases}$$
This result can now be substituted in the expressions for the normalized heat flux (cf. Equation (\[qparallel-mu-v-f1\])) and the current density (Equation (\[jparallel-mu-v-f1\])). Exploiting the hemispherical antisymmetry of Equation (\[f1-r-mu-x\]) in evaluating the integral over $\mu$, we obtain
$$\label{qparallel-x-mu-r-01}
q_{\parallel} = - \, \frac{4 n_e k_B T_e v_{\rm Te} \lambda_{\rm ei}}{\sqrt{\pi} \, R} \int_0^\infty dx \, x^5 \left [ \left ( x^2-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz}+\frac{eE_\parallel}{k_B T_e} \right ] \, e^{-x^2} \, \int_0^1 d\mu \, \mu \ln ( 1+Rx^4 \mu ) \,\,\,$$
and
$$\label{jparallel-x-mu-r-01}
j_{\parallel} = \frac{4 n_e \, e \, v_{\rm Te} \lambda_{\rm ei}}{\sqrt{\pi} \, R} \int_0^\infty dx \, x^3 \left [ \left ( x^2-\frac{5}{2} \right )
\frac{1}{T_e} \, \frac{dT_e}{dz} + \frac{eE_\parallel}{k_B T_e} \right ] \, e^{-x^2} \, \int_0^1 d\mu \, \mu \, \ln ( 1+Rx^4 \mu ) \,\,\, .$$
Making the substitution $y=1+Rx^4 \mu$ and then integrating by parts, the integral over $\mu$ evaluates to
$$\label{integral-evaluation}
\int_0^1 d\mu \, \mu \ln ( 1+Rx^4 \mu ) = \frac{1}{2 (Rx^4)^2} \left \{ \left [ (Rx^4)^2 - 1 \right ] \ln (1 + Rx^4) + Rx^4 \left ( 1 - \frac{1}{2} Rx^4 \right ) \right \} \,\,\, .$$
The values of the thermoelectric transport coefficients $\kappa_\parallel, \alpha_\parallel$, $\beta_\parallel$ and $\sigma_\parallel$ may now be obtained by substituting the expression (\[integral-evaluation\]) for the $\mu$ integral in the expressions (\[qparallel-x-mu-r-01\]) and (\[jparallel-x-mu-r-01\]). These transport coefficients are best represented in relation to their Spitzer ($R=0$) values. We find
$$\label{kappa-ratio-general}
\frac{\kappa_{\parallel}}{\kappa_{\parallel, S}} = \frac{3}{2R} \, \, \frac{\int_0^\infty \frac{1}{(Rx^4)^2} \left \{ \left [ (Rx^4)^2 - 1 \right ] \ln (1 + Rx^4) + Rx^4 \left ( 1 - \frac{1}{2} Rx^4 \right ) \right \} \, x^5 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx}{\int_0^\infty x^9 \, \left ( x^2 - \frac{5}{2} \right ) \, e^{-x^2} \, dx} \,\,\, ;$$
$$\label{alpha-ratio-general}
\frac{\alpha_{\parallel}}{\alpha_{\parallel,S}} = \frac{3}{2R} \, \frac{\int_0^\infty \frac{1}{(Rx^4)^2} \left \{ \left [ (Rx^4)^2 - 1 \right ] \ln (1 + Rx^4) + Rx^4 \left ( 1 - \frac{1}{2} Rx^4 \right ) \right \} \, x^5 \, e^{-x^2} \, dx}{\int_0^\infty x^9 \, e^{-x^2} \, dx} \,\,\, ;$$
$$\label{beta-ratio-general}
\frac{\beta_{\parallel}}{\beta_{\parallel,S}} = \frac{3}{2R} \, \frac{\int_0^\infty \frac{1}{(Rx^4)^2} \left \{ \left [ (Rx^4)^2 - 1 \right ] \ln (1 + Rx^4) + Rx^4 \left ( 1 - \frac{1}{2} Rx^4 \right ) \right \} \, x^3 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx}{\int_0^\infty x^7 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx} \,\,\, ;$$
and
$$\label{sigma-ratio-general}
\frac{\sigma_{\parallel}}{\sigma_{\parallel,S}} = \frac{3}{2R} \, \frac{\int_0^\infty \frac{1}{(Rx^4)^2} \left \{ \left [ (Rx^4)^2 - 1 \right ] \ln (1 + Rx^4) + Rx^4 \left ( 1 - \frac{1}{2} Rx^4 \right ) \right \} \, x^3 \, e^{-x^2} \, dx}{\int_0^\infty x^7 \, e^{-x^2} \, dx} \,\,\, .$$
For $R \ll 1$, we have a collision-dominated regime, and the results (\[kappa-ratio-general\]) through (\[sigma-ratio-general\]) should approach unity. Indeed, for $R \ll 1$, the logarithm may be expanded in a Taylor series $\ln (1+q) = q - q^2/2 + q^3/3 - \ldots$, giving
$$\label{int-r-small}
\frac{1}{2 (Rx^4)^2} \left \{ \left [ (Rx^4)^2 - 1 \right ] \ln (1 + Rx^4) + Rx^4 \left ( 1 - \frac{1}{2} Rx^4 \right ) \right \} \, {\overset{R \ll 1}\longrightarrow} \,\, \frac{1}{3} \, R x^4$$
in place of Equation (\[integral-evaluation\]). (This result may also be obtained quickly by expanding the logarithm to first order in the integral $\int_0^1 d\mu \, \mu \ln ( 1+Rx^4 \mu )$.) Using the limiting form in the numerators of (\[kappa-ratio-general\]) through (\[sigma-ratio-general\]) yields unity for all four ratios, as it should.
On the other hand, for the turbulence-dominated transport regime characterized by $R \gg 1$, we obtain
$$\label{kappa-ratio-rggg1}
\frac{\kappa_{\parallel}}{\kappa_{\parallel, S}} \, {\overset{R \gg 1}\longrightarrow} \,\, \frac{3 \int_0^\infty \, x^5 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^{2}} \, dx }{2 \int_0^\infty x^9 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx} \, \left ( \frac{\ln R}{R} \right ) = \frac{1}{40} \, \left ( \frac{\ln R}{R} \right ) \,\,\, ;$$
$$\label{alpha-ratio-rggg1}
\frac{\alpha_{\parallel}}{\alpha_{\parallel, S}} \, {\overset{R \gg 1}\longrightarrow} \,\, \frac{3 \int_0^\infty \, x^5 \, e^{-x^{2}} \, dx }{2 \int_0^\infty x^9 \, e^{-x^2} \, dx} \, \left ( \frac{\ln R}{R} \right ) = \frac{1}{8} \, \left ( \frac{\ln R}{R} \right ) \,\,\, ;$$
$$\label{beta-ratio-rggg1}
\frac{\beta_{\parallel}}{\beta_{\parallel, S}} \, {\overset{R \gg 1}\longrightarrow} \,\, \frac{3 \int_0^\infty \, x^3 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^{2}} \, dx }{2 \int_0^\infty x^7 \, \left ( x^2-\frac{5}{2} \right ) \, e^{-x^2} \, dx} \, \left ( \frac{\ln R}{R} \right ) = - \frac{1}{12} \, \left ( \frac{\ln R}{R} \right ) \,\,\, ;$$
and
$$\label{sigma-ratio-rggg1}
\frac{\sigma_{\parallel}}{\sigma_{\parallel, S}} \, {\overset{R \gg 1}\longrightarrow} \,\, \frac{3 \int_0^\infty \, x^3 \, e^{-x^{2}} \, dx }{2 \int_0^\infty x^7 \, e^{-x^2} \, dx} \, \left ( \frac{\ln R}{R} \right ) = \frac{1}{4} \, \left ( \frac{\ln R}{R} \right ) \,\,\, .$$
Note both the $\ln R/R$ dependence of these expressions and also the change of sign of $\beta_\parallel$ at large $R$ (which occurs for the same reasons given in the discussion at the end of Section \[modified-spitzer\]).
Application to solar flares {#application}
===========================
Heating and cooling of flare coronal plasma {#coronal-heating-cooling}
-------------------------------------------
A solar flare is ubiquitously characterized by enhanced emission in soft X-rays [e.g., @1969MNRAS.144..375C; @1970MNRAS.151..141C; @1982SoPh...78..107A]; indeed, the commonly-used GOES classification of flares is based on the soft X-ray flux in the $(1-8)$Å soft X-ray waveband. The plasma responsible for emitting these soft X-rays is generally located in the corona and has a temperature ${\, \lower3pt\hbox{$\sim$}\llap{\raise2pt\hbox{$>$}}\,}10^7$ K [e.g., @2011ApJ...727L..52R]. The [*in situ*]{} heating [@1983ApJ...266..383C; @2004ApJ...609..439F; @2007ApJ...659..750R] to these temperatures is generally believed to be due to a combination of ohmic heating associated with current dissipation in the primary site of magnetic reconnection, plus collisional heating by accelerated nonthermal particles, notably electrons. Ohmic heating by passage of the beam-neutralizing return current (see Section \[return-current\]) through the flaring corona may also play a role.
The heat energy transported by both nonthermal electrons and thermal conduction [@2003LNP...612..161H] to the chromosphere causes local heating and a corresponding increase in the emission measure of $10^7$ K gas, enhancing the overall soft X-ray emission. Further, the pressure gradients established by this rapid heating of chromospheric material through the region of radiative instability from $T\simeq 10^5$ K to $T \simeq 10^7$ K [@1969ApJ...157.1157C], drive a significant hydrodynamic response of the solar atmosphere [e.g., @1984ApJ...279..896N; @1989ApJ...341.1067M; @2005ApJ...630..573A], in particular an upward motion of soft-X-ray-emitting plasma into the corona, a process somewhat incorrectly, but nevertheless ubiquitously, termed “chromospheric evaporation.”
The hot $10^7$ K plasma thus created subsequently cools through radiation, conduction to the chromosphere, and flow of enthalpy. Estimating the pertinent cooling times shows that, under the assumption of classical, collisional [@1962pfig.book.....S] conductivity, thermal conduction generally dominates the cooling time at the highest temperatures. However, in many events the thermal plasma is sustained well beyond the duration of the impulsive hard X-ray burst (and hence heating by nonthermal electrons), for times much longer than the conductive cooling time. This has led to the suggestion [e.g., @1980sfsl.work..341M; @2012ApJ...759...71E] that energy is somehow injected, by unknown processes, into the corona after the impulsive phase has ceased. However, it is important to realize that this conclusion is based on thermal conductive losses governed by a purely collisional model of heat transport, which may not be valid if, as suggested above, turbulent processes play a significant role in the region of electron transport.
To illustrate, let us consider a flare volume of $V \simeq 10^{27}$ cm$^3$, with a plasma of density $n_e = 10^{10}$ cm$^{-3}$ embedded in a magnetic field $B \simeq 10^3$ Gauss. The magnetic energy density is $B^{2}/8\pi \simeq 10^{5}$ erg cm$^{-3}$ and the total available magnetic energy is $(B^{2}/8\pi) V \simeq 10^{32}$ erg, which is a good estimate of the amount of energy released by magnetic reconnection in a large flare [e.g., @2012ApJ...759...71E]. Taking the flare duration as $\tau \simeq 10^2$ s, this corresponds to a released power of $10^{30}$ erg s$^{-1}$. This energy is transformed into kinetic energy of the particles, thermal and non thermal. A fraction, say $10\%$, of this energy goes into the kinetic energy of non-thermal electrons [@2012ApJ...759...71E], giving a power $10^{29}$ erg s$^{-1}$ in accelerated electrons, a number close to that inferred from observations of hard X-ray emission in flares [e.g., @2011SSRv..159..107H]. This corresponds to a volumetric heating rate due to fast electrons $Q \simeq 100$ erg cm$^{-3}$ s$^{-1}$.
With these parameters established, we can now consider the heating (by fast electrons) and cooling (by conduction) of the soft X-ray emitting plasma, governed by the energy equation
$$\label{energy-balance}
\frac{\partial (n_e k_B T_e)}{\partial t} = - \frac{\partial q_\parallel}{\partial z} + Q \,\,\, .$$
As we have seen above, the conductive heat flux takes the form
$$\label{qparallel-energy}
q_{\parallel} = - \frac{2 n_e k_B \, (2 k_B T_e)^{1/2}}{m_e^{1/2}} \, \lambda \, \frac{dT_e}{dz} \,\,\, ,$$
where $\lambda$ is the mean free path, representing the combined effect of Coulomb collisions and turbulent scattering,
$$\lambda= \frac{\lambda_{\rm ei}}{1+R} \,\,\,,$$
and where the turbulent reduction factor
$$R =\frac{ \lambda_{ei}}{\lambda_0}=\frac{ 10^4 \, T_e^2 \, (\delta B_\perp/B_0)^2}{n_e \, \lambda_B} \,\,\, .$$
When $R\gg1$, i.e., when the turbulent parameter $\lambda_{0}$ is much smaller than the collisional mean free path $\lambda_{\rm ei}$ (corresponding to $ \lambda_{0}\ll 10^{4}T_{e}^{2}/n_{e}$), heat transport is regulated predominantly by turbulence. We call this a “turbulence-dominated regime,” although we again stress (cf. Section \[scattering-parameters\]) that the role of collisions in maintaining a near Maxwellian distribution of background electrons remains important. On the other hand, turbulence plays a negligible role when $R \ll 1$.
Let us first balance the terms on the right side of Equation (\[energy-balance\]) to obtain an estimate of the steady-state temperature in the presence of a heat source $Q$ balanced by thermal conduction:
$$2 n_e k_B \left ( \frac{2 k_B T_e}{m_e} \right )^{1/2} \lambda \, \left ( \frac{T_e}{L^2} \right ) \simeq Q \,\,\, ,$$
from which follows the scaling law
$$\label{scaling-law}
T_e \simeq \frac{m_e^{1/3}}{2 k_B} \, \left ( \frac{Q L^{2}}{n_e\lambda} \right )^{2/3} \,\,\, .$$
The vast majority of theoretical studies concerning the response of coronal loops to heating use the collisional [@1962pfig.book.....S] conductivity and corresponding mean free path $\lambda_{\rm ei}$; this includes the determination of equilibrium scaling laws and temperature distribution for active region loops and the modelling of evaporative cooling and enthalpy-based response to coronal heating [e.g., @1978ApJ...220.1137A; @1995ApJ...439.1034C; @2010LRSP....7....5R; @2012ApJ...758....5C]. When the mean free path is indeed close to its collisional value $\lambda_{ei}$ (Equation (\[lambda-ei\])), Equation (\[scaling-law\]) gives the scaling law appropriate to a collisional regime of transport [@1978ApJ...220..643R] :
$$\label{scaling-law-collisional}
T_e \simeq \frac{m_e^{1/7}}{2 k_B} \, \left ( 2 \pi e^4 \, \ln \Lambda \right )^{2/7} Q^{2/7} \, L^{4/7} \simeq 50 \, Q^{2/7} \, L^{4/7} \,\,\, .$$
Note that this is independent on the density $n_e$, and that this scaling follows straightforwardly from the familiar balancing relation $Q \propto T_e^{7/2}/L^2$ appropriate to a collisional environment. Substituting the values $Q \simeq 100$ erg cm$^{-3}$ s$^{-1}$ and $L \simeq 10^9$ cm, we obtain
$$\label{scaling-law-collisional-result}
T_e \simeq 3 \times 10^7 \, {\rm K} \,\,\, ,$$
a value nicely consistent with soft X-ray emission.
On the other hand, returning to Equation (\[scaling-law\]) with $R\gg1$, we now obtain the fundamentally different scaling law for the equilibrium temperature in a turbulence dominated regime:
$$\label{scaling-law-turbulent}
T_e \simeq \frac{m_e^{1/3}}{2 k_B} \, \left ( \frac{Q}{n_e} \right )^{2/3}L^{4/3}\lambda_{B}^{-2/3}
\left (\frac{\delta B_{\perp}}{B_{0}} \right)^{4/3} \,\,\, ,$$
which now depends on the density $n_e$. With a magnetic field perturbation ratio $\delta B_\perp /B_0 \simeq 0.1$ and a magnetic correlation length $\lambda_B \simeq 10^6$ cm ($\lambda_0 \simeq 10^8$ cm) gives the significantly larger temperature
$$\label{scaling-law-turbulence-result}
T_e \simeq 1 \times 10^8 \, {\rm K} \,\,\, .$$
Hence, for a given heating rate $Q$ and loop properties $n_e$ and $L$, the turbulent suppression of heat transport associated with large values of $R$ leads to a higher steady-state temperature than that obtained by using Spitzer conductivity. Plasma at temperatures $\sim$$10^8$ K ($\simeq 10$ keV) can make a meaningful contribution to the [*hard*]{} X-ray emission from the flare. Since thermal conduction is inhibited, possibly also by other collective plasma processes [@1979ApJ...228..592B; @1980ApJ...242..799S], such $10^8$ K temperatures will likely be confined rather than extending along the entire magnetic loop. For a given heating rate $Q$, Equations (\[scaling-law-collisional\]) and (\[scaling-law-turbulent\]) give quite different dependencies of the flaring coronal temperature $T_e$ on the loop length $L$ ($L^{4/7}$ and $L^{4/3}$, respectively), a result that should be observationally testable.
We note that the steady state temperature $T_{e}=3\times 10^{7}$ K in Equation (\[scaling-law-collisional-result\]) above was obtained by assuming a collisional transport regime. Such an assumption is valid only under the dual conditions that both the Knudsen number ${\rm Kn} = \lambda_{\rm ei}/L_T$ and the turbulent reduction factor $R$ are $\ll 1$. Consistency therefore demands that we evaluate the validity of these two assumptions. We find that at these temperatures (and densities $n_e \sim 10^{10}$ cm$^{-3}$), the collisional mean free path $\lambda_{\rm ei}$ is actually of the order of the loop length $L \simeq 10^9$ cm; thus the Knudsen number is of order unity and so the use of a collision-related expression for the heat flux is somewhat questionable. In such conditions the heat flux is instead determined [e.g.. @1979ApJ...228..592B] by a flux-limited value equal to a fraction of the free-streaming limit (Equation (\[qparallel\])), leading to a generally higher equilibrium temperature and conductive cooling time than is appropriate for the collisional case. With regard to the second assumption, the main thrust of the present paper is that, whether or not the Knudsen number is small, the heat flux may be further limited (by a factor $\simeq R$) by the presence of collisionless pitch-angle scattering. Thus, if the turbulent mean free path is substantially less than the collisional mean free path (or, for Knudsen numbers of order unity or more, the length of the flaring loop), then the much larger temperatures ($T_{e}\sim 10^{8}$ K; Equation (\[scaling-law-turbulence-result\])) appropriate to a turbulence-dominated regime apply.
We now turn to a consideration of the role of conduction in cooling the coronal plasma after the energy input $Q$ has ceased. Balancing the heating and conductive cooling in Equation (\[energy-balance\])), and using the expression (\[qparallel-energy\]), we obtain an expression for the cooling time-scale $\tau_{c}$:
$$\label{cooling-time}
\tau_{c} \simeq \frac{m_e^{1/2}}{(2 k_B T_e)^{1/2}} \, \frac{L^2}{\lambda} =
\left ( \frac{L}{v_{\rm te}} \right ) \, \left ( \frac{L}{\lambda} \right ) = \left ( \frac{L}{v_{\rm te}} \right ) \, \left ( \frac{L}{\lambda_{\rm ei}} \right ) \, (1+R) \,\,\, .$$
The cooling time $\tau_{c}$ is thus the free-streaming transport time-scale $L/v_{\rm te}$ of thermal electrons multiplied by the inverse of the Knudsen number ${\rm Kn} = \lambda/L$. For a temperature $T_{e} \simeq 3 \times 10^7$ K (Equation (\[scaling-law-collisional-result\])), the electron thermal speed $v_{\rm Te} \simeq 3 \times 10^9$ cm s$^{-1}$, so that for a loop length $L \simeq 10^9$ cm, the free-streaming escape time is $L/v_{\rm te} \simeq$ 0.3 s. The collisional mean free path $\lambda_{\rm ei}\simeq 10^9$ cm and hence $L/\lambda_{ei}\sim 1$, which when $R \ll 1$, yields a cooling time of the order of the free-streaming time scale. A smaller initial temperature $T_{e} \simeq 10^7$ K has the collisional mean free path decreased by an order of magnitude giving a cooling time $\tau_{c}\sim 3s$. These time scales are much shorter than the duration of the soft X-ray emission, which has led to the realization that the coronal plasma will cool very rapidly after the cessation of the energy input term $Q$, to the point where some form of post-impulsive-phase energy input to the corona is needed to sustain the soft X-ray emission for the observed times ${\, \lower3pt\hbox{$\sim$}\llap{\raise2pt\hbox{$>$}}\,}100$ s [see, e.g., @1980sfsl.work..341M; @2012ApJ...759...71E]. However, in the presence of turbulence the cooling time $\tau_{\rm c}$ is further enhanced by a factor $(1+R)$, and so cooling by thermal conduction parallel to the guiding magnetic field can thus be significantly inhibited. This increase in the cooling time significantly reduces the previously-assumed requirement [@1980sfsl.work..341M; @2012ApJ...759...71E] for post-impulsive phase heating of the coronal plasma.
Before moving on, we parenthetically note that similar issues regarding thermal conductivity (and its suppression) in stochastic magnetic fields arise in the context of galaxy cluster formation and in the theory of cooling flows .
Return current effects {#return-current}
----------------------
The very significant electrical currents associated with the injection of non-thermal electrons necessitate a current-neutralizing return current, set up by a combination of electrostatic and inductive processes, the relative role of which has been a matter of some debate . However, irrespective of the detailed physics responsible for establishing this return current, driving it through the finite resistivity of the ambient medium requires a local electric field ${\cal E}_\parallel = j_\parallel/\sigma_\parallel$, which in turn causes both an Ohmic heating rate $Q_{\rm rc} = j_\parallel \, {\cal E}_\parallel = j_\parallel^2/\sigma_\parallel$ and an additional energy loss rate $| dE/dt | = e \, {\cal E}_\parallel \, v = j_\parallel \, {\cal E}_\parallel/n_e = j_\parallel^2/n_e \, \sigma_\parallel$ for each of the accelerated electrons.
Now, the transport of [*non-thermal*]{} electrons is dominated by non-diffusive cold-target energy losses [see, e.g., @1971SoPh...18..489B; @1973SoPh...31..143B; @1978ApJ...224..241E], and hence we expect that the transport of such electrons, and hence the current density $j_\parallel$ that they carry, is largely unaffected by collisionless pitch-angle scattering. (Although @2014ApJ...780..176K have shown that the direct beam current $j_\parallel$ is also reduced somewhat due to the presence of pitch-angle scattering, the reduction factor is not as large as the transport coefficient reduction factors $R$ considered here, so that we may assume that the current density $j_\parallel$ associated with the injected electrons is essentially the same as in the purely collisional case.) Any change in ohmic energy losses is therefore driven primarily by changes in the parallel electrical conductivity $\sigma_\parallel$. Reducing the value of $\sigma_\parallel$ through turbulence results in a greater rate of Ohmic heating $Q_{\rm rc}$ (and hence higher coronal heating rates) and also a greater energy loss rate $|dE/dt|$ for the accelerated electrons.
This enhancement of the return-current electron energy loss rates affects the heating rate as a function of position [@1980ApJ...235.1055E] and hence the hydrodynamic response of the atmosphere [e.g., @1984ApJ...279..896N; @1989ApJ...341.1067M]. It also reduces the amount of energy precipitating into the chromospheric footpoints, thus possibly accounting for the “gentle” evaporation observed by, e.g., @1988ApJ...329..456Z. Enhanced return current energy losses also result in a more effective confinement of hard-X-ray-producing electrons in the corona, which may offer an alternative explanation for loop-top coronal sources .
Estimates of the ratio of return current heating to collisional energy loss in the flaring corona show that, for moderately large flares they are comparable [see Figure 3 of @1980ApJ...235.1055E]. The same figure shows that the ratio of return current heating to collisional energy loss in the chromosphere, where most of the electron heating occurs, can be up to several percent. Thus, enhancing the return current heating/energy loss rate by even an order of magnitude through turbulent modification of the electrical conductivity $\sigma_\parallel$ and could possibly transform the flaring corona into a return-current-dominated regime . This has very significant implications, ranging from the spatial distribution of hard X-ray emission and electron heating, to the total number of accelerated electrons required to produce a given hard X-ray intensity .
A more dominant role for return current losses in the energy loss rate for accelerated electrons has a possibly even more interesting effect. Since the energy loss rate for an individual electron $| dE/dt | = e \, {\cal E}_\parallel \, v = e \, v \, j_\parallel/\sigma_\parallel$, which is proportional to the injected current $j_\parallel$ and hence the electron injection rate, and since the total hard X-ray yield is proportional to the injection rate divided by the energy loss rate [@1988ApJ...331..554B], it follows that the hard X-ray yield in a return-current-loss dominated regime [*is independent of the injected number of electrons*]{} [@1980ApJ...235.1055E]. Such a possible saturation of hard X-ray flux with increasing flare intensity has been reported by @2007ApJ...666.1268A.
Summary and conclusions {#summary}
=======================
Motivated by observations suggesting the presence of magnetic fluctuations in flaring loops and also suggesting that turbulent pitch-angle scattering plays a significant role in the transport of energy by both thermal and non-thermal electrons in solar flares, we have derived formulae for the thermal and electrical conductivities in the presence of both collisions and magnetic turbulence.
The enhanced electron confinement effected by the addition of collisionless pitch-angle scattering can reduce the [*thermal*]{} conductivity of the corona, thus decreasing thermal conductive losses and so increasing coronal temperatures compared to those in a model with collisionally-dominated transport. This may explain the localization of coronal X-ray sources in the apex of the loop [see, e.g., @2011SSRv..159..107H for a review]. It also increases the cooling time for the flare-heated coronal plasma, possibly alleviating the need for post-impulsive-phase heating by unidentified processes [@1980sfsl.work..341M; @2012ApJ...759...71E]. Finally, it means that the corona becomes more of a “warm” target in the calculation of the energy loss rate of accelerated electrons, which has an impact on the relationship between the source-integrated electron spectrum and the injected spectrum [@1988ApJ...331..554B; @2003ApJ...595L.115B; @2003ApJ...595L.119E; @2011SSRv..159..301K; @2015ApJ...809...35K] and hence on the overall energetics associated with accelerated electrons [@1986NASACP...2439..505D; @1997JGR...10214631M; @2003ApJ...595L..97H; @2004JGRA..10910104E; @2005JGRA..11011103E; @2012ApJ...759...71E].
The suppressed value of the [*electrical*]{} conductivity may significantly increase the importance of ohmic heating, both in the thermodynamics of the flare-heated atmosphere and in the propagation of the accelerated electrons themselves. In particular, because the inclusion of return current energy losses affects the “bremsstrahlung efficiency” (energy of hard X-rays produced per electron energy injected in the corona), it may significantly alter the injected electron flux required to produce given hard X-ray flux, with further attendant implications for the overall role of accelerated electrons in flare energetics.
Because of these important implications for quantitative details of the impulsive phase of solar flares and even for its overall viability , we urge workers in the field to consider such anomalous transport effects in their modeling of particle transport, thermal conduction, and the electrodynamics of solar flares.
This work is partially supported by a STFC consolidated grant. Financial support by the European Commission through the “Radiosun” (PEOPLE-2011-IRSES-295272) is gratefully acknowledged. AGE was supported by grant NNX10AT78G from NASA’s Goddard Space Flight Center.
[^1]: This approximation is, for the Coulomb interaction, equivalent to the Lorentz gas approximation which has immobile heavy hard spheres as scattering agents, as in the Drude-Lorentz model of electric conductivity, hence the name “Lorentz plasma” [e.g., @1964PhFl....7..407K], see also [@1966JETP...23..145R] for the analogy between low-frequency electrostatic turbulence and the Lorentz gaz
[^2]: Note that only one term arises in evaluating the velocity-space derivative $\partial f_0(z,v)/\partial v$ that prefixes the electric field $E_\parallel$ term in the expression for $f_1(z,v,\mu)$. Thus only one velocity-moment term appears in each of the expressions for $\alpha_\parallel$ and $\sigma_\parallel$, and the signs of these two thermoelectric coefficients are therefore fixed.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Any nonpositively curved symmetric space admits a topological compactification, namely the Hadamard compactification. For rank $1$ spaces, this topological compactification can be endowed with a differentiable structure such that the action of the isometry group is differentiable. Moreover, the restriction of the action on the boundary leads to a flat model for some geometry (conformal, CR or quaternionic CR depending of the space). One can ask whether such a differentiable compactification exists for higher rank spaces, hopefully leading to some knew geometry to explore. In this paper we answer negatively.'
author:
- Benoît Kloeckner
bibliography:
- 'biblio.bib'
nocite: '[@Serre; @Vinberg; @Knapp; @Guivarch]'
title: Symmetric spaces of higher rank do not admit differentiable compactifications
---
Introduction
============
Let $M$ be a symmetric space of nonpositive curvature, $G$ its group of isometries and $G_0$ the identity component in $G$.
As a Riemannian manifold, $M$ is a Hadamard space and is diffeomorphic to an open ball. Its (or ) is a topological gluing of $M$ and its $M(\infty)$ such that $\adherence{M}=M\cup M(\infty)$ is a closed ball. The group $G$ acts continuously on $\adherence{M}$.
When $M$ is of rank one, that is to say when it is negatively curved, this topological compactification admits “nice” models, carrying an invariant differentiable structure: in these models, the action of $G$ is differentiable on $\adherence{M}$. Moreover, the restriction of this action to the boundary is a flat model for some geometry. The boundary spheres of the real, complex and quaternionic hyperbolic spaces wield the standard conformal, CR and quaternionic CR structures respectively. Concerning the octonionic hyperbolic plane, the corresponding geometry has not been studied yet, as far as we know.
It is natural to ask whether such a differentiable compactification exists when $M$ is of higher rank. One could expect such a model to give birth to a new, luckily interesting, geometry.
In this paper, we give a negative answer to this question, and show that the obstruction comes from the spherical building at infinity. This combinatorial structure is trivial only in the Euclidean and rank one spaces. Thus there is an alternative: a symmetric space of nonpositive curvature admits either an interesting building at infinity or a differentiable compactification, not both.
Differentiable compactifications
--------------------------------
Our goal is to extend the differentiable structure of $M$ to the manifold with boundary $\adherence{M}$, so that we do not lose symmetry in the process. This leads to the following definition.
\[DefComp\] A of $M$ is a differentiable ($\diffb{1}$) structure $\ron{D}$ on $\adherence{M}$ compatible with the differentiable structure of $M$ and such that the action of $G$ is .
We can define a in the same way, where $r$ can be finite, $\infty$ or $\omega$, $\diffb{\omega}$ meaning real analytic. When no precision is given, differentiable means .
By a $\diffb{r}$ action, we mean that the map $G\times M\to M$ is $\diffb{r}$. It implies that $G$ acts by $\diffb{r}$ diffeomorphisms and that the map $G\to\diff{M}$ is continuous in the $\diffb{r}$ topology.
This condition can be greatly relaxed thanks to the Bochner and Montgomery theorem: if $G$ acts continuously by $\diffb{r}$ diffeomorphisms, then its action is in fact $\diffb{r}$ [@Bochner-Montgomery].
For the sake of brevity we will often write “differentiable compactification” instead of “differentiable Hadamard compactification”. However, a symmetric space admits other topological compactifications than the Hadamard one, e.g. Martin, Satake and Furstenberg compactifications. It would be interesting to extend our study to these, but the Hadamard compactification seems to be of utmost importance for our question. First, it is very natural, defined directly by the geometry of the space for a large class of Riemannian manifolds. Second, there is as far as we know little hope to get a manifold with boundary from the other compactifications. Either the infinity does not have the right dimension (e.g. the Poisson boundary) or the most natural differentiable structure is that of a manifold with boundary *and corners* (e.g. the maximal Satake compactification). A detailed account on all classical compactifications can be found in [@Borel-Ji] and [@Guivarch].
Existence of differentiable compactifications
---------------------------------------------
Let us now discuss the existence of differentiable compactifications for the three types of nonpositively curved symmetric spaces.
#### Symmetric spaces of rank 1.
It is well known that the real hyperbolic space $\mH^n$ admits a differentiable Hadamard compactification, given for example by the closure of Klein’s ball: the central projection of the hyperboloid $Q=-1$ (where $Q$ is the canonical Lorentzian metric on $\mR^{n+1}$) gives an embedding of $\mH^n$ into $\mR\mP^n$ where the group $\SOopq{1}{n}$ of isometries of $\mH^n$ acts analytically. This construction can be generalized to all symmetric spaces of nonpositive curvature and rank $1$.
It is worth noticing that $\mH^n$ admits other differentiable Hadamard compactifications. For example, the action of $\SOopq{1}{n}$ on Poincaré’s ball extends analytically to the closed ball and the resulting action is not conjugate to the previous one (this can be seen by looking at asymptotic geodesics: they are tangent one to another in the closure of Poincaré’s ball, not in Klein’s ball.) Details are given in [@Kloeckner2], where it is shown that $\mH^n$ admits an infinite number of nonconjugate analytic compactifications in the sense of definition \[DefComp\].
#### Euclidean spaces.
If $M$ is a Euclidean space, once again it admits a differentiable Hadamard compactification we briefly describe. Identify $\mR^n$ with the affine hyperplane $\{x_0=1\}$ of $\mR^{n+1}$ where $n$ is the dimension of $M$. The projection of center $0$ of $M$ on the open upper unit half-sphere is a diffeomorphism. Pushing forward by this map we get an action of $G$ (the affine group) on the open upper half-sphere whose continuous prolongation to the closed half-sphere is real analytic. This action is a real analytic Hadamard compactification of $M=\mR^n$.
#### Symmetric spaces of higher rank.
The main result of this paper is the following.
\[theoreme\] No noneuclidean symmetric space of rank $k\geqslant 2$ admits a differentiable Hadamard compactification.
#### Structure of the paper.
From now on, $M$ is supposed to be a symmetric space of rank $k\geqslant 2$.
We shall start with a simple remark about the natural projection of a fiber $S_xM$ of the unit tangent bundle of $M$ on $M(\infty)$.
In the second section we prove that $\mH^2\times\mR$ admits no differentiable Hadamard compactification.
Next we generalize this fact to every product $F\times\mR^{k-1}$ where $k\geqslant 2$ and $F$ is a symmetric space of noncompact type of rank $1$.
Finally, we prove Theorem \[theoreme\].
Note that the different parts are more or less independent: the proof of Theorem \[theoreme\] does not make use of preceding results. However some arguments of Section \[FxRn\] will be useful, and Section \[HxR\] gives a good insight of the general phenomenon on the simplest case.
Apartments and the visual projection
====================================
Apartments
----------
We give some basic vocabulary about the building structure of $M(\infty)$. More details can be found in [@BallmannGromov], Appendix 5. Our main reference for the building structure of a symmetric space is [@Eberlein]. For details about general buildings, see [@Brown].
Let $A$ be a maximal flat ( a totally geodesic submanifold isometric to the Euclidean space of maximal dimension) of $M$, $\adherence{A}$ its closure in $\adherence{M}$ and $A(\infty)=\adherence{A}\cap M(\infty)$ its boundary. $A(\infty)$ is called an of $M(\infty)$. It is a topological submanifold.
Every point of $M(\infty)$ belongs to at least one apartment. A point is said to be if it belongs to exactly one apartment, otherwise it is said to be .
Let $x$ be a point of $M(\infty)$. We denote by $a(x)$ the set of all apartments containing $x$. If $x$ is singular, it is said to have $1$ if $a(x)$ is minimal with respect to inclusion among sets $a(y)$ of singular $y$’s.
The connected component of $x$ in the set of points $y$ such that $a(x)=a(y)$ is a . Facets are topological submanifolds. If $x$ is regular, we call its facet a or simply a ; if $x$ is singular of index $1$, we call its facet a .
The dimension of every apartment is $k-1$ (where $k$ is the rank of $M$). Chambers have dimension $k-1$, panels have dimension $k-2$.
Two facets are if their closures intersect. If they are adjacent and of different dimensions, one is contained in the closure of the other.
The facets form a simplicial complex on $M(\infty)$ if $M$ is of noncompact type. If $M$ has a Euclidean factor, some of the cells are spheres rather than simplicies.
This complex has the incidence structure of a spherical thick building, which means:
1. each apartment is a spherical Coxeter complex (see [@Brown] for details),
2. for any two facets, there is an apartment containing both of them,
3. there exists at least three chambers adjacent to any given panel,
4. if there are two apartments $A$, $A'$ containing two facets $F$ and $F'$, then there is an isomorphism $A\mapsto A'$ fixing $F$ and $F'$ pointwise.
The group $G$ acts by isomorphisms on this building: it preserves the adjacency relation and sends facets onto facets of the same dimension.
Non smoothness of the visual projection {#VP}
---------------------------------------
Let $x$ be a point of $M$. A unit vector $v$ tangent to $M$ at $x$ defines a geodesic ray $\gamma_v$, hence a point $\gamma_v(\infty)$ of the Hadamard boundary $M(\infty)$. The map $$\pi_x: \begin{array}{rcl}
S_xM &\to & M(\infty) \\
v &\mapsto & \gamma_v(\infty) \end{array}$$ is called the from the point $x$.
For all $x$, the visual projection from $x$ is a homeomorphism. It seems reasonable to expect the visual projections to be diffeomorphisms for a “good” differentiable Hadamard compactification. However, it cannot be.
\[RemarqueSimple\] If $M$ is a nonpositively curved symmetric space of higher rank, there is no differentiable structure on $M(\infty)$ such that all apartments are submanifolds.
Let $\gamma$ be some geodesic ray in $M$ that is singular of index $1$. As the spherical building at infinity of $M$ is thick, there are at least three (in fact, an infinite number of) chambers $C_1$, $C_2$, $C_3$ adjacent to $P$.
For each pair $C_i,C_j$ ($i\neq j$) there is a flat $A_{ij}$ such that $C_i\subseteq
A_{ij}(\infty)$ and $C_j\subseteq A_{ij}(\infty)$. But $A_{ij}(\infty)$ is an embedded submanifold of $M(\infty)$, thus $C_i$ and $C_j$ have opposite tangent half spaces $E_i$, $E_j$ at $\gamma(\infty)$. See figure \[RS1\].
![Three chambers meeting at a panel.[]{data-label="RS1"}](RangSup13.eps){width="90mm"}
Thus we get three half subspaces $E_1$, $E_2$, $E_3$ of $T_{\gamma(\infty)}M(\infty)$ such that $E_1=-E_2$, $E_1=-E_3$, $E_2=-E_3$, a contradiction.
There is no differentiable structure on $M(\infty)$ such that $\pi_x$ is a diffeomorphism for all $x\in M$.
Suppose there is such a differentiable structure.
Let $A$ be a maximal flat of $M$, $x$ be a point of $A$.
Then $S_xA$ is an embedded submanifold of $S_xM$ and $\pi_x$ is a diffeomorphism. Thus $A(\infty)=\pi_x(S_xA)$ is a submanifold of $M(\infty)$. A contradiction with Proposition \[RemarqueSimple\].
Study of $\mH^2\times\mR$ {#HxR}
=========================
We summarize briefly the building structure of $\mH^2\times\mR$.
The singular geodesics are those of the form $\{x\}\times\mR$ where $x$ is a point of $\mH^2$; they are parallel (asymptotic at both ends) to one another. The maximal flats are the products $\gamma\times\mR$ where $\gamma$ is a geodesic of $\mH^2$ (see figure \[RS2\]).
The boundary of $\mH^2\times\mR$ is a $2$-sphere partitionned into two points and a family of nonintersecting curves joining them. The points are the end points of every singular geodesic, therefore panels of the building. The curves are the Weyl chambers, there is one of them for each point in the boundary of $\mH^2$ (see figure \[RS3\]). The union of any two of them and of the two panels is an apartment.
![A family of parallel singular geodesics lying on the same flat.[]{data-label="RS2"}](RangSup23.eps){width="90mm"}
![Here we show the Weyl chambers in the boundary of $\mH^2\times\mR$. Any choice of two of them corresponds to a geodesic of $\mH^2$, a flat of $\mH^2\times\mR$ and an apartment.[]{data-label="RS3"}](RangSup33.eps){width="90mm"}
\[HDR\] The space $\mH^2\times\mR$ admits no differentiable Hadamard compactification.
Suppose $\ron{D}$ is a differentiable Hadamard compactification of $M=\mH^2\times\mR$ and $\adherence{M}$ is endowed with $\ron{D}$.
Let $\gamma=\{x\}\times\mR^+$ be any singular geodesic ray of unit speed. We denote by $\gamma(\infty)$ the point of $M(\infty)$ defined by $\gamma$.
Since $\gamma(\infty)$ is fixed by all orientation-preserving isometries of $\mH^2$, the derivatives of these isometries give a linear representation of $\PSLDR$ on $T_{\gamma(\infty)}\adherence{M}$. This representation is reducible as $T_{\gamma(\infty)} M(\infty)$ is an invariant subspace. Let $\rho$ be the induced representation. As a representation of a simple Lie group $\rho$ is trivial or faithful.
Let $s_x$ be the geodesic symmetry of $\mH^2$ around $x$. We identify $s_x$ with the isometry $s_x \times \id$ of $\mH^2\times\mR$. For every time $t\in\mR$, $ds_x(\gamma(t))$ has eigenvalues $1$, $-1$, $-1$. By continuity, $ds_x(\gamma(\infty))$ must have the same eigenvalues, thus the restriction to $T_{\gamma(\infty)} M(\infty)$ of $ds_x(\gamma(\infty))$ is $-\id$, that is to say $\rho(s_x)=-\id$.
Let $y$ be any point of $\mH^2$ different from $x$. Then $\rho(s_y)=-\id$ too. Thus $\rho(s_xs_y)=\id$, but $s_xs_y$ is a non-trivial hyperbolic transformation (it is a translation along the geodesic containing $x$ and $y$). Thus $\rho$ is neither faithful nor trivial, a contradiction.
Product of a Euclidean space by a rank $1$ space {#FxRn}
================================================
We now generalize Proposition \[HDR\] to the case when $M=F\times\mR^{k-1}$ is the product of an Euclidean space by a symmetric space $F$ of rank $1$.
However, as we will need it later, we prove something stronger.
A of $M$ is defined as a differentiable Hadamard compactification where we replace $G$ by its identity component $G_0$.
We prove that $M$ admits no weak differentiable Hadamard compactification. Thus, in order to generalize the argument used in the proof of Proposition \[HDR\] we need the geodesic symmetries to belong to $G_0$. Of course this is false if $F$ is of odd dimension, and we shall use another argument in this case.
Let $F$ be a rank $1$ symmetric space of noncompact type. If $\dim{F}$ is even, then the geodesic symmetries belong to $G_0$. If $\dim{F}$ is odd, then $F$ is a real hyperbolic space $\mH^{2m+1}$.
From the classification of symmetric spaces (see for example [@Helgason], Chapter IX) we know that the rank $1$ symmetric spaces of noncompact type are: the real hyperbolic spaces, the complex hyperbolic spaces, the quaternionic hyperbolic spaces and an exceptionnal space, the octonionic hyperbolic plane (therefore the last assertion is clear). The identity components of their isometry groups are respectively $\SOopq{1}{n}$, $\SUpq{1}{n}$, $\Sppq{1}{n}$ and $F_{4(-20)}$, which are simple Lie groups.
We shall use the following criterion. Let $\al{g}=\al{k}+\al{p}$ be a Cartan decomposition of the Lie algebra of $G_0$. Then the geodesic symmetries are in $G_0$ if and only if $\al{k}$ contains a maximal abelian algebra of $\al{g}$ (see [@Helgason], Chapter IX �3).
It is know sufficient to compare the ranks of $\al{g}$ and $\al{k}$ (most of them can be found in [@Knapp2], Appendix C):
- in the real case, $\al{g}=\sopq{n}{1}$ is of rank $\lfloor\frac{1}{2}(n+1)\rfloor$ and $\al{k}=\so{p}$, of rank $\lfloor n/2\rfloor$. These two ranks coincide exactly when $n$ is even,
- in the complex case, $\al{g}=\supq{n}{1}$ and $\al{k}=\al{s}(\up{n}\times\up{1})$ are both of rank $n$,
- in the quaternionic case, $\al{g}=\sppq{n}{1}$ and $\al{k}=\spp{n}\times\spp{1}$ are both of rank $n+1$,
- in the octonionic case, $\al{g}=\al{f}_{4(-20)}$ and $\al{k}=\al{so}(9)$ are both of rank $4$. I wish to thank Fokko du Cloux, Jérôme Germoni and Bruno Sévennec for explaning this case to me.
We can now prove the following.
\[produit\] If $M=F\times\mR^{k-1}$ where $F$ is a rank $1$ symmetric space of noncompact type, then $M$ admits no weak differentiable Hadamard compactification. In particular, $M$ admits no differentiable Hadamard compactification.
Suppose there is such a differentiable structure on $\adherence{M}$.
We identify an isometry $g$ of $F$ with the isometry $g\times \id$ of $M$. We denote the component of identity of the group of isometries of $F$ by $G_0^F$ and consider it a subgroup of $G_0$.
We first suppose that $F$ is of even dimension.
Let $s_x$ and $s_y$ be the geodesic symmetries around two different points $x$ and $y$ of $F$. Let $d$ be a geodesic of $\mR^{k-1}$. Then $\gamma_1=\{x\}\times d$ and $\gamma_2=\{y\}\times d$ are asymptotic and $z=\gamma_1(+\infty)=\gamma_2(+\infty)$ is a singular point of index $1$ of $M(\infty)$.
Differentiation $g\mapsto dg(z)$ gives us a linear representation of $G_0^F$ in $T_{z} \adherence{M}$. This representation is reducible since $T_{z} M(\infty)$ is an invariant subspace. Let $\rho$ be the representation induced on $T_{z} M(\infty)$.
We now decompose the representation $\rho$. For all $t\in\mR$, the eigenvalues of $ds_x(\gamma_1(t))$ are $1$ with multiplicity $k-1$ and $-1$ with multiplicity $\dim{F}$. Thus, $\rho(s_x)$ must have eigenvalues $1$ with multiplicity $k-2$ and $-1$ with multiplicity $\dim{F}$ (an eigendirection transverse to the boundary must have a nonnegative eigenvalue).
We shall decompose $\rho$ using the following lemma.
\[lemm1produit\] The panel $P$ of $\gamma_1(\infty)$ is a differentiable submanifold of $M(\infty)$.
The panel $P$ is pointwise fixed by all isometries of $F$. Thus it is pointwise fixed by $s_x$. In a local chart, it is defined by $p\in P\Rightarrow s_x(p)-p=0$. Since $s_x-\id$ has rank $\dim{F}$ and $\dim{M(\infty)}=\dim{F}+k-2$, the inverse function theorem implies that $P$ is contained in a differentiable submanifold of $M(\infty)$ of dimension $k-2$ (namely the set of fixed points of $s_x$). But, as a panel in the boundary of a rank $k$ symmetric space, it is an open topological manifold of dimension $k-2$. Thus $P$ is a differentiable submanifold of dimension $k-2$ of $M(\infty)$.
The tangent space $T_z P$ is an invariant subspace of $\rho$. Thus, this representation splits in two parts : $\rho=\rho_0\oplus\rho_1$ where $\rho_0$ is the trivial representation of dimension $k-2$ and $\rho_1$ is a representation of dimension $\dim{F}$.
Now we have a representation $\rho_1$ (which, as $G_0^F$ is simple, must be faithful or trivial) with $\rho_1(s_x)=-\id$. So $\rho_1$ cannot be trivial. But $ds_y(\gamma_2(t))$ has the same eigenvalues as $ds_x(\gamma_1(t))$, and thus $\rho_1(s_y)=-\id$ too. Now we have $\rho_1(s_xs_y)=\id$ with $s_xs_y$ a hyperbolic transformation, so $\rho_1$ cannot be faithful, a contradiction.
Suppose now that $F$ is of odd dimension.
We have $F=\mH^{2m+1}$, the real hyperbolic space. The geodesic symmetries are not in $G_0$ (their determinant is $-1$) and we shall use Proposition \[RemarqueSimple\].
Let $A$ be a maximal flat. Then $A$ is the product of a geodesic $\gamma$ of $F$ by $\mR^{k-1}$. Let $r\in G_0^F$ be the rotation of angle $\pi$ around $\gamma$ in $F$; $A$ is the set of fixed points of $r$, and $A(\infty)$ is the set of fixed points of $r$ in $M(\infty)$. Since $r$ is an involution, $A(\infty)$ is a submanifold of $M(\infty)$, a contradiction to Proposition \[RemarqueSimple\].
We give an alternative proof for the odd dimension case, less elegant but more useful for the proof of Theorem \[theoreme\]. We can define the representation $\rho_1$ like in the even dimension case. Then $\rho_1$ is a representation of dimension $\dim{F}$ of $G_0^F$. Since, for all $x\in M$ fixed by $r$, $dr(x)$ has eigenvalues $-1$ with multiplicity $2m$ and $1$ with multiplicity $k$, $\rho_1(r)$ has eigenvalues $-1$ with multiplicity $2m$ and $1$ with multiplicity $1$. Thus $\rho_1(r)\neq\id$, hence $\rho_1$ is not trivial.
But $G_0^F=\SOopq{2m+1}{1}$ admits no non trivial representation of dimension less than $2m+2$, a contradiction.
Proof of Theorem \[theoreme\] {#Stheoreme}
=============================
We shall now prove Theorem \[theoreme\] with the same ideas that we used for the previous propositions.
Let $M$ be a noneuclidean symmetric space of nonpositive curvature of dimension $n$ and rank $k>1$. As before, $G$ is the group of all isometries of $M$, $G_0$ is the identity component of $G$ and $\al{g}$ is the Lie algebra of $G$.
Suppose that there exists a differentiable Hadamard compactification $\ron{D}$ of $M$.
We denote by $\alpha$ the action of $G$ on $\adherence{M}$. We also denote by $\alpha$ the corresponding action of $\al{g}$.
The first step is to find in $M$ an embedded product $F\times \mR^{k-1}$.
Let $\gamma$ be a singular geodesic of index $1$. Let $F_{\gamma}$ be the union of geodesics parallel to $\gamma$ (recall that parallel means that they are both positively and negatively asymptotic). Then, $F_{\gamma}$ is a totally geodesic submanifold of $M$ isometric to a product $F\times\mR^{k-1}$ where $F$ is a symmetric space of rank $1$ (see [@Eberlein] Section 2.11).
Let $\adherence{F}_{\gamma}$ be the closure of $F_{\gamma}$ in $\adherence{M}$ and $F_{\gamma}(\infty)=
\adherence{F}_{\gamma}\cap M(\infty)$.
It would be interesting to prove that $\adherence{F}_{\gamma}$ is a submanifold of $\adherence{M}$, since we could directly use Proposition \[produit\] to get a contradiction, but a weaker statement (namely Lemma \[lemm5theorem\]) will be sufficient.
Up to a change of parametrization we can write $$\gamma(t)=(p,(t,0,\dots,0))$$ where $p\in F$, and $F^t=F\times\{(t,0\dots,0)\}$ is identified with its embedding into $M$.
Since $F$, identified with $F^0$, is a totally geodesic submanifold of $M$, the Lie algebra of the group $G^F$ of isometries of $F$ is a subalgebra of $\al{g}$ and the identity component $G_0^F$ of $G^F$ is a subgroup of $G_0$ (that’s why we needed the stronger statement in Proposition \[produit\]). Thus taking derivatives gives us a representation $\rho$ of $G_0^F$ on $T_{\gamma(\infty)} \adherence{M}$.
Let $\al{k}^t\oplus\al{p}^t$ be the Cartan decomposition of $\al{g}$ at $\gamma(t)$. Since $F_{\gamma}$ is a totally geodesic submanifold we have a further decomposition $\al{p}^t=\al{p}^t_F\oplus\al{p}^t_{eucl}\oplus\al{p}^t_0$ where the terms are pairwise orthogonal (with respect to the Killing form), $\al{p}^t_F$ is mapped by $\alpha$ onto $T_pF^t$, $\al{p}^t_{eucl}$ is mapped onto $T_p\mR^{k-1}$ and $\al{p}^t_0$ is mapped onto $\ortho{(T_pF_{\gamma})}$. We define $\al{p}^t_{\gamma}=\al{p}^t_F\oplus\al{p}^t_{eucl}$; $\alpha$ maps $\al{p}^t_{\gamma}$ onto $T_{\gamma(t)} F_{\gamma}$. Moreover, $\al{p}^t_{\gamma}$ is the set of all Killing fields in $\al{p}$ commuting with that Killing field $X\in\al{p}^t$ such that $\alpha(X)$ is the unit tangent vector of $\gamma$.
We shall split $\rho$ in three parts in correspondence with the splitting $\al{p}^t=\al{p}^t_F\oplus\al{p}^t_{eucl}\oplus\al{p}^t_0$. To achieve this, we use the following stability result.
\[lemmrep\] Let $K$ be any compact group and $(\mu_t)_{t\in\mR}$ be a continuous family of linear representations of $K$ on some finite-dimensional real vector space $V$. Then for all pairs $(t_1,t_2)$ of real numbers, the representations $\mu_{t_1}$ and $\mu_{t_2}$ are conjugate.
As $K$ is compact and $V$ is finite-dimensional, the conjugacy class of a representation $\mu_t$ is determined by its character. More precisely, the multiplicity in $\mu_t$ of some irreducible representation $\nu$ is given by the scalar product of the characters of $\mu_t$ and $\nu$, hence is a continuous map. This multiplicity is an integer and is thus constant.
\[lemm1theorem\] The tangent subspaces $T_{\gamma(t)} F^t$, $T_{\gamma(t)} \mR^{k-1}$ and $\ortho{(T_{\gamma(t)} F_{\gamma})}$ admit limits when $t\rightarrow\infty$, denoted respectively by $V_F$, $V_{eucl}$ and $V_0$. Moreover one has $T_{\gamma(\infty)}\adherence{M}=V_F\oplus V_{eucl}\oplus V_0$.
Let $K_0^F$ be the isotropy group of $p$ in $F$. Taking derivatives gives linear representations $\rho^t$ of $K_0^F$ in $T_{\gamma(t)} M$ for all $t\in\mR$. Then $T_{\gamma(t)} F^t$, $T_{\gamma(t)}\mR^{k-1}$ and $\ortho{(T_{\gamma(t)}F_\gamma)}$ are invariant spaces of $\rho_t$.
By continuity, the restriction $\rho^{\infty}$ of $\rho$ to $K_0^F$ splits into three parts and the conclusion holds.
From Lemma \[lemmrep\] we deduce that the action of $\rho^{\infty}$ is conjugate with that of $\rho^0$. Since $\rho^{t}$ acts trivially on $T_{\gamma(t)}\mR^{k-1}$, $\rho^{\infty}$ acts trivially on $V_{eucl}$.
We shall now prove that $V_F$ is an invariant subspace for $\rho$.
\[lemm2theorem\] Let $\ron{O}$ be the orbit of $\gamma(\infty)$ under the action of $G_0$. Then $V_0=T_{\gamma(\infty)} \ron{O}$ and, for all $t$, the restriction of $\alpha_{\gamma(\infty)}$ to $\al{p}^t_0$ is one-to-one and onto $V_0$.
As an orbit, $\ron{O}$ is a submanifold of $M(\infty)$, invariant under the action of $G_0$, thus $T_{\gamma(\infty)}\ron{O}$ is an invariant space of $\rho$.
The definition of $\ron{O}$ shows that $\alpha$ sends $\al{p}^t_0$ on $T_{\gamma(\infty)}\ron{O}$. We want to prove that this map is one-to-one and onto.
Let $H$ be an element of $\al{p}^t_0$. By definition, $\alpha(\exp(H))(\gamma)$ is not parallel to $\gamma$. If $\alpha(H)_{\gamma(\infty)}=0$, then $\alpha(H)_{\gamma(-\infty)}\neq 0$. But after conjugacy by the geodesic symmetry at $\gamma(t)$ we find $\alpha(-H)_{\gamma(-\infty)}=0$, a contradiction. Thus the restriction of $\alpha_{\gamma(\infty)}$ to $\al{p}^t_0$ is one-to-one.
Since $\dim{\al{p}^t_0}=n-\dim{F_{\gamma}}=
(n-1)-(\dim{F_{\gamma}}-1)=\dim{\ron{O}}$, it is onto and $V_0=T_{\gamma(\infty)}\ron{O}$.
\[lemm3theorem\] The subspace $V_F$ contains no subspace where $\rho$ is trivial.
Let $X$ be that vector of $\al{p}^0_{eucl}$ such that $\gamma'(t) = \alpha(X)_{\gamma(t)}$. The linear operator $d_{\gamma(\infty)} \alpha(\exp X)-\id$ acting on $V_0$ is of maximal rank.
The linear action $\rho^{\infty}$ of $K_0^F$ on $T_{\gamma(\infty)} M(\infty)$ is transitive on $V_F$ and $\rho^{\infty}$ is the restriction of $\rho$ to $K_0^F$ thus $V_F$ contains no trivial part.
To prove the second part of the lemma, we use the root space decomposition $\al{g}=\al{g}_0+\sum \al{g}_{\lambda}$ given by some maximal flat containing $\gamma$. We have $$\al{p}^0_0\subseteq \sum_{\lambda(X)\neq 0} \al{g}_{\lambda}$$ and $Ad(\exp X) = e^{\lambda(X)} I$ on $\al{g}_{\lambda}$.
Since $\alpha$ is onto from $\al{p}^0_0$ to $V_0$, it is onto from $\sum_{\lambda(X)\neq 0} \al{g}_{\lambda}$ to $V_0$.
For all $H\in \al{g}_{\lambda}$, we have $$d_{\gamma(\infty)} \alpha(\exp X)(\alpha(H))
= \alpha(Ad(\exp(X))(H))
= e^{\lambda(X)}\alpha(H),$$ and thus $d_{\gamma(\infty)} \alpha(\exp X)-\id$ is nondegenerate on $V_0$.
\[lemm4theorem\] The panel $P$ of $\gamma(\infty)$ is a submanifold of $M(\infty)$ and its tangent space at $\gamma(\infty)$ equals $V_{eucl}\cap T_{\gamma(\infty)}M(\infty)$.
The panel $P$ is contained in the set of the points of $M(\infty)$ left fixed by the actions of $G$ and of $\exp(X)$. Written in local coordinates, this gives us an infinite system of equations. By Lemma \[lemm3theorem\] we know that this system is of rank at least $d=\dim V_0+\dim V_F$ at $\gamma(\infty)$. We can extract a subsystem of $d$ equations that is of maximal rank at $\gamma(\infty)$. The inverse function theorem implies that this subsystem defines a submanifold of $M(\infty)$ of dimension $k-2=\dim V_{eucl}\cap T_{\gamma(\infty)}M(\infty)$ and containing $P$. But $P$ is topologically a manifold of dimension $k-2$ and thus must be a differentiable submanifold of $M(\infty)$.
Since $K_0^F$ acts trivially on $P$, its tangent space must be $V_{eucl}\cap T_{\gamma(\infty)}M(\infty)$.
Since $G$ acts trivially on $P$, it must preserve its tangent space. We are now ready to prove the following.
\[lemm5theorem\] The subspace $V_F$ is invariant by $\rho$.
From previous lemmas we know that $V_0$ and $V_{eucl}\cap T_{\gamma(\infty)}M(\infty)$ are invariant subspaces for $\rho$. Since $\rho$ is totally reducible, there exists some subspace $V'$ invariant by $\rho$ such that one has the following decomposition $T_{\gamma(\infty)} M(\infty) =
V'\oplus V_{eucl}\cap T_{\gamma(\infty)}M(\infty) \oplus V_0$. But $V'$ must be invariant by $\rho^{\infty}$, and thus $V'=V_F$ and $V_F$ is invariant by $\rho$.
Denote by $\rho_1$ the representation of $G_0^F$ induced by $\rho$ on $V_F$.
Since the restriction of $\rho_1$ to $K_0^F$ is the limit of the restriction of $\rho^t$ to $T_{\gamma(t)} F^t$, we can now use the same arguments as in the proof of Proposition \[produit\].
If $F$ is of even dimension, $\rho_1$ is neither trivial nor faithful, a contradiction.
If $F$ is of odd dimension, $\rho_1$ is a nontrivial representation of dimension $\dim{F}$ of $G_0^F$, a contradiction.
Theorem \[theoreme\] is proved. Note that we actually get something stronger: there exists no differentiable Hadamard compactification of $M$; the obstructions to differentiability appear in the identity component of $G$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we propose a scheme called “beam-nulling" for MIMO adaptation. In the beam-nulling scheme, the eigenvector of the weakest subchannel is fed back and then signals are sent over a generated subspace orthogonal to the weakest subchannel. Theoretical analysis and numerical results show that the capacity of beam-nulling is closed to the optimal water-filling at medium SNR. Additionally, signal-to-interference-plus-noise ratio (SINR) of MMSE receiver is derived for beam-nulling. Then the paper presents the associated average bit-error rate (BER) of beam-nulling numerically which is verified by simulation. Simulation results are also provided to compare beam-nulling with beamforming. To improve performance further, beam-nulling is concatenated with linear dispersion code. Simulation results are also provided to compare the concatenated beam-nulling scheme with the beamforming scheme at the same data rate. Additionally, the existing beamforming and new proposed beam-nulling can be extended if more than one eigenvector is available at the transmitter. The new extended schemes are called multi-dimensional (MD) beamforming and MD beam-nulling. Theoretical analysis and numerical results in terms of capacity are also provided to evaluate the new extended schemes. Simulation results show that the MD scheme with LDC can outperform the MD scheme with STBC significantly when the data rate is high.'
author:
- |
Mabruk Gheryani, Zhiyuan Wu, and Yousef R. Shayan\
Concordia University, Department of Electrical Engineering\
Montreal, Quebec, H4G 2W1, Canada\
email: (m\_gherya, zy\_wu, yshayan)@ece.concordia.ca
title: 'Capacity and Performance of Adaptive MIMO System Based on Beam-Nulling'
---
Introduction
============
Since the discovery of multiple-input-multiple-output (MIMO) capacity[@Telatar:MIMOCapacity][@Foschini:MIMOCapacity], a lot of research efforts have been put into this field. It has been recognized that adaptive techniques proposed for single-input-single-output (SISO) channels[@Cavers:VR_transmission][@Goldsmith:VR_MQAM] can also be applied to improve MIMO channel capacity.
The ideal scenario is that the transmitter has full knowledge of channel state information (CSI). Given this perfect CSI feedback, the original MIMO channel can be converted to multiple uncoupled SISO channels via singular value decomposition (SVD) at the transmitter and the receiver[@Telatar:MIMOCapacity]. In other words, the original MIMO channel can be decomposed into several orthogonal “spatial subchannels" with various propagation gains.
To achieve better performance, various schemes can be implemented depending on the availability of CSI at the transmitter [@Luo:Capacity_Time-Varying]-[@Giannakis:Multiantenna-Beamforming-Constrained-Feedback]. If the transmitter has full knowledge of the channel matrix, i.e., full CSI, the so-called “water-filling" (WF) principle is performed on each spatial subchannel to maximize the channel capacity [@Telatar:MIMOCapacity]. This scheme is optimal in this case. Various WF-based schemes have been proposed, such as [@Shen:Comparison_Water-filling][@Zhang:QoS_WF]. For the WF-based scheme, the feedback bandwidth for the full CSI grows with respect to the number of transmit and receive antennas and the performance is often very sensitive to channel estimation errors.
To mitigate these disadvantages, various beamforming (BF) techniques for MIMO channels have also been investigated intensively. In an adaptive beamforming scheme, the complex weights of the transmit antennas are fed back from the receiver. If only one eigenvector can be fed back, eigen-beamforming [@Giannakis:Optimal-STBC-Channel-Mean] is optimal. The eigen-beamforming scheme only applies to the strongest spatial subchannel but can achieve full diversity and high signal-to-noise ratio (SNR) [@Giannakis:Optimal-STBC-Channel-Mean]. Also, in practice, the eigen-beamforming scheme has to cooperate with the other adaptive parameters to improve performance and/or data rates such as constellation and coding rate. There are also other beamforming schemes based on various criteria. Examples of such schemes are [@Giannakis:Optimal-STBC-Channel-Mean] -[@Cavers:BF]. Note that the conventional beamforming is optimal in terms of maximizing the SNR at the receiver. However, it is sub-optimal from the MIMO capacity perspective, since only a single data stream, as opposed to parallel streams, is transmitted through the MIMO channel [@Zhou:Channel_Prediction].
In this paper, we propose a new technique called “beam-nulling" (BN). This scheme uses the same feedback bandwidth as beamforming, that is, only one eigenvector is fed back to the transmitter. The beam-nulling transmitter is informed by the weakest spatial subchannel and, where both transmitter and receiver know how to generate the same spatial subspace, sends signals over a generated spatial subspace orthogonal to the weakest subchannel. Although the transmitted symbols are “precoded" according to the feedback, beam-nulling is different from the other existing precoding schemes with limited feedback channel, which are independent of the instantaneous channel but the optimal precoding depends on the instantaneous channel [@Love:Beamforming][@Zheng:Capacity_Precode].
Using this new techniques instead of the optimal water-filling scheme, the loss of channel capacity can be reduced. This paper also addresses the performance of beam-nulling. To achieve better performance, beam-nulling can be concatenated with the other space-time (ST) coding schemes, such as space-time trellis codes (STTCs)[@Tarokh:STC_Perf_Const], space-time block codes (STBCs)[@Alamouti:STBC][@Tarokh:STBC] and linear dispersion codes (LDCs)[@Hassibi:LDC]-[@Wu:Design4CSTM], etc. For simplicity and flexibility, LDCs are preferable. We provide numerical and simulation results are provided to demonstrate the merits of the new proposed scheme. Additionally, if more than one eigenvector, e.g. $k$ eigenvectors, can be available at the transmitter, the existing beamforming scheme and the proposed beam-nulling scheme can be further extended, respectively. The extended schemes will exploit or discard $k$ spatial subchannels and they will be referred to as “multi-dimensional (MD)" beamforming and “multi-dimensional" beam-nulling, respectively.
This paper will be organized as follows. Our channel model is presented in Section \[sec\_sys\_mod\]. In Section \[sec\_PA\], four power allocation strategies, i.e., equal power, water-filling, eigen-beamforming, and a new power allocation strategy called “beam-nulling" are studied and compared in terms of channel capacity. In Section \[sec\_perf\_BN\], bit error rate (BER) of the proposed beam-nulling scheme using MMSE detector is studied and verified. The proposed scheme is compared with the eigen-beamforming scheme at various data rates in terms of BER. Beam-nulling concatenated with LDC is proposed and evaluated. In Section \[sec\_ext\_frame\], extended adaptive frameworks, i.e., MD beamforming and MD beam-nulling, are proposed. Capacity and performance of these two schemes are discussed and compared. To improve performance further and maintain reasonable complexity, MD schemes concatenated with linear space-time codes, such as STBC and LDC, are proposed and evaluated. Finally, in Section \[sec\_con\], conclusions are drawn.
Channel Model {#sec_sys_mod}
=============
In this study, the channel is assumed to be a Rayleigh flat fading channel with $N_t$ transmit and $N_r$ ($N_r \geq N_t$) receive antennas. We denote the complex gain from the transmit antenna $n$ to the receiver antenna $m$ by $h_{mn}$ and collect them to form an $N_r \times N_t$ channel matrix $\mathbf{H}=[h_{mn}]$. The channel is known perfectly at the receiver. The entries in $\mathbf{H}$ are assumed to be independent and identically distributed (*i.i.d.*) symmetrical complex Gaussian random variables with zero mean and unit variance.
The symbol vector at the $N_t$ transmit antennas is denoted by $\mathbf{x}=[x_1, x_2, \ldots, x_{N_t}]^T$. According to information theory [@Shannon:info_theory], the optimal distribution of the transmitted symbols is Gaussian. Thus, the elements $\{x_i\}$ of $\mathbf{x}$ are assumed to be *i.i.d.* Gaussian variables with zero mean and unit variance, i.e., $E(x_i)=0$ and $E|x_i|^2=1$.
The singular-value decomposition of $\mathbf{H}$ can be written as $$\label{svd_H}
\mathbf{H}=\mathbf{U}\mathbf{\Lambda}\mathbf{V}^{H}$$ where $\mathbf{U}$ is an $N_r \times N_r$ unitary matrix, $\mathbf{\Lambda}$ is an $N_r \times N_t$ matrix with singular values $\{\lambda_{i}\}$ on the diagonal and zeros off the diagonal, and $\mathbf{V}$ is an $N_t \times N_t$ unitary matrix. For convenience, we assume $\lambda_{1}\geq \lambda_{2} \ldots
\geq \lambda_{N_t}$, $\mathbf{U}=[\mathbf{u}_1 \mathbf{u}_2 \ldots
\mathbf{u}_{N_r} ]$ and $\mathbf{V}=[\mathbf{v}_1 \mathbf{v}_2
\ldots \mathbf{v}_{N_t} ]$. $\{\mathbf{u}_i\}$ and $\mathbf{v}_i$ are column vectors. From equation (\[svd\_H\]), the original channel can be considered as consisting of uncoupled parallel subchannels. Each subchannel corresponds to a singular value of $\mathbf{H}$. In the following context, the subchannel is also referred to as “spatial subchannel". For instance, one spatial subchannel corresponds to $\lambda_{i}$, $\mathbf{u}_i$ and $\{\mathbf{v}_i\}$.
Power Allocation Among Spatial Subchannels {#sec_PA}
==========================================
We assume that the total transmitted power is constrained to $P$. Given the power constraint, different power allocation among spatial subchannels can affect the channel capacity tremendously. Depending on power allocation strategy among spatial subchannels, four schemes are presented which are equal power, water-filling, eigen-beamforming, and the new power allocation which is beam-nulling.
If the transmitter has no knowledge about the channel, the most judicious strategy is to allocate the power to each transmit antenna equally, i.e., equal power. In this case, the received signals can be written as $$\label{sys_mod4eq}
\mathbf{y}=\sqrt{\frac{P}{N_t}}\mathbf{H}\mathbf{x}+\mathbf{z}$$ $\mathbf{z}$ is the additive white Gaussian noise (AWGN) vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma_z^2$. The associated ergodic channel capacity can be written as [@Telatar:MIMOCapacity] $$\label{cap_equal}
\bar{C}_{eq}=E\left[\sum\limits_{i=1}^{N_t}\log\left(1+\frac{\rho}{N_t
}\lambda_{i}^2\right)\right]$$ where $E[\cdot]$ denotes expectation with respect to $\mathbf{H}$ and $\rho=\frac{P}{\sigma^{2}_{z}}$ denotes SNR. If the transmitter has full knowledge about the channel, the most judicious strategy is to allocate the power to each spatial subchannel by water-filling principle [@Telatar:MIMOCapacity]. In water-filling scheme, the received signals can be written as $$\label{WF_subchan}
\tilde{y}_i=\sqrt{P_i} \lambda_{i} x_i+\tilde{z}_i$$ where $\sum\limits_{i=1}^{N_t}P_i=P$ as a constraint and $\tilde{z}_i$ is the AWGN random variable with zero mean and $\sigma_z^2$ variance. Following the method of Lagrange multipliers, $P_i$ can be found [@Telatar:MIMOCapacity] and the total ergodic channel capacity is $$\label{cap_wf}
\bar{C}_{wf}=E\left[\sum\limits_{i=1}^{N_t}\log\left(1+\frac{P_i}{\sigma^{2}_{z}}\lambda_{i}^2\right)\right]$$
To save feedback bandwidth, beamforming can be considered. For the MIMO model, the optimal beamforming is called “eigen-beamforming" [@Giannakis:Optimal-STBC-Channel-Mean], or simply beamforming. We assume one symbol, saying $x_1$, is transmitted. At the receiver, the received vector can be written as $$\label{eigen_rec_y}
\mathbf{y}_1=\sqrt{P} \mathbf{H}\mathbf{v}_{1}x_1+\mathbf{z}_1$$ where $\mathbf{z}_1$ is the additive white Gaussian noise vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma^{2}_{z}$. The associated ergodic channel capacity can be written as $$\label{cap_bf}
\bar{C}_{bf}=E\left[\log\left(1+\rho\lambda_{1}^2 \right)\right]$$
The eigen-beamforming scheme can save feedback bandwidth and is optimized in terms of SNR [@Cavers:BF]. However, since only one spatial subchannel is considered, this scheme suffers from loss of channel capacity [@Zhou:Channel_Prediction], especially when the number of antennas grows.
Beam-Nulling {#subsec_BN_model}
------------
The eigen-beamforming scheme can save feedback bandwidth and is optimized in terms of SNR [@Cavers:BF]. However, since only a single spatial subchannel is considered, this scheme suffers from loss of channel capacity [@Zhou:Channel_Prediction], especially when the number of antennas grows. Inspired by the eigen-beamforming scheme, we will propose a new beamforming-like scheme called “beam-nulling" (BN). This scheme uses the same feedback bandwidth as beamforming, that is, only one eigenvector is fed back to the transmitter. Unlike the eigen-beamforming scheme in which only the best spatial subchannel is considered, the beam-nulling scheme discards only the worst spatial subchannel. Hence, in comparison with the optimal water-filling scheme, the loss of channel capacity can be reduced.
![\[fig:1D\_BN\_scheme\] beam-nulling scheme.](1D_beamnulling.eps){width="3in"}
In this scheme as shown in Fig. \[fig:1D\_BN\_scheme\], the eigenvector associated with the minimum singular value from the transmitter side, i.e., $\mathbf{v}_{N_t}$, is fed back to the transmitter. A subspace orthogonal to the weakest spatial channel is constructed so that the following condition is satisfied. $$\label{orth_subspace}
\mathbf{\Phi}^{H}\mathbf{v}_{N_t}=\mathbf{0}$$ The $N_t \times (N_t-1)$ matrix $\mathbf{\Phi}=[\mathbf{g}_1
\mathbf{g}_2 \ldots \mathbf{g}_{N_t-1}]$ spans the subspace. Note that the method to construct the subspace $\mathbf{\Phi}$ should also be known to the receiver.
Here is an example of construction of the orthogonal subspace. We construct an $N_t \times N_t$ matrix $$\label{A_construction}
\mathbf{A}=[\mathbf{v}_{N_t} \mathbf{I}']$$ where $\mathbf{I}'=[\mathbf{I}_{(N_t-1)\times(N_t-1)}
\mathbf{0}_{(N_t-1)\times 1}]^T$. Applying QR decomposition to $\mathbf{A}$, we have $$\label{QR_Decom}
\mathbf{A}=[\mathbf{v}_{N_t} \mathbf{\Phi}]\cdot\mathbf{\Gamma}$$ where $\mathbf{\Gamma}$ is an upper triangular matrix with the (1,1)-th entry equal to 1. $\mathbf{\Phi}$ is the subspace orthogonal to $\mathbf{v}_{N_t}$.
At the transmitter, $N_t-1$ symbols denoted as $\mathbf{x}'$ are transmitted over the orthogonal subspace $\mathbf{\Phi}$. The received signals at the receiver can be written as
$$\begin{aligned}
\label{BN_rec_y}
% \nonumber to remove numbering (before each equation)
\nonumber
\mathbf{y}' &=& \sqrt{\frac{P}{N_t-1}}\mathbf{H}\mathbf{\Phi}\mathbf{x}'+\mathbf{z}' \\
&=& \widehat{\mathbf{H}}\mathbf{x}'+\mathbf{z}'\end{aligned}$$
where $\mathbf{z}'$ is additive white Gaussian noise vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma^{2}_{z}$ and $\widehat{\mathbf{H}}=\sqrt{\frac{P}{N_t-1}}\mathbf{H}\mathbf{\Phi}$.
Substituting (\[svd\_H\]) into (\[BN\_rec\_y\]) and multiplying $\mathbf{y}'$ by $\mathbf{U}^{H}$, results in $$\label{BN_tld_y}
\widetilde{\mathbf{y}}=\sqrt{\frac{P}{N_t-1}}\mathbf{\Lambda}\left(%
\begin{array}{c}
\mathbf{B} \\
\mathbf{0}^{T} \\
\end{array}%
\right)\mathbf{x}'+\widetilde{\mathbf{z}}$$ where $\widetilde{\mathbf{z}}$ is additive white Gaussian noise vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma^{2}_{z}$. With the condition in (\[orth\_subspace\]), $$\mathbf{V}^{H}\mathbf{\Phi}=\left(%
\begin{array}{c}
\mathbf{B} \\
\mathbf{0}^{T} \\
\end{array}%
\right)$$ where $$\label{matrix_B}
\mathbf{B}=\left(%
\begin{array}{cccc}
\mathbf{v}_{1}^{H}\mathbf{g}_{1} & \mathbf{v}_{1}^{H}\mathbf{g}_{2} & \ldots & \mathbf{v}_{1}^{H}\mathbf{g}_{N_t-1} \\
\mathbf{v}_{2}^{H}\mathbf{g}_{1} & \ddots & \ldots & \vdots \\
\vdots & \vdots & \ddots & \vdots \\
\mathbf{v}_{N_t-1}^{H}\mathbf{g}_{1}& \ldots & \ldots & \mathbf{v}_{N_t-1}^{H}\mathbf{g}_{N_t-1} \\
\end{array}%
\right)$$
$\mathbf{B}$ is an $(N_t-1)\times(N_t-1)$ unitary matrix. From (\[BN\_tld\_y\]), the available spatial channels are $N_t-1$. Since the weakest spatial subchannel is “nulled" in this scheme, power can be allocated equally among the other $N_t-1$ subchannels. Equation (\[BN\_tld\_y\]) can be rewritten as $$\label{BN_tld_y_1}
\widetilde{\mathbf{y}}'=\sqrt{\frac{P}{N_t-1}}\mathbf{\Lambda}'\mathbf{B}\mathbf{x}'+\widetilde{\mathbf{z}}'$$ where $\widetilde{\mathbf{y}}'$ and $\widetilde{\mathbf{z}}'$ are column vectors with the first $(N_r-1)$ elements of $\widetilde{\mathbf{y}}$ and $\widetilde{\mathbf{z}}$, respectively, and $\mathbf{\Lambda}'=\mathrm{diag}[\lambda_{1},
\lambda_{2}, \ldots, \lambda_{(N_t-1)}]$. From (\[BN\_tld\_y\_1\]), the associated ergodic channel capacity can be found as $$\label{cap_bn}
\bar{C}_{bn}=E\left[\sum\limits_{i=1}^{N_t-1}\log\left(1+\frac{\rho}{N_t-1}\lambda_{i}^2\right)\right]$$
As can be seen, the beam-nulling scheme only needs one eigenvector to be fed back. However, since only the worst spatial subchannel is discarded, this scheme can increase channel capacity significantly as compared to the conventional beamforming scheme.
Comparisons Among the Four Schemes
----------------------------------
In this section, we compare the new proposed beam-nulling scheme with the other schemes, i.e., equal power, beamforming and water-filling schemes. Water-filling is the optimal solution among the four schemes for any SNR.
Differentiating the above ergodic capacities with respect to $\rho$ respectively, we have $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
\frac{\partial \bar{C}_{eq}}{\partial \rho} &=& E\left[\sum\limits_{i=1}^{N_t}\frac{1}{\rho+\frac{N_t}{\lambda_{i}^2}}\right] \\
\frac{\partial \bar{C}_{bf}}{\partial \rho} &=& E\left[\frac{1}{\rho+\frac{1}{\lambda_{1}^2}}\right] \\
\frac{\partial \bar{C}_{bn}}{\partial \rho} &=& E\left[\sum\limits_{i=1}^{N_t-1}\frac{1}{\rho+\frac{N_t-1}{\lambda_{i}^2}}\right] \\\nonumber\end{aligned}$$ The differential will also be referred to as “slope". Since the second order differentials are negative, the above ergodic capacities are concave and monotonically increasing with respect to $\rho$.
With the fact that $\lambda_{1} \geq \lambda_{2} \ldots \geq
\lambda_{N_t}$, it can be readily checked that the slopes of ergodic capacities associate with equal power and beam-nulling are bounded as follows. $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
E\left(\frac{N_t}{\rho+\frac{N_t}{\lambda_{1}}}\right) \geq \frac{\partial \bar{C}_{eq}}{\partial \rho} \geq E\left(\frac{N_t}{\rho+\frac{N_t}{\lambda_{N_t}}}\right)\\
E\left(\frac{N_t-1}{\rho+\frac{N_t-1}{\lambda_{1}}}\right) \geq \frac{\partial \bar{C}_{bn}}{\partial \rho} \geq E\left(\frac{N_t-1}{\rho+\frac{N_t-1}{\lambda_{(N_t-1)}}}\right)\\\nonumber\end{aligned}$$
For the case of $N_t=2$, beamforming and beam-nulling have the same capacity for any $\rho$ as can be seen from equations of capacity and slope. If $\rho \rightarrow 0$, equivalently at low SNR, it can be easily found that $$\label{diff_low_SNR}
\frac{\partial \bar{C}_{bf}}{\partial \rho} \geq \frac{\partial
\bar{C}_{bn}}{\partial \rho} \geq \frac{\partial
\bar{C}_{eq}}{\partial \rho}, \rho \rightarrow 0$$ If $\rho \rightarrow \infty$, equivalently at high SNR, it can be easily found that $$\label{diff_high_SNR}
\frac{\partial \bar{C}_{eq}}{\partial \rho} \geq \frac{\partial
\bar{C}_{bn}}{\partial \rho} \geq \frac{\partial
\bar{C}_{bf}}{\partial \rho}, \rho \rightarrow \infty$$ Note that $\bar{C}_{bf}=\bar{C}_{bn}=\bar{C}_{eq}=0$ when $\rho=0$ or minus infinity in dB. Hence, at medium SNR, $\frac{\partial
\bar{C}_{bn}}{\partial \rho}$ has the largest value compared to $\frac{\partial\bar{C}_{bf}}{\partial \rho}$ and $
\frac{\partial\bar{C}_{eq}}{\partial \rho}$. Therefore, for low, medium and high SNRs, beamforming, beam-nulling and equal power have the largest capacities, respectively.
In Fig. \[5x5\_cap\], capacities of water-filling, beamforming, beam-nulling and equal power are compared over $5 \times 5$ Rayleigh fading channels, respectively. Note that since SNR is measured in dB, the curves become convex. In these figures, “EQ" stands for equal power, “WF" stands for water-filling, “BF" stands for beamforming and “BN" stands for beam-nulling. As can be seen, the water-filling has the best capacity at any SNR region. The other schemes perform differently at different SNR regions. At low SNR, the beamforming is the closest to the optimal water-filling, e.g., the SNR region below $3.5$ dB for $5 \times
5$ fading channel. Note that at low SNR, the water-filling scheme may only allocate power to one or two spatial subchannels. At medium SNR, the proposed beam-nulling is the closest to the optimal water-filling, e.g., the SNR region from $3.5$ dB to $16$ dB for $5 \times 5$ fading channel. The beam-nulling scheme only discards the weakest spatial subchannel and allocates power to the other spatial subchannels. As can be seen from the numerical results, the beam-nulling scheme performs better than the other schemes in this case. Note that at high SNR, the equal power scheme will converge with the water-filling scheme.
![\[5x5\_cap\]$5 \times 5$ Rayleigh fading channel.](clear_5x5_cap.eps){width="3.25in"}
Performance of Beam-nulling {#sec_perf_BN}
===========================
MMSE Detector
-------------
The close-form error probability for the optimal ML receiver is difficult to establish. Other suboptimal receivers can also be implemented. The MMSE detector is especially popular due to its low complexity and good performance[@Lupas:LMUD][@Poor:prob_MMSE]. In the following context, BER of the MMSE detector is analyzed for the beam-nulling scheme.
Let us define $\widehat{\mathbf{H}}=\sqrt{\frac{P}{N_t-1}}\mathbf{H}\mathbf{\Phi}$ and $\hat{\mathbf{h}}_i$ is the $i$-th column of $\widehat{\mathbf{H}}$. Equation (\[BN\_rec\_y\]) can also be written as $$\label{sinr_main1_bn}
\mathbf{y}'=\hat{\mathbf{h}}_i
x_i+\sum_{j \neq i}\hat{\mathbf{h}}_j x_j+
\mathbf{z}'$$ where $x_i$ is the i-th element of $\mathbf{x}'$.
Without loss of generality, we consider the detection of one symbol, say $x_i$. We collect the rest of the symbols into a column vector $\mathbf{x}_I$ and denote $\widehat{\mathbf{H}}_I=[\hat\mathbf{h}_{1},..,
\hat\mathbf{h}_{i-1},\hat\mathbf{h}_{i+1}, ...,
\hat\mathbf{h}_{N_t-1}]$ as the matrix obtained by removing the $i$-th column from $\widehat{\mathbf{H}}$.
A linear MMSE detector [@Poor:prob_MMSE][@Bohnke:SINR_Analysis] is applied and the corresponding output is given by $$\label{x_hat_i_bn}
\hat{x}_{i}=\mathbf{w}_{i}^H \mathbf{y}=x_i+ \hat{z}_{i}.$$ where $\hat{z}_{i}$ is the noise term of zero mean. $\hat{z}_{i}$ can be approximated to be Gaussian [@Poor:prob_MMSE]. The corresponding $\mathbf{w}_{i}$ can be found as $$\label{coef_i_bn}
\mathbf{w}_{i}=\frac{\left(\hat{\mathbf{h}}_i
\hat{\mathbf{h}}_i^H+\mathbf{R}_{I}\right)^{-1}
\hat{\mathbf{h}}_i}{\hat{\mathbf{h}}_i^{H}
\left(\hat{\mathbf{h}}_i
\hat{\mathbf{h}}_i^H+\mathbf{R}_{I}\right)^{-1}
\tilde{\mathbf{h}}_i}$$
where $\mathbf{R}_{I}=\widehat{\mathbf{H}}_{I}
\widehat{\mathbf{H}}_{I}^H + \sigma_z^2 \mathbf{I}$. Note that the scaling factor $\frac{1}{\hat{\mathbf{h}}_i^{H}
\left(\hat{\mathbf{h}}_i
\hat{\mathbf{h}}_i^H+\mathbf{R}_{I}\right)^{-1}
\hat{\mathbf{h}}_i}$ in the coefficient vector of the MMSE detector $\mathbf{w}_{i}$ is added to ensure an unbiased detection as indicated by (\[x\_hat\_i\_bn\]). The variance of the noise term $\hat{z}_{i}$ can be found from (\[x\_hat\_i\_bn\]) and (\[coef\_i\_bn\]) as $$\label{var_inf_no_bn}
\hat{\sigma}_{i}^{2}=\mathbf{w}_{i}^H \mathbf{R}_{I}
\mathbf{w}_{i}$$ Substituting the coefficient vector for the MMSE detector in (\[coef\_i\_bn\]) into (\[var\_inf\_no\_bn\]), the variance can be written as $$\label{var_inf_no_mmse_bn}
\hat{\sigma}_{i}^{2}=\frac{1}{\hat{\mathbf{h}}_i^H
\mathbf{R}_{I}^{-1} \hat{\mathbf{h}}_i}$$
Then, the SINR of MMSE associated with $x_i$ is $1/\hat{\sigma}_{i}^{2}$. $$\label{sinr1}
\gamma_{i}=\frac{1}{\hat{\sigma_{i}^{2}}}
=\hat{\mathbf{h}}_i^H
\mathbf{R}_{I}^{-1}
\hat{\mathbf{h}}_i$$
The closed-form BER for a channel model such as (\[x\_hat\_i\_bn\]) can be found in [@Proakis:DigComm]. The average BER over MIMO fading channel for a given constellation can be found for beam-nulling as follows. $$\label{ber_av_bn_num}
BER_{av}=E_{\gamma_{i}}\left[ \frac{1}{N_t-1} \sum_{i}
BER(\gamma_{i})\right]$$
The closed-form formula for the average BER in (\[ber\_av\_bn\_num\]) depends on the distribution of $\gamma_{i}$, which is difficult to determine. Here, the above average BER is calculated numerically. For example, the average BER for $2^\eta$-PSK is $$BER_{av}=E_{\gamma_{i}}\left[ \frac{1}{N_t-1} \sum_{i}
\frac{2}{\eta}Q\left(\sqrt{2
\eta~\gamma_{i}}~\sin(\frac{\pi}{2^{\eta}})\right)\right]$$ and the average BER for rectangular $2^{\eta}$-QAM is $$BER_{av}=E_{\gamma_{i}}\left[ \frac{1}{N_t-1} \sum_{i}
\frac{4}{\eta}Q\left(\sqrt{\frac{3\eta~\gamma_{i}}{2^{\eta}-1}}\right)\right]$$ where $Q(\cdot)$ denotes the Gaussian $Q$-function.
In Fig. \[ber\_BN\_num\], numerical and simulation results are compared for 8PSK over $3 \times 3$ Rayleigh fading channel and QPSK over $4 \times 4$ Rayleigh fading channel, respectively. As can be seen, the numerical and simulation results match well.
Performance Comparison Between Beamforming and Beam-nulling
-----------------------------------------------------------
In Fig. \[ber\_bf\_bn\_4x4\], simulation results are compared for various data rates $R$ over $4 \times 4$ Rayleigh fading channels. In the following simulations, a data rate $R$ is measured in bits per channel use. The beamforming scheme is equivalent to a SISO channel using a maximum ratio combining (MRC) receiver [@Love:Beamforming]. For the beam-nulling scheme, the optimal ML receiver and the suboptimal MMSE receiver are used.
From Fig. \[ber\_bf\_bn\_4x4\], if the data rate is low, i.e., constellation size is low, beamforming outperforms beam-nulling. If the data rate is high, i.e., constellation size is high, beam-nulling outperforms beamforming at low and medium SNR, however at high SNR beamforming outperforms beam-nulling. Also, as can be seen, at the high data rate, even the beam-nulling scheme with suboptimal MMSE receiver outperforms the beamforming scheme.
Concatenation of Beam-nulling and LDC
-------------------------------------
To further improve the performance of beam-nulling with tractable complexity, we propose to concatenate beam-nulling with a linear dispersion code. Note that to meet error-rate requirements, multiple levels of error protection can be implemented. In this study, we focus on space-time coding domain.
In this system, the information bits are first mapped into symbols. The symbol stream is parsed into blocks of length $L=(N_t-1)T$. The symbol vector associated with one modulation block is denoted by $\mathbf{x}=[x_1, x_2, \ldots, x_L]^T$ with $x_i \in \Omega\equiv\{\Omega_m | m=0, 1, \ldots, 2^\eta-1, \eta
\geq1\}$, i.e., a complex constellation of size $2^\eta$, such as $2^\eta$-QAM). The average symbol energy is assumed to be $1$, i.e., $\frac{1}{2^\eta}\sum\limits_{m=0}^{2^\eta-1}
|\Omega_m|^2 =1$. Each symbol in a block will be mapped to a dispersion matrix of size $N_t \times T$ (i.e., $\mathbf{M}_i$) and then combined linearly to form $(N_t-1)$ data streams over $T$ channel uses. The output $(N_t-1)$ data streams are transmitted only over the subspace $\mathbf{\Phi}$ orthogonal to the weakest spatial channel. The generation of the orthogonal subspace $\mathbf{\Phi}$ is described in Section \[subsec\_BN\_model\]. The received signals can be written as $$\label{BN_LDC_rec_y}
\mathbf{y}=\sqrt{\frac{P}{N_t-1}}\mathbf{H}\mathbf{\Phi}\sum\limits_{i=1}^{L}\mathbf{M}_i
x_i+\mathbf{z}$$ where $\mathbf{z}$ is additive white Gaussian noise vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma^{2}_{z}$. It is worthy to note that the traditional beamforming scheme cannot work with space-time coding since it can be viewed as a SISO channel. We compare the concatenated scheme with the original schemes at the same data rate.
In Fig. \[ber\_bf\_bn\_bl\_4x4\], simulation results are compared for various data rates $R$ over $4 \times 4$ Rayleigh flat fading channels. In the figure, “BL" denotes beam-nulling with LDC. As can be seen, beam-nulling with LDC outperforms beam-nulling without LDC using the same receiver. The performance of beam-nulling with LDC using MMSE receiver is close to that of beam-nulling without LDC using the optimal ML receiver.
Also it can be seen, if data rate is low, i.e., constellation size is low, the performance of beam-nulling with LDC can approach that of beamforming at high SNR. If data rate is high, i.e., constellation size is high, beam-nulling with LDC outperforms beamforming even when the suboptimal MMSE receiver is used.
Extended Adaptive Frameworks {#sec_ext_frame}
============================
For the beamforming and beam-nulling schemes, only one eigenvector has been fed back to the transmitter. If more backward bandwidth is available for feedback, e.g. $k$ eigenvectors, can be sent to the transmitter for adaptation. With the feedback of $k$ eigenvectors, we can extend our frameworks, which will be called multi-dimensional (MD) beamforming and MD beam-nulling. The original schemes can be referred to as 1D-beamforming and 1D-beam-nulling. To save bandwidth, $k \leq
\lfloor\frac{N_t}{2}\rfloor$ should be satisfied, where $\lfloor\cdot\rfloor$ denotes rounding towards minus infinity. That is, whether the strongest or the weakest $k$ spatial subchannels will be fed back according to the channel conditions. For example, at low SNR, $k$ strongest spatial subchannels will be fed back. At medium SNR, $k$ weakest spatial subchannels will be fed back.
MD Beamforming
--------------
For MD beamforming, $\mathbf{v}_{1}, \ldots, \mathbf{v}_{k}$ are fed back to the transmitter. $k$ symbols, saying $\mathbf{x}_k=[x_1, x_2, \ldots, x_k]^T$, are transmitted. At the receiver, the received vector can be written as $$\label{eigen_rec_y_k}
\mathbf{y}_k=\sqrt{\frac{P}{k}} \mathbf{H}[\mathbf{v}_{1} \ldots
\mathbf{v}_{k} ]\mathbf{x}_k+\mathbf{z}_k$$ where $\mathbf{z}_k$ is the additive white Gaussian noise vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma^{2}_{z}$.
Consequently, the associated ergodic channel capacity can be found as $$\label{cap_bf_k}
\bar{C}_{k,bf}=E\left[\sum\limits_{i=1}^{k}\log\left(1+\frac{P}{k~\sigma^{2}_{z}}\lambda_{i}^2\right)\right]$$ Let $\rho = {P}/{\sigma^{2}_{z}}$ denote SNR. It is readily checked that the capacity of MD beamforming is also concave and monotonically increasing with respect to SNR $\rho$. Differentiating the above ergodic capacity with respect to $\rho$, we have $$\frac{\partial \bar{C}_{k,bf}}{\partial \rho} = E\left(\sum\limits_{i=1}^{k}\frac{1}{\rho+\frac{k}{\lambda_{i}^2}}\right)$$ If $\rho \rightarrow 0$, equivalently at low SNR, it can be easily found that $$\frac{\partial \bar{C}_{(k-1),bf}}{\partial \rho} > \frac{\partial
\bar{C}_{k,bf}}{\partial \rho}, \rho \rightarrow 0$$ If $\rho \rightarrow \infty$, equivalently at high SNR, it can be easily found that $$\frac{\partial \bar{C}_{k,bf}}{\partial \rho} > \frac{\partial
\bar{C}_{(k-1),bf}}{\partial \rho}, \rho \rightarrow \infty$$ Note that $\bar{C}_{k,bf}=0$ for any $k$ when $\rho=0$ or minus infinity in dB. Hence, at low SNR, the capacity of the $k$-D beamforming scheme is worse than the $(k-1)$-D beamforming scheme and while at high SNR, the capacity of the $k$-D beamforming scheme is better than the $(k-1)$D beamforming scheme at the cost of feedback bandwidth.
MD Beam-nulling
---------------
For MD beam-nulling, similar to $1$D beam-nulling, by a certain rule, a subspace orthogonal to the $k$ weakest spatial channel is constructed. That is, the following condition should be satisfied. $$\label{orth_subspace_k}
\mathbf{v}_{n}^H\mathbf{\Phi}^{(k)}=\mathbf{0}^T, \forall
n=N_t-k+1,\ldots, N_t.$$ The $N_t \times (N_t-k)$ matrix $\mathbf{\Phi}^{(k)}=[\mathbf{g}_1
\mathbf{g}_2 \ldots \mathbf{g}_{N_t-k}]$ spans the $(N_t-k)$-dimensional subspace.
At the transmitter, $N_t-k$ symbols denoted as $\mathbf{x}^{(k)}$ are transmitted only over the orthogonal subspace $\mathbf{\Phi}^{(k)}$. The received signals at the receiver can be written as $$\label{BN_rec_y_k}
\mathbf{y}^{(k)}=\sqrt{\frac{P}{N_t-k}}\mathbf{H}\mathbf{\Phi}^{(k)}\mathbf{x}^{(k)}+\mathbf{z}^{(k)}$$ where $\mathbf{z}^{(k)}$ is additive white Gaussian noise vector with *i.i.d.* symmetrical complex Gaussian elements of zero mean and variance $\sigma^{2}_{z}$. From (\[BN\_rec\_y\_k\]), the associated instantaneous channel capacity with respect to $\mathbf{H}$ can be found as $$\label{cap_bn_k}
\bar{C}_{bn}^{(k)}=E\left[\sum\limits_{i=1}^{N_t-k}\log\left(1+\frac{P}{(N_t-k)\sigma^{2}_{z}}\lambda_{i}^2\right)\right]$$ It is readily checked that the capacity of MD beam-nulling is also concave and monotonically increasing with respect to SNR $\rho$. Let $\rho = {P}/{\sigma^{2}_{z}}$ denote SNR. Differentiating the above ergodic capacity with respect to $\rho$, we have $$\frac{\partial \bar{C}^{(k)}_{bn}}{\partial \rho} = E\left(\sum\limits_{i=1}^{N_t-k}\frac{1}{\rho+\frac{N_t-k}{\lambda_{i}^2}}\right)$$ If $\rho \rightarrow 0$, equivalently at low SNR, it can be easily found that $$\frac{\partial \bar{C}^{(k)}_{bn}}{\partial \rho} > \frac{\partial
\bar{C}^{(k-1)}_{bn}}{\partial \rho}, \rho \rightarrow 0$$ If $\rho \rightarrow \infty$, equivalently at high SNR, it can be easily found that $$\frac{\partial \bar{C}^{(k-1)}_{bn}}{\partial \rho} >
\frac{\partial \bar{C}^{(k)}_{bn}} {\partial \rho}, \rho
\rightarrow \infty$$ Note that $\bar{C}_{k,bn}=0$ for any $k$ when $\rho=0$ or minus infinity in dB. Hence, at low SNR, the capacity of the $k$-D beam-nulling scheme is better than the $(k-1)$-D beam-nulling scheme at the cost of feedback bandwidth and while at high SNR, the capacity of the $k$-D beam-nulling scheme is worse than the $(k-1)$-D beam-nulling scheme.
![ \[md\_bn\_cap\]MD beam-nulling over $5 \times 5$ Rayleigh fading channel.](Cap_MD_BN.eps){width="2.5in"}
For example, in Fig. \[md\_bn\_cap\], capacities of 1D beam-nulling and 2D beam-nulling schemes are compared with WF and equal power scheme over $5 \times 5$ Rayleigh fading channel at different SNR regions. At relatively low SNR, i.e., less than 13dB, the 2D beam-nulling scheme outperforms the 1D beam-nulling scheme in terms of capacity at the price of feedback bandwidth. While at relatively high SNR, i.e., more than 13dB, the 1D-beam-nulling scheme outperforms the 2D beam-nulling scheme as predicted.
Capacity Comparison of MD Schemes
---------------------------------
![\[md\_bf\_bn\_cap\]Comparison over $5 \times 5$ Rayleigh fading channel.](Cap_MD_BF_BN.eps){width="2.5in"}
Here, over $5 \times 5$ Rayleigh fading channel, the MD schemes are compared with WF and equal power schemes as shown in Fig. \[md\_bf\_bn\_cap\]. It can be readily check that, at relatively low SNR, MD beamforming schemes are better than MD beam-nulling schemes; while at relatively high SNR, the results are opposite. Specifically, at very low SNR, i.e. less than 0dB, the 1D beamforming scheme outperforms the other MD schemes. At the SNR region between 0dB and 5.5dB, the 2D beamforming scheme outperforms the other MD schemes. At the SNR region between 5.5dB and 12.7dB, the 2D beam-nulling scheme outperforms the other MD schemes. At the SNR region between 12.7dB and 23dB, the 1D beam-nulling scheme outperforms the other MD schemes. Again, when SNR is more than 23dB, the equal power scheme outperforms the other suboptimal schemes.
MD Schemes Concatenated with Linear Space-Time Code
---------------------------------------------------
MD beamforming scheme and MD beam-nulling scheme make $k$ and $N_t-k$ spatial subchannels available, respectively. As a result, they can concatenate with space-time schemes to improve performance. For simplicity, space-time codes with linear structure, such as high-rate LDCs [@Hassibi:LDC] and STBCs [@Tarokh:STBC] (i.e., orthogonal design), are preferable. It is worthy of noting that the 2D beamforming scheme in [@Giannakis:Optimal-STBC-Channel-Mean] is just a special case of MD beamforming. As shown in Fig. \[MD\_scheme\], we propose to concatenate an MD scheme with an LDC or an STBC. In these figures “OD" stands for orthogonal design.
![\[MD\_scheme\]Concatenated MD scheme.](MD_LDC_OD.eps){width="2.5in"}
Over $5 \times 5$ Rayleigh fading channel, concatenated MD schemes are compared at various data rate. In the simulations, two eigenvectors can be fed back to the transmitter. For an MD scheme with LDC, a suboptimal linear MMSE receiver is applied. Since a MD scheme with STBC are orthogonal, a matched filter is applied, which is also optimal.
In Fig. \[BER\_MD\_BF\], MD beamforming scheme with STBC are compared with MD beamforming scheme with LDC in terms of BER when data rate is $R=2$. Also when $R=6$, Their BERs are shown in Fig. \[BER\_MD\_BF\_BN\]. From these figures, it is shown that at high data rate, MD beamforming with LDC outperform MD beamforming with STBC significantly even though a suboptimal MMSE receiver is applied. Specifically, when BER is $10^{-5}$, the coding gain is about $4$dB. At low data rate, MD beamforming with LDC performs slightly worse than MD beamforming with STBC since the suboptimal receiver is applied. Specifically, when BER is $10^{-5}$, the coding gain is about $1$dB.
In Fig. \[BER\_MD\_BN\], MD beamforming scheme with STBC are compared with MD beamforming scheme with LDC in terms of BER when data rate is $R=3$. Also when $R=6$, Their BERs are shown in Fig. \[BER\_MD\_BF\_BN\]. From these figures, it is shown that at high data rate, MD beam-nulling with LDC outperform MD beam-nulling with STBC significantly even though a suboptimal MMSE receiver is applied. Specifically, when BER is $10^{-5}$, the coding gain is about $6.8$dB. At low data rate, MD beam-nulling with LDC performs slightly worse than MD beam-nulling with STBC since the suboptimal receiver is applied. Specifically, when BER is $10^{-5}$, the coding gain is about $1.5$dB.
In Fig. \[BER\_MD\_BF\_BN\], four schemes are compared when data rate is $R=6$. As shown in the figure, MD beam-nulling with LDC has the best BER performance even suboptimal MMSE receiver is used. In summary, MD scheme with LDC outperforms MD scheme with STBC especially when the data rate is high. At low data rate, the performance will depend on the receiver. At high data rate, MD beam-nulling with LDC perform the best among the four schemes.
![\[BER\_MD\_BF\]BER of concatenated MD beamforming when $R=2$.](MD_BF_LDC_OD_k2_R2.eps){width="2.5in"}
![\[BER\_MD\_BN\]BER of concatenated MD beam-nulling when $R=3$.](MD_BN_LDC_OD_k2_R3.eps){width="2.5in"}
![\[BER\_MD\_BF\_BN\]BER Comparison of concatenated MD schemes when $R=6$.](MD_LDC_OD_k2_R6.eps){width="2.5in"}
Conclusions {#sec_con}
===========
Based on the concept of spatial subchannels and inspired by the beamforming scheme, we proposed a scheme called “beam-nulling". The new scheme exploits all spatial subchannels except the weakest one and thus achieves significantly high capacity that approaches the optimal water-filling scheme at medium signal-to-noise ratio. The performance of beam-nulling with an MMSE receiver has been analyzed and verified by numerical and simulation results. It has been shown that if the data rate is low, beamforming outperforms beam-nulling. If the data rate is high, beam-nulling outperforms beamforming at low and medium SNR but beamforming outperforms at high SNR. To achieve better performance and maintain tractable complexity, beam-nulling was concatenated with a linear dispersion code and it was demonstrated that if the data rate is low, beam-nulling with a linear dispersion code can approach beamforming at high SNR. If the data rate is high, beam-nulling outperforms beamforming even with a suboptimal MMSE receiver. If more than one eigenvector can be fed back to the transmitter, new extended schemes based on the existing beamforming and the proposed beam-nulling are proposed. The new schemes are called multi-dimensional beamforming and multi-dimensional beam-nulling, respectively. The theoretical analysis and numeric results in terms of capacity are also provided to evaluate the new proposed schemes. Both of MD schemes can be concatenated with an LDC or an STBC. It is shown that the MD scheme with LDC can outperform the MD scheme with STBC significantly when the data rate is high. Additionally, at high data rate, MD beam-nulling with LDC outperforms MD beamforming with LDC, MD beamforming with STBC and MD beam-nulling with STBC.
[1]{} I. E. Telatar, “Capacity of multi-antenna Gaussian channels," *Eur. Trans. Telecom.*, vol 10, pp. 585-595, Nov. 1999.
G. J. Foschini, M. J. Gans, “On limits of wireless communications in a fading environment when using multiple antennas," *Wireless Personal Communications*, vol. 6, no. 3, pp. 311-335, 1998.
J. K. Cavers, “Variable-rate transmission for Rayleigh fading channels," *IEEE Transactions on Communications*, COM-20, pp.15-22, 1972.
A. J. Goldsmith and S.-G. Chua, “Variable rate variable power MQAM for fading channels," *IEEE Trans. Commun.*, vol. 45, no. 10, pp. 1218–1230, Oct. 1997.
Z. Luo, H. Gao, Y. Liu and J. Gao “Capacity Limits of Time-Varying MIMO Channels," *IEEE International Conference On Communications* vol.2, pp. 795- 799, May 2005.
W. Yu, W. Rhee, S. Boyd, and J. Cioffi, “Iterative water-filling for Gaussian vector multiple-access channels," *IEEE Trans. Inform. Theory*, vol.50, no. 1, pp. 145–152, Jan. 2004.
N. Jindal, W. Rhee, S. Vishwanath, S. A. Jafar, and A. Goldsmith, “Sum power iterative water-filling for multi-antenna gaussian broadcast channels," *IEEE Trans. Inform. Theory*, vol. 51, no. 4, pp. 1570–1580, April 2005.
M. Demirkol and M. Ingram, “Power-controlled capacity for interfering MIMO links," *in Proc. IEEE Veh. Technol. Conf. (VTC)*, Atlantic City, USA, Oct. 2001, pp. 187–191.
Z. Shen, R. W. Heath, Jr., J. G. Andrews, and B. L. Evans, “Comparison of Space-Time Water-filling and Spatial Water-filling for MIMO Fading Channels," *in Proc. IEEE Int Global Communications Conf.* vol. 1, pp. 431 – 435, Nov. 29-Dec. 3, 2004, Dallas, TX, USA.
Z. Zhou and B. Vucetic “Design of adaptive modulation using imperfect CSI in MIMO systems," *2004 Eelectronics Letters* vol. 40 no. 17, Aug. 2004.
X. Zhang and B. Ottersten, “Power allocation and bit loading for spatial multiplexing in MIMO systems," *IEEE Int. Conf.on Acoustics, Speech, and Signal Processing, 2003. Proceedings (ICASSP ’03)* vol.5 pp. 54-56, Apr. 2003.
S. Zhou and G. B. Giannakis, “Optimal transmitter eigen-beamforming and space-time block coding based on channel mean feedback" *IEEE Transactions on Signal Processing*, vol. 50, no. 10, October 2002.
S. Zhou and G. B. Giannakis, “How accurate channel prediction needs to be for transmit-beamforming with adaptive modulation over Rayleigh MIMO channels," *IEEE Trans. Wireless Comm.*, vol. 3, no. 4, pp. 1285 – 1294, July 2004.
D. J. Love, R. W. Heath, Jr. and T. Strohmer, “Grassmannian beamforming for multiple-input multiple-output wireless systems," *IEEE Trans. Inform. Theory*, vol. 49, no. 10, pp. 2735 – 2747, Oct. 2003.
J. Zheng and B. D. Rao, “Capacity analysis of MIMO systems using limited feedback transmit precoding schemes," *IEEE Trans. on Signal Processing*, vol. 56, no. 7, pp. 2886 - 2901, July 2008.
S. Zhou and G. B. Giannakis, “Adaptive modulation for multiantenna transmissions with channel mean feedback," *IEEE Trans. Wireless Comm.*, vol.3, no.5, pp. 1626-1636, Sep. 2004.
P. Xia and G. B. Giannakis, “Multiantenna adaptive modulation with beamforming based on bandwidth-constrained feedback," *IEEE Transactions on Communications*, vol.53, no.3, March 2005.
B. Mondal and R. W. Heath, Jr., “Performance analysis of quantized beamforming MIMO systems," *IEEE Transactions on Signal Processing* , vol. 54, no. 12, Dec. 2006.
T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast scheduling using zero-forcing beamforming," *IEEE J. Select. Areas in Commun.*, vol. 24, no. 3, pp. 528–541, March 2006.
S. Zhou, Z. Wang, and G. Giannakis, “Quantifying the power loss when transmit beamforming relies on finite rate feedback,” IEEE Trans. on Wireless Commun., vol. 4, no. 4, pp. 1948–1957, 2005.
J. F. Paris and A. J. Goldsmith, “Adaptive Modulation for MIMO Beamforming under Average BER Constraints and Imperfect CSI," *Proc. of Int. Conf. Comm.* ICC 2006, pp.1312-1317, June 2006.
J. K. Cavers, “Single-user and multiuser adaptive maximal ratio transmission for Rayleigh channels," *IEEE Trans. Veh. Technol.*, vol. 49, no. 6, pp. 2043–2050, Nov. 2000.
V. Tarokh, N. Seshadri, and A. Calderbank, “Space-time codes for high data rate wireless communications: Performance criterion and code construction," *IEEE Trans. Inform. Theory*, vol. 44, pp. 744-765, Mar. 1998.
S. Alamouti, “A simple transmitter diversity scheme for wireless communications," *IEEE J. Select. Areas Commun.*, vol. 16, pp. 1451-1458, Oct. 1998.
V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block code from orthogonal designs,” *IEEE Trans. Inform. Theory*, vol. 45, pp. 1456-1467, July 1999.
B. Hassibi and B. Hochwald, “High-rate codes that are linear in space and time," *IEEE Trans. Inform. Theory*, vol. 48, pp. 1804-1824, July 2002.
R. W. Heath and A. Paulraj, “Linear dispersion codes for MIMO systems based on frame theory," *IEEE Trans. on Signal Processing*, vol. 50, No. 10, pp. 2429-2441, October 2002.
X. Ma and G. B. Giannakis, “Full-diversity full-rate complex-field space-time coding," *IEEE Trans. Signal Processing*, vol. 51, no. 11, pp. 2917-2930, July 2003.
Z. Wu and X. F. Wang, “Design of coded space-time modulation," *IEEE International Conference on Wireless Networks, Communications and Mobile Computing*, vol. 2, pp. 1059-1064, Jun. 13-16, 2005.
C. E. Shannon, “A mathematical theory of communication”, *Bell Syst. Tech. J.*, vol. 27, pp. 379–423 (Part one), pp. 623–656 (Part two), Oct. 1948, reprinted in book form, University of Illinois Press, Urbana, 1949.
R. Lupas and S. Verdu, “Linear multiuser detectors for synchronous code-division multiple-access channels," *IEEE Trans. inform. Theory*, vol. 35, pp. 123-136, Jan. 1989.
H. V. Poor and S. Verdu, “Probability of error in MMSE multiuser detection," *IEEE Trans. inform. Theory*, vol. 43, pp. 858-871, May 1997.
R. Bohnke and K. Kammeyer, “SINR Analysis for V-BLAST with Ordered MMSE-SIC Detection," *International Wireless Communications and Mobile Computing Conference*, pp. 623-628, July 2006.
J. Proakis, *Digital Communications*, 4th ed. 1em plus 0.5em minus 0.4emNew York: McGraw-Hill, 2001.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The physical solutions of Lagrangian of octonionics are researched in the paper. It is shown, the gravitational interaction in Friedmann space and in spherically symmetric space in such model is to be described by pair of charged massless vectorial D-bosons of Minkowski space. It is proposed to use the formalism for the description of jet and supernova.'
author:
- |
V. Yu. Dorofeev[^1]\
Laboratory for theoretical physics
title: Octonionic Physics
---
Introduction
============
In [@Dor1] author proposed the generalization of Lagrangian from Weinberg-Salam theory, devoted to electroweak interactions, on octonionic algebra. This paper is a corollary of the conducted research. Here the physical interpretation of the octonionic Lagrangian from [@Dor1] is proposed. The subject of research is the opportunity to describe gravitational interaction on the base of octonionic algebra in the flat Minkowski space.
Algebraic structure of the state space\
of the extended octonionic space
=======================================
The octonionic algebra could be represented by matrices [@Dor1] ($i=1,2,3$) $$\label{Zornpr}
\matrix{\Sigma^0=\left(\matrix{\sigma^0&0\cr0&\sigma^0}\right)&\Sigma^i=
\left(\matrix{0&-i\sigma^i\cr i\sigma^i&0}\right)\cr
\Sigma^4=\left(\matrix{-1&0\cr0&1}\right)&
\Sigma^{4+i}=\left(\matrix{0&-\sigma^i\cr -\sigma^i&0}\right)}$$ where $\sigma^0$ is a unit matrice and $\sigma^i,i=1,2,3$ is a Pauli matrice: $$\label{matDir}
\sigma^1=\left(\matrix{0&1\cr1&0}\right),\quad
\sigma^2=\left(\matrix{0&-i\cr i&0}\right),\quad
\sigma^3=\left(\matrix{1&0\cr0&-1}\right),\quad
\sigma^0=\left(\matrix{1&0\cr0&1}\right)$$
with special multiplication rule $$\u*\u'=\left(\matrix{\lambda I&A\cr B&\xi I}\right)*
\left(\matrix{\lambda' I&A'\cr B'&\xi' I}\right)=$$ $$\label{Dab}
=\left(\matrix{(\lambda\lambda'+\frac12\tr(AB'))I\hfill&\lambda
A'+\xi' A+\frac i2[B,B']\hfill\cr\lambda'B+\xi B'-\frac
i2[A,A']&(\xi\xi'+\frac12\tr(BA'))I\hfill}\right)$$
The matrices $A,A',B,B$ with size $(2\times2)$ were introduced, together with unit matrice $I$ of same size and complex numbers $\lambda,\xi,\lambda',\xi'$.
Hermitian conjugation on those matrices is introduced as follows: $$\label{Dab2}
\u=\left(\matrix{\lambda I&A\cr B&\xi I}\right),\qquad
\u^+=\left(\matrix{\lambda^*I&B^+\cr A^+&\xi^*I}\right)$$
Multiplication (\[Dab\]) is defined for a wider matrice class than for matrices from octonionic algebra representation, $\Sigma^a,a=0,1,2,\dots,7$ introduced earlier. Two more matrices could be introduced $$\label{dmo}
f^8=\left(\matrix{0&I\cr
I&0}\right),\quad f^9=\left(\matrix{0& iI\cr-iI&0}\right)$$ That maintain multiplication (\[Dab\]).
So the set of matrices $$\label{lnbaz}
f^a=\Sigma^a,\quad a=0,1,\dots,7,\quad f^8,\quad f^9$$ Form a basis of linear space upon the field of complex numbers, which is further called the extended octonionic space, and is denoted for $\overline\O$.
Notice $$(f^8*f^1)*(f^1*f^8)=0,\quad f^8*(f^1*f^1)*f^8=1$$
Using the multiplication defined in (\[Dab\]) introduce scalar product in $\overline\O$: $$(\u_1,\u_2)=\frac12\tr(\u_1^+*\u_2)=\frac12\tr(\left(\matrix{\lambda_1^*I&B_1^+\cr A_1^+&\xi_1^*I}\right)*
\left(\matrix{\lambda_2I&A_2\cr B_2&\xi_2I}\right))$$ $$\label{skpro}
=\frac12\tr(\left(\matrix{(\lambda_1^*\lambda_2+\frac12\tr(B_1^+B_2))I\hfill&
\lambda_1^*A_2+\xi_2 B_1^++\frac i2[A_1^+,B_2]\hfill\cr
\lambda_2A_1^++\xi_1^*B_2-\frac
i2[B_1^+,A_2]&(\xi_1^*\xi_2+\frac12\tr(A_1^+A_2))I\hfill}\right))$$ $$=\lambda_1^*\lambda_2+\xi_1^*\xi_2+\frac12\tr(B_1^+B_2))
+\frac12\tr(A_1^+A_2)$$
The norm of vector $\o$ in extended octonionic space $\overline\O$ $$\u=\left(\matrix{\lambda^*I&B^+\cr A^+&\xi^*I}\right)$$ is the number $$||\u||=\sqrt{|\lambda|^2+|\xi|^2+\frac12\tr(A^+A)+\frac12\tr(B^+B))}$$
There is an orthonormal basis with respect to scalar product (\[skpro\]) $$\label{obbaz}
e^0=\left(\matrix{I&0\cr0&0}\right),\quad
e^i=\left(\matrix{0&\sigma^i\cr0&0}\right),\quad
e^4=\left(\matrix{0&0\cr0&I}\right)$$
$$e^{4+i}=\left(\matrix{0&0\cr\sigma^i&0}\right),\quad
e^8=\left(\matrix{0&I\cr0&0}\right),\quad
e^9=\left(\matrix{0&0\cr I&0}\right)$$
Obviously, the trace of the production of any three matrices is associative one, i.e. $$\tr((A*B)*C)=\tr(A*(B*C))=\tr(A*B*C)$$
Introduce a matrice $$\label{mo}
u_0=\frac14\left(\matrix{0&i\sigma^3\cr
0&I}\right)=\frac12(\Sigma^0+\Sigma^4-\Sigma^3-i\Sigma^7)$$
Due to high importance of $u_0$ in the construction of octonionic Lagrangian, find the scalar product of this matrice with generatrices $\Sigma^a$ of octonionic algebra $$\tr((u_0^+*(\Sigma^a*\Sigma^b)*u_0),\qquad\tr((u_0^+*\Sigma^a)*(\Sigma^b*u_0))$$ (Evidently, the last brackets disposition is the corollary of the present ones and the associativity of trace of a three matrices.)
$$\begin{matrix}{
b=1,\dots,7&\tr(u_0^+*\Sigma^b*u_0)=&-\delta^{3b}\cr
a,b=1,2,3&\tr((u_0^+*\Sigma^a)*(\Sigma^b*u_0))=&\delta^{ab}\cr
a,b=1,2,3&\tr(u_0^+*(\Sigma^a*\Sigma^b)*u_0)=&\delta^{ab}-i\varepsilon^{4ab7}\cr
a,b=5,6,7&\tr((u_0^+*\Sigma^a)*(\Sigma^b*u_0))=
&\delta^{ab}\cr
a,b=5,6,7&\tr(u_0^+*(\Sigma^a*\Sigma^b)*u_0)=
&\delta^{ab}-i\varepsilon^{4ab7}\cr
a=4,b=1,2,3&\tr(u_0^+*(\Sigma^4*\Sigma^b)*u_0)=&0\cr
a=4,b=1,2,3&\tr((u_0^+*\Sigma^4)*(\Sigma^b*u_0))=&0\cr
a=4,b=5,6,7&\tr(u_0^+*(\Sigma^4*\Sigma^b)*u_0)=&-i\delta^{7b}\cr
a=4,b=5,6,7&\tr((u_0^+*\Sigma^4)*(\Sigma^b*u_0))=&-i\delta^{7b}\cr
a=1,2,3,b=5,6,7&\tr((u_0^+*\Sigma^a)*(\Sigma^b*u_0))=&0\cr
a=1,2,3,b=5,6,7&\tr(u_0^+*(\Sigma^a*\Sigma^b)*u_0))=
&-i\varepsilon^{4ab3}\cr
a=5,6,7,b=1,2,3&\tr((u_0^+*\Sigma^a)*(\Sigma^b*u_0))=&0\cr
a=5,6,7,b=1,2,3&\tr(u_0^+*(\Sigma^a*\Sigma^b)*u_0))=
&-i\varepsilon^{4ab3}}
\end{matrix}$$
It is clear from the equations if $a$ and $b$ are equal to 4 the multiplication is associative, otherwise with $a\ne b$ the four-member-production is not associative therefore could be rewritten as follows $$\label{moso}
\tr(u_0^+*\Sigma^a*\Sigma^b*u_0)=\delta^{ab}+ic_c^{ab}A^{ab}_c$$
Here, $A^{ab}_c,(c=1,2)$ is antisymmetrical variable equal to $0,\pm1$. In only case when neither $a$ nor $b$ is equal to 4 and special brackets disposition the value of $A^{ab}_c$ differs from zero $$\label{moso2}
a\ne b,\qquad\tr(u_0^+*(\Sigma^a*\Sigma^b)*u_0)=iA^{ab}_2$$
$$\tr((u_0^+*\Sigma^a)*(\Sigma^b*u_0))=iA^{ab}_1=0$$ And its non-zero values are $$\label{moso0}
A^{21}_2=A^{65}_2=A^{25}_2=A^{74}_2=A^{74}_1=1$$ And if $a=4$ then the order of brackets is does not matter $$\label{moso3}
\tr((u_0^+*\Sigma^4)*(\Sigma^b*u_0))=
\tr(u_0^+*(\Sigma^4*\Sigma^b)*u_0)$$ While constants $c_c^{ab}$ are defined by the model of brackets disposition in the production of octonionic operands, i.e. by relief of non-associativity. Three models of it are proposed below.
[**Probability model of non-associativity relieving.**]{} $$\label{vm0}
\tr(A*B*C*D)=p_1\tr(((A*B)*C)*D)+p_2\tr(A*(B*C)*D)$$ $$p_1+p_2=1$$
Here the frequencies $p_1$ and $p_2$ are introduced. Their values depend on the number of brackets permutations leading to the same result in (\[vm0\]). For example, in (\[vm0\]) it is natural to admit $c_1^{ab}=p_1=3/4,c_2^{ab}p_2=1/4$ because three different types of brackets disposition lead to the same result in the first member in the right part of the equation (14) whereas only one type - in the second member of the right part of (\[vm0\]).
Similarly, the probability model of non-associativity relieving for (\[moso\] gives $$\label{vm01}
\tr(u_0^+*\Sigma^a*\Sigma^b*u_0)=\delta^{ab}+\frac{3i}4A^{ab}_1+\frac i4A^{ab}_2$$
The definition given in (\[vm0\]) is also applicable for greater number of elements.
[**Minimal model of non-associativity relieving.**]{} $$\label{vm2}
f(ABCD)\tr(A*B*C*D)=\min_{()}(f(ABCD)\tr(A*B*C*D))$$
The right part of (\[vm2\]) means the choice of brackets disposition is done to minimize $f(ABCD)\tr(A*B*C*D)$. There the coefficient $f(ABCD)$ (which could be equal to 1) is introduced before $\tr$. Notice, the coefficient before $\tr$ does not necessarily require its incorporation in the definition of the minimal model.
[**The maximal model**]{} is to be defined similarly.
Obviously, in associative case the same result is obtained to that of the models defined above.
It is easy to ensure, with $i\ne j,k\ne l$ $$\label{chp5}
\frac14\tr(\Sigma^i*\Sigma^j*\Sigma^k*\Sigma^l)=
-\varepsilon^{ijkl}\delta^{ik}\delta^{jl}+A^{ijkl}$$ where $A^{ijkl}$ equals to $\pm1,0$ regarding to the brackets disposition in production and the values of indices.
Lagrangian of octonionics
=========================
In [@Dor1] the generalization of Weinberg-Salam Lagrangian on octonionic algebra is proposed as follows: ($a,b=0,1,\dots,7,k=1,2,\dots,7$): $$\L_{oct.}=\L_f+(\partial_\mu\bux\Psi\varphi-
\frac i2q^aA_\mu^{a}\bux\Psi\varphi*\Sigma^{a})
*(\partial^\mu\Psi_\varphi+
\frac i2q^{b}A^{\mu(b)} \Sigma^{b}*\Psi_\varphi)$$ $$+\frac i2\overline L*\gamma_\mu(\overrightarrow\partial^\mu
L+\frac i2c_Lq^kA^{\mu(k)}\Sigma^k*L+\frac
i2c_{L_0}q^0A^{\mu(0)}L)$$ $$-\frac i2\overline L*\gamma_\mu(\overleftarrow\partial^\mu L-
\frac i2c_Lq^kA^{\mu(k)}\Sigma^{k}*L-\frac
i2c_{L_0}q^0A^{\mu(0)}L)$$ $$+\frac i2\overline R\gamma_\mu(\overrightarrow\partial^\mu R+
iq^0A^{\mu0}R)-\frac i2\overline
R\gamma_\mu(\overleftarrow\partial^\mu R-iq^0A^{\mu0}R)$$ $$\label{vst}
-\tilde h\overline L*\Psi_\varphi R-\tilde h\overline
R\bux\Psi\varphi*L)+m^2||\Psi_\varphi||^2-\frac f4||\Psi_\varphi||^4$$
There following denotes are done. Free fields Lagrangian $\L_f$ $$\label{svs0}
\L_f=-\frac14F^0_{\mu\nu}F^{\mu\nu(0)}-\frac14\tr(F_{\mu\nu}^k*F^{\mu\nu(k)})
$$ $$+\frac1{16}f^{ijkl}(A_\mu^iA_\nu^j-A_\nu^j
A_\mu^i)(A^{\mu(k)}A^{\nu(l)}-A^{\nu(k)}A^{\mu(l)})$$ $$\label{l04}
f^{ijkl}=q^{ij}q^{kl}\tr(\Sigma^i*\Sigma^j*\Sigma^k*\Sigma^l)$$ $$F^0_{\mu\nu}=\partial_\mu A^0_\nu-\partial_\nu A^0_\mu,\qquad
F_{\mu\nu}^k=\partial_\mu A_\nu^k-\partial_\nu A_\mu^k-\varepsilon^{ijk}q^{ij}(A_\mu^iA_\nu^j-A_\nu^iA_\mu^j)$$
The value $f^{ijkl},i,j,k,l=1,\dots,7$ reflects the non-associate character of free octonionic fields Lagrangian.
Left and right spinor components $$\label{lpm}
\frac12(1+\gamma^5)\Psi=L,\qquad\frac12(1-\gamma^5)\Psi=R,\qquad\gamma^5=i\gamma^0\gamma^1\gamma^2\gamma^3$$
$\Psi,\Psi_\varphi$ are the vectors from the generalized octonionic state space, $q^a,c_L,c_{L_0}$ are some numbers, caused by normalization; $\gamma^\mu$ – are Dirac matrices ($\mu=0,1,2,3$): $$\label{matDir2}
\gamma^0=\left(\matrix{I&0\cr0&-I}\right),\qquad
\gamma^i=\left(\matrix{0&\sigma^i\cr-\sigma^i&0}\right), \quad
i=1,2,3$$
The summing up is carried out with metric tensor of Minkowski space $\eta_{\mu\nu}$ with signature $(1,-1,-1,-1)$, by different-case Greek indices, while the simple summing up is conducted by the same-case indices.
It is also shown in [@Dor1] that the vectors from extended state space $\bar\O$ $$\label{matDir3}
\Psi_\varphi=\Psi_0=\frac{m}{\sqrt 2f}\left(\matrix{0&i\sigma^3\cr
0&I}\right)=\frac{\sqrt2m}{\sqrt f}u_0,\qquad R(x)=e^-_R(x)$$
$$\Psi=\left(\matrix{(\alpha_1\nu+\alpha_2e)I&A_1\nu(x)+A_2e(x)\cr
B_1\nu(x)+B_2e(x)&(\beta_1\nu+\beta_2e)I}\right)$$ and numbers $$c_{L_0}=-1,\quad c_0^2=\frac{32}{257},\quad
c_L=(c_0^2)^{-1},\tilde h=2\sqrt2h/c_0,$$ $$q^1=q^2=q^3=q^{12}=g,\qquad q^0=g'$$ With $k=1,2,3$ in (\[lpm\]) give model from electron-neutrino sector of Weinberg-Salam theory with group charge of $SU(2)$-symmetry $g$ and charge $g'$ of symmetry $U(1)$, Higgs field $\Psi_0$, and the constant of interaction of Higgs field with matter fields $h$, see [@Okun].
Once more referring to [@Dor1], the expression for octonionic Lagrangian in case $\Psi_\varphi=\Psi_0$ is obtained there $$\L_{oct}=\L_f+
\frac{q^{(k)2}m^2}{2f}A_\mu^{k}A^{\mu(k)}+\frac{o^{ij}m^2}{2f}A_\mu^{i}A^{\mu(j)}+
\frac{g^{(1)2}m^2}{2f}B_\mu B^\mu
-\frac{gg^{(1)}m^2}fA_\mu^3B^\mu$$ $$+\frac{g^{(1)}}2\overline\nu_L\gamma_\mu B^\mu\nu_L+
\frac{g^{(1)}}2\overline e_L\gamma_\mu B^\mu e_L+\frac g2\overline
e_L\gamma_\mu A^{\mu3}e_L-\frac g2\overline\nu_L\gamma_\mu
A^{\mu3}\nu_L$$ $$-\frac g2\overline\nu_L\gamma_\mu e_L(A^{\mu1}-iA^{\mu2})-
\frac g2\overline e_L\gamma_\mu\nu_L(A^{\mu1}+iA^{\mu2})$$ $$+\frac i2(\overline e_L\gamma_\mu\partial^\mu e_L-\partial^\mu\overline
e_L\gamma_\mu e_L)+\frac
i2(\overline\nu_L\gamma_\mu\partial^\mu\nu_L-
\partial^\mu\overline\nu_L\gamma_\mu\nu_L)+\frac{m^4}f$$ $$+\frac i2(\overline e_R\gamma_\mu\partial^\mu
e_R-\partial^\mu\overline e_R\gamma_\mu e_R)+g^{(1)}\overline
e_R\gamma_\mu B^\mu e_R-\frac{\sqrt2hm}{\sqrt f}(\overline
e_Le_R+\overline e_Re_L)$$ $$-q^4A^{\mu(4)}
(\kappa_1\overline\nu_L\gamma_\mu\nu_L- \kappa_2\overline
e_L\gamma_\mu e_L)-
\frac32q^6A^{\mu(6)}\overline e_L\gamma_\mu e_L$$ $$\label{plna}
-\frac54(q^6A^{\mu(6)}+iq^5A^{\mu(5)})\overline\nu_L\gamma_\mu
e_L-\frac54(q^6A^{\mu(6)}-iq^5A^{\mu(5)})\overline e_L\gamma_\mu
\nu_L$$ Where the following notation is used $ic_2^{ij}A^{ij}q^iq^j=o^{ij}$.
So, the final Lagrangian contains non-associative elements along with associative.
Octonionic Lagrangian research
==============================
1\. Consider non-associative summands from Lagrangian (24). First of all, it is the quadratic member by the fields $A_\mu^k$ $$\label{ina1}
\frac{o^{ij}m^2}{2f}A_\mu^{i}A^{\mu(j)}=ic_k^{ij}A^{ij}_k \frac{q^iq^jm^2}{2f}A_\mu^{i}A^{\mu(j)}$$
Apply probability model of non-associativity relieving to this member. Assume for non-zero components $c^{ij}_1=c^{ji}_1=3/4, c^{ij}_2=c^{ji}_2=1/4$ Hence due to symmetry of the expression $q^iq^jA_\mu^{i}A^{\mu(j)}$ with respect to $i,j$ (there is no summing up by them) and the anti-symmetry of multiplier $A^{ij}_k$ (again with respect to $i,j$) this member equals zero.
2\. The free fields Lagrangian Lf contains non-associativity of fourth order with respect to fields $A_\mu$. $$\label{ina2}
f^{ijkl}(A_\mu^iA_\nu^j-A_\nu^j
A_\mu^i)(A^{\mu(k)}A^{\nu(l)}-A^{\nu(k)}A^{\mu(l)})$$
Formally the summand in Lagrangian is to be considered as potential energy (similarly to $\lambda\varphi^4$), therefore the maximal scheme of brackets disposition is applicable here (assuming $L=T-V$). When considering that member with minus then the model of brackets would be minimal, therefore implying the physical application of the problem, in what follows the brackets disposition according to $\min$ or $\max$ rule is called [**potential model**]{}.
3\. Lagrangian in (\[plna\]) contains a neutral current $$\label{is3}
-q^4A^{\mu(4)}(\kappa_1\overline\nu_L\gamma_\mu\nu_L- \kappa_2\overline
e_L\gamma_\mu e_L)$$ which interacts with left spinors only. In this manner the neutral vectorial boson $C_\mu$ is defined $$\label{is2}
q_CC_\mu=-q^4A^{\mu(4)}$$
4\. Lagrangian (\[plna\]) contain current $$\label{is1}
-\frac54\overline\nu_L\gamma_\mu(q^6A^{\mu(6)}+iq^5A^{\mu(5)})
e_L-\frac54(q^6A^{\mu(6)}-iq^5A^{\mu(5)})\overline e_L\gamma_\mu\nu_L$$ which indicates the necessity of introduction of two oppositely charged vectorial bosons $$\label{is20}
q_DD_\mu=-\frac54q^6A^{\mu(6)}-i\frac54q^5A^{\mu(5)}$$
$$q_D\buu D_\mu=-\frac54q^6A^{\mu(6)}+i\frac54q^5A^{\mu(5)}$$ With charge $\pm q_D$.
Hence, it is necessary to modify the current $q^6A^{\mu(6)}\overline e_L\gamma_\mu e_L$ in as follows $$\label{is30}
-\frac32q^6A^{\mu(6)}\overline e_L\gamma_\mu e_L=$$
$$-\frac34\overline e_L\gamma_\mu(q^6A^{\mu(6)}+iq^5A^{\mu(5)})e_L) -\frac34(q^6A^{\mu(6)}-iq^5A^{\mu(5)})\overline e_L)\gamma_\mu e_L$$
5\. Lagrangian (\[plna\]) contains Lagrangians of left electron fields, which interact with vector field $A^{\mu(6)}$ $$\label{is4}
\frac i2(\overline e_L\gamma_\mu\partial^\mu e_L-\partial^\mu\overline e_L\gamma_\mu e_L)-\frac32q^6A^{\mu(6)}\overline e_L\gamma_\mu e_L$$ which is equivalent (in analogue to electro-magnetic field) to the expression with long derivative $$\label{is5}
\partial^\mu-i\frac34q^6A^{\mu(6)}$$ And given Remark 4 the long derivative transforms to
$$\label{is6}
\partial^\mu-i\frac34q^6A^{\mu(6)}-\frac34q^5A^{\mu(5)}$$
To summarize, the Lagrangian of matter fields, which interact with vector fields can be rewritten as follows $$\label{is7}
\frac i2\overline e_L\gamma_\mu\partial^\mu e_L-\frac i2\partial^\mu\overline e_L\gamma_\mu e_L-\frac32q^6A^{\mu(6)}\overline e_L\gamma_\mu e_L=$$
$$=\frac i2\overline e_L\gamma_\mu((\partial^\mu-i\frac34q^6A^{\mu(6)}-\frac34q^5A^{\mu(5)}) e_L)-$$
$$-\frac i2((\partial^\mu+i\frac34q^6A^{\mu(6)}-\frac34q^5A^{\mu(5)})\overline e_L)\gamma_\mu e_L$$
Tetrad representation
=====================
Consider a differentiable Riemannian manifold $M$ with local coordinates $x^\mu(p)$, $p\in M,\mu=0,1,2,3$ and metrics $$\label{mpr}
ds^2=g_{\mu\nu}dx^\mu dx^\nu$$ Covariant derivative is defined as follows $$\label{kp}
A^\mu_{;\nu}=A^\mu_{,\nu}+\Gamma^\mu_{\nu\lambda}A^\lambda$$ And it is a vector. Christoffel symbols of second sort $$\label{sk}
\Gamma^\mu_{\nu\lambda}=\frac12g^{\mu\kappa}(g_{\mu\kappa,\nu}+
g_{\nu\kappa,\lambda}-g_{\lambda\nu,\kappa})$$ Do not complete a tensor. However, what is subjected to tensors transformation law is the Christoffel symbols combination - Riemannian tesnor, which is defined according to [@Eisenhart] $$\label{tkr}
R^\tau_{\mu\nu\lambda}=\Gamma^\tau_{\mu\lambda,\nu}-
\Gamma^\tau_{\mu\nu,\lambda}+\Gamma^\tau_{\sigma\nu}\Gamma^\sigma_{\mu\lambda}-
\Gamma^\tau_{\sigma\lambda}\Gamma^\sigma_{\mu\nu}$$
Quadratic form (\[mpr\]) could be diagonalized in some new coordinate system. Associating surface parameters with physical space-time parameters, and assuming in some coordinates metrics is transformed to the Minkowski space metric type: $$\label{mpm}
ds^2=c^2dt^2-dx^2-dy^2-dz^2=\eta_{ab}dx^adx^b$$
In each point of space-time introduce the tetrads $h^\mu_a=\partial x^\mu/\partial x^a,a=0,1,2,3$ which implement such transformation. Tetrads comply with a natural orthogonality property $$\label{v010}
h^b_\mu h^\mu_{a,\nu}=\delta^b_a,\qquad h^{\mu(a)}h^\nu_a=g^{\mu\nu}$$ and with the properties to lift and to sink indices. $$\label{v12}
\delta A^\mu=\delta(A^ah^\mu_a)=\delta A^ah^\mu_a+
A^ah^\mu_{a,\nu}\delta x^\nu=\Gamma^\mu_{\nu\lambda}A^\nu\delta x^\lambda$$
In each point of Riemannian space contravariant vector A... could be considered in diagonal coordinates system $A^a=A^\mu
h^a_\mu$. ”читыва€ $h^b_\mu h^\mu_{a,\nu}+h^b_{\mu,\nu} h^\mu_a=0$, find $$\label{t2}
\delta A^b=\gamma^b_{ac}A^a\delta x^c,\qquad\gamma^b_{ac}=\eta^{bd}
\gamma_{dac}=h_{\mu(d);\nu}h^\mu_ah^\nu_c$$ where $\gamma^b_{ac}$ are Ricci coefficients.
Well-known spinor transformation $\Psi(x)$ in case of one likes Lorentz transform [@Ahieser] $$\label{v012}
\Psi\to\Psi+\frac12\omega^{ab}\sigma^{ab}\Psi,\qquad\sigma^{ab}=\frac14[\gamma^a,\gamma^b],\qquad a,b=0,1,2,3$$ Where the constant matrix of infinitesimal transformation $\omega^{ab}$, defining Lorentz rotation, is introduced. Dirac matrices comply with the following multiplication rule on matrices $\sigma^{ab}$: $$\label{v012d}
\gamma^a\sigma^{bc}=\frac14\gamma^a[\gamma^b,\gamma^c]=\frac12\eta^{ab}\gamma^a-\frac12\eta^{ac}\gamma^b-\frac i2\varepsilon^{dabc}\gamma^5\gamma_d$$
It is well-known, $\overline\Psi\gamma^\mu\Psi$ is a vector regarding to overall-coordinates transformation [@Ahieser]. The latter imposes restrictions on spinors transformations. To concord overall-coordinate transformation with Lorentz transform replace the infinitesimal matrix $\omega^{ab}$ with the matrix depending on $x,\omega^{ab}=\omega^{ab}(x)$ in analogue with local gauge symmetry [@Utiyama]. As a result, find the general expression for derivative in locally flat Minkowski space, which would be a vector to overall-coordinates transformation $$\label{ivsok1}
\partial_a\Psi\to D_a\Psi=\partial\Psi/\partial x^a-i\Phi_a\psi-\Gamma_a\Psi$$
This transformation first appears in the work of Fock-Ivanenko [@FokIvanenko]. It was shown there, that $\Gamma_a$ is a real value depending on rotation coefficients $\gamma_{abc}$ and Dirac matrices and $\Phi_a$ is a real value proportional to a unit matrix. Furthermore in [@FokIvanenko] $\Phi_a$ is identified exceptionally with electromagnetic field. Matrix $\Gamma_a$$$\label{v013}
\Gamma_a=-\frac12\gamma_{abc}\sigma_{bc}$$
Particularly, considered in orthogonal coordinates, gravitational field in Riemannian space with metrics $$\label{ivsok2}
ds^2=H^{(0)2}dx^{(0)2}-
H^{(1)2}dx^{(1)2}-H^{(2)2}dx^{(2)2}-H^{(3)2}dx^{(3)2}$$ And given (\[v012d\]) lead to the following $$\label{ivsok20}
-\gamma^a\Gamma_a=
\frac14h^\mu_ah^\nu_bh_{(c)\nu;\mu}(\eta^{ab}\gamma^a-\eta^{ac}\gamma^b- i\varepsilon^{dabc}\gamma^5\gamma_d)$$
Recall to the metrics (\[ivsok2\]) is diagonal therefore tetrads $h_{\mu(a)}$ are also diagonal, then with different $a,b,c$ it takes $$\label{ivsok201}
\Gamma^\lambda_{\mu\nu}h_{c\lambda}h^\mu_bh^\nu_c=\frac12g^{\mu\kappa}(g_{\mu\kappa,\nu}+
g_{\nu\kappa,\lambda}-g_{\lambda\nu,\kappa})h_{c\lambda}h^\mu_bh^\nu_c=0$$ hence $$\label{ivsok202}
-\gamma^a\Gamma_a=\frac14h^\mu_ah^{\nu(a)}h_{(c)\nu;\mu}\gamma^c-\frac14
h^{\mu(a)}h^{\nu(b)}h_{(a)\nu;\mu}\gamma^b$$
$$=\frac14h^\mu_{c;\mu}\gamma^c+\frac14h^{\mu(a)}h^\nu_{b;\mu}h_{(a)\nu;\mu}\gamma^b=\frac12h^\mu_{c;\mu}\gamma^c$$ and Dirac equation is converted to Ivanenko-Sokolov equation [@SokolovIvanenko]: $$\label{ivsok3}
(i\gamma^a(H^a)^{-1}(\partial_a-i\Phi_a+\frac12
\partial_a\left(\ln\frac{\sqrt{-g}}{H^a}\right))+m)\psi=0$$ Assume Dirac equation is rewritten in Cartesian coordinate system. In that case we identify field $\Phi_a$ с $A^6_a$ with $A^6_a$ and field $\Gamma_a$ with $A^5_a$ (in absence of EMF). This analogy follows from Dirac equations (\[is7\]) and (\[ivsok3\]) comparison
So, in orthogonal coordinates it takes (\[ivsok2\]): $$\label{pdg}
q_DA_a^5(x)=\frac12\partial_a\left(\ln\frac{\sqrt{-g}}{H^a}\right),
\qquad \sqrt{-g}=H_0H_1H_2H_3$$
Therefore $$\label{a52}
H^a=e^{\frac23q_D\oint A_b^5dx^b-2\int A_a^5dx^a},\qquad
\sqrt{-g}=e^{\frac23q_D\oint A_b^5dx^b}$$
(In the integral of kind $\int$ it is assumed the absence of summing up by index $a$, and in the integral of kind $\oint$ vice versa is assumed the absence of summing up by $b$.)
Actually curvilinear space appears as a solution method for the Dirac equation in external field $A_\mu^5$, but as free variables $x^k$ have clear physical meaning, it is so proved the generalization of Weinberg-Salam theory on Caley octaves algebra, by means proposed in this paper, is equal to introduction of curvilinear space which is the background for matter fields consideration.
Friedmannian space
==================
In this section it is shown in the flat Friedmannian space there is a self-consistent solution of octonionics Lagrangian.
Consider homogeneous and isotropic Universe
$$\label{oiv}
ds^2=dx^{(0)2}-a^2(t)(dx^{(1)2}+dx^{(2)2}+dx^{(3)2})=
a^2(\eta)(d\eta^2-dl^2),$$
with the conformal time $dt=a(\eta)d\eta$, and wit $$\label{f1}
g_{00}=a^2(\eta),\quad g_{\alpha\beta}=a^2(\eta)\eta_{\alpha\beta}$$
By means of (\[sk\]) find all the nonzero components of Christoffel symbol: $$\label{f2}
\Gamma^0_{00}=\frac{a'}{a^3},\quad\Gamma^0_{\alpha\beta}=-\frac{a'}{a^3}g_{\alpha\beta},\quad\Gamma^\alpha_{0\beta}=\frac{a'}a\delta^\alpha_\beta$$ and estimate the value $$\label{f3}
G=g^{\mu\nu}(\Gamma^\lambda_{\mu\nu}\Gamma^\kappa_{\lambda\kappa}- \Gamma^\lambda_{\mu\kappa}\Gamma^\kappa_{\nu\lambda})
=\frac{6{a'}^2}{a^4},$$ where the stroke means the derivative with respect to conformal time.
From (\[pdg\]) it takes $$\label{f30}
q_DA_\mu^5=(\frac{da}{a^2d\eta},\vec 0)$$
Given $A_\mu^5$ and $A_\mu^6$ are united in one vectorial field $D_\mu$, deduce $A_\mu^5=A_\mu^6$ in case under consideration.
Assume $$\label{f4}
\frac{q^{47}q^{56}}{q_D^2}A^{(4)\mu}A^{(7)}_\mu=Constant=-\frac32\gamma$$ then non-associative on fields Lagrangian member (\[plna\]) can rewritten $$\label{f6}
\frac1{16}f^{ijkl}(A_\mu^iA_\nu^j-A_\nu^j
A_\mu^i)(A^{\mu(k)}A^{\nu(l)}-A^{\nu(k)}A^{\mu(l)})$$
$$=4q^{47}q^{56}A^{(4)\mu}A^{(7)}_\mu A^{(5)\nu}A^{(6)}_\nu=
-6\gamma\left(\frac{da}{a^2d\eta}\right)^2=-6\gamma\left(\frac{a'}{a^2}\right)^2$$ Estimated using potential scheme of non-associativity relieving. Associate this expression with gravitation field Lagrangian$$\label{f7}
\L_{grav.}=4q^{47}q^{56}A^{(6)\mu}A^{(6)}_\mu A^{(5)\nu}A^{(5)}_\nu=
-\gamma G$$ While the gravitation field action is $$\label{f8}
S=\int d^4x(-\gamma G+\L_{mat.})=\int d^4x(-\gamma G+\L_{mat.})-\gamma\int d^4xdiv\Gamma$$
$$=\int d^4x(-\gamma R+\L_{mat.})$$ Where $R$ is the curvature. Here the divergent members are introduced, which complete Lagrangian $G$ to curvature, the surface member is assumed to be zero. As a corollary Einstein equation is obtained $$\label{uren}
R_{\mu\nu}=\frac1\gamma T_{\mu\nu}$$
However, the question of legitimacy of (\[f4\]) and (\[f30\]) is still left untouched, that is why further research of the free octonionic Lagrangian L is needed. Because only the long-range part of it, without EMF, is important for the research, so eliminate $A_\mu^k,k=01,2,3$. Let $q^5A^5_\mu=q^6A^6_\mu=q_DD_\mu$ then $$\label{uel1}
F_{\mu\nu}^D=D_{\nu,\mu}-D_{\mu,\nu}$$
And Euler-Lagrange equation performs $$\label{uel10}
F^{\mu\nu(D)}_{,\mu}+
(m^2_D+q^{47}q^{56}A^{\mu(4)}A^{\nu(7)})D^\nu=0$$
Assume mass $m_D$ to be that abolishing the expression in brackets (\[uel10\]) $$\label{uel1d}
m^2_D=-q^{47}q^{56}A^{\mu(4)}A^{\nu(7)})=\frac32q^2_D\gamma$$
Once expression in brackets equals zero find the solution for vectorial $D$-boson $$\label{uel2}
F^{\mu\nu(5)}_{,\mu}=F^{\mu\nu(6)}_{,\mu}=0,\qquad A^5_\mu=A^6_\mu=(g(t),\vec0)$$ where $g(t)$ could be any time-dependent function. So $D$-bosons are proved to be massless vectorial charged particles.
Consider equation for $C$- and $E$-bosons (denote $\dot a=\frac{da}{dt}$) $$\label{uel3}
F^{\mu\nu(C)}_{,\mu}+m_C^2C^\nu+q^{47}q^{56}\frac{\dot a^2}{a^2}E^\nu=0$$
$$F^{\mu\nu(E)}_{,\mu}+m_E^2E^\nu+q^{47}q^{56}\frac{\dot a^2}{a^2}C^\nu=0$$
Assuming $C$ and $E$ - boson masses so that $$\label{uel40}
\matrix{\displaystyle m_C^2C^\nu+ q^{47}q^{56}\frac{\dot a^2}{a^2}E^\nu\approx m_C^2C^\nu\cr\cr\displaystyle
m_E^2E^\nu+q^{47}q^{56}\frac{\dot a^2}{a^2}C^\nu\approx m_E^2E^\nu}$$ come to the expression for free vectorial massive particles. For solutions $$\label{uel5}
C_\mu=e^{ikx}c_\mu,\qquad E_\mu=e^{-ikx}e_\mu$$ with constant polarization vectors $c_\mu$ and $e_\mu$ obtain $$\label{uel50}
C_\mu E_\mu=c_\mu e_\mu=Constant$$ come to condition (\[f4\]).
From octonionics Lagrangian it is inferred $E_\mu$-bosons decompose to particles and anti-particles almost immediately (assumed their mass is large), however, the proposed model deals with the pair $C\bar E$ - a particle and an anti-particle in the bound state. Moreover the given state defines vacuum, this is why it is assumed the decomposition is denied in ordinary terms. Presumably, special terms are early cosmology and closely to black holes particles birthes.
For the initially posed question was the Friedmann type solutions existence on the octonionics Lagrangian, the problem could be deemed as solved.
Schwarzchild solution
=====================
Consider the space where there is a massive spherically symmetrical object. Let this object be the source of octonionic field. On great distances from the object the electroweak interaction could be eliminated therefore the object could be merely the source of the senior octonionic fields. Due to symmetry of the problem it is acceptable if on the great distances there are octonionic fields, caused by the object, then they are produced by vector-potential $A_\mu^k=A_\mu^k(r),k=4,5,6,7$.
Let, e.g. the electron, moving in the space. The space where the electron is moving is by definition the Minkowski space. In spherically symmetrical coordinates the metrics in the space forms $$\label{Mprs}
ds^2=dt^2-dr^2-r^2(\sin^2\theta d\varphi^2+d\theta^2)$$
Restrict the consideration by left spinors. Rewrite its the equation of moving in massive source octonionic field, assumed it does not interact with the fields $A^{4,7}_\mu$ $$\label{ivsok31}
(i\gamma^0\partial_0+i\gamma^r(\partial_r-\vec\Sigma\cdot\hat{\vec L}-\frac{3i}4q^6A^6_r-\frac34q^5A^5_r)+m)\psi=0$$ here denote $$\gamma^r=\gamma^1\sin\theta\cos\varphi+
\gamma^2\sin\theta\sin\varphi+\gamma^3\cos\theta,\quad
\vec\Sigma=\left(\matrix{\vec\sigma&0\cr0&\vec\sigma}\right)$$ where $\vec L=\vec r\times\vec p$ is the angular moment operator. [@Wheeler].
In solutions (\[ivsok31\]) the spherically-symmetrical part is selected, that responses to the proper value of angular momentum, hence comparison of (\[ivsok31\]) with Ivanenko-Sokolov equation (\[ivsok3\]) gives the idea that formally its solution could be found in the curvilinear space-time $$\label{Shwm00}
ds^2=H_0^2(r)dt^2-H^2_1(r)dr^2-r^2(\sin^2\theta d\varphi^2+d\theta^2)$$ (denoted $\bar g=-H_0^2H_1^2$) $$\label{ivsok32}
-\frac34q^5A^5_r=\frac12\partial_a\left(\ln\frac{\sqrt{-\bar g}}{H^a}\right))=\frac{H_{0,r}}{2H_0},\qquad A^5=(0,A^5_r,0,0)$$
For a weak field it is well-known $$\label{slp}
g_{00}=1-r_g/r=f^2$$ So assume $H_0$ to be known and find $H_1(r)$. Approximately, with great $r$, assume $H_1=1+C/r^n,n>1$. Find gravitational field Lagrangian $G_{gr}$ in metrics (\[Shwm00\]) $$\label{gshv10}
G_{gr}=g^{\mu\nu}(\Gamma^\lambda_{\mu\nu}\Gamma^\kappa_{\lambda\kappa}- \Gamma^\lambda_{\mu\kappa}\Gamma^\kappa_{\nu\lambda})
=\frac2{r^2H_1^2}+\frac{4H_{0,r}}{rH_1^2H_0}-\frac{H_{0,r}H_{1,r}}{H_1^3H_0}$$
On the other hand according to the octonionic theory the gravitational field Lagrangian is defined as $$\label{gshv3}
-\gamma \L_{gr}=G_{pl}+C_0q^{(5)2}A_\mu^5A^{(5)\mu}= \frac2{r^2}+C_0\frac{4H_{0,r}^2}{9H_0^2}$$ (Here the member with zero-curvature, $G_{pl}$, is added, which responses to the zero curvature of the value $G$ in curvilinear coordinates and the constant $C_0$ responds to the convolution $A^{(4)\mu}A^7_\mu$.)
To obtain the same infinitesimal order of (\[gshv10\]) and (\[gshv3\]) when $r$ it is required (accurate within next infinitesimal orders in decomposition by $r$ powers) $$\label{gshv4}
H_0H_1=1,\qquad g_{11}=(1-r_g/r)^{-1}$$
Finally the spherically-symmetrical Schwarzchild metrics is obtained $$\label{Shwm0}
ds^2=(1-\frac{r_g}r)dt^2-\frac{dr^2}{1-\frac{r_g}r}-
r^2(\sin^2\theta d\varphi^2+d\theta^2)$$ where $r_g=2M/r$ is gravitational radius. Then from the formal, built up in the paper, point of view, it is necessary to consider complementary vectorial field $D_\mu$ so that $$\label{DDm}
q_DD_\mu=q_6A_\mu^6+iq_5A_\mu^5,\qquad q_5A_\mu^5=q_6A_\mu^6=f_{,r}/f
\approx \frac{r_g}{2r^2}$$
However, the question of admissibility of (\[DDm\]) arose therefore research the free Lagrangian $\L_f$. Once only its long-range piece without $A_\mu^k,k=0,1,2,3$ is under interest, then eliminate $q^5A^5_\mu=q^6A^6_\mu=q_DD_\mu$ and it takes $$\label{uel12}
F_{\mu\nu}^D=D_{\nu,\mu}-D_{\mu,\nu}$$
The Euler-Lagrange equations $$\label{uel13}
F^{\mu\nu(D)}_{,\mu}+
(m^2_D+q^{47}q^{56}A^{\mu(4)}A^{\nu(7)})D^\nu=0$$
Let $m_D$ mass is that abolishing the expression in brackets in (\[uel13\]) $$\label{uel1d5}
m^2_D+q^{47}q^{56}A^{\mu(4)}A^{\nu(7)}=0$$
Once the expression in brackets equals zero find the solution for vctorial $D$-boson
$$\label{uel20}
F^{\mu\nu(D)}_{,\mu}=0,\qquad A^5_\mu=A^6_\mu=(0,g(r),0,0)$$
where $g(r)$ is any coordinates-dependent function. So $D$-bosons are proved to be massless vectorial charged particles.
Consider equations for $C$- and $E$-bosons$$\label{uel30}
F^{\mu\nu(C)}_{,\mu}+m_C^2C^\nu+q^{47}q^{56}\frac{r_g^2}{4r^4}E^\nu=0$$
$$F^{\mu\nu(E)}_{,\mu}+m_E^2E^\nu+q^{47}q^{56}\frac{r_g^2}{4r^4}C^\nu=0$$
Assume $C$- and $E$-bosons’ masses are so that $$\label{uel4}
m_C^2C^\nu+ q^{47}q^{56}\frac{r_g^2}{4r^4}E^\nu\approx m_C^2C^\nu,\qquad m_E^2E^\nu+q^{47}q^{56}\frac{r_g^2}{4r^4}C^\nu\approx m_E^2E^\nu$$ then come to the free vectorial massive particles equation. For solutions $$\label{uel501}
C_\mu=e^{ikx}c_\mu,\qquad E^\mu=e^{-ikx}e^\mu$$ With constant polarization vectors $c_\mu$ and $e_\mu$ it takes $$\label{uel57}
C_\mu E^\mu=c_\mu e^\mu=Constant$$ and condition (\[f4\]) is obtained.
As it was earlier assumed, the pairs $C\bar E$ in ordinal conditions comprise a stable state.
For the initially posed question was the Friedmann type solutions existence on the octonionics Lagrangian, the problem could be deemed as solved.
Formally, in the paper it is shown the existence of the approximate octonionic field, which could be appropriate one for physical reality. These considerations are deemed by author as deserving the most scrupulous attention and further analysis of octonionic gravity approach, proposed above.
Jet phenomenology
=================
In astronomy the beautiful phenomenon is well-known, when from the center of a galaxy, at right angle to its disk, the radiant flux is observed. It appears such behavior is consistent with the proposed theoretical gravity model. Indeed, consider a massive galaxy with a dense nucleus in the center of it. (According to octonionic theory there are no black holes in nature, because in strong gravitational fields the D-bosons pairs are born and their gravitational description is unacceptable. Hence in case of a strong field the consideration could be restricted by flat Minkowski space, although assuming major vectors-potentials $A^c_\mu,c=4,5,6,7$ to be the main fields. From the representation proposed here $A^4$ and $A^7$ complete bound states, vacuum, and are therefore stable. So, restrict the consideration and research only $D$-bosons.)
For a particle falling ti the star interaction with $D$-boson is represented a gravitational field with Schwarzchild metrics (\[Shwm0\]). Formally, denote particle mass by $m$, impulse by $p$ and energy by $p^0=E$ then there is the invariant $$\label{inve1}
g_{\mu\nu}p^{\mu\nu}=(1-\frac{r_g}r)E^2-\frac{p^2}{1-\frac{r_g}r}=m^2$$ which means $$\label{inve10}
E^2=\frac{p^2}{(1-\frac{r_g}r)^2}+\frac{m^2}{1-\frac{r_g}r}$$
It is evident from latter when reaching the gravitational radius its energy tends to infinity. Let $D$-boson mass be equal to Planck’s mass. Then with $r\approx r_g+r_{pl}$ $D$-boson birth could happen in case the mass of the star piece enclosed into gravitational radius is estimated by $10^{12}$ sun mass.
Consider the massive vector field $A_\mu$ $$\label{v1}
F^{\mu\nu}_{;\nu}-m^2A^\mu=0,\qquad F_{\mu\nu}=A_{\nu,\mu}-A_{\mu,\nu}.$$
Rewrite using vector-potential $A_\mu$ $$\label{v2}
\frac1{\sqrt{-g}}((A_{\sigma,\rho}-A_{\rho,\sigma})g^{\rho\mu} g^{\sigma\nu}\sqrt{-g}),\nu-m^2A^\mu=0.$$
Introduce the scalar functions of radial coordinates and time $f^{lm}(r,t)$ and $h^{lm}(r,t)$, $k^{lm}(r,t), a^{lm}(r,t)$ [@WheelerRegger] $$\label{sph0}
A^0=f^{lm}(r,t)Y^{lm}(\theta,\phi),\qquad
A^1=h^{lm}(r,t)Y^{lm}(\theta,\phi)$$ $$\label{sph2}
A^2=k^{lm}(r,t)Y^{lm}(\theta,\phi)+a^{lm}(r,t)Y^{lm}(\theta,\phi)/\sin\theta$$ $$\label{sph3}
A^3=k^{lm}(r,t)Y^{lm}(\theta,\phi)-a^{lm}(r,t)Y^{lm}(\theta,\phi)\sin\theta$$
And for the radial component of $A_r$ it takes $$\label{urar1}
\lambda(h^{lm}-k^{lm}_{,r})+((f^{lm}_{,r}-h^{lm}_t)r^2)_{,t}f^{-1}
-m^2r^2h^{lm}=0$$ where $\lambda=l(l+1)$.
Assume the solution to be stable, then AЕ In this case solution exists with $A^0=A^2=A^3=0,A^1=A^1(t,\theta,\phi)$ and $$\label{urar2}
\omega^2(r)f^{-1}(r)+\frac\lambda{r^2}-m^2=0$$
Hence, there is a solution $D=(0,C\cdot Y^{lm},0,0)$. That is centrally-symmetrical solution with a constant radius $r=r_g$. Which means the $D^{\pm}$-bosons pair birth happens. From the angular momentum conservation law they are born symmetrically relatively to the center of the star. Because of their nature, on the boundary of Schwarzchild radius the bosons form the boson condensate. Let the galaxy material be distributed like a rotating disk. If in the field of that system appears considerably large and slow galaxy and it is caught then the great thread of matter tends to the center of supermassive galaxy. As result is the increase in gravitational radius. The conversion of boson condensate from the level $\omega(r=r_{g_1})$ to the level $\omega(r=r_{g_2})$ happens. This conversion is accompanied with radiation because gravitational $D$-particles are charged. The angular distribution of radiation is estimated by the formula [@Davidov] $$\label{v9}
I=I_0\sum_{k=1}^la_k\cos^k\theta$$
Once here $l=mpv\to\infty$ and the galaxy has a form of a disk then only the radiation orthogonally to the disk surface is possible. Hence, appears a radiation precisely at the right angle to the galaxy.
Along with discussed another scenario could take place. Let the supermassive galaxy be star-like. Vectorial $D$-bosons are born around the massive nucleus of the star, completing the bound state, which is close by its nature to the hydrogen atom. As a result the mass accumulation, as a boson cover around baryon nucleus, takes place. Due to the increase in the gravitational radius after the increase in mass the completion of more and more remote from center energy levels of the star happens. As a corollary of cooling process, higher levels of matter fall to the center. At the expense of increase in gravitational radius the radiation in all directions is emited, because no disk symmetry is assumed. As a result of it explosion and, perhaps, the birth of supernova happens.
Conclusion
==========
In the physical review of the paper some phenomenological solutions of octonionic Lagrangian are proposed because the problem of physical interpretation of the octonionic Lagrangian deserves attention. The scope of the paper does not allow to research those solutions. Moreover, there are some more interesting octonionic Lagrangian physical corollaries. Particularly, it is amazing that vacuum state $u_0=\Psi_0$ and bound state $A^4_\mu A^{(7)\mu}$ are close by their structures and meanings. The author hopes the research of that fact would be useful in Higgs boson nature problem.
[99]{} V. Yu. Dorofeev. arxiv 0908.3247v1 \[math-ph\] (2009) L. B. Okun. [*Leptons and qurks.*]{} M:Nauka, 1990. L. P. Eisenhart. *Riemannian geometry.* 1926. A. E. Ahieser, V. B. Berestetski. *Quantum electrodinamics.* M:Nauka, 1981. R. Utiyama. Phys. Rev., [**101**]{}, 1597, 1956. V. Fock, D. Ivanenko. C. R., Paris 188, 1470 (1929). A. Sokolov, D. Ivanenko. *Quantum theory field.* 1952. D. R. Brill, J. A. Wheeler. Rev. Mod. Phys. D. [**29**]{}, (1957), 465-479; A. O. Barut and I. H. Duru, Phys. Rev. [**D 36**]{}, (1987), 3705-3711; F. Finelly, A. Gruppuso and G. Venturi, Class. Quantum Grav. [**16**]{}, (1999), 3923-3937. T. Regge and J. A. Wheeler, Phys. Rev. 108 (1957) 1063, R. A. Konoplya Phys. Rev D, 73 044002 (2006), R. A. Konoplya, A. Zidenko, C. Molina gr-qc/0602047 A S. Davidov. *Quantum mehanics.* M., 1963.
[^1]: Dep. of Math., SPb SUEF, Sadovaya 21, 191023, St.Petersburg, Russia, e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study dually epi-translation invariant valuations on cones of convex functions containing the space of finite-valued convex functions. The existence of a homogeneous decomposition is used to associate a distribution to every valuation of this type similar to the Goodey-Weil embedding for translation invariant valuations on convex bodies. The relation between the valuation and its associated distribution is used to establish a notion of support for valuations. As an application, we show that there are no $\mathrm{SL}(n)$ or translation invariant valuations except constant valuations in this class and we discuss which valuations on finite-valued convex functions can be extended to larger cones. In addition, we examine some topological properties of spaces of valuations with compact support.'
author:
- Jonas Knoerr
bibliography:
- 'literature.bib'
title: 'The support of dually epi-translation invariant valuations on convex functions'
---
Introduction
============
General background
------------------
Let $V$ be a finite dimensional real vector space and let $\mathcal{K}(V)$ denote the space of convex, compact subsets of $V$. Then a valuation on $\mathcal{K}(V)$ with values in an abelian semigroup $(F,+)$ is a map $\mu:\mathcal{K}(V)\rightarrow F$ that satisfies $$\begin{aligned}
\mu(K)+\mu(L)=\mu(K\cup L)+\mu(K\cap L)
\end{aligned}$$ for all $K,L\in\mathcal{K}(V)$ such that $K\cup L\in \mathcal{K}(V)$. Valuations on convex bodies are a classical part of convex geometry and many structural results have been established (for an overview see for example [@Schneider:convex_bodies_Brunn-Minkowski] Chapter 6). For a Hausdorff real topological vector space $F$ let ${\mathrm{Val}}(V,F)$ denote the space of all valuations $\mu:\mathcal{K}(V)\rightarrow F$ that are translation invariant and continuous with respect to the Hausdorff metric. A valuation $\mu\in{\mathrm{Val}}(V,F)$ is called *$k$-homogeneous* or homogeneous of degree $k\in{\mathbb{R}}$, if $\mu(tK)=t^k\mu(K)$ for all $K\in\mathcal{K}(V)$, $t\ge 0$. Let ${\mathrm{Val}}_k(V,F)$ denote the subspace of $k$-homogeneous valuations. Then we have the following homogeneous decomposition:
\[theorem\_McMullen\_decomposition\_for\_Val\] $$\begin{aligned}
{\mathrm{Val}}(V,F)=\bigoplus\limits_{i=0}^{\dim V}{\mathrm{Val}}_i(V,F).
\end{aligned}$$
In other words, $\mu(tK)$ is a polynomial in $t\ge 0$ for $\mu\in{\mathrm{Val}}(V,F)$ and $K\in\mathcal{K}(V)$ and the degree of this polynomial is bounded by the dimension of $V$. Polarizing this polynomial for a $k$-homogeneous valuation, one obtains the notion of mixed valuations. For example, by polarizing the volume we obtain functionals known as mixed volumes, which play an important role in convex geometry.\
In this paper we are interested in a functional version of this homogeneous decomposition for a special class of valuations on convex functions and additional structures derived from this decomposition.\
A valuation on some class of real-valued functions $X$ with values in some abelian semi-group $(G,+)$, is a map $\mu:X\rightarrow G$ satisfying $$\begin{aligned}
\mu(f)+\mu(h)=\mu(\max(f,h))+\mu(\min(f,h))
\end{aligned}$$ for all $f,h\in X$ such that the pointwise maximum $\max(f,h)$ and minimum $\min(f,h)$ belong to $X$.\
To see the link with the notion of valuations on convex bodies, assume that the functions in $X$ are defined on some set $A$. Consider the *epi-graph* ${\mathrm{epi}}(f):=\{(a,t)\in A\times{\mathbb{R}}: f(a)\le t\}$ for $f\in X$. Then $$\begin{aligned}
&{\mathrm{epi}}(\max(f,h))={\mathrm{epi}}(f)\cap {\mathrm{epi}}(h), && {\mathrm{epi}}(\min(f,h))={\mathrm{epi}}(f)\cup {\mathrm{epi}}(h),
\end{aligned}$$ for all $f,h\in X$. Thus a valuation on a space of functions can also be considered as a valuation on their epi-graphs.\
Although the study of valuations on functions was started only recently, a number of classification results have been established for different spaces of functions, for example Sobolev-spaces [@Ludwig:Fisher_information_valuations; @Ludwig:valuations_sobolev; @Ma:valuations_sobolev], $\mathrm{L}^p$-spaces [@Ludwig:covariance_matrices_valuations; @Ober:Minkowski_valuations_on_Lq_spaces; @Tsang:valuations_Lp; @Tsang:minkowski_val_Lp], quasi-concave functions [@Bobkov_Colesanti:quermassintegrals_quasi-concave_functions; @Colesanti_Lombardi:valuations_quasi-concave; @Colesanti_Lombardi_Parapatits:translation_invariant_valuations_quasi-concave], Orlicz-spaces [@Kone:valuations_orlicz], and functions of bounded variation [@Wang:semivaluations_bounded_variations]. Due to their intimate relation to convex bodies, valuations on convex functions have been the object of intense research: Monotone or $\mathrm{SL}(n)$-invariant valuations were classified in [@Cavallina_Colesanti:monotone_valuations_convex_functions; @Colesanti_Ludwig_Mussnig:minkowski; @Colesanti_Ludwig_Mussnig:Valuations_convex_functions; @Mussnig:SLn_invariant_super_coercive; @Mussnig:volume_polar_volume_euler] and a connection between certain valuations on convex functions and translation invariant valuations on convex bodies was studied by Alesker in [@Alesker:valuations_convex_functions_Monge-Ampere] with the help of Monge-Ampère operators.\
Let us introduce the general framework for this paper. We will be interested in valuations on convex functions on a finite dimensional real vector space $V$. The spaces of functions under consideration are subspaces of $$\begin{aligned}
{\mathrm{Conv}}(V):=\{f:V\rightarrow(-\infty,\infty] \ : \ f \text{ convex, lower semi-continuous and proper}\},
\end{aligned}$$ where a function $f:V\rightarrow(-\infty,\infty]$ is called *proper*, if it is not identical to $+\infty$. This space carries a natural topology, the topology of *epi-convergence*, which we will recall in Section \[section\_topology\_convex\_functions\]. Especially valuations on the subspaces of finite or super-coercive convex functions $$\begin{aligned}
{\mathrm{Conv}}(V,{\mathbb{R}})&:=\{f\in{\mathrm{Conv}}(V) \ : \ f<+\infty\},\\
{\mathrm{Conv}}_{s.c.}(V)&:=\{f\in{\mathrm{Conv}}(V) \ :\ \lim\limits_{x\rightarrow\infty}\frac{f(x)}{|x|}=\infty\}
\end{aligned}$$ have been of interest, see for example [@Colesanti_Ludwig_Mussnig:Hessian_valuations; @Colesanti_Ludwig_Mussnig:Valuations_convex_functions; @Colesanti_Ludwig_Mussnig:homogeneous_decomposition]. From a valuation theoretic point of view, these two spaces are identical: It was shown in [@Colesanti_Ludwig_Mussnig:Hessian_valuations] that the Legendre transform establishes a continuous involution between the two spaces, that exchanges minimum and maximum. Thus any continuous valuation on one space induces a unique continuous valuation on the other.\
Following the terminology in [@Colesanti_Ludwig_Mussnig:homogeneous_decomposition], let us call a valuation $\mu:{\mathrm{Conv}}_{s.c.}(V)\rightarrow G$ *translation invariant*, if $$\begin{aligned}
\mu(f\circ \tau)=\mu(f) \quad\forall f\in {\mathrm{Conv}}_{s.c.}(V)
\end{aligned}$$ for all translations $\tau:V\rightarrow V$. We will call it *vertically translation invariant*, if $$\begin{aligned}
\mu(f+c)=\mu(f) \quad \forall f\in {\mathrm{Conv}}_{s.c.}(V), c\in{\mathbb{R}},
\end{aligned}$$ and *epi-translation invariant*, if $\mu$ is both translation and vertically translation invariant. These invariance properties are equivalent to the invariance of the valuation with respect to translations of the epigraph of its argument in $V\times{\mathbb{R}}$. In the dual setting (i.e. considering the composition of an epi-translation invariant valuation with the Legendre transform), this leads to the following notion:\
A valuation $\mu:{\mathrm{Conv}}(V,{\mathbb{R}})\rightarrow G$ is called *dually epi-translation invariant*, if $$\begin{aligned}
\mu(f+\lambda+c)=\mu(f) \quad \forall f\in{\mathrm{Conv}}(V,{\mathbb{R}}), \lambda\in V^*, c\in{\mathbb{R}}.
\end{aligned}$$ Real-valued valuations on ${\mathrm{Conv}}(V,{\mathbb{R}})$ invariant under the addition of linear functionals were considered by Alesker in [@Alesker:valuations_convex_functions_Monge-Ampere]. The examples constructed by him are extensions of functionals of the form $$\begin{aligned}
f\mapsto \int_{{\mathbb{R}}^n} B(x)\det( H_f(x)[k], A_{1}(x),...,A_{n-k}(x)) dx\quad \text{for }f\in{\mathrm{Conv}}({\mathbb{R}}^n,{\mathbb{R}})\cap C^2({\mathbb{R}}^n).
\end{aligned}$$ Here $H_f$ denotes the usual Hessian of a twice differentiable function $f$, $B\in C_c({\mathbb{R}}^n)$, $A_1,...,A_k\in C_c({\mathbb{R}}^n,{\mathbb{R}}^{n\times n})$ are compactly supported functions with values in the space of symmetric $n\times n$-matrices, and $\det$ denotes the mixed determinant of $n$ symmetric matrices ($H_f(x)$ is taken with multiplicity $k$ in the example above). From this representation it is easy to deduce that the valuations Alesker used in his proofs are actually dually epi-translation invariant, so his results also apply to the type of valuations considered in this paper. Similar functionals were also examined in [@Colesanti_Ludwig_Mussnig:Hessian_valuations] by Colesanti, Ludwig and Mussnig.\
Let us introduce another example: Let $\nu$ be a compactly supported signed Radon measure on ${\mathbb{R}}^n$ with the property $\int_{V}d\mu=\int_{V}\lambda d\mu=0$ for all $\lambda\in V^*$. Then $$\begin{aligned}
\mu(f):=\int_{V}fd\mu \quad \forall f\in{\mathrm{Conv}}(V,{\mathbb{R}})
\end{aligned}$$ defines a dually epi-translation invariant valuation. This class includes valuations such as $$\begin{aligned}
\mu_1(f)&:=f(x)+f(-x)-2f(0) &&\text{for } f\in {\mathrm{Conv}}(V,{\mathbb{R}}),\\
\mu_2(f)&:=\frac{1}{{\mathrm{vol}}_{n-1}(S^{n-1})}\int_{S^{n-1}} f d\sigma -f(0) && \text{for } f\in{\mathrm{Conv}}({\mathbb{R}}^n,{\mathbb{R}}),
\end{aligned}$$ where $x\in V\setminus \{0\}$ is an arbitrary point and $\sigma$ denotes the usual measure on the unit sphere.\
Results of the present paper
----------------------------
We will consider valuations with values in some Hausdorff topological vector space $F$ on certain cones $C\subset{\mathrm{Conv}}(V)$ containing the space ${\mathrm{Conv}}(V,{\mathbb{R}})$ of finite-valued convex functions. Let ${\mathrm{VConv}}(C;V,F)$ denote the space of continuous, dually epi-translation invariant valuations on $C$ with values in $F$ (see Section \[section\_valuations\_convex\_functions\] for the precise definition). A valuation $\mu\in{\mathrm{VConv}}(C;V,F)$ will be called *$k$-homogeneous*, if $\mu(tf)=t^k\mu(f)$ for all $f\in C$ and $t\ge 0$ and the subspace of $k$-homogeneous valuations will be denoted by ${\mathrm{VConv}}_k(C;V,F)$. Then we have the following decomposition:
\[maintheorem\_homogeneous\_decomposition\] If $C\subset{\mathrm{Conv}}(V)$ is a cone containing ${\mathrm{Conv}}(V,{\mathbb{R}})$, then $$\begin{aligned}
{\mathrm{VConv}}(C;V,F)=\bigoplus\limits_{i=0}^{\dim V}{\mathrm{VConv}}_i(C;V,F).
\end{aligned}$$
In [@Colesanti_Ludwig_Mussnig:homogeneous_decomposition], this was proved by Colesanti, Ludwig and Mussnig for real-valued valuations on ${\mathrm{Conv}}(V,{\mathbb{R}})$. The more general statement can be deduced from their proof with only minor modifications. Instead of repeating their arguments, we will present a slightly different proof in Section \[section\_homogeneous\_decomposition\] using an embedding of ${\mathrm{VConv}}(C;V,F)$ into ${\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ using support functionals of convex bodies.\
Similar to the classical McMullen decomposition (and to [@Colesanti_Ludwig_Mussnig:homogeneous_decomposition] Theorem 23), this allows us to define the polarization of any homogeneous valuation $\mu\in {\mathrm{VConv}}_k(C;V,F)$. We obtain a symmetric functional $\bar{\mu}:C^k\rightarrow F$ which satisfies $\bar{\mu}(f,...,f)=\mu(f)$ for all $f\in C$ and which is a continuous, $1$-homogeneous and *additive* valuation in each coordinate, i.e. it satisfies $$\begin{aligned}
\bar{\mu}(f+h,f_2,\dots,f_k)=\bar{\mu}(f,f_2,\dots,f_k)+\bar{\mu}(h,f_2,\dots,f_k)
\end{aligned}$$ for all $f,h, f_2,...,f_k\in C$. Goodey and Weil used the polarization of a valuation in ${\mathrm{Val}}_k({\mathbb{R}}^n)$ to define a distribution on the $k$-fold product of the unit sphere in ${\mathbb{R}}^n$ (see [@Goodey_Weil:Distributions_and_valuations]). Alesker proved in [@Alesker:McMullenconjecture] that the support of this distribution is contained in the diagonal of this Cartesian product of spheres, which plays a crucial role in the proof of his Irreducibility Theorem (see [@Alesker:McMullenconjecture; @Alesker:IrreducibilityThm]).\
Following the ideas of Goodey-Weil and Alesker, we establish the following version of the Goodey-Weil embedding for valuations on convex functions in Section \[section\_Goodey-Weil-embedding\] and \[section\_support\_GW\_diagonal\]:
\[maintheorem\_GW\] Let $F$ be a locally convex vector space admitting a continuous norm, $C\subset{\mathrm{Conv}}(V)$ a cone containing ${\mathrm{Conv}}(V,{\mathbb{R}})$. For every $\mu\in{\mathrm{VConv}}_k(C;V,F)$ there exists a unique distribution $GW(\mu)\in\mathcal{D}'(V^k,\bar{F})$ with compact support, which satisfies the following property: If $f_1,...,f_k\in {\mathrm{Conv}}(V,{\mathbb{R}})\cap C^\infty(V)$ then $$\begin{aligned}
GW(\mu)(f_1\otimes...\otimes f_k)=\bar{\mu}(f_1,...,f_k),
\end{aligned}$$ where $\bar{\mu}$ is the polarization of $\mu$.\
Moreover, the support of $GW(\mu)$ is contained in the diagonal in $V^k$.
Here $\bar{F}$ denotes the completion of $F$ and $\mathcal{D}'(V^k,\bar{F})$ denotes the space of all distributions on $V^k$ with values in $\bar{F}$, i.e. the space of all continuous functionals $\phi:C^\infty_c(V^k)\rightarrow \bar{F}$, where $C^\infty_c(V^k)$ is equipped with the usual topology on test functions. We will call the distribution $GW(\mu)$ the Goodey-Weil distribution of $\mu$. We thus obtain an injective map $GW:{\mathrm{VConv}}_k(C;V,F)\rightarrow \mathcal{D}'(V^k,\bar{F})$, which will be called the Goodey-Weil embedding.\
More generally, we also define a version of the Goodey-Weil embedding for valuations with values in an arbitrary locally convex vector space. Although the Goodey-Weil distribution is still uniquely determined by its underlying valuation in this case, the support is, in general, not compact. To illustrate this fact, we examine the Hessian measures (see [@Colesanti_Ludwig_Mussnig:Hessian_valuations]), which can be considered as continuous, dually epi-translation invariant valuations with values in the space of signed Radon measures on $V$ (equipped with the vague topology). This example was also our main motivation to examine valuations with values in arbitrary locally convex vector spaces.\
In Section \[section\_support\_of\_a\_valuation\] we use the support of the Goodey-Weil distribution to obtain a notion of support for the corresponding valuation and we discuss how the support can be defined intrinsically. As an application, we show that there are no non-trivial real-valued dually epi-translation invariant valuations (except constant valuations) that are also invariant under the operation of the special linear group or translations. We also characterize the image of the embedding used in the proof of Theorem \[maintheorem\_homogeneous\_decomposition\].\
\
It is natural to consider the subspace ${\mathrm{VConv}}_A(C;V,F)$ of valuations with support contained in a fixed subset $A\subset V$. In Section \[section\_subspaces\_compact\_support\] we discuss some topological properties of ${\mathrm{VConv}}_A(V,F)$ in the case where $A$ is compact and we show the following topological result:
\[mainproposition\_approximation\_using\_sequences\] Let $F$ be a locally convex vector space admitting a continuous norm. If a sequence $(\mu_j)_j$ converges to $\mu$ in ${\mathrm{VConv}}(V,F)$, then there exists a compact set $A\subset V$ such that the support of $\mu$ and $\mu_j$ is contained in $A$ for all $j\in\mathbb{N}$. In particular, $(\mu_j)_j$ converges to $\mu$ in ${\mathrm{VConv}}_A(V,F)$.
In Section \[section\_description\_valuations\_on\_cones\_with\_support\] we relate the valuations in ${\mathrm{VConv}}(C;V,F)$ to the common domains of $f\in C$, i.e. the sets ${\mathrm{dom }}f:=\{x\in V \ : \ f(x)<\infty\}$. This also gives a partial answer to the question, which valuations on finite-valued convex functions can be extended to larger cones in ${\mathrm{Conv}}(V)$. Together with Corollary \[corollary\_existence\_non\_trivial\_valuations\], this refines [@Colesanti_Ludwig_Mussnig:homogeneous_decomposition] Theorem 30, where valuations on the maximal cone $C={\mathrm{Conv}}(V)$ were considered.\
These results apply in particular to cones of the form $C_U:=\{f\in{\mathrm{Conv}}(V) \ : \ f|_U<\infty\}$ for some open and convex subset $U\subset V$. In Section \[section\_valuations\_on\_open\_sets\], we considering valuations on the space ${\mathrm{Conv}}(U,{\mathbb{R}}):=\{f:U\rightarrow\mathbb{R} \ : \ f\text{ convex }\}$ with values in a locally convex vector space $F$ that are dually epi-translation invariant and continuous with respect to the topology of uniform convergence on compact subsets in $U$. Denoting the space of these functionals by ${\mathrm{VConv}}(U,F)$, we can identify this space of valuations with valuations on the cone $C_U$:
\[maintheorem\_isomorphism\_cone\_open\_subset\] If $U\subset V$ is an open convex subset and $F$ is a locally convex vector space admitting a continuous norm, then the map $$\begin{aligned}
{\mathrm{res}}^*:{\mathrm{VConv}}(U,F)&\rightarrow {\mathrm{VConv}}(C_U;V,F)\\
\mu&\mapsto[f\mapsto \mu(f|_U)]
\end{aligned}$$ is a topological isomorphism.
Here both spaces are equipped with the topology of uniform convergence on compact subsets in ${\mathrm{Conv}}(U,{\mathbb{R}})$ and $C_U$, respectively.
#### Acknowledgments
Part of this paper was written during a stay at the Università degli Studi di Firenze and I want to thank the university and especially Andrea Colesanti for the hospitality. I also want to thank Andreas Bernig for many useful discussions, suggestions and encouragement during this project, as well as his comments and remarks on the first draft of this paper.
Convex functions {#section_convex_functions}
================
In this section, we collect some facts about the space of convex functions and its topology. For simplicity we will assume that $V$ is a euclidean vector space. Let $B_R:=B_R(0)$ denote the closed ball of radius $R>0$ in $V$.\
Let $U\subset V$ be a convex subset. A function $f:U\rightarrow{\mathbb{R}}\cup\{+\infty\}$ is called convex, if for any $x_0,x_1\in U$ and every $t\in[0,1]$ the following inequality holds: $$\begin{aligned}
f(tx_0+(1-t)x_1)\le tf(x_0)+(1-t)f(x_1).
\end{aligned}$$ Equivalently, $f:U\rightarrow{\mathbb{R}}\cup\{+\infty\}$ is convex if and only if its *epigraph* $\mathrm{epi}(f):=\{(x,t)\in U\times{\mathbb{R}}\ : \ f(x)\le t\}$ is a convex subset in $U\times{\mathbb{R}}$. Note that $f$ is lower semi-continuous if and only if ${\mathrm{epi}}(f)$ is closed in $U\times{\mathbb{R}}$. In this paper we are mostly interested in subsets of $$\begin{aligned}
{\mathrm{Conv}}(V):=\{f:V\rightarrow{\mathbb{R}}\cup\{+\infty\} \ : \ f\text{ convex, lower semi-continuous}, f\not\equiv +\infty\}.
\end{aligned}$$ For any $f\in{\mathrm{Conv}}(V)$, we define the domain of $f$ $$\begin{aligned}
\mathrm{dom}(f):=\{x\in V \ : \ f(x)<+\infty\}.
\end{aligned}$$ By definition, $\mathrm{dom}(f)$ is a non-empty convex subset of $V$. $f$ is always continuous on the interior of $\mathrm{dom}(f)$. In particular, the space of finite-valued convex functions $$\begin{aligned}
{\mathrm{Conv}}(V,{\mathbb{R}}):=\{f:V\rightarrow{\mathbb{R}}\ : \ f \text{ convex}\}\subset{\mathrm{Conv}}(V)
\end{aligned}$$ contains only continuous functions. Note that ${\mathrm{Conv}}(V,{\mathbb{R}})$ is closed under the formation of the pointwise maximum, while the maximum of two elements of ${\mathrm{Conv}}(V)$ may be identical to $+\infty$.
Topology on the space of convex functions {#section_topology_convex_functions}
-----------------------------------------
A sequence $(f_j)_j$ in ${\mathrm{Conv}}(V)$ *epi-converges* to $f\in{\mathrm{Conv}}(V)$ if and only if for every $x\in V$ the following conditions hold:
1. For every sequence $(x_j)_j$ in $V$ converging to $x$: $f(x)\le \liminf\limits_{j\rightarrow\infty}f_j(x_j)$.
2. There exists a sequence $(x_j)_j$ converging to $x$ such that $f(x)= \lim\limits_{j\rightarrow\infty}f_j(x_j)$.
It is known that this notion of convergence is induced by a metrizable topology on ${\mathrm{Conv}}(V)$ (see for example [@Rockafellar_Wets:Variational_analysis] Theorem 7.58).\
In the constructions used in this paper, the limit function $f\in{\mathrm{Conv}}(V)$ will usually be finite on some open subset of $V$. In this case epi-convergence, pointwise convergence and locally uniform convergence are compatible in the following sense:
[@Rockafellar_Wets:Variational_analysis Theorem 7.17] \[proposition\_convergence\_finite\_convex\_functions\] For a function $f\in{\mathrm{Conv}}(V)$ such that ${\mathrm{dom }}f$ has non-empty interior and a sequence $(f_j)_j$ in ${\mathrm{Conv}}(V)$ the following are equivalent:
1. $(f_j)_j$ epi-converges to $f$.
2. $(f_j)_j$ converges pointwise to $f$ on a dense subset.
3. $(f_j)_j$ converges uniformly to $f$ on all compact subsets that do not contain a boundary point of ${\mathrm{dom }}f$.
In particular, a sequence $(f_j)_j$ in ${\mathrm{Conv}}(V,{\mathbb{R}})$ epi-converges to $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ if and only if it converges uniformly on compact subsets.
Compactness in ${\mathrm{Conv}}(V,{\mathbb{R}})$
------------------------------------------------
\[proposition\_convex\_functions\_local\_lipschitz\_constants\] Let $U\subset V$ be a convex open subset and $f:U\rightarrow{\mathbb{R}}$ a convex function. If $X\subset U$ is a set with $X+\epsilon B_1\subset U$ and $f$ is bounded on $X+ B_\epsilon$, then $f$ is Lipschitz continuous on $X$ with Lipschitz constant $\frac{2}{\epsilon}\|f|_{X+ B_\epsilon}\|_\infty$.
This is a special case of [@Rockafellar_Wets:Variational_analysis] 9.14.
\[proposition\_compactness\_Conv\] A subset $U\subset {\mathrm{Conv}}(V,{\mathbb{R}})$ is relatively compact if and only if it is bounded on compact subsets.
As the topology on ${\mathrm{Conv}}(V,{\mathbb{R}})$ is metrizable, we only have to show that the closure of such a subset is sequentially compact. Let $(f_k)_k$ be a sequence in ${\mathrm{Conv}}(V,{\mathbb{R}})$ that is bounded on compact subsets of $V$. Then the Lipschitz constants of these functions are also uniformly bounded on $B_j$ for all $j\in\mathbb{N}$ by Proposition \[proposition\_convex\_functions\_local\_lipschitz\_constants\]. In particular, the set $\{f_k|_{B_j} \ : \ k\in\mathbb{N}\}\subset C(B_j)$ is equicontinuous. By the theorem of Arzelà-Ascoli, we can choose a subsequence $f_{j,k}$ that converges uniformly on $B_j$ to some function $f_{j,\infty}\in C(B_j)$ for $k\rightarrow\infty$. Iterating this argument for all $j\in\mathbb{N}$ and taking an appropriate diagonal series, we find a subsequence that converges uniformly on $B_j$ for all $j\in\mathbb{N}$ to some function $f\in C(V)$. It is easy to see that $f$ is convex. Now the claim follows from Proposition \[proposition\_convergence\_finite\_convex\_functions\].
The Legendre transform, subdifferentials and some density results
-----------------------------------------------------------------
The *Legendre transform* or *convex dual* of a function $f\in{\mathrm{Conv}}(V)$ is the function $f^*:V^*\rightarrow (-\infty,\infty]$ given by $$\begin{aligned}
f^*(y)=\sup\limits_{x\in V}\langle y,x\rangle -f(x) \quad \text{for }y\in V^*,
\end{aligned}$$ where $\langle \cdot,\cdot\rangle:V^*\times V\rightarrow{\mathbb{R}}$ denotes the canonical pairing. As a consequence of [@Rockafellar:Convex_Analysis] Theorem 12.2 and Corollary 12.2.1, we have
For $f\in{\mathrm{Conv}}(V)$, $f^*\in{\mathrm{Conv}}(V^*)$ and $f^{**}:=(f^*)^*=f$.
Let $f\in{\mathrm{Conv}}(V)$. An element $y\in V^*$ is called a *subgradient* of $f$ in $x_0\in V$, if $$\begin{aligned}
f(x_0)+\langle y,x-x_0\rangle\le f(x) \quad \text{for all }x\in V.
\end{aligned}$$ The set of all subgradients of $f$ in a point $x_0\in V$ is called the subdifferential of $f$ in $x_0$ and will be denoted by $\partial f(x_0)$.\
We recall the following basic properties of the subdifferential:
\[lemma\_properties\_subgradients\] For $f\in{\mathrm{Conv}}(V)$, $x\in V$, $y\in V^*$ the following are equivalent:
1. $y\in\partial f(x)$,
2. $\langle y,x\rangle =f(x)+f^*(y)$,
3. $y\in \mathrm{argmax}_{x\in V}\langle y,x\rangle -f(x)$.
\[lemma\_compactness\_face\_epigraph\_conjugate-function\] Let $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$. Then $\partial f(x)\in\mathcal{K}(V^*)$ for all $x\in V$ and if $f$ is $L$ Lipschitz continuous on a neighborhood of $x\in V$, then $\partial f(x)\subset B_L(0)$.
This is a special case of [@Clarke:optimization_non_smooth_analysis] 2.1.2.
Also note that $\partial f(x)\ne \emptyset$ for any $x\in V$, if $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$.\
The support function $h_K\in{\mathrm{Conv}}(V,{\mathbb{R}})$ of a convex body $K\in\mathcal{K}(V^*)$ is defined as $$\begin{aligned}
h_K(y):=\sup\limits_{x\in K}\langle y,x\rangle.
\end{aligned}$$ These functions will play an important role in Section \[section\_relation\_to\_val\_convex\_bodies\], where we relate valuations on convex functions to valuations on convex bodies. This construction relies on a density results contained in Corollary \[corollary\_density\_various\_families\_of\_convex\_functions\] below, which we will deduce from the following proposition.
\[proposition\_epi\_graph\_support\_function\_convex\_body\] Let $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ be a finite-valued convex function, $f^*$ its convex dual. If $\|f\|_{C(B_{j+2})}\le c$, then the set $K_{f,j}:={\mathrm{epi}}(f^*)\cap \{(y,t)\in V^*\times{\mathbb{R}}\ : \ |y|\le 2c,|t|\le (2j+3)c\}$ is a convex body in $V^*$ and satisfies $$\begin{aligned}
f(x)=h_{K_{f,j}}(x,-1) \quad\text{for all }x\in B_{j+1}.
\end{aligned}$$
Consider the set $C:=\{y\in V^* \ : \ y\in\partial f(x) \text{ for some }x\in B_{j+1}\}$. As $f|_{B_{j+1}}$ is Lipschitz continuous with Lipschitz constant $L=2\|f\|_{C(B_{j+2})}$ by Proposition \[proposition\_convex\_functions\_local\_lipschitz\_constants\], Lemma \[lemma\_compactness\_face\_epigraph\_conjugate-function\] implies that $C$ is contained in the ball of radius $L$ around the origin.\
Now for any $y\in C$ we have $f^*(y)=\langle x,y\rangle -f(x)$ for some $x\in B_{j+1}$ due to Lemma \[lemma\_properties\_subgradients\]. Thus $$\begin{aligned}
|f^*(y)|\le |\langle y,x\rangle| +|f(x)|\le L(j+1)+\|f\|_{C(B_{j+2})}\le (2j+3)\|f\|_{C(B_{j+2})}\le (2j+3)c.
\end{aligned}$$ Let us show that $f(x)=h_{K_{f,j}}(x,-1)$ for all $x\in B_{j+1}$. Obviously, the left hand side is equal to or larger than the right hand side. By Lemma \[lemma\_compactness\_face\_epigraph\_conjugate-function\], we know that for any $x\in B_{j+1}$ there exists $y\in V^*$ such that $f(x)=\langle y,x\rangle-f^*(y)$, i.e. $y\in C$. Then $$\begin{aligned}
(y,f^*(y))\in {\mathrm{epi}}(f^*)\cap \{(y,t)\in V^*\times{\mathbb{R}}\ : \ |y|\le 2c,|t|\le (2j+3)c\}=K_{f,j}
\end{aligned}$$ by the previous discussion, so $$\begin{aligned}
f(x)=\langle y,x\rangle-f^*(y)\le \sup_{(\tilde{y},t)\in K_{f,j}}\langle \tilde{y},x\rangle -t= h_{K_{f,j}}(x,-1).
\end{aligned}$$ Is $f^*$ is lower semi-continuous, the set ${\mathrm{epi}}(f^*)\cap \{(y,t)\in V^*\times{\mathbb{R}}\ : \ |y|\le 2(j+2)c,|t|\le 3(j+2)c\}$ is convex, closed and bounded, i.e. it belongs to $\mathcal{K}(V^*\times{\mathbb{R}})$.
We thus obtain the following density results.
\[corollary\_density\_various\_families\_of\_convex\_functions\] The following families of functions are dense in ${\mathrm{Conv}}(V,{\mathbb{R}})$:
1. $\{h_K(\cdot,-1) \ : \ K\in\mathcal{K}(V^*\times{\mathbb{R}})\}$,
2. $\{h_P(\cdot,-1) \ : \ P\in\mathcal{K}(V^*\times{\mathbb{R}}) \text{ polytope}\}$,
3. $\{h_K(\cdot,-1) \ : \ K\in\mathcal{K}(V^*\times{\mathbb{R}})\text{ smooth and strictly convex}\}$,
4. ${\mathrm{Conv}}(V,{\mathbb{R}})\cap C^\infty(V)$.
For the first set, this follows directly from Proposition \[proposition\_epi\_graph\_support\_function\_convex\_body\] and the continuity of the map $\mathcal{K}(V^*\times{\mathbb{R}})\rightarrow{\mathrm{Conv}}(V,{\mathbb{R}})$, $K\mapsto h_K(\cdot,-1)$ (see Lemma \[lemma\_continuity\_support\_function\]). As the two sets $$\begin{aligned}
&\{P\in\mathcal{K}(V^*\times{\mathbb{R}}) \ : \ P \text{ polytope}\}, &&\{K\in\mathcal{K}(V^*\times{\mathbb{R}}) \ : \ K \text{ smooth and strictly convex}\}
\end{aligned}$$ are dense subsets of $\mathcal{K}(V^*\times{\mathbb{R}})$, this implies the density of the second and third set. For the last set, observe that the support function of any smooth and strictly convex body is smooth, so the last set contains a dense subset and is thus dense itself.
Lipschitz regularization
------------------------
Most of our results are actually results on valuations on ${\mathrm{Conv}}(V,{\mathbb{R}})$, which generalize to more general subspaces of ${\mathrm{Conv}}(V)$ by approximation.\
For $r>0$ the *Lipschitz regularization* or *Pasch-Hausdorff envelope* of a convex function $f\in{\mathrm{Conv}}(V)$ is defined as $$\begin{aligned}
{\mathrm{reg}}_r(f):=(f^*+1^\infty_{B_{1/r}})^*.
\end{aligned}$$ We will need the following properties:
\[proposition\_properties\_LIpschitz\_regularization\] For $f,h\in{\mathrm{Conv}}(V)$ and $r>0$, the Lipschitz regularization has the following properties:
1. There exists $r_0>0$ such that ${\mathrm{reg}}_rf\in{\mathrm{Conv}}(V,{\mathbb{R}})$ for all $0<r\le r_0$.
2. ${\mathrm{reg}}_r f$ epi-converges to $f$ for $r\rightarrow 0$.
3. If $x\in{\mathrm{dom }}(f)$ and $\partial f(x)\cap B_{1/r}\ne\emptyset$, then ${\mathrm{reg}}_r f(x)=f(x)$ and $\partial {\mathrm{reg}}_r f(x)=\partial f(x)\cap B_{1/r}$.
4. If $(f_j)_j$ is a sequence in ${\mathrm{Conv}}(V)$ that epi-converges to $f$, then there exists $r_0>0$ such that $({\mathrm{reg}}_rf_j)_j$ epi-converges to ${\mathrm{reg}}_rf$ for all $0<r\le r_0$.
5. If $\min(f,h)$ is convex, then there exists $r_0>0$ such that $$\begin{aligned}
&{\mathrm{reg}}_r(\max(f,h))=\max({\mathrm{reg}}_r f,{\mathrm{reg}}_r h), && {\mathrm{reg}}_r(\min(f,h))=\min({\mathrm{reg}}_r f, {\mathrm{reg}}_r h)
\end{aligned}$$ for all $0<r\le r_0$.
Note that ii. implies that ${\mathrm{Conv}}(V,{\mathbb{R}})\subset {\mathrm{Conv}}(V)$ is dense. Thus the sets considered in Corollary \[corollary\_density\_various\_families\_of\_convex\_functions\] are also dense in ${\mathrm{Conv}}(V)$.
Dually epi-translation invariant valuations on convex functions {#section_valuations_convex_functions}
===============================================================
Let $C\subset{\mathrm{Conv}}(V)$ be a non-empty subset and let $(G,+)$ be an abelian semigroup.
A map $\mu:C\rightarrow G$ is called a valuation, if $$\begin{aligned}
\mu(f)+\mu(h)=\mu(\max(f,h))+\mu(\min(f,h))
\end{aligned}$$ for all $f,h\in C$ such that the pointwise maximum $\max(f,h)$ and minimum $\min(f,h)$ belong to $C$.
A valuation $\mu:C\rightarrow G$ is called *dually epi-translation invariant*, if $$\begin{aligned}
\mu(f+\lambda+c)=\mu(f)\quad \forall f\in C, \lambda\in V^*, c\in{\mathbb{R}}\text{ s.t. } f+\lambda+c\in C.
\end{aligned}$$
If $(F,+)$ is a topological vector space, we let ${\mathrm{VConv}}(C;V,F)$ denote the space of all valuations $\mu:C\rightarrow F$ that are
1. continuous with respect to epi-convergence,
2. dually epi-translation invariant.
If $C={\mathrm{Conv}}(V,{\mathbb{R}})$, we will use the notation ${\mathrm{VConv}}(V,G)$ instead. For $F={\mathbb{R}}$, we will write ${\mathrm{VConv}}(C;V,{\mathbb{R}})={\mathrm{VConv}}(C;V)$ and ${\mathrm{VConv}}(V,{\mathbb{R}})={\mathrm{VConv}}(V)$ for brevity.\
We also equip ${\mathrm{VConv}}(C;V,F)$ with the compact-open topology, which is generated by the open sets $$\begin{aligned}
\mathcal{M}(K,O):={\{\mu\in{\mathrm{VConv}}(C;V,F) \ : \ \mu(f)\in O\quad\forall f\in K\}}
\end{aligned}$$ for $K\subset C$ compact and $O\subset F$ open. If $F$ is locally convex, ${\mathrm{VConv}}(C;V,F)$ is a locally convex vector space equipped with the family of semi-norms $$\begin{aligned}
\|\mu\|_{F;K}:=\sup_{f\in K}|\mu(f)|_F,
\end{aligned}$$ where $K\subset C$ is compact and $|\cdot|_F$ is a continuous semi-norm on $F$. It is easy to see that ${\mathrm{VConv}}(C;V,F)$ is complete, if $F$ is complete. Also note that Proposition \[proposition\_compactness\_Conv\] provides a characterization of all compact subsets for $C={\mathrm{Conv}}(V,{\mathbb{R}})$.
Relation to valuations on convex bodies {#section_relation_to_val_convex_bodies}
---------------------------------------
As noted in Section \[section\_convex\_functions\], the support function $h_K$ of $K\in\mathcal{K}(V^*)$ is defined by $h_K(y):=\sup_{x\in K}\langle y,x\rangle$. It has the following well known properties:
1. $h_K\in{\mathrm{Conv}}(V^*,{\mathbb{R}})$.
2. $h_{tK}=th_K$ for all $t\ge 0$.
3. If $K,L$ are convex bodies such that $K\cup L$ is convex, then $h_{K\cup L}=\max(h_K,h_L)$ and $h_{K\cap L}=\min(h_K, h_L)$.
4. A sequence $(K_j)_j$ of convex bodies converges to $K$ with respect to the Hausdorff metric if and only if $(h_{K_j})_j$ converges to $h_K$ uniformly on compact subsets.
The last property implies
\[lemma\_continuity\_support\_function\] The map $$\begin{aligned}
P:\mathcal{K}(V^*\times{\mathbb{R}})\rightarrow{\mathrm{Conv}}(V,{\mathbb{R}})\\
K\mapsto h_K(\cdot,-1)
\end{aligned}$$ is continuous.
Here we have used the canonical isomorphism $(V\times{\mathbb{R}})^*\cong V^*\times{\mathbb{R}}$.
\[theorem\_embedding VConv->Val(VxR)\] Let $F$ be a Hausdorff real topological vector space. For $\mu\in{\mathrm{VConv}}(C;V,F)$ and $L\in\mathcal{K}(V^*\times{\mathbb{R}})$ define $Q(\mu)(L):=\mu(h_L(\cdot,1))=\mu(P(L))$. Then $Q:{\mathrm{VConv}}(C;V,F)\rightarrow {\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ is well defined, continuous and injective. Here ${\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ is equipped with the compact-open topology.
It is clear that $Q(\mu)=\mu\circ P\in{\mathrm{Val}}(V^*\times{\mathbb{R}})$.\
Let us show that $Q$ is injective: If $Q(\mu)=0$, then $\mu(h_L(\cdot,-1))=0$ for all $L\in\mathcal{K}(V^*\times{\mathbb{R}})$. By Corollary \[corollary\_density\_various\_families\_of\_convex\_functions\], these functions form a dense subspace of ${\mathrm{Conv}}(V,{\mathbb{R}})$, which is dense in $C$, so the continuity of $\mu$ implies $\mu=0$, as $F$ is Hausdorff. Thus $Q$ is injective.\
A basis for the topology of ${\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ is given by the open sets $$\begin{aligned}
\mathcal{M}(B,O)=\{\mu\in{\mathrm{Val}}(V^*\times{\mathbb{R}},F) \ : \ \mu(L)\in O\quad\forall L\in B\},
\end{aligned}$$ where $O\subset F$ is open and $B\subset\mathcal{K}(V^*\times{\mathbb{R}})$ is compact. Then $$\begin{aligned}
Q^{-1}(\mathcal{M}(B,O))&=\{\mu\in{\mathrm{VConv}}(C;V,F) \ : \ \mu(h_L)\in O\quad\forall L\in B\}\\
&=\{\mu\in{\mathrm{VConv}}(C;V,F) \ : \ \mu(f)\in O\quad\forall f\in P(B)\}\\
&=\mathcal{M}(P(B),O).
\end{aligned}$$ As $P$ is continuous, $P(B)$ is compact in $C$, so $Q^{-1}(\mathcal{M}(B,O))=\mathcal{M}(P(B),O)$ is open in ${\mathrm{VConv}}(C;V,F)$.
Homogeneous decomposition {#section_homogeneous_decomposition}
=========================
Proof of Theorem 1
------------------
In this section we prove Theorem \[maintheorem\_homogeneous\_decomposition\]. Let $F$ be a Hausdorff real topological vector space and $C\subset{\mathrm{Conv}}(V)$.
A continuous valuation $\mu:C\rightarrow F$ is called $k$-homogeneous, if $\mu(tf)=t^k\mu(f)$ for all $f\in C$ and for all $t\ge0$ such that $tf\in C$. We will denote the space of $k$-homogeneous valuations in ${\mathrm{VConv}}(C;V,F)$ by ${\mathrm{VConv}}_k(C;V,F)$
\[proposition\_McMullen\_for\_Conv(V)\_with\_n+1\_degree\] Let $F$ be a Hausdorff real topological vector space, $C\subset{\mathrm{Conv}}(V)$ a cone containing ${\mathrm{Conv}}(V,{\mathbb{R}})$, and $\mu\in {\mathrm{VConv}}(C;V,F)$. Then there exist valuations $\mu_i\in{\mathrm{VConv}}_i(C;V,F)$, $i=0,...,n+1$ such that $$\mu=\sum_{i=0}^{n+1}\mu_i.$$
Consider the injective map $Q:{\mathrm{VConv}}(C;V,F)\rightarrow{\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ from Theorem \[theorem\_embedding VConv->Val(VxR)\], given by $Q(\mu)[K]=\mu(h_K(\cdot,-1))$.\
For $t\ge 0$ define $\mu^t\in{\mathrm{VConv}}(C;V,F)$ by $\mu^t(f):=\mu(t\cdot f)$ for $f\in C$. Then $Q(\mu^t)[K]=Q(\mu)[tK]$ as $h_{tK}=th_K$ for $t\ge 0$. Using the McMullen decomposition (Theorem \[theorem\_McMullen\_decomposition\_for\_Val\]) for ${\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ we see that $Q(\mu_t)=\sum_{i=0}^{n+1}t^i\tilde{\mu}_i$ for homogeneous elements $\tilde{\mu}_i\in {\mathrm{Val}}_i(V^*\times{\mathbb{R}},F)$. Plugging in $0<t_0<...<t_{n+1}$ and using the inverse of the Vandermonde matrix, we obtain constants $c_{ij}\in {\mathbb{R}}$ such that $\tilde{\mu}_i=\sum_{j=0}^{n+1}c_{ij}Q(\mu^{t_j})$.\
Now define $\mu_i\in{\mathrm{VConv}}(C;V,F)$ by $\mu_i:=\sum_{j=0}^{n+1}c_{ij}\mu^{t_j}$. Then obviously $Q(\mu_i)=\tilde{\mu}_i$ and for any $K\in\mathcal{K}(V^*\times{\mathbb{R}})$ $$\begin{aligned}
Q(\mu_i^t)(K)=Q(\mu_i)(tK)=\tilde{\mu}_i(tK)=t^i\tilde{\mu}_i(K)=t^iQ(\mu_i)(K)=Q(t^i\mu_i)(K).
\end{aligned}$$ The injectivity of $Q$ implies $t^i\mu_i=\mu_i^t$, i.e. $\mu_i$ is $i$-homogeneous.\
To see that $\mu=\sum_{i=0}^{n+1}\mu_i$, observe that $$\begin{aligned}
Q(\mu)=\sum_{i=0}^{n+1}\tilde{\mu}_i=\sum_{i=0}^{n+1}Q(\mu_i)=Q(\sum_{i=0}^{n+1}\mu_i).
\end{aligned}$$ Thus the injectivity of $Q$ implies $\mu=\sum_{i=0}^{n+1}\mu_i$.
To show that the $n+1$-homogeneous component is trivial, we need the following lemma, which is a direct consequence of [@Colesanti_Ludwig_Mussnig:Hessian_valuations] Lemma 16:
\[lemma\_injectivity\_characteristic\_function\] Let $\mu\in{\mathrm{VConv}}(V,F)$. If $\mu(h_K(\cdot-y))=0$ for all $K\in\mathcal{K}(V^*)$ and $y\in V$, then $\mu=0$.
\[proposition\_n+1-degree=0\] ${\mathrm{VConv}}_{n+1}(C;V,F)=0$
Let $\mu\in{\mathrm{VConv}}_{n+1}(C;V,F)$. As ${\mathrm{Conv}}(V,{\mathbb{R}})$ is dense in $C$, we only need to show that $\mu$ vanishes on finite-valued convex functions. Using Lemma \[lemma\_injectivity\_characteristic\_function\], it is thus sufficient to show $\mu(h_K(\cdot-y))=0$ for all $K\in\mathcal{K}(V^*)$ and $y\in V$. However, $K\mapsto \mu(h_K(\cdot-y))$ defines an element of ${\mathrm{Val}}_{n+1}(V^*,F)=0$. The claim follows.
This follows directly from Proposition \[proposition\_McMullen\_for\_Conv(V)\_with\_n+1\_degree\] and Proposition \[proposition\_n+1-degree=0\].
Polynomiality and polarization
------------------------------
From now on we will assume that $F$ is a Hausdorff real topological vector space and that $C\subset{\mathrm{Conv}}(V)$ is a cone containing ${\mathrm{Conv}}(V,{\mathbb{R}})$. Then we can consider the question of polynomiality for elements of ${\mathrm{VConv}}(C;V,F)$. From Theorem \[maintheorem\_homogeneous\_decomposition\] we deduce
\[corollary\_polynomiality\_valuations\] For $\mu\in {\mathrm{VConv}}(C;V,F)$, and $f_1,...,f_m\in C$, $\mu(\sum_{j=1}^{m}\lambda_j f_j)$ is a polynomial in $\lambda_j\ge 0$.
We will proceed by induction. For $m=1$, this is just Theorem \[maintheorem\_homogeneous\_decomposition\]. Assume we have shown the statement for $m$. The valuation $$\begin{aligned}
f\mapsto \mu(\sum\limits_{j=1}^{m}\lambda_j f_j+f)
\end{aligned}$$ obviously belongs to ${\mathrm{VConv}}(C;V,F)$. Using Theorem \[maintheorem\_homogeneous\_decomposition\], we obtain $$\begin{aligned}
\mu(\sum\limits_{j=1}^{m}\lambda_j f_j+tf)=\sum\limits_{i=0}^{n}t^i\mu_i(\sum\limits_{j=1}^{m}\lambda_j f_j,f)
\end{aligned}$$ where $\mu_i:C^2\rightarrow F$ is $i$-homogeneous in the second component and a continuous dually epi-translation invariant valuation in the first. The induction assumption implies that $(\lambda_1,...,\lambda_m)\mapsto \mu_i(\sum_{j=1}^{m}\lambda_j f_j,f)$ is a polynomial in $\lambda_j\ge0$, $1\le j\le m$. The claim follows.
A valuation $\mu\in {\mathrm{VConv}}(C;V,F)$ is called additive if $\mu(f+g)=\mu(f)+\mu(g)$ for all $f,g\in C$.
By continuity, any additive valuation is $1$-homogeneous.
\[theorem\_polarization\_VConv\] For $\mu\in {\mathrm{VConv}}_k(C;V,F)$, there exists a unique map $\bar{\mu}:C^k\rightarrow F$, called the *polarization* of $\mu$, with the following properties:
1. $\bar{\mu}$ is additive and $1$-homogeneous in each coordinate,
2. $\bar{\mu}$ is symmetric,
3. $\mu(f)=\bar{\mu}(f,...,f)$ for all $f\in C$.
We start by showing uniqueness: Using 1. and 3. we obtain $$\begin{aligned}
\mu(\sum\limits_{j=1}^{k}\lambda_jf_j)=\bar{\mu}(\sum\limits_{j=1}^{k}\lambda_jf_j,...,\sum\limits_{j=1}^{k}\lambda_jf_j)=\sum\limits_{j_1,...,j_k=1}^k\lambda_{j_1}...\lambda_{j_k}\bar{\mu}(f_{j_1},...,f_{j_k})
\end{aligned}$$ Differentiating and using 2., we obtain the formula $$\begin{aligned}
\label{equation_definition_polarization}
k!\bar{\mu}(f_1,...,f_k)=\frac{\partial}{\partial\lambda_1}\Big|_0...\frac{\partial}{\partial\lambda_k}\Big|_0\mu(\sum\limits_{j=1}^{k}\lambda_jf_j)
\end{aligned}$$ This shows uniqueness. To prove the existence of $\bar{\mu}$, we use Corollary \[corollary\_polynomiality\_valuations\] to see that the right-hand side of is actually well defined, so we can use this equation to define $\bar{\mu}$.\
Obviously the definition is symmetric in $f_1,...,f_n$. To see that $\bar{\mu}$ is additive in each coordinate, we thus only need to consider one coordinate: Setting $$\begin{aligned}
f(t,s):=&\frac{1}{k!}\frac{\partial }{\partial \lambda_1}\big|_0...\frac{\partial }{\partial \lambda_{k-1}}\big|_0\mu(\sum\limits_{j=1}^{k-1}\lambda_j f_j+tf+sg),\\
g(t):=& f(t,t)
\end{aligned}$$ we obtain $$\begin{aligned}
\bar{\mu}(f_1,...,f_{k-1},f+g)&=g'(0)=\frac{\partial f}{\partial t}|_{(0,0)}+\frac{\partial f}{\partial s}|_{(0,0)}
=\bar{\mu}(f_1,...,f_{k-1},f)+\bar{\mu}(f_1,...,f_{k-1},g).
\end{aligned}$$ Let us see that we can recover $\mu(f)$: $$\begin{aligned}
\bar{\mu}(f,...,f)=\frac{1}{k!}\frac{\partial}{\partial\lambda_1}\Big|_0...\frac{\partial}{\partial\lambda_k}\Big|_0\mu(\sum\limits_{j=1}^{k}\lambda_jf)=\frac{1}{k!}\frac{\partial}{\partial\lambda_1}\Big|_0...\frac{\partial}{\partial\lambda_k}\Big|_0\left(\sum\limits_{j=1}^{k}\lambda_j\right)^k\cdot\mu(f)
\end{aligned}$$ as $\mu$ is $k$-homogeneous. Thus $\bar{\mu}(f,...,f)=\mu(f)$.
Note that the construction shows that $\bar{\mu}$ is a dually epi-translation invariant valuation in each coordinate. We will now show that $\bar{\mu}$ is jointly continuous. From the defining properties of $\bar{\mu}$ we deduce the following corollary.
\[corollary\_boundedness\_degree\_polynomial\_valuation\] For $\mu\in {\mathrm{VConv}}_k(C;V,F)$, $m\in\mathbb{N}$ and $f_1,...,f_m\in C$, $\mu(\sum_{j=1}^{m}\lambda_jf_j)$ is a polynomial of degree at most $k$ in $\lambda_j\ge 0$.
\[corollary\_continuity\_polarization\] $\bar{\mu}:C^k\rightarrow{\mathbb{R}}$ is continuous for $\mu\in{\mathrm{VConv}}_k(C;V,F)$.
Assume we are given sequences $(f_{i,j})_j$ $1\le i\le k$ in $C$ such that each sequence $(f_{i,j})_j$ converges to some $f_i\in C$. Then the polynomials $P_j(\lambda_1,...,\lambda_k):=\mu(\sum_{i=1}^{k}\lambda_if_{i,j})$ converge pointwise to $P(\lambda_1,...,\lambda_k):=\mu(\sum_{i=1}^{k}\lambda_if_{i})$ for $\lambda_i\ge0$. Note that the degree of $P_j$ is bounded by $k$ due to Corollary \[corollary\_boundedness\_degree\_polynomial\_valuation\], so we see that the coefficient in front of $\lambda_1\cdot...\cdot\lambda_k$ converges. Now the claim follows from the definition of $\bar{\mu}$ in the proof of Theorem \[theorem\_polarization\_VConv\].
We close this section with an inequality that will be used in the construction of the Goodey-Weil embedding. It also shows that the map, which associates the polarization to a given valuation, is continuous with respect to the natural topologies.
\[lemma\_continuity\_valuation->polarization\] There exists a constant $C_{k}>0$ such that the following holds: If $K\subset {\mathrm{Conv}}(V,{\mathbb{R}})$ is compact, then $$\begin{aligned}
\|\bar{\mu}\|_{F;K}:=\sup_{f_1,...,f_k\in K}|\bar{\mu}(f_1,...,f_k)|_F\le C_{k}\|\mu\|_{F;K'}
\end{aligned}$$ for every semi-norm $|\cdot |_F$ on $F$ and all $\mu\in{\mathrm{VConv}}_k(V,F)$, where $$\begin{aligned}
K':=\sum_{j=1}^k \bigcup\limits_{i=1}^jiK=\{\sum\limits_{j=1}^{k}f_j \ : \ f_j\in \bigcup\limits_{i=1}^jiK\}\subset {\mathrm{Conv}}(V,{\mathbb{R}})
\end{aligned}$$ is compact. If $K$ is convex with $0\in K$, there exists a constant $C_k'>0$ independent of $K$ such that $$\begin{aligned}
\|\bar{\mu}\|_{F;K}\le C_{k}'\|\mu\|_{F;K}.
\end{aligned}$$
For $f,g\in {\mathrm{Conv}}(V,{\mathbb{R}})$ and $\lambda\ge 0$, we have $$\begin{aligned}
\mu(f+\lambda g)=\sum\limits_{i=0}^{k}\lambda^{k-i}i!(k-i)!\bar{\mu}(f[i],g[k-i])=\sum\limits_{i=0}^{k}\lambda^{i}i!(k-i)!\bar{\mu}(f[k-i],g[i]).
\end{aligned}$$ For the polarization we are only interested in the linear term. Plugging in $\lambda=0,...,k$ and setting $\mu_i'(f):=\mu(f+ig)$ we obtain a valuation $\mu'=(\mu_0',...,\mu_{k}')$ with values in $F^{k+1}$. Denoting the $k+1$ components of $F^k$ with subindices $0,...,k$ and the Vandermonde matrix with entries corresponding to $(0,...,k)$ by $S_k$, we see that the $i$-th entry of $S_k^{-1}\mu'(f)\in F^{k+1}$ is exactly $i!(k-i)!\bar{\mu}(f[k-i],g[i])$, i.e. the entry $[S_k^{-1}\mu'(f)]_{1}$ is $\bar{\mu}(f[k-1],g)$ up to scaling: $$\begin{aligned}
\bar{\mu}(f[k-1],g)=\frac{1}{(k-1)!}[S_k^{-1}\mu'(f)]_{1}.
\end{aligned}$$ If we equip $F^{k+1}$ with the family of semi-norms $|(v_0,...,v_k)|_F:=\max_{i=0,...,k}|v_i|_F$, and denote by $\|S_k\|_\infty$ the operator norm of $S_k:{\mathbb{R}}^{k+1}\rightarrow {\mathbb{R}}^{k+1}$ with respect to the maximum norm on ${\mathbb{R}}^{k+1}$, we obtain $$\begin{aligned}
|\bar{\mu}(f[k-1],g)|_F\le\frac{1}{(k-1)!}\|S_k^{-1}\|_\infty \ |\mu'(f)|_F.
\end{aligned}$$ For $f\in {\mathrm{Conv}}(V,{\mathbb{R}})$, $g\in K$ the value of $\mu'(f)$ only depends on $\mu$ evaluated in an element of $f+\bigcup_{i=0}^kiK$, i.e. we obtain $$\begin{aligned}
|\bar{\mu}(f[k-1],g)|_F\le \|S_k^{-1}\|_\infty \|\mu\|_{F;f+\bigcup\limits_{i=0}^kiK}
\end{aligned}$$ Iterating this construction starting with the $k-1$-homogeneous valuation $\nu(f):=\bar{\mu}(f[k-1],g)$, we see that there exists $C_{k}'>0$ depending on $k$ only such that for $f_1,...,f_k\in K$ $$\begin{aligned}
|\bar{\mu}(f_1,...,f_k)|_F\le C_{k}\sup\{|\mu(g)|_F: g\in \sum_{j=1}^k \bigcup\limits_{i=0}^jiK\}= C_{k}\|\mu\|_{F,K'}
\end{aligned}$$ for every semi-norm $|\cdot|_F$ on $F$. Also note that $K'$ is compact, as it is a sum of finitely many compact subsets of ${\mathrm{Conv}}(V,{\mathbb{R}})\subset C(V)$.\
If $K$ is convex and $0\in K$ we have $K'\subset k^2K$, so $$\begin{aligned}
\|\mu\|_{F;K'}=\sup\limits_{g\in K'}|\mu(g)|_F\le \sup\limits_{g\in k^2K}|\mu(g)|_F\le (k^2)^k\sup\limits_{g\in K}|\mu(g)|=k^{2k}\|\mu\|_{F;K},
\end{aligned}$$ i.e. we can choose $C_k':=k^{2k}C_k$.
Goodey-Weil embedding
=====================
Construction and basic properties {#section_Goodey-Weil-embedding}
---------------------------------
In this section, we will assume that $V$ carries an euclidean structure. Let $C^2_b(V)$ denote the Banach space of twice differentiable functions with finite $C^2$-norm $$\begin{aligned}
\|\phi\|_{C^2_b(V)}:=&\|\phi\|_\infty+\|\nabla \phi\|_\infty+\|H_\phi\|_\infty
=\sup\limits_{x\in V}|\phi(x)|+\sup\limits_{x\in V}|\nabla f(x)|+\sup\limits_{x\in V,v\in S_V}|\langle H_\phi(x)v,v\rangle|,
\end{aligned}$$ where $H_\phi$ denotes the Hessian matrix of a twice differentiable function $\phi:V\rightarrow {\mathbb{R}}$ and $S_V\subset V$ is the unit sphere.
\[lemma\_difference\_C2\_convex\_functions\] There exists constants $c(A)>0$ for $A\subset V$ compact with the following property: For every $\phi\in C_b^2(V)$ there exist two convex functions $f,h\in{\mathrm{Conv}}(V,{\mathbb{R}})$ such that $f-h=\phi$ and such that $\|f|_A\|_{\infty},\|h|_A\|_{\infty}\le c(A)\|\phi\|_{C_b^2(V)}$ for all compact subsets $A\subset V$.
Simply take $f(x):=c\frac{|x|^2}{2}+\phi(x)$, $h(x)=c\frac{|x|^2}{2}$, where $c:= \|\phi\|_{C_b^2(V)}$. Then $f$ and $h$ are convex, as their Hessians are positive semidefinite. In addition $$\begin{aligned}
\|h|_A\|_{\infty},\|f|_A\|_{\infty}\le c\sup_{x\in A}\frac{|x|^2}{2}+\|\phi\|_\infty\le \left(\sup_{x\in A}\frac{|x|^2}{2}+1\right)\|\phi\|_{C_b^2(V)}.
\end{aligned}$$ Setting $c(A):=\sup_{x\in A}\frac{|x|^2}{2}+1$ we obtain the inequalities.
To every $\mu\in {\mathrm{VConv}}_k(C;V,F)$ we can associate a $k$-multilinear continuous functional $$\begin{aligned}
\tilde{\mu}:C_b^2(V)^k\rightarrow F
\end{aligned}$$ as follows: Given functions $\phi_1,...,\phi_k\in C^2_b(V)$ we take the polarization $\bar{\mu}$ of $\mu$ and convex functions $f_1$,..., $f_k$, $h_1$,...,$h_k$ in ${\mathrm{Conv}}(V,{\mathbb{R}})$ with $\phi_i=f_i-h_i$ satisfying the inequality from Lemma \[lemma\_difference\_C2\_convex\_functions\] and define inductively for arbitrary convex functions $g_1,...,g_k\in{\mathrm{Conv}}(V,{\mathbb{R}})$ and $1\le i\le k-1$: $$\begin{aligned}
\mu^{(1)}(\phi_1,g_2,...,g_k):=&\bar{\mu}(f_1,g_2,...,g_k)-\bar{\mu}(h_1,g_2,...,g_k),\\
\mu^{(i+1)}(\phi_1,...,\phi_{i+1},g_{i+2},...,g_k):=&\mu^{(i)}(\phi_1,...,\phi_i,f_{i+1},g_{i+1},...,g_k)-\mu^{(i)}(\phi_1,...,\phi_i,h_{i+1},g_{i+1},...,g_k).
\end{aligned}$$ Then we set $\tilde{\mu}(\phi_1,...,\phi_k):=\mu^{(k)}(\phi_1,...,\phi_k)$. It is easy to see that this is equivalent to $$\begin{aligned}
\label{equation_representation_GW_using_polarization}
\tilde{\mu}(\phi_1,..,\phi_k)=\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f_{\sigma(1)},...,f_{\sigma(l)},h_{\sigma(l+1)},...,h_{\sigma(k)}),
\end{aligned}$$ where $\bar{\mu}$ denotes the polarization of $\mu$ from Theorem \[theorem\_polarization\_VConv\].\
Using the additivity of $\bar{\mu}$ in each component, one readily verifies that this definition only depends on the functions $\phi_1,...,\phi_k$ (and not the special choices of $f_i$ and $h_i$) and that this functional is multilinear. To see that it is continuous, observe that the set $K$ of convex functions that are bounded by $c(A)$ on every compact set $A$ (as defined above) is compact in ${\mathrm{Conv}}(V,{\mathbb{R}})$ by Proposition \[proposition\_compactness\_Conv\], so it is also compact in $C$. In particular, $\mu$ is bounded on this set. As $K$ is also convex with $0\in K$, Lemma \[lemma\_continuity\_valuation->polarization\] and Equation imply $$\begin{aligned}
\label{equation_continuity_GW}
|\tilde{\mu}(\phi_1,...,\phi_k)|_F\le c_{k}\|\mu\|_{F;K}\cdot \prod\limits_{i=1}^ k\|\phi_i\|_{C_b^2}(V)
\end{aligned}$$ for any continuous semi-norm $|\cdot|_F$ on $F$ for some constant $c_{k}>0$ depending on $k$ only (we can choose $2^k$ times the constant $C'_k$ from Lemma \[lemma\_continuity\_valuation->polarization\]). Thus $\tilde{\mu}$ is continuous.\
We will need the well known L. Schwartz kernel theorem:
Let $F$ be a complete locally convex vector space and let $V,W$ be finite dimensional real vector spaces. For every continuous bilinear map $$\begin{aligned}
b:C^\infty_c(V)\times C^\infty_c(W)\rightarrow F
\end{aligned}$$ there exists a unique continuous linear map $$\begin{aligned}
B:C^\infty_c(V\times W)\rightarrow F
\end{aligned}$$ such that $B(f\otimes h)=b(f,h)$ for all $f\in C^\infty_c(V)$, $h\in C^\infty_c(W)$.
Let us denote by $\mathcal{D}'(V,F)$ the space of all distributions on $V$ with values in a locally convex vector space $F$, i.e. the space of all continuous linear functionals $\phi:C^\infty_c(V)\rightarrow F$. Applying the L. Schwartz kernel theorem to the functional $\tilde{\mu}$ for $\mu\in {\mathrm{VConv}}_k(C;V,F)$ leads to the following notion:
\[definition\_Goddey\_Weil\_embedding\_VCONV\] Let $F$ be a locally convex vector space, $1\le k\le n$. To every $\mu\in{\mathrm{VConv}}_k(C;V,F)$ we associate the distribution $GW(\mu)\in \mathcal{D}'(V^k,\bar{F})$ determined by $$\begin{aligned}
GW(\mu)(\phi_1\otimes...\otimes\phi_k)=\tilde{\mu}(\phi_1,...,\phi_k)
\end{aligned}$$ for $\phi_1,...,\phi_k\in C_c^\infty(V)$. The distribution will be called the *Goodey-Weil distribution* of $\mu$.
To show that this distribution has compact support in $V^k$ if $F$ admits a continuous norm, we need the following lemma.
\[lemma\_smooth function difference of convex functions\] Let $\phi\in C^2(V)$. Then there exists a convex function $h\in{\mathrm{Conv}}(V,{\mathbb{R}})$ with the following property: If $\psi\in C^\infty(V)$ is a function with $\|\psi\|_{C^2(B_j)}\le \|\phi\|_{C^2(B_j)}$ for all $j\in\mathbb{N}$, then $h+\psi$ is convex.
Assume we are given $\phi$ and let $\psi$ be an arbitrary function with the property stated above. Let us inductively define a sequence of convex functions $h_j\in{\mathrm{Conv}}(V,{\mathbb{R}})$. Set $c_j:=\|\phi\|_{C^2(B_{j+1})}$, such that $c_j\ge\sup_{x\in B_{j+1},v\in S_V}\langle H_\psi(x)v,v\rangle $.\
For $j=1$ define $h_1$ by $h_1(x):=c_1\frac{|x|^2}{2}$. Then obviously $h_1+\psi$ is convex on $B_2$.\
Assume we have already constructed $h_j$. Then $c_{j+1}\frac{|x|^2}{2}+\psi$ is convex on $B_{j+2}$ and we set $$\begin{aligned}
h_{j+1}(x):=&\max\{c_{j+1}\frac{|x|^2-j^2}{2},0\}+h_j(x).
\end{aligned}$$ Thus $h_{j+1}$ is a finite-valued convex function for all $\in\mathbb{N}$ that coincides with $h_j$ on $B_j$. We deduce that for each point $x\in V$ the sequence $(h_j(x))_j$ becomes constant for large $j$ and thus $(h_j)_j$ converges pointwise to a function $h\in{\mathrm{Conv}}(V,{\mathbb{R}})$.\
To see that $h+\psi$ is convex, observe that for every point $x\in V$, there exists an open convex neighborhood such that the restriction of $h+\psi$ to this neighborhood is convex, i.e. $h+\psi$ is a locally convex function: By construction, $$\begin{aligned}
h(x)+\psi(x)=&\max\{c_{j+1}\frac{|x|^2-j^2}{2},0\}+h_j(x)+\psi\\
=&\max\{c_{j+1}\frac{|x|^2-j^2}{2},0\}+h_{j-1}+c_{j}\frac{|x|^2-(j-1)^2}{2}+\psi
\end{aligned}$$ on $U_{j+1}\setminus B_{j-1}$, where $c_{j}\frac{|x|^2-(j-1)^2}{2}+\psi$ is locally convex on this set. Obviously the other two functions are locally convex on this set as well, so the same applies to $h+\psi$.\
As any locally convex function defined on $V$ is convex, we see that $h+\psi$ is convex.
We will now prove the first part of Theorem \[maintheorem\_GW\].
\[theorem\_compact\_support\_GW\] Let $F$ be a locally convex vector space admitting a continuous norm. Then the following holds: For every $\mu\in{\mathrm{VConv}}_k(C;V,F)$ the distribution $GW(\mu)\in\mathcal{D}'(V^k,\bar{F})$ has compact support and is uniquely defined by the following property: If $f_1,...,f_k\in {\mathrm{Conv}}(V,{\mathbb{R}})\cap C^\infty(V)$ then $$\begin{aligned}
\label{equation_defining_equality_GW}
GW(\mu)(f_1\otimes...\otimes f_k)=\bar{\mu}(f_1,...,f_k),
\end{aligned}$$ where $\bar{\mu}$ is the polarization of $\mu$.
Uniqueness follows directly from Equation , as every function $\phi\in C^\infty_c(V)$ can be written as a difference of two smooth convex functions due to Lemma \[lemma\_difference\_C2\_convex\_functions\] and a distribution on $V^k$ is uniquely defined by its values on functions of the form $\phi_1\otimes...\otimes\phi_k$ for $\phi_1,...,\phi_k\in C^\infty_c(V)$ by the L. Schwartz kernel theorem.\
Let us assume that $GW(\mu)$ does not have compact support and let $\|\cdot\|$ denote a continuous norm on $F$. Then we can inductively define a sequence of functions $(\phi_i^j)_j$ in $C_c^\infty(V)$ for each $1\le i\le k$ and a strictly increasing sequence $(r_j)_j$ of positive real numbers with $\lim\limits_{j\rightarrow\infty} r_j=\infty$ with the properties
1. For each $1\le i\le k$ the functions $(\phi_i^j)_j$ have pairwise disjoint support.
2. The support of $\phi_i^j$ is contained in $V\setminus B_{r_j}(0)$ for all $j\in\mathbb{N}$, $1\le i\le k$.
3. $\|GW(\mu)(\phi_1^j\otimes...\otimes \phi_k^j)\|=\|\tilde{\mu}(\phi_1^j,...,\phi_k^j)\|=1$.
as follows: Assume we have constructed $\phi_1^j,...\phi_k^j$ as well as $r_j>0$. First choose $r_{j+1}>0$ such that the restriction of $GW(\mu)$ to $[U_{r_{j+1}}\setminus B_{r_j}]^k\subset V^k$ does vanish. Then take $\phi_1^{j+1},...,\phi_k^{j+1}\in C^\infty(U_{r_{j+1}}\setminus B_{r_j})$ with $GW(\mu)[\phi_1^{j+1}\otimes\dots\otimes\phi_k^{j+1}]\ne0$ and rescale one function by an appropriate constant.\
Note that $\phi_i:=\sum_{j=1}^{\infty}\phi_{i}^j\in C^\infty(V)$ is well defined as this sum is locally finite. More precisely, the supports of the functions $\phi_i^j$ are disjoint for each $1\le i\le k$, so $\|\phi_i^j\|_{C^2(B_N)}\le \|\phi_i\|_{C^2(B_N)}$ for all $N\in\mathbb{N}$ and $j\in\mathbb{N}$. Applying Lemma \[lemma\_smooth function difference of convex functions\] to the functions $\phi_i$, we find convex functions $h_1,..., h_k\in{\mathrm{Conv}}(V,{\mathbb{R}})$ such that for $1\le i\le k$ the function $f_i^j:=h_i+\phi_i^j$ is convex for all $j\in\mathbb{N}$. By Equation $$\begin{aligned}
\tilde{\mu}(\phi_1^j,...,\phi_k^j)=&\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f^j_{\sigma(1)},...,f^j_{\sigma(l)},h_{\sigma(l+1)},...,h_{\sigma(k)}).
\end{aligned}$$ As $f_i^j\rightarrow h_i$ uniformly on compact subsets for all $1\le i\le k$, the joint continuity of the polarization $\bar{\mu}$ from Corollary \[corollary\_continuity\_polarization\] implies $$\begin{aligned}
\lim\limits_{j\rightarrow\infty}\|\tilde{\mu}(\phi_1^j,...,\phi_k^j)\|
=&\|\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(h_{\sigma(1)},...,h_{\sigma(l)},h_{\sigma(l+1)},...,h_{\sigma(k)})\|\\
=&\|(-1)^k\sum\limits_{l=0}^{k}(-1)^l\frac{k!}{(k-l)!l!}\bar{\mu}(h_1,...,h_k)\|\\
=&\|(-1)^k\bar{\mu}(h_1,...,h_k)\sum\limits_{l=0}^{k}(-1)^l\binom{k}{l}\|\\
=&\|\bar{\mu}(h_1,...,h_k)\|\cdot0=0.
\end{aligned}$$ We arrive at a contradiction to $\|\tilde{\mu}(\phi_1^j,...,\phi_k^j)\|=1$ for all $j\in\mathbb{N}$, so the distribution $GW(\mu)$ must have compact support.\
It remains to see that $GW(\mu)(f_1\otimes...\otimes f_k)=\bar{\mu}(f_1,...,f_k)$ for all convex functions $f_i\in {\mathrm{Conv}}(V,{\mathbb{R}})\cap C^\infty(V)$. Take a sequence of functions $\phi_j\in C_c^\infty(V)$ with $\phi_j\equiv 1$ on $B_j(0)$ and such that $\|\phi_j\|_{C^2(V)}\le C$ for all $j\in\mathbb{N}$ and some $C>0$. Such a sequence can be constructed by setting $\phi_j(x):=\psi(\frac{x}{j})$ for $\psi\in C^\infty_c(V)$ with $\psi\equiv 1$ on $B_1(0)$. As the support of $GW(\mu)$ is compact, we obtain $N\in\mathbb{N}$ such that $$\begin{aligned}
GW(\mu)(f_1\otimes...\otimes f_k)=GW(\mu)(\phi_j f_1\otimes...\otimes \phi_j f_k)\quad\forall j\ge N.
\end{aligned}$$ Using the Leibnitz-rule, we see that there exists $C'>0$ such that for any compact set $A\subset V$ and all $j\in\mathbb{N}$ the inequality $\|\phi_jf_i\|_{C^2(A)}\le C'\|f_i\|_{C^2(A)}$ holds. Now take the function $h_i$ from Lemma \[lemma\_smooth function difference of convex functions\] for the function $\phi=C'f_i$. Then $h_i+\phi_j f_i$ is convex for all $j\in\mathbb{N}$ and $h_i+\phi_jf_i$ converges to $h_i+f_i$ uniformly on compact subsets, i.e. in ${\mathrm{Conv}}(V,{\mathbb{R}})$. Plugging in the definition of $\tilde{\mu}$ and using the joint continuity of the polarization $\bar{\mu}$, we obtain $$\begin{aligned}
& GW(\mu)(f_1\otimes...\otimes f_k)=\lim\limits_{j\rightarrow\infty}GW(\mu)(\phi_j f_1\otimes...\otimes\phi_j f_k)\\
=&\lim\limits_{j\rightarrow\infty}\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(h_{\sigma(1)}+\phi_jf_{\sigma(1)},...,h_{\sigma(l)}+\phi_jf_{\sigma(l)},h_{\sigma(l+1)},...,h_{\sigma(k)})\\
=&\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(h_{\sigma(1)}+f_{\sigma(1)},...,h_{\sigma(l)}+f_{\sigma(l)},h_{\sigma(l+1)},...,h_{\sigma(k)})\\
=&\bar{\mu}(f_1,...,f_k),
\end{aligned}$$ where we have used the additivity of $\bar{\mu}$ in the last step.
\[corollary\_GW\_injective\] $GW:{\mathrm{VConv}}_k(C;V,F)\rightarrow \mathcal{D}'(V^k,\bar{F})$ is injective.
Assume first that $F$ admits a continuous norm. As ${\mathrm{Conv}}(V,{\mathbb{R}})\cap C^\infty(V)$ is dense in $C$ due to Proposition \[proposition\_properties\_LIpschitz\_regularization\], the claim follows from Theorem \[theorem\_compact\_support\_GW\] and the continuity of $\mu$.\
If $F$ is an arbitrary locally convex vector space, then the definition of $GW$ implies $$\begin{aligned}
\lambda\circ GW(\mu)=GW(\lambda\circ\mu) \quad \forall \lambda\in \bar{F}'\cong F',
\end{aligned}$$ where $F'$ denotes the topological dual of $F$. In particular, $GW(\mu)=0$ if and only if $GW(\lambda\circ\mu)=0$ for all $\lambda\in F'$. By the previous discussion, $GW(\lambda\circ\mu)=0$ implies $\lambda\circ\mu=0$. If this holds for all $\lambda\in F'$, we obtain $\mu=0$, as $F$ is locally convex.
Let us contrast the compactness of the support of $GW(\mu)$ for valuations with values in a locally convex vector space admitting a continuous norm with the more general case. Consider the following example: As a special case of the Hessian measures examined by Colesanti, Ludwig and Mussnig in [@Colesanti_Ludwig_Mussnig:Hessian_valuations], we can consider the valuation $\Phi_n\in{\mathrm{VConv}}_n(V,\mathcal{M}(V))$ with values in the space $\mathcal{M}(V)$ of signed Radon measures on $V$ that extends $$\begin{aligned}
\tilde{\Phi}_n(f)[B]:=\int_V 1_B\det(H_f(x))dx \quad \forall f\in{\mathrm{Conv}}(V,{\mathbb{R}})\cap C^2(V), B\subset V \text{ Borel set},
\end{aligned}$$ where $\mathcal{M}(V)$ is equipped with the vague topology, i.e. the topology induced by the semi-norms $|\mu|_\phi:=|\int_V\phi d\mu|$ for $\phi\in C_c(V)$. Then $\mathcal{M}(V)$ is a complete locally convex vector space that does not admit a continuous norm. The Goodey-Weil distribution $GW(\Phi_n)[\phi_1\otimes...\otimes\phi_n]$ is the signed measure given by integrating the mixed determinant of the Hessians of the functions $\phi_1,...,\phi_n\in C^\infty_c(V)$. In particular, $GW(\Phi_n)$ does not have compact support.
Diagonality of the support of the Goodey-Weil distributions {#section_support_GW_diagonal}
-----------------------------------------------------------
The following theorem is an adaption of the proof of the corresponding statement for the Goodey-Weil embedding for translation invariant valuations on convex bodies (see [@Alesker:McMullenconjecture]). It also proves the second part of Theorem \[maintheorem\_GW\].
\[theorem\_support\_GW\_diagonal\] Let $F$ be a locally convex vector space. For $\mu\in{\mathrm{VConv}}_k(C;V,F)$ the support of $GW(\mu)$ is contained in the diagonal in $V^k$.
Let us assume that $V$ carries a euclidean structure. Using a partition of unity, it is sufficient to show that $GW(\mu)(h_1\otimes... \otimes h_k)=0$ if $h_1,...,h_k\in C^\infty_c(V)$ are smooth functions satisfying ${\mathrm{supp\ }}h_i\subset U_\epsilon (a_i)$ where $a_i\in V$, $1\le i\le k$ are points with $U_{\epsilon}(a_i)\cap U_{\epsilon}(a_j)=\emptyset$ for $i\ne j$ and some fixed $\epsilon>0$. Noting that $\lambda\circ GW(\mu)=GW(\lambda\circ\mu)$ for all $\lambda\in \bar{F}'$, we only need to show this claim for $F={\mathbb{R}}$.\
First, there exists $\delta>0$ such that $1+|x|^2+\sum_{i=1}^{k}\delta_i h_i$ is convex and non-negative for all $|\delta_i|\le \delta$. Set $$\begin{aligned}
H(x):=1+|x|^2+ \sum\limits_{i=3}^{k}\delta_i h_i
\end{aligned}$$ and choose an affine hyperplane that separates $U_\epsilon(a_1)$ and $U_\epsilon(a_2)$. This plane is given by the equation $\langle y,x-x_0\rangle=0$ for some $y,x_0\in V$. We can choose $y$ such that $U_\epsilon(a_1)$ is contained in the positive half space with respect to the normal $y$. Define the convex functions $$\begin{aligned}
G_\pm(x):=&\max(0,\pm\langle y,x-x_0\rangle)
=\begin{cases}
0 & \pm\langle y,x-x_0\rangle\le 0,\\
\pm \langle y,x-x_0\rangle & \pm\langle y,x-x_0\rangle>0.
\end{cases}
\end{aligned}$$ As $G_\pm$ is positive on the supports of $h_1$ and $h_2$ respectively, we can rescale $y$ such that $G_+$ is larger than $H+\delta_1h_1$ on the support of $h_1$ and $G_-$ is larger than $H+\delta_2h_2$ on the support of $h_2$ for all $|\delta_i|\le\delta$. Now set $\tilde{H}_+:=\max(H+\delta_1h_1,G_+)$ and $\tilde{H}_-:=\max(H+\delta_1h_1,G_-)$. Then $\tilde{H}_+$ and $\tilde{H}_-$ are convex functions with $$\begin{aligned}
\tilde{H}_+(x)=&\begin{cases}
H(x)+\delta_1h_1(x) & \langle y,x-x_0\rangle\le 0,\\
\max(H(x)+\delta_1h_1(x),\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle > 0,
\end{cases}\\
\tilde{H}_-(x)=&\begin{cases}
H(x)+\delta_1h_1(x) & \langle y,x-x_0\rangle\ge 0,\\
\max(H(x)+\delta_1h_1(x),-\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle < 0.
\end{cases}
\end{aligned}$$ In particular $\min (\tilde{H}_+,\tilde{H}_-)=\tilde{H}:=H+\delta_1 h_1$ is convex. Because the support of $h_1$ is contained in the positive half space and because of the definition of $y$, we see that in fact $$\begin{aligned}
\tilde{H}_+(x)=&\begin{cases}
H(x) & \langle y,x-x_0\rangle\le 0,\\
\max(H(x),\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle > 0,
\end{cases}\\
\tilde{H}_-(x)=&\begin{cases}
H(x)+\delta_1h_1(x) & \langle y,x-x_0\rangle\ge 0,\\
\max(H(x),-\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle < 0.
\end{cases}
\end{aligned}$$ Thus $\max(\tilde{H}_+,\tilde{H}_-)=\max (H,|\langle y,\cdot-x_0\rangle|)$.\
Let us also define $H_\pm:=\max(H,G_\pm)$. Then $H_+=\tilde{H}_+$ and using the non-negativity of $H$ it is easy to see that $\min(H_+,H_-)=H$ and $\max(H_+,H_-)=\max (H,|\langle y,\cdot-x_0\rangle|)=\max(\tilde{H}_+,\tilde{H}_-)$.\
Using the valuation property, we obtain the two equations $$\begin{aligned}
\mu(\tilde{H}_+)+\mu(\tilde{H}_-)=\mu(\max(\tilde{H}_+,\tilde{H}_-))+\mu(\min (\tilde{H}_+,\tilde{H}_-)),\\
\mu(H_+)+\mu(H_-)=\mu(\max(H_+,H_-))+\mu(\min (H_+,H_-)).
\end{aligned}$$ Thus using $\max (H_+,H_-)=\max (\tilde{H}_+,\tilde{H}_-)$, and $\tilde{H}_+=H_+$, we obtain $$\begin{aligned}
\mu(\tilde{H}_-)-\mu(H_-)=\mu(\min(\tilde{H}_+,\tilde{H}_-))-\mu(\min(H_+,H_-))
\end{aligned}$$ by subtracting the two equations. Plugging in the relations for the two minima we arrive at $$\begin{aligned}
\mu(\tilde{H}_-)-\mu(H_-)=\mu(\tilde{H})-\mu(H).
\end{aligned}$$ Set $$\begin{aligned}
\Delta(x):=& H_-(x)=\begin{cases}
H(x) & \langle y,x-x_0\rangle\ge0,\\
\max(H(x),-\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle < 0,
\end{cases}\\
\tilde{\Delta}(x):=& \tilde{H}_-(x)=\begin{cases}
H(x) +\delta_1h_1(x)& \langle y,x-x_0\rangle\ge0,\\
\max(H(x),-\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle < 0,
\end{cases}
\end{aligned}$$ to rewrite the previous equation as $$\begin{aligned}
\mu(\tilde{\Delta})-\mu(\Delta)=\mu(\tilde{H})-\mu(H)=\mu(H+\delta_1h_1)-\mu(H).
\end{aligned}$$ Now if we replace $H$ by $H+\delta_2h_2$ and repeat the argument, we obtain convex functions $\Delta'$ and $\tilde{\Delta}'$ satisfying $$\begin{aligned}
\mu(\tilde{\Delta}')-\mu(\Delta')=\mu(H+\delta_1h_1+\delta_2h_2)-\mu(H+\delta_2h_2)
\end{aligned}$$ given by $$\begin{aligned}
\Delta'(x)=& \begin{cases}
H(x) +\delta_2h_2(x)& \langle y,x-x_0\rangle\ge0,\\
\max(H(x)+\delta_2h(x),-\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle < 0,
\end{cases}\\
\tilde{\Delta}'(x)=& \begin{cases}
H(x) +\delta_1h_1(x)+\delta_2h_2(x)& \langle y,x-x_0\rangle\ge0,\\
\max(H(x)+\delta_2h_2(x),-\langle y,x-x_0\rangle )& \langle y,x-x_0\rangle < 0.
\end{cases}
\end{aligned}$$ However, the support of $h_2$ is contained in the negative halfspace and $H+\delta_2 h_2$ is smaller than $-\langle y,x-x_0\rangle$ on the support of $h_2$. Thus we see that $\Delta'=\Delta$ and $\tilde{\Delta}'=\tilde{\Delta}$ and we obtain the equation $$\begin{aligned}
\mu(H+\delta_1h_1)-\mu(H)=\mu(\tilde{\Delta})-\mu(\Delta)=\mu(H+\delta_1h_1+\delta_2h_2)-\mu(H+\delta_2h_2).
\end{aligned}$$ for all $\delta_i$ with $|\delta_i|<\delta$. As the arguments are smooth convex functions and the valuation under consideration is real-valued, Theorem \[theorem\_compact\_support\_GW\] implies $$\begin{aligned}
GW(\mu)((H+\delta_1h_1)^{\otimes k})-GW(\mu)(H^{\otimes k})=GW(\mu)((H+\delta_1h_1+\delta_2 h_2)^{\otimes k})-GW(\mu)((H+\delta_2h_2)^{\otimes k}),
\end{aligned}$$ i.e. $$\begin{aligned}
& GW(\mu)((1+|x|^2+\sum\limits_{i=1,i\ne 2}^{k}\delta_ih_i)^{\otimes k})-GW(\mu)(((1+|x|^2+\sum\limits_{i=3}^{k}\delta_ih_i)^{\otimes k})\\
=& GW(\mu)((1+|x|^2+\sum\limits_{i=1}^{k}\delta_ih_i)^{\otimes k})-GW(\mu)((1+|x|^2+\sum\limits_{i=2}^{k}\delta_ih_i)^{\otimes k})
\end{aligned}$$ for all $\delta_i$ with $|\delta_i|\le \delta$. Because $GW(\mu)$ is linear, both sides are polynomials in $\delta_1,...,\delta_k$. Consider the coefficient in front of $\delta_1\cdot...\cdot \delta_k$. Clearly the left-hand side is independent of $\delta_2$, so this coefficient vanishes, whereas we obtain $k!GW(\mu)(h_1\otimes...\otimes h_k)$ on the right. Thus $GW(\mu)(h_1\otimes...\otimes h_k)=0$.
A notion of support for dually epi-translation invariant valuations {#section_support_of_a_valuation}
===================================================================
Throughout this section, let $F$ be a locally convex vector space. Motivated by Theorem \[theorem\_support\_GW\_diagonal\] we make the following definition:
\[defition\_support\_valuation\] For $1\le k\le n$ and $\mu\in{\mathrm{VConv}}_k(C;V,F)$ let the support ${\mathrm{supp\ }}\mu\subset V$ be the set $$\begin{aligned}
{\mathrm{supp\ }}\mu:=\bigcap\limits_{A\subset V\text{ closed, } {\mathrm{supp\ }}GW(\mu)\subset \Delta A}A.
\end{aligned}$$ Here $\Delta:V\rightarrow V^k$ is the diagonal embedding. For $\mu\in{\mathrm{VConv}}_0(C;V,F)$, we set ${\mathrm{supp\ }}\mu=\emptyset$. If $\mu=\sum_{i=0}^{n}\mu_i$ is the homogeneous decomposition of $\mu\in {\mathrm{VConv}}(C;V,F)$ we set ${\mathrm{supp\ }}\mu:=\bigcup_{i=0}^n{\mathrm{supp\ }}\mu_i$.
Theorem \[theorem\_compact\_support\_GW\] implies
If $F$ admits a continuous norm, then any $\mu\in{\mathrm{VConv}}(C;V,F)$ has compact support.
Let us justify the terminology:
\[proposition\_support\_convex\_valuation\] The support of $\mu\in{\mathrm{VConv}}(V)$ is minimal (with respect to inclusion) amongst the closed sets $A\subset V$ with the following property: If $f,g\in{\mathrm{Conv}}(V,{\mathbb{R}})$ satisfy $f=g$ on an open neighborhood of $A$, then $\mu(f)=\mu (g)$.
Let us first show that any closed set $A$ satisfying the property contains the support of $\mu$. Using the homogeneous decomposition, we can assume that $\mu$ is $k$-homogeneous.\
We will argue by contradiction. Assume the support was not contained in $A$. Then $ {\mathrm{supp\ }}GW(\mu)\setminus \Delta A\ne \emptyset$. In particular, we find functions $\phi_1,...,\phi_k$ with support in $V\setminus A$ such that $$\begin{aligned}
GW(\mu)(\phi_1\otimes...\otimes \phi_k)\ne0.
\end{aligned}$$Using the Hahn-Banach theorem, we can choose $\lambda\in F'$ with $$\begin{aligned}
\lambda(GW(\mu)(\phi_1\otimes\dots\otimes \phi_k))\ne0.
\end{aligned}$$ Choose an euclidean structure on $V$ and let $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ be given by $f(x):=|x|^2$. Then $f+\sum_{i=1}^{k}\delta_i\phi_i$ is convex for all $\delta_i$ small enough. Let us compare $\lambda(\mu(f))$ and $\lambda(\mu(f+\sum_{i=1}^{k}\delta_i\phi_i))$. By construction, the two functions coincide on an open neighborhood of $A$ and thus the assumption implies $\lambda(\mu(f))=\lambda(\mu(f+\sum_{i=1}^{k}\delta_i\phi_i))$ for all $\delta_i$ small enough. Applying Theorem \[theorem\_compact\_support\_GW\], we see that the right hand side is a polynomial in $\delta_i$ for all $\delta_i$ small enough and that the coefficient in front of $\delta_1\cdot\dots\cdot\delta_k$ is exactly $k!GW(\lambda\circ \mu)(\phi_1\otimes\dots\otimes\phi_k)=k!\lambda(GW(\mu)(\phi_1\otimes...\otimes\phi_k))$. As the left hand side is independent of $\delta_i$, this coefficient has to vanish, so we obtain a contradiction.\
It remains to see that ${\mathrm{supp\ }}\mu$ actually satisfies the property. Again, we can assume that $\mu$ is $k$-homogeneous. As $F$ is locally convex, it is sufficient to show the claim for all valuations $\lambda\circ\mu\in{\mathrm{VConv}}(C;V)$ for $\lambda\in F'$. As this is a real-valued valuation, its support is compact, so under the assumptions above the mollified functions $f_\epsilon,g_\epsilon\in {\mathrm{Conv}}(V,{\mathbb{R}})\cap C^\infty(V)$ satisfy $f_\epsilon=g_\epsilon$ on an open neighborhood of the support of $\lambda\circ\mu$ for all $\epsilon>0$ small enough. In particular, $f_\epsilon^{\otimes k}=g_\epsilon^{\otimes k}$ on a neighborhood of ${\mathrm{supp\ }}GW(\lambda\circ\mu)$ and using Theorem \[theorem\_compact\_support\_GW\] we obtain $$\begin{aligned}
\lambda(\mu(f))=&\lim\limits_{\epsilon\rightarrow0}\lambda(\mu(f_\epsilon))=\lim\limits_{\epsilon\rightarrow0}GW(\lambda\circ\mu)(f^{\otimes k}_\epsilon)=\lim\limits_{\epsilon\rightarrow0}GW(\lambda\circ\mu)(g^{\otimes k}_\epsilon)\\
=&\lim\limits_{\epsilon\rightarrow0}\lambda(\mu(g_\epsilon))=\lambda(\mu(g)).
\end{aligned}$$
As a first application of this notion of support, we will discuss the (non-) existence of invariant real-valued valuations for non-compact groups $G\subset Aff(V)$, where a valuation $\mu\in{\mathrm{VConv}}(V)$ is called $G$-invariant, if $\mu(f\circ g)=\mu(f)$ for all $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ and $g\in G$. We need the following preparatory proposition:
\[proposition\_no\_concentration\_of\_support\_in\_1\_point\] If the support of $\mu\in{\mathrm{VConv}}(C;V,F)$ is contained in a one-point set, then it is empty and $\mu$ is constant.
By considering $\lambda\circ\mu$ for $\lambda\in F'$ again, it is enough to consider the case $F={\mathbb{R}}$. Let us also assume $V={\mathbb{R}}^n$ and without loss of generality, let the support of $\mu$ be contained in $\{0\}$. By taking the homogeneous decomposition of $\mu$, we can assume that $\mu$ is homogeneous of degree $k$. We thus only need to show that the assumptions imply $\mu=0$ for $k>0$.\
If $\mu$ is $1$-homogeneous, $GW(\mu)$ is a distribution with compact support of order at most $2$ due to inequality , so there exist constants $c_\alpha\in{\mathbb{R}}$ such that $$\begin{aligned}
GW(\mu)=\sum\limits_{|\alpha|\le 2}c_\alpha\partial^\alpha\delta_0.
\end{aligned}$$ Plugging in linear and constant functions, we see that $c_\alpha=0$ for $|\alpha|<2$. Thus for any $f\in C^\infty(V)$: $$\begin{aligned}
GW(\mu)(f)=\sum\limits_{|\alpha|=2}c_\alpha\partial^\alpha f(0).
\end{aligned}$$ Fix $1\le i\le n$ and consider the functions $f_\epsilon(x)=\sqrt{\epsilon^2+x_i^2}$ for $\epsilon>0$. Then $$\begin{aligned}
\partial^\alpha f_\epsilon(x)=\begin{cases}
\frac{\epsilon^2}{\sqrt{\epsilon^2+x_i^2}^3} & \alpha=(i,i),\\
0 & \text{else}.
\end{cases}
\end{aligned}$$ Moreover $f_\epsilon\rightarrow|x_i|$, so the continuity of $\mu$ implies $$\begin{aligned}
\mu(|x_i|)=\lim\limits_{\epsilon\rightarrow0}\mu(f_\epsilon)=\lim\limits_{\epsilon\rightarrow0}GW(\mu)(f_\epsilon) =\lim\limits_{\epsilon\rightarrow0}c_{(i,i)} \frac{1}{\epsilon}.
\end{aligned}$$ Thus we must have $c_{(i,i)}=0$. In total we are left with an expression of the form $$\begin{aligned}
GW(\mu)=\sum\limits_{i< j}c_{ij}\partial_i\partial_j\delta_0.
\end{aligned}$$ Now consider $f_\epsilon(x)=\sqrt{\epsilon^2+(x_i+x_j)^2}$ for $i\ne j$, which converges to $|x_i+x_j|$ for $\epsilon\rightarrow0$. Then $\partial_i\partial_jf_\epsilon(x)=\frac{\epsilon^2}{\sqrt{\epsilon^2+(x_i+x_j)^2}^3}$, so the same argument as before shows that $$\begin{aligned}
\mu(|x_i+x_j|)=\lim\limits_{\epsilon\rightarrow0}\mu(f_\epsilon)=\lim\limits_{\epsilon\rightarrow0}GW(\mu)(f_\epsilon)=\lim\limits_{\epsilon\rightarrow0}c_{ij} \frac{1}{\epsilon}.
\end{aligned}$$ Thus $c_{ij}=0$ for all $1\le i,j\le n$, i.e. $GW(\mu)=0$. The injectivity of $GW$ from Corollary \[corollary\_GW\_injective\] implies $\mu=0$.\
If $\mu$ is $k$-homogeneous, we consider the valuation $$\begin{aligned}
\mu_f:=\bar{\mu}(\cdot,f[k-1])
\end{aligned}$$ for $f\in C$ obtained from $\bar{\mu}$ by setting the last $k-1$ arguments equal to $f$. Then $\mu_f$ is a $1$-homogeneous valuation. Using Proposition \[proposition\_support\_convex\_valuation\], it is easy to see that the support of $\mu_f$ is contained in the support of $\mu$, so we deduce $\mu_f=0$ from the case $k=1$. In particular $\mu(f)=\bar{\mu}(f,f[k-1])=\mu_f(f)=0$.
Let $G\subset Aff(V)$ be a subgroup such that either
1. there exists no compact orbit in $V$, or
2. the only compact orbit in $V$ consists of a single point.
Then any $G$-invariant valuation in ${\mathrm{VConv}}(V)$ is constant. In particular, any translation or $\mathrm{SL}(V)$-invariant valuation (for $\dim V\ge 2$) is constant.
Without loss of generality we can assume that $\mu$ is homogeneous of degree $k$ and $G$-invariant. We will show that $\mu$ has to vanish identically if $k>0$.\
Suppose $k>0$. It is easy to see that $GW:{\mathrm{VConv}}_k(V)\rightarrow\mathcal{D}'(V^k)$ is equivariant with respect to the operation of the affine group. In particular, any $G$-invariant valuation induces a $G$-invariant distribution. As the support of any such distribution must be invariant with respect to the group, the same holds true for the support of $\mu$. However, the support of $\mu$ is compact, so we directly see that the support of $\mu$ is either empty or consists of a single point. Due to Proposition \[proposition\_no\_concentration\_of\_support\_in\_1\_point\] the second case cannot occur, so the support of $\mu$ is empty, i.e. $\mu=0$.
Subspaces of valuations with compact support {#section_subspaces_compact_support}
--------------------------------------------
The goal of this section is to establish some useful results on the topology of spaces of valuations with support contained in a fixed (compact) set. An application of these results will be presented in an upcoming work.\
Throughout this section let $F$ be a locally convex vector space and let us assume for simplicity that $V$ carries some euclidean structure.
For $A\subset V$ we denote by ${\mathrm{VConv}}_A(V,F)$ the space of valuations that have support in $A$.
\[lemma\_spaces\_with\_fixed\_support\_closed\_in\_VConv\] If $A\subset V$ is closed, then ${\mathrm{VConv}}_A(V,F)$ is a closed subspace of ${\mathrm{VConv}}(V,F)$.
If $(\mu_\alpha)_\alpha$ is a net in ${\mathrm{VConv}}_A(V,F)$ converging to $\mu$ in ${\mathrm{VConv}}(V,F)$ and $f,h\in{\mathrm{Conv}}(V,{\mathbb{R}})$ are two functions with $f=h$ on a neighborhood of $A$, we deduce $\mu_\alpha(f)=\mu_\alpha(h)$ for all $\alpha$ using Proposition \[proposition\_support\_convex\_valuation\]. Taking the limit, we obtain $\mu(f)=\mu(h)$. As this is true for any $f,h\in{\mathrm{Conv}}(V,{\mathbb{R}})$ with $f=h$ on a neighborhood of $A$, the support of $\mu$ has to be contained in $A$ by Proposition \[proposition\_support\_convex\_valuation\]. Thus ${\mathrm{VConv}}_A(V,F)$ is closed in ${\mathrm{VConv}}(V,F)$.
To illustrate the relevance of these spaces, let us prove Proposition \[mainproposition\_approximation\_using\_sequences\]:
Let us denote the continuous norm by $\|\cdot\|$ and let $U_R:=U_R(0)$ denote the open ball with radius $R>0$ in $V$.\
Using the homogeneous decomposition, we can assume that all valuations are $k$-homogeneous.\
Assume that the supports of the valuations $\mu_j$ are not bounded. Choosing a subsequence if necessary, we can assume that the following holds: There exists a strictly increasing sequence $(r_j)_j$ of positive real numbers converging to $+\infty$ such that
1. ${\mathrm{supp\ }}\mu\subset U_{r_0}$,
2. ${\mathrm{supp\ }}\mu_j\subset U_{r_j}$ for all $j\ge 1$,
3. ${\mathrm{supp\ }}\mu_{j+1}\setminus B_{r_j}\ne \emptyset$ for all $j\ge 1$.
In particular, for every $j\in\mathbb{N}$ we can inductively define functions $\phi^j_1,...,\phi^j_k\in C^\infty_c(V)$ with the properties
1. ${\mathrm{supp\ }}\phi^j_i\subset U_ {r_j}\setminus B_{r_{j-1}}$ for all $j\ge 1$,
2. $\|\sum\limits_{i=1}^{j}GW(\mu_j)(\phi_1^i\otimes...\otimes\phi^i_k)\|\ge1$ for all $j\ge 1$,
as follows: Assume that we have constructed the functions $\phi_i^l$ for all $1\le i\le k$ and $l\le j-1$. If $\|\sum\limits_{i=1}^{j-1}GW(\mu_j)(\phi_1^i\otimes...\otimes\phi^i_k)\|\ge1$, choose $\phi_1^j=...=\phi_k^j=0$.\
If $\|\sum\limits_{i=1}^{j-1}GW(\mu_j)(\phi_1^i\otimes...\otimes\phi^i_k)\|< 1$, choose $\phi_i^j\in C_c^\infty(U_{r_j}\setminus B_{j-1})$ such that $GW(\mu_j)(\phi_1^j\otimes\dots\otimes\phi_k^j)\ne 0$. Then $$\begin{aligned}
\|\sum\limits_{i=1}^{j}GW(\mu_j)(\phi_1^i\otimes...\otimes\phi^i_k)\|
\ge& \|GW(\mu_j)(\phi_1^j\otimes...\otimes\phi^j_k)\|-\|\sum\limits_{i=1}^{j-1}GW(\mu_j)(\phi_1^i\otimes...\otimes\phi^i_k)\|\\
>&\|GW(\mu_j)(\phi_1^j\otimes...\otimes\phi^j_k)\|-1
\end{aligned}$$ Scaling one of the functions $\phi_i^j$ appropriately for $1\le i\le k$, we can make the right hand side equal to $1$.\
In any case, we obtain functions satisfying $\|\sum_{i=1}^{j}GW(\mu_j)(\phi_1^i\otimes...\otimes\phi^i_k)\|\ge 1$ for all $j\ge 1$.\
For $1\le i\le k$ define $\phi_i:=\sum_{j=1}^{\infty}\phi_i^j$. By construction, this is a locally finite sum, so we obtain an element in $C^\infty(V)$. Using Lemma \[lemma\_smooth function difference of convex functions\] we can find $k$ functions $f_i\in{\mathrm{Conv}}(V,{\mathbb{R}})$, $1\le i\le k$, such that $f_i+\sum_{j\in A}\phi_i^j$ is convex for all $A\subset\mathbb{N}$, as the supports of the functions $\phi_i^j$ are pairwise disjoint for each $1\le i\le k$. Set $f_i^j:=f_i+\sum_{i=1}^j\phi_i^j$. Then $(f_i^j)_j$ converges to $f_i+\phi_i$ uniformly on compact subsets, i.e. in ${\mathrm{Conv}}(V,{\mathbb{R}})$. Furthermore, $f_i^j=f_i$ on an open neighborhood of the support of $\mu$, so $\mu(f_i)=\mu(f^j_i)$ for all $j$. Now the polarization $\bar{\mu}$ is a linear combination of $\mu$ evaluated in positive linear combinations of the arguments. In particular, exchanging $f_i$ and $f_i^j$ does not change the value of $\bar{\mu}$. Thus for any $j\in\mathbb{N}$:
$$\begin{aligned}
0=&\|GW(\mu)(0\otimes...\otimes0)\|\\
=&\|\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f_{\sigma(1)},...,f_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})\|\\
=&\|\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})\|,\\
\end{aligned}$$
i.e. $\sum_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})=0$.\
Set $K:=\{f_i^j \ : \ j\in\mathbb{N},1\le i\le k\}\cup\{f_1+\phi_1,...,f_k+\phi_k,f_1,...,f_k\}$. Then $K\subset{\mathrm{Conv}}(V,{\mathbb{R}})$ is compact, so $(\mu_j)_j$ converges to $\mu$ uniformly on $K$. By Lemma \[lemma\_continuity\_valuation->polarization\] the same holds for the polarizations $(\bar{\mu}_j)_j$. In particular, there exists $N\in\mathbb{N}$ such that $$\begin{aligned}
&\|\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}_j(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})\|\\
=&\|\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})\\
&-\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}_j(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})\|<\frac{1}{2}
\end{aligned}$$ for all $j\ge N$. By definition $$\begin{aligned}
\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}_j(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})
= GW(\mu_j)(\sum_{i=1}^j\phi_1^i\otimes...\otimes \sum_{i=1}^j\phi_k^i).
\end{aligned}$$ As $GW(\mu_j)$ has support contained in the diagonal and the functions belonging to different superscripts $i$ have disjoint support, we obtain $$\begin{aligned}
&\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}_j(f^j_{\sigma(1)},...,f^j_{\sigma(l)},f_{\sigma(l+1)},...,f_{\sigma(k)})=\sum\limits_{i=1}^{j}GW(\mu_j)(\phi_1^i\otimes...\otimes \phi_k^i).
\end{aligned}$$ Thus we arrive at $$\begin{aligned}
\|\sum\limits_{i=1}^{j}GW(\mu_j)(\phi_1^i\otimes...\otimes \phi_k^i)\|<\frac{1}{2}
\end{aligned}$$ for all $j\ge N$, which is a contradiction.
In the rest of this section, we introduce special continuous semi-norms on ${\mathrm{VConv}}_A(V,F)$ for compact subsets $A\subset V$. The main goal for the introduction of these semi-norms is to simplify convergence arguments. Together with Proposition \[mainproposition\_approximation\_using\_sequences\] this gives a rather effective set of tools for approximation problems.
\[proposition\_family\_of\_continuous\_norms\_on\_VConv\] Let $A\subset V$ be compact and convex. Let $|\cdot|_F$ denote a continuous semi-norm on $F$ and choose $s>0$. For $\mu\in {\mathrm{VConv}}_{A}(V,F)$ define $$\begin{aligned}
\|\mu\|_{F;A,s}:=\sup\{|\mu(f)| \ : \ f\in{\mathrm{Conv}}(V,{\mathbb{R}}), \|f\|_{C(A+2s B_1)}\le 1 \}.
\end{aligned}$$ This defines a continuous semi-norm on ${\mathrm{VConv}}_{A}(V)$. If $|\cdot|_F$ is a norm, so is $\|\cdot\|_{F;A,s}$. In addition, the topology induced by the family $\|\cdot\|_{F;A,s}$( for all continuous semi-norms $|\cdot|_F$ on $F$) on ${\mathrm{VConv}}_{A}(V,F)$ coincides with the relative topology.
If $\|\mu\|_{F;A,s}$ is finite, it is clear that it defines a semi-norm. Let $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ with $\|f\|_{C(A+2sB_1)}\le 1$ be given. By Proposition \[proposition\_convex\_functions\_local\_lipschitz\_constants\], $f$ is Lipschitz continuous on $B_{A+sB_1}$ with Lipschitz constant $L=\frac{2}{s}\|f|_{A+2sB_1}\|_{\infty}\le \frac{2}{s}$. Consider the following function $$\begin{aligned}
\tilde{f}(x):=\begin{cases}
\sup\limits_{x=\lambda y+(1-\lambda)z,\lambda\ge1}\lambda f(y)+(1-\lambda)f(z) & x\in V\setminus (A+sB_1),\\
f(x) & x\in A+sB_1.
\end{cases}
\end{aligned}$$ By the proof of Theorem 4.1 in [@Yan:Extension_convex_function], $\tilde{f}$ is a finite-valued convex extension of the Lipschitz continuous function $f|_{A+sB_1}$. Now for any $\lambda\ge 1$, $y,z\in A+sB_1$ with $x=\lambda y+(1-\lambda)z$: $$\begin{aligned}
\lambda f(y)+(1-\lambda)f(z)\le&|\lambda[f(y)-f(z)]|+|f(z)|\le \frac{2}{s}\lambda |y-z|+\|f\|_{C(A+sB_1)}\\
\le& \frac{2}{s}|\lambda y-\lambda z|+1=\frac{2}{s}|x-z|+1.
\end{aligned}$$ Thus for $x\in V\setminus (A+sB_1)$ we obtain $$\begin{aligned}
\tilde{f}(x)\le \frac{2}{s} \sup_{z\in A+sB_1}|x-z|+1\le \frac{2}{s} (\mathrm{dist}(x,A+sB_1)+{\mathrm{diam}}(A+sB_1))+1.
\end{aligned}$$ Choosing $\lambda=\frac{|z-x|}{s}$, $y=z+s\frac{x-z}{|z-x|}$ and $z\in A$, we also obtain the inequality $$\begin{aligned}
\frac{|z-x|}{s}f(z+s\frac{x-z}{|z-x|})+(1-\frac{|z-x|}{s})f(z)\le\tilde{f}(x)
\end{aligned}$$ As $$\begin{aligned}
&|\frac{|z-x|}{s}f(z+s\frac{x-z}{|z-x|})+(1-\frac{|z-x|}{s})f(z)|\le \frac{|z-x|}{s}+|(1-\frac{|z-x|}{s})|\\
\le& 2\frac{|z-x|}{s}+1\le \frac{2}{s} (\mathrm{dist}(x,A+sB_1)+{\mathrm{diam}}(A+sB_1))+1,
\end{aligned}$$ $|\tilde{f}(x)|\le \frac{2}{s} (\mathrm{dist}(x,A+sB_1)+{\mathrm{diam}}(A+sB_1))+1$ for all $x\in V$, so the set $K:=\{f\in{\mathrm{Conv}}(V,{\mathbb{R}}) \ : \ f=\tilde{h} \text{ for some }h\in{\mathrm{Conv}}(V,{\mathbb{R}}) \text{ with }\ \|h\|_{C(A+2sB_1)}\le 1\}$ is uniformly bounded on compact subsets and therefore relatively compact in ${\mathrm{Conv}}(V,{\mathbb{R}})$ due to Proposition \[proposition\_compactness\_Conv\]. In particular, $\mu$ is bounded on $C$, as it is continuous.\
For any function $f\in {\mathrm{Conv}}(V,{\mathbb{R}})$ we have $\tilde{f}=f$ on $A+sB_1$, i.e. these functions coincide on an open neighborhood of the support of $\mu$. Proposition \[proposition\_support\_convex\_valuation\] implies $\mu(f)=\mu(\tilde{f})$ and therefore $$\begin{aligned}
\|\mu\|_{F;A,s}=\sup\{\left|\mu(f)\right|_F \ : \ f\in{\mathrm{Conv}}(V,{\mathbb{R}}), \|f\|_{A+sB_1}\le 1 \}=\sup_{\tilde{f}\in K}|\mu(\tilde{f})|_F<\infty.
\end{aligned}$$ In addition, we see that the compact subset $\bar{K}\subset{\mathrm{Conv}}(V,{\mathbb{R}})$ satisfies $$\begin{aligned}
\|\mu\|_{F;K,s}\le \|\mu\|_{F;\bar{K}} \quad \text{for all }\mu\in {\mathrm{VConv}}_{A}(V,F).
\end{aligned}$$ On the other hand, any $f\in \bar{K}$ satisfies $\|f\|_{C(A+2sB_1)}\le \sup_{x\in A+2sB_1}\frac{2}{s} (\mathrm{dist}(x,A+sB_1)+{\mathrm{diam}}(A+sB_1))+1\le c_{A,s}:=\frac{2}{s} ({\mathrm{diam}}(A)+3s)+1=\frac{2}{s} {\mathrm{diam}}(A)+4$. By considering the $k$-homogeneous component $\mu_k$ of $\mu$, we obtain $$\begin{aligned}
\|\mu_k\|_{F;\bar{K}}=\sup_{f\in\bar{K}}\left|\mu_k(f)\right|=c_{A,s}^k\sup_{f\in\bar{K}}\left|\mu_k\left(\frac{f}{c_{A,s}}\right)\right| \le c_{A,s}^k\|\mu_k\|_{F;A,s}.
\end{aligned}$$ Thus $\|\cdot\|_{F;A,s}$ is continuous on ${\mathrm{VConv}}_A(C;V,F)$.\
More generally, any compact set $D\subset {\mathrm{Conv}}(V,{\mathbb{R}})$ satisfies $t:=\sup_{f\in D,x\in A+2sB_1}|f(x)|<\infty$. Assuming $t>0$ this implies $$\begin{aligned}
\|\mu_k\|_{F;D}=\sup_{f\in D}\left|\mu_k(f)\right|_F=t^k\sup_{f\in D}\left|\mu_k\left(\frac{f}{t}\right)\right|_F \le t^k\|\mu_k\|_{F;A,s}.
\end{aligned}$$ If $t=0$, then any $f\in D$ coincides with the zero function on a neighborhood of the support of $\mu$, so $\mu_k(f)=\mu(0)$ for all $f\in D$ due to Proposition \[proposition\_support\_convex\_valuation\], i.e. $\|\mu_k\|_{F;D}\le \|\mu_k\|_{F;A,s}$.\
In any case, we see that $\|\cdot\|_{F;A,s}$ defines a continuous semi-norm on ${\mathrm{VConv}}_{A}(V,F)$ and that the family of these semi-norms generates the subspace topology.\
Now let us assume that $|\cdot|_F$ is a norm. Repeating the argument above for $D=\{f\}$, we see that $\|\mu\|_{F;D}>0$ for $\mu\in{\mathrm{VConv}}_A(V,F)$ implies $\|\mu\|_{F;A,k}>0$. Thus $\|\cdot\|_{F;A,k}$ is indeed a norm.
Let us observe that the semi-norms constructed from $|\cdot|_F$ are equivalent for different values of $s$:
Let $A\subset V$ be a compact convex subset. For $0<s<t$ we have $$\begin{aligned}
\|\mu\|_{F;A,t}\le\|\mu\|_{F;A,s}\le (\frac{2}{s} (2t+{\mathrm{diam}}A)+1)^k\|\mu\|_{F;A,t}
\end{aligned}$$ for all $k$-homogeneous $\mu\in{\mathrm{VConv}}_A(V,F)$.
The first inequality is obvious. For the second inequality, let $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ be a function with $\|f\|_{C(A+2sB_1)}\le 1$. Considering the function $\tilde{f}\in{\mathrm{Conv}}(V,{\mathbb{R}})$ $$\begin{aligned}
\tilde{f}(x):=\begin{cases}
\sup\limits_{x=\lambda y+(1-\lambda)z,\lambda\ge1}\lambda f(y)+(1-\lambda)f(z) & x\in V\setminus (A+sB_1),\\
f(x) & x\in A+sB_1
\end{cases}
\end{aligned}$$ from the previous proof, we see that $|\tilde{f}(x)|\le \frac{2}{s} (\mathrm{dist}(x,A+sB_1)+{\mathrm{diam}}(A+sB_1))+1$, so $\|\tilde{f}\|_{C(A+2tB_1)}\le \frac{2}{s} (2t-s +{\mathrm{diam}}A+s)+1=\frac{2}{s} (2t+{\mathrm{diam}}A)+1$. As $f=\tilde{f}$ on a neighborhood of the support of $\mu$, we see that $$\begin{aligned}
|\mu(f)|_F=(\frac{2}{s} (2t+{\mathrm{diam}}A)+1)^k|\mu\left(\frac{f}{\frac{2}{s} (2t+{\mathrm{diam}}A)+1}\right)\le (\frac{2}{s} (2t+{\mathrm{diam}}A)+1)^k\|\mu\|_{F;A,s}.
\end{aligned}$$ The claim follows.
\[corollary\_Frechet-topology\_for\_compactly\_supported\_valuations\] If $A$ is compact and $F$ is a Banach or Fréchet space, then ${\mathrm{VConv}}_A(V,F)$ is also a Banach or Fréchet space, respectively.
By Lemma \[lemma\_spaces\_with\_fixed\_support\_closed\_in\_VConv\], ${\mathrm{VConv}}_A(V,F)$ is a closed subspace of the complete locally convex space ${\mathrm{VConv}}(V,F)$ and so it is also complete.\
If $A$ is compact and convex, we can take one of the semi-norms from Proposition \[proposition\_family\_of\_continuous\_norms\_on\_VConv\], which generates the subspace topology, so the space ${\mathrm{VConv}}_A(V,F)$ is complete with respect to this semi-norms. If $F$ is a Banach space, we only obtain one norm, while we get a sequence of norms if $F$ is a Fréchet space. In both cases, the claim follows\
If $A$ is not convex, choose $R>0$ such that $A\subset B_R(0)$. Using the same argument as in Lemma \[lemma\_spaces\_with\_fixed\_support\_closed\_in\_VConv\], we see that ${\mathrm{VConv}}_A(V,F)\subset{\mathrm{VConv}}_{B_R(0)}(V,F)$ is a closed subspace of a Banach or Fréchet space. The claim follows.
Vertical support of valuations on convex bodies and the image of the embedding $T:{\mathrm{VConv}}(V,F)\rightarrow {\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ {#section_vertical_support_and_image_embedding}
-----------------------------------------------------------------------------------------------------------------------------------------------------------
Similar to the definition of support of valuations in ${\mathrm{VConv}}(C;V,F)$, we will give a notion of *vertical support* for elements of ${\mathrm{Val}}(V)$. Starting point is the Goodey-Weil embedding for translation invariant valuations on convex bodies. Consider the space $\mathbb{P}_+(V^*)$ of oriented lines in $V^*$ and the line bundle $L$ over $\mathbb{P}_+(V^*)$ with fiber over $l\in\mathbb{P}_+(V^*)$ given by $$\begin{aligned}
P_l:=\{h:l^+\rightarrow{\mathbb{R}}\ 1\text{-homogeneous}\}.
\end{aligned}$$ Note that every support function defines a continuous section of $L$. For $y\in V^*\setminus\{0\}$ we will write $[y]$ for the corresponding oriented line in $\mathbb{P}_+(V^*)$.
Let $F$ be a locally convex vector space. For every $\mu\in{\mathrm{Val}}_k(V,F)$ there exists a unique distribution $GW(\mu)\in \mathcal{D}'(\mathbb{P}_+(V^*)^k,L^{\boxtimes k}),\bar{F})$ such that $$\begin{aligned}
\mu(h_{K_1}\otimes\dots\otimes h_{K_k})=\bar{\mu}(K_1,\dots,K_k).
\end{aligned}$$ Here $\bar{\mu}$ denotes the polarization of $\mu\in{\mathrm{Val}}_k(V,F)$. Furthermore, the support of this distribution is contained in the diagonal.
In [@Goodey_Weil:Distributions_and_valuations], the existence of such a distribution was shown in the case $F={\mathbb{R}}$. The same construction can be done for arbitrary locally convex vector spaces $F$, similar to our construction of the Goodey-Weil embedding for ${\mathrm{VConv}}(V,F)$. The diagonality of the support was first shown in [@Alesker:McMullenconjecture] for real-valued valuations, but as in the proof of Theorem \[theorem\_support\_GW\_diagonal\], this implies the more general statement.
Following the approach in the previous section, we define the *vertical support*:
For $1\le k\le n$, we define the vertical support of $\mu\in{\mathrm{Val}}_k(V,F)$ to be the set $$\begin{aligned}
{\mathrm{v}\text{-}{supp }}\mu:=\bigcap\limits_{{\mathrm{supp\ }}GW(\mu)\subset \Delta A,\ A\subset \mathbb{P}_+(V^*)\text{ compact}} A.
\end{aligned}$$ For $k=0$, we set ${\mathrm{v}\text{-}{supp }}\mu=\emptyset$. If $\mu=\sum_{i=0}^{n}\mu_i$ is the homogeneous decomposition, we set ${\mathrm{v}\text{-}{supp }}\mu:=\cup_{i=0}^n{\mathrm{v}\text{-}{supp }}\mu_i$.
As before, the vertical support can be characterized without a reference to the Goodey-Weil embedding.
\[lemma\_property\_vertical\_support\] If $K,L\in\mathcal{K}(V)$ are two convex bodies with $h_K=h_L$ on an open neighborhood of ${\mathrm{v}\text{-}{supp }}\mu$, then $\mu(K)=\mu(L)$.
Using the homogeneous decomposition, we can assume that $\mu$ is $k$-homogeneous.\
Let us choose a metric on $V$ to identify $\mathbb{P}_+(V^*)\cong S_V$, which also trivializes $L$. Take a sequence of positive functions $\phi_j\in C^\infty(S_V)$ with $\int_{SO(V)}\phi_j(g)dg=1$, such that the diameter of the support of $\phi_j$ converges to zero for $j\rightarrow\infty$. It is easy to see that $f_j(v):=\int_{SO(n)}\phi_j(g)f(g^{-1}v)dg$ defines a sequence of smooth functions on $S_V$ for every $f\in C(S_V)$, that converges uniformly to $f$. Moreover, if $f=h_K$, then $f_j$ is the restriction of a support function of some convex body $K_j$. The uniform convergence $h_{K_j}\rightarrow h_K$ on $S_V$ implies that $K_j\rightarrow K$ in the Hausdorff metric. Similarly, we obtain convex bodies $L_j$ from $h_L$.\
Note that $f_j(v)$ only depends on the values of $f$ in a neighborhood of $v$ depending on the diameter of the support of $\phi_j$. As the diameter of $\phi_j$ converges to zero and $h_K=h_L$ on a neighborhood of $A$, we see that there exists $N\in\mathbb{N}$ such that $h_{K_j}=h_{L_j}$ on a neighborhood of $A$ for all $j\ge N$. Thus $h_{K_j}^{\otimes k}=h_{L_j}^{\otimes k}$ on a neighborhood of the support of $GW(\mu)$ and we obtain $$\begin{aligned}
\mu(K)=\lim\limits_{j\rightarrow\infty}\mu(K_j)=\lim\limits_{j\rightarrow\infty}GW(\mu)(h_{K_j}^{\otimes k})=\lim\limits_{j\rightarrow\infty}GW(\mu)(h_{L_j}^{\otimes k})=\lim\limits_{j\rightarrow\infty}\mu(L_j)=\mu(L).
\end{aligned}$$
\[proposition\_characterization\_support\_Goodey\_Weil\_for\_Val\] Let $\mu\in {\mathrm{Val}}(V)$ and $A\subset \mathbb{P}_+(V^*)$ be a compact subset with the following property: If $K,L\in\mathcal{K}(V)$ are two convex functions with $h_K=h_L$ on an open neighborhood of $A$, then $\mu(K)=\mu (L)$. Then the vertical support of $\mu$ is contained in $A$.
Again let us choose a metric on $V$ and identify $\mathbb{P}_+(V^*)\cong S_V$, trivializing $L$.\
Using the homogeneous decomposition, we can assume that $\mu$ is $k$-homogeneous. Now assume that the claim was false. Then we would find functions $\phi_1,...,\phi_k\in C^\infty(S_V)$ with support contained in $S_V\setminus A$ such that $GW(\mu)(\phi_1\otimes...\otimes\phi_k)=1$. Consider the function $1+\sum_{i=1}^{k}\delta_i\phi_i$ on $S_V$. For small $\delta_i>0$, it is the support function of a convex body $K_\delta$ and by definition, $h_{K_\delta}=1=h_B$ on a neighborhood of $A$, where $B$ is the unit ball in $V$, so by assumption, $\mu(K_\delta)=\mu(B)$. Note that $\mu(K_\delta)$ is a polynomial in $\delta_i>0$. The coefficient before $\delta_1...\delta_k$ is exactly $k!GW(\mu)(\phi_1\otimes....\otimes \phi_k)=k!$, while the right side does not depend on $\delta_i>0$, i.e. the coefficient has to vanish. Thus we obtain a contradiction.
For $A\subset \mathbb{P}_+(V^*)$ let ${\mathrm{Val}}_{A}(V)$ denote the subspace of valuations with vertical support contained in $A$.
\[corollary\_Val\_subspaces\_compact\_support\_are\_Banach\] Let $A\subset \mathbb{P}_+(V^*)$ be closed. Then ${\mathrm{Val}}_{A}(V,F)$ is closed in ${\mathrm{Val}}(V,F)$.
As in Lemma \[lemma\_spaces\_with\_fixed\_support\_closed\_in\_VConv\].
We are now able to describe the image of $T:{\mathrm{VConv}}(V,F)\rightarrow{\mathrm{Val}}(V^*\times{\mathbb{R}},F)$ in the case, that $F$ admits a continuous norm. Note that by Theorem \[maintheorem\_GW\], all valuations $\mu\in{\mathrm{VConv}}(V,F)$ have compact support in this case. We start with the following observation:
\[proposition\_compatibility\_supports\_under\_embedding\_VCONV-VAL\] For $\mu\in{\mathrm{VConv}}(V)$, ${\mathrm{v}\text{-}{supp }}(T(\mu))\subset P({\mathrm{supp\ }}\mu)$, where $$\begin{aligned}
P:V\rightarrow& \mathbb{P}_+(V\times{\mathbb{R}})\\
v\mapsto& [(v,-1)].
\end{aligned}$$
By Proposition \[proposition\_characterization\_support\_Goodey\_Weil\_for\_Val\], we only need to show that $T(\mu)[K]=T(\mu)[L]$ whenever $h_K$ and $h_L$ coincide on an open neighborhood of $P({\mathrm{supp\ }}\mu)$. Considering $h_K$ and $h_L$ as $1$-homogeneous functions on $V\times{\mathbb{R}}$, the equality $h_K=h_L$ on an open neighborhood $U$ of $P({\mathrm{supp\ }}\mu)$ implies that they coincide on the open set $\pi^{-1}(U)\subset V\times{\mathbb{R}}$, where $\pi:(V\times{\mathbb{R}})\setminus\{0\}\rightarrow \mathbb{P}_+(V\times{\mathbb{R}})$ is the natural projection. Obviously, this is an open neighborhood of ${\mathrm{supp\ }}\mu\times\{-1\}$, so we can apply Proposition \[proposition\_support\_convex\_valuation\], to obtain $\mu(h_K(\cdot,-1))=\mu(h_L(\cdot,-1))$, i.e. $T(\mu)(K)=T(\mu)(L)$. The claim follows.
\[theorem\_image\_VCONV->Val\] The image of $T:{\mathrm{VConv}}_k(V,F)\rightarrow{\mathrm{Val}}_k(V^*\times{\mathbb{R}},F)$ consists precisely of the valuations $\mu\in{\mathrm{Val}}_k(V^*\times{\mathbb{R}},F)$ that have vertical support contained in the lower half sphere $\mathbb{P}_+(V\times{\mathbb{R}})^-:=\{[(y,s)]\in \mathbb{P}_+(V\times{\mathbb{R}}):s<0\}$. If $F$ is a Fréchet space, $T:{\mathrm{VConv}}_A(V,F)\rightarrow{\mathrm{Val}}_{P(A)}(V^*\times{\mathbb{R}},F)$ is a topological isomorphism for any compact subset $K\subset V$.
Starting with $\mu\in{\mathrm{VConv}}_k(V,F)$, Proposition \[proposition\_compatibility\_supports\_under\_embedding\_VCONV-VAL\] shows that $T(\mu)$ has vertical support contained in $\mathbb{P}_+(V\times{\mathbb{R}})^-$.\
Conversely, let $\nu\in{\mathrm{Val}}_k(V^*\times{\mathbb{R}})$ be a valuation with vertical support contained in $\mathbb{P}_+(V\times{\mathbb{R}})^-$. As $P:V\rightarrow \mathbb{P}_+(V\times{\mathbb{R}})^-$ is a diffeomorphism, $P^{-1}({\mathrm{v}\text{-}{supp }}\nu)$ is compact.\
Let us construct a functional $\mu$ on ${\mathrm{Conv}}(V,{\mathbb{R}})$ as follows: Given $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$, let $K_f\in\mathcal{K}(V^*\times{\mathbb{R}})$ be a convex body with $h_{K_f}(\cdot,-1)=f$ on some neighborhood of $P^{-1}({\mathrm{v}\text{-}{supp }}\nu)$ (which exists by Proposition \[proposition\_epi\_graph\_support\_function\_convex\_body\]). Now set $$\begin{aligned}
\mu(f):=\nu(K_f).
\end{aligned}$$ Note that this does not depend on a special choice of $K_f$: If $K$ is another convex body with $h_K(\cdot,-1)=f$ on some neighborhood of $P^{-1}({\mathrm{v}\text{-}{supp }}\nu)$, then $h_K=f=h_{K_f}$ on a neighborhood of ${\mathrm{v}\text{-}{supp }}\nu$, so Lemma \[lemma\_property\_vertical\_support\] implies $\nu(K)=\nu(K_f)$.\
The functional constructed this way is also a valuation: Choose a metric on $V$ and let $N\in\mathbb{N}$ be such that $P^{-1}({\mathrm{v}\text{-}{supp }}\nu)$ is contained in $B_N$. If $\min(f,h)$ is convex, then $$\begin{aligned}
& {\mathrm{epi}}\max(f,h)^*={\mathrm{epi}}\min(f^*,h^*)={\mathrm{epi}}(f^*)\cup {\mathrm{epi}}(h^*),\\
& {\mathrm{epi}}\min(f,h)^*={\mathrm{epi}}\max(f^*,h^*)={\mathrm{epi}}(f^*)\cap {\mathrm{epi}}(h^*).
\end{aligned}$$ For $c=\max \{||f||_{C(B_{N+2})},||h||_{C(B_{N+2})},||\min \{f,h\}||_{C(B_{N+2})},||\max\{f,h\}||_{C(B_{N+2})}\}$ choose $$\begin{aligned}
K_f&={\mathrm{epi}}(f^*)\cap \{|y|\le 2(n+2)c,|t|\le 3(n+2)c\},\\
K_h&={\mathrm{epi}}(h^*)\cap \{|y|\le 2(n+2)c,|t|\le 3(n+2)c\}.
\end{aligned}$$ Then Proposition \[proposition\_epi\_graph\_support\_function\_convex\_body\] implies $$\begin{aligned}
&\max(f,h)= h_{K_f\cup K_h}(\cdot,-1), &&\min(f,h)=h_{K_f\cap K_h}(\cdot,-1)\quad \text{on } B_{N+1}.
\end{aligned}$$ Thus by the definition of $\mu$, $$\begin{aligned}
\mu(\max(f,h))+\mu(\min(f,h))=\nu(K_f\cup K_h)+\nu(K_f\cap K_h)=\nu(K_f)+\nu(K_h)=\mu(f)+\mu(h).
\end{aligned}$$ Furthermore, $\mu$ is obviously invariant under the addition of linear or constant functions. It remains to show that $\mu$ is continuous. We will argue by contradiction.\
Let $(f_j)_j$ be a sequence in ${\mathrm{Conv}}(V,{\mathbb{R}})$ converging to $f\in{\mathrm{Conv}}(V,{\mathbb{R}})$ uniformly on compact subsets and assume that there exists $\epsilon>0$ such that $|\mu(f_j)-\mu(f)|>\epsilon$ for all $j\in\mathbb{N}$ for some continuous semi-norm $|\cdot|$ on $F$. Recall that we have chosen $N\in\mathbb{N}$ such that $P^{-1}({\mathrm{v}\text{-}{supp }}\nu)\subset B_N$. As the set $\{f_j|j\in\mathbb{N}\}\cup\{f\}$ is compact, the functions are bounded on $B_{N+2}$ by some constant $c>0$. Using Proposition \[proposition\_epi\_graph\_support\_function\_convex\_body\], we see that if we choose $$\begin{aligned}
K_{f_j}&={\mathrm{epi}}(f_j^*)\cap \{|y|\le 2(N+2)c,|t|\le 3(N+2)c\},\\
K_f&={\mathrm{epi}}(f^*)\cap \{|y|\le 2(N+2)c,|t|\le 3(N+2)c\},
\end{aligned}$$ then $h_{K_{f_j}}=f_j$ and $h_{K_f}=f$ on $B_{N+1}$. By construction, the sequence $(K_{f_j})_j$ of convex bodies is bounded, so by the Blaschke selection theorem we find a subsequence $K_{f_{j_k}}$ converging to some convex body $K\in\mathcal{K}(V^*\times{\mathbb{R}})$. Then $h_K(\cdot,-1)=h_{K_f}(\cdot,-1)$ on $B_{N+1}$, as $h_{K_{f_{j_k}}}(\cdot,-1)=f_{j_k}$ on $B_{N+1}$ and $f_{j}\rightarrow f$. As $\mu(f)$ does not depend on the special choice of the convex body, we deduce that $$\begin{aligned}
\lim\limits_{k\rightarrow\infty}\mu(f_{j_k})=\lim\limits_{k\rightarrow\infty}\nu(K_{f_{j_k}})=\nu(K)=\nu(K_f)=\mu(f).
\end{aligned}$$ This is a contradiction to $|\mu(f_j)-\mu(f)|>\epsilon$ for all $j\in\mathbb{N}$. Thus $\mu$ has to be continuous.\
We have constructed $\mu\in{\mathrm{VConv}}(V,F)$ with $T(\mu)=\nu$ and the support of $\mu$ is contained in $P^{-1}({\mathrm{v}\text{-}{supp }}\nu)$ by construction.\
Now let $A\subset V$ be compact, $F$ a Fréchet space. Observe that the restriction $T:{\mathrm{VConv}}_A(V,F)\rightarrow{\mathrm{Val}}_{P(A)}(V^*\times{\mathbb{R}})$ is a well defined, injective, and continuous map between Fréchet spaces by Corollary \[corollary\_Frechet-topology\_for\_compactly\_supported\_valuations\] and Theorem \[theorem\_embedding VConv->Val(VxR)\]. By the preceding discussion it is also surjective, so Banach’s inversion theorem implies that $T^{-1}:{\mathrm{Val}}_{P(A)}(V^*\times{\mathbb{R}},F)\rightarrow{\mathrm{VConv}}_A(V,F)$ is continuous, i.e. $T:{\mathrm{VConv}}_A(V,F)\rightarrow{\mathrm{Val}}_{P(A)}(V^*\times{\mathbb{R}},F)$ is a topological isomorphism.
Valuations on special cones of convex functions
===============================================
Restrictions on the support {#section_description_valuations_on_cones_with_support}
---------------------------
In this section we are going to relate valuations on a cone $C\subset {\mathrm{Conv}}(V)$ to valuations on finite-valued convex functions that satisfy certain restrictions on their support. As usual, let $F$ denote a locally convex vector space.
\[maintheorem\_support\_valuations\_non-finite\_functions\] Let $C\subset{\mathrm{Conv}}(V)$ be a cone containing ${\mathrm{Conv}}(V,{\mathbb{R}})$. Consider the sets $B_C:=\bigcap\limits_{f\in C}\overline{{\mathrm{dom }}f}$, $U_C:=\overset{\circ}{B_C}$. Then the following holds:
1. The support of any valuation in ${\mathrm{VConv}}(C;V,F)$ is contained in $B_C$.
2. If $F$ admits a continuous norm, then every valuation in ${\mathrm{VConv}}(V,F)$ with support contained in $U_C$ extends uniquely to an element of ${\mathrm{VConv}}(C;V,F)$.
If $F$ admits a continuous norm, we thus have inclusions $$\begin{aligned}
{\mathrm{VConv}}_{U_C}(V,F)\hookrightarrow{\mathrm{VConv}}(C;V,F)\hookrightarrow {\mathrm{VConv}}_{B_C}(V,F).
\end{aligned}$$
For the first statement, consider the Goodey-Weil distribution of $\mu$ and let $\phi_1,\dots,\phi_k\in C^\infty_c(V\setminus B_C)$. We need to show $GW(\mu)(\phi_1\otimes \dots\otimes \phi_k)=0$. Using a partition of unity, we can assume that ${\mathrm{supp\ }}\phi_i\subset U_\epsilon(x_i)$ for some $x_i\in V\setminus B_C$ and that $B_{\epsilon}(x_i)\subset V\setminus B_C$. We claim that every point $y\in B_\epsilon(x_i)$ has a neighborhood where some $f_{x_i}\in C$ is identical to $\infty$. Indeed, if $y\in B_\epsilon(x_i)$ is a point where the assertion is violated, then $y\in \overline{{\mathrm{dom }}f}$ for all $f\in C$. Thus $y\in B_C$, which is a contradiction to $y\in B_{\epsilon}(x_i)\subset V\setminus B_C$. As $B_C$ is compact, we can thus find a finite number of functions $f_{1,i},...,f_{j,i}\in C$, such that $f_i:=\sum_{l=1}^{j}f_{l,i}=\infty$ on $B_\epsilon(x_i)$.\
The Lipschitz regularization $f_{i,r}:={\mathrm{reg}}_rf_i$ belongs to ${\mathrm{Conv}}(V,{\mathbb{R}})$ for all $r>0$ small enough. Let $h_i\in{\mathrm{Conv}}(V,)$ be a convex function such that $h_i+\phi_i\in{\mathrm{Conv}}(V,{\mathbb{R}})$. Then $\tilde{h}_{i,r}:=f_{i,r}+h_i\in C$ satisfies $\tilde{h}_{i,r}+\phi_i\in C$ as well, so $$\begin{aligned}
&GW(\mu)(\phi_1\otimes\dots\phi_k)\\
=&\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(\tilde{h}_{\sigma(1),r}+\phi_{\sigma(1)},...,\tilde{h}_{\sigma(l),r}+\phi_{\sigma(l)},\tilde{h}_{\sigma(l+1),r},...,\tilde{h}_{\sigma(k),r})
\end{aligned}$$ for all $r>0$ sufficiently small. Of course, $\tilde{h}_{i,r}$ epi-converges to $f_i+h_i$ for $r\rightarrow0$ and $\tilde{h}_{i,r}+\phi_i$ epi-converges to $f_i+h_i+\phi_i=f_i+h_i$, as $f_i=\infty$ on the support of $\phi_i$. The joint continuity of $\bar{\mu}$ implies $$\begin{aligned}
GW(\mu)(\phi_1\otimes\dots\phi_k)=& \sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f_{\sigma(1)}+h_{\sigma(1)},...,f_{\sigma(k)}+h_{\sigma(k)})\\
=&\bar{\mu}(f_1+h_1,...,f_k+h_k)\sum\limits_{l=0}^k(-1)^{k-l}\binom{k}{l}=0.
\end{aligned}$$\
For the second statement, let $\mu\in{\mathrm{VConv}}_k(V)$ be a valuation with support in $U_C$. If $f\in C$ is any function, it is finite and thus continuous on $U_C$. In particular, it is bounded on a compact neighborhood $A\subset U_C$ of the support of $\mu$ by definition of $U_C$. Taking a smaller neighborhood $U$ of the support of $\mu$ with $\bar U\subset \overset{\circ}{A}$, Proposition \[proposition\_convex\_functions\_local\_lipschitz\_constants\] implies that $f$ is Lipschitz continuous on $U$. In particular, any subgradient of $f$ on $U$ has norm bounded by the Lipschitz constant. Proposition \[proposition\_properties\_LIpschitz\_regularization\] iv. implies that there exists $r_0>0$ such that ${\mathrm{reg}}_rf=f$ on $U$ for all $0<r\le r_0$. By Proposition \[proposition\_support\_convex\_valuation\] this shows that $\mu({\mathrm{reg}}_rf)$ does not depend on $0<r\le r_0$ and thus $$\begin{aligned}
\mu'(f):=\lim\limits_{r\rightarrow 0}\mu({\mathrm{reg}}_rf)
\end{aligned}$$ defines an extension of $\mu$ to $C$. Due to Proposition \[proposition\_properties\_LIpschitz\_regularization\] v., it is a valuation. We need to show that this extension is continuous. As the topology on $C$ is metrizable, we only need to show that $\mu'$ is sequentially continuous. Let $(f_j)_j$ be a sequence in $C$ epi-converging to $f\in C$. Then all functions are finite on $U_C$ and thus they converge uniformly on the compact set $A\subset U_C$. Now the estimate in Proposition \[proposition\_convex\_functions\_local\_lipschitz\_constants\] shows that $\{f_j \ : \ j\in\mathbb{N}\}\cup \{f\}$ is uniformly Lipschitz continuous on $U$, so Proposition \[proposition\_properties\_LIpschitz\_regularization\] iii. implies that there exists $r_0>0$ such that ${\mathrm{reg}}_rf_j=f_j$ and ${\mathrm{reg}}_rf=f$ on $U$ for all $0<r\le r_0$ independent of $j\in\mathbb{N}$. In particular, using Proposition \[proposition\_support\_convex\_valuation\] we see that there exists $r_0>0$ such that $\mu({\mathrm{reg}}_r f)$ and $\mu({\mathrm{reg}}_r f_j)$ do not depend on $0<r\le r_0$. As ${\mathrm{reg}}_r f_j\rightarrow {\mathrm{reg}}_r f$ for $j\rightarrow\infty$ and all $r$ sufficiently small, we obtain $$\begin{aligned}
\mu({\mathrm{reg}}_rf)=\lim\limits_{j\rightarrow\infty}\mu({\mathrm{reg}}_rf_j)\quad \text{for all $r$ sufficiently small}.
\end{aligned}$$ However, $\mu({\mathrm{reg}}_r f_j)$ and $\mu({\mathrm{reg}}_r f)$ are constant in $r$ for $0<r\le r_0$ independent of $j\in\mathbb{N}$, so we conclude $\mu'(f)=\lim\limits_{j\rightarrow\infty}\mu'(f_j)$.\
Obviously, the inclusion constructed this way is injective.\
Let us show that the both inclusions in Theorem \[maintheorem\_support\_valuations\_non-finite\_functions\] are strict in general:\
Define $\mu(f):=f(0)+f(2)-2f(1)$ for $f\in{\mathrm{Conv}}({\mathbb{R}},{\mathbb{R}})$. It is easy to see that $\mu$ is a dually epi-translation invariant valuation with support contained in $\{0,1,2\}$.\
For the first inclusion, let $C$ be the cone generated by ${\mathrm{Conv}}({\mathbb{R}},{\mathbb{R}})$ and the *convex indicators* $I^\infty_{[-\frac{1}{n},\infty)}$ for all $n\in\mathbb{N}$, where the convex indicator of a closed convex set $K\subset {\mathbb{R}}$ is defined by $$\begin{aligned}
I^\infty_K(x):=\begin{cases}
0 & x\in K\\
\infty & x\notin K.
\end{cases}
\end{aligned}$$ Then $U_C=(0,\infty)$ and any $f\in C$ satisfies ${\mathrm{supp\ }}\mu\subset\overset{\circ}{{\mathrm{dom }}(f)}$. Now let $(f_j)_j$ be a sequence in $C$ epi-converging to $f$. Due to Proposition \[proposition\_convergence\_finite\_convex\_functions\], the sequence converges locally uniformly on the interior of ${\mathrm{dom }}f$, so in particular on $\{0,1,2\}$, i.e. $\mu(f_j)=f_j(0)+f_j(2)-2f_j(1)\rightarrow f(0)+f(2)-2f(1)$. Thus we can extend $\mu$ continuously to $C$ by setting $\mu(f):=f(0)+f(2)-2f(1)$ for $f\in C$.\
For the second inclusion, let $C\subset{\mathrm{Conv}}({\mathbb{R}})$ be the cone generated by ${\mathrm{Conv}}({\mathbb{R}},{\mathbb{R}})$ and the convex indicator $I^\infty_{[0,\infty)}$. Consider the sequence $(f_j)_j$ in $C$ given by $$\begin{aligned}
f_j(x)=j^2\max(\frac{1}{j}-x,0)= \begin{cases}
j-j^2x & x\le\frac{1}{j},\\
0 & x>\frac{1}{j}.
\end{cases}
\end{aligned}$$ Using Proposition \[proposition\_convergence\_finite\_convex\_functions\] again, we see that $(f_j)_j$ epi-converges to $I^\infty_{[0,\infty)}$, but $\mu(f_j)=j$ for all $j\in\mathbb{N}$, so $\mu$ does not extend to $C$ by continuity.\
The restrictions on the support apply in particular to cones that are invariant under large subgroups of the affine group.
\[corollary\_existence\_non\_trivial\_valuations\] Let $C\subset{\mathrm{Conv}}(V)$ be a cone containing ${\mathrm{Conv}}(V,{\mathbb{R}})$ such that $C$ is invariant with respect to translations or $\mathrm{SL}(V)$ for $\dim V\ge 2$. If $C$ contains a non-finite convex function, then any dually epi-translation invariant valuation is constant.
In both cases $B_C$ is either empty or contains only the origin. But for any $1\le k\le n$ there are no non-trivial valuations with this support due to Proposition \[proposition\_no\_concentration\_of\_support\_in\_1\_point\]. Thus the only valuations are the constant valuations.
Let us see that in special cases we have an equality for the first inclusion in Theorem \[maintheorem\_support\_valuations\_non-finite\_functions\].
\[proposition\_support\_valuations\_on\_open\_set\] Let $U\subset V$ be an open convex set, $C_U:=\{f\in{\mathrm{Conv}}(V) \ : \ f|_U<\infty\}$, and $F$ a locally convex vector space. Then the support of any valuation $\mu\in {\mathrm{VConv}}(C_U;V,F)$ is contained in $U$
This is trivial for $U=V$, thus let us assume $U\ne V$. Due to Theorem \[maintheorem\_support\_valuations\_non-finite\_functions\], it is enough to show that the support of any valuation $\mu\in{\mathrm{VConv}}(C_U;V,F)$ does not contain any point $x_0\in\partial U$. By considering $\lambda\circ\mu$ for all $\lambda\in F'$, it is also sufficient to consider real-valued valuations. Let us assume that $\mu\in{\mathrm{VConv}}(C_U;V,F)$ is $k$-homogeneous for $1\le k\le n$ and that $x_0\in{\mathrm{supp\ }}\mu\cap\partial U$. Let us identify $V\cong{\mathbb{R}}^n$. By taking a supporting hyperplane through $x_0$ and using translations as well as rotations, we can assume that $x_0=0$ and that $\bar{U}\subset [0,\infty)\times{\mathbb{R}}^{n-1}$.\
As $0\in{\mathrm{supp\ }}\mu$, we can choose functions $\phi_1^j,...,\phi_k^j\in C^\infty_c({\mathbb{R}}^n)$ with ${\mathrm{supp\ }}\phi_i^j\subset U_{\frac{1}{j}}(0)$ such that $$\begin{aligned}
GW(\mu)(\phi_1^j\otimes...\otimes\phi_k^j)=1 \quad\forall j\in\mathbb{N}.
\end{aligned}$$ Consider the function $f_j\in C_U$ given by $$\begin{aligned}
h_j(x)=\begin{cases}
\infty & x\in (-\infty,0)\times\mathbb{R}^{n-1},\\
\max(\frac{(x_1-(j+\frac{1}{j}))^2}{2}+\sum\limits_{i=2}^{n}\frac{x_i^2}{2}, \frac{j^2}{2}))-\frac{j^2}{2} & x\in [0,\infty)\times\mathbb{R}^{n-1}.
\end{cases}
\end{aligned}$$ Then $h_j\equiv 0$ on $B_{j}(j+\frac{1}{j},0,\dots,0)$. Setting $x_j:=(j+\frac{1}{j},0,\dots,0)$, we see that $x\in B_{j}(x_j)$ implies $$\begin{aligned}
|x-x_{j+1}|\le |x-x_j|+|x_j-x_{j+1}|\le j+\frac{1}{j}-\frac{1}{j+1}\le j+1,
\end{aligned}$$ so $B_j(x_j)\subset B_{j+1}(x_{j+1})$. If $y=(y_1,...,y_n)\in (0,\infty)\times\mathbb{R}^{n-1}$ is given, $$\begin{aligned}
|y-x_j|^2-j^2=-2(j+\frac{1}{j})y_1+\frac{1}{j^2}+2+\sum_{i=1}^{n}y_i^2\rightarrow-\infty \quad \text{for }j\rightarrow\infty,
\end{aligned}$$ so we see that $\bigcup_{j\in\mathbb{N}}B_j(x_j)=(0,\infty)\times{\mathbb{R}}^{n-1}$. In particular, the sequence $(h_j)_j$ converges pointwise to $h:=I^\infty_{[0, \infty)\times\mathbb{R}^{n-1}}$ for all $x\notin \{0\}\times {\mathbb{R}}^{n-1}$. By Proposition \[proposition\_convergence\_finite\_convex\_functions\] this implies that $(h_j)_j$ epi-converges to $h$.\
Now set $c_j:=\max_{i=1,...,k}\|\phi_i^j\|_{C^2(V)}$ and define $f_i^j:= c_jh_j+\phi_i^j$. Then $f_i^j\in C_U$ for all $1\le i\le k$, $j\in\mathbb{N}$. By construction $\lim_{j\rightarrow\infty}f_i^j(x)=I^\infty_{[0, \infty)\times\mathbb{R}^{n-1}}(x)=h(x)$ for $x\notin \{0\}\times\mathbb{R}^{n-1}$, so Proposition \[proposition\_convergence\_finite\_convex\_functions\] shows that $(f_i^j)_j$ epi-converges to $h$ for $j\rightarrow\infty$. Using the definition of the Goodey-Weil embedding and the joint continuity of the polarization $\bar{\mu}$, we obtain the contradiction $$\begin{aligned}
1=&\lim\limits_{j\rightarrow\infty}GW(\mu)(\phi_1^j\otimes...\otimes\phi_k^j)\\
=&\lim\limits_{j\rightarrow\infty}\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(f^j_{\sigma(1)},...,f^j_{\sigma(l)},h_{\sigma(l+1)},\dots,h_{\sigma(k)})\\
=&\sum\limits_{l=0}^k(-1)^{k-l}\frac{1}{(k-l)!l!}\sum_{\sigma\in S_k}\bar{\mu}(h[l],h[k-l])\\
=& (-1)^k\mu(h)\sum\limits_{l=0}^k(-1)^{l}\frac{k!}{(k-l)!l!}=0.
\end{aligned}$$ Thus $0\notin{\mathrm{supp\ }}\mu$.
Valuations on convex functions on open convex sets {#section_valuations_on_open_sets}
--------------------------------------------------
For an open and convex subset $U\subset V$ let us denote by ${\mathrm{Conv}}(U,{\mathbb{R}})$ the space of all convex functions $f:U\rightarrow{\mathbb{R}}$. Equipped with the topology of uniform convergence on compact subset of $U$, ${\mathrm{Conv}}(U,{\mathbb{R}})$ becomes a metrizable topological space.
For $f\in{\mathrm{Conv}}(U,{\mathbb{R}})$, define $\tilde{f}$ by $$\begin{aligned}
\tilde{f}(x_0)=\begin{cases}
f(x_0) & x_0\in U,\\
\liminf\limits_{x\rightarrow x_0,x\in U}f(x) & x_0\in\partial U,\\
\infty & x_0\in V\setminus\bar{U}.
\end{cases}
\end{aligned}$$ Then $\tilde{f}\in{\mathrm{Conv}}(V)$.
Observe that $\tilde{f}(x_0)=\liminf\limits_{x\rightarrow x_0,x\in U}f(x)$ for all $x\in\bar{U}$, as $f$ is continuous on $U$.\
Obviously, $\tilde{f}$ is lower semi-continuous. We need to show that $\tilde{f}>-\infty$ and that $\tilde{f}$ is convex.\
Let $x\in\partial U$ be any point, $(x_j)_j$ a sequence in $U$ converging to $x$ such that $\lim\limits_{j\rightarrow\infty}f(x_j)=\tilde{f}(x)$. For $y\in U$ and $\lambda\in (0,1)$ the convexity of $f$ implies $$\begin{aligned}
f(\lambda y+(1-\lambda)x_j)\le \lambda f(y)+(1-\lambda)f(x_j).
\end{aligned}$$ As $U$ is open, $\lambda y+(1-\lambda)x_j\rightarrow \lambda y+(1-\lambda)x$ in $U$ for all $\lambda\in (0,1)$, so the continuity of $f$ implies $$\begin{aligned}
f(\lambda y+(1-\lambda)x)\le \lambda f(y)+(1-\lambda)\tilde{f}(x).
\end{aligned}$$ In particular, $\tilde{f}(x)>-\infty$. In addition, we see that $\tilde{f}$ is convex along line segments $[x,y]$, where $x\in\partial U$ and $y\in U$. To see that $\tilde{f}$ is convex, the only non-trivial case remaining is a line segment $[x,y]$ where $x,y\in\partial U$. Take a sequence $(y_j)_j$ in $U$ converging to $y$ such that $\lim\limits_{j\rightarrow\infty}f(y_j)=\tilde{f}(y)$. Using the inequality above we see that for $\lambda\in (0,1)$ $$\begin{aligned}
f(\lambda y_j+(1-\lambda)x)\le \lambda f(y_j)+(1-\lambda)\tilde{f}(x).
\end{aligned}$$ Now $\lambda y_j+(1-\lambda)x\in U$ defines a sequence converging to $\lambda y+(1-\lambda)x\in\bar{U}$. Thus taking limits and using the remark we obtain $$\begin{aligned}
\tilde{f}(\lambda y+(1-\lambda)x)\le\liminf\limits_{j\rightarrow\infty}f(\lambda y_j+(1-\lambda)x)\le \lambda \tilde{f}(y)+(1-\lambda)\tilde{f}(x).
\end{aligned}$$
\[proposition\_continuity\_extension\_from\_open\_convex\_set\] The extension $f\mapsto \tilde{f}$ defines a continuous, injective map $i_U:{\mathrm{Conv}}(U,{\mathbb{R}})\rightarrow C_U:=\{f\in{\mathrm{Conv}}(V) \ : \ f|_U<\infty\}$. The inverse map is given by restricting the map $$\begin{aligned}
{\mathrm{res}}:C_U&\rightarrow {\mathrm{Conv}}(U,{\mathbb{R}})\\
f&\mapsto f|_U
\end{aligned}$$ to the image of ${\mathrm{Conv}}(U,{\mathbb{R}})$ in $U_C$ and is also continuous. In addition, $i_U$ and ${\mathrm{res}}$ are compatible with the formation of pointwise maximum and minimum of two convex functions.
It is clear that $i_U$ is injective. To see that it is continuous, it is enough to show that it is sequentially continuous, as both spaces are metrizable.\
Let $(f_j)_j\subset{\mathrm{Conv}}(U,{\mathbb{R}})$ be a sequence converging to $f\in{\mathrm{Conv}}(U,{\mathbb{R}})$. Then $(\tilde{f}_j)_j$ converges pointwise on the dense subset $V\setminus\partial U$ to $\tilde{f}$, so the claim follows from Proposition \[proposition\_convergence\_finite\_convex\_functions\]. Of course, the restriction map defines the inverse to this extension procedure. The continuity follows again from Proposition \[proposition\_convergence\_finite\_convex\_functions\].\
Obviously, the restriction map is compatible with the formation of the pointwise maximum and minimum. If $f,h\in{\mathrm{Conv}}(U,{\mathbb{R}})$, then $i_U(\max(f,h))=\max (i_U(f),i_U(h))$ on $V\setminus \partial U$. Thus $i_U(\max(f,h))(x_0)\le\max(i_U(f),i_U(h))(x_0)$ for $x_0\in\partial U$ by definition of $i_U$. For the reverse inequality, take a sequence $(x_j)_j$ in $U$ converging to $x_0$ such that $$\begin{aligned}
\lim\limits_{j\rightarrow\infty}\max(f(x_j),h(x_j))=i_U(\max(f,h))(x_0).
\end{aligned}$$ As $i_U(f)$ and $i_U(h)$ are lower semi-continuous, given $\epsilon>0$ there exists $N\in\mathbb{N}$ such that $i_U(f)(x_0)\le f(x_j)+\epsilon$ and $i_U(h)(x_0)\le h(x_j)+\epsilon$ for all $j\ge N$ and thus $$\begin{aligned}
\max(i_U(f),i_U(h))(x_0)\le\max(f(x_j),h(x_j))+\epsilon \quad\forall j\ge N.
\end{aligned}$$ Thus $\max(i_U(f),i_U(h))(x_0)\le i_U(\max(f,h))(x_0)$. The same argument can be applied to the minimum.
Let $U$ be an open convex subset, $F$ a real locally convex vector space. We will denote the space of all continuous valuations $\mu:{\mathrm{Conv}}(U,{\mathbb{R}})\rightarrow F$ that are dually epi-translation invariant by ${\mathrm{VConv}}(U,F)$.
As usual, we equip ${\mathrm{VConv}}(U,F)$ with the topology of uniform convergence on compact subsets, which is generated by the semi-norms $\|\mu\|_{F;K}:=\sup_{f\in K}|\mu(f)|_F$ for all continuous semi-norms $|\cdot|_F$ of $F$ and compact subsets $K\subset {\mathrm{Conv}}(U,{\mathbb{R}})$.\
Let $C_U:=\{f\in{\mathrm{Conv}}(V) \ : \ f|_U<\infty\}$ be the cone of convex functions that are finite on $U$. Using Proposition \[proposition\_continuity\_extension\_from\_open\_convex\_set\], we can consider the map $$\begin{aligned}
{\mathrm{res}}^*:{\mathrm{VConv}}(U,F)&\rightarrow {\mathrm{VConv}}(C_U;V,F)\\
\mu&\mapsto [f\mapsto \mu(f|_U)].
\end{aligned}$$
\[lemma\_properties\_inclusion\_open\_subset\] ${\mathrm{res}}^*:{\mathrm{VConv}}(U,F)\rightarrow {\mathrm{VConv}}(C_U;V,F)$ is injective and continuous.
Assume that ${\mathrm{res}}^*(\mu)=0$. Let $f\in{\mathrm{Conv}}(U,{\mathbb{R}})$. The Lipschitz regularization ${\mathrm{reg}}_r \tilde{f}$ belongs to ${\mathrm{Conv}}(V,{\mathbb{R}})$ for $r>0$ small enough, Proposition \[proposition\_properties\_LIpschitz\_regularization\] together with Proposition \[proposition\_continuity\_extension\_from\_open\_convex\_set\] implies that $\mu([{\mathrm{reg}}_r\tilde{f}]|_U)$ converges to $\mu(f)$. However $\mu([{\mathrm{reg}}_r\tilde{f}]|_U)=0$, so $\mu(f)=0$. As this holds for arbitrary $f\in{\mathrm{Conv}}(U,{\mathbb{R}})$, $\mu=0$.\
To see that the map is continuous, let $K\subset C_U$ be a compact subset. As the restriction ${\mathrm{res}}:C_U\rightarrow{\mathrm{Conv}}(U,{\mathbb{R}})$ is continuous due to Proposition \[proposition\_continuity\_extension\_from\_open\_convex\_set\], ${\mathrm{res}}(K)\subset{\mathrm{Conv}}(U,{\mathbb{R}})$ is compact, and $$\begin{aligned}
\|{\mathrm{res}}^*\mu\|_{F;K}=\|\mu\|_{F;{\mathrm{res}}(K)}.
\end{aligned}$$ Thus ${\mathrm{res}}^*$ is continuous.
In addition to ${\mathrm{res}}^*$, we can also consider $$\begin{aligned}
i_{U}^*:{\mathrm{VConv}}(C_U;V,F)&\rightarrow{\mathrm{VConv}}(U,F)\\
\mu&\mapsto [f\mapsto \mu(i_U(f))].
\end{aligned}$$ Using the same argument as in Lemma \[lemma\_properties\_inclusion\_open\_subset\], we see that this is well defined and continuous.
\[proposition\_isormorphism\_restriction\] Let $F$ be a locally convex vector space admitting a continuous norm. Then $i_{U}^*$ and ${\mathrm{res}}^*$ are topological isomorphisms and inverse to each other.
It is easy to see that $i_{U}^*\circ {\mathrm{res}}^*=Id_{{\mathrm{VConv}}(U,F)}$, so $i_U^*$ is surjective. Let us show that $i_U^*$ is injective. Assume that $\mu\in{\mathrm{VConv}}(C_U;V,F)$ satisfies $\mu(i_U(f))=0$ for all $f\in{\mathrm{Conv}}(U,{\mathbb{R}})$. Due to Proposition \[proposition\_support\_valuations\_on\_open\_set\], the support of $\mu$ is compactly contained in $U$. Given $h\in{\mathrm{Conv}}(V,{\mathbb{R}})$, the function $i_U(h|_U)$ coincides with $h$ on $U$, i.e. they coincide on a neighborhood of the support of $\mu$. Proposition \[proposition\_support\_convex\_valuation\] implies $\mu(h)=\mu(i_U(h|_U))=0$. Thus $\mu$ vanishes on the dense subset ${\mathrm{Conv}}(V,{\mathbb{R}})\subset C_U$, i.e. $\mu=0$.\
We obtain $(i_U^*)^{-1}={\mathrm{res}}^*$, which is continuous. The same applies to $({\mathrm{res}}^*)^{-1}=i_U$.
This is just a reformulation of Proposition \[proposition\_isormorphism\_restriction\].
Jonas Knoerr, <span style="font-variant:small-caps;">Institut für Mathematik, Goethe-Universität Frankfurt am Main, Robert-Mayer-Str. 10, 60054 Frankfurt, Germany</span>
*E-mail address*: `[email protected]`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper studies a distributed collision avoidance control problem for a group of rigid bodies on a sphere. A rigid body network, consisting of multiple rigid bodies constrained to a spherical surface and an interconnection topology, is first formulated. In this formulation, it is shown that motion coordination on a sphere is equivalent to attitude coordination on the 3-dimensional Special Orthogonal group. Then, an angle based control barrier function that can handle a geodesic distance constraint on a spherical surface is presented. The proposed control barrier function is then extended to a relative motion case and applied to a collision avoidance problem for a rigid body network operating on a sphere. Each rigid body chooses its control input by solving a distributed optimization problem to achieve a nominal distributed motion coordination strategy while satisfying constraints for collision avoidance. The proposed collision-free motion coordination law is validated through simulation.'
author:
- 'Tatsuya Ibuki$^{1}$, Sean Wilson$^{2}$, Aaron D. Ames$^{3}$, and Magnus Egerstedt$^{2}$[^1][^2][^3][^4]'
title: |
**Distributed Collision-Free Motion Coordination on a Sphere:\
A Conic Control Barrier Function Approach**
---
INTRODUCTION
============
Safe and distributed motion coordination of individual robots within a multi-robot collective is required to solve many tasks, like formation, flocking, and coverage control [@book:RN]-[@martinez_07CSM]. While many studies focused on motion coordination of a multi-robot system consider the 3-dimensional (3D) Euclidean space or a 2-dimensional (2D) plane as a workspace, a spherical surface is also often required [@muralidharan_19TAC]-[@saber_06CDC]. This motion coordination on a sphere is motivated not only by theoretical interests but also by some industrial application such as planetary-scale motion coordination/localization in the space/ocean, and vehicle coordination on the surface with a large radius of curvature. Moreover, spherical motion constraints can be considered for motion control of single arm manipulators or attitude control of pan-tilt cameras, and recently, the constraints are also utilized in dynamics of cooperative transportation of a payload with multiple flying vehicles [@lee_14CDC; @wu_14CDC]. No matter what motion constrained workspace a multi-robot system operates in, effective collision avoidance is an essential requirement to guarantee hardware safety. A potential function based approach is one common collision avoidance strategy for multi-agent systems [@li_14TAC], [@verginis_19CSL]-[@dimarogonas_06Automatica]. This technique introduces a nonnegative scalar function that increases as a robot approaches obstacles, like other robots or environmental hazards. Then, collision-free motion coordination methods incorporate its negative gradient to guarantee a safe operating distance. However, the potential function often needs to be infinite at the obstacle boundary, which causes overcaution about safety, i.e., less control performance.
More recently, constraint based optimization methods have been used to guarantee robot safety during operation [@ames_17TAC; @kolathaya_19CSL]. Here, a scalar function describing a safe set, called a *control barrier function (CBF)*, is introduced, and the forward invariance of the dynamics within the safe set is guaranteed via the constraints derived by the CBF. In this approach, the control input is given by solving an optimization problem to achieve a control task *as much as possible* while guaranteeing the safety. This technique is also applied to collision-free motion coordination problems for multi-agent systems as in [@ibuki_20RAL; @wang_17TRO; @panagou_16TAC]. Most of the existing studies, however, consider collision avoidance problems with standard Euclidean distances, i.e., in 3D space or on a 2D plane. This work extends the CBF based approach to a collision-free motion coordination method on a spherical surface, where the safety is defined with geodesic distances.
This paper first formulates a rigid network consisting of multiple rigid bodies with their motion dynamics constrained to a spherical surface and an interconnection topology. We show that motion coordination on a sphere is equivalent to attitude coordination on the 3D Special Orthogonal group: $SO(3)$. Then, as a bridge to a collision avoidance problem on a sphere, we develop a CBF based safe control technique on $SO(3)$. This approach is first applied to an angle based (conic) constraint satisfaction problem, and by extending it to a relative motion case, we propose a collision-free motion coordination law for a rigid body network on a sphere. In the proposed method, each rigid body selects its control input by solving a distributed optimization problem to achieve a given motion coordination task as much as possible while guaranteeing collision avoidance. The effectiveness of the proposed approach is demonstrated via simulation.
The main contributions of this paper are twofold. First, we develop a new CBF to handle constraints on $SO(3)$ by extending the classical CBF methods for vector fields presented in [@ames_17TAC; @kolathaya_19CSL]. Here, we also provide an example of safe attitude control for a single rigid body with a conic constraint. Second, we extend this kind of CBFs to a relative motion case, and propose a novel distributed collision-free motion coordination method for a rigid body network on a spherical surface.
PROBLEM SETTINGS {#sec:rbm}
================
Rigid Body Motion
-----------------
As a preliminary, the motion dynamics of multiple rigid bodies in general 3D space are first introduced. Let us consider a set of $n$ rigid bodies. Each rigid body $i\in\{1,\dots,n\}$ has a body fixed frame $\Sigma_i$ in a world frame $\Sigma_w$. The position and attitude of rigid body $i$ in $\Sigma_w$ are represented by $(p_{wi},e^{\hat\xi_{wi}\theta_{wi}})\in SE(3)$. Here, $e^{\hat\xi\theta}\in SO(3)$ is the exponential coordinate of the rotation matrix with the rotation axis $\xi\in{{\mathbb R}}^3$ $(\|\xi\|=1)$ and angle $\theta\in[-\pi,\pi)$ [@book:robomani]. The operator $\wedge:{{\mathbb R}}^3\to so(3)$ gives $\hat ab=a\times b$ for any 3D vectors $a,b\in{{\mathbb R}}^3$, and $\vee: so(3)\to{{\mathbb R}}^3$ is its inverse operator. For the ease of representation, $\hat\xi_{wi}\theta_{wi}$ is written as $\hat\xi\theta_{wi}$ throughout this paper. The body translational and rotational velocity of rigid body $i$ relative to $\Sigma_w$ is denoted by $v_i\in{{\mathbb R}}^3$ and $\omega_i\in{{\mathbb R}}^3$, respectively. Then, for each rigid body $i$, we have the following *rigid body motion* [@book:robomani]: $$\begin{aligned}
\dot p_{wi}=e^{\hat\xi\theta_{wi}}v_i,~\dot e^{\hat\xi\theta_{wi}}=e^{\hat\xi\theta_{wi}}\hat\omega_i. \label{eq:rbm}\end{aligned}$$
Rigid Body Motion on a Sphere {#subsec:rbms}
-----------------------------
![Multiple rigid bodies on sphere. The origin of the world frame $\Sigma_w$ is located at the center of a sphere with radius $\rho$, and the position of each rigid body is constrained on the surface of the sphere.[]{data-label="fig:rbm_sphere"}](fig/rbm_sphere.eps){width="1\linewidth"}
Let us next consider the motion dynamics of multiple rigid bodies constrained to the surface of a sphere with radius $\rho>0$. Without loss of generality, let the origin of the world frame $\Sigma_w$ be located at the center of the sphere, and the direction of the $z$-axis of each body frame $\Sigma_i,~i\in\{1,\dots,n\}$ coincide with the radial direction (see Fig. \[fig:rbm\_sphere\]). In this case, each rigid body $i$ has the spherical constraint with the basis axis $e_3:=[0~0~1]^{{\mathrm T}}\in{{\mathbb S}}^2:=\{s\in{{\mathbb R}}^3\mid \|s\|=1\}$ as $$\begin{aligned}
p_{wi}=\rho e^{\hat\xi\theta_{wi}}e_3. \label{eq:s_constraint}\end{aligned}$$ Substituting (\[eq:s\_constraint\]) into (\[eq:rbm\]), we obtain the following *rigid body motion on a sphere* for each rigid body $i$:
\[eq:rbm\_sphere\] [ $$\begin{aligned}
\dot p_{wi}&=&-\rho e^{\hat\xi\theta_{wi}}\hat e_3\omega_i, \\
\dot e^{\hat\xi\theta_{wi}}&=&e^{\hat\xi\theta_{wi}}\omega_i. \label{eq:rbm_sphere_a}\end{aligned}$$]{}
\[remark:sphere\]
From the observation in Remark \[remark:sphere\], motion coordination on a sphere, like formation and collision avoidance, is equivalent to attitude coordination on $SO(3)$. Therefore, for the convenience of introducing control barrier functions in the subsequent discussion, this paper focuses on attitude control on $SO(3)$ and considers the rotational body velocity $\omega_i$ as the control input of each rigid body $i$. The actual control input on the sphere is then given by the first two elements of $v_i=-\rho\hat e_3\omega_i=[\rho\omega_{i,y}~-\rho\omega_{i,x}~0]^{{\mathrm T}}$ and the third element of $\omega_i$ (i.e., $\omega_{i,z}$) for the notation $\omega_i=[\omega_{i,x}~\omega_{i,y}~\omega_{i,z}]^{{\mathrm T}}$.
Rigid Body Network on a Sphere and Research Objective
-----------------------------------------------------
In this paper, we suppose that a motion coordination strategy to achieve a control task is given *a priori*, and mainly focus on a distributed collision avoidance problem. Here, the interconnection topology between rigid body pairs for the given motion coordination strategy is represented by a directed graph ${{\mathcal G}}=({{\mathcal V}},{{\mathcal E}})$, composed of the rigid body set ${{\mathcal V}}:=\{1,\dots,n\}$ and the edge set ${{\mathcal E}}\subset {{\mathcal V}}\times{{\mathcal V}}$. We also define the neighbor set of each rigid body $i$ for the strategy as ${{\mathcal N}}_i:=\{j\in{{\mathcal V}}\mid(j,i)\in{{\mathcal E}}\}$. Then, $j\in{{\mathcal N}}_i$ means that rigid body $i$ obtains information about rigid body $j$.
Throughout this work, a group of $n$ rigid bodies with the rigid body motion on a sphere (\[eq:rbm\_sphere\]) and the interconnection topology ${{\mathcal G}}$ is called a *rigid body network on a sphere*. In this formulation, the main objective of this work is to develop a distributed collision avoidance method for a rigid body network on a sphere.
CONIC CONTROL BARRIER FUNCTIONS {#sec:conic}
===============================
As a bridge to a collision avoidance problem for a rigid body network on a sphere, this section presents a geometric control barrier function (CBF) on $SO(3)$ (refer to [@ames_17TAC] for more details about CBFs). We note that only a single rigid body $i$ with the attitude dynamics described in (\[eq:rbm\_sphere\_a\]) is considered in this section.
Control Barrier Functions on $SO(3)$
------------------------------------
Consider the attitude dynamics (\[eq:rbm\_sphere\_a\]), for rigid body $i$ in the world frame $\Sigma_w$, with $e^{\hat\xi\theta_{wi}}\in{{\mathcal S}}\subset SO(3)$, $\omega_i\in\Omega\subset{{\mathbb R}}^3$, and the constraint set $C_o$ defined as $$\begin{aligned}
C_o:=\{e^{\hat\xi\theta_{wi}}\in SO(3)\mid h(e^{\hat\xi\theta_{wi}})\ge0\}. \label{eq:general_constraint}\end{aligned}$$ Here, $h: SO(3)\to{{\mathbb R}}$ is a continuously differentiable function. Then, we provide a CBF definition on $SO(3)$ similarly to [@ames_17TAC] as follows:
\[definition:ZCBF\] The function $h(e^{\hat\xi\theta_{wi}})$ is called a *zeroing CBF (ZCBF)* defined on the set ${{\mathcal S}}$ with $C_o\subseteq{{\mathcal S}}\subset SO(3)$, if there exists an extended class ${{\mathcal K}}$ function $\alpha:{{\mathbb R}}\to{{\mathbb R}}$ satisfying $$\begin{aligned}
\sup_{\omega_i\in\Omega}(\dot h(e^{\hat\xi\theta_{wi}})+\alpha(h(e^{\hat\xi\theta_{wi}})))\ge0~\forall e^{\hat\xi\theta_{wi}}\in{{\mathcal S}}. \nonumber\end{aligned}$$
Notice in Definition \[definition:ZCBF\] that the control input $\omega_i$ appears in $\dot h(e^{\hat\xi\theta_{wi}})$ from (\[eq:rbm\_sphere\_a\]). Then, we have the following theorem:
\[theorem:general\] If the function $h(e^{\hat\xi\theta_{wi}})$ is a ZCBF on ${{\mathcal S}}$, then any Lipschitz continuous controller $\omega_i:{{\mathcal S}}\to\Omega$ satisfying $\dot h(e^{\hat\xi\theta_{wi}})+\alpha(h(e^{\hat\xi\theta_{wi}}))\ge0$ will render the set $C_o$ forward invariant.
By appropriately considering the subspace ${{\mathcal S}}\subset SO(3)$, the attitude dynamics (\[eq:rbm\_sphere\_a\]) can be rewritten by local Lipschitz continuous dynamics in vector form (see Appendix). Corollary 2 in [@ames_17TAC] can be thus applied.
Theorem \[theorem:general\] means that the attitude $e^{\hat\xi\theta_{wi}}$ remains in the set $C_o$ for all time.
Definitions of Conic Control Barrier Functions
----------------------------------------------
![Conic constraints. Constraint (\[eq:conic\_l\]): The basis axis $e^{\hat\xi\theta_{wi}}a_i$ is constrained inside and on the boundary of the conic region determined by $a_w$ and $\theta_c$ in $\Sigma_w$. Constraint (\[eq:conic\_s\]): The basis axis $e^{\hat\xi'\theta'_{wi}}a_i$ is constrained outside and on the boundary of the conic region (shown by the semitransparent illustration).[]{data-label="fig:constraint"}](fig/conic.eps){width=".7\linewidth"}
Let us now provide explicit definitions of $h(e^{\hat\xi\theta_{wi}})$ to represent cone-type (conic) attitude constraints used in this work, which are motivated by [@weiss_14ACC]-[@nakano_18CDC]. Let $a_w,a_i\in{{\mathbb S}}^2$ be the basis axes in the world frame $\Sigma_w$ and the body frame $\Sigma_i$, respectively. Notice then that $e^{\hat\xi\theta_{wi}}a_i$ means the direction of the basis axis of $\Sigma_i$ viewed from $\Sigma_w$.
Consider two kinds of inequality constraints as follows:
\[eq:conic\] $$\begin{aligned}
&&\hspace{-4ex}a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}a_i\ge\cos\theta_c, \label{eq:conic_l} \\
&&\hspace{-4ex}a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}a_i\le\cos\theta_c, \label{eq:conic_s}\end{aligned}$$
where $\theta_c\in[0,\pi/2]$ is the constraint parameter to determine the size of the conic region. These constraints are formed by the inner product of the basis axes $a_w$ and $e^{\hat\xi\theta_{wi}}a_i$. The constraint (\[eq:conic\_l\]) (constraint (\[eq:conic\_s\])) thus means that the head of the vector $e^{\hat\xi\theta_{wi}}a_i$ is constrained inside (outside) and on the boundary of the conic region determined by $a_w$ and $\theta_c$ in $\Sigma_w$ (see Fig. \[fig:constraint\]). These kinds of constraints are called *conic constraints* in this paper.
We next develop a ZCBF to guarantee the conic constraint (\[eq:conic\_l\]). Based on (\[eq:conic\_l\]), an angle based ZCBF, referred to as a *conic CBF* in this work, is defined as $h(e^{\hat\xi\theta_{wi}}):=a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}a_i-\cos\theta_c$. This enables us to represent the attitude set satisfying the conic constraint (\[eq:conic\_l\]) by (\[eq:general\_constraint\]).
Then, we have the following corollary from Theorem \[theorem:general\]:
\[theorem:conic\] Any Lipschitz continuous controller $\omega_i:{{\mathcal S}}\to\Omega$ satisfying $$\begin{aligned}
-a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}\hat a_i\omega_i+\alpha(a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}a_i-\cos\theta_c)\ge0 \label{eq:condition}\end{aligned}$$ will render the set $C_o$ forward invariant.[^5]
The following corollary also holds for the constraint (\[eq:conic\_s\]):
\[corollary:conic\] Any Lipschitz continuous controller $\omega_i:{{\mathcal S}}\to\Omega$ satisfying $$\begin{aligned}
a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}\hat a_i\omega_i-\alpha(a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}a_i-\cos\theta_c)\ge0 \nonumber $$ will render the set $C_o$ with $h(e^{\hat\xi\theta_{wi}})=-a^{{\mathrm T}}_we^{\hat\xi\theta_{wi}}a_i+\cos\theta_c$ forward invariant.
Safe Control with Conic Control Barrier Functions
-------------------------------------------------
Corollary \[theorem:conic\] enables us to propose the attitude control input provided by the solution of the quadratic program to guarantee the conic constraint (\[eq:conic\_l\]) as follows: $$\begin{aligned}
\omega^\ast_i=\arg\min_{\omega_i\in\Omega}\|\omega_i-\omega_{nom,i}\|^2~\mathrm{s.t.}~(\ref{eq:condition}). \label{eq:QP}\end{aligned}$$ Here, $\omega_{nom,i}\in{{\mathbb R}}^3$ is the nominal controller to achieve a given control task. The optimization in (\[eq:QP\]) implies that rigid body $i$ tries to achieve the given control task as much as possible while guaranteeing the conic constraint (\[eq:conic\_l\]).
![Time trajectory of attitude. The blue line shows the time trajectory of the head of the vector $e^{\hat\xi\theta_{wi}}a_i$ in $\Sigma_w$. The attitude is constrained inside or on the boundary of the conic constraint (\[eq:conic\_l\]) represented by the red line.[]{data-label="fig:trajectory"}](fig/sphere.eps){width=".8\linewidth"}
As verification, we apply the control input (\[eq:QP\]) to the attitude dynamics (\[eq:rbm\_sphere\_a\]) with $a_w=a_i=e_3$, $\theta_c=\pi/6$, and a geometric trajectory tracking law as the nominal controller $\omega_{nom,i}$. Here, the desired trajectory is intentionally set so that the conic constraint (\[eq:conic\_l\]) is violated if the nominal controller is directly applied. Fig. \[fig:trajectory\] depicts the time trajectory of the head of the vector $e^{\hat\xi\theta_{wi}}a_i$ in $\Sigma_w$ by the blue line and the boundary of the conic constraint (\[eq:conic\_l\]) by the red one. We see from this figure that the conic constraint (\[eq:conic\_l\]) is satisfied.
COLLISION-FREE MOTION COORDINATION {#sec:ca}
==================================
Definitions of Collisions on a Sphere
-------------------------------------
As stated in Section \[subsec:rbms\], this paper focuses on attitude control on $SO(3)$ to deal with motion coordination of a rigid body network on a sphere. We first define the relative attitude of rigid body $j$ to rigid body $i$ as $e^{\hat\xi\theta_{ij}}:=e^{-\hat\xi\theta_{wi}}e^{\hat\xi\theta_{wj}}\in SO(3)$. Then, we extend the conic CBF approach presented in Section \[sec:conic\] to a collision avoidance problem by considering the relative attitude case of Corollary \[corollary:conic\] with $a_i=e_3~\forall i\in{{\mathcal V}}$.
Under the spherical constraint (\[eq:s\_constraint\]), the geodesic distance between rigid body $i$ and rigid body $j$ is defined as the arc length of the spherical surface (see Fig. \[fig:definitions\](a)): $$\begin{aligned}
d_g(p_{wi},p_{wj}):=\rho\cos^{-1}\left(\frac{p^{{\mathrm T}}_{wi}p_{wj}}{\rho^2}\right). \label{eq:g_distance0}\end{aligned}$$ Then, substituting (\[eq:s\_constraint\]) into (\[eq:g\_distance0\]) can rewrite $d_g(p_{wi},p_{wj})$ as $$\begin{aligned}
d_g(e^{\hat\xi\theta_{ij}})=\rho\cos^{-1}(e^{{\mathrm T}}_3e^{\hat\xi\theta_{ij}}e_3), \label{eq:g_distance}\end{aligned}$$ that is, the geodesic distance $d_g$ is formed by the relative attitude $e^{\hat\xi\theta_{ij}}$.
\[remark:g\_distance\]
From the geodesic distance definition (\[eq:g\_distance\]), we define collisions between rigid bodies and collision avoidance for a rigid body network on a sphere as follows (see Fig. \[fig:definitions\](b)):
The collision between rigid body $i$ and rigid body $j$ occurs when $d_g(e^{\hat\xi\theta_{ij}})<D_c$, where $D_c>0$ is their geodesic diameter. Then, *collision avoidance* is achieved for a rigid body network on a sphere if $$\begin{aligned}
d_g(e^{\hat\xi\theta_{ij}}(t))\ge D_c~\forall i,j\in{{\mathcal V}}~(i\neq j),~t\ge0. \label{eq:ca}\end{aligned}$$
In this formulation, the objective of this paper is to design conic CBFs, derive conditions, and propose a distributed method for rigid body $i\in{{\mathcal V}}$ in order to achieve the collision avoidance (\[eq:ca\]) for a rigid body network on a sphere.
\[fig:g\_distance\]
\[fig:collision\]
Distributed Collision Avoidance on a Sphere {#subsec:ca}
-------------------------------------------
Define the following safe set $C$ for a rigid body network on a sphere: $$\begin{aligned}
C\hspace{-.2ex}:=\hspace{-.2ex}\{e^{\hat\xi\theta_{wi}},~\hspace{-.2ex}i\in{{\mathcal V}}\hspace{-.2ex}\mid\hspace{-.2ex} d_g(e^{\hat\xi\theta_{ij}})\hspace{-.2ex}\ge\hspace{-.2ex} D_c~\hspace{-.2ex}\forall i,j\in{{\mathcal V}}~\hspace{-.2ex}(i\neq j)\}.\hspace{-.2ex}\label{eq:safe_set}\end{aligned}$$ Then, the collision avoidance (\[eq:ca\]) is equivalent to the forward invariance of the safe set $C$. Let us now assume that the collision distance $D_c$ satisfies $D_c<(\rho\pi)/2$. This assumption is reasonable since this inequality means that the geodesic diameter of each rigid body is less than one fourth of the circumference of the sphere, i.e., the size of each rigid body is not too large compared with that of the sphere.
We note that each rigid body is required to take collision avoidance behaviors only when it approaches other rigid bodies. Besides the graph ${{\mathcal G}}$ for a given motion coordination strategy, therefore, we introduce another geodesic distance based undirected graph ${{\mathcal G}}'=({{\mathcal V}},{{\mathcal E}}'),~{{\mathcal E}}':=\{(i,j)\in{{\mathcal V}}~(i\neq j)\mid d_g(e^{\hat\xi\theta_{ij}}) \le D_a~\forall i,j\in{{\mathcal V}}\}$. Here, $D_a>D_c$ is the geodesic distance within which rigid bodies take account of collision avoidance behaviors. According to ${{\mathcal G}}'$, we also define a new neighbor set of rigid body $i$ for the collision avoidance (\[eq:ca\]), called *distance neighbors*, as ${{\mathcal N}}_{d,i}:=\{j\in{{\mathcal V}}\mid(j,i)\in{{\mathcal E}}'\}$. Notice now that by employing the reasonable assumption $D_a<(\rho\pi)/2$, we can avoid the undesired special case stated in Remark \[remark:g\_distance\], i.e., the existence of a distance neighbor perfectly at the opposite position on the sphere, in the collision avoidance process.
Let us define the conic CBF candidates as $$\begin{aligned}
h_{ij}:=-e^{{\mathrm T}}_3e^{\hat\xi\theta_{ij}}e_3+\cos\left(\frac{D_c}{\rho}\right)\in{{\mathbb R}}\nonumber\end{aligned}$$ for $i,j\in{{\mathcal V}}$, where we take $\cos(\cdot)$ for the geodesic distance. Then, by rewriting (\[eq:safe\_set\]) as $$\begin{aligned}
C=\{e^{\hat\xi\theta_{wi}},~i\in{{\mathcal V}}\mid h_{ij}\ge0~\forall i,j\in{{\mathcal V}}~(i\neq j)\}, \nonumber\end{aligned}$$ the forward invariance of the safe set $C$ for a rigid body network on a sphere is analogous to collision-free motions. We now have the following theorem showing the achievement of the collision avoidance (\[eq:ca\]):
\[theorem:ca\] Suppose that collisions do not occur in a rigid body network on a sphere at the initial time, i.e., $\{e^{\hat\xi\theta_{wi}}(0)\}_{i\in{{\mathcal V}}}\in C$. Then, any Lipschitz continuous controllers $\omega_i,~i\in{{\mathcal V}}$ satisfying [ $$\begin{aligned}
e^{{\mathrm T}}_3e^{-\hat\xi\theta_{ij}}\hat e_3\omega_i&\ge& k\left(e^{{\mathrm T}}_3e^{\hat\xi\theta_{ij}}e_3-\cos\left(\frac{D_c}{\rho}\right)\right) \nonumber\vspace{1.5ex} \\
&&\hspace{12ex}\forall j\in{{\mathcal N}}_{d,i},~k>0 \label{eq:ca_condition}\end{aligned}$$]{} will render the safe set $C$ forward invariant.
The following condition is first derived from $\dot h_{ij}(e^{\hat\xi\theta_{ij}})+2kh_{ij}(e^{\hat\xi\theta_{ij}})\ge0$ for each rigid body $i\in{{\mathcal V}}$: $$\begin{aligned}
&&\hspace{-8ex}e^{{\mathrm T}}_3e^{-\hat\xi\theta_{ij}}\hat e_3\omega_i+e^{{\mathrm T}}_3e^{\hat\xi\theta_{ij}}\hat e_3\omega_j \nonumber \\
&&\ge2k\left(e^{{\mathrm T}}_3 e^{\hat\xi\theta_{ij}}e_3-\cos\left(\frac{D_c}{\rho}\right)\right)~\forall j\in{{\mathcal N}}_{d,i}. \label{eq:proof}\end{aligned}$$ Here, $\alpha(h)=2kh,~k>0$ is employed as an extended class ${{\mathcal K}}$ function, and only the distance neighbors $j\in{{\mathcal N}}_{d,i}$ are considered because we have $h_{ij}(e^{\hat\xi\theta_{ij}})>0$ for any $j\in{{\mathcal V}}\setminus{{\mathcal N}}_{d,i}~(j\neq i)$ from $D_a>D_c$.
The condition (\[eq:proof\]) for each rigid body $i$ is not distributed since it requires input information of distance neighbors, i.e., $\omega_j,~j\in{{\mathcal N}}_{d,i}$. We thus employ the distributed condition (\[eq:ca\_condition\]) to satisfy (\[eq:proof\]). Then, because $j\in{{\mathcal N}}_{d,i}\Leftrightarrow i\in{{\mathcal N}}_{d,j}$ and $e^{{\mathrm T}}_3e^{\hat\xi\theta_{ij}}e_3=e^{{\mathrm T}}_3e^{\hat\xi\theta_{ji}}e_3$ hold, the satisfaction of (\[eq:ca\_condition\]) for all $i\in{{\mathcal V}}$ guarantees (\[eq:proof\]) for all $i\in{{\mathcal V}}$. Here, considering (\[eq:ca\_condition\]) can be regarded as sharing (\[eq:proof\]) equally[^6] between rigid body $i$ and rigid body $j$. Theorem \[theorem:general\] can be thus applied.
Collision-Free Motion Coordination on a Sphere {#subsec:input}
----------------------------------------------
Based on Theorem \[theorem:ca\], we propose, for rigid body $i\in{{\mathcal V}}$ in a rigid body network on a sphere, the following control input provided by the solution of the quadratic program to guarantee the collision avoidance (\[eq:ca\]): $$\begin{aligned}
\omega^\ast_i=\arg\min_{\omega_i\in{{\mathbb R}}^3}\|\omega_i-\omega_{nom,i}\|^2~\mathrm{s.t.}~(\ref{eq:ca_condition}). \label{eq:input}\end{aligned}$$ Here, $\omega_{nom,i}\in{{\mathbb R}}^3,~i\in{{\mathcal V}}$ are the nominal control inputs to achieve a given motion coordination strategy. The optimization in (\[eq:input\]) implies that each rigid body tries to achieve the given motion coordination task as much as possible while guaranteeing the collision avoidance (\[eq:ca\]).
Any motion coordination strategy can be applied as $\omega_{nom,i}$ in (\[eq:input\]). In the simulation verification presented in Section \[sec:simulation\], we apply the following attitude synchronization law [@book:hatanaka]: $$\begin{aligned}
\omega_{nom,i}=k_c\sum_{j\in{{\mathcal N}}_i}{\mathrm{sk}}(e^{\hat\xi\theta_{ij}})^\vee. \label{eq:as_input}\end{aligned}$$ Here, $k_c>0$ is the controller gain, and ${\mathrm{sk}}(e^{\hat\xi\theta}):=(1/2)(e^{\hat\xi\theta}-e^{-\hat\xi\theta})=\hat\xi\theta\in so(3)$. Then, it is shown in [@book:hatanaka] that if the initial attitudes of a rigid body network satisfy $|\theta_{ij}(0)|<\pi~\forall i,j\in{{\mathcal V}}$ and the interconnection topology ${{\mathcal G}}$ is fixed and strongly connected, the control input $\omega_i=\omega_{nom,i}$ given by (\[eq:as\_input\]) achieves the *attitude synchronization* defined as follows: $$\begin{aligned}
\lim_{t\to\infty}\|e^{\hat\xi\theta_{wi}}(t)-e^{\hat\xi\theta_{wj}}(t)\|_F=0~\forall i,j\in{{\mathcal V}}. \label{eq:as}\end{aligned}$$ Here, $\|\cdot\|_F$ is the Frobenius norm.
In the case of a rigid body network on a sphere under the spherical constraint (\[eq:s\_constraint\]), the attitude synchronization (\[eq:as\]) also implies the *position synchronization* defined as $$\begin{aligned}
\lim_{t\to\infty}\|p_{wi}(t)-p_{wj}(t)\|=0~\forall i,j\in{{\mathcal V}}. \nonumber $$ Then, by applying the control input (\[eq:input\]) with the nominal input (\[eq:as\_input\]), we can expect the achievement of a flocking-like behavior: cohesion; alignment; and separation [@reynolds_87CG], on a sphere. Here, the final control input of rigid body $i\in{{\mathcal V}}$ is distributed and based only on relative attitudes with respect to $j\in{{\mathcal N}}_i\cup{{\mathcal N}}_{d,i}$, which can be implemented in a distributed manner using onboard sensors, e.g., vision or infrared, without any other communication or global information.
SIMULATION {#sec:simulation}
==========
Simulation is conducted to demonstrate the validity of the proposed collision-free motion coordination method (\[eq:input\]), (\[eq:as\_input\]). Here, we slightly modify each control input by adding the common rotational body velocity $\omega_c=[0.1~0.2~-0.4]\in{{\mathbb R}}^3$ to make it easy to see the final configuration of a rigid body network on a sphere. Even with this modification, the relative attitude dynamics are not changed, i.e., the same behavior in the sense of the relative states can be seen.
Consider a rigid body network on a sphere with 20 rigid bodies and strongly connected interconnection topology ${{\mathcal G}}$. The simulation parameters are set as $\rho=1$, (i.e., $D_c=\theta_c$), $\theta_c=\pi/150$, $D_a=2D_c$, $k=1$, and $k_c=5$. The initial positions are set so that each rigid body $i\in{{\mathcal V}}$ exists in the upper half of the sphere in the $z$-axis direction of $\Sigma_w$.
\[fig:trajectory2\] \[fig:distance\]
The simulation results are shown in Fig. \[fig:simulation\]. Fig. \[fig:simulation\](a) illustrates the position trajectories of the rigid body network on a sphere in $\Sigma_w$, which demonstrates the proposed control method achieves the cohesion and alignment behaviors. The collision avoidance (\[eq:ca\]) can be confirmed by Fig \[fig:simulation\](b) depicting the time responses of the minimum geodesic distance between the rigid body pairs $(i,j)\in{{\mathcal V}}\times{{\mathcal V}}~(i\neq j)$.
CONCLUSIONS
===========
This paper presented a distributed collision avoidance control method for a group of multiple rigid bodies on a sphere. Based on the fact that the rigid body motion constrained to a spherical surface is equivalent to attitude motion on $SO(3)$, the collision avoidance law is derived with conic control barrier functions on $SO(3)$ that can handle geodesic distance constraints on a spherical surface. In the proposed method, each rigid body chooses its control input by solving a distributed optimization problem to achieve a given motion coordination strategy while satisfying constraints for collision avoidance derived by the conic control barrier functions. The validity of the proposed approach was demonstrated via simulation.
APPENDIX {#appendix .unnumbered}
========
Local Lipschitz Continuity of Attitude Dynamics (\[eq:rbm\_sphere\_a\])
-----------------------------------------------------------------------
It is well known that according to the subspace ${{\mathcal S}}\subset SO(3)$, the attitude dynamics (\[eq:rbm\_sphere\_a\]) can be rewritten in vector form using the Euler angles, quaternions, Cayley parameters, etc. [@chaturvedi_11CSM]-[@huynh_09JMIV]. We here introduce the XYZ Euler angle case.
Consider the XYZ Euler angle representation: $(\phi_i,\psi_i,\eta_i)$, e.g., in the region $\phi_i,\psi_i,\eta_i\in(-\pi/2,\pi/2)$. Then, the rotation matrix $e^{\hat\xi\theta_{wi}}$ can be uniquely given as $e^{\hat\xi\theta_{wi}}=R_x(\phi_i)R_y(\psi_i)R_z(\eta_i)$ in the subspace ${{\mathcal S}}$ according to the region of the Euler parameters. Here, $R_x,R_y,R_z\in SO(3)$ are respectively the basis rotation matrices with respect to $x$-, $y$-, and $z$-axes [@book:robomani]. Then, with $\zeta_i:=[\phi_i~\psi_i~\eta_i]^{{\mathrm T}}\in{{\mathbb R}}^3$, the attitude dynamics (\[eq:rbm\_sphere\_a\]) are analogous to $$\begin{aligned}
\hspace{1ex}\dot\zeta_i=\begin{bmatrix}
\frac{\cos\eta_i}{\cos\psi_i} & -\frac{\sin\eta_i}{\cos\psi_i} & 0\\
\sin\eta_i & \cos\eta_i & 0 \\
-\cos\eta_i\tan\psi_i & \sin\eta_i\tan\psi_i & 1
\end{bmatrix}\omega_i=:g(\zeta_i)\omega_i. \nonumber$$ Since $g(\zeta_i)$ consists of smooth trigonometric functions, the attitude dynamics are locally Lipschitz continuous on the subspace ${{\mathcal S}}$.
[99]{} F. Bullo, J. Cortés, and S. Martínez, *Distributed Control of Robotic Networks: A Mathematical Approach to Motion Coordination Algorithms*. Princeton, NJ, USA: Princeton Univ. Press, 2009. M. Mesbahi and M. Egerstedt, *Graph Theoretic Methods for Multiagent Networks*. Princeton, NJ, USA: Princeton University Press, 2010. W. Ren and R. W. Beard, *Distributed Consensus in Multi-vehicle Cooperative Control: Theory and Applications*. London, U.K.: Springer-Verlag, 2008. T. Ibuki, S. Wilson, J. Yamauchi, M. Fujita, and M. Egerstedt, “Optimization-based distributed flocking control for multiple rigid bodies,” *IEEE Robot. Autom. Lett.*, vol. 5, no. 2, pp. 1891–1899, Apr. 2020. S. Martínez, J. Cortés, and F. Bullo, “Motion coordination with distributed information," *IEEE Control Syst. Mag.*, vol. 27, no. 4, pp. 75–88, Aug. 2007. V. Muralidharan, A. D. Mahindrakar, and A. Saradagi, “Control of a driftless bilinear vector field on $n$-sphere,” *IEEE Trans. Autom. Control*, vol. 64, no. 8, pp. 3226–3228, Aug. 2019. J. Markdahl, J. Thunberg, and J. Goncalves, “Almost global consensus on the $n$-sphere,” *IEEE Trans. Autom. Control*, vol. 63, no. 6, pp. 1664–1675, Jun. 2018. S. Al-Abri and F. Zhang, “Consensus on a sphere for a 3-dimensional speeding up and slowing down strategy,” in *Proc. 56th IEEE Conf. Decis. Control*, 2017, pp. 1503–1508. C. Lageman and Z. Sun, “Consensus on spheres: Convergence analysis and perturbation theory,” in *Proc. 55th IEEE Conf. Decis. Control*, 2016, pp. 19–24. W. Li, “Collective motion of swarming agents evolving on a sphere manifold: A fundamental framework and characterization,” *Sci. Rep.*, vol. 5, 2015, Art. no. 13603. W. Li and M. W. Spong, “Unified cooperative control for multiple agents on a sphere for different spherical patterns,” *IEEE Trans. Autom. Control*, vol. 59, no. 5, pp. 1283–1289, May 2014. S. Hernandez and D. A. Paley, “Three-dimensional motion coordination in a spatiotemporal flowfield,” *IEEE Trans. Autom. Control*, vol. 55, no. 12, pp. 2805–2810, Dec. 2010. R. Olfati-Saber, “Swarms on sphere, A programmable swarm with synchronous behaviors like oscillator networks,” in *Proc. 45th IEEE Conf. Decis. Control*, 2006, pp. 5060–5066. T. Lee, “Geometric control of multiple quadrotor UAVs transporting a cable suspended rigid body,” in *Proc. 53rd IEEE Conf. Decis. Control*, 2014, pp. 6155–6160. G. Wu and K. Sreenath, “Geometric control of multiple quadrotors transporting a rigid-body load,” in *Proc. 53rd IEEE Conf. Decis. Control*, 2014, pp. 6141–6148. C. K. Verginis and D. V. Dimarogonas, “Closed-form barrier functions for multi-agent ellipsoidal systems with uncertain Lagrangian dynamics,” *IEEE Control Syst. Lett.*, vol. 3, no. 3, pp. 727–732, Jul. 2019. T. Hatanaka, N. Chopra, M. Fujita, and M. W. Spong, *Passivity-based Control and Estimation in Networked Robotics*. Cham, Switzerland: Springer, 2015. L. Sabattini, C. Secchi, and N. Chopra, “Decentralized connectivity maintenance for networked Lagrangian dynamical systems with collision avoidance,” *Asian J. Control*, vol. 17, no. 1, pp. 111-123, Jan. 2015. D. M. Stipanović, P. F. Hokayem, M. W. Spong, and D. D. $\check{\textrm{S}}$iljak, “Cooperative avoidance control for multiagent systems,” *J. Dyn. Syst., Meas., Control*, vol. 129, no. 5, pp. 699–707, Apr. 2007. D. V. Dimarogonas, S. G. Loizou, K. J. Kyriakopoulos, and M. M. Zavlanos, “A feedback stabilization and collision avoidance scheme for multiple independent non-point agents,” *Automatica*, vol. 42, pp. 229–243, Feb. 2006. A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,” *IEEE Trans. Autom. Control*, vol. 62, no. 8, pp. 3861–3876, Aug. 2017. S. Kolathaya and A. D. Ames, “Input-to-state safety with control barrier functions,” *IEEE Control Syst. Lett.*, vol. 3, no. 1, pp. 108–113, Jan. 2019. L. Wang, A. D. Ames, and M. Egerstedt, “Safety barrier certificates for collisions-free multirobot systems,” *IEEE Trans. Robot.*, vol. 33, no. 3, pp. 661–674, Jun. 2017. D. Panagou, D. M. Stipanović, and P. G. Voulgaris, “Distributed coordination control for multi-robot networks using Lyapunov-like barrier functions,” *IEEE Trans. Autom. Control*, vol. 61, no. 3, Mar. 2016. R. M. Murray, Z. Li, and S. S. Sastry, *A Mathematical Introduction to Robotic Manipulation*. Boca Raton, FL, USA: CRC Press, 1994. N. A. Chaturvedi, A. K. Sanyal, and N. H. McClamroch, “Rigid-body attitude control,” *IEEE Control Syst. Mag.*, vol. 31, no. 3, pp. 30–51, Jun. 2011. H. Schaub, P. Tsiotras, and J. L. Junkins “Principal rotation representations of proper $N\times N$ orthogonal matrices,” *Int. J. Eng. Sci.*, vol. 33, no. 15, pp. 2277–2295, Dec. 1995. D. Q. Huynh, “Metrics for 3D rotations: Comparison and analysis,” *J. Math. Imaging Vis.*, vol. 35, pp. 155–164, Jun. 2009. A. Weiss, F. Leve, M. Baldwin, J. R. Forbes, and I. Kolmanovsky, “Spacecraft constrained attitude control using positively invariant constraint admissible sets on $SO(3)\times{{\mathbb R}}^3$,” in *Proc. Amer. Control Conf.*, 2014, pp. 4955–4960. S. Kulumani and T. Lee, “Constrained geometric attitude control on $SO(3)$,” *Int. J. Control, Autom. Syst.*, vol. 15, no. 6, pp. 2796–2809, Dec. 2017. S. Nakano, T. W. Nguyen, E. Garone, T. Ibuki, and M. Sampei, “Attitude constrained control on $SO(3)$: An explicit reference governor approach,” in *Prof. 57th IEEE Conf. Decis. Control*, 2018, pp. 1833–1838. C. W. Reynolds, “Flocks, herds and schools: A distributed behavioral model,” *Comput. Graph.*, vol. 21, no. 4, pp. 25–34, Jul. 1987.
[^1]: \*This work was supported in part by JSPS KAKENHI under Grant 18K13775 and in part by the US National Science Foundation through Grant 1531195.
[^2]: $^{1}$T. Ibuki is with Department of Systems and Control Engineering, School of Engineering, Tokyo Institute of Technology, Tokyo 152-8550, Japan [[email protected]]{}
[^3]: $^{2}$S. Wilson and M. Egerstedt are with the School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA [{sean.t.wilson, magnus}@gatech.edu]{}
[^4]: $^{3}$A. D. Ames is with the Department of Mechanical and Civil Engineering, California Institute of Technology, Pasadena, CA 91125, USA [[email protected]]{}
[^5]: The property $\hat ab=-\hat ba$ for any 3D vectors $a,b\in{{\mathbb R}}^3$ is used to obtain the condition (\[eq:condition\]).
[^6]: As generalization of the equally sharing, we can also introduce weights $w_{ij}\in{{\mathbb R}},~(j,i)\in{{\mathcal E}}'$ satisfying $w_{ij}+w_{ji}=1$ to share the condition (\[eq:proof\]).
|
{
"pile_set_name": "ArXiv"
}
|
**Quantum multiplexing with optical coherent states**
Juan Carlos García-Escartín
*e-mail: [email protected]*
Pedro Chamorro-Posada
*Dep. Teoría de la Señal y Comunicaciones e Ingenería Telemática.*
*E.T.S.I. Telecomunicación, Campus Miguel Delibes, Camino del Cementerio s/n,*
*47011 Valladolid, Spain.*
Introduction {#intro}
============
Quantum information promises secure communications, faster algorithms and new ways to think about computation and physics [@NC00]. The recent developments in quantum computation and communication have reached a point where architectural problems need to be taken into account. As quantum processing units and networks grow in complexity, it is increasingly important to give efficient ways to connect the different elements inside quantum processors and to be able to accommodate multiple users in the same quantum network.
One particularly important problem is how a finite amount of resources is assigned to a number of competing users. In most communication systems, many different users share the available channel capacity and a series of multiple access techniques are employed. Inspired by one of these techniques, Code Division Multiple Access, CDMA, we will put forward a coherent state quantum multiplexer. The proposed scheme is easily scaled and the only limit is given by the desired transmission quality. Coherent states allow to reproduce some of the most interesting features of CDMA, like the use of almost orthogonal signals.
The given multiplexing architecture will be an extension of the ideas of optical coherent state quantum computation. We will present the building blocks needed in such an extension and suggest simple physical implementations for them. Finally, an analysis of the efficiency is presented and future possible lines of work are commented.
Section \[multiple\] introduces the problem of multiple access and gives a brief review of the most extended techniques both in classical and quantum communication systems. One of them, CDMA, is covered in particular detail as the inspiration for the proposed multiple access scheme. Section \[QCoherent\] reviews the fundamental aspects of quantum computation with coherent states and why it is an adequate representation for a multiple access system. Section \[tools\] presents the optical elements that are the basic building blocks of our proposal. Section \[cohmux\] gives a description of the multiplexing system and presents the block diagrams of the quantum multiplexer and demultiplexer. It also discusses how the number of users can be dynamically increased. The section is completed with an example of operation. Section \[Conclusion\] concludes with a review of the obtained results and a hint of future possibilities.
Multiple users in communications {#multiple}
================================
Channel capacity is a limited resource. In many situations, the same transmission resources need to be shared by more than one user. *Multiple access* methods are concerned with how $N$ users can share 1 channel. This scenario arises naturally in many classical communication systems [@Skl83; @Skl01]. For instance, there is only a limited amount of frequencies that can be used to transmit efficiently through the atmosphere and usually limited frequency bands are allocated to each service. Additionally, each band must be offered to users that have to make a new division. In those cases, Frequency Division Multiple Access, FDMA, schemes are applied to coordinate the division of the spectrum. Another usual situation is the one of optical networks, where optical fibre deployment is costly and physical links are scarce. There are different techniques to accommodate many users into the same fibre link. Users can transmit in turns, one user at a time, as in Time Division Multiple Access, TDMA. It is also possible to associate each user to a different wavelength, so that different data do not interfere, as in Wavelength Division Multiple Access, WDMA. These are some of the multiple options that are in widespread use in current communication systems.
CDMA {#CDMA}
----
One multiple access alternative that has been especially successful in modern wireless networks is Code Division Multiple Access, CDMA [@PSM82; @Sta01]. In this form of multiple access, all users emit at the same time and their signals spread through all the available bandwidth. In order to avoid interference, the system takes advantage of the algebraic properties of the transmitted signals, which are encoded in orthogonal codes. In CDMA, each user, $U$, is assigned a code $c_U$. User $U$ will transmit the code $-c_U$ to send a 0 and $c_U$ to send 1. The resulting signal at the receiver, $d$, will be the linear combination of all the transmitted data. The receiver can select a channel by applying the corresponding decoding function $s_U(d)$, which is calculated taking the scalar product of $d$ by the code $c_U$ of the user the receiver is listening to. The result of $s_U(d)$ is then compared to a certain threshold to recover the data. If all the users have orthogonal codes, the original 0s and 1s are perfectly recovered (see the example of Figure \[CDMA1\]).
![[**Example of CDMA multiple access:**]{} CDMA for three users A, B and C that transmit 0, 1 and 0, respectively. At the receiver, 1s will result in a positive weight, 8, equal to the length of the code, and 0s will produce the corresponding negative result, -8. The given codes, $c_A$, $c_B$ and $c_C$ are orthogonal and perfect symbol recovery is possible in an ideal transmission.\[CDMA1\]](CDMA1.eps)
Interestingly, the scheme is still valid for almost orthogonal codes (see Figure \[CDMA2\]). If two user codes have a small scalar product, the decoding process can still be valid if there is low noise and all the signals arrive with the same power. As the number of non-orthogonal codes grows, it becomes increasingly difficult to decode the information of each channel correctly. Similarly, the smaller the overlap between codes, the better the quality of the transmission is.
![[**Example of CDMA with non-orthogonal codes:**]{} A new user, D, is added to the transmission scenario of Figure \[CDMA1\]. Its code, $c_D$, is not orthogonal to the previous ones, but has a small overlap with them. In this case, where D is sending a 1, decoding the received superposed signal, $d$, still gives weights that can be easily recognized as 1s, the positive values, or as 0s, the negative ones.\[CDMA2\]](CDMA2.eps)
The ability to use non-orthogonal codes provides CDMA systems a unique characteristic. Unlike other multiple access schemes, users can be added above the perfect operation limit. For instance, in an FDMA system, if all the spectrum divisions are taken, new users are not allowed to join the channel. Already established connections are given preference and newcomers must wait until one frequency band is liberated. In CDMA, when the orthogonal codes are all assigned, new users can transmit using new non-orthogonal codes, often chosen so as to minimize interference. The quality of the communication will decrease proportionally to the number of the new users. This degradation will be equally suffered by all the participants. As a result, the negative consequences of channel overloading are diffused through all the users and no connection is refused. This gentle degradation of transmission quality can be useful when the number of users is only slightly over the limit. In that case, users do not see an appreciable decrease in the system performance. The system can absorb temporary peaks at a lower efficiency and still treat all the users equally. If the number of users continues increasing, it will eventually cause the system to stop working. The same mechanisms make it robust against noise that is picked up during the transmission.
This is the behaviour we would like to reproduce with our quantum system.
Quantum Multiple Access {#HDMA}
-----------------------
Multiple access techniques can also be applied to quantum information. Some of the existing classical infrastructure has already been used for the transmission of quantum data coming from different users. The wavelength division multiplexing experiments [@BBG03] and the proposal for FDMA of microwave frequency qubits [@OC06] are only some examples of how quantum data can be multiplexed with adapted classical architectures.
The concepts behind CDMA can be generalized to the quantum case, including intrinsically quantum aspects like entanglement, correlation erasure and superpositions. Starting from basic state swapping concepts, generic quantum multiplexer, QMUX, and demultiplexer, QDEMUX, circuits can be derived [@GC07].
These circuits convert between the two usual quantum information units, qubits and qudits. Like in the classical communication and computation domains, the most important results and implementations of quantum information processing are based on digital systems, with a discrete set of possible values. The basic quantum information unit, analogue to the classical bit, the qubit, is described as a two-level system that can be in any superposition ${|\psi\rangle}={\mu{|0\rangle}+\nu{|1\rangle}}$. $\mu$ and $\nu$ are the complex probability amplitudes associated to each value. A measurement on the qubit basis will give ${|0\rangle}$ and ${|1\rangle}$ with probabilities $|\mu|^2$ and $|\nu|^2$ respectively. These probabilities must sum to 1. An alternative encoding uses multilevelled systems. A qudit is a d-levelled system that can be in any superposition of $d$ possible values $$\sum_{i=0}^{d-1}\mu_i {|i\rangle}^d,$$ where the superscript in ${|i\rangle}^d$ gives the dimensionality of the qudit space. Again, when a measurement is taken, the probability of finding ${|i\rangle}^d$ is determined by $|\mu_i|^2$ and $$\sum_{i=0}^{d-1}|\mu_i|^2 =1.$$ Qubits are a particular case of qudits where $d=2$.
The QMUX circuit takes $N$ qubits and expresses them in a single qudit of dimension $d=2^N$ that lives in a Hilbert space larger than those of any of the individual qubits. It is this qudit which is sent. The demultiplexer will perform the inverse operation to recover the user information. The architecture adapts the generic classical model with due attention to entanglement and correlation erasure. Deleting information is not trivial in the quantum world and some precautions need to be taken before disposing of the original $N$ qubits. Figure \[MUXes\] contrasts the classical and quantum multiplexing blocks.
![Classical and quantum multiplexing.\[MUXes\]](MUXescoh.eps)
In the QMUX model, the spare qubits are taken to a known state, considered here to be ${|0\ldots0\rangle}$ without a loss of generality, which is then measured without affecting the qudit. These steps are required to avoid unwanted state reductions. The same known states can be generated later at the receiver to apply the inverse gate and recover the initial qubits at the other side of the channel.
Quantum computation with optical coherent states {#QCoherent}
================================================
Optical coherent states
-----------------------
The study of the quantum states of light has been an active field of research since the second half the twentieth century. Optical coherent states were among the first quantum states to be described [@Gla63] and have been extensively treated ever since.
### Description and properties
Some of the properties of coherent states will be important for our application. In particular [@Lou00]:
- [*Definition:*]{} Coherent states can be defined by a complex number, $\alpha$, which determines the probability amplitude coefficients of the state when written as a superposition of photon number states. $$\label{cohnum}
{|\alpha\rangle}=e^{-\frac{1}{2}|\alpha|^2}\sum_{n=0}^{\infty}\frac{\alpha^n}{\sqrt{n!}}{|n\rangle}.$$
- [*Phase and quadrature operators:*]{} An illustrative representation of coherent states can be made in terms of the complex arguments of $\alpha$. For $\alpha=|\alpha|e^{i\varphi_{\alpha}}$, we can speak of an amplitude $|\alpha|$ and a $\varphi_{\alpha}$ phase similar to those of a classical field.
Coherent states can be represented on the complex XY plane, where X and Y are the real and imaginary parts of $\alpha$ respectively. The real part is in phase with the reference phase and the Y component represents the quadrature part of the signal. Homodyne detection allows to measure these operators by adjusting the phase of a reference local oscillator.
![Representation of the ${{|\alpha\rangle}}$ coherent state in the XY complex plane.\[coherent\]](coherent.eps)
Figure \[coherent\] shows the quadrature representation of the ${{|\alpha\rangle}}$ coherent state. The state is represented as a circle due to the inherent uncertainty of quadrature measurements. The points inside the circle correspond to measurement results within the standard deviation. The centre of the circle is given by the mean values for each measurement. The measurement statistics for the $\hat{X}$ and $\hat{Y}$ observables are given by the well-known relations: $$\begin{aligned}
{\langle\alpha|}\hat{X}{{|\alpha\rangle}}&=&Re( \alpha)=|\alpha|\cos{\varphi_{\alpha}}.\\
{\langle\alpha|}\hat{Y}{{|\alpha\rangle}}&=&Im( \alpha)=|\alpha|\sin{\varphi_{\alpha}}.\\
(\Delta X)^2&=&(\Delta Y)^2=\frac{1}{4}.\end{aligned}$$ The uncertainty is constant for all coherent states, which are represented with circles of the same diameter.
- [*Overlap:*]{} Coherent states form an overcomplete set. Different coherent states are not completely orthogonal to one another, but, for a large enough distance in the XY plane, their overlap can be arbitrarily small. The overlap between two coherent states is $${\langle\alpha|\beta\rangle}=e^{-\frac{1}{2}|\alpha|^2-\frac{1}{2}|\beta|^2+\alpha^{*}\beta},$$ and the probability of taking a coherent state ${{|\alpha\rangle}}$ for another state ${|\beta\rangle}$ is related to the value of $$|{\langle\alpha|\beta\rangle}|^2=e^{-|\alpha-\beta|^2}.
\label{overl}$$
### Why coherent states?
There are a variety of reason why optical coherent states result attractive for quantum information:
- They are relatively *easy to produce* with single-mode lasers. In fact, many optical quantum computing proposals approximate number states by a coherent state.
- There exist efficient *experimental techniques* to manipulate coherent states and to perform X and Y measurement (like homodyne detection [@CLG87]). Standard optical elements such as beamsplitters and photodetectors allow for a wide range of operations.
- There are *many available almost orthogonal states* to choose from. There are infinitely many coherent states, although there will be practical restrictions like noise and power considerations that will limit the usable set. To that respect, coherent states can be related to CDMA codes.
- They have *well-known properties* and a clear interpretation. They also show quantum behaviour while retaining a classical flavour that simplifies the study.
Coherent states as qubits
-------------------------
Coherent states can be used for universal quantum computation [@JK02; @RGM03]. We will follow the ingenious model of [@RGM03] all through the paper, assuming perfect operation of all the gates described there. We will only treat superficially the implementation details, referring the interested reader to the literature on the topic. In this Section, we review the more important aspects that will be essential for our proposal and assume all the other gates and procedures of the model work perfectly.
In this version of coherent state quantum computation, logical states are encoded into coherent states so that ${|0\rangle}_L\equiv{{|-\alpha\rangle}}$ and ${|1\rangle}_L\equiv{{|\alpha\rangle}}$, with $\alpha$ being real. Although ${|0\rangle}_L$ and ${|1\rangle}_L$ are not orthogonal, the probability of a wrong identification can be made arbitrarily close to zero. Figure \[cohqubit\] shows the qubit representation in the XY plane and the values of the overlap for different choices of $\alpha$, as described by Equation (\[overl\]). Qubits can then be distinguished by homodyne in-phase detection. A negative X measurement will indicate ${|0\rangle}_L$ and a positive result can be interpreted as a ${|1\rangle}_L$ state. This classification will be correct with a high probability and errors will be more unlikely as $\alpha$ grows. This symbol detection is very similar to the 0 and 1 discrimination we saw for CDMA in Section \[CDMA\].
![Qubit representation in the XY plane (left) and the evolution of the overlap between logical ${|0\rangle}_L$ and ${|1\rangle}_L$ as a function of $\alpha$ (right).\[cohqubit\]](cohqubits.eps)
Even for relatively small values of $\alpha$, the probability of confusion is so small that can be neglected and will probably not be the dominant source of error when compared with the effects of implementation imperfections or decoherence.
The available operations on coherent state qubits have also been shown to be able to produce a universal set of quantum gates [@RGM03], i.e. a set of operations that can be combined to produce any desired quantum computation. In many cases, these gates involve going out of the qubit space.
Teleportation {#tele}
-------------
Teleportation is the basic operation in coherent state quantum computation. Many of the coherent state logic gates are based on variations of the basic scheme which can teleport a transformed version of the original qubit. Quantum teleportation [@BBC93] is one of the basic elements of quantum computation. It allows to reproduce the state of an arbitrary qubit at another location by means of shared entangled states and Bell measurements.
Figure \[teleportation\] depicts the basic coherent qubit teleportation configuration. At the input we have the qubit that is to be teleported and a Bell coherent state $\frac{{{|-\alpha\rangle}}{{|-\alpha\rangle}}+{{|\alpha\rangle}}{{|\alpha\rangle}}}{\sqrt{2}}$[^1].
![Teleportation of a coherent state qubit.\[teleportation\]](teleportation.eps)
In the scheme, Bell measurement of the joint state of the qubit and the first state of the Bell pair allows teleportation up to single qubit operations that can be later applied depending on the measured values. The X operation only requires a $\pi$ phase shifter. The classically controlled Z correction is achieved by repeated teleportation until a second sign change cancels the first one, for a probability of one half at each attempt. The coherent Bell basis is not a complete set and there will be a residual small probability of failure. In the basic Bell measurement, which is based on mixing the states in a 50% beamsplitter, this probability of error is of the order of $e^{-\alpha^2}$. Failures are heralded by a null measurement result and are related to the probability of finding zero photons in the two output ports of the beamsplitter.
Teleportation of coherent states is not only a state transfer mechanism, but can also have an error correction function. Teleportation can take the states back into the qubit space with high probability. Imagine a qubit that has deviated from the qubit space and is in the ${|(1\pm\epsilon)\alpha\rangle}$ state. Teleportation can take it back to ${{|\alpha\rangle}}$ with probability $e^{-\frac{|\alpha|^2|\epsilon|^2}{2}}$ [@RGM03]. For small deviations, there is a high probability of recovering the original coherent state.
Basic tools {#tools}
===========
In the following Section, we will describe the multiplexer and demultiplexer setups in terms of functional blocks that are independent on the actual implementation. This Section suggests some possible ways to build these blocks. The list is not exhaustive and alternative elements could be used instead.
Entangled cat states
--------------------
The most important resource in the whole coherent state computation model is a reliable source of entangled coherent states of the different required types. Production of cat states $\frac{{{|-\alpha\rangle}}+{{|\alpha\rangle}}}{\sqrt{2}}$ is far from trivial, especially for a high $\alpha$, and a perfect source has not been found yet. Nevertheless, extensive research is being conducted and good enough approximations are already available. Some of the proposals include using cross-phase modulation in dense media [@PKH03] or creating cat states from squeezed photons and detectors [@JLR05] (see [@GV08] for more alternatives). Probabilistic generation methods are also valid. The entangled resources can be generated besides the multiplexing process and be introduced from whichever source succeeds when needed.
Beamsplitters
-------------
The effect of a beamsplitter, BS, on two input coherent states can be described as ${|\alpha\rangle}{|\beta\rangle}\longrightarrow {|\cos{\theta}\alpha+\sin{\theta}\beta\rangle}{|\sin{\theta}\alpha-\cos{\theta}\beta\rangle}$. The angle $\theta$ determines the transmission and reflection coefficients, which are $\sin{\theta}$ and $\cos{\theta}$, respectively. Reflection on the lower side produces a $\pi$ phase shift. Inside optical fibre systems, optical couplers can replace beamsplitters and perform the same functions.
Beamsplitters will be the basic element in our divider, multiplier and adder proposals and can be used to transform between the different cat states we need as a resource. From a $\frac{{{|-\alpha\rangle}}+{{|\alpha\rangle}}}{\sqrt{2}}$ cat state and the vacuum, we can generate $\frac{{|-\cos{\theta}\alpha\rangle}{|-\sin{\theta}\alpha\rangle}+{|\cos{\theta}\alpha\rangle}{|\sin{\theta}\alpha\rangle}}{\sqrt{2}}$ cats. The same procedure can be repeated as many times as needed to create cats made of more than two coherent states.
Dividers
--------
In a first order approximation, a BS with a parameter $\theta=\frac{1}{M} \ll 1$ and an empty input port will induce the evolution ${|\alpha\rangle}{|0\rangle}\longrightarrow {|\alpha\rangle}{|\frac{\alpha}{M}\rangle}$. We consider that the small error taken in the approximation can be neglected. The first state has, however, an amplitude slightly smaller than $\alpha$, as energy conservation requires.
It must also be remarked that the divider is not a cloning machine. Although the outputs are independent for a single coherent state input, both output ports will become correlated for a general input that is a superposition of coherent states. In order to recover an independent evolution, we will need to introduce later additional elements to disentangle the ports.
Multipliers
-----------
A qubit can be amplified to any desired amplitude by teleportation through $\frac{{|-\alpha\rangle}{|-M\alpha\rangle}+{|\alpha\rangle}{|M\alpha\rangle}}{\sqrt{2}}$ states, which can be easily produced using a $\frac{1}{M}$ divider on a $\frac{{|-M\alpha\rangle}+{|M\alpha\rangle}}{\sqrt{2}}$ cat state. Experimentally, large cat states are more demanding to create, though.
It is important to notice that, unlike dividers, multipliers are only valid for known states with a given value of $\alpha$. The multiplier block is not a general amplifier for arbitrary superpositions of coherent states. In this case, we are applying a variation on teleportation. The setup can be adapted to multiply qudits, but, the more the possible states, the more complex the teleportation circuit and the necessary resource entangled state will be.
Adders {#adder}
------
Beamsplitters can also be used to add two coherent states. An adder is a block that produces the transformation ${|\alpha'\rangle}{|\beta'\rangle}\rightarrow{|\alpha'+\beta'\rangle}$. This operation, as we need it, is irreversible.
We will use a beamsplitter similar to the divider that we will call sum BS. The sum BS has a $\theta=\frac{\pi}{2}-\frac{1}{M^m}$, with $M,m>1$. In the range of values of $m$ and $M$ we are going to use, we can approximate the reflection coefficient of the sum BS as $\frac{1}{M^m}$ and the transmission coefficient as $1-\frac{1}{2M^{2m}}\approx1$, where we neglect all the terms above the first order. As it happened in the divider, this does not imply a violation of the conservation of energy, as the exact calculations show.
We will add two states of different amplitudes so that $\alpha'>\beta'$. The addition will have two steps. In the first one, we apply an $M^m$ multiplier on the first half ${|\alpha'\rangle}{|\beta'\rangle}$ input to produce the state ${|M^m\alpha'\rangle}{|\beta'\rangle}$. The result is mixed in a sum BS in the second step.
For a general input $$\mu_{00}{|-M^m\alpha'\rangle}{|-\beta'\rangle}+\mu_{01}{|-M^m\alpha'\rangle}{|\beta'\rangle}+\mu_{10}{|M^m\alpha'\rangle}{|-\beta'\rangle}+\mu_{11}{|M^m\alpha'\rangle}{|\beta'\rangle}$$ the sum BS presents an output
[1]{} [@clclcl]{} &$\phantom{+}\mu_{00}$&${|-\alpha'-\beta'\rangle}{|-M^m\alpha'+M^{-m}\beta'\rangle}$&$+\mu_{01}$&${|-\alpha'+\beta'\rangle}{|-M^m\alpha'-M^{-m}\beta'\rangle}$\
&$+\mu_{10}$&${|\alpha'-\beta'\rangle}{|M^m\alpha'+M^{-m}\beta'\rangle}$&$+\mu_{11}$&${|\alpha'+\beta'\rangle}{|M^m\alpha'-M^{-m}\beta'\rangle}$&,
$$\label{addereq}$$ which already has the desired sum state at the first port. Before proceeding, we need to erase the state of the second port without altering the superposition of the sum states. Erasure will be based on teleportation error correction.
The state of the second port we want to erase is ${|\pm M^m\alpha'\pm M^{-m}\beta'\rangle}$, which can be written as ${|M^m\alpha'(\pm 1\pm \frac{\beta'}{\alpha'}M^{-2m})\rangle}$ and taken as a ${|\pm M^{m}\alpha'\rangle}$ state with an error $$\label{errorsum}
\epsilon=\pm \frac{\beta'}{\alpha'} M^{-2m}.$$ For the erasure, we mix the second port with the cat state $\frac{{|-M^{m}\alpha'\rangle}+{|M^{m}\alpha'\rangle}}{\sqrt{2}}$ using a 50% beamsplitter and then measure the photon number in each port.
After the 50% beamsplitter, we obtain coherent states ${|\pm \frac{M^m\alpha'}{\sqrt{2}}(2\pm \epsilon)\rangle}{|\pm \frac{M^m\alpha'}{\sqrt{2}}\epsilon\rangle}$ and ${|\pm \frac{M^m\alpha'}{\sqrt{2}}\epsilon\rangle}{|\pm \frac{M^m\alpha'}{\sqrt{2}}(2\pm \epsilon)\rangle}$. The error correction happens when we detect zero photons at one of the detectors $D_1$ (up) or $D_2$ (down). If no photon is found, we can fairly assume that we originally had the ${|\pm \frac{M^m\alpha'}{\sqrt{2}}\epsilon\rangle}$ state. The probability of finding zero photons for the ${|\pm \frac{M^m\alpha'}{\sqrt{2}}(2\pm \epsilon)\rangle}$ state is negligibly small for our choice of $\alpha$ and $M$.
The protocol succeeds if we find zero photons in one of the detectors, which happens with probability $e^{-\frac{|M^m\alpha'|^2|\epsilon|^2}{2}}$, similarly to what we had for the teleportation error correction of Section \[tele\]. In the adder, $\alpha'>\beta'$ and we have a bounded error with $|\epsilon|< M^{-2m}$. In this case, the correction succeeds with a probability of, at least, $e^{-\frac{{\alpha'}^2M^{-2m}}{2}}$, which can be made arbitrarily close to 1 by choosing an appropriate value of $m$.
The measurement can introduce a phase shift between the sum terms. The amplitude of the ${|\pm \frac{M^m\alpha'}{\sqrt{2}}(2\pm \epsilon)\rangle}$ state has a phase of $0$ or $\pi$, depending on the sign. The detector that counts a number of photons $n_p\ne 0$ will associate a $(-1)^{n_p}$ phase to the sum states correlated to the ${|-\frac{M^m\alpha'}{\sqrt{2}}(2\pm \epsilon)\rangle}$ terms. The final state can be written, up to a global phase, as $$\mu_{00}{|-\alpha'-\beta'\rangle}+\mu_{01}{|-\alpha'+\beta'\rangle}+(-1)^{n_p}\mu_{10}{|\alpha'-\beta'\rangle}+(-1)^{n_p}\mu_{11}{|\alpha'+\beta'\rangle}.$$ If $n_p$ is odd, there will be a relative $\pi$ phase shift between the components. On average, this sign shift will happen half of the times we perform an erasure.
In the multiplexer, the result will be equivalent to a sum in which the first qubit has suffered a Z operation. Z correction can be used with a caveat: the sum state cannot be targeted by the usual coherent state teleportation Z gate lest we destroy the information of the smaller amplitude inputs, which will also carry the user’s qubits. However, correlated states can be extracted with a divider with a negligible alteration of the original state. For the values of $\alpha'$ and $\beta'$ we will have at the multiplexer, the extracted state will be dominated by the $\alpha'$ component. We can repeat the erasing procedure for the extracted state using a resource cat state of the amplitude of the leading term. Again, on an odd photon count, we will have a Z operation that corrects the first one. This is achieved with probability one half. The procedure can be repeated as many times as needed until the sum superposition has been corrected.
The adder we have presented is also valid for inputs that are, in turn, sums from previous adder stages. The results will have a leading term and smaller deviations. The approximations we have taken in the discussion are still correct as long as the leading terms in the lower port are of an amplitude smaller than the smallest term in the upper port.
![Representation for the divider (a), multiplier (b) and combiner (c) blocks.\[symbol\]](symbols.eps)
Combiners
---------
Adders can be cascaded to provide a combiner. We call combiner to any element that merges up to $N$ inputs ${|\alpha'\rangle}{|\beta'\rangle}\ldots{|\omega'\rangle}$ into a single channel with a coherent state ${|\alpha'+\beta'+\ldots+\omega'\rangle}$, where the constant amplitude correction terms needed for the conservation of energy have been dropped from the discussion for simplicity. Different arrangements inside the combiner will give different correction terms for the input states. We will consider inputs of decreasing amplitudes $\alpha'>\beta'>\ldots>\omega'$, in which we assign an index from $N-1$ to $0$ to the inputs. The $i$-th input will have amplitudes $\pm M^i\alpha$.
One way to create a combiner is chaining adders in a series configuration. First we need to provide the additional energy that will be lost in the erasure. We can include a premultiplication stage in all the qubits but the last one. A coherent qubit of order $i$, for $i$ from $1$ to $N-1$, will go through a multiplier of factor $M^{i+m}$, with an $m$ that is determined by the sum adder. Now the coherent qubits can be added starting from the qubits of order 0 and 1. Each stage produces a sum state in which the amplitude is no longer multiplied by $M^m$. The higher amplitude terms that have not been added yet still have the $M^m$ factor and enter their adders by the upper port. The series combination will sum the inputs in an increasing amplitude order. The new qubits are incorporated into the sum state one by one, loosing the $M^m$ factor in the process. After $N-1$ adding stages, we will have the desired sum state at the output.
Combination can also be done in parallel. We can pair the inputs into adders with the same strategy employed in knockout tournaments to determine the matches. Each adder corresponds to one match. This configuration has a maximum number of $\log_2{N}$ stages, with $N-1$ adders [@Knu98]. In this case, the sum states that go through more than one adder need to be premultiplied by a factor $M^{N_Am}$, where $N_A$ is the number of adders it enters through the upper port. In each adder, the sum terms of the upper port must be of a higher index than the sum terms in the lower port. Here, we have needed a more restrictive premultiplication because, after the first sum has been performed, our multiplier cannot be applied. The multiplier is not a coherent state amplifier, but specific to the concrete qubits of a determined user of index $i$. It cannot be used on sum states.
These two groups of cascaded adders form a suitable combiner. The setup can be simplified by an $N\times N$ Multi-Mode Interference, MMI, device with the same transfer matrix as the whole beamsplitter system. MMI devices can give low loss power combiners for any number of input ports [@SP95]. The $N-1$ spare ports can be dealt with in the same way as the extra output of the adder scheme. Undesired residual correlation will be erased after the teleportation-like measurement.
Quantum multiplexing with optical coherent states {#cohmux}
=================================================
Using these blocks, we can propose an architecture for quantum multiplexing with many of the characteristics of CDMA. We will show that coherent quantum multiplexers are feasible with existing technology, at least for a reasonable rate of success.
In the following discussion, we will assume that a working teleportation circuit is available with high probability and that we have an unlimited amount of resource Bell states. We will also temporarily forget of any requisite on the maximum value of $\alpha$, assuming any amplitude is possible. This assumption can later be dropped at the cost of a reduction in the global efficiency. Finally, we will usually take first order approximations and establish the order of magnitude of the involved probabilities instead of presenting rigorous, detailed calculations that would obscure the proposal.
Multiplexing
------------
The task of the multiplexer is to combine the information in a way that allows an easy separation at the receiver. The obvious solution is to use a combiner to merge the qubits into one channel. This approach has an important shortcoming: more than one state end in the same multiplexed state in the transmitted qudit. For instance, for the combiner of the previous section, the result of merging two logical ${|0\rangle}_L$ states and a logical ${|1\rangle}_L$ state would result in the same coherent state, ${{|-\alpha\rangle}}$, irrespective of which qubit contains the ${|1\rangle}_L$. Logical ${|100\rangle}_L$ would interfere with ${|010\rangle}_L$ and ${|001\rangle}_L$, which can be a problem at the reception, where separation should be straightforward. The confusion comes from the irreversibility of the combiner step. A small amount of preprocessing and planning can correct this. In our scheme, distinguishability is achieved by multiplying the coherent states of the qubits by a different factor for every user. Alternatively, users could have different values of $\alpha$ for their qubits from the start.
![Quantum multiplexer for coherent state qubits.\[cQMUX\]](cQMUX.eps)
Figure \[cQMUX\] shows the complete architecture of the multiplexer. Each qubit is expressed as $\mu_i{|-M^i\alpha\rangle}+\nu_i{|M^i\alpha\rangle}$, where $i$ is the index of the qubit position. Different qubits are encoded at different levels of amplitude and can be later separated. Now, each logical combination belongs to a different state in the qudit. For $M\ge 3$ those states are at least as separated in the XY plane as the original qubits and confusion is highly unlikely.
The resulting state will be encoded in $d=2^N$ almost orthogonal coherent states. These states define the coherent qudit ${|\psi\rangle}^{2^N\!}\!$. The superposition can be written with the formula: $${|\psi\rangle}^{2^N}=\sum_{l=0}^{2^N-1}\left( \prod_{j=0}^{N-1} \mu_j^{b_j^l\oplus 1}\nu_j^{b_j^l} \right)\left|\left( \sum_{k=0}^{N-1} (-1)^{b_k^l\oplus 1}M^k\right) \alpha \right\rangle,$$ where $b_k^l$ is the bit that describes the term $2^k$ in the binary representation of the number $l$ that is encoded by the logical qubit states and $\oplus$ represents the logical XOR operation.
The probability of success of each addition of the multiplexing process depends on the kind of combiner we use. For a series configuration, the probability of success for the addition of user $i$ is limited by the erasure procedure. We sum states $\alpha'=M^i\alpha$ and $\beta'=M^{i-1}\alpha$. From Equation (\[errorsum\]), we can see that the error we correct in the erasure is $\epsilon=\pm \frac{\beta'}{\alpha'}M^{-2m}=\pm M^{-2m-1}$. The new user will be added with probability $e^{-\frac{M^{2i-2m-1}\alpha^2}{2}}$. We can adjust the $m$ factor of each adder so that $m\ge i$. In this case, the probability of success is at least as high as the probability of error correction of an ${|\alpha\rangle}$ state with an error $\epsilon=\frac{1}{M}$, which will be a reference figure in the demultiplexer.
Demultiplexing
--------------
At the receiver, the qubits must be extracted. The main difference between quantum and classical transmission is that information cannot be copied. In our scheme, we cannot generate multiple independent copies like in CDMA. This fact, when combined with our encoding, imposes a sequential ordered recovery. We must know in advance the order in which the qubits are going to be extracted.
Figure \[cDEMUX\] shows the demultiplexing cell for the $k$-th receiver. The input is a qudit ${|\psi\rangle}^{2^{k+1}}$ with the information of all the qubits with indices from 0 to k. After the processing, the $k$-th qubit is extracted and only the information of the qubits with indices from 0 to k-1 is kept in the output qudit ${|\psi\rangle}^{2^k}$.
![Quantum demultiplexer cell for the $k$-th qubit. The qubit is decoupled from the coherent qudit, which is then put back into the channel.\[cDEMUX\]](cDEMUX.eps)
The procedure is divided into two stages. In the first one, the coherent state superposition goes through a splitter so that the state is reproduced, divided by $M^k$, in another port. This state is projected to ${|\pm\alpha\rangle}$ using a teleportation procedure that will also produce an additional output to take care of residual correlations. The first stage ends with three entangled ports, one with the original qudit and, associated to each qudit value, the corresponding coherent state of the $k$-th qubit and an ancillary state that will be used to erase entanglement. The second stage removes all correlation between qubit and qudit.
After the division by $M^k$, the terms of the superposition on the divided port are of the form $$\left|\pm\alpha+\frac{\alpha}{M^k}(\pm M^{k-1}\pm{M^{k-2}\pm\ldots\pm1})\right\rangle,$$ where the concrete signs depend on the encoded state. Each of these states is still entangled to a different qudit value. Attending to the contents of the extracted state, there are two dominant groups of states. One group corresponds to the logical ${|0\rangle}_L$, with amplitudes close to $-\alpha$ and the other corresponds to logical ${|1\rangle}_ L$, with amplitudes closer to $\alpha$. All of them can be seen as ${|\pm (1\pm \epsilon )\alpha\rangle}$ states, states in which an error has corrupted the original information. Notice that, in the worst, maximum distance, case $$\label{demuxerror}
\epsilon=\frac{1}{M^k}\sum_{i=0}^{k-1}M^i=\frac{1}{M^k}\frac{1-M^k}{1-M}=\frac{1}{M^k(1-M)}-\frac{1}{1-M}.$$ If $M\gg1$, $\epsilon$ is of the order $\frac{1}{M}$, which is also small. After the teleportation, all the states within the same group will be projected to the same final state with probability $e^{-\frac{\alpha^2}{2M^2}}$, which is close to 1. The information of the additional qubits has been treated as an error to be corrected. The effects of noise and qudit transmission errors will also be cleaned if they are of a magnitude similar to $M^k\alpha\epsilon$.
Here, the only difference with respect to basic coherent qubit teleportation is the presence of an output to undo correlations. Teleportation is based on a joint Bell measurement of the incoming state and the first element of the entangled state $$\frac{{{|-\alpha\rangle}}{|M^k\alpha\rangle}{{|-\alpha\rangle}}+{{|\alpha\rangle}}{|-M^k\alpha\rangle}{{|\alpha\rangle}}}{\sqrt{2}}.$$ After teleportation, states related to the ${|0\rangle}_L$ of the $k$-th qubit will be entangled to ${|M^k\alpha\rangle}{{|-\alpha\rangle}}$ states and qudit states related to the ${|1\rangle}_L$ will be entangled to ${|-M^k\alpha\rangle}{{|\alpha\rangle}}$ states. The ${|\pm M^k\alpha\rangle}$ states will erase the trace of the $k$-th qubit after the combiner with a probability $e^\frac{M^{2k-2m}\alpha^2}{2}$, which, for $m=i+1$, will be similar to the probability of extraction[^2].
By repeating this procedure, the $n$ qubits can be extracted from the qudit.
Scalability and users above the limit
-------------------------------------
This scheme can be easily scaled. A new user can be added at any moment just by assigning it a new power of $M$ to multiply its qubits. Theoretically, this could be done for any number of users, but practical limits will set a maximum amplitude for coherent states and a maximum $M^k\alpha$ value that can be reached.
If there is a limit around $\beta$ on the amplitude, a possible heuristic to accommodate $n$ users would be to choose $M=\left(\frac{\beta}{\alpha}\right)^{\frac{1}{N-1}}$, allowing for the smallest available value of $\epsilon$. If, later, new users must come into the system, they can still be added by defining a new $M'=\sqrt{M}$. This will create new slots to place the qubits of the new users. The price to pay is a general reduction of the efficiency. Now, some of the teleportations will reduce their probability of success from the order of $e^{-\frac{|\alpha|^2}{2M^2}}$ to $e^{-\frac{|\alpha|^2}{2M}}$.
This behaviour reproduces the gentle degradation in transmission quality of CDMA. In CDMA, almost orthogonal codes created noise that masked the legitimate signal. Here, states are closer in the XY plane and separation is progressively more difficult as new users occupy the new slots.
Example
-------
To better illustrate the protocol, we can study the multiplexing of three qubits. In this example, the values of $\alpha$ and $M$ are chosen to be round numbers for clarity, which will give coherent amplitudes too large for a real world application. We will discuss later how $\alpha$ and $M$ have to be chosen. In the example, the starting point are three independent users, with coherent states qubits with $\alpha=2$. $M$ is set to be 10.
### Multiplexing:
First, we multiply the qubits by factors of 10: $$\begin{aligned}
\nonumber
{|\psi_0\rangle}=\mu_0{|-2\rangle}+\nu_0{|2\rangle}& \stackrel{}{\longrightarrow}&\mu_0{|-2\rangle}+\nu_0{|2\rangle}.\\
\nonumber
{|\psi_1\rangle}=\mu_0{|-2\rangle}+\nu_0{|2\rangle}& \stackrel{\times 10}{\longrightarrow}&\mu_1{|-20\rangle}+\nu_1{|20\rangle}.\\
\nonumber
{|\psi_2\rangle}=\mu_0{|-2\rangle}+\nu_0{|2\rangle}& \stackrel{\times 10^2}{\longrightarrow}&\mu_2{|-200\rangle}+\nu_2{|200\rangle}.\end{aligned}$$
The first adder of the combiner will sum the qubits of users 0 and 1. If we choose a sum BS with $m=1$, we need to apply a M=10 multiplier to the qubit of user 1. At the input of the sum BS, we will have $\mu_1{|-200\rangle}+\nu_1{|200\rangle}$ at the upper port and $\mu_0{|-2\rangle}+\nu_0{|2\rangle}$ at the lower port. The output will be $$\mu_1\mu_0{|-22\rangle}{|-199.8\rangle}+\mu_1\nu_0{|-18\rangle}{|-200.2\rangle}+\nu_1\mu_0{|18\rangle}{|200.2\rangle}+\nu_1\nu_0{|22\rangle}{|199.8\rangle}.$$ The erasure succeeds with a 98% probability. The resulting state $$\mu_1\mu_0{|-22\rangle}+\mu_1\nu_0{|-18\rangle}+\nu_1\mu_0{|18\rangle}+\nu_1\nu_0{|22\rangle}$$ carries the data of both qubits.
We can now add the state of user 2. If the maximum amplitude for the coherent states is not restricted, we can directly take a sum BS of $m=2$ to create the state
[1]{} [@lclclc]{} $\mu_2\mu_1\mu_0$&${|-222\rangle}{|-19999.78\rangle}$&$+\mu_2\mu_1\nu_0$&${|-218\rangle}{|-19999.82\rangle}$&$+\mu_2\nu_1\mu_0$&${|-182\rangle}{|-20000.18\rangle}$\
+$\mu_2\nu_1\nu_0$&${|-178\rangle}{|-20000.22\rangle}$&$+\nu_2\mu_1\mu_0$&${|178\rangle}{|20000.22\rangle}$&$+\nu_2\mu_1\nu_0$&${|182\rangle}{|20000.18\rangle}$\
$+\nu_2\nu_1\mu_0$&${|218\rangle}{|19999.82\rangle}$&$+\nu_2\nu_1\nu_0$&${|222\rangle}{|-19999.78\rangle}$&.&
$$$$
The probability of success of the erasure is, for the worst case, of almost the 97,61%. However, we have needed to produce coherent state superpositions of amplitudes of the order of $10^4$. We can reduce the required amplitude if we allow for a higher rate of failure.
For instance, for $m=1.4$, the multiplication factor would be $M^m=10^{1.4}\approx 25.12$. Now, we have at the upper port of the sum BS coherent states of amplitude 5024, an order of magnitude below the amplitudes of the previous case. The probability of success in this case would be around the 68.14% for the maximum error.
In both cases we have commented, the qubits are combined into the state
---------------------------------- ------------------ ------------------- ------------------ ------------------- ------------------ ------------------- ------------------ --
$\mu_2\mu_1\mu_0$ ${|-222\rangle}$ $\mu_2\mu_1\nu_0$ ${|-218\rangle}$ $\mu_2\nu_1\mu_0$ ${|-182\rangle}$ $\mu_2\nu_1\nu_0$ ${|-178\rangle}$
$+\hspace{0.3ex}\nu_2\mu_1\mu_0$ ${|178\rangle}$ $\nu_2\mu_1\nu_0$ ${|182\rangle}$ $\nu_2\nu_1\mu_0$ ${|218\rangle}$ $\nu_2\nu_1\nu_0$ ${|222\rangle}$,
---------------------------------- ------------------ ------------------- ------------------ ------------------- ------------------ ------------------- ------------------ --
$$$$ which encodes the original tensor product. There is a one-to-one correspondence with the original logical state
-------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ---
$\mu_2\mu_1\mu_0$ ${|000\rangle}_L$ $\mu_2\mu_1\nu_0$ ${|001\rangle}_L$ $\mu_2\nu_1\mu_0$ ${|010\rangle}_L$ $\mu_2\nu_1\nu_0$ ${|011\rangle}_L$
$+\nu_2\mu_1\mu_0$ ${|100\rangle}_L$ $\nu_2\mu_1\nu_0$ ${|101\rangle}_L$ $\nu_2\nu_1\mu_0$ ${|110\rangle}_L$ $\nu_2\nu_1\nu_0$ ${|111\rangle}_L$ .
-------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ------------------- ---
$$$$
The probability of success of this reencoding depends on the maximum amplitude for which we can keep a coherent state superposition. We can adjust the parameter $m$ of the sum beamsplitters to reduce the maximum required amplitude at the cost of a lower rate of success.
### Demultiplexing:
At the receiver, we first extract ${|\psi_2\rangle}$. We start by splitting the coherent states of the superposition with a 100 divider. The resulting state is, approximately,
[1]{} [@lclclc]{} $\mu_2\mu_1\mu_0$&${|-222\rangle}{|-2.22\rangle}$&$+\mu_2\mu_1\nu_0$&${|-218\rangle}{|-2.18\rangle}$&$+\mu_2\nu_1\mu_0$&${|-182\rangle}{|-1.82\rangle}$\
$+\mu_2\nu_1\nu_0$&${|-178\rangle}{|-1.78\rangle}$&$+\nu_2\mu_1\mu_0$&${|178\rangle}{|1.78\rangle}$&$+\nu_2\mu_1\nu_0$&${|182\rangle}{|1.82\rangle}$\
$+\nu_2\nu_1\mu_0$&${|218\rangle}{|2.18\rangle}$&$+\nu_2\nu_1\nu_0$&${|222\rangle}{|2.22\rangle}$&.&
$$$$
The second coherent state can be projected into either ${|-2\rangle}$ or ${|2\rangle}$ by teleportation with a probability over the 97.60% for these values. The entangled resource state $\frac{{|-2\rangle}{|200\rangle}{|-2\rangle}+{|2\rangle}{|-200\rangle}{|2\rangle}}{\sqrt{2}}$, will guarantee that, along with the projection, an accompanying ${|\pm200\rangle}$ state is created. This state is merged with the multiplexed state using a combiner. If we choose a sum BS with $m=3$, the erasure step of the combination succeeds with a probability above the 97.56%, or above the 67.66% for a more moderate $m=2.4$.
After merging the new state with the original channel qudit, we have
-------------------- ----------------------------- ------------------- ----------------------------- ------------------- ---------------------------- ------------------- ---------------------------- --
$\mu_2\mu_1\mu_0$ ${|-22\rangle}{|-2\rangle}$ $\mu_2\mu_1\nu_0$ ${|-18\rangle}{|-2\rangle}$ $\mu_2\nu_1\mu_0$ ${|18\rangle}{|-2\rangle}$ $\mu_2\nu_1\nu_0$ ${|22\rangle}{|-2\rangle}$
$+\nu_2\mu_1\mu_0$ ${|-22\rangle}{|2\rangle}$ $\nu_2\mu_1\nu_0$ ${|-18\rangle}{|2\rangle}$ $\nu_2\nu_1\mu_0$ ${|18\rangle}{|2\rangle}$ $\nu_2\nu_1\nu_0$ ${|22\rangle}{|2\rangle}$.
-------------------- ----------------------------- ------------------- ----------------------------- ------------------- ---------------------------- ------------------- ---------------------------- --
$$$$
![Progressive reduction of the qudit space as the individual qubits are extracted. The states that encode the same logical qubit that is extracted at each step are circled by an ellipse.\[example\]](example.eps)
This state can be factored as $$(\mu_1\mu_0{|-22\rangle}+\mu_1\nu_0{|-18\rangle}+\nu_1\mu_0{|18\rangle}+\nu_1\nu_0{|22\rangle})\otimes(\mu_2{|-2\rangle}+\nu_2{|2\rangle}),$$ where the qubit ${|\psi_2\rangle}$ has been separated from the superposition.
We can proceed in the same way to extract ${|\psi_1\rangle}$. This time, we divide by 10. The resulting state is $$\mu_1\mu_0{|-22\rangle}{|-2.2\rangle}+\mu_1\nu_0{|-18\rangle}{|-1.8\rangle}+\nu_1\mu_0{|18\rangle}{|1.8\rangle}+\nu_1\nu_0{|22\rangle}{|2.2\rangle}.$$ Teleportation will correct deviations from ${|\pm2\rangle}$ with a probability slightly above 98%.
After the projection and combination of the ${|\pm20\rangle}$ state and the qudit, the global state becomes $$\mu_1\mu_0{|-2\rangle}{|-2\rangle}+\mu_1\nu_0{|2\rangle}{|-2\rangle}+\nu_1\mu_0{|2\rangle}{|2\rangle}+\nu_1\nu_0{|2\rangle}{|2\rangle}$$ which is, exactly, $$(\mu_1{|-2\rangle}+\nu_1{|2\rangle})\otimes(\mu_0{|-2\rangle}+\nu_0{|2\rangle}).$$ The combination that erases the remains of the extracted qubit is performed with a probability over the 97.60%, for an $m=1$ adder. At this point, all the original qubits are again separated.
Figure \[example\] summarizes the whole procedure in the XY plane. The multiplexed qudit can be in any superposition of a certain number of different coherent states. The set of the possible coherent states is reduced after each extraction. The qubit value is recovered by identifying logical ${|0\rangle}_L$ and ${|1\rangle}_L$ with two different sets of states that, after the qubit has been recovered, interfere to produce a superposition in the new smaller qudit space.
It is also interesting to see the effects of the addition of a new user after the system has been designed. Imagine that, once an upper limit around $\alpha=222$ has been defined, a new user wants to enter the channel. Its qubit ${|\psi_3\rangle}=\mu_3{|-2\rangle}+\nu_3{|2\rangle}$ can be multiplied by $M'=\sqrt{10}$, resulting in $\mu_3{|-\sqrt{40}\rangle}+\nu_3{|\sqrt{40}\rangle}$. The whole procedure could be repeated reserving the second channel for this new user. The extraction of qubit ${|\psi_2\rangle}$ would be slightly impaired. Now the most separated state from ${|2\rangle}$ would be ${|2.2832\rangle}$ and the probability of correction goes down from 97.60% to 96%. The probability of correct erasure at the combiner similarly reduces from the 97.56% to the 97.42%, for $m=3$, and from the 67.66% to the 66.15% for $m=2.4$. In the next steps, erasure in the combiner is not affected much by the existence of an additional user. More dangerous is the decoding of the new state. The ${|\psi_0\rangle}$ and ${|\psi_3\rangle}$ states are combined as $$\begin{aligned}
\nonumber\mu_3\mu_0{|-8.3246\rangle}{|-2.6325\rangle}&+&\mu_3\nu_0{|-4.3246\rangle}{|-1.3675\rangle}\\
+\hspace{5ex}\nu_3\mu_0{|4.3246\rangle}{|1.3675\rangle}&+&\nu_3\nu_0{|8.3246\rangle}{|2.6325\rangle}\end{aligned}$$ after the division by $M'$. Now, $\epsilon\approx 0.3162$ in the worst case and the probability of success is reduced to only the 81.87%.
With this example, it becomes clear that the new user slots should be filled first from the lower amplitude states. If an error occurs, at least the upper channels will have been transmitted correctly.
Discussion {#Conclusion}
==========
We have presented a quantum version of the CDMA multiple access technique that can be used to send $N$ coherent state qubits together as a superposition of $2^N$ coherent states. Different channels are encoded at different levels of detail in the amplitude of the transmitted state, so that all the other channels can be taken as noise. The users employ encodings that are separated by factors of a parameter $M$. The methods of coherent state qubit error correction by means of teleportation allow us to separate the users at the receiver. User information must be extracted sequentially as a consequence of the strong correlations between the qubits that form the transmitted qudit. New users can be easily added to the channel, even above first design limits, at the cost of smaller success rates. The basic resource are entangled cat states.
Coherent state multiplexing permits to send the quantum data of many users with a single transmission. Instead of sending each qubit at different times, or at different frequencies, so that they do not interfere, we can make a better use of the channel. Apart from this advantage, we have to consider the effects of decoherence on the new encoding. Transmission through a real medium, such as a lossy optical fibre, can destroy the coherence of the quantum data. In particular, superpositions of coherent states can be very sensitive to decoherence and show a decay rate which is proportional to their separation in the XY plane [@WM94]. This imposes a limit on the maximum amplitude of our multiplexed states.
Nevertheless, coherent state multiplexing can also have some benefits. The proposed scheme gives a partial protection against decoherence. The two most important effects of decoherence during the transmission of superpositions of coherent states of a real $\alpha$ are amplitude reduction and $\pi$ phase shifts, or Z errors [@GVR04]. The effect of the channel on the transmission of such states can be modelled as a beamsplitter in which the superposition is mixed with the vacuum. The transmission coefficient of the beamsplitter is determined by the channel losses. The situation is equivalent to the problem of erasure we found in Section \[adder\] when we discussed the adder.
In this model, amplitude reduction is determined by the part of the field that leaks into the medium. The teleportation extraction of the demultiplexer provides an effective regeneration. In the same way that the extraction of a qubit cleans the traces of the other channels, it will also correct the presence of noise or amplitude reductions of the same order of amplitude as the discarded data. Z errors are also related to the field that the loss beamsplitter extracts from the channel. As we have discussed for the adder, if this field is measured and an odd number of photons is found, there will be a sign shift between the terms with a positive and a negative amplitude. In the case of decoherence, we cannot know whether this Z error has occurred or not. But, as it happened in the adder, this error will only affect the leading term. The sign of the amplitude depends only on the qubit of order $N-1$. All the other qubits are protected by the highest amplitude qubit, which will take all the Z errors. After the extraction, the new leading term will be exposed to Z errors. However, in many scenarios, the multiplexed state will travel most of the distance and decoherence between the extractions at the location of each user will probably be less important. With coherent state multiplexing, we only have to worry for the dephasing of one qubit.
In the design of the system there are some tradeoffs that need to be taken into account. On one hand, large cats should be avoided, as coherent states of high intensity are challenging to obtain. Furthermore, while individual coherent states are relatively robust against losses, large cat states are very sensitive to loss and inefficient detection [@GV08]. From that point of view, the maximum amplitude value of $M^{N-1}\alpha$ should be kept small. On the other hand, it is desirable to minimize errors in the extraction. Efficiency is limited by the error correction technique, which succeeds with probability $e^\frac{-|\alpha|^2|\epsilon|^2}{2}$. $\epsilon$ has been shown to be of the order of $\frac{1}{M}$. Extraction will be more efficient for values of $M\gg \alpha$. Additionally, $\alpha$ should be high in order to easily distinguish between ${|0\rangle}_L={{|-\alpha\rangle}}$ and ${|1\rangle}_L={{|\alpha\rangle}}$ logical values of the qubit. From that point of view, large values of $\alpha$ and $M$ are preferred.
Future work must be done to carefully analyse the election of these parameters. In particular, it would be interesting to find balanced error probabilities. The given examples present almost null qubit confusion error probabilities, but appreciable error correction failure probabilities in qubit extraction. An optimum balance between these values needs to be found so that the limiting probability factor can be raised. Adaptive multiplexing, where each channel is multiplied by a different value $M_k$, which needs not to be the power of a fixed value, can also help to take the most out of the existing coherent state technology. Other adaptations of the model that could be explored are multiplexers with squeezed states or data encodings that make full use of the available state space instead of confining the data to the X quadrature.
Among the next tasks to be done in coherent multiplexing, it is an exhaustive calculation of the fidelity of the received qubits. In this paper, we have proved such a multiplexing is possible, but detailed numerical analysis of fidelity and behaviour against photon loss and other imperfections is left for another occasion. It would also be worth trying to build a simple experimental prototype, for instance for a Quantum Key Distribution application.
The presented multiplexing technique can be an interesting addition to the coherent state quantum information model. In particular, it is a useful complement to fibre transmission of quantum information in coherent states. With the present technology two or three qubits could be sent together with a reasonable probability of success. Further advances will be constrained to the creation of larger entangled coherent state resources.
Acknowledgements {#acknowledgements .unnumbered}
================
J.C. García-Escartín would like to thank Barry Sanders and Aggie Brańczyk for interesting coffee-break conversations and advice at QTS5 and EYSCQI in Valladolid and Vienna. This work has been funded by MEC and FEDER project Ref. TEC2007-67429-C02, and by JCyL, Grant No. VA001A08.
[10]{}
M. A. Nielsen and I. L. Chuang (2000) [*Quantum Computation and Quantum Information*]{}. First Ed. Cambridge University Press, Cambridge, UK.
B. Sklar (1983) [*A Structured Overview of Digital Communications- A Tutorial Review- Part II*]{}. , 21(7), 6.
B. Sklar (2001) [*Digital Communications*]{}. Second Ed. Prentice Hall, Upper Saddle River, New Jersey.
R. L. [Pickholtz]{}, D. L. [Schilling]{}, and L. B. [Milstein]{} (1982) [ *[Theory of spread-spectrum communications - A tutorial]{}*]{}. IEEE Transactions on Communications, 30, 855–884.
W. Stallings (2001) [*Wireless Communications and Networks*]{}. Prentice Hall Professional Technical Reference.
G. Brassard, F. Bussieres, N. Godbout and S. Lacroix (2003) [ *Multiuser quantum key distribution using wavelength division multiplexing*]{}. Proceedings of the SPIE Applications of Photonic Technology 6, 5260(1), 149–153.
A. Ortigosa-Blanch and J. Capmany (2006) [*Subcarrier multiplexing optical quantum key distribution*]{}. Physical Review A, 73(2), 024305.
J. C. García-Escartín and P. Chamorro-Posada (2007) [*Quantum Multiplexing for Quantum Computer Networks*]{}. quant-ph/0701145.
R. J. Glauber (1963) [*Coherent and Incoherent States of the Radiation Field*]{}. Physical Review, 131(6), 2766–2788.
R. Loudon (2000) [*The Quantum Theory of Light*]{}. Third Ed. Oxford University Press, Oxford, UK.
M. Collett, R. Loudon and C. Gardiner (1987) [*[Quantum Theory of Optical Homodyne and Heterodyne Detection]{}*]{}. Journal of Modern Optics, 34(6), 881–902.
H. Jeong and M. S. Kim (2002) [*Efficient quantum computation using coherent states*]{}. Physical Review A, 65(4), 042305.
T. C. Ralph, A. Gilchrist, G. J. Milburn, W. J. Munro and S. Glancy (2003) [*Quantum computation with optical coherent states*]{}. Physical Review A, 68(4), 042319.
C. Bennett, G. Brassard, C. Cr[é]{}peau, R. Jozsa, A. Peres and W. Wootters (1993) [*[Teleporting an unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels]{}*]{}. Physical Review Letters, 70(13), 1895–1899.
M. Paternostro, M. S. Kim and B. S. Ham (2003) [*Generation of entangled coherent states via cross-phase-modulation in a double electromagnetically induced transparency regime*]{}. Physical Review A, 67(2), 023811.
H. Jeong, A. P. Lund and T. C. Ralph (2005) [*Production of superpositions of coherent states in traveling optical fields with inefficient photon detection*]{}. Physical Review A, 72(1), 013801.
S. Glancy and H. M. de Vasconcelos (2008) [*Methods for producing optical coherent state superpositions*]{}. Journal of the Optical Society of America B, No. 5, Vol. 25, 712-733.
D. E. Knuth (1998) [*The art of computer programming, volume 3: sorting and searching*]{}. Second Ed. Addison Wesley Longman Publishing Co., Inc., Redwood City, CA, USA.
L. Soldano and E. Pennings (1995) [*[Optical multi-mode interference devices based on self-imaging: principles and applications]{}*]{}. Lightwave Technology, Journal of, 13(4), 615–627.
D. F. Walls and G. J . Milburn (1994) [*Quantum Optics*]{}. Springer-Verlag, Berlin, Germany.
S. Glancy, H. M. Vasconcelos and T. C. Ralph (2004) [*Transmission of optical coherent-state qubits*]{}. Physical Review A, 70(2), 022317.
[^1]: Strictly, this state is not a unit vector. Its components are not completely orthogonal and the cross terms will not be zero. Therefore, the normalization factor is not exactly $\sqrt{2}$. Even so, it will be a good approximation for our range of values of $\alpha$ and a $\sqrt{2}$ factor will be used in the rest of the paper.
[^2]: In the adder, the ${|\pm M^k \alpha\rangle}$ states will be multiplied by $M^m$ and added to the multiplexed state. The probability of success can be deduced from Equation (\[errorsum\]) considering $\alpha'=M^k\alpha$ and $\beta'=M^k(1\pm \epsilon)$, for the $\epsilon$ of Equation (\[demuxerror\]). In this case, $\frac{\beta'}{\alpha'}$ tends to 1 and the error to be corrected in the erasure is, approximately, $M^{-2k}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
bibliography:
- 'fromFase19-2.bib'
title: Assurance via Workflow$^+$ Modelling and Conformance
---
[^1]
Assurance, Safety case, Workflow modelling and conformance, Metamodelling
[^1]: Paper deadline is April 29 “Foundations” and “Practice and Innovation” papers must not exceed 10 pages (including figures and appendices, excluding references). All papers may include 1 additional page for references.
|
{
"pile_set_name": "ArXiv"
}
|
---
address: |
Instituut-Lorentz, Leiden University, P.O. Box 9506, 2300 RA Leiden, The Netherlands\
[()]{} \
author:
- 'P. W. Brouwer, P. G. Silvestrov,[^1] and C. W. J. Beenakker'
title: Theory of directed localization in one dimension
---
In a recent Letter,[@HatanoNelson] Hatano and Nelson have demonstrated the existence of a mobility edge in a disordered ring with an imaginary vector potential. A non-Hermitian Hamiltonian containing an imaginary vector potential arises from the study of the pinning of vortices by columnar defects in a superconducting cylinder.[@NelsonVinokur] Their discovery of a delocalization transition in one and two-dimensional systems has generated considerable interest,[@Efetov; @FeinbergZee; @Chalker] since all states are localized by disorder in one and two dimensions if the vector potential is real. Localization in this specific kind of non-Hermitian quantum mechanics is referred to as “directed localization”,[@Efetov] because the imaginary vector potential breaks the symmetry between left-moving and right-moving particles, without breaking time-reversal symmetry.
The analytical results of Ref. consist of an expression for the mobility edge plus a solution of the one-dimensional problem with a single impurity. Here we go further, by solving the many-impurity case in one dimension. Most of the technical results which we will need were derived previously in connection with the problem of localization in the presence of an imaginary [*scalar*]{} potential. Physically, these two problems are entirely different: an imaginary vector potential singles out a direction in space, while an imaginary scalar potential singles out a direction in time: A negative imaginary part of the scalar potential corresponds to absorption and a positive imaginary part to amplification. One might surmise that amplification could cause a delocalization transition, but in fact all states remain localized in one dimension in the presence of an imaginary scalar potential.[@Zhang; @Paasschens]
Following Ref. we consider a disordered chain with single-particle Hamiltonian $${\cal H} = -{w \over 2} \sum_{j} \left(e^{ha} c^{\dagger}_{j+1} c^{\vphantom{\dagger}}_{j} + e^{-ha} c^{\dagger}_{j} c^{\vphantom{\dagger}}_{j+1} \right) + \sum_{j} V_{j}^{\vphantom{\dagger}} c^{\dagger}_{j} c^{\vphantom{\dagger}}_{j}. \label{eq:Hlattice}$$ The operators $c^{\dagger}_{j}$ and $c^{\vphantom{\dagger}}_{j}$ are creation and annihilation operators, $a$ is the lattice constant, and $w$ the hopping parameter. The random potential $V_{j}$ is chosen independently for each site, from a distribution with zero mean and variance $u^2$. For weak disorder (mean free path much larger than the wavelength), higher moments of the distribution of $V_{j}$ are not relevant. The Hamiltonian is non-Hermitian because of the real parameter $h$, corresponding to the imaginary vector potential. The chain of length $L$ is closed into a ring, and the problem is to determine the eigenvalues $\varepsilon$ of ${\cal H}$. If $\varepsilon$ is an eigenvalue of ${\cal H}$, than also $\varepsilon^{*}$ is one — because ${\cal H}$ is real. Real $\varepsilon$ correspond to localized states, while complex $\varepsilon$ correspond to extended states.[@HatanoNelson]
To solve this problem, we reformulate it in terms of the $2 \times 2$ transfer matrix $M_{h}(\varepsilon)$ of the chain, which relates wave amplitudes at both ends.[@DorokhovAB] The energy $\varepsilon$ is an eigenvalue of ${\cal H}$ if and only if $M_{h}(\varepsilon)$ has an eigenvalue $1$. The use of the transfer matrix is advantageous, because the effect of the imaginary vector potential is just to multiply $M$ with a scalar, $$M_h(\varepsilon) = e^{h L} M_0(\varepsilon). \label{eq:Mh}$$ The energy spectrum is therefore determined by $$\det \left[1 - e^{h L} M_0(\varepsilon) \right] = 0. \label{eq:detM}$$ Time-reversal symmetry implies $\det M_0 = 1$. Hence the determinantal equation (\[eq:detM\]) is equivalent to[@footnoteHJT] $$\mbox{tr}\, M_0(\varepsilon) = 2 \cosh hL. \label{eq:Mcond}$$ We seek the solution in the limit $L \to \infty$.
Since $M_0$ is the transfer matrix in the absence of the imaginary vector potential ($h=0$), we can use the results in the literature on localization in conventional one-dimensional systems (having an Hermitian Hamiltonian).[@Pendry] The four matrix elements of $M_0$ are given in terms of the reflection amplitudes $r$, $r'$ and the transmission amplitude $t$ by $$\begin{array}{ll}
(M_0)_{11} = -(1/t) \det S, & (M_0)_{12} = r'/t, \\
(M_0)_{21} = -r/t, & (M_0)_{22} = 1/t,
\end{array}$$ where $\det S = r r' - t^2$ is the determinant of the scattering matrix. (There is only a single transmission amplitude because of time-reversal symmetry, so that transmission from left to right is equivalent to transmission from right to left.) The transmission probability $T = |t|^2$ decays exponentially in the large-$L$ limit, with decay length $\xi$: $$- \lim_{L \to \infty} L^{-1} \ln T = \xi^{-1}. \label{eq:TL}$$ The energy dependence of $\xi$ is known for weak disorder, such that $|k|\xi \gg 1$, where the complex wavenumber $k$ is related to $\varepsilon$ by the dispersion relation $$\varepsilon = -w \cos ka. \label{eq:dispersion}$$ For real $k$, the decay length is the localization length $\xi_0$, given by [@DorokhovPRB] $$\xi_0 = a (w/u)^2 \sin^2 (\mbox{Re}\, ka). \label{eq:xi0}$$ (Since $\xi_0$ is of the order of the mean free path $\ell$, the condition of weak disorder requires $\ell$ large compared to the wavelength.) For complex $k$, the decay length is shorter than $\xi_0$, regardless of the sign of $\mbox{Im}\, k$, according to [@Paasschens] $$\xi^{-1} = \xi_0^{-1} + 2 |\mbox{Im}\, k|. \label{eq:xicomplex}$$
We use these results to simplify Eq. (\[eq:Mcond\]). Upon taking the logarithm of both sides of Eq. (\[eq:Mcond\]), dividing by $L$ and taking the limit $L \to \infty$, one finds $$|h| - \case{1}{2}\xi^{-1} = \lim_{L \to \infty} L^{-1} \ln|1 - \det S|, \label{eq:trM}$$ where we have used that $L^{-1} \ln f \to L^{-1} \ln |f|$ as $L \to \infty$ for any complex function $f(L)$. For complex $k$, the absolute value of $\det S$ is either $< 1$ (for $\mbox{Im}\, k > 0$) or $> 1$ (for $\mbox{Im}\, k < 0$). As a consequence, $\ln|1-\det S|$ remains bounded for $L \to \infty$, so that the right-hand-side of Eq. (\[eq:trM\]) vanishes. Substituting Eq. (\[eq:xicomplex\]), we find that complex wavenumbers $k$ satisfy $$|\mbox{Im}\, k| = |h| - \case{1}{2} \xi_0^{-1}. \label{eq:Imk}$$ Together with the expression (\[eq:xi0\]) for the localization length $\xi_0$, this is a relation between the real and imaginary parts of the wavenumber. Using the dispersion relation (\[eq:dispersion\]), and noticing that the condition $|k| \xi \gg 1$ for weak disorder implies $|\mbox{Im}\, k| \ll |\mbox{Re}\, k|$, we can transform Eq. (\[eq:Imk\]) into a relation between the real and imaginary parts of the energy, $$|\mbox{Im}\, \varepsilon| =
% {w^2 |h| a - |h| a (\mbox{Re}\, \varepsilon)^2 - \case{1}{2} u^2 \over
% \left[ w^2 - (\mbox{Re}\, \varepsilon)^2 \right]^{{1/2}}}
|h| a \sqrt{ w^2 - (\mbox{Re}\, \varepsilon)^2 } - {u^2 \over 2 \sqrt{ w^2 - (\mbox{Re}\, \varepsilon)^2}}.
\label{eq:spectrum}$$
The support of the density of states in the complex plane consists of the closed curve (\[eq:spectrum\]) plus two line segments on the real axis,[@footnote] extending from the band edge $\pm w$ to the mobility edge $\pm \varepsilon_c$. The real eigenvalues are identical to the eigenvalues at $h=0$, up to exponentially small corrections. The energy $\varepsilon_c$ is obtained by putting $\mbox{Im}\, \varepsilon = 0$ in Eq. (\[eq:spectrum\]), or equivalently be equating[@HatanoNelson] $2 \xi_0$ to $1/|h|$, hence $$\varepsilon_c = (w^2 - u^2/2 |h| a)^{1/2}.$$ The delocalization transition at $\varepsilon_c$ exists for $|h| > h_c = \case{1}{2} u^2/w^2 a$.
=0.89
=0.89
In Fig. \[fig:1\]a, the analytical theory is compared with a numerical diagonalization of the Hamiltonian (\[eq:Hlattice\]). The numerical finite-$L$ results are consistent with the large-$L$ limit (dashed curve). To leading order in $1/L$, fluctuations of $\mbox{Im}\, \varepsilon$ around the large-$L$ limit (\[eq:spectrum\]) are governed by fluctuations of the transmission probability $T$. (Fluctuations of $L^{-1} \ln T$ are of order $L^{-1/2}$, while the other fluctuating contributions to Eq. (\[eq:Mcond\]) are of order $L^{-1}$.) The variance of $\ln T$ for large $L$ is known,[@Misirpashaev] $$\mbox{var}\, \ln T = {2L \over \xi_0} + 8L |\mbox{Im}\, k| e^{4 \xi_0 |{\rm Im}\, k|} \mbox{Ei}\,(-4 \xi_0 |\mbox{Im}\, k|),$$ where $\mbox{Ei}$ is the exponential integral. Equating $|\mbox{Im}\, k| = |h| + \case{1}{2} L^{-1} \ln T$, we find $\mbox{var}\, |\mbox{Im}\, k| = \case{1}{4} L^{-2}\, \mbox{var}\, \ln T$ and thus $$\mbox{var}\, |\mbox{Im}\, \varepsilon| = {a^2 \over 2 L \xi_0}
\left[ {1 + 2 \gamma e^{2 \gamma} \mbox{Ei}\,(-2 \gamma)} \right] \left[ {w^2 - (\mbox{Re}\, \varepsilon)^2} \right], \label{eq:var}$$ where $\gamma = 2 |h| \xi_0 - 1$. In Fig. \[fig:1\]b we see that Eq. (\[eq:var\]) agrees well with the results of the numerical diagonalization. The fluctuations $\Delta\, \mbox{Im}\, \varepsilon$ are correlated over a range $\Delta \mbox{Re}\, \varepsilon$ which is large compared to $\Delta \mbox{Im}\, \varepsilon$ itself,[@footnoteP] their ratio $\Delta\, \mbox{Im}\, \varepsilon/\Delta\, \mbox{Re}\, \varepsilon$ decreasing $\propto L^{-1/2}$. This explains why the complex eigenvalues for a specific sample appear to lie on a smooth curve (see Fig. \[fig:1\]a). This curve is sample-specific and fluctuates around the large-$L$ limit (\[eq:spectrum\]).
In conclusion, we have presented an analytical theory for the delocalization transition in a single-channel disordered wire with an imaginary vector potential. We find good agreement with numerical diagonalizations, both for the relation between the real and imaginary parts of the energy in the infinite-length limit and for the finite-size effects. In the absence of the imaginary vector potential, the transfer matrix approach used in this paper has been very successful for the study of localization in disordered wires with more than one scattering channel. We expect that such an extension of the theory is possible for non-Hermitian systems as well.
Discussions with K. B. Efetov and D. R. Nelson motivated us to do this work. We have also benefitted from discussions with T. Sh. Misirpashaev. Support by the Dutch Science Foundation NWO/FOM is gratefully acknowledged.
[*Note added:*]{} Just before submitting our paper, we learned of a different analytical approach to this problem by Janik et al.[@Janik]
N. Hatano and D. R. Nelson, Phys. Rev. Lett. [**77**]{}, 570 (1996). D. R. Nelson and V. Vinokur, Phys. Rev. B [**48**]{}, 13060 (1993). K. B. Efetov, preprint (cond-mat/9702091). J. Feinberg and A. Zee, preprint (cond-mat/9704191). J. T. Chalker and Z. J. Wang, preprint (cond-mat/9704198). Z.-Q. Zhang, Phys. Rev. B [**52**]{}, 7960 (1995). J. C. J. Paasschens, T. Sh. Misirpashaev, and C. W. J. Beenakker, Phys. Rev. B [**54**]{}, 11887 (1996). O. N. Dorokhov, Zh. Eksp. Teor. Fiz. [**101**]{}, 966 (1992) \[Sov. Phys. JETP [**74**]{}, 518 (1992)\]. J. B. Pendry, Adv. Phys. [**43**]{}, 461 (1994). O. N. Dorokhov, Phys. Rev. B [**37**]{}, 10526 (1988). Eq. (\[eq:Mcond\]) is equivalent to $$\prod_{j} \left[ {2 \over w} (\varepsilon_j^0 - \varepsilon)\right] =
2 \cosh hL - 2,$$ where the product is over all eigenvalues $\varepsilon_j^0$ of ${\cal H}$ at $h=0$. This formula can be used for an alternative derivation of our results, and can be derived from a site representation of the transfer matrix $M_0$. See D. J. Thouless, J. Phys. C [**5**]{}, 77 (1972). One might wonder whether the line segments at real $\varepsilon$ could extend into the interior of the curve (\[eq:spectrum\]) of complex $\varepsilon$. This is not possible in the large-$L$ limit since $|h| - 1/2 \xi_0 > 0$ for complex $\varepsilon$ \[see Eq. (\[eq:Imk\])\], while real eigenvalues of ${\cal H}$ appear only if $|h| - 1/2 \xi_0 < 0$. This last inequality follows from Eq. (\[eq:trM\]), which for real $\varepsilon$ equates $|h| - 1/2 \xi_0$ to a quantity $L^{-1} \ln|1 - \det S|$ which oscillates between $-\infty$ and $0$ as $L \to \infty$. T. Sh. Misirpashaev, J. C. J. Paasschens, and C. W. J. Beenakker, Physica A [**236**]{}, 189 (1997). An estimate for $(\Delta\, \mbox{Im}\, \varepsilon/\Delta\, \mbox{Re}\, \varepsilon)^2$ follows from the variance of the slope $d\, \mbox{Im}\, \varepsilon/d\, \mbox{Re}\, \varepsilon$, which for large $L$ is given by $$\mbox{var}\, {d\, \mbox{Im}\, \varepsilon \over d\, \mbox{Re}\, \varepsilon} = {a \over 2 \pi^2 L\, |\mbox{Im}\, \varepsilon|} \sqrt{w^2 - (\mbox{Re}\, \varepsilon)^2}.$$ R. Janik, M. A. Nowak, G. Papp, and I. Zahed, preprint (cond-mat/9705098).
[^1]: On leave of absence from Budker Institute of Nuclear Physics, Novosibirsk, Russia.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'When a single cell senses a chemical gradient and chemotaxes, stochastic receptor-ligand binding can be a fundamental limit to the cell’s accuracy. For clusters of cells responding to gradients, however, there is a critical difference: even genetically identical cells have differing responses to chemical signals. With theory and simulation, we show collective chemotaxis is limited by cell-to-cell variation in signaling. We find that when different cells cooperate the resulting bias can be much larger than the effects of ligand-receptor binding. Specifically, when a strongly-responding cell is at one end of a cell cluster, cluster motion is biased toward that cell. These errors are mitigated if clusters average measurements over times long enough for cells to rearrange. In consequence, fluid clusters are better able to sense gradients: we derive a link between cluster accuracy, cell-to-cell variation, and the cluster rheology. Because of this connection, increasing the noisiness of individual cell motion can actually [*increase*]{} the collective accuracy of a cluster by improving fluidity.'
author:
- 'Brian A. Camley'
- 'Wouter-Jan Rappel'
title: 'Cell-to-cell variation sets a tissue-rheology-dependent bound on collective gradient sensing'
---
Many cells follow signal gradients to survive or perform their functions, including white blood cells finding a wound, cells crossing a developing embryo, and cancerous cells migrating from tumors. Chemotaxis, sensing and responding to chemical gradients, is crucial in all of these examples [@swaney2010eukaryotic; @levine2013physics]. Chemotaxis is traditionally studied by exposing single cells to gradients – but cells often travel in groups, not singly [@hakim2017collective; @camley2017physical]. Collective cell migration is essential to development and metastasis [@friedl2009collective], and can have remarkable effects on chemotaxis. Even when single cells cannot sense a gradient, a cluster of cells may cooperate to sense it. While collective chemotaxis is our primary focus, this “emergent” gradient sensing is found in response to many signals, including soluble chemical gradients (chemotaxis) [@theveneau2010collective; @malet2015collective; @ellison2016cell], conditioned substrates (haptotaxis) [@winklbauer1992cell], substrate stiffness gradients (durotaxis) [@sunyer2016collective] and electrical potential (galvanotaxis) [@li2012cadherin; @lalli2015collective].
Cells can cooperate to sense gradients – but the physical principles limiting a cluster’s sensing accuracy are not settled. For single cells, the fundamental bounds on sensing chemical concentrations and gradients are well-studied [@berg1977physics; @kaizu2014berg; @hu2010physical; @hu2011geometry; @endres2009maximum; @endres2008accuracy; @fuller2010external; @ueda2007stochastic; @andrews2007information; @bialek2005physical], showing unavoidable stochasticity in receptor-ligand binding limits chemotactic accuracy. Is this true for cell clusters? Is a cell cluster simply equivalent to a larger cell? No! There is an essential difference between many clustered cells and a single large cell: even clonal populations of cells can have highly variable responses to signals, due to many factors, including intrinsic variations in regulatory protein concentrations [@swain2002intrinsic; @niepel2009non; @sigal2006variability]. These cell-to-cell variations (CCV) can be persistent over timescales much larger than the typical motility timescale of the cell [@sigal2006variability]. CCV has not been addressed in models of collective chemotaxis and it is not clear whether collective gradient sensing is limited by CCV or by stochastic receptor-ligand binding [@malet2015collective; @camley2016emergent; @camley2016collective; @varennes2016collective; @ellison2016cell; @mugler2016limits; @cai2016modeling].
Using a combination of analytics and simulations, we show that unless CCV is tightly controlled, collective guidance of a cluster of cells is limited by these variations: gradient sensing is biased toward cells with intrinsically strong responses. This bias swamps the effects of stochastic ligand-receptor binding. Cell clusters may reduce this error by time-averaging their gradient measurements only if the cells rearrange their positions, creating an unavoidable link between the [*mechanics*]{} of the cell cluster and its [*gradient sensing ability*]{}. As a result, surprising new tradeoffs arise: clusters must balance using motility to follow a biased signal with using motility to reduce error, and compromise between reducing noise and increasing cluster fluidity.
Gradient sensing error is dominated by cell-to-cell variation, not receptor noise
=================================================================================
{width="180mm"}
We study a two-dimensional model of gradient sensing with CCV and ligand-receptor dynamics where cells sense a chemoattractant with concentration gradient ${\ensuremath{{ {\mathbf g}}}\xspace}$. Each cell at position ${\ensuremath{{ {\mathbf r}}}\xspace}$ measures local concentration, $c({\ensuremath{{ {\mathbf r}}}\xspace}) = c_0 (1 + {\ensuremath{{ {\mathbf g}}}\xspace}\cdot {\ensuremath{{ {\mathbf r}}}\xspace})$, via ligand-receptor binding, which is stochastic. This noise leads to unavoidable errors in the cluster’s estimate of ${\ensuremath{{ {\mathbf g}}}\xspace}$. In addition, even if concentration is perfectly sensed, each cell responds differently to a fixed $c$, which models known CCV in signal response [@wang2012diverse; @samadani2006cellular]. As a result, when the cluster combines measurements from its cells, it may develop a drift in the direction of stronger-responding cells ([Fig. ]{}\[fig:ccv\_illus\]). To combine these effects, we specify the “measured” signal in cell $i$, $M_i$, which is what the cluster believes the chemoattractant signal in cell $i$ to be, including ligand-receptor binding and CCV: $$M^i = \left[c({\ensuremath{{ {\mathbf r}}}\xspace}^i) + \delta c^i \eta^i\right]/\bar{c} + \Delta^i \label{eq:M}$$ where $\bar{c}$ is the mean concentration over the cluster, $\bar{c} = N^{-1} \sum_i c({\ensuremath{{ {\mathbf r}}}\xspace}^i)$, $\eta^i$ are uncorrelated Gaussian noises with zero mean and unit variance and $\Delta^i$ are uncorrelated Gaussian noises with zero mean and variance ${\ensuremath{\sigma_{\Delta}}}^2$. Stochastic fluctuations in ligand-receptor binding are taken into account in the term $\delta c^i$, where $(\delta c^i / c({\ensuremath{{ {\mathbf r}}}\xspace}^i))^2 = \frac{1}{n_r} \frac{(c^i+K_D)^2}{c^i K_D}$. This is the error in concentration sensing from a single snapshot of $n_r$ receptors with simple ligand-receptor kinetics and dissociation constant $K_D$ ([@kaizu2014berg], Appendix \[app:conc\]). [Eq. ]{}\[eq:M\] assumes that cell-cell variance additively corrupts the measurement of the concentration $c({\ensuremath{{ {\mathbf r}}}\xspace})$ [*after*]{} an adaptation to the overall level of signal across the cluster $\bar{c}$. This is natural if the primary cell-to-cell variation is [*downstream*]{} of adaptation, as found to be a reasonable model in [@wang2012diverse].
To determine gradient sensing accuracy, we perform maximum likelihood estimation (MLE) of ${\ensuremath{{ {\mathbf g}}}\xspace}$ in [Eq. ]{}\[eq:M\], as in past approaches for single cell gradient sensing [@hu2011geometry]. We obtain the MLE estimator ${\ensuremath{\hat{{ {\mathbf g}}}}\xspace}$ numerically ([*Methods*]{}, Appendix \[app:mle\]), and thus the uncertainty ${\ensuremath{\sigma_{\mathbf{g}}}}^2 \equiv {\ensuremath{\left\langle |\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} - {\ensuremath{{ {\mathbf g}}}\xspace}|^2 \right\rangle}}$ ([Fig. ]{}\[fig:ccv\_illus\]b, symbols), where $\langle \cdots \rangle$ is an average over CCV and ligand-receptor binding. For fixed and roughly circular (isotropic) cluster geometry, if the concentration change across the cluster is small, $g R_\textrm{cluster} \ll 1$, ${\ensuremath{\sigma_{\mathbf{g}}}}^2$ can be approximated by assuming $\delta c^i$ is constant across the cluster, resulting in $${\ensuremath{\left\langle |\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} - {\ensuremath{{ {\mathbf g}}}\xspace}|^2 \right\rangle}} \approx \frac{2}{\chi} \left({\ensuremath{\sigma_{\Delta}}}^2 + \frac{1}{n_r} \frac{(\bar{c}+K_D)^2}{\bar{c} K_D} \right) \label{eq:combined_error}$$ Here, $\chi = \frac{1}{2}\sum_i |{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}|^2$ is a shape parameter, and ${\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}= { {\mathbf r}}^i - { {\mathbf r}}_{\textrm{cm}}$ is cell position relative to cluster center of mass. Evaluating this expression reveals that it is an excellent approximation to the numerically-obtained uncertainty (dashed lines, [Fig. ]{}\[fig:ccv\_illus\]b).
The approximate expression for the uncertainty, Eq. \[eq:combined\_error\], allows us to quantify the relative contribution of receptor-ligand fluctuations and CCV to the gradient sensing error. For background concentrations $\bar{c}$ near the receptor-ligand equilibrium constant $K_D$ and for typical receptor numbers in eukaryotic cells ($n_r \sim 10^5$ [@hesselgesser1998identification; @macdonald2008heterogeneity]), $\delta c / \bar{c}$ can be smaller than 0.01. Protein concentrations, on the other hand, often vary between cells to 10%-60% of their mean [@niepel2009non] – hence we estimate ${\ensuremath{\sigma_{\Delta}}}\approx 0.1-0.6$. Thus, we expect CCV to dominate gradient sensing error and that the error from concentration sensing and receptor binding can be neglected completely if ${\ensuremath{\sigma_{\Delta}}}> 0.1$ ([Fig. ]{}\[fig:ccv\_illus\]b). [Eq. ]{}\[eq:combined\_error\] also reveals that CCV masks the impact of changing background concentration. When ${\ensuremath{\sigma_{\Delta}}}= 0$, gradient sensing is limited by ligand-receptor fluctuations, and increases as $\bar{c}$ moves away from $K_D$ ([Fig. ]{}\[fig:ccv\_illus\]d) – accuracy decreases if either few receptors are bound, or if receptors are saturated. As CCV increases, ${\ensuremath{\sigma_{\mathbf{g}}}}^2$ no longer depends strongly on $\bar{c}$ ([Fig. ]{}\[fig:ccv\_illus\]d). Finally, Eq. \[eq:combined\_error\] shows that gradient sensing accuracy depends on the shape parameter $\chi$ and, therefore, on cluster size. For hexagonally packed clusters of cells with unit spacing[^1], a cluster with $Q$ layers has $N = 1 + 3Q + 3Q^2$ cells and $\chi(Q) = (5/8) Q^4 + (5/4) Q^3 + (7/8) Q^2 + (1/4) Q$ (Appendix \[app:Q\]), i.e. $\chi(Q) \sim Q^4 \sim N^2$. Clusters of increasing size then have an error that decreases as $1/N^2$ ([Fig. ]{}\[fig:ccv\_illus\]c); [this scaling is similar to earlier results for single cells (Appendix \[app:Q\]).]{}
Reducing estimation error by time-averaging {#sec:timeaverage}
===========================================
If a cluster made $n$ independent measurements, it could reduce ${\ensuremath{\sigma_{\mathbf{g}}}}^2$ by a factor of $n$. In single-cell gradient sensing, independent measurements can be made by averaging over time – improving errors by a factor $\sim T/\tau_{\textrm{corr}}$, where $T$ is the averaging time, and $\tau_{\textrm{corr}}$ the measurement correlation time. At first glimpse, time averaging seems unlikely to help with CCV, when correlation times for protein levels can be longer than cell division times, reaching 48 hours in human cells [@sigal2006variability]. However, since gradient sensing bias from CCV depends on the locations of strong- and weak-signaling cells within the cluster, time averaging can be successful if it is over a time long enough for the cluster to re-arrange. This is true even if, as we initially assume, CCV biases $\Delta^i$ are time-independent. We expect gradient sensing error with time averaging, ${\ensuremath{\sigma_{\mathbf{g},T}}}^2$, will decrease by a factor of $T/\tau_r$ from ${\ensuremath{\sigma_{\mathbf{g},0}}}^2$, where $\tau_r$ is a correlation time related to cell positions ([Fig. ]{}\[fig:timeaverage\]). Is this true, and how should we define $\tau_r$?
Our earlier results suggest that CCV dominates the gradient sensing error. Ligand-receptor noise will also be even less relevant in the presence of time averaging, as the receptor relaxation time (seconds to minutes [@wang2007quantifying]) is much faster than that for cluster re-arrangement (tens of minutes or longer). We therefore completely neglect ligand-receptor binding fluctuations, allowing an analytical solution for the MLE estimator ${\ensuremath{\hat{{ {\mathbf g}}}}\xspace}$ ([*Methods*]{}, Appendix \[app:mle\]).
![. a) Schematic drawing of how cell-cell rearrangement can change bias due to CCV. Shades of gray indicate measured signal $M$; a cell with strong response (marked with $X$) moves through the cluster, leading to biases in gradient estimate (blue arrow). The characteristic relaxation time for this bias is $\tau_r$ (see text). b) This leads to a link between the timescale $\tau_r$, which is a measure of the cluster’s rheology, and chemotactic accuracy ${\ensuremath{\sigma_{\mathbf{g},T}}}$ (box). c) Different re-arrangement mechanisms will depend on cluster size in different ways (see text and Appendix \[app:mechanisms\]). []{data-label="fig:timeaverage"}](rearrangement_variation_drawing){width="90mm"}
How much does time-averaging reduce error? If we average the MLE estimate $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}}$ over a time $T$ by applying a kernel $K_T(t)$, i.e., we define ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}(t) \equiv \int_{-\infty}^{\infty} {\ensuremath{\hat{{ {\mathbf g}}}}\xspace}(t') K_T(t-t') dt'$ and ${\ensuremath{\sigma_{\mathbf{g},T}}}^2 \equiv {\ensuremath{\left\langle |{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}-{ {\mathbf g}}|^2 \right\rangle}}$, we can derive (Appendix \[app:timeaverage\]) $${\ensuremath{\sigma_{\mathbf{g},T}}}^2 = {\ensuremath{\sigma_{\mathbf{g},0}}}^2 \times \int_{-\infty}^\infty \frac{d\omega}{2\pi} |K_T(\omega)|^2 C_{rr}(\omega) \label{eq:snrt}$$ where $C_{rr}(t'-t'') \equiv \langle \boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}(t')\cdot\boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}(t'') \rangle / \langle |\boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}|^2 \rangle$ is the normalized cell position-position correlation function, $C_{rr}(\omega)$ its Fourier transform, and ${\ensuremath{\sigma_{\mathbf{g},0}}}^2 = 2 {\ensuremath{\sigma_{\Delta}}}^2/\chi$ is the error in the absence of time-averaging. To derive [Eq. ]{}\[eq:snrt\], we make two approximations: 1) the cluster has a constant and isotropic shape, and 2) re-arrangement of cell positions relative to the center of mass is independent of the particular values of $\Delta$. The first approximation is not necessary, but is a useful simplification; a generalized result is given in Appendix \[app:timeaverage\]. [The second approximation assumes that averaging over CCV and averaging over cell positions are independent. This decoupling approximation is necessary to characterize cluster fluidity and mechanics separately from the details of signaling. It excludes, e.g. models where cells with larger-than-average $\Delta$ sort out from the cluster. We will discuss potential errors due to this approximation later in the paper.]{}
For exponential position-position correlation functions and averaging, $C_{rr}(t) = \exp(-t/\tau_r)$ and $K_T(t) = \theta(t) \frac{1}{T} e^{-t/T}$, where $\theta(t)$ is the Heaviside step function, [Eq. ]{}\[eq:snrt\] is simple: $${\ensuremath{\sigma_{\mathbf{g},T}}}^2 = \frac{{\ensuremath{\sigma_{\mathbf{g},0}}}^2}{1+T/\tau_r} \label{eq:snrt_simp}$$ In other words, gradient sensing accuracy can be improved by taking $T/\tau_r$ independent measurements in a time $T$. Crucial in this reduction is the position-position correlation time $\tau_r$ which depends on the cluster rearrangement mechanism. Two natural mechanisms are persistent cluster rotation and neighbor re-arrangements within the cluster ([Fig. ]{}\[fig:timeaverage\]c). These mechanisms may coexist, as when cells slide past one another during cluster rotation [@rappel1999self]. $\tau_r$ can depend on cluster size; for diffusive rearrangements, we expect $\tau_r \sim R^2/D_{\textrm{eff}}$, and for persistently rotating clusters, $\tau_{\textrm{r}} \sim R/v_{\textrm{cell}}$ (Appendix \[app:mechanisms\]).
We have assumed that the CCV is time-independent over our scale of interest – consistent with the long memory found in [@sigal2006variability]. If $\Delta$ changes faster than the cluster re-arranges, our results can be straightforwardly modified. Generalizations of [Eq. ]{}\[eq:snrt\] and [Eq. ]{}\[eq:snrt\_simp\] to this case are provided in Appendix \[app:timeaverage\].
{width="180mm"}
Tradeoffs in collective accuracy and motility: cluster rotation {#sec:rotate}
===============================================================
Our central result ([Eq. ]{}\[eq:snrt\]) shows that clusters can improve their chemotactic accuracy by changing cell positions. The simplest mechanism to do this is cluster rotation, which occurs in border cell clusters [@combedazou2016myosin] and transiently in leukocyte clusters [@malet2015collective]. When should a cluster actively rotate in order to increase its accuracy? Rotation creates an important tradeoff: more work must be put into rotating, and therefore less into crawling up the gradient [^2]
For constant work of motility, the maximum speed of a cluster of radius $R$ that rotates with angular speed $\Omega$ is ([*Methods*]{}) $$v(\Omega) = v_\textrm{max}\sqrt{1 - \frac{1}{2}\Omega^2\tau_{\textrm{rot}}^2} \label{eq:rot_tradeoff_main}$$ where $\tau_\textrm{rot} = R/v_\textrm{max}$ and $v_\textrm{max}$ the maximum cluster speed absent rotation. The cluster cannot rotate faster than $\Omega_\textrm{max} = \sqrt{2}/\tau_\textrm{rot}$.
If a cluster follows its best estimate ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ with speed $v(\Omega)$ given by [Eq. ]{}\[eq:rot\_tradeoff\_main\], it can improve its velocity in the gradient direction by rotating when the averaging time $T$ is long compared with $\tau_\textrm{rot}$ ([Fig. ]{}\[fig:rot\]a) [^3] We find that the optimal rotation speed $\Omega$ that maximizes the upgradient speed depends only on the signal-to-noise ratio without rotation, ${\ensuremath{\textrm{SNR}}}_0 \equiv \frac{1}{2}g^2/\sigma_{{\ensuremath{{ {\mathbf g}}}\xspace},0}^2$ and $T/\tau_\textrm{rot}$ ([Fig. ]{}\[fig:rot\]b, [*Methods*]{}). Mammalian cells have speeds in the range of microns/minute and radii of tens of microns, so to benefit from averaging ($T \gtrsim\tau_\textrm{rot}$), $T$ must be longer than tens of minutes. The timescale $\tau_\textrm{rot}$ and the signal-to-noise-ratio ${\ensuremath{\textrm{SNR}}}_0$ both depend on cluster size – larger clusters with more cells are both better gradient sensors and more difficult to drive to large angular speeds. As a consequence, the optimal $\Omega$ is highly cluster-size-dependent: as cluster size decreases, there is a continuous transition to nonzero optimal $\Omega$ if $T$ is sufficiently long ([Fig. ]{}\[fig:rot\]c).
{width="170mm"}
Linking chemotactic accuracy and fluidity {#sec:fluidity}
=========================================
Clusters with more cell rearrangement are more accurate by [Eq. ]{}\[eq:snrt\_simp\]. To further quantify the consequences of cell rearrrangements, we model a cluster of cells as self-propelled particles that follow the cluster estimate ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ with a noise characterized by angular diffusion $D_\psi$ and with cell-cell connections modeled as springs of strength $\kappa$ between Delaunay neighbors ([*Methods*]{}). We emphasize that the angular diffusion parameterized by $D_\psi$ is an additional source of noise: as $D_\psi$ increases, cells are less accurate in following the cluster’s estimate of the gradient. These two parameters are systematically varied to study the effects of cluster fluidity on chemotactic accuracy.
Cluster fluidity improves cluster chemotaxis
--------------------------------------------
Increasing cell-cell adhesion $\kappa$ makes clusters more ordered, moving between fluidlike and crystalline states ([Fig. ]{}\[fig:kappa\]a). As a consequence, rearrangement slows significantly ([Fig. ]{}\[fig:kappa\]c) with $\tau_r \sim \exp(\kappa/2)$ ($C_{rr}(t)$ is single-exponential).
Cluster structure and size change when clusters fluidize ([Fig. ]{}\[fig:kappa\]a), which may in principle affect the shape parameter $\chi$, which also strongly affects the chemotactic accuracy ([Eq. ]{}\[eq:combined\_error\]). However, in our simulations $\chi$ is not significantly dependent on $\kappa$, changing by under 10% ([Fig. ]{}\[fig:kappa\]d). Averaging time $T$ also has only a weak effect on cluster shape and dynamics – changes in $\tau_r$ and $\chi$ when the averaging time $T$ is increased by orders of magnitude are small ([Fig. ]{}\[fig:kappa\]). This is consistent with our assumption decoupling the gradient estimate and cell rearrangements, suggesting clusters should obey the bound [Eq. ]{}\[eq:snrt\].
We can, using the results of Section \[sec:timeaverage\], predict the cluster chemotactic index, $\textrm{CI} \equiv \langle V_x / |{ {\mathbf V}}| \rangle$, where ${ {\mathbf V}}$ is the cluster velocity. Assuming ${ {\mathbf V}} \sim {\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$, we can compute $\textrm{CI}$ from ${\ensuremath{\sigma_{\mathbf{g},T}}}$ given by [Eq. ]{}\[eq:snrt\_simp\] ([*Methods*]{}). This requires parameters $\tau_r$ and $\chi$ (measured from simulations), and ${ {\mathbf g}}$, $T$, and ${\ensuremath{\sigma_{\Delta}}}$ (known). [We note that our approach, which extracts $\tau_r$ and $\chi$ from cell trajectories, could also be applied to experimental data; in that case, ${ {\mathbf g}}$ would still be known, but the extent of time-averaging ($T$) and the error due to CCV (${\ensuremath{\sigma_{\Delta}}}$) would have to be determined by fitting to the data.]{} This prediction should be an upper bound to the measured $\textrm{CI}$, because our model includes additional noise beyond the assumptions of [Eq. ]{}\[eq:snrt\_simp\], via $D_\psi$. As expected, cluster CI decreases significantly as clusters solidify and the relaxation time $\tau_r$ increases. The simulation data qualitatively follows the predicted upper bound ([Fig. ]{}\[fig:kappa\]b). When the averaging time $T$ is reduced below typical relaxation times, the CI significantly decreases. In addition, for this short time-averaging, changing cluster stiffness no longer strongly affects CI.
Increasing single-cell stochasticity can increase cluster accuracy
------------------------------------------------------------------
{width="180mm"}
Any mechanism that fluidizes the cluster can decrease the correlation time $\tau_r$. Because of this, [*increasing*]{} noise can improve cluster chemotactic accuracy ([Fig. ]{}\[fig:Dpsi\]). We increase single-cell angular noise $D_\psi$, and see an initial sharp increase in cluster chemotactic index as $D_\psi > 0$ ([Fig. ]{}\[fig:Dpsi\]b). At larger values of $D_\psi$, cluster CI decreases below the bound set by [Eq. ]{}\[eq:snrt\_simp\], as the additional noise added degrades the gradient-following behavior. Without significant time-averaging ($T = 0.2$), additional noise primarily impedes chemotactic accuracy.
Why can extra noise $D_\psi$ help sensing? For $D_\psi = 0$, all cells follow the best estimate ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ precisely, leading to an ordered cluster ([Fig. ]{}\[fig:Dpsi\]a) with $\tau_r$ effectively infinite. As $D_\psi$ is increased, the cluster fluidizes, and the relaxation time decreases strongly ([Fig. ]{}\[fig:Dpsi\]c) resulting in more independent measurements. As in [Fig. ]{}\[fig:kappa\], this fluidization is only relevant if the averaging time $T$ exceeds the relaxation time, so when $T = 0.2$, the effect of increasing $D_\psi$ is solely detrimental to chemotaxis.
[In deriving our bound, we made two key approximations: cluster isotropy and decoupling. These approximations are exact for the rigid cluster rotation in Sec. \[sec:rotate\], but only approximate for this self-propelled particle model. As a consequence, at small $D_\psi$, simulated clusters have chemotactic indices slightly exceeding our predictions ([Fig. ]{}\[fig:Dpsi\]b). This error likely arises from emergent couplings between cluster shape and $\Delta_i$ – clusters may spread perpendicular to ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$, weakening the decoupling approximation ([Fig. ]{}\[fig:Dpsi\]a). The approximation of cluster isotropy can be removed (Appendix \[app:timeaverage\]), and does not resolve the violation of the bound ([Fig. ]{}\[fig:Dpsi\]b). Despite this potential error source, the model captures CI variation over a broad range of parameters (Appendix \[app:params\]).]{}
Discussion
==========
Our study results in several predictions and suggestions for experiments that investigate collective chemotaxis. For example, we predict that, when CCV is large, gradient sensing error is insensitive to background concentration ([Fig. ]{}\[fig:ccv\_illus\]c). This is consistent with recent measurements on developing organoids that show that the up-gradient bias is not strongly dependent on mean concentration [@ellison2016cell], though in contrast with results on lymphocyte clusters [@malet2015collective]. Furthermore, if CCV limits collective chemotaxis, clusters gradient sensing [*in vivo*]{} should have tightly regulated expression of proteins relevant to the signal response. Interestingly, measurements of zebrafish posterior lateral line primordium [@venkiteswaran2013generation] show tightly-controlled Sdf1 signaling, as measured by Cxcr4b internalization, suggesting that CCV may be small enough to allow for accurate gradient sensing.
We also show that there is a direct link between fluidity and chemotaxis as shown by [Eq. ]{}\[eq:snrt\]. Verifying this expression in experiments requires simultaneous measurement of several quantities, including cluster size, cluster re-arrangement, and signal gradient. Therefore, care has to be taken when modifying experimental conditions as these might change several of these quantities simultaneously. Altering adhesion, for example, changes both cluster fluidity and spreading as shown in a recent study using neural crest clusters [@kuriyama2014vivo], creating a confounding factor. Nevertheless, these types of experiments may be successful in setting bounds on possible time-averaging and the link between fluidity and chemotaxis.
Our results suggest that many recent experiments may need reinterpretation. Measured chemotactic accuracies can depend on cluster size [@theveneau2010collective; @malet2015collective; @cai2016modeling]; these results have been modeled without time averaging or CCV [@camley2016emergent; @malet2015collective; @cai2016modeling; @camley2016collective]. Our results show that rearrangement times $\tau_r$ also influence chemotaxis – and that $\tau_r$ depends on cluster size. Cluster relaxation dynamics are therefore an unexplored potential issue for interpreting collective gradient sensing experiments.
Essential in the reduction of gradient sensing errors due to CCV is the existence of a biochemical or mechanical memory that can perform a time average over tens of minutes. There are several possibilities. First, memory could be external to the cluster – e.g. stored in extracellular matrix structure, or a long-lived trail [@lim2015neutrophil]. Secondly, supracellular structures like actin cables influence cell protrusion and leader cell formation [@reffay2014interplay], suggesting that collective directional memory could be kept by regulating actin cable formation and maintenance. Third, memory may be kept at the individual cell level by cells attempting to estimate their own bias level $\Delta^i$ and compensating for it. This contrasts with our straightforward average of the collective estimate ${\ensuremath{\hat{{ {\mathbf g}}}}\xspace}$, but could be an important alternative mechanism.
Our results are critical for understanding the ubiquitous phenomenon of collective gradient sensing. The importance of CCV provides a new design principle: CCV must either be tightly controlled or mitigated by time-averaging. We also established a surprising link between a central [*mechanical*]{} property of a cluster – its rheology – and its sensing ability. This connects mechanical transitions like unjamming [@bi2016motility] to sensing, opening up new areas of study. In addition, our results show cluster accuracy depends strongly on cluster rearrangement mechanism. Finally, our results show that noise in cell motility can be beneficial for collective sensing.
Methods
=======
Maximum likelihood estimation of gradient direction in the presence of cell-cell variation and ligand-receptor noise {#maximum-likelihood-estimation-of-gradient-direction-in-the-presence-of-cell-cell-variation-and-ligand-receptor-noise .unnumbered}
--------------------------------------------------------------------------------------------------------------------
We compute the maximum likelihood estimate of gradient direction given the measured signal at cell $i$, $M^i$, given by [Eq. ]{}\[eq:M\]. If the cluster of cells is in a shallow linear gradient, with concentration $c_0$ at the cluster’s center of mass ${\ensuremath{{ {\mathbf r}}}\xspace}_{\textrm{cm}} = N^{-1} \sum_i {\ensuremath{{ {\mathbf r}}}\xspace}^i$, then $c({\ensuremath{{ {\mathbf r}}}\xspace}) = c_0 \left[1 + { {\mathbf g}} \cdot ({\ensuremath{{ {\mathbf r}}}\xspace}-{\ensuremath{{ {\mathbf r}}}\xspace}_{\textrm{cm}}) \right]$ and thus $\bar{c} = c_0$. $M_i$ is then $M^i = 1 + {\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}+ (\delta c^i/c_0) \eta^i + \Delta^i$ with ${\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}= {\ensuremath{{ {\mathbf r}}}\xspace}-{\ensuremath{{ {\mathbf r}}}\xspace}_{\textrm{cm}}$ and $(\delta c^i/c^i)^2 = \frac{1}{n_r} \frac{(c^i+K_D)^2}{c^i K_D}$, i.e. $(\delta c^i/c_0)^2 = \frac{1}{n_r} (1+{\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}+K_D/c_0)^2\frac{1+{\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}}{K_D/c_0}$.
$\Delta^i$ are uncorrelated between cells, with a Gaussian distribution of zero mean and standard deviation [$\sigma_{\Delta}$]{}, i.e. ${\ensuremath{\left\langle \Delta^i\Delta^j \right\rangle}} = {\ensuremath{\sigma_{\Delta}}}^2 \delta^{ij}$ with $\delta^{ij}$ the Kronecker delta function. As $\eta^i$ and $\Delta^i$ are both Gaussian, the sum of these variables is also Gaussian, and the likelihood of observing a configuration of measured signals $\{M^i\}$ as ${\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}}) = P({\ensuremath{\{M^{i}\}\xspace}}| {\ensuremath{{ {\mathbf g}}}\xspace})$, where $P({\ensuremath{\{M^{i}\}\xspace}}| {\ensuremath{{ {\mathbf g}}}\xspace})$ is the probability density function of observing the configuration [$\{M^{i}\}\xspace$]{}given parameters ${\ensuremath{{ {\mathbf g}}}\xspace}$. $${\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}}) = \prod_{i} \frac{1}{\sqrt{2\pi h^i}} \exp\left[-\frac{(M^i - \mu^i)^2}{2 h^i} \right]$$ where $\mu^i = 1 + {\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$ is the mean value of $M^i$ and $h^i = (\delta c^i / c_0)^2 + {\ensuremath{\sigma_{\Delta}}}^2$ its variance. We want to apply the method of maximum likelihood by finding the gradient parameters $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}}$ that maximize this likelihood, i.e. $$\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = \textrm{arg} \, \max\limits_{{\ensuremath{{ {\mathbf g}}}\xspace}} {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}})$$ However, because of the complex dependence of $h^i$ on the gradient ${\ensuremath{{ {\mathbf g}}}\xspace}$, this is not possible analytically. We perform this optimization numerically using a Nelder-Mead method (Matlab’s fminsearch), with an initial guess set by the maximum for $n_r \to \infty$ (i.e. neglecting concentration sensing noise), which can be found exactly. For numerical convenience, we maximize the log-likelihood $\ln {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}})$, $\ln {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}}) = -\frac{1}{2}\sum_{i} \ln h^i - \sum_i \frac{(M^i - \mu^i)^2}{2 h^i}$ up to an additive constant.
In the limit of $n_r \to \infty$ (neglecting concentration noise), our model becomes a simple linear regression, and the log likelihood can be maximized analytically by finding [$\hat{{ {\mathbf g}}}$]{}such that $\partial_{{\ensuremath{{ {\mathbf g}}}\xspace}} \ln {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace}; {\ensuremath{\{M^{i}\}\xspace}})|_{{\ensuremath{\hat{{ {\mathbf g}}}}\xspace}} = 0$ (Appendix \[app:mle\]). The result is $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = \mathcal{A}^{-1}\cdot \sum_i M^i {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$, where $\mathcal{A}_{\alpha\beta} \equiv \sum_i\delta r^i_\alpha \delta r^i_\beta$. This estimator is simplest in the limit of roughly circular (isotropic) clusters, where $\sum_i (\delta x^i)^2 \approx \sum_i (\delta y^i)^2 \gg \sum_i \delta x^i \delta y^i$. In this case, $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = (\chi^{-1}) \sum_i M^i{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$ where $\chi = \frac{1}{2}\sum_i |{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}|^2$.
Cluster rotation dynamics {#cluster-rotation-dynamics .unnumbered}
-------------------------
How much speed does a cluster lose by rotating? One possibility is to assume the power expended in generating motility is constant. Consider a circular cluster propelling itself over a surface, with the cells having velocity ${ {\mathbf v}}({\ensuremath{{ {\mathbf r}}}\xspace})$; we expect that the frictional force per unit area between the cluster and substrate will be ${ {\mathbf f}}_{\textrm{drag}} = -\xi { {\mathbf v}}$, where $\xi$ is a friction coefficient with the substrate. If all of the power available for motility is going into driving the cluster over the substrate, then we can write: $P = -\int d^2 r \, { {\mathbf v}}\cdot{ {\mathbf f}}_\textrm{drag} = \xi \int d^2 r |{ {\mathbf v}}|^2$. If the cluster is traveling as a rigid, circular cluster with its maximum possible velocity, ${ {\mathbf v}} = v_{\textrm{max}} \hat{{ {\mathbf x}}}$ then $P = \xi \pi R^2 v_\textrm{max}^2 \equiv \gamma_t v_\textrm{max}^2$ where $\gamma_t = \xi \pi R^2$ is the translational drag coefficient of the cluster. If, instead, the cluster puts its entire power into rigid-body rotation with ${ {\mathbf v}}({\ensuremath{{ {\mathbf r}}}\xspace}) = \Omega_{\textrm{max}} r (-\sin \theta,\cos \theta)$ (in polar coordinates), then $P = \xi \Omega_\textrm{max}^2 \int d^2 r r^2 = \xi\frac{\pi}{2} R^4 \Omega_{\textrm{max}}^2 \equiv \gamma_r \Omega_\textrm{max}^2$ where $\gamma_r = \frac{\xi \pi}{2} R^4$ is the rotational drag coefficient of the cluster. In general, the power dissipated if the cluster is moving rigidly with velocity ${ {\mathbf v}}$ and angular speed $\Omega$ is $P = \gamma_t v^2 + \gamma_r \Omega^2$ and hence, we find that the speed $v(\Omega$) that a cluster rotating with angular velocity $\Omega$ is able to travel to obtain is $$v(\Omega) = \sqrt{v_\textrm{max}^2 - \frac{\gamma_r}{\gamma_t} \Omega^2} \label{eq:tradeoff}$$ This quantifies one reasonable tradeoff between speed and angular velocity for a cluster. If the power available for cell motility is a small amount of the cell’s energy budget [@purcell1977life; @flamholz2014quantified; @katsu2009substantial], other tradeoffs may be more important and additional modeling will be necessary.
We consider a circular cluster traveling towards its best estimate of the gradient with speed $v(\Omega)$ given by [Eq. ]{}\[eq:tradeoff\], and traveling in the direction of the estimator ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$. We can then determine when the cluster maximizes its mean velocity in the direction of the increasing gradient, which we choose to be $x$, ${\ensuremath{\left\langle v_x \right\rangle}}$, as a function of $\Omega$. This average is $${\ensuremath{\left\langle v_x \right\rangle}} = \sqrt{v_\textrm{max}^2 - \frac{\gamma_r}{\gamma_t} \Omega^2} \times {\ensuremath{\left\langle \frac{{\ensuremath{\hat{{ {\mathbf g}}}}_{T,x}\xspace}}{|{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}|} \right\rangle}}$$ We know from our results above and in Appendix \[app:timeaverage\] that, for a fixed configuration, ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ has a Gaussian distribution with mean ${ {\mathbf g}} = g \hat{{ {\mathbf x}}}$ and variance given by [Eq. ]{}\[eq:snrt\]. The average of $\frac{{\ensuremath{\hat{{ {\mathbf g}}}}_{T,x}\xspace}}{|{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}|}$ depends only on the signal-to-noise ratio ${\ensuremath{\textrm{SNR}}}_T \equiv \frac{1}{2}(g^2/\sigma_{{ {\mathbf g}},T}^2)$, with ${\ensuremath{\left\langle \frac{{\ensuremath{\hat{{ {\mathbf g}}}}_{T,x}\xspace}}{|{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}|} \right\rangle}} = C({\ensuremath{\textrm{SNR}}}_T^{-1/2})$.
Given the angular velocity $\Omega$, we can work out the distribution of ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ by [Eq. ]{}\[eq:snrt\]. We know ${\ensuremath{{ {\mathbf r}}}\xspace}(t)\cdot{\ensuremath{{ {\mathbf r}}}\xspace}(0) = |{\ensuremath{{ {\mathbf r}}}\xspace}(0)|^2\cos \Omega t$, and hence $C_{rr}(t) = \cos \Omega t$ and its Fourier transform $C_{rr}(\omega) = \pi \left[\delta(\omega-\Omega) + \delta(\omega+\Omega)\right]$, and thus ${\ensuremath{\sigma_{\mathbf{g},T}}}^2 = {\ensuremath{\sigma_{\mathbf{g},0}}}^2 \times |K_T(\Omega)|^2 = {\ensuremath{\sigma_{\mathbf{g},0}}}^2/(1+\Omega^2T^2)$.
By rescaling to unitless parameters, we then find that $${\ensuremath{\left\langle v_x \right\rangle}}/v_\textrm{max} = \sqrt{1 - \frac{1}{2} \omega^2} \times C(\left[{\ensuremath{\textrm{SNR}}}_0 (1+\omega^2\widetilde{T}^2)\right]^{-1/2})$$ where $\omega = \Omega R / v_{\textrm{max}}$ is the unitless angular velocity, ${\ensuremath{\textrm{SNR}}}_0 = {\ensuremath{\sigma_{\Delta}}}^{-2} g^2 \chi$ is the usual SNR with no averaging, and $\widetilde{T} = T v_{\textrm{max}}/R$ is the ratio of the averaging time to the characteristic rotational time $R/v_{\textrm{max}}$, and $C(\sigma)$ is the function given by [Eq. ]{}\[eq:csigma\]. When ${\ensuremath{\textrm{SNR}}}_0$ is sufficiently small, and $\widetilde{T}$ sufficiently large, ${\ensuremath{\left\langle v_x \right\rangle}}/v_\textrm{max}$ has a maximum at finite $\omega$ ([Fig. ]{}\[fig:rot\]).
In the limit of low [$\textrm{SNR}$]{}, $C(\sigma) \approx \sqrt{\pi/8} \sigma^{-1}$, and we find ${\ensuremath{\left\langle v_x \right\rangle}}/v_\textrm{max}$ is maximized by $\omega = \pm\sqrt{1-\frac{1}{2}\widetilde{T}^{-2}}$ when $\widetilde{T} > 1/\sqrt{2}$ and $\omega = 0$ otherwise. For the large ${\ensuremath{\textrm{SNR}}}$ limit, $C(\sigma) \approx 1-\sigma^2/2$ and rotation will increase the mean velocity in the direction of the gradient when $\widetilde{T}^2 > {\ensuremath{\textrm{SNR}}}_0/2-1/4$. More generally, it is possible to find the value of $\omega$ that maximizes ${\ensuremath{\left\langle v_x \right\rangle}}$ numerically. We show the complete phase diagram in [Fig. ]{}\[fig:rot\]b.
Particle-based model of collective cell migration {#particle-based-model-of-collective-cell-migration .unnumbered}
-------------------------------------------------
We use a minimal model of collective cell migration, describing cells as self-propelled particles connected by springs: $$\begin{aligned}
\label{eq:velocity}
\frac{d}{dt}{ {\mathbf r}}^i &= { {\mathbf p}}^i + \sum_{j\sim i} { {\mathbf F}}^{ij} \\
{ {\mathbf p}}^i &= (\cos\theta^i,\sin\theta^i) \\
\theta^i &= \arctan(\hat{g}_{T,y}/\hat{g}_{T,x}) +\psi^i \\
\frac{d}{dt} \psi^i &= -\tau^{-1} \sin \psi^i+ \sqrt{2 D_\psi} \xi^i(t) \label{eq:psi}\end{aligned}$$ where $\xi^i(t)$ is a Gaussian Langevin noise with zero mean and $\langle \xi^i(t)\xi^j(t') \rangle = \delta^{ij}\delta(t-t')$, with $\delta^{ij}$ the Kronecker delta. In this model, the orientation of an individual cell $\theta^i$ is the cluster’s best estimate of the gradient direction, $\arctan(\hat{g}_{T,y}/\hat{g}_{T,x})$, plus a noise $\psi^i$ which varies from cell to cell. $\tau$ here controls the persistence of this noise, and $D_\psi$ its amplitude; when $D_\psi$ is increased, each individual cell is worse at following the estimate ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$. The cell-cell forces are $${ {\mathbf F}}^{ij} = -\kappa (|{\ensuremath{d^{ij}}\xspace}-\ell|){\ensuremath{\hat{{ {\mathbf r}}}^{ij}}\xspace}$$ where ${\ensuremath{d^{ij}}\xspace}= |{ {\mathbf r}}^i-{ {\mathbf r}}^j|$ and ${\ensuremath{\hat{{ {\mathbf r}}}^{ij}}\xspace}= ({ {\mathbf r}}^i-{ {\mathbf r}}^j)/{\ensuremath{d^{ij}}\xspace}$. The forces are only between neighboring cells $j\sim i$, where we define neighboring cells as any cells connected by the Delaunay triangulation of the cell centers ([Fig. ]{}\[fig:kappa\]a); this approach resembles that of [@meineke2001cell]. We use the Euler-Maruyama method to integrate Eqns. \[eq:velocity\]-\[eq:psi\].
Simulation units {#simulation-units .unnumbered}
----------------
We have chosen our parameters in the simulation and throughout the paper to be measured in units where the equilibrium cell-cell separation $\ell = 1$ (i.e. the cell diameter is unity), and the velocity of a single cell in the absence of cell-cell forces ${ {\mathbf v}} = { {\mathbf p}} = (\cos \theta,\sin\theta)$ has unit magnitude. For, e.g. neural crest cells, the cell diameters are of order 20 microns, and the cell speeds on the order of microns/minute – so a unitless time of $T$ corresponds to roughly $\textrm{20 minutes} \times T$ in real time. However, cell size and speed varies strongly from cell type to cell type, so we prefer to present these results in their unitless form so that they can be more easily converted.
Computing chemotactic indices {#computing-chemotactic-indices .unnumbered}
-----------------------------
![. $C(\sigma)$ plotted numerically from definition in [Eq. ]{}\[eq:csigma\][]{data-label="fig:csigma"}](cos_sigma){width="90mm"}
If we use the maximum likelihood method to make an estimate for the direction in which the cell moves, how do we translate between the uncertainty ${\ensuremath{\sigma_{\mathbf{g}}}}^2$ and the distribution of velocities? We found that the MLE estimate for the gradient is ${\ensuremath{\hat{{ {\mathbf g}}}}\xspace}= { {\mathbf g}} + { {\mathbf \Lambda}}$, with ${ {\mathbf \Lambda}}$ a Gaussian random variable with zero mean and variance ${\ensuremath{\left\langle \Lambda_x^2 \right\rangle}} = {\ensuremath{\left\langle \Lambda_y^2 \right\rangle}} = {\ensuremath{\sigma_{\mathbf{g}}}}^2/2$ – and similar results for the time-average ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$. One measure of this estimate’s accuracy is the [*instantaneous chemotactic index*]{} - or the cosine of the angle between the estimate and the gradient direction. To compute this, if ${ {\mathbf g}} = g \hat{{ {\mathbf x}}}$ without loss of generality, we find $$\begin{aligned}
{\ensuremath{\left\langle \frac{{\ensuremath{\hat{{ {\mathbf g}}}}\xspace}_x}{|{\ensuremath{\hat{{ {\mathbf g}}}}\xspace}|} \right\rangle}} &= {\ensuremath{\left\langle \frac{g + \Lambda_x}{(g+\Lambda_x)^2 + \Lambda_y^2} \right\rangle}} \\
&= \int \frac{dx dy}{2\pi\sigma} \frac{1 + x}{\left[(1+x)^2+y^2\right]^{1/2}} e^{-\frac{(x^2+y^2)}{2\sigma^2}} \\
&\equiv C(\sigma) \label{eq:csigma}\end{aligned}$$ where $\sigma = {\ensuremath{\textrm{SNR}}}^{-1/2}$, with ${\ensuremath{\textrm{SNR}}}= \frac{1}{2} (g^2/{\ensuremath{\sigma_{\mathbf{g}}}}^2)$. These results carry over naturally to the time-averaged case if ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ remains Gaussian – we find ${\ensuremath{\left\langle \frac{{\ensuremath{\hat{{ {\mathbf g}}}}\xspace}_{T,x}}{|{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}|} \right\rangle}} = C({\ensuremath{\textrm{SNR}}}_T^{-1/2})$.
The integral for $C(\sigma)$ can’t be solved analytically, but we can find asymptotic forms for $C(\sigma)$ or evaluate it numerically. For $\sigma \gg 1$, we find $C(\sigma) \approx \sqrt{\pi/8} \sigma^{-1}$, and $C(\sigma) \approx 1 - \frac{1}{2} \sigma^2$ for $\sigma \ll 1$.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Albert Bae and Monica Skoge for useful discussions, and many scientists from the Gordon Research Conference on Directed Cell Motility for interesting questions and reference suggestions. BAC also thanks Kristen Flowers for several useful suggestions. This work was supported by NIH Grant No. P01 GM078586.
[52]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [**** ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{}
[SI Appendix]{}
Review of concentration sensing accuracy {#app:conc}
========================================
With simple ligand-receptor kinetics, i.e. an on-rate of $k_\textrm{on} = k_+ c$ and an off-rate of $k_\textrm{off} = k_i$, the mean probability that each receptor will be occupied is $P_\textrm{on} = k_\textrm{on}/(k_\textrm{on}+k_\textrm{off}) = c({\ensuremath{{ {\mathbf r}}}\xspace})/(c({\ensuremath{{ {\mathbf r}}}\xspace})+K_D)$, where $K_D = k_-/k_+$ is the dissociation constant. The variance in the occupation probability for an individual receptor is then $P_\textrm{on} - P_\textrm{on}^2 = c({\ensuremath{{ {\mathbf r}}}\xspace}) K_D / (c({\ensuremath{{ {\mathbf r}}}\xspace})+K_D)^2$. By the central limit theorem, as the number of receptors $n_r$ becomes large, the number of [*occupied*]{} receptors on a cell will be a Gaussian distribution with mean $\bar{n} = n_r c({\ensuremath{{ {\mathbf r}}}\xspace})/(c({\ensuremath{{ {\mathbf r}}}\xspace})+K_D)$ and variance $\delta n^2 = n_r c({\ensuremath{{ {\mathbf r}}}\xspace}) K_D / (c({\ensuremath{{ {\mathbf r}}}\xspace})+K_D)^2$. Translating this number of occupied receptors into an uncertainty in the local concentration via $\delta c = \frac{dc}{d\bar{n}} \delta n = \frac{(c+K_D)^2}{K_D n_r} \delta n$ [@kaizu2014berg], we find $(\delta c / c)^2 = \frac{1}{n_r} \frac{(c+K_D)^2}{c K_D}$.
Maximum likelihood estimates of gradient direction via collective guidance in the presence of cell-cell variation and ligand-receptor noise {#app:mle}
===========================================================================================================================================
We begin with our model for the measured signal at cell $i$, $M^i$: $$M^i = \left[c({\ensuremath{{ {\mathbf r}}}\xspace}^i) + \delta c^i \eta^i\right]/\bar{c} + \Delta^i$$ where $\bar{c}$ is the mean concentration over the cluster of cells, $\bar{c} = N^{-1} \sum_i c({\ensuremath{{ {\mathbf r}}}\xspace}^i)$. If we assume that the cluster of cells is in a shallow linear gradient, with the concentration measured at the cluster’s center of mass ${\ensuremath{{ {\mathbf r}}}\xspace}_{\textrm{cm}} = N^{-1} \sum_i {\ensuremath{{ {\mathbf r}}}\xspace}^i$ being $c_0$, we have $c({\ensuremath{{ {\mathbf r}}}\xspace}) = c_0 \left[1 + { {\mathbf g}} \cdot ({\ensuremath{{ {\mathbf r}}}\xspace}-{\ensuremath{{ {\mathbf r}}}\xspace}_{\textrm{cm}}) \right]$ and thus $\bar{c} = c_0$. We can then write the measured signal $M^i$ as $$M^i = 1 + {\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}+ (\delta c^i/c_0) \eta^i + \Delta^i$$ with ${\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}= {\ensuremath{{ {\mathbf r}}}\xspace}-{\ensuremath{{ {\mathbf r}}}\xspace}_{\textrm{cm}}$ and $(\delta c^i/c^i)^2 = \frac{1}{n_r} \frac{(c^i+K_D)^2}{c^i K_D}$, i.e. $(\delta c^i/c_0)^2 = \frac{1}{n_r} (1+{\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}+K_D/c_0)^2\frac{1+{\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}}{K_D/c_0}$.
We have assumed that $\Delta^i$ are uncorrelated between cells, with a Gaussian distribution of zero mean and standard deviation [$\sigma_{\Delta}$]{}, i.e. ${\ensuremath{\left\langle \Delta^i\Delta^j \right\rangle}} = {\ensuremath{\sigma_{\Delta}}}^2 \delta^{ij}$ with $\delta^{ij}$ the Kronecker delta. $M^i$, as the sum of the Gaussian variables $\eta^i$ and $\Delta^i$, is also Gaussian, and we can then write the likelihood of observing a configuration of measured signals $\{M^i\}$ as ${\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}}) = P({\ensuremath{\{M^{i}\}\xspace}}| {\ensuremath{{ {\mathbf g}}}\xspace})$, where $P({\ensuremath{\{M^{i}\}\xspace}}| {\ensuremath{{ {\mathbf g}}}\xspace})$ is the probability density function of observing the configuration [$\{M^{i}\}\xspace$]{}given parameters ${\ensuremath{{ {\mathbf g}}}\xspace}$. $${\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}}) = \prod_{i} \frac{1}{\sqrt{2\pi h^i}} \exp\left[-\frac{(M^i - \mu^i)^2}{2 h^i} \right]$$ where $\mu^i = 1 + {\ensuremath{{ {\mathbf g}}}\xspace}\cdot{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$ is the mean value of $M^i$ and $h^i = (\delta c^i / c_0)^2 + {\ensuremath{\sigma_{\Delta}}}^2$ its variance. We want to apply the method of maximum likelihood by finding the gradient parameters $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}}$ that maximize this likelihood, i.e. $$\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = \textrm{arg} \, \max\limits_{{\ensuremath{{ {\mathbf g}}}\xspace}} {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}})$$ However, because of the complex dependence of $h^i$ on the gradient ${\ensuremath{{ {\mathbf g}}}\xspace}$, analytically maximizing the likelihood is intractable. We instead perform this optimization numerically using a Nelder-Mead method (Matlab’s fminsearch), with an initial guess set by the maximum for $n_r \to \infty$ (i.e. neglecting concentration sensing noise), given by [Eq. ]{}\[eq:ghat\_simple\]. For numerical stability and convenience, we will usually instead maximize the log-likelihood $\ln {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}})$, which is $$\ln {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace};{\ensuremath{\{M^{i}\}\xspace}}) = -\frac{1}{2}\sum_{i} \ln h^i - \sum_i \frac{(M^i - \mu^i)^2}{2 h^i}$$ up to an additive constant.
In the limit of $n_r \to \infty$ (neglecting concentration noise), our model becomes a simple linear regression, and the log likelihood can be maximized analytically by finding [$\hat{{ {\mathbf g}}}$]{}such that $\partial_{{\ensuremath{{ {\mathbf g}}}\xspace}} \ln {\ensuremath{\mathcal{L}}\xspace}({\ensuremath{{ {\mathbf g}}}\xspace}; {\ensuremath{\{M^{i}\}\xspace}})|_{{\ensuremath{\hat{{ {\mathbf g}}}}\xspace}} = 0$. This straightforwardly yields $$\sum_i (M^i - 1 - {\ensuremath{\hat{{ {\mathbf g}}}}\xspace}\cdot {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}){\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}= 0$$ or, if we write ${\ensuremath{\hat{{ {\mathbf g}}}}\xspace}= {\ensuremath{\hat{g}_x}\xspace}\hat{{ {\mathbf x}}} + {\ensuremath{\hat{g}_y}\xspace}\hat{{ {\mathbf y}}}$, $$\nonumber
\begin{pmatrix} \sum_{i} (\delta x^i)^2 & \sum_i \delta x^i \delta y^i \\ \sum_i \delta x^i \delta y^i & \sum_i (\delta y^i)^2 \end{pmatrix} \begin{pmatrix} {\ensuremath{\hat{g}_x}\xspace}\\ {\ensuremath{\hat{g}_y}\xspace}\end{pmatrix} = \begin{pmatrix} \sum_i (M^i - 1) \delta x^i \\ \sum_i (M^i - 1) \delta y^i \end{pmatrix}$$
We define $\mathcal{A}$ to be a matrix with elements $\mathcal{A}_{\alpha\beta} = \sum_i \delta r_\alpha \delta r_\beta$, where $\alpha,\beta$ run over the Cartesian coordinates $x,y$, $$\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = (\mathcal{A}^{-1})\cdot \sum_i (M^i - 1) {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}\label{eq:ghat_simple}$$ where we note that, as $\sum_i {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}= 0$, we can also simply use $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = (\mathcal{A}^{-1})\cdot \sum_i M^i{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$.
These estimators are simpler for roughly symmetric clusters, where $\sum_i (\delta x^i)^2 \approx \sum_i (\delta y^i)^2 \gg \sum_i \delta x^i \delta y^i$. In this case, $\mathcal{A}_{\alpha\beta} \approx \chi \delta_{\alpha\beta}$, where $\chi = \frac{1}{2}\sum_i |{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}|^2$ and $\delta_{\alpha\beta}$ is the Kronecker delta, and $$\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = (\chi^{-1}) \sum_i (M^i - 1) {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$$
We can also compute the asymptotic covariance of these estimators, which arises from the Fisher information matrix $$\begin{aligned}
\mathcal{I}_{\alpha\beta} &= {\ensuremath{\left\langle \left(\frac{\partial \ln {\ensuremath{\mathcal{L}}\xspace}}{\partial g_{\alpha}}\right) \left(\frac{\partial \ln {\ensuremath{\mathcal{L}}\xspace}}{\partial g_{\beta}}\right) \right\rangle}}\\
&= -{\ensuremath{\left\langle \frac{\partial^2 \ln {\ensuremath{\mathcal{L}}\xspace}}{\partial g_{\alpha} \partial g_{\beta}} \right\rangle}} \\
&= {\ensuremath{\sigma_{\Delta}}}^{-2} \sum_i \delta r^i_\alpha \delta r^i_\beta = {\ensuremath{\sigma_{\Delta}}}^{-2} \mathcal{A}_{\alpha\beta}\end{aligned}$$ where we use ${\ensuremath{\left\langle \cdots \right\rangle}}$ to indicate the average over the cell-to-cell systematic errors $\Delta_i$.
The Fisher information matrix controls the variance of our maximum-likelihood estimator [@kay1993fundamentals], which is given by $$\begin{aligned}
{\ensuremath{\left\langle ({\ensuremath{\hat{{ {\mathbf g}}}}\xspace}- { {\mathbf g}})_\alpha ({\ensuremath{\hat{{ {\mathbf g}}}}\xspace}- { {\mathbf g}})_\beta \right\rangle}} &= \left(\mathcal{I}^{-1}\right)_{\alpha\beta} \\
&= {\ensuremath{\sigma_{\Delta}}}^2 (\mathcal{A}^{-1})_{\alpha\beta}\end{aligned}$$ In particular, for ${\ensuremath{\sigma_{\mathbf{g}}}}^2 \equiv {\ensuremath{\left\langle |{\ensuremath{\hat{{ {\mathbf g}}}}\xspace}-{ {\mathbf g}}|^2 \right\rangle}}$, $${\ensuremath{\sigma_{\mathbf{g}}}}^2 = {\ensuremath{\sigma_{\Delta}}}^2 \textrm{tr} A^{-1} \label{eq:traceerr}$$
In the symmetric cluster limit, the matrix is diagonal, and the result is simply $\sigma_{\alpha}^{2} = {\ensuremath{\sigma_{\Delta}}}^{2}/\chi$ where $\sigma_{\alpha}$ is the standard deviation of the estimator $\hat{g}_\alpha$, and hence ${\ensuremath{\sigma_{\mathbf{g}}}}^2 = \sigma_x^2 + \sigma_y^2$, $${\ensuremath{\sigma_{\mathbf{g}}}}^{2} = 2{\ensuremath{\sigma_{\Delta}}}^{2}/\chi \label{eq:sigx}$$
Computation of $\chi$ for cells in hexagonally-packed cluster {#app:Q}
=============================================================
![ Illustration of $Q$-layer hexagonally packed cell clusters and computation of $\chi(Q)$.[]{data-label="fig:Q"}](qmer_chi_calculation){width="90mm"}
When we keep a fixed cluster geometry, we choose to work with hexagonally-packed cell clusters, following our earlier work [@camley2016emergent]. We illustrate clusters with $Q = 1, 2,$ and $3$ layers in [Fig. ]{}\[fig:Q\].
How can we calculate $\chi(Q)$? Because of the isotropy of the cluster, it is easiest to work with $\chi = \frac{1}{2} \sum |{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}|^2 = \sum (\delta x^i)^2 = \sum (x^i)^2$. We know that for a single cell ($Q = 0$), $\chi(0) = 0$. We can then determine $\chi(Q)$ in terms of $\chi(Q-1)$ by computing $x^2$ for each of the cells in the outer layer, which we call $G(Q) = \sum_{\textrm{i in outside layer}} (x^i)^2$. Then, $\chi(Q) = \chi(Q-1) + G(Q) = \sum_{q=1}^{Q} G(q)$.
To compute $G(Q)$, there are three relevant portions of the cluster, as drawn in [Fig. ]{}\[fig:Q\]. These are the two left and rightmost extreme cells at $x = \pm Q^2$ (two blue dashed boxes), the sides (four red boxes, solid lines), and the top and bottom edges (black dashed boxes). We then find $$\begin{aligned}
\nonumber G(Q) &= 2 \times Q^2 + 4 \times \left[\sum_{j = 1}^{Q-1} (Q-j/2)^2\right] + 2 \times \left[\sum_{k = 0}^Q (Q/2-k)^2 \right] \\
\nonumber &= 2 \times Q^2 + 4 \times \left[ \frac{14 Q^3 - 15 Q^2 + Q}{24} \right] + 2 \times \left[ \frac{Q^3 + 3 Q^2 + 2 Q}{12}\right] \\
\nonumber &= \frac{5}{2} Q^3 + \frac{1}{2} Q\end{aligned}$$ and hence $\chi(Q) = \sum_{q=1}^{Q} G(q) = (5/8) Q^4 + (5/4) Q^3 + (7/8) Q^2 + (1/4) Q$, as we state in the main paper.
How does $\chi$ scale with the cluster size? A cluster with $Q$ layers has $N(Q) = 1 + 3 Q + 3 Q^2$ cells, so $\chi \sim Q^4 \sim N^2$. For a roughly circular cluster of radius $R$, we’d then expect that $\chi \sim R^4$. This will obviously depend on the precise details of the cluster shape and cell-cell spacing, but if we can approximate the sum in $\chi = \frac{1}{2}\sum_i |{\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}|^2$ as an integral, we find $\chi \approx \frac{1}{2}\rho_c \int d^2 r r^2 = \frac{\pi}{4} \rho_c R_{\textrm{cluster}}^{4}$, where $\rho_c$ is the number of cells per unit area in the cluster.
[How does this scaling compare with earlier results for single cells? Ref. [@hu2011geometry] uses a maximum likelihood method to determine the accuracy limit for a single cell with $n_r$ receptors spaced around its radius, finding $\sigma_p^2 \sim \gamma / n_r$, where $\gamma = \frac{(\bar{c}+K_D)^2}{\bar{c}K_D}$ and $p \sim g \times R_\textrm{cell}$, with $R_\textrm{cell}$ the cell radius. Changing variables to ${ {\mathbf g}}$, this is $\sigma_g^2 \sim \frac{\gamma}{n_r R_\textrm{cell}^2}$. Our results, ignoring cell-to-cell variation, are that ${\ensuremath{\sigma_{\mathbf{g}}}}^2 \sim \frac{\gamma}{n_r \chi}$, with $\chi \sim \rho_c R^4$ for a circular cluster of radius $R$, and $n_r$ the number of receptors per cell. Then, as there are $N \sim \rho_c R^2$ cells in the cluster, the total number of receptors within the cluster is $n_t \sim N\times n_r$. Our result for ${\ensuremath{\sigma_{\mathbf{g}}}}$ (in the absence of time averaging and CCV) is then ${\ensuremath{\sigma_{\mathbf{g}}}}^2 \sim \frac{\gamma}{n_t R^2}$. This shows that, up to geometric factors, a cluster’s sensing bound is the same as a giant cell with the same [*total*]{} number of receptors and radius.]{}
Detailed derivation of time-averaged gradient sensing error {#app:timeaverage}
===========================================================
We will, in this section, completely neglect the ligand-receptor fluctuations in gradient sensing, as appropriate for the physically likely case ${\ensuremath{\sigma_{\Delta}}}> 0.1$. We show in Appendix \[app:mle\] that the maximum likelihood estimator of the gradient is then $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} = (\mathcal{A}^{-1})\cdot \sum_i (M^i - 1) {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}$, where $\mathcal{A}_{\alpha\beta} = \sum_i \delta r_\alpha \delta r_\beta$, with $\alpha,\beta$ the Cartesian coordinates $x,y$.
We now evaluate $\hat{{\ensuremath{{ {\mathbf g}}}\xspace}}$ with the signal $M^i = 1 + {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}\cdot{\ensuremath{{ {\mathbf g}}}\xspace}+ \Delta^i$: $$\begin{aligned}
\hat{{\ensuremath{{ {\mathbf g}}}\xspace}} &= {\ensuremath{{ {\mathbf g}}}\xspace}+ (\mathcal{A}^{-1})\cdot\sum_i \Delta^i {\ensuremath{\boldsymbol\delta { {\mathbf r}}^i}\xspace}\\
&\equiv {\ensuremath{{ {\mathbf g}}}\xspace}+ {\ensuremath{\boldsymbol\Lambda\xspace}}\end{aligned}$$ We will treat a more general case than in the main body, allowing $\Delta_i$ to vary in a time-dependent manner, ${\ensuremath{\left\langle \Delta^i(t) \Delta^j(0) \right\rangle}} = {\ensuremath{\sigma_{\Delta}}}^2 C_{\Delta\Delta}(t) \delta^{ij}$ where $\delta^{ij}$ is the Kronecker delta function and $C_{\Delta\Delta}(t)$ characterizes the correlation of the CCV; within the main paper, we take $C_{\Delta\Delta} \to 1$, assuming CCV is persistent over all relevant time scales of the motion.
Suppose the cell time-averages its maximum likelihood estimator with a time window [$T$]{}: $$\begin{aligned}
{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}(t) &= \int {\ensuremath{\hat{{ {\mathbf g}}}}\xspace}(t') K_T(t-t') \\
&= {\ensuremath{{ {\mathbf g}}}\xspace}+ {\ensuremath{\boldsymbol\Lambda_T\xspace}}\end{aligned}$$ where $K_T(t)$ is an averaging function with $K(t<0) = 0$ and ${\ensuremath{\boldsymbol\Lambda_T\xspace}}\equiv \int_{-\infty}^{\infty} {\ensuremath{\boldsymbol\Lambda\xspace}}(t') K_T(t-t') dt'$. We will often use a simple exponential average with $K(t) = \theta(t) \frac{1}{T} e^{-t/T}$, where $\theta(t)$ is the Heaviside step function.
Clearly, for a single configuration of cells, ${\ensuremath{\left\langle {\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}\right\rangle}} = {\ensuremath{{ {\mathbf g}}}\xspace}$. We then would like to compute how much the variations in ${\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}$ are reduced by the time-averaging, i.e. we compute
$$\begin{aligned}
{\ensuremath{\left\langle |{\ensuremath{\hat{{ {\mathbf g}}}}_T\xspace}-{\ensuremath{{ {\mathbf g}}}\xspace}|^2 \right\rangle}} \equiv {\ensuremath{\sigma_{\mathbf{g},T}}}^2 &= {\ensuremath{\left\langle |{\ensuremath{\boldsymbol\Lambda_T\xspace}}|^2 \right\rangle}} \\
&= {\ensuremath{\left\langle \int_{-\infty}^{\infty}dt'\int_{-\infty}^{\infty}dt'' {\ensuremath{\boldsymbol\Lambda\xspace}}(t')\cdot{\ensuremath{\boldsymbol\Lambda\xspace}}(t'') K_T(t-t') K_T(t-t'') \right\rangle}} \\
&= \int_{-\infty}^{\infty}dt''\int_{-\infty}^{\infty}dt'' K_T(t-t') K_T(t-t'') {\ensuremath{\left\langle \sum_{i,j} {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t') \Delta^i(t') \delta r_\gamma^i(t') {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t'') \Delta^j(t'') \delta r_\beta^j(t'') \right\rangle}} \end{aligned}$$
where we have used Einstein summation notation in the last equation. We also note that the average ${\ensuremath{\left\langle \cdots \right\rangle}}$ now includes an average over time – the cell configurations are changing. We now want to perform the average over the $\Delta^i$. This is possible if the re-arrangement of the cell positions are independent of the particular values of $\Delta$, i.e. ${\ensuremath{\left\langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t') \Delta^i(t') \delta r_\gamma^i(t') {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t'') \Delta^j(t'') \delta r_\beta^j(t'') \right\rangle}} \approx {\ensuremath{\left\langle \Delta^i(t')\Delta^j(t'') \right\rangle}} \langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t') \delta r_\gamma^i(t') {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t'') \delta r_\beta^j(t'') \rangle$. This would be natural if, e.g. the cluster collectively chooses an estimated direction, but the re-arrangements are only due to local fluctuations, independent of $\Delta$. However, if each cell has a motility related to $\Delta^i$, this assumption may not be accurate. This approximation is also slightly violated if the cell cluster takes on a different shape in response to its estimate of the gradient location. This is an important approximation, but one that we suspect is unavoidable to create a measure of the correlation ${\ensuremath{\left\langle {\ensuremath{\boldsymbol\Lambda\xspace}}(t) {\ensuremath{\boldsymbol\Lambda\xspace}}(t') \right\rangle}}$ that does not depend on $\Delta$. With this decoupling approximation, we find: $$\begin{aligned}
{\ensuremath{\sigma_{\mathbf{g},T}}}^2 &\approx \int_{-\infty}^{\infty}dt'\int_{-\infty}^{\infty}dt'' K_T(t-t') K_T(t-t'') \sum_{i,j}{\ensuremath{\left\langle \Delta^i(t')\Delta^j(t'') \right\rangle}} \langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t') \delta r_\gamma^i(t') {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t'') \delta r_\beta^j(t'') \rangle\\
&= {\ensuremath{\sigma_{\Delta}}}^2 \int_{-\infty}^{\infty}dt'\int_{-\infty}^{\infty}dt'' K_T(t-t') K_T(t-t'') C_{\Delta\Delta}(t'-t'') \sum_i \langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t') \delta r_\gamma^i(t') {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t'') \delta r_\beta^i(t'') \rangle \label{eq:inprogress}\end{aligned}$$ We emphasize that in the absence of averaging, $K_T(t-t') \to \delta(t-t')$, our result agrees with the results of Appendix \[app:mle\]. In this case, the right-hand-side of [Eq. ]{}\[eq:inprogress\] becomes ${\ensuremath{\sigma_{\Delta}}}^2 C_{\Delta\Delta}(0) \sum_i \langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t) \delta r_\gamma^i(t) {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t) \delta r_\beta^i(t) \rangle = C_{\Delta\Delta}(0) {\ensuremath{\sigma_{\Delta}}}^2 \langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\alpha} \rangle$, as $\sum_i \delta r_\gamma^i \delta r_\beta^i = \mathcal{A}_{\gamma\beta}$. $C_{\Delta\Delta}(0) = 1$ by definition. This suggests we write $$\begin{aligned}
\nonumber {\ensuremath{\sigma_{\mathbf{g},T}}}^2 = {\ensuremath{\sigma_{\mathbf{g},0}}}^2 &\times \\ \int_{-\infty}^{\infty}dt''&\int_{-\infty}^{\infty}dt'' K_T(t-t') K_T(t-t'') C_{\Lambda\Lambda}(t'-t'') \label{eq:conv}\end{aligned}$$ where $C_{\Lambda\Lambda}(t)$ is the normalized correlation function $C_{\Lambda\Lambda}(t'-t'') = C_{\Delta\Delta}(t-t')\times \sum_i \langle {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\gamma}(t') \delta r_\gamma^i(t') {\ensuremath{\mathcal{A}^{-1}}}_{\alpha\beta}(t'') \delta r_\beta^i(t'') \rangle / \langle {\ensuremath{\mathcal{A}^{-1}}}_{\mu\mu} \rangle$.
The correlation function $C_{\Lambda\Lambda}$ can be calculated readily from the cell trajectories relative to the cluster center of mass $\delta {\ensuremath{{ {\mathbf r}}}\xspace}^i(t)$. However, in the limit of isotropic clusters of roughly constant shape, we can significantly simplify this form. For isotropic clusters of constant shape, $\mathcal{A}_{\alpha\beta}(t) = \chi \delta_{\alpha\beta}$ independent of time. Given this assumption, $C_{\Lambda\Lambda}(t'-t'') / C_{\Delta\Delta}(t'-t'') = \chi^{-1} \sum_i \langle \delta r_\alpha^i(t') \delta r_\alpha^i(t'') \rangle$ – i.e. $C_{\Lambda\Lambda}(t'-t'') = C_{\Delta\Delta}(t'-t'')C_{rr}(t'-t'')$, where $C_{rr}(t'-t'') \equiv \langle \boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}(t')\cdot\boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}(t'') \rangle / \langle |\boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}|^2 \rangle$.
The double convolution in [Eq. ]{}\[eq:conv\] is simpler in Fourier space, $${\ensuremath{\sigma_{\mathbf{g},T}}}^2 = {\ensuremath{\sigma_{\mathbf{g},0}}}^2 \times \int_{-\infty}^\infty \frac{d\omega}{2\pi} |K_T(\omega)|^2 C_{\Lambda\Lambda}(\omega) \label{eq:gtvar}$$ where $K_T(\omega)$ is the Fourier transform of $K_T(t)$, $K_T(t) = \int \frac{d\omega}{2\pi} e^{i\omega t} K_T(\omega)$ and $C_{\Lambda\Lambda}(\omega)$ the Fourier transform of $C_{\Lambda\Lambda}(t)$. For $K_T(t-t') = \theta(t-t') \frac{1}{T} e^{-(t-t')/T}$, $K_T(\omega) = \frac{1}{1+i\omega T}$.
In the common case that $C_{\Lambda\Lambda}(t) = \exp(-t/\tau_\Lambda)$, where $\tau_\Lambda$ is a characteristic correlation time, and $K_T(t-t') = \theta(t-t') \frac{1}{T} e^{-(t-t')/T}$, this is even simpler: $${\ensuremath{\sigma_{\mathbf{g},T}}}^2 = \frac{{\ensuremath{\sigma_{\mathbf{g},0}}}^2}{1 + T/\tau_\Lambda}$$ When we can approximate $C_{\Lambda\Lambda} \approx C_{rr}$, this is consistent with our intuition: it takes roughly a time of $\tau_r$ for the cells to re-arrange, and so in a time of $T$, the cluster can make $T/\tau_r$ independent measurements in an averaging time, and so it can decrease the measurement error by $T/\tau_r$. If the amount of CCV varies over time, the characteristic timescale is then $\tau_c = \frac{\tau_r \tau_\Delta}{\tau_r + \tau_\Delta}$ – the relaxation timescale that is relevant is the faster of the two timescales. Within the main article, we have assumed that this is always cell position rearrangement.
Because of the complexity of the different results and regimes in this section, we provide a summary in Table \[tab:summary\].
[**Formula for ${\ensuremath{\sigma_{\mathbf{g},T}}}^2$**]{} [**Assumptions made**]{} [**Associated quantities**]{}
---------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
${\ensuremath{\sigma_{\mathbf{g},0}}}^2 \times \int_{-\infty}^\infty \frac{d\omega}{2\pi} |K_T(\omega)|^2 C_{\Lambda\Lambda}(\omega)$ Decoupling of averages over CCV and configurations
${\ensuremath{\sigma_{\mathbf{g},0}}}^2 \times \int_{-\infty}^\infty \frac{d\omega}{2\pi} |K_T(\omega)|^2 \left\{C_{\Delta\Delta}(t) C_{rr}(t)\right\}_\omega$ Decoupling of averages over CCV and configurations; isotropic clusters of roughly constant shape $C_{rr}(t'-t'') \equiv \langle \boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}(t')\cdot\boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}(t'') \rangle / \langle |\boldsymbol\delta{\ensuremath{{ {\mathbf r}}}\xspace}|^2 \rangle$, $\{f(t)\}_\omega$ is the Fourier transform of $f(t)$ to $\omega$.
${\ensuremath{\sigma_{\mathbf{g},0}}}^2/\left[1+T/\tau_\Lambda\right]$ Decoupling of averages over CCV and configurations, exponential time averaging, $C_{\Lambda\Lambda} = e^{-t/\tau_\Lambda}$
${\ensuremath{\sigma_{\mathbf{g},0}}}^2/\left[1+T/\tau_{c}\right]$ Decoupling of averages over CCV and configurations, exponential time averaging, isotropic clusters of roughly constant shape, $C_{\Delta\Delta} = e^{-t/\tau_\Delta}$, $C_{rr} = e^{-t/\tau_r}$ $\tau_c = \frac{\tau_r \tau_\Delta}{\tau_r + \tau_\Delta}$
Characteristic time scales of different re-arrangement mechanisms {#app:mechanisms}
=================================================================
The characteristic time of the position-position correlation, $\tau_r$ is critical in calculating chemotactic accuracy via [Eq. ]{}\[eq:snrt\_simp\]. What is $\tau_r$, and what does it depend on? This depends on the mechanism of re-arrangement.
[**Persistent cluster rotation.**]{} If the cluster rigidly rotates with angular velocity $\Omega$, we can see that ${\ensuremath{{ {\mathbf r}}}\xspace}(t)\cdot{\ensuremath{{ {\mathbf r}}}\xspace}(0) = |{\ensuremath{{ {\mathbf r}}}\xspace}(0)|^2 \cos \Omega t$, and time-averaging over a timescale $T > \Omega^{-1}$ can improve the signal-to-noise ratio. How does $\Omega$ depend on the cluster and cell properties? As, for a cluster to actively rotate with velocity $\Omega$, cells at the edge must be able to crawl with speed $R \Omega$, where $R$ is the cluster radius, we expect that $\Omega \sim v_{\textrm{max}}/R$, where $v_{\textrm{max}}$ is the maximum speed a cell can crawl. Hence, the characteristic time over which ${\ensuremath{{ {\mathbf r}}}\xspace}$ changes is $\tau_{\textrm{rot}} \sim R/v_{\textrm{max}}$, as studied explicitly for rotating clusters in the main paper.
[**Rotational diffusion of the cluster.**]{} If the cluster diffuses rotationally as a rigid body with angular diffusion rate $D_r$, $\langle {\ensuremath{{ {\mathbf r}}}\xspace}(t)\cdot{\ensuremath{{ {\mathbf r}}}\xspace}(0)\rangle = \langle |{\ensuremath{{ {\mathbf r}}}\xspace}(0)|^2\rangle e^{-D_r t}$. However, the scaling of the rotational diffusion coefficient with cluster size is not obvious, and will depend on the model. Rotational diffusion could arise from clusters that undergo collectively-driven rotation, but occasionally switch between moving in different directions over a timescale $\tau_{\textrm{switch}}$ – this is observed in small clusters of cells on micropatterns [@segerer2015emergence]. If this is the case, the effective rotational diffusion coefficient is just $D_r \sim \Omega^2 \tau_{\textrm{switch}}$ – so $D_r \sim v_{\textrm{max}}^2 \tau_{\textrm{switch}}/R^2$. By contrast, we have seen that in the absence of a collective aligning effect, rotational diffusion can be small or absent, depending on certain details about the underlying cell motility model [@camley2016emergent].
[**Cell re-arrangements.**]{} Cells within a tissue can often be described as undergoing a persistent random walk: they maintain a direction over a timescale $\tau_{\textrm{persist}}$, but lose their orientation beyond this time, leading to an effectively diffusive motion with effective diffusion coefficient $D_{\textrm{eff}}$ at long time scales [@szabo2010collective]; these numbers can be on the order of 10 microns$^2$/min. For cells to move from one half of the cluster to another by diffusion will then take a timescale $\tau_\textrm{diff}$ where $D_{\textrm{eff}} \tau_{\textrm{diff}} \sim R^2$. However, if cells are persistent over the timescale required to cross the cluster, re-arrangements could be accelerated – in this case, $\tau_{\textrm{rearrange}} \sim R/v_\textrm{cell}$. Naturally, if cell motion becomes subdiffusive, this will also change the dynamics of the cluster re-arrangement. This would all be captured in [Eq. ]{}\[eq:snrt\] once the cell-cell correlation function is determined.
Bounds capture variation of CI over large range of parameters {#app:params}
=============================================================
{width="180mm"} {width="180mm"}
In this section, we show a larger range of variations in parameters, showing that the predicted CI from the bound captures simulated CI in our models well. While our primary results on computing the upper bound in Section \[sec:fluidity\] were performed with [Eq. ]{}\[eq:snrt\_simp\], it is also possible to compute ${\ensuremath{\sigma_{\mathbf{g}}}}^2$ without any assumptions about cluster isotropy, using the correlation function $C_{\Lambda\Lambda}(t)$ as seen in Appendix \[app:timeaverage\], and the no-time-averaging result of $\sigma_{\mathbf{g},0}^2$ via [Eq. ]{}\[eq:traceerr\]. We find that typically, in our simple collective cell motility simulations, that $\tau_{\Lambda}$ and $\tau_r$ are very close, as are $\textrm{tr} \mathcal{A}^{-1}$ and $2/\chi$ – unless the cluster becomes elongated or otherwise anisotropic. However, we do note that $\textrm{tr} \mathcal{A}^{-1} \ge 2/\chi$ for any configuration of cells, so this implies a larger uncertainty than the isotropic approximation. We compare the simulated CI and our predictions from the isotropic and full theories in [Fig. ]{}\[fig:compareall\].
[^1]: We measure in units of the cell diameter; see [*Methods*]{}
[^2]: We note that this is most relevant if motility is a large portion of the cluster’s energy budget, a complex, cell-type-dependent question [@purcell1977life; @flamholz2014quantified; @katsu2009substantial]
[^3]: The cluster’s directionality is always improved by rotating, so there is no tradeoff unless speed of motion matters.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A systematic search for axial octupole deformation in the actinides and superheavy nuclei with proton numbers $Z=88-126$ and neutron numbers from two-proton drip line up to $N=210$ has been performed in covariant density functional theory (DFT) using four state-of-the-art covariant energy density functionals representing different model classes. The nuclei in the $Z\sim 96, N\sim 196$ region of octupole deformation have been investigated in detail and the systematic uncertainties in the description of their observables have been quantified. Similar region of octupole deformation exists also in Skyrme DFT and microscopic+macroscopic approach but it is centered at somewhat different particle numbers. Theoretical uncertainties in the predictions of the regions of octupole deformation are increasing on going to superheavy nuclei with $Z\sim 120, N\sim 190$. There are no octupole deformed nuclei for $Z=112-126$ in covariant DFT calculations. This agrees with Skyrme DFT calculations, but disagrees with Gogny DFT and microscopic+macroscopic calculations which predict extended $Z\sim 120, N\sim 190$ region of octupole deformation.'
author:
- 'S. E. Agbemava'
- 'A. V. Afanasjev'
bibliography:
- 'references17.bib'
title: 'Octupole deformation in the ground states of even-even $Z\sim 96, N\sim 196$ actinides and superheavy nuclei'
---
Introduction
============
Reflection asymmetric (or octupole deformed) shapes represent an interesting example of symmetry breaking of the nuclear mean field [@BN.96]. They are present in the ground and rotating states of the lanthanides with $Z\sim 58, N\sim 90$ and light actinides with $Z\sim 90, N\sim 136$ (see Refs. [@BN.96; @MBCOISI.08; @RR.12; @AAR.16] and references quoted therein). These shapes also affect the outer fission barriers in actinides and superheavy nuclei [@AAR.12; @SR.16; @Zhou.16] and cluster radioactivity [@WR.11]. The first significant wave of the studies of octupole deformed shapes took place in the 80ies and first half of 90ies of the last century (see review of Ref. [@BN.96]). The interest to the study of such shapes has significantly increased during this decade (see references on theoretical and experimental works quoted in Ref. [@AAR.16]).
Different theoretical frameworks have been used for the study of octupole deformed shapes (see Refs. [@BN.96; @AAR.16; @SR.16] and references quoted therein). Here we employ the covariant density functional theory (CDFT) [@VALR.05]. Its previous applications to the investigation of such shapes have been overviewed and compared with the results of non-relativistic studies in Ref. [@AAR.16]. Built on Lorentz covariance and the Dirac equation, CDFT provides a natural incorporation of spin degrees of freedom [@Rei.89; @Ring1996_PPNP37-193] and a good parameter free description of spin-orbit splittings [@Ring1996_PPNP37-193; @BRRMG.99; @LA.11], which have an essential influence on the underlying shell structure. In addition, in CDFT the time-odd components of the mean fields are given by the spatial components of the Lorentz vectors. Therefore, because of Lorentz invariance, these fields are coupled with the same constants as the time-like components [@AA.10] which are fitted to ground state properties of finite nuclei (which are affected only by time-even mean fields) and nuclear matter properties.
Starting from pioneering work of Ref. [@RMRG.95], the CDFT has been extensively used in the study of reflection asymmetric shapes especially during last decade. Most of these applications have been focused on reflection symmetric shapes with axial symmetry; they have been reviewed in Ref. [@AAR.16]. Let us mention some of these studies performed in the actinides. At the mean field level, the ground state properties of the actinides have been studied in Refs. [@RMRG.95; @GMT.07; @LSYVM.13; @NVL.13; @NVNL.14; @AAR.16]. Some axial octupole deformed nuclei have been studied also in the beyond mean field approaches based on CDFT. For example, simultaneous quadrupole and octupole shape phase transitions in the Th isotopes have been studied in Ref. [@LSYVM.13] employing microscopic collective Hamiltonian. Using Interacting Boson Model Hamiltonian with parameters determined by mapping the microscopic potential energy surfaces, obtained in the relativistic Hartree-Bogoliubov calculations, to the expectation value of the Hamiltonian in the boson condensate the microscopic analysis of the octupole phase transition has been performed in Refs. [@NVL.13; @NVNL.14]. The generator coordinate method studies taking into account dynamical correlations and quadrupole-octupole shape fluctuations have been undertaken in $^{224}$Ra employing the PC-PK1 functional in Ref. [@YZL.15]. They revealed rotation-induced octupole shape stabilization.
Nonaxial-octupole ${Y}_{32}$ correlations in the $N=150$ isotones and tetrahedral shapes in neutron-rich Zr isotopes have been studied in Refs. [@ZLEZ.12; @ZLZZ.17] employing multidimensional constrained CDFT. Although the energy gain due to $\beta_{32}$ distortion exceeds 300 keV in $^{248}$Cf and $^{250}$Fm in model calculations, it is not likely that static deformation of this type is present in nature in these two nuclei. This is because their rotational features are well described in the cranked relativistic Hartree-Bogoliubov framework with no octupole deformation [@AO.13; @A.14]. Despite theoretical predictions and substantial experimental efforts a clear experimental signal for tetrahedral shapes is still absent (see the discussion in the introduction of Ref. [@ZLZZ.17]). In addition, symmetry unrestricted multidimensional constrained CDFT calculations are extremely time-consuming. Because of these reasons only reflection symmetric shapes with axial symmetry are considered in the present paper.
The most comprehensive study of octupole deformed shapes at the mean field level within the CDFT framework has been performed in Ref. [@AAR.16]. In this manuscript the global search for such shapes has been carried out in all $Z\leq 106$ even-even nuclei located between two-proton and two-neutron drip lines with two covariant energy density functionals (CEDFs) NL3\* and DD-PC1. As a result, a new region of octupole deformation, centered around $Z\sim 98, N\sim 196$ has been found in the CDFT framework for the first time. Based on the results obtained with these two functionals it was concluded that in terms of its size in the $(Z,N)$ plane and the impact of octupole deformation on binding energies this region is similar to the best known region of octupole deformed nuclei centered at $Z\sim 90, N\sim 136$. In addition, the systematic uncertainties in the description of the ground states of octupole deformed nuclei in the $Z\sim 58, N\sim 90$ lanthanides and $Z\sim 90, N\sim 136$ actinides have been defined for the first time in the CDFT framework using five state-of-the-art CEDFs representing different classes of the CDFT models.
However, the number of questions still remains unresolved in Ref. [@AAR.16]. The search for the answers on these questions is the main goal of this manuscript. First, there are the indications that octupole deformation can be present in the ground states of superheavy elements (SHE) with $Z\geq 108, N\sim 190$. They come from the results of the calculations within the microscopic+macroscopic (mic+mac) approach (Ref. [@MNMS.95]) and non-relativistic Hartree-Fock-Bogoliubov (HFB) method based on finite range Gogny D1S force (Ref. [@WE.12]). To our knowledge no search of octupole deformation in the ground states of superheavy $Z\geq 108$ nuclei has been performed within the CDFT framework so far. To fill this gap in our knowledge we will perform such a search in the region of proton numbers $108 \leq Z \leq 126$ and in the region of neutron numbers from the two-proton drip line up to neutron number $N = 210$. This region almost coincides with the region used in recent reexamination of the properties of SHE in the CDFT framework in Ref. [@AANR.15].
Second, we will establish systematic theoretical uncertainties in the predictions of the properties of the octupole deformed nuclei in the $Z\sim 98,
N\sim 196$ mass region and in superheavy nuclei. This is important since these nuclei will not be accessible with future facilities such as FRIB. However, the accounting of octupole deformation in the ground states of these nuclei is essential for the modeling of fission recycling in neutron star mergers [@GBJ.11; @JBPAGJ.15] since the gain in binding energy of the ground states due to octupole deformation will increase the fission barrier heights as compared with the case when octupole deformation is neglected.
To achieve these goals we use the four most up-to-date covariant energy density functionals of different types, with a nonlinear meson coupling (NL3\* [@NL3*]), with density-dependent meson couplings (DD-ME2 [@DD-ME2]), and with density-dependent zero-range interactions (DD-PC1 [@DD-PC1] and PC-PK1 [@PC-PK1]). They represent different classes of CDFT models (see discussion in Ref. [@AARR.14]). The functional DD-ME$\delta$ used in our previous studies of the global performance of CDFT [@AARR.13; @AARR.14; @AANR.15; @AAR.16; @AA.16; @AAR.17] is not employed here since it fails to reproduce octupole deformation in light actinides [@AAR.16] and inner fission barriers in superheavy nuclei [@AAR.17].
The paper is organized as follows. Section \[theory\_details\] describes the details of the solutions of the relativistic Hartree-Bogoliubov equations. Sec. \[oct-def-dep\] is devoted to the discussion of the ground state properties of octupole deformed nuclei and their dependence on the covariant energy density functional. The evolution of potential energy surfaces with proton and neutron numbers is discussed in Sec. \[PES-evolution\]. The assesment of systematic theoretical uncertainties in the predictions of ground state properties of octupole deformed nuclei and the comparison with other model predictions are performed in Sec. \[uncertainties\]. Finally, Sec. \[conclusions\] summarizes the results of our work.
The details of the theoretical calculations {#theory_details}
===========================================
The calculations have been performed in the Relativistic-Hartree-Bogoliubov (RHB) approach using parallel computer code RHB-OCT developed in Ref. [@AAR.16]. Note that only axial reflection asymmetric shapes are considered in this code.
The calculations in the RHB-OCT code perform the variation of the function $$\begin{aligned}
E_{RHB} + \sum_{\lambda=2,3} C_{\lambda 0}
(\langle\hat{Q}_{\lambda 0}\rangle-q_{\lambda 0})^2
\label{constr}\end{aligned}$$ employing the method of quadratic constraints. Here $E_{RHB}$ is the total energy (see Ref. [@AARR.14] for more details of its definition) and $\langle\hat{Q}_{\lambda 0}\rangle$ denote the expectation value of the quadrupole ($\hat{Q}_{20}$) and octupole ($\hat{Q}_{30}$) moments which are defined as $$\begin{aligned}
\hat{Q}_{20}&=&2z^2-x^2-y^2,\\
\hat{Q}_{30}&=&z(2z^2-3x^2-3y^2).\end{aligned}$$ $C_{20}$ and $C_{30}$ in Eq. (\[constr\]) are corresponding stiffness constants [@RS.80] and $q_{20}$ and $q_{30}$ are constrained values of the quadrupole and octupole moments. In order to provide the convergence to the exact value of the desired multipole moment we use the method suggested in Ref. [@BFH.05]. Here the quantity $q_{\lambda 0}$ is replaced by the parameter $q_{\lambda 0}^{eff}$, which is automatically modified during the iteration in such a way that we obtain $\langle\hat{Q}_{\lambda 0}\rangle = q_{\lambda 0}$ for the converged solution. This method works well in our constrained calculations. We also fix the (average) center-of-mass of the nucleus at the origin with the constraint $$\begin{aligned}
<\hat{Q}_{10}>=0\end{aligned}$$ on the center-of-mass operator $\hat{Q}_{10}$ in order to avoid a spurious motion of the center-of-mass.
![(Color online) The dependence of calculated quadrupole and octupole deformations, total binding energy and the $|\Delta E^{oct}|$ quantity on the number of fermionic shells employed in the RHB calculations for $^{290}$Cm with the DD-PC1 functional. The results obtained in octupole and quadrupole RHB codes in respective local minima with $(\beta_2 \neq 0, \beta_3 \neq 0)$ and $(\beta_2 \neq 0, \beta_3 = 0)$ are shown by solid black and dashed red curves, respectively. []{data-label="truncation"}](fig-1.eps){width="8.8cm"}
The charge quadrupole and octupole moments are defined as $$\begin{aligned}
Q_{20} &=& \int d^3r \rho({\bm r})\,(2z^2-r^2_\perp), \\
Q_{30} &=& \int d^3r \rho({\bm r})\,z(2z^2-3r^2_\perp)\end{aligned}$$ with $r^2_\perp=x^2+y^2$. In principle these values can be directly compared with experimental data. However, it is more convenient to transform these quantities into dimensionless deformation parameters $\beta_2$ and $\beta_3$ using the relations $$\begin{aligned}
Q_{20}&=&\sqrt{\frac{16\pi}{5}} \frac{3}{4\pi} Z R_0^2 \beta_2,
\label{beta2_def} \\
Q_{30}&=&\sqrt{\frac{16\pi}{7}}\frac{3}{4\pi} Z R_0^3 \beta_3
\label{beta4_def}\end{aligned}$$ where $R_0=1.2 A^{1/3}$. These deformation parameters are more frequently used in experimental works than quadrupole and octupole moments. In addition, the potential energy surfaces (PES) are plotted in this manuscript in the ($\beta_2,\beta_3$) deformation plane.
In order to avoid the uncertainties connected with the definition of the size of the pairing window [@KALR.10], we use the separable form of the finite range Gogny pairing interaction introduced by Tian et al [@TMR.09]. Its matrix elements in $r$-space have the form $$\begin{aligned}
\label{Eq:TMR}
V({\bm r}_1,{\bm r}_2,{\bm r}_1',{\bm r}_2') &=& \nonumber \\
= - G \delta({\bm R}-&\bm{R'}&)P(r) P(r') \frac{1}{2}(1-P^{\sigma})
\label{TMR}\end{aligned}$$ with ${\bm R}=({\bm r}_1+{\bm r}_2)/2$ and ${\bm r}={\bm r}_1-{\bm r}_2$ being the center of mass and relative coordinates. The form factor $P(r)$ is of Gaussian shape $$\begin{aligned}
P(r)=\frac{1}{(4 \pi a^2)^{3/2}}e^{-r^2/4a^2}\end{aligned}$$ The two parameters $G=728$ MeV$\cdot$fm$^3$ and $a=0.644$ fm of this interaction are the same for protons and neutrons and have been derived in Ref. [@TMR.09] by a mapping of the $^1$S$_0$ pairing gap of infinite nuclear matter to that of the Gogny force D1S [@D1S]. This pairing provides a reasonable description of pairing properties in the actinides (see Refs. [@AO.13; @AARR.14; @DABRS.15]) and has been used in our previous studies of octupole deformation in Ref. [@AAR.16][^1].
The potential energy surfaces are calculated in constrained calculations in the ($\beta_2,\beta_3$) plane for the $\beta_2$ values ranging from $-0.2$ up to 0.4 (ranging from $-0.6$ up to 0.2) if the ground state has prolate (oblate) deformation in the calculations of Ref. [@AANR.15]) and for the $\beta_3$ values ranging from 0.0 up to 0.3 with a deformation step of 0.02 in each direction. The energies of the local minima are defined in unconstrained calculations.
The effect of octupole deformation can be quantitatively characterized by the quantity $\Delta E_{oct}$ defined as $$\begin{aligned}
\Delta E_{oct} = E^{oct}(\beta_2, \beta_3) - E^{quad}(\beta'_2,\beta'_3=0)\end{aligned}$$ where $E^{oct}(\beta_2, \beta_3)$ and $E^{quad}(\beta'_2, \beta'_3=0)$ are the binding energies of the nucleus in two local minima of potential energy surface; the first minimum corresponds to octupole deformed shapes and second one to the shapes with no octupole deformation. The quantity $|\Delta E_{oct}|$ represents the gain of binding due to octupole deformation. It is also an indicator of the stability of the octupole deformed shapes. Large $|\Delta E_{oct}|$ values are typical for well pronounced octupole minima in the PES; for such systems the stabilization of static octupole deformation is likely. On the contrary, small $|\Delta E_{oct}|$ values are characteristic for soft (in octupole direction) PES typical for octupole vibrations. In such systems beyond mean field effects can play an important role (see Ref. [@AAR.16]) and references quoted therein).
The truncation of the basis is performed in such a way that all states belonging to the major shells up to $N_F=16$ ($N_F=18$ for superheavy $Z > 106$ nuclei) fermionic shells for the Dirac spinors and up to $N_B=20$ bosonic shells for the meson fields in the case of meson exchange functionals are taken into account. The dependence of the calculated quantities on $N_F$ is illustrated in Fig. \[truncation\]. One can see that all physical quantities of interest saturate with increasing of $N_F$. The comparison of the results shows that the calculations with $N_F=16$ reproduce the results of the $N_F=20$ truncation scheme with an accuracy of 0.007% or better for binding energies, 1.6% for the $|\Delta E^{oct}|$ quantity, 1.56% or better for quadrupole deformations and 2.3% for octupole deformation. Somewhat increased errors for deformations are the consequences of the softness of potential energy surface; for such PES some drift in the calculated equilibrium deformation is possible with little impact on total binding energy. Note that larger basis with $N_F = 18$ is used for superheavy nuclei with $Z > 106$. This increase of the basis fully compensates the increase of the proton number in the system. As a result, similar or better accuracy of the description of physical observables is obtained in superheavy nuclei. Thus, one conclude that employed truncation of the basis provides sufficient numerical accuracy of the calculations in the vicinity of the normal deformed minimum.
The properties of octupole deformed nuclei and their dependence on the covariant energy density functional {#oct-def-dep}
==========================================================================================================
The global search for octupole deformed nuclei has been performed for all even-even $Z=88-126$ nuclei from two-proton drip line up to either neutron number $N=210$ or two-neutron drip line (whichever comes first in neutron number) employing CEDFs NL3\*, DD-ME2, DD-PC1 and PC-PK1. Note that we use here the results obtained with CEDFs NL3\* and DD-PC1 in Ref. [@AAR.16] for the $Z=88-106$ nuclei. Contrary to the results obtained within the microscopic+macroscopic approach in Ref. [@MNMS.95] and HFB calculations with Gogny D1S force in Ref. [@WE.12], our calculations do not reveal the presence of octupole deformation in the ground states of superheavy nuclei with $Z\geq 110$. This issue will be discussed later in detail in Sec. \[uncertainties\].
{width="5.8cm"} {width="5.8cm"} {width="5.8cm"} {width="5.8cm"} {width="5.8cm"} {width="5.8cm"} {width="5.8cm"} {width="5.8cm"} {width="5.8cm"}
Fig. \[fig-global-CDFT\] shows the summary of the nuclei which possess octupole deformation in the ground state. The $Z \sim 92, N\sim 136$ actinides have been studied previously in detail in Ref. [@AAR.16] and they are shown here only for comparison with the $Z\sim 96, N\sim 196$ region of octupole deformation. In both regions, the number of even-even nuclei with calculated non-zero octupole deformation depends on the employed functional. There are 47 (44), 57 (38), 47 (31) and 64 (46) of such nuclei in the $Z\sim 96, N\sim 196$ ($Z \sim 92, N\sim 136$) region of octupole deformation in the calculations with the NL3\*, DD-PC1, PC-PK1 and DD-ME2 functionals, respectively. Thus, the calculations with CEDFs DD-ME2 and PC-PK1 confirm earlier CDFT predictions on the existence of new region of octupole deformation centered around $Z\sim 96, N\sim 196$ obtained with the CEDF NL3\* and DD-PC1 in Ref. [@AAR.16]. Most of the functionals predict that this region is substantially larger than the one around $Z\sim 92, N\sim 136$. Moreover, the maximum gain in binding due to octupole deformation is comparable in the $Z\sim 96, N\sim 196$ and $Z\sim 92, N\sim 136$ regions. This strongly suggests the stabilization of octupole deformation in the nuclei belonging to the central part of the $Z\sim 96, N\sim 196$ region.
The detailed information on calculated equilibrium quadrupole ($\beta_2$) and octupole ($\beta_3$) deformations as well as the gains ($\Delta E^{oct}$) in binding due to octupole deformation is summarized in Fig. \[Bet23-oct\]. These results show large similarities between the NL3\* and PC-PK1 functionals on the one hand and the DD-ME2 and DD-PC1 functionals on the other hand. The first pair of the functionals typically shows somewhat smaller gain in binding due to octupole deformation as compared with second one. This is likely due to the fact that the pairing is stronger in neutron rich nuclei for the first pair of the functionals as compared with second one (see Ref. [@AA.16]); strong pairing leads to the reduction of $|\Delta E^{oct}|$ (see Sec. V of Ref. [@AAR.16]). The differences/similarities in underlying shell structure could be another source of observed features.
For all functionals the maximum of the gain in binding energy due to octupole deformation takes place around $Z\sim 96, N\sim 196$. For nuclei in the vicinity of these particle numbers there is very little dependence of calculated equilibrium deformations on employed functional. However, on going away from these particle numbers the differences in calculated deformations increase because the nuclei become more soft in octupole deformation and thus more transitional in nature (see discussion in Sec. \[PES-evolution\]). In particular, the particle numbers at which the transition from quadrupole deformed to octupole deformed shapes takes place become strongly dependent on the employed functional.
Two $Z=108$ (two $Z=108$ and one $Z=110$) nuclei have non-zero octupole deformation in the calculations with CEDF DD-PC1 (DD-ME2) (see Figs. \[fig-global-CDFT\]b and c). They are not shown in Fig. \[Bet23-oct\] since all these nuclei are extremely soft in octupole deformation with very small gain in binding energy due to octupole deformation ($|\Delta E^{oct}| <0.1$ MeV).
Evolution of potential energy surfaces with particle numbers: an example of the DD-PC1 functional. {#PES-evolution}
==================================================================================================
In order to better understand the evolution and development of octupole deformation with particle number the potential energy surfaces (PES) of the Cm ($Z=96$) isotopes and $N=198$ isotones obtained in the RHB calculations with CEDF DD-PC1 are shown in Figs. \[Cm\_DD-PC1\] and \[N198\_DD-PC1\]. The center of this cross in the $(Z,N)$ plane represented by the $^{294}$Cm nucleus is located in the region of maximum gain of binding due to octupole deformation (see Fig. \[fig-global-CDFT\]).
{width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"}
The PES of the $^{286}$Cm nucleus are rather soft in the $\beta_3$ direction with the gain in binding due to octupole deformation being $|\Delta E_{oct}|=0.271$ MeV. The addition of the neutrons leads to the stabilization of octupole deformation in the $^{288-294}$Cm isotopes with largest gains in binding due to octupole deformation being 1.994 and 1.790 MeV in the $^{290}$Cm and $^{292}$Cm nuclei, respectively. Subsequent increase of the neutron number leads to the softening of potential energy surfaces so that $|\Delta E_{oct}|$ is rather small (0.049 MeV) in the $^{298}$Cm nucleus.
{width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"}
The PES for the $N=198$ isotones are displayed in Fig. \[N198\_DD-PC1\]. One can see that the lowest $Z$ nucleus ($^{288}$Th with $Z=88$) shown in this figure already has well pronounced minimum in octupole deformation which is characterized by $|\Delta E_{oct}|=1.084$ MeV. This is because the $^{286}$Rn nucleus with lower $Z$ value ($Z=86$), which is expected to be more octupole soft, is located beyond two-neutron drip line (see Fig. \[fig-global-CDFT\]). The $^{290}$U, $^{292}$Pu, $^{294}$Cm, $^{296}$Cf and $^{298}$Fm nuclei have well pronounced octupole minima in the PES. The largest gain in binding due to octupole deformation $|\Delta E_{oct}|=1.419$ MeV is reached in the $^{292}$Pu nucleus. Subsequent increase of proton number above $Z=100$ gradually decreases $|\Delta E_{oct}|$ so that PES surface becomes very soft in $^{302}$Rf.
Assessing systematic uncertainties in model predictions {#uncertainties}
=======================================================
All theoretical approaches to nuclear many body problem are based on some approximations. For example, in the DFT framework, there are two major sources of these approximations, namely, the range of interaction and the form of the density dependence of the effective interaction [@BHP.03; @BB.77]. In the non-relativistic case one has zero range Skyrme and finite range Gogny forces and different density dependencies [@BHP.03]. A similar situation exists also in the relativistic case: point coupling and meson exchange models have an interaction of zero and of finite range, respectively [@VALR.05; @DD-ME2; @NL3*; @DD-PC1]. The density dependence is introduced either through an explicit dependence of the coupling constants [@TW.99; @DD-ME2; @DD-PC1] or via non-linear meson couplings [@BB.77; @NL3*]. This ambiguity in the definition of the range of the interaction and its density dependence leads to several major classes of the covariant energy density functionals which were discussed in detail in Ref. [@AARR.14].
These approximations lead to theoretical uncertainties in the description of physical observables. While in known nuclei these uncertainties could be minimized by benchmarking the model description to experimentally known nuclei (for example, via the fitting protocol), they grow in magnitude when we extrapolate beyond known regions [@DNR.14; @AARR.14]. In such a situation, the estimate of theoretical uncertainties is needed. This issue has been discussed in detail in Refs. [@RN.10; @DNR.14] and in the context of global studies within CDFT in the introduction of Ref. [@AARR.14] and in Ref. [@AAR.17]. In the CDFT framework, systematic theoretical uncertainties and their sources have been studied globally for the ground state masses, deformations, charge radii, neutrons skins, positions of drip lines etc in Refs. [@AARR.13; @AARR.14; @AARR.15; @AANR.15; @AAR.16; @AA.16] and for inner fission barriers in superheavy nuclei in Ref. [@AAR.17].
In the present manuscript, we focus on the uncertainties related to the choice of the energy density functional. Similar to our previous studies ([@AARR.14; @AANR.15; @AAR.16; @AA.16; @AAR.17]), we define systematic theoretical uncertainty for a given physical observable (which we call in the following “spreads”) via the spread of theoretical predictions as [@AARR.14] $$\Delta O(Z,N) = |O_{max}(Z,N) - O_{min}(Z,N)|,$$ where $O_{max}(Z,N)$ and $O_{min}(Z,N)$ are the largest and smallest values of the physical observable $O(Z,N)$ obtained within the set of CEDFs under investigation for the $(Z,N)$ nucleus.
These spreads for the calculated quadrupole and octupole deformations as well as for the $|\Delta E^{oct}|$ quantity are shown in Fig. \[fig-spreads\]. One can see that the spreads for the $\beta_2$ and $\beta_3$ deformations in the central parts of the $Z\sim 96, N\sim 196$ and $Z\sim 92, N\sim 136$ regions are small. They increase at the boundaries of these regions where the PES of the nuclei are soft in octupole deformation. As a result, model predictions become strongly dependent on fine details of underlying single-particle structure so that the same $(Z,N)$ nucleus could be octupole deformed in one functional but only quadrupole deformed in another functional (see Fig. \[fig-global-CDFT\]). Similar situation with low reliability of theoretical predictions in some parts of nuclear chart has been seen earlier in the transitional regions between quadrupole deformed and spherical shapes (see Figs. 18 and 20 in Ref. [@AARR.14]) in the axial RHB calculations restricted to reflection symmetric shapes. The $Z=108, 110$ nuclei with $N=188$ show very large spreads in quadrupole deformation (Fig. \[fig-spreads\]a). These two nuclei are octupole deformed with $\beta_2 \sim -0.045, \beta_3 \sim 0.07$ only in the calculations with the DD-ME2 functional. However, they are spherical in the calculations with CEDFs NL3\* and PC-PK1 but oblate (with $\beta_2 \sim -0.36$) in the calculations with DD-PC1 (see Fig. 6 in Ref. [@AANR.15]).
Systematic theoretical uncertainties for the energy gain due to octupole deformation are shown in Fig. \[fig-spreads\]c. These uncertainties show different pattern in the $(Z,N)$ plane as compared with the uncertainties for the $\beta_2$ and $\beta_3$ deformations (Figs. \[fig-spreads\]a and b). The maximum uncertainties for the $|\Delta E^{oct}|$ quantity exists in the left bottom corners of the $Z\sim 96, N\sim 196$ and $Z\sim 92, N\sim 136$ regions of octupole deformation. Theoretical uncertainties gradually decrease on going away from these corners and become quite small at the boundaries of the regions of octupole deformation. This is not surprising considering the fact that the nuclei at these boundaries are octupole soft with rather small gain in binding due to octupole deformation.
It is important to compare the CDFT predictions for the $Z\sim 96, N\sim 196$ region of octupole deformation with the ones obtained in non-relativistic theories. Such a comparison is presented in Fig. \[fig-global\] where two extreme CDFT predictions for octupole deformed region \[the largest (smallest) $Z\sim 96, N\sim
196$ region of octupole deformation is obtained in the calculations with DD-ME2 (PC-PK1) functional\] in the indicated part of nuclear chart are compared with the predictions obtained in the Skyrme and Gogny DFTs and macroscopic+microscopic approach. The Skyrme DFT calculations with the SLy6 functional predict such a region with the center located around $Z=100, N=190$ [@ELLMR.12] (see Fig. \[fig-global\]c). Similar region of octupole deformation (but with smaller gain in binding energy due to octupole deformation) has also been obtained in the calculations with the SV-min EDF [@ELLMR.12]. The Gogny DFT calculations are limited in the $(Z,N)$ plane (see Fig. \[fig-global\]d); even then they do not indicate the presence of octupole deformation in the nuclei located in the upper parts of the regions of octupole deformation obtained in the Skyrme and CDFT calculations. However, the extension of the Gogny DFT calculations to the $Z=90-108, N=180-210$ region of nuclear chart is needed to clarify the question of the existence of the $Z\sim 96,
N\sim 196$ region of octupole deformation in this type of the EDFs. On the contrary, the mic+mac calculations of Ref. [@MNMS.95] predict the existence of octupole deformation in this region (Fig. \[fig-global\]e). However, the island of octupole deformation is smaller than the one obtained in the CDFT or Skyrme DFT calculations and it is centered around $Z=100, N=184$. It is necessary to mention that that these results have been obtained more than twenty years ago. Newer mic+mac calculations of Ref. [@MBCOISI.08] do not cover this part of nuclear chart. However, in the $Z\sim 92, N\sim 134$ region of octupole deformation, the number of octupole deformed even-even nuclei is increased from 20 in Ref. [@MNMS.95] to 27 in Ref. [@MBCOISI.08]. It would be interesting to see how the number of octupole deformed nuclei in the $Z\sim 100, N\sim 184$ region would be modified if newer formalism of mic+mac approach of Ref. [@MBCOISI.08] with improved model parameters would be applied to this region.
Although placing the center of the island of octupole deformed nuclei at different particle numbers (at $Z\sim 96, N\sim 196$ in CDFT, at $Z\sim 100,
N\sim 190$ in Skyrme DFT and at $Z\sim 100, N\sim 184$ in mic+mac approach), modern theories agree on the existence of such island in neutron-rich actinides and low-$Z$ superheavy nuclei. However, their predictions diverge for the $Z \geq 110$ superheavy nuclei. The CDFT calculations of the present manuscript and the Skyrme DFT calculations of Ref. [@ELLMR.12] do not predict the existence of octupole deformation in the ground states of the $110 \leq Z \leq 126$ and $110 \leq Z \leq 120$ superheavy nuclei, respectively. On the contrary, the Gogny DFT (Fig. \[fig-global\]d and Ref. [@WE.12]) and mic+mac (Fig. \[fig-global\]e and Ref. [@MNMS.95]) calculations predict the existence of such nuclei. The HFB calculations based on the Gogny D1S force predict octupole deformation in the ground states of the $(Z=108-126, N=186-190)$ even-even nuclei (see Fig. 3 in Ref. [@WE.12]). These nuclei either do not have quadrupole deformation (the $N=186$ and some $N=188$ nuclei) or this deformation is rather small ($\beta_2 <0.1$) for $N=190$ and some $N=188$ nuclei. The octupole deformation is rather small for most of these nuclei apart of few $N=188$ nuclei and the majority of the $N=190$ nuclei which have substantial octupole deformation $\beta_3$ exceeding 0.1. Note that these calculations cover only nuclei with $N\leq 190$. More extensive mic+mac calculations of Ref. [@MNMS.95] indicate larger region of octupole deformation in the superheavy nuclei (see Fig. \[fig-global\]e).
The existence of octupole deformed shapes is dictated by the underlying shell structure. Strong octupole coupling exists for particle numbers associated with a large $\Delta N=1$ interaction between intruder orbitals with $(l,j)$ and normal-parity orbitals with $(l-3,j-3)$ [@BN.96]. Thus, the discussed above differences in the model predictions are traced back to the differences in the underlying single-particle structure. For normal deformed nuclei not far away from beta stability the tendency towards octupole deformation or strong octupole correlations occurs just above closed shells. For example, in the CDFT the maximum of octupole correlations takes place in the $A\sim 230$ region of octupole deformation at proton number $Z\sim 92$ (the coupling between the proton 1$i_{13/2}$ and 2$f_{7/2}$ orbitals) and 136 (the coupling between the neutron 1$j_{15/2}$ and 2$g_{9/2}$ orbitals). In the $Z\sim 96, N\sim 196$ region, the presence of octupole deformation is due to the interaction of the 2$h_{11/2}$ and 1$k_{17/2}$ neutron orbitals and of the 1$i_{13/2}$ and 2$f_{7/2}$ proton orbitals. Note that the maximum of the interaction of proton orbitals occurs at a higher proton number $Z$ as compared with the well known $A\sim 230$ region of octupole deformation in actinides. In the $Z\sim 120, N\sim 190$ region, the interaction of the 2$h_{11/2}$ and 1$k_{17/2}$ neutron orbitals and of the 1$j_{15/2}$ and 2$g_{9/2}$ proton orbitals are responsible for strong octupole correlations in the Gogny DFT and mic+mac calculations. However, the energies of these states and their positions with respect of the Fermi level are described differently in different models (see, for example, Figs. 1, 4, 9 and 15 in Ref. [@BRRMG.99], Fig. 4 in Ref. [@BNR.01], and Fig. 1 in Ref. [@AANR.15]).
The predictive power of above discussed models in the description of these energies and, as a consequence, of the regions of octupole deformation decreases on going away from known region of nuclear chart. Some differences in the predictions of the region of octupole deformation do already exist for known $A \sim 230$ region of octupole deformation (see Fig. \[fig-global\] and discussion in Ref. [@AAR.16]). However, they become magnified with increasing of neutron number up to $N \sim 196$ on going to the $Z\sim 96, N\sim 196$ region of octupole deformation and especially pronounced with an additional increase of neutron number up to $Z\sim 120$. In the $Z\sim 120, N\sim 190$ region, there is a substantial discrepancies in model predictions. Note that in this region of nuclear chart the state-of-the-art theories disagree even in the prediction of large spherical shell gaps and thus of the properties of superheavy nuclei [@BRRMG.99; @SP.07; @AANR.15].
The impact of pairing strength changes {#pairing}
======================================
The extrapolations beyond the known region of nuclei are associated with theoretical uncertainties. The systematic uncertainties related to the form of the functional were quantified in Sec. \[uncertainties\]; note that they are related to the particle-hole channel of the DFTs. In addition, there are the uncertainties in the particle-particle (pairing) channel; they are expected to become especially large in the vicinity of the two-neutron drip line (see Refs. [@PMSV.13; @AARR.15]). The study of $^{218-234}$Th isotopes in Sect. V of Ref. [@AAR.16] showed that in general pairing counteracts the shell effects. As a result, the strongest trend towards octupole deformation is seen in the systems with no pairing, while the increase of pairing suppresses it. The modification of the pairing strength may also lead to the changes in the topology of potential energy surfaces.
As illustrated in Figs. \[Cm286\_pairing\] and \[Cm290\_pairing\] these features are also present in neutron-rich actinides. The $^{286}$Cm and $^{290}$Cm nuclei are used here as the examples and the scaling factor $f$ of the pairing strength is varied in indicated range. This is a factor by which the matrix elements of Eq. (\[TMR\]) are multiplied. Based on previous studies of the pairing in the CDFT framework in Refs. [@AO.13; @AARR.15] the variations of the scaling factor in the range of $\pm 3\%$ with respect of $f=1.0$ should be considered as most reasonable, but still larger variations could not be excluded. The $^{286}$Cm nucleus, located at the borderline of the octupole deformed region (see Fig. \[fig-global-CDFT\]c), is characterized by PES which is extremely soft in octupole direction (Fig. \[Cm286\_pairing\]c). The $^{290}$Cm is located at the center of the island of octupole deformation (Fig. \[fig-global-CDFT\]c) and is characterized by deep octupole minimum with large $|\Delta E^{oct}| \sim 2.0$ MeV (see Fig. \[truncation\]d). The impacts of the scaling factor $f$ changes on the gain in binding due to octupole deformation and on equilibrium deformations are summarized in Tables \[table-f-impact-binding\] and \[table-f-impact-deformation\], respectively. Similar to the results presented in Sec. V of Ref. [@AAR.16], the reduction of pairing strength leads to more pronounced octupole minimum in both nuclei. On the contrary, the increase of pairing strength reduces the depth of octupole minimum in $^{290}$Cm and makes the $^{286}$Cm nucleus spherical. Thus, one can conclude that weaker (stronger) pairing would make the island of octupole deformation broader (narrower) with more (less) pronounced gains in binding due to octupole deformation in nuclei. The impact of the modification of the pairing strength on the equilibrium deformation is small in $^{290}$Cm. Similar situation exists also in $^{286}$Cm for $f=0.94-1.00$. However, further increase of $f$ triggers transition to spherical shape.
Nucleus $f=0.94$ $f=0.97$ $f=1.00$ $f=1.03$ $f=1.06$
------------ ---------- ---------- ---------- ---------- ----------
$^{286}$Cm 1.089 0.696 0.271 0.0 0.0
$^{290}$Cm 2.680 2.363 1.994 1.735 1.434
: The gain in binding $|\Delta E^{oct}|$ (in MeV) due to octupole deformation calculated for different values of scaling factor $f$ of the pairing.[]{data-label="table-f-impact-binding"}
Nucleus $f=0.94$ $f=0.97$ $f=1.00$ $f=1.03$ $f=1.06$
------------ -------------- -------------- -------------- -------------- --------------
$^{286}$Cm 0.100, 0.113 0.099, 0.110 0.095, 0.105 0.00, 0.00 0.00, 0.00
$^{290}$Cm 0.131, 0.127 0.132, 0.127 0.131, 0.126 0.134, 0.124 0.135, 0.121
{width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"}
{width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"} {width="5.9cm"}
Conclusions
===========
A systematic search for axial octupole deformation has been performed in the actinides and superheavy nuclei for proton numbers $Z=88-126$ and neutron numbers from two-proton drip line up to $N=210$ using four state-of-the-art covariant energy density functionals. Systematic theoretical uncertainties in the description of physical observables of octupole deformed nuclei have been estimated. The main results can be summarized as follows:
- The present CDFT investigation confirms our earlier predictions on the existence of the region of octupole deformation centered around $Z\sim 96, N\sim 196$ obtained with the DD-PC1 and NL3\* functionals [@AAR.16]. Most of the CEDFs predict the size of this region in the $(Z,N)$ plane larger than the one at $Z\sim 92,
N\sim 136$. On the other hand, the impact of octupole deformation on the binding energies of the nuclei in these two regions are comparable. Similar region of octupole deformation is predicted also in Skyrme DFT [@ELLMR.12] and mic+mac [@MNMS.95] calculations. However, it is centered at $Z\sim 100,
N\sim 190$ in the Skyrme DFT calculations and at $Z\sim 100, N\sim 184$ in mic+mac calculations.
- Systematic theoretical uncertainties in the predictions of quadrupole ($\beta_2$) and octupole ($\beta_3$) deformations as well as the gain in binding due to octupole deformation $|\Delta E^{oct}|$ have been quantified within the CDFT framework. They are comparable in the $Z\sim 96, N\sim 196$ and $Z\sim 92, N\sim 136$ regions of octupole deformation.
- The search for octupole deformation in the ground states of even-even superheavy $Z=108-126$ nuclei has been performed in the CDFT framework for the first time. With exception of two $Z=108$ (two $Z=108$ and one $Z=110$) octupole deformed nuclei in the calculations with CEDF DD-PC1 (DD-ME2), we do not find octupole deformed shapes in the ground states of these nuclei. These results are in agreement with the ones obtained in the Skyrme DFT but disagree with the ones obtained in Gogny DFT and mic+mac calculations. The latter calculations indicate the presence of large island of octupole deformed $Z>110$ nuclei centered around $N\sim 190$. These differences in the location of the islands of octupole deformed nuclei are due to the differences in the underlying single-particle structure.
Acknowledgements
================
This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics under Award No. DE-SC0013037.
[^1]: By mistake the parameters $G=738$ MeV$\cdot$fm$^3$ and $a=0.636$ fm, derived from the D1 Gogny force [@TMR.09], are quoted in Ref. [@AAR.16]. In reality, the same parameters $G=728$ MeV$\cdot$fm$^3$ and $a=0.644$ fm as the ones employed in the present manuscript are used in the calculations of Ref. [@AAR.16].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We give a discrete characterization of the trace of a class of Sobolev spaces on the Sierpinski gasket to the bottom line. This includes the $L^2$ domain of the Laplacian as a special case. In addition, for Sobolev spaces of low orders, including the domain of the Dirichlet form, the trace spaces are Besov spaces on the line.'
address:
- 'Department of Mathematics, Cornell University, Ithaca 14853, USA'
-
- 'Department of Mathematics, Cornell University, Ithaca 14853, USA'
- 'Department of Mathematics, Cornell University, Ithaca 14853, USA'
author:
- Shiping Cao
- Shuangping Li
- 'Robert S. Strichartz'
- Prem Talwai
title: A trace theorem for Sobolev spaces on the Sierpinski gasket
---
[^1]
[^2]
Introduction {#intro}
============
This work deals with the restriction problem for functions in Sobolev spaces on the Sierpinski gasket ($\SG$) to the bottom line. A special case was studied by A. Jonsson in [@jonsson1], where the trace for the Dirichlet form was characterized.
![the Sierpinski gasket.[]{data-label="SG"}](SGlimit.pdf){width="5cm"}
(0,0) (-5,0)[$q_2$]{} (-152,0)[$q_1$]{} (-79,125)[$q_0$]{}
Let’s briefly review Jonsson’s result here. $\SG$ is the attractor of the iterated function system (i.f.s.) in the plane $\mathbb{R}^2$ $$F_i(x)=\frac{1}{2}x+\frac{q_i}{2},\quad i=0,1,2,$$ where $q_0,q_1,q_2$ are vertices of an equilateral triangle. See Figure \[SG\] for a picture of $\SG$. We identify $I=[0,1]$ with the bottom line $\overline{q_1q_2}$ by $$\label{eqn11}
t=tq_1+(1-t)q_2,\quad \forall t\in I.$$ Consider the standard Dirichlet form $(\mathcal{E},dom\mathcal{E})$ on $\SG$. This form was studied in connection with the Brownian motion on $\SG$ (see [@BP; @jonsson2; @Lindstrom]), and was constructed in a pure analytically approach by J. Kigami in [@k1; @k2]. A. Jonsson showed the following theorem.
\[thm11\] Let $\alpha=\frac{\log2+\log5-\log3}{2\log2}(\approx0.868483)$ and $B_\alpha^{2,2}(I)$ be the Besov space on $I$. Then $dom\mathcal{E}|_{I}=B_\alpha^{2,2}(I)$.
The above theorem was extended to a wide class of self-similar sets [@HK], where the trace theorem for Dirichlet forms to self-similar subsets were established. Also, read [@HK; @kumagai] for an application of the trace theorems to penetrating processes.
Recently, many related works emerge, including the trace theorem on the middle line of $\SG$ (see [@cq4; @bvphalf]), and boundary value problems on the upper half domain of $\SG$ (see [@cq2; @bvp2; @Hua; @bvp1]). However, there have not been further results telling us what is the trace of the domain of the Lapalcian and other Sobolev spaces on $\SG$ to $I$. In this work, we will give an answer to the above question.
Below, we briefly introduce our results. We choose $\mu$ to be the Hausdoff measure on $\SG$, satisfying $\mu(\SG)=1$ and $\mu(A)=\sum_{i=0}^2\frac{1}{3}\mu(F_i^{-1}A)$ for each Borel set $A$. For $u\in dom\mathcal{E}$, we say $u\in dom_{L^2}\Delta(\SG)$ with $\Delta u=f$ if $$\mathcal{E}(u,v)=-\int_{\SG} fvd\mu$$ holds for each $v\in dom_0\mathcal{E}:=\{v\in dom\mathcal{E}:v(q_i)=0,\text{ for } i=0,1,2\}$. In our work, we consider Sobolev spaces $L^2_\sigma(\SG)$ with $0\leq\sigma\leq 2$, which can be defined as follows.
\[def12\] Define the Sobolev space $L^2_0(\SG)=L^2(\SG)$ with norm $\|u\|_{L^2_0(\SG)}=\|u\|_{L^2(\SG)}$, and define $L^2_2(\SG)=dom_{L^2}\Delta(\SG)$ with norm $\|u\|^2_{L^2_2(\SG)}=\|u\|^2_{L^2(SG)}+\|\Delta u\|^2_{L^2(SG)}$.
For $0\leq \sigma\leq 2$, we define Sobolev spaces to be $L^2_\sigma(\SG)=[L^2_0(\SG),L^2_2(\SG)]_{\sigma/2}$, where $[X,Y]_\theta$ denotes the complex interpolation space of $X,Y$.
In [@s2], Strichartz gave systematic discussions on Sobolev spaces and other function spaces, where Sobolev spaces were defined in more general settings. Related works on properites of Sobolev spaces can be found in [@cq1; @cq4; @GL; @HW; @HM; @pseudo]. Also read [@HKM; @sw] for recent developments on $p-$Laplacian and the corresponding $L^p$ Sobolev spaces.
In our first main result, we have two critical orders $$b_1=\frac{\log3}{\log5}(\approx 0.682606),\quad b_2=\frac{\log\frac{25(17-\sqrt{73})}{36}}{\log5}(\approx 1.09991).$$ Also, we define the function $\alpha(\sigma):=\frac{\log2+\sigma\log5-\log3}{2\log2}(\approx1.16096\sigma-0.292481)$, which is the unique number such that $5^{\sigma}3^{-1}=2^{2\alpha-1}$. Noticing that $L^2_1(\SG)=dom\mathcal{E}$ (see [@cq4]) is included as a special case, the following Theorem \[thm13\] can be viewed as a direct extension of Jonsson’s theorem.
\[thm13\] Let $b_1<\sigma<b_2$ and $\alpha=\alpha(\sigma)$. Then $L^2_\sigma(\SG)|_{I}=B^{2,2}_\alpha(I)$.
$b_1$ in the above theorem is the critical order for the continuity of functions in Sobolev spaces (see [@cq1; @GL; @s2]), and one can check $\alpha(b_1)=\frac{1}{2}$, the well known critical order for Sobolev spaces on the line. The complicated upper bound $b_2$ has an explanation in Corollary \[coro33\], where the trace of harmonic functions is studied.
On the other hand, Besov spaces on the line segment are no longer the trace spaces of Sobolev spaces for higher orders. To describe the trace spaces, we will define a difference operator $D$. To be more precise, we define $Df(n,k)$ to be a linear combination of the values of $f\in C(I)$ at $\frac{k}{2^n}$ and some neighbouring points. The space $\mathcal{T}_\sigma$ will be discretely characterized as $$\mathcal{T}_{\sigma}=\big\{f\in C(I):\sum_{n=2}^\infty\sum_{k=1}^{2^n-1} 5^{\sigma n}3^{-n}|Df(n,k)|^2<\infty\big\}.$$ Details can be found in Definition \[def35\] and \[def45\]. We will prove the following trace theorem.
\[thm14\] Let $b_1<\sigma\leq2$. Then $L^2_\sigma(\SG)|_{I}=\mathcal{T}_\sigma$.
As supplement, we will show that $B^{2,2}_\alpha(I)\subset \mathcal{T}_\sigma$ for $b_1<\sigma<\frac{\log6}{\log5}(\approx 1.11328)$, and $B^{2,2}_\alpha(I)= \mathcal{T}_\sigma$ for $b_1<\sigma<b_2$. In addition, $\mathcal{T}_\sigma$ is stable under complex interpolation.
In the end, we briefly introduce the structure of this paper. In section 2, we will review the Dirichlet form and harmonic functions on $\SG$, and introduce some notations and tools. In section 3, we will prove Theorem \[thm13\]. Some preparations for Theorem \[thm14\] will be included. In section 4, we will construct the trace space $\mathcal{T}_\sigma$, and prove Theorem \[thm14\]. In Section 5, we will talk about some related results.
Throughout the paper, we always use the notation $f\lesssim g$ if there is a constant $C>0$ such that $f\leq Cg$, and write $f\asymp g$ if $f\lesssim g$ and $g\lesssim f$. Also, we will keep using the critical numbers $b_1,b_2$ and the function $\alpha(\sigma)$ without further specifying.
The Dirichlet form and harmonic functions
=========================================
For convenience of readers, we briefly reivew the Dirichlet form and the harmonic functions on $\SG$ in this section. Some easy lemmas and important tools will also be given. More details can be found in books [@k3; @s1].
Recall that $\SG$ is the attractor of the i.f.s $\{F_i\}_{i=0}^2$, i.e. $$\SG=\bigcup_{i=0}^2F_i\SG.$$ We call each $F_i\SG$ a level-$1$ cell. More generally, define $W_n=\{0,1,2\}^n$ for $n\geq 1$, and set $W_0=\{\emptyset\}$ for uniformity. For each finite word $w=w_1w_2\cdots w_n\in W_*=\bigcup_{m=0}^\infty W_m$, we denote $|w|=n$ the length of the word, and write $F_w=F_{w_1}F_{w_2}\cdots F_{w_n}$ for short. In particular, $F_\emptyset=Id$ is the identity map. We call $F_w\SG$ a level-$n$ cell if $|w|=n$.
We call $V_0=\{q_0,q_1,q_2\}$ the set of boundary vertices of $\SG$, and define the set of level-n vertices $V_n=\bigcup_{w\in W_n} F_wV_0$. For convenience, for $n\geq 1$, let $$\tilde{V}_n=V_n\setminus V_{n-1},$$ and set $\tilde{V}_0=V_0$. The set of vertices $V_*=\bigcup_{n=0}^\infty V_m=\bigsqcup_{n=0}^\infty \tilde{V}_n$ is a dense subset of $\SG$.
On $\SG$, J. Kigami [@k1; @k2] constructed the self-similar energy form by defining it as the limit of a sequence of discrete Dirichlet forms on $V_n$. For each $u\in l(V_n)$, define $$\mathcal{E}_n(u)=(\frac{5}{3})^n\sum_{w\in W_n}\sum_{i\neq j}(u(F_wq_i)-u(F_wq_j))^2.$$ $\{\mathcal{E}_n(u)\}$ is a nondecreasing sequence for each $u\in C(\SG)$, so we can define $\mathcal{E}(u)=\lim\limits_{n\to\infty} \mathcal{E}_n(u)$. Set $dom\mathcal{E}=\{u\in C(\SG):\mathcal{E}(u)<\infty\}$. For $u,v\in dom\mathcal{E}$, we can use polarization to give a bilinear form $$\mathcal{E}(u,v)=\frac{1}{4}\big(\mathcal{E}(u+v)-\mathcal{E}(u-v)\big).$$ It is well known that $(\mathcal{E},dom\mathcal{E})$ is a local regular Dirichlet form on $\SG$ with the Hausdoff measure $\mu$.
Given any boundary value $h_0\in l(V_0)$, there is a unique extension $h_1\in l(V_1)$ that minimizes the energy $\mathcal{E}_1$, i.e. $\mathcal{E}_1(h_1)=\min\{\mathcal{E}_1(u):u\in l(V_1),u|_{V_0}=h_0\}$. The extension algorithm is shown in Figure \[extensionalgorithm\].
![The harmonic function with $h(q_0)=a,h(q_1)=b,h(q_2)=c$.[]{data-label="extensionalgorithm"}](G1.pdf){width="5cm"}
(0,0) (-78,122)[$a$]{} (-150,0)[$b$]{} (-4,0)[$c$]{} (-143,63)[$\frac{2a+2b+c}{5}$]{} (-40,63)[$\frac{2a+b+2c}{5}$]{} (-92,-8)[$\frac{a+2b+2c}{5}$]{}
The above algorithm is local, which means it can be applied to each cell. So we get a sequence of extensions $h_n\in l(V_n)$ that minimize the energy $\mathcal{E}_n$, and $h_n$ converges to $h\in dom\mathcal{E}$. See [@s1] for details. $h$ is called a harmonic function, and we denote by $\mathcal{H}_0$ the space of harmonic functions on $\SG$. Clearly, $\mathcal{H}_0$ is of three dimension, since each harmonic function is uniquely determined by its boundary values. The following lemma can be derived from direct computation.
\[lemma21\]Let $h$ be a harmonic funciton on $\SG$. For $n\geq 2$ and $1\leq k\leq 2^{n-1}$, we have $$h(\frac{2k-1}{2^n})=\begin{cases}
\frac{4}{5}h(\frac{k}{2^n})+\frac{8}{25}h(\frac{k-1}{2^n})-\frac{3}{25}h(\frac{k+1}{2^n}),\text{ if $k$ is odd},\\
\frac{4}{5}h(\frac{k-1}{2^n})+\frac{8}{25}h(\frac{k}{2^n})-\frac{3}{25}h(\frac{k-2}{2^n}),\text{ if $k$ is even}.
\end{cases}$$
*Proof.* By using the harmonic extension algorithm twice, we have $$\begin{aligned}
h(\frac{1}{2})=\frac{2}{5}h(0)+&\frac{2}{5}h(1)+\frac{1}{5}h(q_0),\\
h(\frac{1}{4})=\frac{16}{25}h(0)+\frac{1}{5}h(1)+\frac{4}{25}h(q_0),&\quad h(\frac{3}{4})=\frac{1}{5}h(0)+\frac{16}{25}h(1)+\frac{4}{25}h(q_0),
\end{aligned}$$ where $0=q_1,1=q_2,\frac{1}{2}=F_1q_2,\frac{1}{4}=F^2_1q_2$ and $\frac{3}{4}=F_2^2q_1$ as we set in equation (\[eqn11\]). For larger $n$, we can do the same computation locally on the level-$(n-2)$ cell containing $\frac{2k-1}{2^n}$. Then, it is direct to check the lemma. $\square$
Analogously to the definition of harmonic functions, for each $m\geq 1$ and $x\in \tilde{V}_m=V_m\setminus V_{m-1}$, we can define the tent function $\varphi_x$ by giving the initial value on $V_m$ as follows $$\varphi_x(y)=\begin{cases}
0,\text{ if }y\neq x,\\
1,\text{ if }y=x,
\end{cases}$$ and taking harmonic extension in $\SG\setminus V_m$. Clearly, $\varphi_x$ is harmonic in each level-$m$ cell, and Lemma \[lemma21\] holds for $\varphi_x$ when $n\geq m+2$.
In our work, we will use the following characterization of Sobolev spaces. For the full version and proof, see Theorem 7.11 in [@cq1].
\[th22\] Let $b_1=\frac{\log3}{\log5}<\sigma<2-\frac{\log3}{\log5}$, the series $f=h+\sum_{n=1}^\infty\sum_{x\in \tilde{V}_n} c_x\varphi_x$ belongs to $L^2_\sigma(SG)$ if and only if $$\sum_{n=1}^\infty \sum_{x\in\tilde{V}_n} 5^{n\sigma}3^{-n}|c_x|^2<\infty.$$ In addition, each $f\in L^2_\sigma(SG)$ has a unqiue expanison of the above form, with $\|f\|_{L^2_\sigma(SG)}\asymp (\|h\|^2_{L^2(SG)}+\sum_{n=1}^\infty\sum_{x\in \tilde{V}_n} 5^{n\sigma}3^{-n}|c_x|^2)^{\frac{1}{2}}$.
An extension of A. Jonsson’s Theorem
====================================
In this section, we study the trace theorem for Sobolev spaces of low orders. The result, Theorem \[thm13\], is a direct extension of A. Jonsson’s trace theorem. In the following, we will study the restriction map and the extension map seperately. The two parts together imply Theorem \[thm13\].
A restriction theorem
---------------------
In this part, we follow A. Jonsson’s idea to show a restriction theorem. Recall the fact from [@Kamont] that for $\frac{1}{2}<\alpha<1$, a function $f$ belong to $B^{2,2}_\alpha(I)$ if and only if the following expression $$(|f(0)|^2+|f(1)|^2)^{1/2}+\big(\sum_{n=1}^\infty 2^{2n\alpha}2^{-n}\sum_{k=1}^{2^n}|f(\frac{k}{2^n})-f(\frac{k-1}{2^n})|^2\big)^{1/2}$$ is finite and the norm of $f$ in $B^{2,2}_\alpha(I)$ is equivalent to this expression.
We introduce the following notation to shorten the above expression.
\[def31\] Let $f\in C(I)$. Define $A_n(f)$ to be a vector of length $2^n$, such that $$A_n(f)_k=f(\frac{k}{2^n})-f(\frac{k-1}{2^n}),\forall 1\leq k\leq 2^n.$$
With the above notation, we have $$\|f\|_{B^{2,2}_\alpha(I)}\asymp \big(|f(0)|^2+|f(1)|^2+\sum_{n=1}^\infty 2^{2n\alpha}2^{-n}\|A_n(f)\|^2_{l^2}\big)^{1/2}.$$
We begin with harmonic functions.
\[prop32\] Let $h$ be a harmonic function on $\SG$. Then for $n\geq 0$, we have $$\label{eqn31}
\|A_{n+2}(h|_I)\|^2_{l^2}=\frac{17}{25}\|A_{n+1}(h|_I)\|^2_{l^2}-\frac{54}{625}\|A_{n}(h|_I)\|^2_{l^2}.$$ As a consequence, there exist constants $C_1,C_2$ such that $$\|A_n(h|_I)\|^2_{l^2}=C_1(\frac{17+\sqrt{73}}{50})^n+C_2(\frac{17-\sqrt{73}}{50})^n.$$
*Proof.* By direct computation and using Lemma \[lemma21\], we can verify $$\begin{aligned}
\|A_2(h|I)\|^2_{l^2}&=\sum_{k=1}^{4}\big(h(\frac{k}{4})-h(\frac{k-1}{4})\big)^2\\
&=\frac{17}{25}\Big(\big(h(\frac{1}{2})-h(0)\big)^2+\big(h(1)-h(\frac{1}{2})\big)^2\Big)-\frac{54}{625}\big(h(1)-h(0)\big)^2\\
&=\frac{17}{25}\|A_{1}(h|_I)\|^2_{l^2}-\frac{54}{625}\|A_{0}(h|_I)\|^2_{l^2}.
\end{aligned}$$ This shows (\[eqn31\]) for $n=0$. For larger $n$, we can do the same computation locally on each $n$ cell and add up to get (\[eqn31\]).
The second half of the proposition directly follows (\[eqn31\]), where $\frac{17\pm\sqrt{73}}{50}$ are zeros of the polynomial $x^2-\frac{17}{25}x+\frac{54}{625}$.$\square$
The critical order $b_2$ introduced before Theorem \[thm13\] is the solution of the equation $5^{b_2}3^{-1}=(\frac{17+\sqrt{73}}{50})^{-1}$. Noticing that $2^{2\alpha(b_2)-1}=5^{b_2}3^{-1}$, we have the following Corollary.
\[coro33\] Let $h$ be a harmonic funciton on $\SG$. Then $h|_I\in B^{2,2}_\alpha(I)$ if and only if $\alpha<\alpha(b_2)=\frac{\log\frac{25(17-\sqrt{73})}{54}}{2\log2}(\approx 0.984472)$.
Using Proposition \[prop32\] and Theorem \[th22\], we can prove the following restriction theorem.
\[thm34\] Let $b_1<\sigma<b_2$ and $\alpha=\alpha(\sigma)$. Then, the restriction map $u$ to $u|_I$ is continuous $L^2_\sigma(\SG)\to B^{2,2}_\alpha(I)$.
*Proof.* By Theorem \[th22\], each $u\in L^2_\sigma(\SG)$ admits a unique expansion $$u=h+\sum_{n=1}^\infty\sum_{x\in \tilde{V}_n} c_x\varphi_x=\sum_{n=0}^{\infty} \psi_n,$$ where we write $\psi_n=\sum_{x\in \tilde{V}_n}\varphi_x, n\geq 1$ and $\psi_0=h$ for convenience. Write $C_n=(\sum_{x\in \tilde{V}_n}c^2_x)^{1/2}$ for short.
Then, obviously $A_m(\psi_n|_I)=0$ if $m<n$ and $\|A_n(\psi_n|_I)\|_{l^2}\lesssim C_n$. In addition, by Proposition \[prop32\], for $m>n$ $$\|A_m(\psi_n|_I)\|_{l^2}\lesssim \lambda^{m-n}\|A_{n+1}(\psi_n|_I)\|_{l^2}\lesssim \lambda^{m-n}C_n,$$ where $\lambda=5^{-b_2/2}3^{1/2}=2^{-\alpha(b_2)+1/2}$. Then, we have the estimate $$\begin{aligned}
&(\sum_{m=1}^\infty 2^{2m(\alpha-1/2)}\|A_m(u|_I)\|^2_{l^2})^{1/2}
=\big\|2^{m(\alpha-1/2)}\|A_m(u|_I)\|_{l^2}\big\|_{l^2}\\
\leq& \big\|2^{m(\alpha-1/2)}\sum_{n=0}^m\|A_m(\psi_n|_I)\|_{l^2}\big\|_{l^2}\lesssim \|2^{m(\alpha-1/2)}\sum_{n=0}^m \lambda^{m-n}C_n\|_{l^2}\\
=&\|2^{m(\alpha-1/2)}\sum_{n=0}^m \lambda^{n}C_{m-n}\|_{l^2}\leq \sum_{n=0}^\infty (2^{\alpha-1/2}\lambda)^n\|2^{m(\alpha-1/2)}C_m\|_{l^2}\\
\leq& \big(\sum_{n=0}^\infty (2^{\alpha-1/2}\lambda)^n\big)\big(\|h\|^2_{L^2(SG)}+\sum_{n=1}^\infty\sum_{x\in \tilde{V}_n} 5^{n\sigma}3^{-n}|c_x|^2\big)^{\frac{1}{2}}\lesssim \|u\|_{L^2_\sigma(\SG)},
\end{aligned}$$ where we use Theorem \[th22\] in the last step. The theorem then follows. $\square$
An extension theorem
--------------------
In the rest of this section, we develop an extension map as the right inverse of the restriction map. It suffices to modify A. Jonsson’s idea. However, we provide another extension map here, as preparation for further developments in Section 4.
We introduce some new notations here.
\[def35\] Let $f\in C(I)$. For $n\geq 1$ and $1\leq k\leq 2^n$, define $\tilde{D}f(n,k)$ as $$\tilde{D}f(n,k)=\begin{cases}
f(\frac{2k-1}{2^{n+1}})-\frac{4}{5}f(\frac{k}{2^n})-\frac{8}{25}f(\frac{k-1}{2^n})+\frac{3}{25}f(\frac{k+1}{2^n}),\text{ if $k$ is odd},\\
f(\frac{2k-1}{2^{n+1}})-\frac{4}{5}f(\frac{k-1}{2^n})-\frac{8}{25}f(\frac{k}{2^n})+\frac{3}{25}f(\frac{k-2}{2^n}),\text{ if $k$ is even}.
\end{cases}$$
(a). Let $U_n=V_n\cap I$, $\tilde{U}_n=\tilde{V}_n\cap I$ and $U_*=V_*\cap I$. Clearly, $U_*$ is the set of dyadic rationals on $I$.
(b). For each pair $(n,k)$ where $n\geq 0$ and $1\leq k\leq 2^n$, let $w(n,k)$ be the unique word in $\{1,2\}^n$ such that $$[\frac{k-1}{2^n},\frac{k}{2^n}]=F_{w(n,k)}([0,1]).$$ For example, $F_{w(n,1)}=11\cdots 1$, and $F_{w(n,2^n)}=22\cdots2$.
(c). Let $x_{(n,k)}=F_{w(n,k)}q_0$, and define $NU_n=\{x_{(n,k)}:1\leq k\leq 2^n\}$. See Fiugre \[fig1\] for an illustration.
{width="5.92cm"}
(0,0) (-113,147)[$x_{(0,1)}=q_0$]{} (-150,76)[$x_{(1,1)}$]{} (-49,76)[$x_{(1,2)}$]{} (-171,40)[$x_{(2,1)}$]{} (-110,44)[$x_{(2,2)}$]{} (-88,44)[$x_{(2,3)}$]{} (-27,40)[$x_{(2,4)}$]{}
.\[fig1\]
With the above definitions and notations, we introduce the following space along with the extension map.
\[def36\] (a). Let $f\in C(I)$. For $n\geq 1$ and $1\leq k\leq 2^n$, define $c_{x_{(n,k)}}=5\tilde{D}f(n,k)$. Define the extension map $\tilde{E}$ as follows, $$\tilde{E}f=h+\sum_{n=1}^\infty\sum_{x\in NU_n}c_x\varphi_x,$$ where $h$ is the unique harmonic function on $\SG$ such that $h(0)=f(0),h(1)=f(1)$ and $h(\frac{1}{2})=f(\frac{1}{2})$.
(b). For $\sigma>b_1$, define the space of functions on $I$ $$\tilde{\mathcal{T}}_{\sigma}=\{f\in C(I): \sum_{n=1}^\infty\sum_{k=1}^{2^n} 5^{n\sigma}3^{-n}|\tilde{D}f(n,k)|^2<\infty\}$$ with norm $\|f\|_{\tilde{\mathcal{T}}_{\sigma}}=\big(\|f\|^2_{L^2(I)}+\sum_{n=1}^\infty\sum_{k=1}^{2^n} 5^{n\sigma}3^{-n}|\tilde{D}f(n,k)|^2\big)^{1/2}$.
Immediately from the definition, we have the following proposition.
\[prop37\] Let $b_1<\sigma<2-\frac{\log3}{\log5}$. We have $L^2_\sigma(\SG)|_I=\tilde{\mathcal{T}}_\sigma$, and $\tilde{E}$ is a continuous map from $\tilde{\mathcal{T}}_\sigma$ to $L^2_\sigma(\SG)$ such that $(\tilde{E}f)|_I=f$.
*Proof.* Let $u\in L^2_\sigma(\SG)$ with the unique expansion $u=h+\sum_{m=1}^\infty\sum_{x\in \tilde{V}_m} c_x\varphi_x$ as shown in Theorem \[th22\]. As in the proof of Theorem \[thm34\], denote $\psi_0=h$ and $\psi_m=\sum_{x\in \tilde{V}_m} c_x\varphi_x$. Clearly, for $m>n+1$, we have $\tilde{D}(\psi_m|_I)(n,k)=0$ for any $1\leq k\leq 2^n$, as $\psi_m|_{U_{n+1}}=0$. In addition, for $m<n$, $\tilde{D}(\psi_m|_I)(n,k)=0$ by Lemma \[lemma21\]. As a consequence, we have $$\tilde{D}(u|_I)(n,k)=\tilde{D}(\psi_n|_I)(n,k)+\tilde{D}(\psi_{n+1}|_I)(n,k).$$ Thus, $$\begin{aligned}
\sum_{k=1}^{2^n}|\tilde{D}(u|_I)(n,k)|^2&\leq 2\sum_{k=1}^{2^n}|\tilde{D}(\psi_n|_I)(n,k)|^2+2\sum_{k=1}^{2^n}|\tilde{D}(\psi_{n+1}|_I)(n,k)|^2\\
&\lesssim \sum_{x\in \tilde{V}_n}|c_x|^2+\sum_{x\in \tilde{V}_{n+1}}|c_x|^2.
\end{aligned}$$ Summing over the above estimate, we get $u|_I\in \tilde{\mathcal{T}}_\sigma$. Obviously $\|u|_I\|_{L^2(I)}\leq \|u|_I\|_{L^\infty(I)}\lesssim \|u\|_{L^2_\sigma(\SG)}$, so the restriction map is continuous.
Next, we show $\tilde{E}$ is the desired extension map. It is not hard to see that $$\tilde{D}\varphi_{x_{(n',k')}}(n,k)=\frac{1}{5}\delta_{n'n}\delta_{k'k},$$ where $\delta_{ij}$ denotes the Kronecker delta. As a consequence, $$\tilde{D}\big((\tilde{E}f)|_I\big)(n,k)=\frac{1}{5}c_{x_{(n,k)}}=\tilde{D}f(n,k).$$ In addition, $\tilde{E}f(0)=h(0)=f(0), \tilde{E}f(\frac{1}{2})=h(\frac{1}{2})=f(\frac{1}{2})$ and $\tilde{E}f(1)=h(1)=f(1)$. Combining the above observations, we conclude $(\tilde{E}f)|_I=f$. It is easy to check the continuity of $\tilde{E}$ with Theorem \[th22\]. $\square$
The following lemma shows the relationship between two spaces $\tilde{\mathcal{T}}_\sigma$ and $B^{2,2}_\alpha(I)$.
\[lemma38\] Let $\alpha=\alpha(\sigma)$. For $b_1<\sigma<\frac{\log6}{\log5}$, we have $B^{2,2}_\alpha(I)\subset \tilde{\mathcal{T}}_\sigma$; for $b_1<\sigma<b_2$, we have $\tilde{\mathcal{T}}_\sigma=B^{2,2}_\alpha(I)$.
*Proof.* It is clear that $B^{2,2}_\alpha(I)\subset \tilde{\mathcal{T}}_\sigma$ for $b_1<\sigma<\frac{\log6}{\log5}$, as $$\tilde{D}f(n,k)=\begin{cases}
\frac{1}{5}A_{n+1}(f)_{2k-1}-\frac{4}{5}A_{n+1}(f)_{2k}+\frac{3}{25}A_n(f)_k+\frac{3}{25}A_n(f)_{k+1},\text{ if $k$ is odd},\\
\frac{4}{5}A_{n+1}(f)_{2k-1}-\frac{1}{5}A_{n+1}(f)_{2k}-\frac{3}{25}A_n(f)_{k-1}-\frac{3}{25}A_n(f)_k,\text{ if $k$ is even}.
\end{cases}$$ On the other hand, by Theorem \[thm34\] and Proposition \[prop37\], we have $\tilde{\mathcal{T}}_\sigma=L^2_\sigma(SG)|_{I}\subset B^{2,2}_\alpha(I)$ for $b_1<\sigma<b_2$.$\square$
**Remark.** One can check that the linear function $f(t)=t$ on $I$ is not in $\tilde{\mathcal{T}}_\sigma$ for $\sigma\geq \frac{\log6}{\log5}$. So the bound for the range of $\sigma$ in Lemma 3.8 is sharp.
Combining Proposition \[prop37\] and Lemma \[lemma38\], we get the extension theorem as follows.
\[thm39\] Let $b_1<\sigma<b_2$ and $\alpha=\alpha(\sigma)$. The extension map $\tilde{E}$ is a continuous map from $B^{2,2}_\alpha(I)$ to $L^2_\sigma(\SG)$ such that $(\tilde{E}f)|_I=f$.
A trace theorem for higher order
================================
In Section 3, we developed an extension of A. Jonsson’s theorem. However, for Sobolev spaces of higher orders, the Besov spaces are no longer the trace spaces. In this section, we work on a discrete characterization of $L^2_\sigma(\SG)|_I$ for $b_1<\sigma\leq 2$. This includes $L^2_2(\SG)=dom_{L^2}\Delta(SG)$ as a special case. We still study the restriction theorem and the extension theorem seperately, and prove theorem \[thm14\] at the end.
A restriction theorem
---------------------
In this subsection, we will introduce the trace space (see Definition \[def45\]) and prove a restriction theorem.
We would like to study the space $\tilde{\mathcal{T}}_\sigma$ first, and try to modify it.
![An illustration for $Z_{n,k}$ and $\tilde{Z}_{n,k}$.[]{data-label="fig2"}](figure2.pdf "fig:"){width="5.2cm"} ![An illustration for $Z_{n,k}$ and $\tilde{Z}_{n,k}$.[]{data-label="fig2"}](figure1.pdf "fig:"){width="5.2cm"}
(0,0) (-280,28)[$Z_{1,1}$]{} (-210,28)[$Z_{1,2}$]{}
(-122,41)[$\tilde{Z}_{1,1}$]{} (-51,41)[$\tilde{Z}_{1,2}$]{}
Recall that we define $w(n,k)\in \{1,2\}^n$ such that $F_{w(n,k)}([0,1])=[\frac{k-1}{2^n},\frac{k}{2^n}]$ for $n\geq 0$ and $1\leq k\leq 2^n$.
(a). Write $Z_{(n,k)}=F_{w(n,k)}SG$ and $\tilde{Z}_{(n,k)}=F_{w(n,k)}F_0SG$. See Figure \[fig2\] for an illustration of $Z_{(n,k)}$ and $\tilde{Z}_{(n,k)}$.
(b). Say $(n',k')\geq (n,k)$ if and only if $F_{w(n',k')}(I)\subset F_{w(n,k)}(I)$. It is easy to see that $$Z_{(n,k)}=\bigcup_{(n',k')\geq (n,k)}\tilde{Z}_{(n',k')}.$$
(c). Define non-abelian ‘+’ on the pairs with the following equation $$w\big((n,k)+(n',k')\big)=w(n,k)w(n',k')\in \{1,2\}^{n+n'}.$$ Clearly, $(n,k)+(n',k')\geq(n,k)$.
As an example of (c), readers can check that $$(n,k)+(1,1)=(n+1,2k-1) \text{ and } (n,k)+(1,2)=(n+1,2k).$$
The idea of the following lemma and Lemma \[lemma47\] can be found in [@cq3], where pointwise approximations of Laplacians were discussed.
\[lemma41\] There exist $J,J'\in C(\SG)$ such that for each $u\in dom_{L^2}\Delta(SG)$ and $n\geq 0,1\leq k\leq 2^n$, we have $$\begin{cases}
\tilde{D}(u|_I)(n+1,2k-1)=(\frac{3}{5})^n\int_{Z_{(n,k)}}\Delta f(x)J\circ F^{-1}_{w(n,k)}(x)d\mu(x),\\
\tilde{D}(u|_I)(n+1,2k)=(\frac{3}{5})^n\int_{Z_{(n,k)}} \Delta f(x)J'\circ F^{-1}_{w(n,k)}(x)d\mu(x).
\end{cases}$$
*Proof.* First, by the Riesz representation theorem on Hilbert spaces, we can find $J\in dom_0\mathcal{E}$ such that $\tilde{D}v(1,1)=-\mathcal{E}(v,J)$ for each $v\in dom_0\mathcal{E}$.
Define $u_0=u-h$, where $h$ is the unique harmonic function that $h|_{V_0}=u|_{V_0}$. Using the weak formula of the Laplacian and the fact that $\tilde{D}(h|_I)(1,1)=0$, we get the following desired formula $$\begin{aligned}
\tilde{D}(u|_I)(1,1)=\tilde{D}(u_0|_I)(1,1)=-\mathcal{E}(u_0,J)=\int_{\SG} \Delta u_0\cdot Jd\mu=\int_{\SG} \Delta u\cdot Jd\mu.
\end{aligned}$$ A same idea works for $\tilde{D}u(1,2)$. The lemma then follows by scaling.$\square$
\[lemma42\] The restriction map $R:L^2_\sigma(K)\to \tilde{\mathcal{T}}_{\sigma}$ is a continuous linear map for $b_1<\sigma\leq 2$.
*Proof.* It suffices to prove the argument for $1\leq \sigma\leq 2$. First, for $\sigma=1$, it is an immediate consequence of Theorem \[thm34\] and Lemma \[lemma38\].
Next, we show the lemma for $\sigma=2$. By using Lemma \[lemma41\], we get $$|\tilde{D}(u|_I)(n+1,2k-1)|\lesssim (\frac{3}{5})^n\int_{Z_{(n,k)}}|\Delta u|d\mu\\
\lesssim (\frac{3}{5})^n\sum_{(n',k')\geq (n,k)}3^{-n'/2}\|\Delta u\|_{L^2(\tilde{Z}_{(n',k')})}.$$ The same estimate holds for $\tilde{D}(u|_I)(n+1,2k)$. Using the above estimate and the Minkowski inequality, we have $$\begin{aligned}
&\big(\sum_{n=1}^\infty\sum_{k=1}^{2^n} 5^{2n}\cdot 3^{-n}|\tilde{D}(u|_I)(n,k)|^2\big)^{1/2}\\
\lesssim& \big\|3^{n/2}\sum_{(n',k')\geq (n,k)}3^{-n'/2}\|\Delta u\|_{L^2(\tilde{Z}_{(n',k')})}\big\|_{l^2(n,k)}
\\=&\big\|\sum_{(n',k')\geq (0,1)}3^{-n'/2}\|\Delta u\|_{L^2(\tilde{Z}_{(n,k)+(n',k')})}\big\|_{l^2(n,k)}\\
\leq& \sum_{(n',k')\geq (0,1)}3^{-n'/2}\|\Delta u\|_{L^2(\bigcup_{(n,k)\geq (0,1)}\tilde{Z}_{(n,k)+(n',k')})}\\
\leq &\sum_{n'=0}^\infty (\frac{2}{3})^{n'/2} \|\Delta u\|_{L^2(\bigcup_{k=1}^{2^{n'}}Z_{(n',k)})}\lesssim \|\Delta u\|_{L^2(\SG)},
\end{aligned}$$ where we use the notation $\|a_{(n,k)}\|_{l^2(n,k)}=\big(\sum_{(n,k)\geq (0,1)}|a_{(n,k)}|^2\big)^{1/2}$ for convenience. This proves the argument for $\sigma=2$. For general $1\leq \sigma\leq 2$, we can use complex interpolation to deduce the lemma.$\square$
However, $\tilde{\mathcal{T}}_{2}$ is actually a larger space than the trace space of $L^2_2(SG)$. For example, we will see in Corollary \[coro33\] that $\varphi_x|_{I}\in \tilde{\mathcal{T}}_2\setminus \big(L^2_2(SG)|_I\big)$, where $\varphi_x$ is a tent function.
Recall that on $\SG$, for a function $u\in C(\SG)$, we define the normal derivative at a boundary point to be $$\partial_n u(q_i)=\lim_{n\to\infty}(\frac53)^n \big(2u(q_i)-u(F_i^nq_j)-u(F_i^nq_k)\big),\text{ with }\{i,j,k\}=\{0,1,2\}.$$ The definition can be localized to any vertex in $V_*$ by scaling, and we use $\uparrow,\rightarrow,\leftarrow$ to show the direction, i.e. $$\begin{aligned}
\partial^\uparrow_n u(F_wq_0)=(\frac{5}{3})&^{|w|}\partial_n (u\circ F_w)(q_0),\\ \partial^\leftarrow_n u(F_wq_1)=(\frac{5}{3})^{|w|}\partial_n(u\circ F_w)(q_1),&\quad \partial^\rightarrow_n u(F_wq_2)=(\frac{5}{3})^{|w|}\partial_n(u\circ F_w)(q_2).
\end{aligned}$$
\[lemma43\] Let $u\in dom_{L^2}\Delta(\SG)$. We have $$\partial_n u(q_1)=\lim_{n\to\infty}(\frac{5}{3})^n\big(4u(q_1)-5u(F_1^{n+1}q_2)+u(F^n_1q_2)\big).$$ For general cases, for $n\geq 0$ and $1\leq k\leq 2^n$, $$\partial^\rightarrow_nu(\frac{k}{2^n})=\lim_{m\to\infty}(\frac{5}{3})^m\big(4u(\frac{k}{2^n})-5u(\frac{k}{2^n}-\frac{1}{2^{m+1}})+u(\frac{k}{2^n}-\frac{1}{2^{m}})\big);$$ for $n\geq 0$ and $0\leq k\leq 2^n-1$ $$\partial^\leftarrow_nu(\frac{k}{2^n})=\lim_{m\to\infty}(\frac{5}{3})^m\big(4u(\frac{k}{2^n})-5u(\frac{k}{2^n}+\frac{1}{2^{m+1}})+u(\frac{k}{2^n}+\frac{1}{2^{m}})\big).$$
*Proof.* We only need to show the special case for $q_1=0$, since general cases can be proven by using scaling and symmetry.
First, the equation holds for harmonic functions without taking the limit, since $$u(F_1^{n+1}q_2)=\frac{1}{5}u(F^n_1q_0)+\frac{2}{5}u(q_1)+\frac{2}{5}u(F^n_1q_2).$$ For general $u\in dom_{L^2}\Delta(\SG)$, we only need to notice that $$u(F_1^{n+1}q_2)=\frac{1}{5}u(F^n_1q_0)+\frac{2}{5}u(q_1)+\frac{2}{5}u(F^n_1q_2)+\frac{1}{5^n}\int_{SG}G(F_1q_2,y)\Delta u(F_1^ny)d\mu(y),$$ where $G$ is the Green’s function on $\SG$. $\square$
Let $x=\frac{1}{2}$, then $R\varphi_x\notin RL^2_2(SG)$.
*Proof*. Assume there exists $u\in dom_{L^2}\Delta(\SG)$ such that $u|_{I}=\varphi_x|_{I}$. Then by Lemma \[lemma43\], we have $$\partial^\rightarrow_n u(\frac{1}{2})=\partial^\rightarrow_n \varphi_x(\frac{1}{2}),\quad \partial^\leftarrow_n u(\frac{1}{2})=\partial^\leftarrow_n \varphi_x(\frac{1}{2}).$$ Thus $u$ does not satisfies the matching condition at $x$, i.e. $\partial^\rightarrow_n u(\frac{1}{2})+\partial^\leftarrow_n f(\frac{1}{2})\neq 0$, which contradicts the fact that $u\in dom_{L^2}\Delta(\SG)$.$\square$
Inspired by the above observation, we need to include the information of matching condition into the desired trace space.
\[def45\] Let $f\in C(I)$, and let $n\geq 2,1\leq k\leq 2^n-1$.
(a). Define $Df(n,k),$ as follows. For $k$ odd, define $$Df(n,k)=\tilde{D}f(n-1,\frac{k+1}{2});$$ for $k$ even, define $$Df(n,k)=f(\frac{k}{2^n})-\frac{5}{8}\big(f(\frac{k-1}{2^n})+f(\frac{k+1}{2^n})\big)+\frac{1}{8}\big(f(\frac{k-2}{2^n})+f(\frac{k+2}{2^n})\big).$$
(b). Define $$\mathcal{T}_{\sigma}=\big\{f\in C(I):\sum_{n=2}^\infty\sum_{k=1}^{2^n-1} 5^{\sigma n}3^{-n}|Df(n,k)|^2<\infty\big\},$$ with norm $\|f\|_{\mathcal{T}_\sigma}=\big(\|f\|^2_{L^2(I)}+\sum_{n=2}^\infty\sum_{k=1}^{2^n-1} 5^{\sigma n}3^{-n}|Df(n,k)|^2\big)^{1/2}$.
**Remark.** We can also characterize $\mathcal{T}_\sigma$ with $$\mathcal{T}_\sigma=\{f\in\tilde{\mathcal{T}}_\sigma:\sum_{n=2}^\infty\sum_{k=1}^{2^{n-1}-1}5^{\sigma n}3^{-n}|Df(n,2k)|^2<\infty\},$$ which means we additionally require the matching condition on $\mathcal{\tilde{T}}_\sigma$. In addition, for small $\sigma$, the two spaces coincide as stated by the following lemma.
\[lemma46\] For $b_1<\sigma<b_2$, we have $\mathcal{T}_\sigma=\tilde{\mathcal{T}}_\sigma$.
*Proof.* By the above remark and using Lemma \[lemma38\], we can easily check $\mathcal{T}_\sigma\subset \tilde{\mathcal{T}}_\sigma=B^{2,2}_\alpha(I)\subset \mathcal{T}_\sigma$, where $\alpha=\alpha(\sigma)$.$\square$
Lemma \[lemma46\] can be polished, see Corollary \[coro411\]. Parellel to Lemma \[lemma41\], we have the following lemma \[lemma47\].
\[lemma47\] There exists $J_1,J_2\in C(\SG)$ such that for each $u\in dom_{L^2}\Delta(\SG)$ and $n\geq 1,1\leq k\leq 2^n-1$, the following equality holds $$\begin{aligned}
D(u|_I)(n+1,2k)=&(\frac{3}{5})^n\int_{F_{w(n,k)}SG}J_1(F^{-1}_{w(n,k)}x)\Delta u(x)d\mu(x)\\&+(\frac{3}{5})^n\int_{F_{w(n,k+1)}SG}J_2(F^{-1}_{w(n,k+1)}x)\Delta u(x)d\mu(x).
\end{aligned}$$
*Proof.* The proof is very similar to that of Lemma \[lemma41\]. For any function $h$ that is harmonic on $F_1\SG\cup F_2\SG$, it is direct to check that $D(h|_I)(2,2)=0$. Let $\tilde{u}=u-h$ on $F_1\SG\cup F_2\SG$, where $h$ is harmonic on $F_1\SG\cup F_2\SG$ with boundary values $h(q_1)=u(q_1),h(F_0q_1)=u(F_0q_1),h(F_0q_2)=u(F_0q_2)$ and $h(q_2)=u(q_2)$. We can find $J_0$ on $F_1\SG\cup F_2\SG$ such that $\mathcal{E}_{F_1\SG\cup F_2\SG}(\tilde{u},J)=-D(\tilde{u}|_I)(2,2)$. By a same argument as in the proof of Lemma \[lemma41\], we have $$D(u|_I)(2,2)=\int_{F_1\SG\cup F_2\SG} J_0\Delta ud\mu.$$ Take $J_1=\frac{5}{3}J_0\circ F_1$ and $J_2=\frac{5}{3}J_0\circ F_2$, then we get the desired equation for $D(u|_I)(2,2)$. For general cases, we only need to use scaling.$\square$
Following a same proof of Lemma \[lemma42\], we finally get the following restriction theorem.
\[thm48\] The restriction map $R:L^2_\sigma(K)\to \mathcal{T}_\sigma$ is a continuous linear map for $b_1<\sigma\leq 2$.
*Proof.* For $b_1<\sigma<b_2$, the result is an easy consequence of Lemma \[lemma46\] and Proposition \[prop37\]. For $\sigma=2$, the result follows from a similar estimate as the proof of Lemma \[lemma42\], using Lemma \[lemma47\] and Lemma \[lemma42\]. The theorem then follows by using complex interpolation.$\square$
An extension theorem
--------------------
Next, we construct an extension map by modifying $\tilde{E}$.
\[extension\] Choose $v_0\in dom_{L^2}\Delta(\SG)$ such that $v_0|_{V_0}=0$ and $\partial_n v_0(q_j)=\delta_{0,j}$. Recall that in Definition \[def36\] $$\tilde{E}f=h+\sum_{n=1}^\infty\sum_{x\in NU_n} c_x\varphi_x,$$ where $c_{x_{(n,k)}}=5\tilde{D}f(n,k)=5Df(n+1,2k-1)$ and $h\in \mathcal{H}_0$ with $h|_{U_1}=f|_{U_1}$. Define the extension map $E$ as follows, $$Ef=\tilde{E}f-\sum_{n=1}^\infty\sum_{k=1}^{2^n}\frac{12}{5}c_{x_{(n,k)}}v_0\circ F^{-1}_0\circ F^{-1}_{w(n,k)}.$$
\[thm410\] For $b_1<\sigma\leq2$, the extension map $E:\mathcal{T}_\sigma\to L^2_\sigma(SG)$ is a continuous linear map such that $(Ef)|_{I}=f$.
*Proof.* First, we show $E:\mathcal{T}_\sigma\to L^2_\sigma(\SG)$ is bounded for $1\leq\sigma\leq2$. Let $$\tilde{E}_mf=h+\sum_{n=1}^m\sum_{x\in NU_n} c_x\varphi_x.$$ Choose $v_2\in dom\Delta$ such that $v_2|_{V_0}=0$ and $\partial_n v_2(q_j)=\delta_{2,j}$, and define $$\begin{aligned}
E_mf(x)=&\tilde{E}_mf(x)-\sum_{n=1}^m\sum_{k=1}^{2^n}\frac{12}{5}c_{x_{n,k}}v_0\circ F^{-1}_0\circ F^{-1}_{w(n,k)}\\
&-\sum_{k=1}^{2^m-1}\big(\frac{24}{5}Df(m+1,2k)\big)v_2\circ F^{-1}_{w(m+1,2k)}.
\end{aligned}$$ $E_mf$ satisfies the matching conditions at all vertices, which implies that $E_mf\in L^2_2(\SG)\subset L^2_\sigma(\SG)$.
Notice that $v_0\circ F^{-1}_0\circ F^{-1}_{w(n,k)}$ supports on $\tilde{Z}_{(n,k)}$, and $v_2\circ F^{-1}_{w(m+1,2k)}$ supports on $F_{w(m+1,2k)}SG$, which are disjoint sets. We can easily get the following estimates $$\begin{cases}
\|E_mf\|_{L^2_1}\lesssim \big(f^2(0)+f^2(1)+f^2(\frac{1}{2})+\sum_{n=1}^m\sum_{k=1}^{2^n}(\frac{5}{3})^n|\tilde{D}f(n,k)|^2\\\qquad\qquad\qquad+\sum_{k=1}^{2^m-1}(\frac{5}{3})^m|Df(m+1,2k)|^2\big)^{1/2},\\
\|E_mf\|_{L^2_2}\lesssim \big(f^2(0)+f^2(1)+f^2(\frac{1}{2})+\sum_{n=1}^m\sum_{k=1}^{2^n}(\frac{25}{3})^n|\tilde{D}f(n,k)|^2\\\qquad\qquad\qquad+\sum_{k=1}^{2^m-1}(\frac{25}{3})^m|Df(m+1,2k)|^2\big)^{1/2}.
\end{cases}$$ Clearly, the above estimates holds uniformly for any $m\geq 1$. Using complex interpolation, we then get $$\begin{aligned}
\|E_mf\|_{L^2_\sigma}\lesssim \big(f^2(0)+f^2(1)+f^2(\frac{1}{2})+\sum_{n=1}^m\sum_{k=1}^{2^n}5^{\sigma n}3^{-n}|\tilde{D}f(n,k)|^2\\\qquad\qquad\qquad+\sum_{k=1}^{2^m-1}5^{\sigma m}3^{-m}|Df(m+1,2k)|^2\big)^{1/2}
\end{aligned}$$ holds uniformly for any $m$. In other words, $\|E_mf\|_{L^2_\sigma}\lesssim \|f\|_{\mathcal{T}_\sigma}$ uniformly for any $m\geq 1$.
Similarly, the following estimate holds uniformly for any $1\leq\sigma\leq2$ and $m\geq 1$, $$\begin{aligned}
\|E_{m'}f-E_mf\|^2_{L^2_\sigma(SG)}&\lesssim \sum_{n=m+1}^\infty\sum_{k=1}^{2^n}5^{2n}3^{-n}|\tilde{D}f(n,k)|^2+\sum_{k=1}^{2^{m}-1}5^{2m}3^{-m}\big(Df(m+1,2k)\big)^2\\
&+\sum_{k=1}^{2^{m'}-1}5^{2m'}3^{-m'}\big(Df(m'+1,2k)\big)^2.
\end{aligned}$$ As a result, $E_mf$ converges in $L^2_\sigma(\SG)$. Noticing that $E_mf$ converges pointwise to $Ef$, we conclude that $E_mf$ converges to $Ef$ in $L^2_\sigma(\SG)$ sense. Thus $Ef\in L^2_\sigma(\SG)$ and $\|Ef\|_{L^2_\sigma(\SG)}\lesssim \|f\|_{\mathcal{T}_\sigma}$.
Next, for $b_1<\sigma<1$, we have the scaling property that $\|v_0\circ F_w^{-1}\|_{L^2_\sigma(\SG)}\asymp (5^\sigma3^{-1/2})^{|w|}\|v_0\|_{L^2_\sigma(\SG)}$, as a consequence of Theorem \[th22\]. In addition, using Theorem \[th22\], we can check that $$\|\sum_{n=1}^\infty\sum_{k=1}^{2^n}\frac{12}{5}c_{x_{(n,k)}}v_0\circ F^{-1}_0\circ F^{-1}_{w(n,k)}\|^2_{L^2_\sigma(\SG)}\asymp \sum_{n=1}^\infty\sum_{k=1}^{2^n}\|\frac{12}{5}c_{x_{(n,k)}}v_0\circ F^{-1}_0\circ F^{-1}_{w(n,k)}\|^2_{L^2_\sigma(\SG)},$$ as $v_0\circ F^{-1}_0\circ F^{-1}_{w(n,k)}$ have disjoint supports. Combining the above two facts, it is direct to see that $E:\mathcal{T}_\sigma\to L^2_\sigma(\SG)$ is continuous for small $\sigma$.
Lastly, $(Ef)|_{I}=(\tilde{E}f)|_I=f$, since $v_0\circ F_0^{-1}\circ F_{w(n,k)}^{-1}$ is supported away from $I$. $\square$
Combining Theorem \[thm48\] and \[thm410\], we finally get Theorem \[thm14\]. Also, the following corollary shows the relationship of the different traces spaces discussed in this paper.
\[coro411\] Let $\alpha=\alpha(\sigma)$. Then we have the following relationships.
\(a) For $b_1<\sigma<b_2$, we have $B^{2,2}_\alpha(I)=\mathcal{T}_\sigma$; for $b_1<\sigma<\frac{\log6}{\log5}$, we have $B^{2,2}_\alpha(I)=\mathcal{T}_\sigma$; for $b_1<\sigma<2-\frac{\log3}{\log5}$, we have $\tilde{\mathcal{T}}_\sigma=\mathcal{T}_\sigma$.
\(b) For any $b_1<\sigma_1<\sigma_2\leq 2$, we have $[\mathcal{T}_{\sigma_1},\mathcal{T}_{\sigma_2}]_\theta=\mathcal{T}_{(1-\theta)\sigma_1+\theta\sigma_2}$.
*Proof.* (a) is immediately from Lemma \[lemma38\], Proposition \[prop37\] and Theorem \[thm14\].
(b). Using complex interpolation, the restriction map maps from $L^2_{(1-\theta)\sigma_1+\theta\sigma_2}(\SG)$ to $[\mathcal{T}_{\sigma_1},\mathcal{T}_{\sigma_2}]_\theta$, and the extension map maps from $[\mathcal{T}_{\sigma_1},\mathcal{T}_{\sigma_2}]_\theta$ to $L^2_{(1-\theta)\sigma_1+\theta\sigma_2}(\SG)$. Thus $\mathcal{T}_{(1-\theta)\sigma_1+\theta\sigma_2}=L^2_{(1-\theta)\sigma_1+\theta\sigma_2}(\SG)|_I=[\mathcal{T}_{\sigma_1},\mathcal{T}_{\sigma_2}]_\theta$. $\square$
Related observations and further questions
==========================================
In this last section, we provide some related results and question that worth further study.
Another space that we are interested in is $$dom_{L^\infty}\Delta(\SG)=\{u\in dom\mathcal{E}:\Delta u\in L^\infty(\SG)\}.$$ With a same method as in the last section, we can derive the following result.
\[thm51\] Define $$\mathcal{T}^\infty_2=\{f\in C(\SG):\sup_{n\geq 2}\sup_{1\leq k\leq 2^n-1}5^n|Df(n,k)|<\infty\}.$$ Then, $dom_{L^\infty}\Delta(\SG)|_{I}=\mathcal{T}^\infty_2$.
Consider the symmetric derivative of the functions. Let $f\in C(I)$ and fix $x\in U_m\setminus U_0$, we define $\delta_nf(x)=f(x-\frac{1}{2^n})-2f(x)+f(x+\frac{1}{2^n}).$ The symmetric derivative at $x$ is defined to be the renormalized limit of $\delta_n f(x)$, $$f''_s(x)=\lim_{n\to\infty} 4^n\delta_nf(x).$$
\[prop51\] Let $u\in dom_{L^\infty}\Delta(\SG)$ and $f=u|_I$. Then for all $x\in U_*\setminus \{0,1\}$, we have $g''_s(x)=0$.
*Proof.* Let $x=\frac{k}{2^n}$ for some $n\geq 1$ and $1\leq k\leq 2^n-1$. By direct computation and using Theorem \[thm51\], we have $$|\delta_{m+1}f(x)-\frac{1}{5}\delta_mf(x)|=|\frac{8}{5}Df(m+1,2^{m+1-n}k)|\leq c(\frac{1}{5})^{m-n},$$ where $c$ is a constant depends on $\|\Delta u\|_\infty$. Summing over the above estimate, we get $$|\delta_{m}f-(\frac{1}{5})^{m-n+1}\delta_{n+1}f|\leq c(m-n-1)(\frac15)^{m-n}.$$ As an immediate consequence, we get $f''_s(x)=\lim_{m\to\infty} 4^m\delta_{m}f(x)=0$.$\square$
On the other hand, $\delta_nf(x)$ should not converge uniformly to $0$ in general cases. Since otherwise, it would imply that $f\in dom\Delta (I)$ with $\Delta f=0$, which means $f$ is a linear function. In fact, the following result shows that $f'(x)=\lim\limits_{n\to\infty}2^n(f(x)-f(x+\frac{1}{2^n}))$ diverges when $\partial^\rightarrow_nu(x)\neq 0$.
Let $u\in dom_{L^\infty}\Delta(\SG)$ and let $f=u|_I$. Let $x\in U_*$ and suppose $\partial_n u(x)\neq 0$. Then $|f'(x)|=\infty$.
*Proof.* Without loss of generality, assume $x\neq 1$, and we denote $\delta^+_mf(x)=f(x+\frac{1}{2^m})-f(x)$. Then by Lemma \[lemma43\], we have $$\begin{aligned}
\lim_{m\to\infty} (\frac{5}{3})^m \big(\delta^+_mf(x)-5\delta^+_{m+1}f(x)\big)=\partial_n^\rightarrow u(x).
\end{aligned}$$ Clearly, $$\begin{aligned}
\delta_m^+f(x)&=\sum_{k=n}^{m-1} \big((\frac{1}{5})^{m-k-1}\delta^+_{k+1}f(x)-(\frac{1}{5})^{m-k}\delta^+_kf(x)\big)+(\frac{1}{5})^{m-n}\delta^+_n f(x)\\
&\asymp -\sum_{k=n}^{m-1} (\frac{3}{5})^{k}(\frac{1}{5})^{m-k}\big(\partial_n^\rightarrow u(x)+o(1)\big)+(\frac{1}{5})^{m-n}\delta^+_n f(x).
\end{aligned}$$ As a result, we have $\lim\limits_{m\to\infty}(\frac{5}{3})^m\delta^+_m f(x)=-\frac{1}{2}\partial^\rightarrow_n u(x)$. Similarly, for $x\neq 0$, let $\delta^-_mf(x)=f(x)-f(x-\frac{1}{2^m})$, we have $\lim\limits_{m\to\infty}(\frac{5}{3})^m\delta^-_m f(x)=\frac{1}{2}\partial^\leftarrow_n u(x)$. The proposition is immediate from the above observation. $\square$
We are also interested in Sobolev spaces of higher orders. We believe that a similar idea would work, but more complicated differences will occur in the discrete characterization. For example, for $1<\sigma\leq 2$, we will need to study extension algorithm of biharmonic functions, and find suitable difference operators. The computations are getting messy, so we do not go further in this direction. Hopefully, readers may get new ideas dealing with this.
Readers may have noticed that $b_2$ plays the important role in that it is the highest index that $B^{2,2}_{\alpha}(I)\supset L^2_\sigma(\SG)|_{I}$ for each $\alpha<\alpha(b_2)$ as long as $\sigma\geq\alpha(b_2)$. Noticing that $b_2$ is uniquely characterized by harmonic functions, we wonder whether biharmonic functions play similarly important roles in higher order cases. It is also of interest to find the largest index $\alpha$ such that $h|_{I}$ lies in $B^{2,2}_{\alpha}(I)$ for multi-harmonic functions, and what kind of role these indexes will play. We hope to find a systematic way to deal with these.
[10]{} M.T. Barlow and E.A. Perkins, [*Brownian motion on the Sierpinski gasket,*]{} Probab. Theory Related Fields, 79 (1988) 543-623.
S. Cao and H. Qiu, *Atomic decompositions and Besov type characterizations of Sobolev spaces on p.c.f. fractals,* arXiv:1904.00342.
S. Cao and H. Qiu, *Boundary Value Problems for harmonic functions on domains in Sierpinski gaskets,* arXiv:1702.02419.
S. Cao and H. Qiu, *Higher order tangents and Higher order Laplacians on Sierpinski Gasket Type Fractals,* arXiv:1607.07544.
S. Cao and H. Qiu, [*Sobolev spaces and trace theorem on the Sierpinski gasket,*]{} arXiv:1903.07012.
Z. Guo, R. Kogan, H. Qiu and R.S. Strichartz, *Boundary value problems for a family of domains in the Sierpinski gasket,* Illinois J. Math. 58 (2014), 497-519.
M. Hinz and T. Kumagai, *A trace theorem for Dirichlet forms on fractals,* J. Func. Anal. 238 (2006), 578-611.
M. Hinz, D. Koch and M. Meinert, *Sobolev spaces and calculus of variations on fractals,* arXiv:1805.04456.
Q. Gu and K. Lau, [*Dirichlet forms and critical exponents on fractals,*]{} arXiv:1703.07061.
J. Hu and X. Wang, *Domains of Dirichlet forms and effective resistance estimates on p.c.f. fractals*, Studia Math. 177:2 (2006), 153-172.
J. Hu and M. Zähle, *Potential spaces on fractals*, Studia Math. 170:3 (2005), 259-281.
H. Qiu, *Exact spectrum of the Laplacian on a domain in the Sierpinski gasket,* arXiv:1206.1381.
M. Ionescu, L.G. Rogers and R.S. Strichartz, *Pseudo-differential operators on fractals and other metric measure spaces,* Rev. Mat. Iberoam. 29 (2013) 1159–1190.
A. Jonsson, *A trace theorem for the Dirichlet form on the Sierpinski gasket,* Math. Z. 250 (2005), 599-609.
A. Jonsson, *Brownian motion on fractals and function spaces,* Math. Z. 222 (1996), 495-504.
A. Kamont, *A discrete characterization of Besov Spaces,* Approx. Th. Appl. 13 (1997) 63-77.
J. Kigami, *A harmonic calculus on the Sierpinski spaces,* Jpn. J. Appl. Math. 6 (1989), 259-290.
J. Kigami, *A harmonic calculus on p.c.f. self-similar sets,* Trans. Amer. Math. Soc. 335 (1993), 721-755.
J. Kigami, *Analysis on Fractals.* Cambridge University Press, 2001.
T. Kumagai, *Brownian Motion Penetrating Fractals: An Application of the Trace Theorem of Besov Spaces,* J. Func. Anal. 170 (2000) 69-92.
W. Li and R.S. Strichartz, *Boundary value problems on a half Sierpinski gasket,* J. Fractal Geom. 1 (2014) 1-43.
T. Lindstr[ø]{}m, *Brownian motion on nested fractals,* Mem. Amer. Math. Soc. 83 (1990), 420.
J. Owen and R.S. Strichartz, *Boundary value problems for harmonic functions on a domain in the Sierpinski gasket,* Indiana Univ. Math. J. 61(2012), 319-335.
R.S. Strichartz and C. Wong, *The p-Laplacian on the Sierpinski gasket,* Nonlinearity 17 (2004), 595-616.
R.S. Strichartz, *Differential Equations on Fractals: A Tutorial.* Princeton University Press, 2006.
R.S. Strichartz, *Function spaces on fractals,* J. Funct. Anal. 198 (2003), 43-83.
[^1]:
[^2]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We reexamine the dynamical generation of mass for fermions charged under various Lie groups with equal charge and mass at a high Grand Unification scale, extending the Renormalization Group Equations in the perturbative regime to two-loops and matching to the Dyson-Schwinger Equations in the strong coupling regime.'
address: 'Depto. Física Teórica, Universidad Complutense de Madrid, Plaza de las Ciencias 1, 28040 Madrid, Spain.'
author:
- 'Felipe J. Llanes-Estrada and Alexandre Salas-Bernárdez'
title: Chiral symmetry breaking for fermions charged under large Lie groups
---
Introduction
============
The Standard Model is a gauge field theory based on the gauged symmetry $$\label{eq:SM}
SU(3)_C\otimes SU(2)_L\otimes U(1)_Y\;.$$ Here $SU(3)_C$ denotes the color interaction responsible for the strong force, $SU(2)_L$ the isospin coupling of left-handed fermions and $U(1)_Y$ the hypercharge group. The spontaneous breaking of the electroweak symmetry by the Higgs mechanism suggested the possibility of higher symmetries at yet higher scales that would also be spontaneously broken, providing strong and electroweak force unification at higher scales; these symmetries would also have to be spontaneously broken [^1].
In the SM, the Higgs vacuum expectation value breaks the global symmetry $SU(2)_L\times U(1)_Y$ of the Higgs sector in the SM down to $U(1)_{em}$ [@SymSM] (or, considering the $U(1)$ as a perturbation, and the approximate global custodial $SU(2)$, it breaks $SU(2)\times SU(2)\to SU(2)_c$). This generates masses for the $W^\pm$ and $Z$ bosons, and for fermions, leaving us the symmetry $$SU(3)_C\otimes U(1)_{em}\;$$ (and the approximate custodial $SU(2)$). A feature of the symmetry group of the Standard Model that stands out is the small size of the numbers 1-2-3. Why are we confronted by such symmetry groups? Why not larger groups like $SU(6)$ or $Sp(10)$?
To address these question we study in Section \[FromGUT\] how hypothetical quarks colored under different groups acquire masses from a Grand Unified Theory (GUT) scale where all groups under consideration are chosen to have the same couplings and quark masses, down to lower energies where the interaction becomes strong. For this task we will use the Renormalization Group Equations (RGE).
Then, section \[sec:DSE\] treats the Dyson-Schwinger Equations (DSE) for the lowest scales when the interactions become strong. Any workable truncation of the DSE typically fails to satisfy local gauge invariance, while respecting global symmetry. This is however enough to discuss its breaking in view of Elitzur’s theorem. While realistic models [@GG] that embed the SM such as $SU(5)$ or $SO(10)$ are often discussed [^2], we are here less ambitious and keep the discussion at a general level, considering multiple groups.
In addition to a brief discussion in section \[sec:outlook\], the article has an appendix addressing the computation of color factors for almost all of the continuous Lie groups (results for E8 are not at hand). We have kept the article as short as is compatible with its being self-contained, since the theory behind our approach has already been laid out in a previous publication [@Llanes]. We have striven to extend that calculation as explained next.
From Grand Unification to strong interaction scale with the Renormalization Group Equations {#FromGUT}
===========================================================================================
Our motivation in this work is to extend the one-loop RGE computation of [@Llanes] to two loops. This was the aspect that introduced the most uncertainty to predict the mass of the fermions charged under large groups. In doing so we have unveiled partial errors in the original publication that we here correct. An erratum has also been issued to warn the reader of the earlier article.
We evolve the masses of one single color-charged fermion for the different color groups from the Grand Unification scale of $\mu_{GUT}=10^{15}\;GeV$ to the point where interactions become strong (at a scale $\sigma$) for each group, that is, when $C_F\,\alpha_s(\sigma)=0.4$. Once this happens we use Dyson-Schwinger equations for the non-perturbative regime in order to obtain the constituent masses for these fermions: this step is explained in the next section \[sec:DSE\].
An efficient way of keeping track of the parameter evolution needed for the physical predictions of a theory to be invariant under $\mu$ scale choice is the use of RGEs. We generalize those of Quantum Chromodynamics to an arbitrary gauge group $G$. The running of the coupling constant $g_s$ with $\mu$ is [@TMuta] determined by the $\beta(g_s)$ function, $$\beta(g_s)\equiv -\mu\frac{dg_s}{d\mu}=\beta_1g_s^3+\beta_2 g_s^5+...\;.$$ The one-loop correction $\beta_1$ is $$\beta_1=\frac{1}{(4\pi)^2}\left(\frac{11C_G-2T_RN_f}{3}\right)\;,$$ where $C_G$ is the adjoint Casimir [^3], $T_R$ the normalization of the generators $T^a$ of the group $G$ defined as $Tr(T^aT^b)\equiv T_R\delta^{ab}$ and $N_f$ the number of colored fermions [^4].
The two loop contribution to the $\beta(g_s)$ function, $\beta_2$, entails a larger effort in perturbation theory, but can also be easily found in the literature [@TMuta], $$\beta_2=\frac{1}{(4\pi)^4}\left(\frac{34}{3}C_G^2-4\left(\frac{5}{3}C_G+C_F\right)T_RN_f\right)\;,$$ where $C_F$ is the Casimir of the fundamental representation(see appendix). Using the color coefficients listed there, we obtain the running couplings of $SU(N)$, $SO(N)$, $Sp(N)$ and the exceptional groups $G2$, $F4$, $E6$ and $E7$, shown in Figure \[fig:runningcoupling\].
![Running couplings for the families $SU(N)$, $SO(N)$, $Sp(N)$ and most of the Exceptional Lie groups. \[fig:runningcoupling\]](FIGS.DIR/SU_N_CA.png "fig:"){width="1\columnwidth"} ![Running couplings for the families $SU(N)$, $SO(N)$, $Sp(N)$ and most of the Exceptional Lie groups. \[fig:runningcoupling\]](FIGS.DIR/SO_N_CA.png "fig:"){width="1\columnwidth"} ![Running couplings for the families $SU(N)$, $SO(N)$, $Sp(N)$ and most of the Exceptional Lie groups. \[fig:runningcoupling\]](FIGS.DIR/Sp_N_CA.png "fig:"){width="1\columnwidth"} ![Running couplings for the families $SU(N)$, $SO(N)$, $Sp(N)$ and most of the Exceptional Lie groups. \[fig:runningcoupling\]](FIGS.DIR/ConstanteAgruposespeciales.png "fig:"){width="1\columnwidth"}
The result of [@Llanes], that for small groups and one flavor $\sigma \propto e^{N}$ stands out. The very large groups have strongly interacting scales $\sigma$ clustering around the GUT scale, since they run very fast. We are then ready to start employing the DSEs down from the scale $\sigma$.
Simultaneously, running of the current mass $m_c$ is set by the self energy correction to the quark propagator that implies an anomalous mass dimension $\gamma_m$ $$\gamma_m(g_s)\equiv -\frac{\mu}{m}\frac{dm}{d\mu}=\gamma_1g_s^2+\gamma_2 g_s^4+...\;.$$ The one loop contribution to the $\gamma_2(g_s)$ function for the quarks, $\gamma_1$, amounts to $$\gamma_1=\frac{6C_F}{(4\pi)^2}\;.$$
The two loop contribution to $\gamma_m(g_s)$ (see [@Tarrach]), $\gamma_2$, is $$\gamma_2=\frac{C_F}{(4\pi)^4}\Big(3C_F+\frac{97}{3}C_G-\frac{20}{3}T_R N_f\Big)\label{eq:gamma2}\;.$$
At the GUT starting scale of the RGEs we choose a fermion mass $m_{c}(\mu_{GUT})=1\;MeV$ and fix the coupling $\alpha_{s}(\mu_{GUT})\equiv g_s({\mu_{GUT}})^2/4\pi=0.0165$ to broadly reproduce the isospin average mass for the $SU(3)_C$ quarks of the first generation at the scale $\mu=2\;GeV$, $$\overline{m}(2\,GeV)=\frac{m_u(2\;GeV)+m_d(2\;GeV)}{2}\simeq3.5\;\;MeV\;.$$ These initial conditions are taken to be the same for all Lie groups, as suggested by the concept of GUT. Then, the mass running for the various Lie groups, with color factors taken from \[sec:CF\] is plotted in Figure \[fig:runningmass\].\
![Running masses for the families $SU(N)$, $SO(N)$, $Sp(N)$ and four out of five Exceptional Lie groups.\[fig:runningmass\]](FIGS.DIR/SU_N_Masas.png "fig:"){width="1\columnwidth"} ![Running masses for the families $SU(N)$, $SO(N)$, $Sp(N)$ and four out of five Exceptional Lie groups.\[fig:runningmass\]](FIGS.DIR/SO_N_Masas.png "fig:"){width="1\columnwidth"} ![Running masses for the families $SU(N)$, $SO(N)$, $Sp(N)$ and four out of five Exceptional Lie groups.\[fig:runningmass\]](FIGS.DIR/Sp_N_Masas.png "fig:"){width="1\columnwidth"} ![Running masses for the families $SU(N)$, $SO(N)$, $Sp(N)$ and four out of five Exceptional Lie groups.\[fig:runningmass\]](FIGS.DIR/GruposEspecialesMasas.png "fig:"){width="1\columnwidth"}
$ $
Running at the strong interaction scale with the Dyson-Schwinger Equations {#sec:DSE}
==========================================================================
Once the interactions become strong, perturbation theory breaks down and resummation becomes necessary: we thus adopt the simplest possible DSE for the quark self energy. The free propagator of a fermion with current mass $m_c$ [@TMuta], $ S_0(p^2)=\frac{1}{m_c-\slashed p}\;,$ becomes a fully dressed one $ \tilde{S}(p^2)=\frac{1}{B(p^2)-A(p^2)\slashed p}\;$. Being only interested in qualitative features of spontaneous mass generation, we can approximate $A(p^2)=1$ which leaves the physical mass as $M(p^2)\equiv B(p^2)$. Denoting $\Sigma(p)$ as the sum of all one-particle irreducible diagrams, the DSE takes the form (omitting the $p$ dependence) $$\tilde{S}(p^2)=S_0(p^2)\,(1-\Sigma(p) S_0(p^2))^{-1}\;.$$ Inverting, we see that $ \tilde{S}^{-1}(p^2)=S_0(p^2)^{-1}-\Sigma(p) \Rightarrow M(p^2)=m_c-\Sigma(p)\;. $
To illustrate the possibilities, we will employ the *rainbow truncation* that sums only “rainbow shaped” diagrams, with great simplification (Fig. \[fig:rainbow\]).
![Rainbow DSE for the full quark propagator (filled circle).\[fig:rainbow\]](FIGS.DIR/DysonSchwinger1loop.png "fig:"){width="0.9\columnwidth"}\
![Rainbow DSE for the full quark propagator (filled circle).\[fig:rainbow\]](FIGS.DIR/Rainbowselfenergy.png "fig:"){width="0.7\columnwidth"}
\
The one-loop self energy is then, passing to Euclidean space with $k^0\to ik^0$, $p^0\to ip^0$, given by $$\begin{aligned}
\Sigma_{\rm rainbow}(p)=&g^2_s\int\frac{d^4k}{i(2\pi)^4}\gamma^\mu(T^a)\frac{1}{M(k^2)-\slashed k}\gamma^\nu(T^a) \frac{\eta_{\mu\nu}}{(k-p)^2}\nonumber\\
= C_F&g^2_s\int_0^\infty\frac{dk_E \,k_E^3}{\pi^3}\frac{-M(k^2)}{M^2(k^2)+k_E^2}\nonumber\times\\
\times&\int_{-1}^{+1}\sqrt{1-x^2}\frac{dx}{(k_E^2-2|k_E||p_E|\,x+p^2_E)}\;.\end{aligned}$$ We define the last integral in $x$ as the averaged gluon propagator $D^0_{k-p}$ (in the Feynman Gauge) over the four dimensional polar angle. Hence, we conclude that the Dyson-Schwinger equation in the rainbow approximation for the quark propagator is $$\label{eq:DSE}
M(p^2)=m_c+C_Fg^2_s\int_0^\infty\frac{dq \,q^3}{\pi^3}\frac{M(q^2)}{M^2(q^2)+q^2}D^0_{q-p}\;.$$
Note that the integral in (\[eq:DSE\]) is divergent and must be regularized. We could employ a simple cutoff regularization cutting this integral at a scale $\Lambda$; instead we would like to preserve Lorentz invariance and exhibit renormalizability. Following again [@Llanes], we introduce renormalization constants $Z(\Lambda^2,\mu^2)$ to absorb infinities and any dependence on the cutoff $\Lambda$, $$\tilde{S}^{-1}(p^2,\mu^2)\equiv Z_2 S_0^{-1}(p^2)-\Sigma(p^2,\mu^2)\;,$$ where the dependence of $\Sigma$ on $\mu$ is given by the fermion and gluon propagators. Apart from the wavefunction renormalization $Z_2$ we introduce $Z_m$ for the bare quark mass. The relation between the (cutoff dependent) unrenormalized mass $m_c(\Lambda^2)$ and the renormalized mass at the renormalization scale $\mu$, $m_R(\mu^2)$, is $$m_c(\Lambda^2)=Z_m(\Lambda^2,\mu^2)m_R(\mu^2).$$ Since we will maintain the restriction $A(p^2)=1$, renormalization of the quark wavefunction is not necessary, therefore $Z_2=1$. The only renormalization condition is to fix the mass function at $p^2=\mu^2$. The DSE is then $$M(p^2)=Z_m m_R(\mu^2)-\Sigma(p^2,\mu^2)\;.\label{EQ:RMM}$$ Evaluating (\[EQ:RMM\]) at $p^2=\mu^2$ and subtracting it again to (\[EQ:RMM\]) we obtain, $$\begin{aligned}
\label{eq:DSEMOM}
M(p^2)&=M(\mu^2)\nonumber\\
&+C_Fg^2_s\int_0^\infty\frac{dq \,q^3}{\pi^3}\frac{M(q^2)}{M^2(q^2)+q^2}\Big(D^0_{q-p}-D^0_{q-\mu}\Big)\;.\end{aligned}$$ It is easy to show, taking $\mu$ parallel to $p$, that asymptotically [@Llanes], $$\frac{\partial M(p^2)}{\partial \Lambda}\propto\frac{M(\Lambda^2)(p-\mu)}{\Lambda^2}\;.$$ Therefore, for large $\Lambda$, $M(p^2)$ stops depending on the cutoff, which can be taken [*e.g.*]{} to $\Lambda=10^{10}\; GeV$ and renormalization is achieved.
Now we are ready to obtain the quark constituent masses for all the groups studied. We match the RGE solution (high scales) to the DSE solution (low scales) at the matching energy $\sigma$ where interactions become strong, $C_F\alpha_s(\sigma)=0.4$ for each group, as advertised. For SU(3) ($C_F=\frac{4}{3}$), the scale where $\alpha_s(\sigma)=0.3$ is $\sigma=2.09\; GeV$. From this point down in scale we freeze $\alpha_s$. A constant vertex factor of order 7 is applied to the DSE to guarantee sufficient chiral symmetry breaking at low scales, requiring the constituent quark mass $M(0)$ to be close to $300\;MeV$ using the substracted DSE (\[eq:DSEMOM\]). This is supposed to mock up the effect of vertex-corrections not included, and is known to scale with $N$ [@Alkofer:2008tt] for large $N$, the group’s fundamental dimension. Finally, the $M(p)$ obtained is plotted in Figure \[fg:DSERGE\].
![Matching of RGE and DSE solutions of the Mass Running for $SU(3)$.[]{data-label="fg:DSERGE"}](FIGS.DIR/RGEDSEsolutionscopia.png){width="1\columnwidth"}
To obtain the constituent fermion masses for the different Lie Groups we use a trick presented in [@Llanes]: to perform a scale transformation $$p^2\to \lambda^2p^2 \ ; \ \sigma^2\to \lambda^2\sigma^2$$ on the DSE (\[eq:DSEMOM\]). Changing the integration variable $q^2\to \lambda^2q^2$, giving $d^4q\to \lambda^4d^4q$, the modified DSE equation is satisfied by a modified $\tilde{M}$ and the relation between the constituent masses is simply $ M(0)=\frac{\tilde{M}(0)}{\lambda}\;.$ Now, taking $\lambda$ as the ratio of the scales where interactions become strong for $SU(3)$ and another group, $$\frac{\sigma_{group}}{\sigma_{SU(3)}}=\lambda\;,$$ the mass function scales in the same way, $$\frac{M_{group}(0)}{M_{SU(3)}(0)}=\lambda\;.$$ Hence, eliminating the auxiliary $\lambda$, we find $$\frac{M_{group}(0)}{M_{SU(3)}(0)}=\frac{\sigma_{group}}{\sigma_{SU(3)}}\,.$$ Using these results we compute the constituent masses for the quarks charged under the different groups (Fig. \[constituent\]).
![Constituent Masses for the groups which break chiral symmetry in RGE before $10^{-5}\;GeV$.[]{data-label="constituent"}](FIGS.DIR/MasasConstituyentes1.png){width="1.1\columnwidth"}
The outcome is that the special Lie groups examined do not spontaneously generate fermion mass at a high scale: their interactions, running at two loops from the GUT, are too weak to do so. This is because the $C_G$ Casimir of the adjoint representation, though proportional to the group dimension, carries a small numerical factor that reduces the intensity of coupling running.
Large $SU(N)$ and $Sp(N)$ groups, on the other hand, behave as advanced in [@Llanes], and generate a mass for the fermions that puts them beyond reach of past accelerators. The exceptions are $Sp(4)$, for which the mass generation is similar to QCD; and $Sp(2)$ , which is too weak. As for the special orthogonal groups, for $SO(N>10)$, once more the fermion mass generated is too large to be accessible at accelerators.
Conclusions and Outlook {#sec:outlook}
=======================
We have examined mass generation for different Lie groups with an arbitrary number of colours. As a definite starting point, we have adopted the philosophy of Grand Unification in which fermion masses as well as coupling constants, for all groups, coincide at a high scale, namely $10^{15}$ GeV.
We have run the couplings and masses for each group to lower scales employing two-loop Renormalization Group Equations, using as an input the Cuadratic Casimirs obtained in \[sec:CF\]. We chose the initial conditions at $\mu=\mu_{\rm GUT}$ to be the same for all groups and selected so that $SU(3)_C$ at the scale of 2 GeV yields a rough approximation of the strong force coupling and first-generation isospin-averaged quark mass.
Typically, for all but the smallest groups, a scale arises where interactions become strong (discernable as a Landau pole in perturbation theory). We stop running at the scale $\mu$ such that $\alpha_s(\mu)C_F=0.4$; below that, we employ a non-perturbative treatment, namely Dyson-Schwinger Equations in the rainbow approximation to assess the masses down to yet lower scales.
Combining the methods of RGE and DSE and requiring that the constituent masses of $SU(3)_C$ colored quarks to be 300 MeV has allowed us to obtain the constituent masses of hypothetical fermions charged under different groups from a Grand Unification Scale of $10^{15}\;\text{GeV}$. From this treatment we can conclude that groups belonging to the $SU(N)$ and $Sp(N)$ families, with $N>4$, generate masses of order or above the few TeV. Notwithstanding the crude approximations we have employed, our computation gives about 5 TeV to $SU(4)$-charged fermions, which would not be far out of reach of mid-future experiments provided the GUT conditions apply. It appears from our simple work that larger groups (except the Exceptional Groups and $SO(N)$ with $N<10$) might endow fermions with a mass too high to make them detectable in the foreseeable future. In case these superheavy fermions would have been coupled to the Standard Model, they would have long decayed in the early universe due to the enormous phase space available. If they existed and be decoupled from the SM, they would appear to be some form of dark matter. We have also provided a partial answer to the question “*Why the symmetry group of the Standard Model, $S U (3)_C \otimes SU(2)_L \otimes U(1)_Y$ , contains only small-dimensional subgroups?*” It happens that, upon equal conditions at a large Grand Unification scale, large-dimensioned groups in the classical $SO(N)$, $SU(N)$ and $Sp(N)$ families force dynamical mass generation at higher scales because their coupling runs faster. Should fermions charged under these groups exist, they would appear in the spectrum at much higher energies than hitherto explored [@Odense].\
Interestingly for collider phenomenology, we find the masses of fermions charged under the following groups are within reach of the energy frontier: $M_{\rm SU(4)}\simeq 5$ TeV; $M_{Sp(6)}\simeq4.4$ TeV; $M_{\rm SO(10)}\simeq 7$ TeV. The LHC might be able to exclude those [^5].
However, the following groups $SO(N<10)$, $E_6$, $E_7$, $G2$ and $F4$ yield masses that are below the TeV scale and should already have been seen if they coupled to the rest of the Standard Model (one could argue that those isomorphic to groups present in the SM have already been sighted). Their absence from phenomenology thus suggests that fermions charged under any of those groups , if at all existing, belong to a decoupled dark sector.
Color Factors {#sec:CF}
=============
We here present some of the calculations carried out to obtain the quadratic Casimirs needed in both RGE and DSE. Such quadratic Casimirs are elements in the Lie Algebra which commute with all the other elements (See [@Cvit; @GTN; @FFS] for the necessary group theory).\
We will focus on the Casimir invariant in the fundamental representation of the group $G$, $C_F\delta^{ij}=(T^a T^a)^{ij}$, and the Casimir invariant in the adjoint representation, $C_G\delta^{ab}=f^{acd}f^{bcd}$. Normalization of the algebra generators is chosen as $Tr(T^aT^b)=\frac{1}{2}\delta^{ab}$.
Special Unitary Groups $SU(N)$
------------------------------
We start with the special unitary family $SU(N)$. Its generators $T^{a}$ are traceless hermitian. Therefore every Hermitean $N\times N$ matrix $A$ can be written as, $$A=A^\dagger=c_0\mathbb{I}+c_a T^a\,.$$ From this we find $$c_0=\frac{Tr(A)}{N}\;\;\;\;\;\;c_a=2\, Tr(AT^a)\,.$$ Having then $$\begin{aligned}
&A_{ij}=A_{lk}\delta^{li}\delta^{kj}=A_{lk}\Big(2(T^a)_{ij}(T^a)^{kl}+\frac{1}{N}\delta^{kl}\delta_{ij}\Big)\\
&\Rightarrow A_{lk}\Big(2(T^a)_{ij}(T^a)^{kl}+\frac{1}{N}\delta^{kl}\delta_{ij}-\delta^{li}\delta^{kj}\Big)=0\;.\end{aligned}$$ Since $A$ is arbitrary, we find for the generators the useful relation $$(T^a)_{ij}(T^a)_{kl}=\frac{1}{2}\Big(\delta_{li}\delta_{kj}-\frac{1}{N}\delta_{kl}\delta_{ij}\Big)\;.$$ Contracting $j$ and $k$ we obtain the fundamental representation Casimir or Color Factor $$(T^aT^a)_{ij}=\frac{1}{2}\Big(\frac{N^2-1}{N}\Big)\delta_{ij}=C_F\delta_{ij}\;.$$ Now we compute the following combination, $$\begin{aligned}
(T^a)_{i}^j(T^b)_{jk}(T^a)^k_{l}=\frac{1}{2}\Big((T^b)_{jk}\delta_{li}\delta^{kj}-\frac{1}{N}(T^b)_{jk}\delta^k_{l}\delta_{i}^j\Big)\nonumber\\
=-\frac{1}{2N}(T^b)_{il}\;.\end{aligned}$$ Noting the following identity and using the results already computed, we obtain the adjoint Casimir for $SU(N)$, $$\begin{aligned}
\label{eq:CG}
f^{acd}f^{bcd}&=-2\,Tr\Big([T^a,T^c][T^b,T^c]\Big)\nonumber\\
&=-2 Tr\Big(2T^aT^cT^bT^c-(T^aT^b+T^bT^a)T^cT^c\Big)\nonumber\\
&=N\,\delta^{ab}=C_G\delta^{ab}\,.\end{aligned}$$
Special Orthogonal Groups $SO(N)$
---------------------------------
We will follow now the same steps for the Special Orthogonal family $SO(N)$. Its generators are antisymmetric and traceless and they form a basis for the antisymmetric $N\times N$ matrices. Thus, taking an antisymmetric matrix $A$, we have $$A=-A^T=c_a T^a\;\;\;\; \Rightarrow\;\;\;\; c_a=2 Tr\Big(A T^a\Big)\;.$$ Then we have $$A_{ij}=A_{kl} \delta^k_i\delta^l_j=\frac{1}{2}A_{kl}\Big( \delta^k_i\delta^l_j-\delta^k_j\delta^l_i\Big)=A_{lk}\Big(2 (T^a)_{ij}(T^a)^{kl}\Big)\;.$$ Finding $$A_{kl}\Big[ \frac{1}{2}\Big( \delta^k_i\delta^l_j-\delta^k_j\delta^l_i\Big)+\Big(2 (T^a)_{ij}(T^a)^{kl}\Big)\Big]=0\;,$$ and since A is an arbitrary antisymmetric matrix we get $$(T^a)_{ij}(T^a)^{kl}=\frac{1}{4}\Big(\delta^k_j\delta^l_i- \delta^k_i\delta^l_j\Big)\;.$$ Here, since the group is real there is no need for distinction between upper and lower indices. Contracting in the previous expression $j$ with $k$ we obtain the Color Factor $$(T^a)_{ij}(T^a)^{jl}=\frac{1}{4}\Big(\delta^j_j\delta^l_i- \delta^j_i\delta^l_j\Big)=\frac{N-1}{4}\delta_i^l=C_F\delta_i^l\;.$$ As before, we compute $$\begin{aligned}
(T^a)_{ij}(T^b)^{jk}(T^a)_{kl}&=\frac{1}{4}\Big((T^b)^{jk}\delta_{il}\delta_{kj}-(T^b)^{jk}\delta_{ik}\delta_{lj}\Big)\nonumber\\
&=-\frac{1}{4}(T^b)_{li}=\frac{1}{4}(T^b)_{il}\;.\end{aligned}$$ We are able now to obtain the adjoint Casimir for $SO(N)$. Similar to (\[eq:CG\]) $$\begin{aligned}
f^{acd}f^{bcd}&=2\,Tr\Big(\frac{1}{2}T^aT^b-(T^aT^b+T^bT^a)\frac{N-1}{4}\Big)\nonumber\\
&=\frac{1}{2}(N-2)\delta^{ab}=C_G\delta^{ab}\,.\end{aligned}$$
Simplectic Groups $Sp(N)$
-------------------------
The elements $M\in Sp(N)$ (with $N$ even) are $N\times N$ matrices which preserve the antisymmetric tensor $$\Omega=
\begin{pmatrix}
0 & \mathbb{I}_{\frac{N}{2}\times\frac{N}{2}}\\
-\mathbb{I}_{\small{\frac{N}{2}\times\frac{N}{2}}} & 0
\end{pmatrix},$$ in the sense $$\Omega=M^T\Omega \,M\Rightarrow\;\;M^{-1}=\Omega^TM^T\Omega\,.$$ Using this relation it is possible to prove that the generators of the group take the form $$\label{eq:Spgen}
-T^a=\Omega^T(T^a)^T\Omega\;\;\Rightarrow\;\;T^a=
\begin{pmatrix}
A & B\\
C & -A^T
\end{pmatrix},$$ where $B$ and $C$ are symmetric matrices. It is now possible to show that the generators satisfy $$(T^a)_{ij}(T^a)_{kl}=\frac{1}{4}\Big(\delta_{il}\delta_{jk}+\Omega_{ik}(\Omega^{-1})_{jl}\Big)\;.$$ Therefore $$(T^a)_{ij}(T^a)^j_{l}=\frac{1}{4}(N+1)\delta_{ij}=C_F\delta_{ij}\;.$$ Noticing $\Omega^T=\Omega^{-1}=-\Omega$, we compute the usual combination $$\begin{aligned}
(T^a)_{ij}(T^b)^{jk}(T^a)_{kl}&=\frac{1}{4}\Big(\delta_{il}(T^b)^{jk}\delta_{jk}+\Omega_{ik}(T^b)^{jk}(\Omega^{T})_{jl}\Big)\nonumber\\
&=\frac{1}{4}(\Omega^{T})_{ik}((T^b)^T)^{kj}\Omega_{jl}=-\frac{1}{4}(T^b)_{il}\;,\end{aligned}$$ where in the last equality we have used (\[eq:Spgen\]). The adjoint Casimir now falls down easily $$\begin{aligned}
f^{acd}f^{bcd}&=-2\,Tr\Big(-\frac{1}{2}T^aT^b-(T^aT^b+T^bT^a)\frac{N+1}{4}\Big)\nonumber\\
&=\frac{1}{2}(N+2)\delta^{ab}=C_G\delta^{ab}\end{aligned}$$ To obtain the Color Factors and adjoint Casimirs for $G2$, $F4$, $E6$ and $E7$ we refer to the article of P. Cvitanović [@Cvit]. The results obtained are presented in Table I.\
Group Color Factor $(C_F)$ Adjoint Casimir $(C_G)$ $N_c$
--------- ---------------------------------------- ------------------------------ ---------------------------------
$SU(N)$ $\frac{1}{2}\Big(N-\frac{1}{N}\Big)$ $N$ $\forall N\in\mathbb{N}$
$SO(N)$ $\frac{1}{4}\Big(N-1\Big)$ $\frac{1}{2}\Big(N-2\Big)$ $\forall N\in\mathbb{N}$
$Sp(N)$ $\frac{1}{4}\Big(N+1\Big)$ $\frac{1}{2}\Big(N+2\Big)$ $N=2n$ $\forall n\in\mathbb{N}$
$G2$ $\frac{1}{4}\Big(N-3\Big)$ $\frac{1}{2}\Big(N-3\Big)$ $N=7$
$F4$ $\frac{1}{18}\Big(N-8\Big)$ $\frac{1}{18}\Big(N+1\Big)$ $N=26$
$E6$ $\frac{1}{12}\Big(N-\frac{29}{3}\Big)$ $\frac{1}{12}\Big(N-3\Big)$ $N=27$
$E7$ $\frac{1}{48}\Big(N+1\Big)$ $\frac{1}{48}\Big(N+16\Big)$ $N=56$
[21]{} S. Elitzur, Phys. Rev. D. [**12**]{} (1975) 3978-3982.
S. Willenbrock, Physics in D >= 4 proceedings, Theoretical Advanced Study Institute in elementary particle physics 2004, Boulder, USA, June 6-July 2, 2004, pages 3-38. Published by World Scientific, Hackensack, USA. hep-ph/0410370.
H. Georgi and S. L. Glashow, Phys. Rev. Lett. [**32**]{} (1974) 438. doi:10.1103/PhysRevLett.32.438 B. Fornal and B. Grinstein, arXiv:1808.00953 \[hep-ph\].
G. García Fernández, J. Guerrero Rojas and F. J. Llanes-Estrada, Nucl. Phys. B [**915**]{} (2017) 262 doi:10.1016/j.nuclphysb.2016.12.010 \[arXiv:1507.08143 \[hep-ph\]\].
T. Muta, *Foundations of Quantum Chromodynamics*, World Scientific Lecture notes in Physics-vol. 57, (1984).
R. Alkofer [*et al.*]{}, Annals Phys. [**324**]{} (2009) 106 doi:10.1016/j.aop.2008.07.001; A. Kizilersu [*et al.*]{}, Eur. Phys. J. C [**50**]{} (2007) 871 doi:10.1140/epjc/s10052-007-0250-6; A. C. Aguilar [*et al.*]{}, Phys. Rev. D [**90**]{} (2014) no.6, 065027 doi:10.1103/PhysRevD.90.065027.;
P. Cvitanovic, Phys. Rev. D [**14**]{} (1976) 1536. doi:10.1103/PhysRevD.14.1536 A. Zee, *Group Theory in a Nutshell for Physicists*, Princeton University Press (2016). F. Wilzeck, A. Zee, *Families from Spinors*, Physical Review D 25 2 (1982).
R. Tarrach, Nucl. Phys. B [**183**]{} (1981) 384. doi:10.1016/0550-3213(81)90140-1
This work was presented at the Odense $CP^3$ Origin of Mass at the High intensity Frontier conference in May 2018.\
http://cp3-origins.dk/content/uploads/2018/05/parallel program22052018.pdf
[^1]: It is usually and superficially stated that the gauge symmetry $SU(2)_L\times U(1)_Y$ is spontaneously broken. However, Elitzur’s theorem [@Elitz] states that gauge symmetries cannot be spontaneously broken. First they must be broken explicitly by a gauge fixing term leaving only the global symmetry and then this remaining symmetry can be spontaneously broken. The modern viewpoint is that gauge symmetries are just a redundancy in the description of the theory on which expectation values of observables must not depend. The actual symmetry from which consecuences such as degeneracies in the spectrum, couplings or conserved currents appear is the true global symmetry. We will continue using “spontaneous symmetry breaking” without specifying, though in the understanding that it is the global group which is affected.
[^2]: While the absence of proton decay rules out some classic implementations of the GUT idea, models keep being constructed that evade the constraints [@Fornal:2018aqc]
[^3]: For $SU(N)$, $C_G=N$, the group dimension. But in general, $C_G=aN+b$ with $a,b$ depending on the particular group, as listed in the appendix. This detail was in error in [@Llanes] and is being corrected.
[^4]: In this article we take $N_f=1$, but a brief discussion in [@Llanes] reminds us that there is a critical number of colors $N_f^c$ that shuts off the vacuum antiscreening and thwarts spontaneous symmetry breaking.
[^5]: For comparison, the one-loop results are $M_{\rm SU(4)}\simeq 2$ TeV; $M_{Sp(6)}\simeq1.5$ TeV; $M_{\rm SO(10)}\simeq 3$ TeV., which indicates fair convergence.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The main decoding algorithms for Reed-Solomon codes are based on a bivariate interpolation step, which is expensive in time complexity. Lot of interpolation methods were proposed in order to decrease the complexity of this procedure, but they stay still expensive. Then Koetter, Ma and Vardy proposed in 2010 a technique, called re-encoding, which allows to reduce the practical running time. However, this trick is only devoted for the Koetter interpolation algorithm. We propose a reformulation of the re-encoding for any interpolation methods. The assumption for this reformulation permits only to apply it to the Welch-Berlekamp algorithm.'
author:
- |
Morgan Barbier\
ENSICAEN – GREYC\
`[email protected]`
bibliography:
- 'biblio.bib'
title: 'Re-encoding reformulation and application to Welch-Berlekamp algorithm'
---
#### Keywords:
Reed-Solomon codes, Welch-Berlekamp algorithm, Re-encoding.
Introduction
============
Interpolation based decoding algorithms {#sec:decodingAlgo}
=======================================
Bivariate interpolation for the decoding
----------------------------------------
Original re-encoding
--------------------
Revisited re-encoding {#sec:reencoding}
=====================
Re-encoding and interpolation algorithm
---------------------------------------
Revisited re-encoding {#revisited-re-encoding}
---------------------
Application to the Welch-Berlekamp algorithm {#sec:application}
============================================
Conclusion and perspective
==========================
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Social networks are discrete systems with a large amount of heterogeneity among nodes (individuals). Measures of centrality aim at a quantification of nodes’ importance for structure and function. Here we ask to which extent the most central nodes can be found by purely local search. We find that many networks have close-to-optimal searchability under eigenvector centrality, outperforming searches for degree and betweenness. Searchability of the strongest spreaders in epidemic dynamics tends to be substantially larger for supercritical than for subcritical spreading.'
author:
- Konstantin Klemm
bibliography:
- 'epiwalks.bib'
date: 'Received: date / Accepted: date'
title: 'Searchability of central nodes in networks [^1] '
---
Introduction
============
The science of networks [@Newman:2010] is an interdisciplinary enterprise with application areas ranging from systems biology [@Alon:2006] and systems chemistry [@Nitschke:2009] to ecology [@Bascompte:2007] and the social sciences [@Wasserman:1994]. Methods and models have been traditionally developed in the fields of graph theory [@Diestel:2010] and algorithms [@Thulasiraman:1991; @Cormen:2001]. Dealing with large disordered and heterogeneous systems, statistical physics now has significant influence on network science [@Albert:2002; @Dorogovtsev:2010]. The analysis of networks further benefits from the growing connections between statistical physics and computational complexity [@Moore:2011].
On the empirical side, advances in technology for observation and measurement of large systems and for compilation and storage of the resulting data sets have triggered a revolution. In the social sciences, interaction networks typically comprising less than 100 individuals used to be obtained by surveys [@Wasserman:1994] or by direct long-term observation of social groups, e.g. [@Zachary:1977]. Nowadays digital communication by mobile phones [@Guimera:2003; @Onnela:2007] and e-mail [@Guimera:2003; @Onnela:2007] and through web-based social media such as [*Facebook*]{}, [*Twitter*]{} and [*Flickr*]{} [@Grabowicz:2012] automatically generates data sets covering millions of individuals.
An important property of social networks is searchability, i.e. the possibility of finding a target node by iteratively following connections (edges) and using information on properties of nodes in the direct neighborhood for deciding which connection to follow [@Adamic:2001; @Kleinberg:2000]. This locally available information, such as the geographical location of an individual and other personal attributes, may be represented by placing nodes in a Euclidean space [@Boguna:2007a]. A searcher then iteratively chooses the edge that takes it closest to the target. The success and efficiency of such network navigation crucially depends on the distribution of long- and short-range connections in the network [@Kleinberg:2000]. Social experiments indicate that these conditions are fulfilled because personal messages tend to reach the addressee in a few steps when iteratively forwarded by one person to a chosen acquaintance [@Milgram:1967; @Travers:1969; @Dodds:2003]. An information theoretic approach characterizes the reachability between nodes by the minimal description length of the search path [@Rosvall:2005; @Sneppen:2005].
A related task of large practical relevance is the search for the most important nodes in a given context. These are, for instance, those that would contribute most to the spreading of an epidemic contagion and thus should be vaccinated first [@Pastor-Satorras:2001; @Cohen:2001]. In spreading and other dynamical processes, the importance of a node strongly depends on the network context [@Kitsak:2010; @Klemm:2012]. It may be captured by [*centrality*]{} measures such as the number of neighbours, the number of shortest communication paths running through, or the number of walks emanating from the node [@Wasserman:1994; @Koschutzki:2005]. Given a centrality measure, a [*landscape*]{} is obtained by labeling each node with its centrality value [@Wuchty:2003; @Axelsen:2006].
Here we investigate which network structures and centrality measures give rise to searchable landscapes where the most central nodes are eventually reached by iteratively jumping to the most central neighbour, i.e. by local search. We quantitatively answer this question by introducing the [*smoothness*]{} as the expectation value of the centrality eventually encountered, normalized by the maximum centrality in the whole network.
In practical scenarios, local search can be applied efficiently only when the centrality of a node can be obtained locally at that node. Eigenvector centrality, for instance, is of global nature. The explicit and exact computation of eigenvector centrality necessarily involves the whole network and then yields the centrality values of all nodes. Once all these values are known, the maximum is readily identified. In this case there is no need for local search on the centrality landscape.
Local search is useful when the network as a whole is not known but centrality of an encountered node can be obtained indirectly, e.g. by local measurement of system dynamics. The degree of a node is readily obtained by measurement of the local density from diffusion. The stationary fraction of random walkers at a node is proportional to the node’s number of neighbours. Eigenvector centrality is measured approximately as the frequency with which a node is involved in critical spreading or percolation clusters [@Klemm:2012]. We term this node’s property the spreading centrality. The smoothness of the corresponding landscape quantifies the success of identifying the globally strongest spreaders by iteratively building a path towards stronger spreaders.
Definitions and Methods
=======================
Networks and landscapes
-----------------------
A network (also called a graph) is a tuple $G=(V,E)$ with $V$ a finite set and $E$ a set of unordered tuples in $V$. The elements of $V$ are called nodes, elements of $E$ are called edges. For a node $v \in V$ of a network $(V,E)$, the neighbourhood $N(v)$ is the set $\{w \in V: \{v,w\} \in E \}$. The closed neighbourhood of $v$ is $\bar{N}(v)= N(x) \cup \{v\}$.
If $G=(V,E)$ is a network and $f:V \rightarrow \mathbb{R}$ is an arbitrary mapping, then $L=(V,E,f)$ is called a landscape (over $G$). The max-neighbourhood of a node $v$ is the set $N_{\max}(v) = \{w \in N(v) : f(w) = \max_{x \in N(v)} f(x) \}$. A node $v \in V$ is a \[strict\] local maximum of the landscape if, for all $w\in N(v)$, $f(v) \ge f(w)$ \[$f(v) > f(w)$\].
Search dynamics and smoothness
------------------------------
An adaptive walk is a local search dynamics. At each time step $t$, a transition from the current node $v$ to a neighbour is proposed by drawing $w \in N(v)$ uniformly. If $f(w) \ge f(v)$, $w$ is accepted as the node for the next time step, otherwise the search stays at $v$. Formally, an adaptive walk on $G$ is a homogeneous (time-independent) Markov chain with state set $V$ and transition probability $$\pi(v \rightarrow w) = \left\{ \begin{array}{ll}
\frac{1}{d(v)} & \text{if } w \in N(v) \text{ and } f(w)\ge f(v)\\
0 & \text{otherwise}
\end{array}\right.$$ for $w \neq v$, and $\pi(v\rightarrow v)$ given by normalization of probabilities. An adaptive walk is equivalent to kinetics at zero temperature where $-f$ plays the role of energy and the network encodes the allowed transitions between configurations (nodes). If $v$ is a strict local maximum of $L$, then $\pi(v\rightarrow v) =1$, so $v$ is an absorbing state for the adaptive walk.
A gradient walk is similar to an adaptive walk, but concentrating on $f$-maximal neighbours. The dynamics proceeds to a neighbour with the maximal value for $f$ in the neighbourhood, provided that this maximal value is strictly larger than the one at the current node. A gradient walk is thus a Markov chain with transition probability $$\pi(v \rightarrow w) = \left\{ \begin{array}{ll}
\frac{1}{N_{\max}(v)} & \text{if } w \in N_{\max}(v) \text{ and } f(w)>f(v)\\
0 & \text{otherwise}
\end{array}\right.$$ for $w \neq v$, and $\pi(v\rightarrow v)$ given by normalization. A node $v$ is an absorbing state of the gradient walk ($\pi(v\rightarrow v) =1$), if and only if $v$ is a local maximum of the landscape.
Now we characterize the landscape by the success of the search dynamics. By $\langle f \rangle(t)$ we denote the expectation value of $f$ under a given search dynamics with the uniform distribution on the node set $V$ as initial condition. For finite time $t$, we measure the [*$t$-smoothness*]{} of $L$, $$\label{eq:smoothfinite}
s_t(L) = \langle f \rangle (t) / \max_{v \in V} f(v)$$ to quantify how close the search dynamics approaches nodes with maximum centrality. The long-term success of the search is quantified by $$\label{eq:smoothdef}
s(L) = \lim_{t\rightarrow\infty} s_t(L)~,$$ the [*smoothness*]{} of $L$. The limit in Eq. (\[eq:smoothdef\]) exists because none of the search steps decreases the $f$-value (so this holds also for the expectation value) and $f$ is of finite support, thus upper-bounded. For non-negative functions $f$ — such as the node centralities considered in the following — the smoothness takes values in the unit interval. In particular, $s(L)=1$ indicates that the search is certain to reach a node with maximum $f$-value, regardless of initial condition.
The opposite case, $s(L)<1$, occurs exactly when the Markov chain has an ergodic set including a node on which $f$ is not maximal. For the gradient walk, this means that there is a local but not global maximum of $f$ because here the ergodic sets are exactly the singletons formed by local maxima. In the adaptive walk, each strict local maximum $v^\ast$ gives rise to an ergodic set $\{v^\ast\}$; further ergodic sets are formed by interconnected nodes of the same centrality value, all without neighbours of larger centrality.
Centrality measures
-------------------
The [*degree*]{} centrality of a node $v$ is defined as the number of neighbours $$d(v) = | N(v)| ~.$$ The degree is a purely local measure of node importance, taking into account only the neighbouring nodes of the node considered. The degree distribution $P:\mathbb{N}\cup\{0\} \rightarrow \mathbb{R}$ is defined by $$P(k)= \frac{|\{v \in V : d(v)=k \}|} {|V|}~.$$
The $k$-core of a network $G$ is the largest subnetwork of $G$ in which all nodes have degree at least $k$. It is computed by iteratively removing nodes with less than $k$ neighbours until no such nodes remain in the network. The removal of a node deletes all its edges so the loss of node $v$ may suppress the degree of the neighbours of $v$ below $k$ as well. Therefore the node removal is done iteratively.
The [*shell index*]{} $h(v)$ of node $v$ is the largest number $k$ such that $v$ is in the $k$-core of the network. A large shell index indicates that the node is part of a densely connected subnetwork.
The [*eigenvector centrality*]{} is obtained by finding a non-negative centrality value $e(v)$ for each $v \in V$ to solve the set of equations $$\label{eq:def_eigenv}
\lambda_\text{max} e(v) = \sum_{w \in N(v)} e(w)$$ with $\lambda_\text{max} \in \mathbb{R}$ the maximum value for which such a solution exists. The underlying idea is that the importance of a node is obtained self-consistently as the sum of its neighbours’ importances discounted by a node-independent factor $\lambda_\text{max}^{-1}$. Equation (\[eq:def\_eigenv\]) can be written as the eigenvector equation $\lambda_\text{max} e = Ae$ for matrix $A$, hence the name. The adjacency matrix $A$ has entry $a_{vw}=1$ when $\{v,w\}$ is an edge, and $a_{vw}=0$ otherwise. The Perron-Frobenius theorem guarantees that the solution is unique up to a common scaling factor if the network is connected, i.e. for any two nodes $v$ and $w$, there is a path between $v$ and $w$.
The [*betweenness centrality*]{} describes node importance as the property of being contained in shortest paths between other nodes. We denote by $\sigma_{uw}$ the number of shortest paths between nodes $u$ and $w$, by $\sigma_{uw}(v)$ the number of such paths passing through node $v$. Then $$b(v) = \sum_{u,w \in V \setminus \{v\}} \frac{\sigma_{uw}(v)}{\sigma_{uw}}$$ is the betwenness centrality of node $v$. Computation of betweenness centrality for all nodes in an unweighted network $(V,E)$ requires ${\cal O}(|V| |E|)$ computational time steps and ${\cal O}(|V| + |E|)$ memory.
Spreading centrality {#subsec:spreadc}
--------------------
Bond percolation [@Grimmett:1999] is a theory for the description of spreading of material and information in disordered media. On arbitrary networks, the statistics of clusters generated by bond percolation coincides with those of the basic Susceptible-Infected-Removed (SIR) model of epidemic spreading [@Newman:2002; @Keeling:2005].
A realization of bond percolation on a network $G=(V,E)$ at parameter $\beta$ is a network $R=(V,E^\prime)$ obtained as follows. For each edge $e \in E$ independently, include $e \in E^\prime$ with probability $\beta$, omit $e$ with probability $1-\beta$. Such a realization defines a partition of the node set $V$ into clusters $C_1,C_2,\dots$. Each node is in exactly one cluster. Two nodes $v$ and $w$ are in the same cluster if and only if $E^\prime$ contains edges by which one can walk from $v$ to $w$. By $z_\beta(v)$ we denote the expected size of the cluster containing $v$ in bond percolation at parameter $\beta$. The spreading centrality $\phi_\beta(v)$ is obtained by normalization according to $$\phi_\beta(v) = \frac{z_\beta(v) - \bar{z_\beta}}
{\max_{w\in V} z_\beta(w) - \bar{z_\beta}}$$ with the mean value $\bar{z_\beta} = \sum_{w \in V} z_\beta(w) / |V|$. In simulations, $z_\beta(v)$ is obtained from $10^4$ independent realizations of bond percolation for each network and value of $\beta$.
For each node $v$, $z_\beta(v)$ monotonically increases with $\beta$. However, the increase may differ from node to node. Therefore the ranking of nodes with respect to spreading centrality substantially changes with varying $\beta$ [@Klemm:2012]. Then the landscape of spreading centrality $L=(V,E,\phi_\beta)$ is qualitatively different for different values of $\beta$.
Random networks {#subsec:randgraphs}
---------------
Classic random models of networks are by Gilbert [@Gilbert:1959] and Erdős and Renyi [@Erdos:1959]. Gilbert’s random graph model has two parameters, being the number of nodes $n \in \mathbb{N}$ and the edge probability $p \in[0,1]$. Among all graphs on $n$ vertices, each graph with $\eta$ edges is obtained with probability $p^\eta (1-p)^{n(n-1)/2 -\eta}$. In other words, a realization of the model decides independently for each of the $n(n-1)/2$ possible edges if the edge is present (with probability $p$) or absent (with probability $1-p$). Throughout this contribution we refer to Gilbert’s model as [*random graph*]{}.
The model by Erdős and Renyi is defined as the uniform distribution on all graphs with $n$ nodes and exactly $N$ edges, $n \in \mathbb{N}$ and $N \in
\{0,n(n-1)/2\}$ being the model parameters. Asymptotically, for $n \rightarrow
\infty$, the models by Gilbert and by Erdős and Renyi become arbitrarily similar when taking $\eta = p n(n-1) / 2$.
Another, more versatile form of statistical ensembles are those with a given degree sequence. The degree of each of the $n$ nodes is a parameter value of the ensemble. The ensemble consists of all graphs with nodes having the prescribed degrees, having uniform probability. Here we use Markov chain Monte Carlo sampling with edge switching [@Rao:1996] in order to randomize networks under conservation of degrees. In a step of edge switching, two node-disjoint randomly chosen edges $\{v,w\}$ and $\{x,y\}$ are drawn (proposal) and replaced by edges $\{v,x\}$ and $\{w,y\}$ (acceptance) unless these edges already exist (rejection). In order to obtain a randomization of a network with $n$ nodes, we run the Markov chain until switching has been performed $n^2$ times (counting only accepted steps). The configuration model [@Molloy:1995] is an alternative method for the same purpose.
Stochastic network growth {#subsec:stochgrowth}
-------------------------
Let us introduce four procedures for building up networks by iterative addition of nodes and edges. The following scenario is common to all these procedures. The network is initialized with $m$ nodes fully interconnected, i.e. all $m(m-1)/2$ edges are present, where $m$ is a parameter. Then in each step $t$ of growth, a subset $S$ of size $m$ is chosen stochastically from the current node set; a new node $v(t)$ and $m$ new edges are added to the network, where each edge connects one node in $S$ with $v(t)$. The growth step is iterated until the network reaches the desired size. Each specific procedure is defined by the stochastic choice of the set of nodes that the new edges are attached to.
The [*preferential attachment*]{} rule [@Barabasi:1999; @Krapivsky:2000; @Dorogovtsev:2000] (also termed [*cumulative advantage*]{} [@Price:1976]) adds edges preferably to nodes having large degree already. Starting with an initially empty set $S$, a node $v$ is drawn from the distribution $$\pi(v) = \frac{d(v)+a}{\sum_{w \in V \setminus S} (d(w)+a)}$$ on $V\setminus S$. Then this node is included in $S$. This drawing of nodes without replacement is repeated until $|S|=m$. The parameter $a$ is a constant bias with $m < a < \infty$. [*Uniform attachment*]{}, called model A in refs.[@Barabasi:1999; @Barabasi:1999b], is obtained in the limit $a \rightarrow \infty$. Here $S(t)$ is drawn uniformly from the set of all $m$-node subsets.
The [*edge attachment*]{} rule by Dorogovtsev et al. [@Dorogovtsev:2001] is defined specifically for $m=2$ edges to be added per node. Rather than selecting nodes separately, the set $S$ is determined by drawing an [*edge*]{} $e \in E$ uniformly and setting $S=e$. Thus each new node is attached to the end nodes of an existing edge, thereby forming a triangle.
In the [*deactivation model*]{} by Klemm and Eguíluz [@Klemm:2002a; @Klemm:2002b; @Eguiluz:2002], the set of $S$ of nodes receiving edges—thus called active—evolves over time. Upon initialization, $S$ comprises all $m$ nodes present initially. In each growth each step, after adding a node $v$ and its $m$ edges, $v$ is included in $S$. From the $|S|=m+1$ then contained in $S$, one node $u$ is chosen for removal (deactivation) by the distribution $$\pi (u) = \frac{d(u)^{-1}}{\sum_{w \in S} d(w)^-1}~.$$
The expected degree distribution from the deactivation model and from edge attachment decay as a power law $P(k) \sim k^{-\gamma}$ with $\gamma \approx 3$. Preferential attachment yields $P(k) \sim k^{-\gamma}$ with $\gamma = 3 + a/m$. The asymptotic degree distribution from uniform attachment decays geometrically. The deactivation model generates networks with average distance between nodes increasing linearly with size; the other three rules generate networks with distances increasing logarithmically or sublogarithmically.
Note that all these procedures generate networks with shell index $h(v) = m$ for all nodes $v$, as shown by induction over the number of nodes. The initial condition is a fully interconnected network of $m$ nodes. After adding a node and $m$ links, each node $v$ now has exactly $d(v)=m$ neighbours and therefore $h(v)=m$. After an arbitrary number of node additions, each node has at least $m$ neighbours so that the $m$-core is the whole network. In the computation of the $m+1$-core, nodes with degree $m$ or smaller are iteratively removed, which amounts to disintegrating the whole network in node order similar to its generation. Thus the $m+1$-core is empty. Each node has the same shell index $m$, which is the global maximum. It follows trivially that all these networks have smoothness 1 with respect to shell index.
Numerical results
=================
![\[fig:land\_0\] Smoothness of centrality landscapes on stochastically grown networks. Symbols distinguish centrality measures (see legend). Large open symbols are for search by adaptive walk, small filled symbols for gradient walk. Panels (a-f) distinguish procedures for network generation. (a) preferential attachment with bias $a=0$, degree exponent $\gamma =3$ (b) preferential attachment with bias $a=-m/2$, $\gamma =2.5$. (c) edge attachment, (d) uniform attachment, (e) deactivation model. (f) random graph with edge probability $p=4/(n-1)$. Error bars indicate standard deviation over 10 independent realizations of network generation. In the growing networks (a)-(e), $m=2$ edges per node are attached in the growth process; in (f) edge probability is $p=4/(n-1)$, resulting in asymptotic average degree of $4$ in all cases. ](land_1.pdf){width="\textwidth"}
We first study smoothness of centrality landscapes for stochastically generated network instances, as described in sections \[subsec:randgraphs\] and \[subsec:stochgrowth\]. The dependence of smoothness on the number of nodes is displayed in Figure \[fig:land\_0\]. Smoothness depends both on the type of network and the centrality measure under consideration. Except for the random graphs and the networks grown by the deactivation model, eigenvector landscapes have a maximum smoothness $s=1$ with deviations only for small networks. Betweenness centrality landscapes have smoothness close to 1 in networks grown by edge attachment. For all other combinations of network generation and centrality measure, a decrease of smoothness with the size of the network is observed. In those cases, centrality landscapes become less searchable with increasing size.
The aforementioned results are qualitatively the same for both types of local search. Except for the networks from uniform attachment, smoothness values from adaptive and from gradient walks are also quantitatively the same up to statistically insignificant deviations.
network $|V|, |E|$ degree eigenvector shell index betw.ness
----------------- --------------- ------------- ------------- ------------- -------------
\(a) e-mail 1133, 5451 0.882 0.949 1.000 1.000 0.999 1.000 0.867 0.908
0.859 0.892 0.905 0.939 0.992 0.992 0.791 0.851
\(b) jazz 198, 2742 1.000 1.000 1.000 1.000 1.000 1.000 0.830 0.934
1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000
\(c) cond-mat 40421, 175692 0.766 0.773 0.921 0.921 0.791 0.794 0.627 0.671
0.951 0.967 0.957 0.971 0.995 0.995 0.926 0.955
\(d) pgp 10680, 24340 0.569 0.607 1.000 1.000 0.848 0.852 0.736 0.783
0.832 0.853 0.861 0.870 0.898 0.898 0.813 0.846
\(e) Internet 22963, 48436 0.973 0.981 1.000 1.000 1.000 1.000 0.996 0.997
0.869 0.872 0.892 0.892 0.895 0.895 0.888 0.889
\(f) power grid 4941, 6594 0.401 0.410 0.230 0.233 0.651 0.651 0.411 0.437
0.474 0.495 0.623 0.630 0.970 0.970 0.395 0.426
: \[tab:netsmooth\] Smoothness of centrality landscapes for six empirical networks. In each column, the first value gives the smoothness for search by adaptive walk. The second value is for search by gradient walk. For each network, the first (upper) row of smoothness values is for the original network. The second row is for randomized surrogate networks with the same degree sequence (see section \[subsec:randgraphs\]); each value is the mean of the smoothness values of 10 independent randomizations. Networks are (a) the e-mail contacts from University Rovira i Virgili, restricted to the largest connected component [@Guimera:2003]; (b) jazz bands connected by an edge if they share a musician [@gleiser:2003]; (c) authors in the cond-mat e-print archive (arXiv) where $\{v,w\}$ is an edge if $v$ and $w$ have co-authored a paper [@Newman:2001], updated version including data until March 2005; (d) users of the Pretty-Good-Privacy algorithm for secure information interchange [@Boguna:2004]; (e) Internet at Autnomous Systems level, snapshot taken on July 06, 2006. (f) electric power grid with generators, transformers and substations as nodes, edges being high-voltage transmission lines [@Watts:1998]. Network data (a), (b) and (d) were downloaded from <http://deim.urv.cat/~aarenas/data/welcome.htm> ; (c), (e) and (f) from <http://www-personal.umich.edu/~mejn/netdata/>
Now we turn to centrality landscapes based on empirical networks. Table \[tab:netsmooth\] provides an overview of smoothness values for four social networks (a-d) and two technological networks (e-f) further described in the caption. We find that eigenvector centrality induces the smoothest landscapes on the social networks and on the Internet snapshot (e). On the power grid (f), however, eigenvector centrality reaches lower smoothness than the other three centralities. This network is also the minimum of each of the eight smoothness columns, i.e. the power grid yields the lowest smoothness among the networks under a given centrality measure and walk type.
Randomization of these networks under conservation of the degree sequence drastically changes the smoothness values. The searchability of centrality landscapes depends on properties of the empirical networks beyond the degree distribution.
![\[fig:epi\_0\] Smoothness of spreading centrality landscapes under adaptive walks (open squares) and gradient walks (filled squares). Vertical dashed lines indicate the parameter value $\beta_c= 1 / \lambda_\text{max}$ as an estimate of the transition point from local to global spreading [@Chakrabarti:2008]. Curves with number symbols $t \in \{1,2,3\}$ show $s_(L)$, the [*$t$-smoothness*]{} of the adaptive walk at time $t$, cf. Equation (\[eq:smoothfinite\]). See the caption of Table \[tab:netsmooth\] for details on the networks. ](epi_0.pdf){width="\textwidth"}
For landscapes of spreading centrality (see section \[subsec:spreadc\]), the smoothness as a function of the spreading parameter $\beta$ is plotted in Figure \[fig:epi\_0\]. For small $\beta$, smoothness values are close to those of the corresponding degree landscapes, see Table \[tab:netsmooth\]. This finding is in accordance with the large rank order correlation between spreading centrality and degree below the transition [@Klemm:2012]. In the transition regime ($\beta\approx \beta_c$), smoothness increases with $\beta$. The rank order correlation between degree and spreading centrality is also large in the supercritical regime. Here, however, smoothness for spreading centrality does not coincide with smoothness for degree. The $\beta$-dependence of smoothness varies across networks in the supercritical regime.
Above the transition, the $t$-smoothness values $s_t(L)$ for small number of time steps $t$ strongly increases, indicating an accelerated success of the search for more central nodes. Typically in the supercritical regime, $t=2$ or $t=3$ steps are sufficient for the adaptive walk to saturate, i.e. to reach $s_t(L) \approx s(L)$.
Estimates for a given degree distribution
=========================================
Let us estimate the smoothness of degree centrality landscapes on the basis of the degree distribution alone. To this end, we perform an annealed approximation. At each time step, the network is drawn uniformly at random from the set of networks with the given degree distribution.
From the given degree distribution $P(k)$ we obtain the probability of encountering a node of degree $k$ by following a uniformly chosen edge as $$P^\prime(k) = k P(k) / \bar{k}$$ with the average degree $\bar{k} = \sum_{k=0}^\infty k P(k)$. We use the cumulative of $P^\prime$ being $$Q^\prime(k) = \sum_{i=k}^\infty P^\prime(k)~.$$
We define $x(k,t)$ as the fraction of searches being active at a node of degree $k$ at time $t$. A search is active at initialization and turns inactive when reaching a node that is a strict local maximum w.r.t. degree. By $y(x,t)$ we denote the fraction of searches being inactive at a node of degree $k$ at time $t$. We have $\sum_{k=0}^\infty x(k,t)+y(k,t)=1$ as a normalization at all times $t$. Initially, all searches are active and distributed uniformly across nodes, so $x(k,0) = P(k)$ and $y(k,0) =0$ for all $k$.
If all neighbours of a node with degree $k$ have degree strictly smaller than $k$, the adaptive walk stays there, so the search becomes inactive. This happens with probability $$\alpha(k)=[1-Q^\prime(k)]^k~.$$ Thus we have $$\label{eq:ratey}
y(k,t+1) = y(k,t) + \alpha(k) x(k,t)$$ Otherwise, the searcher eventually finds a neighbouring node with degree at least $k$ and jumps there. In order to simplify the equations, we assume that this jump happens immediately in one time step. Compared to the real dynamics, where one or several rejections may occur before the jump, this simplification modifies the transients but not the asymptotic distribution of searches across degrees. When following a uniformly chosen edge, we find a node of degree $k$ with probability $P^\prime(k) / Q^\prime(l)$ under the constraint $k \ge l$. Therefore the active searches propagate as $$\label{eq:ratex}
x(k,t+1) = \sum_{l=1}^{k} (1-\alpha(l)) [P^\prime(k) / Q^\prime(l)] x(l,t)~.$$ When iterating Equations (\[eq:ratey\], \[eq:ratey\]) over time steps, the fraction of active searches converges to zero. The estimate $s^\ast(L)$ of the smoothness is obtained from the asymptotic distribution of inactive searches $$s^\ast(L) = \lim_{t \rightarrow \infty}
\sum_{k=0}^\infty y(k,t) k / k_\text{max}$$ with the maximum degree $k_\text{max}$.
![\[fig:simconf\_0\] Smoothness estimates $s^\ast$ compared to the real smoothness for adaptive walks on degree centrality landscapes. Symbols distinguish networks as follows. Squares: 10 independent realizations of random graphs with $n=1000$ nodes and parameter $p=4/(n-1)$. Cycles: Same as for squares but $n=9000$. Stars: empirical networks randomized by switching under conserved degrees, cf. subsection \[subsec:randgraphs\]. The abscissa of each star is the smoothness averaged over 10 independent realizations of randomization (see also Table \[tab:netsmooth\]). Error bars indicate the standard deviation over the 10 realizations. ](simconf_0.pdf){width="\textwidth"}
Figure \[fig:simconf\_0\] provides a comparison between actual smoothness value $s$ and the estimates $s^\ast$ for several networks. For realizations of random graphs, the degree distribution provides almost complete information on smoothness. The typical difference in smoothness $s$ across realizations is significantly larger than the deviation of the estimate $s^\ast$ from the true value. Applied to randomizations of empirical networks, there is reasonable agreement between $s$ and $s^\ast$ in most of the cases.
Discussion and outlook
======================
We have defined a framework for assessing the searchability of centrality landscapes arising from a network. The [*smoothness*]{} of a network combined with a centrality quantifies the extent to which nodes of large centrality are found by searching the network locally. Local search means that the sequence of nodes visited is a walk following edges on the network.
This local perspective is motivated by the limited information on the network. Rather than knowing the whole network at the outset, a searching agent explores the system by iteratively following connections. In the present framework, we have assumed that the centrality value of each node becomes available ad hoc when encountering the node. This is the case when the centrality values themselves involve local information only, e.g. the degree. Alternatively, the centrality of interest results from local measurement on dynamics taking place on the network. We have investigated the spreading centrality as an example.
Analytic insight into smoothness $s$ of centrality landscapes is desirable, e.g. lower bounds on $s$ depending on network properties. As a step in this direction, we have defined rate equations for estimating smoothness of degree landscapes in an annealed approximation.
Numerically we find that eigenvector centrality typically generates maximally ($s=1$) or almost maximally smooth landscapes. As a heuristic explanation, smoothness arises due to the summation in the eigenvector equations that amounts to an averaging over neighbouring nodes. On the stochastically generated networks lacking small-world property [@Klemm:2002b] and the extremely sparse electric power grid, none of the centrality measures under consideration generates smooth landscapes. Small neighbourhoods and long distances in search space tend to create obstacles to local search.
[^1]: The author acknowledges funding from Volkswagenstiftung.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Kanit Wongsuphasawat[^1]\
Apple Inc.
- |
Yang Liu[^2]\
University of Washington
- 'Jeffrey Heer[^3]\'
bibliography:
- 'eda-interview.bib'
title: |
Goals, Process, and Challenges of\
Exploratory Data Analysis: An Interview Study
---
Conclusion
==========
[^1]: e-mail: [email protected]. This work was done when the first author was at the University of Washington.
[^2]: e-mail: [email protected]
[^3]: e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Assume that $E$ and $B$ are complex manifolds and that $\pi\colon E\to B$ is a holomorphic Serre fibration such that $E$ admits a finite dominating family of holomorphic fiber-sprays over a small neighborhood of any point in $B$. We show that the parametric Oka property (POP) of $B$ implies POP of $E$; conversely, POP of $E$ implies POP of $B$ for contractible parameter spaces. This follows from a parametric Oka principle for holomorphic liftings which we establish in the paper.'
address: 'Faculty of Mathematics and Physics, University of Ljubljana, and Institute of Mathematics, Physics and Mechanics, Jadranska 19, 1000 Ljubljana, Slovenia'
author:
- Franc Forstnerič
title: Invariance of the parametric Oka property
---
[^1]
*Dedicated to Linda P. Rothschild*
The Oka properties {#Sec:Oka}
==================
The main result of this paper is that a subelliptic holomorphic submersion $\pi\colon E\to B$ between (reduced, paracompact) complex spaces satisfies the [*parametric Oka property*]{}. [*Subellipticity*]{} means that $E$ admits a finite dominating family of holomorphic fiber-sprays over a neighborhood of any point in $B$ (Def. \[def:SSF\]). The conclusion means that for any Stein source space $X$, any compact Hausdorff space $P$ (the parameter space), and any continuous map $f\colon X\times P\to B$ which is $X$-holomorphic (i.e., such that $f_p=f(\cdotp,p)\colon X\to B$ is holomorphic for every $p\in P$), a continuous lifting $F\colon X\times P\to E$ of $f$ (satisfying $\pi\circ F=f$) can be homotopically deformed through liftings of $f$ to an $X$-holomorphic lifting. (See Theorem \[SES:lifting-maps\] for a precise statement.)
-5mm $$\xymatrix{ & E \ar[d]^{\pi} \\
X\times P \ar[r]^{f} \ar[ur]^{F} & B }$$
The following result is an easy consequence. Suppose that $E$ and $B$ are complex manifolds and that $\pi\colon E\to B$ is a subelliptic submersion which is also a Serre fibration (such map is called a [*subelliptic Serre fibration*]{}), or is a holomorphic fiber bundle whose fiber satisfies the parametric Oka property. Then the parametric Oka property passes up from the base $B$ to the total space $E$; it also passes down from $E$ to $B$ if the parameter space $P$ is contractible, or if $\pi$ is a weak homotopy equivalence (Theorem \[SES:ascend-descend\]).
We begin by recalling the relevant notions. Among the most interesting phenomena in complex geometry are, on the one hand, [*holomorphic rigidity*]{}, commonly expressed by Kobayashi-Eisenman hyperbolicity; and, on the other hand, [*holomorphic flexibility*]{}, a term introduced in [@FF:flexibility]. While Kobayashi hyperbolicity of a complex manifold $Y$ implies in particular that there exist no nonconstant holomorphic maps ${\mathbb{C}}\to Y$, flexibility of $Y$ means that it admits many nontrivial holomorphic maps $X\to Y$ from any Stein manifold $X$; in particular, from any Euclidean space ${\mathbb{C}}^n$.
The most natural flexibility properties are the [*Oka properties*]{} which originate in the seminal works of Oka [@Oka1] and Grauert [@Grauert2; @Grauert3]. The essence of the classical [*Oka-Grauert principle*]{} is that a complex Lie group, or a complex homogeneous manifold, $Y$, enjoys the following:
**Basic Oka Property (BOP) of $Y$**: [*Every continuous map $f \colon X\to Y$ from a Stein space $X$ is homotopic to a holomorphic map. If in addition $f$ is holomorphic on (a neighborhood of) a compact ${\mathcal{O}}(X)$-convex subset $K$ of $X$, and if $f|_{X'}$ is holomorphic on a closed complex subvariety $X'$ of $X$, then there is a homotopy $f^t\colon X\to Y$ $(t\in [0,1])$ from $f^0=f$ to a holomorphic map $f^1$ such that for every $t\in [0,1]$, $f^t$ is holomorphic and uniformly close to $f^0$ on $K$, and $f^t|_{X'}=f|_{X'}$.*]{}
All complex spaces in this paper are assumed to be reduced and paracompact. A map is said to be holomorphic on a compact subset $K$ of a complex space $X$ if it is holomorphic in an open neighborhood of $K$ in $X$; two such maps are identified if they agree in a (smaller) neighborhood of $K$; for a family of maps, the neighborhood should be independent of the parameter.
When $Y={\mathbb{C}}$, BOP combines the Oka-Weil approximation theorem and the Cartan extension theorem. BOP of $Y$ means that, up to a homotopy obstruction, the same approximation-extension result holds for holomorphic maps $X\to Y$ from any Stein space $X$ to $Y$.
Denote by ${\mathcal{C}}(X,Y)$ (resp. by ${\mathcal{O}}(X,Y)$) the space of all continuous (resp. holomorphic) maps $X\to Y$, endowed with the topology of uniform convergence on compacts. We have a natural inclusion $$\label{incl}
{\mathcal{O}}(X,Y) \hookrightarrow {\mathcal{C}}(X,Y).$$ BOP of $Y$ implies that every connected component of ${\mathcal{C}}(X,Y)$ contains a component of ${\mathcal{O}}(X,Y)$. By [@FF:CAP Theorem 5.3], BOP also implies the following
[**One-parametric Oka Property:**]{} [*A path $f\colon [0,1] \to {\mathcal{C}}(X,Y)$ such that $f(0)$ and $f(1)$ belong to ${\mathcal{O}}(X,Y)$ can be deformed, with fixed ends at $t=0,1$, to a path in ${\mathcal{O}}(X,Y)$. Hence (\[incl\]) induces a bijection of the path connected components of the two spaces.*]{}
$Y$ enjoys the [**Weak Parametric Oka Property**]{} if for each finite polyhedron $P$ and subpolyhedron $P_0\subset P$, a map $f\colon P\to {\mathcal{C}}(X,Y)$ such that $f(P_0)\subset {\mathcal{O}}(X,Y)$ can be deformed to a map ${\widetilde}f\colon P\to {\mathcal{O}}(X,Y)$ by a homotopy that is fixed on $P_0$: -3mm $$\xymatrix{
P_0 \ar[d]_{incl} \ar[r] & {\mathcal{O}}(X,Y) \ar[d]^{incl} \\
P \ar[r]^{\!\!\!\!\!\! f} \ar[ur]^{{\widetilde}f} & {\mathcal{C}}(X,Y) }$$
This implies that (\[incl\]) is a weak homotopy equivalence [@FPrezelj:OP1 Corollary 1.5].
\[POP\] **(Parametric Oka Property (POP))** Assume that $P$ is a compact Hausdorff space and that $P_0$ is a closed subset of $P$. A complex manifold $Y$ enjoys [POP]{} for the pair $(P,P_0)$ if the following holds. Assume that $X$ is a Stein space, $K$ is a compact ${\mathcal{O}}(X)$-convex subset of $X$, $X'$ is a closed complex subvariety of $X$, and $f\colon X\times P\to Y$ is a continuous map such that
- the map $f_p=f(\cdotp,p)\colon X\to Y$ is holomorphic for every $p\in P_0$, and
- $f_p$ is holomorphic on $K\cup X'$ for every $p\in P$.
Then there is a homotopy $f^t\colon X\times P\to Y$ $(t\in [0,1])$ such that $f^t$ satisfies properties (a) and (b) above for all $t\in[0,1]$, and also
- $f^1_p$ is holomorphic on $X$ for all $p\in P$,
- $f^t$ is uniformly close to $f$ on $K\times P$ for all $t\in [0,1]$, and
- $f^t=f$ on $(X\times P_0)\cup (X'\times P)$ for all $t\in [0,1]$.
The manifold $Y$ satisfies [[POP]{}]{} if the above holds for each pair $P_0\subset P$ of compact Hausdorff spaces. Analogously we define [[POP]{}]{} for sections of a holomorphic map $Z\to X$.
Restricting POP to pairs $P_0\subset P$ consisting of finite polyhedra we get Gromov’s Ell$_\infty$ property [@Gromov:OP Def. 3.1.A.]. By Grauert, all complex homogeneous manifolds enjoy POP for finite polyhedral inclusions $P_0\subset P$ [@Grauert2; @Grauert3]. A weaker sufficient condition, called [*ellipticity*]{} (the existence of a dominating spray on $Y$, Def. \[fiber-spray\] below), was found by Gromov [@Gromov:OP]. A presumably even weaker condition, [*subellipticity*]{} (Def. \[def:SS\]), was introduced in [@FF:subelliptic].
If $Y$ enjoys BOP or POP, then the corresponding Oka property also holds for sections of any holomorphic fiber bundle $Z\to X$ with fiber $Y$ over a Stein space $X$ [@FF:Kohn]. See also Sect. \[subelliptic\] below and the papers [@FF:surveyOka; @Larusson1; @Larusson2; @Larusson3].
It is important to know which operations preserve Oka properties. The following result was stated in [@FF:CAP] (remarks following Theorem 5.1), and more explicitly in [@FF:EOP Corollary 6.2]. (See also [@Gromov:OP Corollary 3.3.C’].)
\[SES:ascend-descend\] Assume that $E$ and $B$ are complex manifolds. If $\pi \colon E\to B$ is a subelliptic Serre fibration (Def. \[def:SSF\] below), or a holomorphic fiber bundle with [[POP]{}]{} fiber, then the following hold:
- If $B$ enjoys the parametric Oka property [([[POP]{}]{})]{}, then so does $E$.
- If $E$ enjoys [[POP]{}]{} for contractible parameter spaces $P$ (and arbitrary closed subspaces $P_0$ of $P$), then so does $B$.
- If in addition $\pi \colon E\to B$ is a weak homotopy equivalence then $${\rm POP\ of\ } E \Longrightarrow \rm{POP\ of\ } B.$$
All stated implications hold for a specific pair $P_0\subset P$ of parameter spaces.
The proof Theorem \[SES:ascend-descend\], proposed in [@FF:EOP], requires the parametric Oka property for sections of certain continuous families of subelliptic submersions. When Finnur Lárusson asked for explanation and at the same time told me of his applications of this result [@Larusson4] (personal communication, December 2008), I decided to write a more complete exposition. We prove Theorem \[SES:ascend-descend\] in Sec. \[AD\] as a consequence of Theorem \[SES:lifting-maps\]. This result should be compared with Lárusson’s [@Larusson4 Theorem 3] where the map $\pi \colon E\to B$ is assumed to be an [*intermediate fibration*]{} in the model category that he constructed.
\[c1.6\] Let $Y=Y_m\to Y_{m-1}\to \cdots\to Y_0$, where each $Y_j$ is a complex manifold and every map $\pi_j\colon Y_j\to Y_{j-1}$ $(j=1,2,\ldots,m)$ is a subelliptic Serre fibration, or a holomorphic fiber bundle with [[POP]{}]{} fiber. Then the following hold:
- If one of the manifolds $Y_j$ enjoys [[BOP]{}]{}, or [[POP]{}]{} with a contractible parameter space, then all of them do.
- If in addition every $\pi_j$ is acyclic (a weak homotopy equivalence) and if $Y$ is a Stein manifold, then every manifold $Y_j$ in the tower satisfies the implication [[BOP]{}]{} $\Longrightarrow$ [[POP]{}]{}.
Part (i) is an immediate consequences of Theorem \[SES:ascend-descend\]. For (ii), observe that BOP of $Y_j$ implies BOP of $Y$ by Theorem \[SES:ascend-descend\] (i), applied with $P$ a singleton. Since $Y$ is Stein, BOP implies that $Y$ is elliptic (see Def. \[def:SS\] below); for the simple proof see [@FPrezelj:OP3 Proposition 1.2] or [@Gromov:OP 3.2.A.]. By Theorem \[SES:OPS\] below it follows that $Y$ also enjoys POP. By part (iii) of Theorem \[SES:ascend-descend\], POP descends from $Y=Y_m$ to every $Y_j$.
\[BiP\] A main remaining open problem is whether the implication $$\label{BOPimpliesPOP}
{{\rm BOP}}\Longrightarrow {{\rm POP}}$$ holds for all complex manifolds. By using results of this paper and of his earlier works, F. Lárusson proved this implication for a large class of manifolds, including all quasi-projective manifolds [@Larusson4 Theorem 4]. The main observation is that there exists an affine holomorphic fiber bundle $\pi\colon E\to {\mathbb{P}}^n$ with fiber ${\mathbb{C}}^n$ whose total space $E$ is Stein; since the map $\pi$ is acyclic and the fiber satisfies POP, the implication (\[BOPimpliesPOP\]) follows from Corollary \[c1.6\] (ii) for any closed complex subvariety $Y\subset {\mathbb{P}}^n$ (since the total space $E|_Y=\pi^{-1}(Y)$ is Stein). The same applies to complements of hypersurfaces in such $Y$; the higher codimension case reduces to the hypersurface case by blowing up.
Subelliptic submersions and Serre fibrations {#subelliptic}
============================================
A holomorphic map $h \colon Z\to X$ of complex spaces is a [*holomorphic submersion*]{} if for every point $z_0\in Z$ there exist an open neighborhood $V\subset Z$ of $z_0$, an open neighborhood $U\subset X$ of $x_0=h(z_0)$, an open set $W$ in a Euclidean space ${\mathbb{C}}^p$, and a biholomorphic map $\phi\colon V\to U\times W$ such that $pr_1\circ \phi= h$, where $pr_1 \colon U\times W\to U$ is the projection on the first factor.
-3mm $$\xymatrix{
\!\!\!\!\!\!\!\!\!\!\!\! Z\supset V \ar[d]^{h} \ar[r]^{\phi} & U\times W \ar[d]^{pr_1} \\
\!\!\!\!\!\!\!\!\!\!\!\!\! X\supset U \ar[r]^{id} & U}$$
Each fiber $Z_x=h^{-1}(x)$ $(x\in X)$ of a holomorphic submersion is a closed complex submanifold of $Z$. A simple example is the restriction of a holomorphic fiber bundle projection $E\to X$ to an open subset $Z$ of $E$.
We recall from [@Gromov:OP; @FF:subelliptic] the notion of a holomorphic spray and domination.
\[fiber-spray\] Assume that $X$ and $Z$ are complex spaces and $h\colon Z\to X$ is a holomorphic submersion. For $x\in X$ let $Z_x=h^{-1}(x)$.
- A [*fiber-spray*]{} on $Z$ is a triple $(E,\pi,s)$ consisting of a holomorphic vector bundle $\pi\colon E\to Z$ together with a holomorphic map $s\colon E\to Z$ such that for each $z\in Z$ we have $$s(0_z)=z,\quad s(E_z) \subset Z_{h(z)}.$$
- A spray $(E,\pi,s)$ is [*dominating*]{} at a point $z\in Z$ if its differential $$({\mathrm{d}}s)_{\, 0_z} \colon {\mathrm{T}}_{0_z} E \to {\mathrm{T}}_z Z$$ at the origin $0_z \in E_z=\pi^{-1}(z)$ maps the subspace $E_z$ of ${\mathrm{T}}_{0_z} E
$ surjectively onto the [*vertical tangent space*]{} ${\mathrm{VT}}_z Z=\ker {\mathrm{d}}h_z$. The spray is [*dominating*]{} (on $Z$) if it is dominating at every point $z\in Z$.
- A family of $h$-sprays $(E_j,\pi_j,s_j)$ $(j=1,\ldots,m)$ on $Z$ is dominating at the point $z\in Z$ if $$({\mathrm{d}}s_1)_{0_z}(E_{1,z}) + ({\mathrm{d}}s_2)_{0_z}(E_{2,z})\cdots
+ ({\mathrm{d}}s_m)_{0_z}(E_{m,z})= {\mathrm{VT}}_z Z.$$ If this holds for every $z\in Z$ then the family is [*dominating*]{} on $Z$.
- A spray on a complex manifold $Y$ is a fiber-spray associated to the constant map $Y\to point$.
The simplest example of a spray on a complex manifold $Y$ is the flow $Y\times {\mathbb{C}}\to Y$ of a ${\mathbb{C}}$-complete holomorphic vector field on $Y$. A composition of finitely many such flows, with independent time variables, is a dominating spray at every point where the given collection of vector fields span the tangent space of $Y$. Another example of a dominating spray is furnished by the exponential map on a complex Lie group $G$, translated over $G$ by the group multiplication.
The following notion of an [*elliptic submersion*]{} is due to Gromov [@Gromov:OP Sect. 1.1.B]; [*subelliptic submersions*]{} were introduced in [@FF:subelliptic]. For examples see [@FF:subelliptic; @FF:CAP; @Gromov:OP].
\[def:SS\] A [holomorphic]{} submersion $h\colon Z\to X$ is said to be [*elliptic*]{} (resp. [*subelliptic*]{}) if each point $x_0\in X$ has an open [neighborhood]{} $U\subset X$ such that the restricted submersion $h\colon Z|_U \to U$ admits a dominating fiber-spray (resp. a finite dominating family of fiber-sprays). A complex manifold $Y$ is elliptic (resp. subelliptic) if the trivial submersion $Y\to point$ is such.
The following notions appear in Theorem \[SES:ascend-descend\].
\[def:SSF\] (a) A continuous map $\pi\colon E\to B$ is [*Serre fibration*]{} if it satisfies the homotopy lifting property for polyhedra (see [[@Whitehead p. 8]]{}).
\(b) A holomorphic map $\pi\colon E\to B$ is an [*elliptic Serre fibration*]{} (resp. a [*subelliptic Serre fibration*]{}) if it is a surjective elliptic (resp. subelliptic) submersion and also a Serre fibration.
The following result was proved in [@FF:Kohn] (see Theorems 1.4 and 8.3) by following the scheme proposed in [@FPrezelj:OP3 Sect. 7]. Earlier results include Gromov’s Main Theorem [@Gromov:OP Theorem 4.5] (for elliptic submersions onto Stein manifolds, without interpolation), [@FPrezelj:OP3 Theorem 1.4] (for elliptic submersions onto Stein manifolds), [@FF:subelliptic Theorem 1.1] (for subelliptic submersion), and [@FF:CAP Theorem 1.2] (for fiber bundles with POP fibers over Stein manifolds).
\[SES:OPS\] Let $h\colon Z\to X$ be a holomorphic submersion of a complex space $Z$ onto a Stein space $X$. Assume that $X$ is exhausted by Stein Runge domains $D_1\Subset D_2\Subset \cdots \subset X=\bigcup_{j=1}^\infty D_j$ such that every $D_j$ admits a stratification $$\label{strat}
D_j = X_0\supset X_{1}\supset\cdots\supset X_{m_j} =\emptyset$$ with smooth strata $S_k=X_k\bs X_{k+1}$ such that the restriction of $Z\to X$ to every connected component of each $S_k$ is a subelliptic submersion, or a holomorphic fiber bundle with [[POP]{}]{} fiber. Then sections $X\to Z$ satisfy [[POP]{}]{}.
In previous papers [@FPrezelj:OP1; @FPrezelj:OP2; @FPrezelj:OP3; @FF:subelliptic; @FF:CAP; @FF:EOP] POP was only considered for pairs of parameter spaces $P_0\subset P$ such that
(\*) $P$ is a nonempty compact Hausdorff space, and $P_0$ is a closed subset of $P$ that is a strong deformation neighborhood retract (SDNR) in $P$.
Here we dispense with the SDNR condition by using the Tietze extension theorem for maps into Hilbert spaces (see the proof of Proposition \[local-ext\]).
Theorem \[SES:OPS\] also hold when $P$ is a locally compact and countably compact Hausdorff space, and $P_0$ is a closed subspace of $P$. The proof requires only a minor change of the induction scheme (applying the diagonal process).
On the other hand, all stated results remain valid if we restrict to pairs $P_0\subset P$ consisting of finite polyhedra; this suffices for most applications.
Theorem \[SES:OPS\] implies the following result concerning holomorphic liftings.
\[Lift\] Let $\pi \colon E\to B$ be a holomorphic submersion of a complex space $E$ onto a complex space $B$. Assume that $B$ admits a stratification $B=B_0\supset B_1\supset \cdots \supset B_m=\emptyset$ by closed complex subvarieties such that the restriction of $\pi$ to every connected component of each difference $B_j\bs B_{j+1}$ is a subelliptic submersion, or a holomorphic fiber bundle with [[POP]{}]{} fiber.
Given a Stein space $X$ and a holomorphic map $f\colon X\to B$, every continuous lifting $F\colon X\to E$ of $f$ $(\pi\circ F=f)$ is homotopic through liftings of $f$ to a holomorphic lifting.
-4mm $$\xymatrix{ & E \ar[d]^{\pi} \\
X \ar[r]^{f} \ar[ur]^{F} & B }$$
Assume first that $X$ is finite dimensional. Then there is a stratification $ X=X_0\supset X_{1}\supset\cdots\supset X_l=\emptyset$ by closed complex subvarieties, with smooth differences $S_j=X_j\bs X_{j+1}$, such that each connected component $S$ of every $S_j$ is mapped by $f$ to a stratum $B_k\bs B_{k+1}$ for some $k=k(j)$. The pull-back submersion $$f^*E= \{(x,e)\in X\times E\colon f(x)=\pi(e)\} \longmapsto X$$ then satisfies the assumptions of Theorem \[SES:OPS\] with respect to this stratification of $X$. Note that liftings $X\to E$ of $f\colon X\to B$ correspond to sections $X\to f^*E$, and hence the result follows from Theorem \[SES:OPS\]. The general case follows by induction over an exhaustion of $X$ by an increasing sequence of relatively compact Stein Runge domains in $X$.
A fascinating application of Theorem \[Lift\] has recently been found by Ivarsson and Kutzschebauch [@Ivarsson-Kutzschebauch; @Ivarsson-Kutzschebauch2] who solved the following [*Gromov’s Vaserstein problem*]{}:
\[Ivar-Kut\] [(Ivarsson and Kutzschebauch [@Ivarsson-Kutzschebauch; @Ivarsson-Kutzschebauch2])]{} Let $X$ be a finite dimensional reduced Stein space and let $f\colon X\to {\rm SL}_m(\mathbb{C})$ be a null-homotopic holomorphic mapping. Then there exist a natural number $N$ and holomorphic mappings $G_1,\dots, G_{N}\colon X\to \mathbb{C}^{m(m-1)/2}$ (thought of as lower resp. upper triangular matrices) such that $$f(x) = \left(\begin{array}{cc} 1 & 0 \cr G_1(x) & 1 \cr \end{array} \right)
\left(\begin{array}{cc} 1 & G_2(x) \cr 0 & 1 \cr \end{array} \right) \ldots
\left(\begin{array}{cc} 1 & G_N(x)\cr 0 & 1 \cr \end{array} \right)$$ is a product of upper and lower diagonal unipotent matrices. (For odd $N$ the last matrix has $G_N(x)$ in the lower left corner.)
In this application one takes $B=SL_m({\mathbb{C}})$, $E=({\mathbb{C}}^{m(m-1)/2})^N$, and $\pi\colon E\to B$ is the map $$\pi(G_1,G_2,\ldots,G_N)=
\left(\begin{array}{cc} 1 & 0 \cr G_1 & 1 \cr \end{array} \right)
\left(\begin{array}{cc} 1 & G_2 \cr 0 & 1 \cr \end{array} \right) \ldots
\left(\begin{array}{cc} 1 & G_N\cr 0 & 1 \cr \end{array} \right).$$ Every null-homotopic holomorphic map $f\colon X\to B=SL_m({\mathbb{C}})$ admits a continuous lifting $F\colon X\to E$ for a suitably chosen $N\in {\mathbb{N}}$ (Vaserstein [@Vaser]), and the goal is to deform $F$ to a holomorphic lifting $G=(G_1,\ldots,G_N) \colon X\to E$. This is done inductively by applying Theorem \[Lift\] to auxiliary submersions obtained by composing $\pi$ with certain row projections. Stratified elliptic submersions naturally appear in their proof.
Convex Approximation Property {#Sec:CAP}
=============================
In this section we recall from [@FF:CAP] a characterization of Oka properties in terms of an Oka-Weil approximation property for entire maps ${\mathbb{C}}^n\to Y$.
Let $z=(z_1,\ldots,z_n)$, $z_j=x_j+{\mathrm{i}}\, y_j$, be complex coordinates on ${\mathbb{C}}^n$. Given numbers $a_j,b_j >0$ $(j=1,\ldots,n)$ we set $$\label{cube}
Q =\{z\in {\mathbb{C}}^n\colon |x_j| \le a_j,\ |y_j|\le b_j,\ j=1,\ldots,n\}.$$
\[special\] A [*special convex set*]{} in ${\mathbb{C}}^n$ is a compact convex set of the form $$\label{special-convex}
K =\{z\in Q \colon y_n \le \phi(z_1,\ldots,z_{n-1},x_n)\},$$ where $Q$ is a cube (\[cube\]) and $\phi$ is a continuous concave function with values in $(-b_n,b_n)$. Such $(K,Q)$ is called a [*special convex pair*]{} in ${\mathbb{C}}^n$.
\[def:CAP\] A complex manifold $Y$ enjoys the [*Convex Approximation Property*]{} [(CAP)]{} if every holomorphic map $f\colon K\to Y$ on a special convex set $K \subset Q \subset {\mathbb{C}}^n$ (\[special-convex\]) can be approximated, uniformly on $K$, by holomorphic maps $Q \to Y$.
$Y$ enjoys the [*Parametric Convex Approximation Property*]{} [(PCAP)]{} if for every special convex pair $(K,Q)$ and for every pair of parameter spaces $P_0\subset P$ as in Def. \[POP\], a map $f\colon Q \times P\to Y$ such that $f_p=f(\cdotp,p)\colon Q\to Y$ is holomorphic for every $p\in P_0$, and is holomorphic on $K$ for every $p\in P$, can be approximated uniformly on $K\times P$ by maps ${\widetilde}f\colon Q\times P\to Y$ such that ${\widetilde}f_p$ is holomorphic on $Q$ for all $p\in P$, and ${\widetilde}f_p=f_p$ for all $p\in P_0$.
The following characterization of the Oka property was found in [@FF:CAP; @FF:EOP] (for Stein source manifolds), thereby answering a question of Gromov [@Gromov:OP p. 881, 3.4.(D)]. For the extension to Stein source spaces see [@FF:Kohn].
\[CAP\] For every complex manifold we have $${{\rm BOP}}\ \Longleftrightarrow {{\rm CAP}}, \qquad {{\rm POP}}\Longleftrightarrow {{\rm PCAP}}.$$
The implication ${{\rm PCAP}}\Longrightarrow {{\rm POP}}$ also holds for a specific pair of (compact, Hausdorff) parameter spaces as is seen from the proof in [@FF:CAP]. More precisely, if a complex manifold $Y$ enjoys PCAP for a certain pair $P_0\subset P$, then it also satifies POP for that same pair.
A Parametric Oka Principle for liftings {#SES;liftings}
=======================================
In this section we prove the main result of this paper, Theorem \[SES:lifting-maps\], which generalizes Theorem \[Lift\] to families of holomorphic maps. We begin by recalling the relevant terminology from [@FPrezelj:OP3].
\[P-section\] Let $h\colon Z\to X$ be a holomorphic map of complex spaces, and let $P_0 \subset P$ be topological spaces.
- A [*$P$-section*]{} of $h \colon Z\to X$ is a continuous map $f\colon X\times P\to Z$ such that $f_p=f(\cdotp,p)\colon X\to Z$ is a section of $h$ for each $p\in P$. Such $f$ is [*holomorphic*]{} if $f_p$ is holomorphic on $X$ for each fixed $p\in P$. If $K$ is a compact set in $X$ and if $X'$ is a closed complex subvariety of $X$, then $f$ is [*holomorphic on $K\cup X'$*]{} if there is an open set $U\subset X$ containing $K$ such that the restrictions $f_p|_U$ and $f_p|_{X'}$ are holomorphic for every $p\in P$.
- A [*homotopy of $P$-sections*]{} is a continuous map $H\colon X\times P\times [0,1]\to Z$ such that $H_t=H(\cdotp,\cdotp,t) \colon X\times P\to Z$ is a $P$-section for each $t\in [0,1]$.
- A [*$(P,P_0)$-section*]{} of $h$ is a $P$-section $f\colon X\times P\to Z$ such that $f_p=f(\cdotp,p)\colon X\to Z$ is holomorphic on $X$ for each $p\in P_0$. A $(P,P_0)$-section is holomorphic on a subset $U\subset X$ if $f_p|_U$ is holomorphic for every $p\in P$.
- A $P$-map $X\to Y$ to a complex space $Y$ is a map $X\times P\to Y$. Similarly one defines $(P,P_0)$-maps and their homotopies.
\[SES:lifting-maps\] Assume that $E$ and $B$ are complex spaces and $\pi \colon E\to B$ is a subelliptic submersion (Def. \[def:SSF\]), or a holomorphic fiber bundle with [POP]{} fiber (Def. \[POP\]). Let $P$ be a compact Hausdorff space and $P_0$ a closed subspace of $P$. Given a Stein space $X$, a compact ${\mathcal{O}}(X)$-convex subset $K$ of $X$, a closed complex subvariety $X'$ of $X$, a holomorphic $P$-map $f \colon X\times P \to B$, and a $(P,P_0)$-map $F\colon X\times P \to E$ that is a $\pi$-lifting of $f$ $(\pi\circ F=f)$ and is holomorphic on (a neighborhood of) $K$ and on the subvariety $X'$, there exists a homotopy of liftings $F^t\colon X\times P \to E$ of $f$ $(t\in [0,1])$ that is fixed on $(X\times P_0)\cup (X'\times P)$, that approximates $F=F^0$ uniformly on $K\times P$, and such that $F^1_p$ is holomorphic on $X$ for all $p\in P$.
If in addition $F$ is holomorphic in a neighborhood of $K\cup X'$ then the homotopy $F^t$ can be chosen such that it agrees with $F^0$ to a given finite order along $X'$.
-6mm $$\xymatrix{ & E \ar[d]^{\pi} \\
X\times P \ar[r]_{f} \ar[ur]^{F_t} & B }$$
\[POPmaps\] A map $\pi \colon E\to B$ satisfying the conclusion of Theorem \[SES:lifting-maps\] is said to enjoy the parametric Oka property (c.f. Lárusson [@Larusson2; @Larusson3; @Larusson4]).
The first step is a reduction to the graph case. Set $Z=X\times E$, ${\widetilde}Z=X\times B$, and let ${\widetilde}\pi \colon Z\to {\widetilde}Z$ denote the map $${\widetilde}\pi(x,e)=(x,\pi(e)),\quad x\in X,\ e\in E.$$ Then ${\widetilde}\pi$ is a subelliptic submersion, resp. a holomorphic fiber bundle with POP fiber. Let ${\widetilde}h\colon {\widetilde}Z= X\times B\to X$ denote the projection onto the first factor, and let $h={\widetilde}h\circ {\widetilde}\pi\colon Z\to X$. To a $P$-map $f\colon X\times P \to B$ we associate the $P$-section ${\widetilde}f(x,p)=(x,f(x,p))$ of ${\widetilde}h\colon {\widetilde}Z\to X$. Further, to a lifting $F\colon X\times P\to E$ of $f$ we associate the $P$-section ${\widetilde}F(x,p)=(x,F(x,p))$ of $h\colon Z\to X$. Then ${\widetilde}\pi \circ {\widetilde}F = {\widetilde}f$. This allows us to drop the tilde’s on $\pi$, $f$ and $F$ and consider from now on the following situation:
- $Z$ and ${\widetilde}Z$ are complex spaces,
- $\pi \colon Z \to {\widetilde}Z$ is a subelliptic submersion, or a holomorphic fiber bundle with [[POP]{}]{} fiber,
- ${\widetilde}h \colon {\widetilde}Z\to X$ is a holomorphic map onto a Stein space $X$,
- $f\colon X\times P\to {\widetilde}Z$ is a holomorphic $P$-section of ${\widetilde}h$,
- $F\colon X\times P\to Z$ is a holomorphic $(P,P_0)$-section of $h= {\widetilde}h\circ \pi \colon Z \to X$ such that $\pi\circ F=f$, and $F$ is holomorphic on $K\cup X'$.
We need to find a homotopy $F^t\colon X\times P\to Z$ $(t\in[0,1])$ consisting of $(P,P_0)$-sections of $h \colon Z \to X$ such that $\pi \circ F^t=f$ for all $t\in[0,1]$, and
- $F^0=F$,
- $F^1$ is a holomorphic $P$-section, and
- for every $t\in [0,1]$, $F^t$ is holomorphic on $K$, it is unifomly close to $F^0$ on $K\times P$, and it agrees with $F^0$ on $(X\times P_0)\cup (X'\times P)$.
-3mm $$\xymatrix{ & Z \ar[d]^{\pi} \\
X \times P \ar[r]_{f} \ar[ur]^{F^t} & {\widetilde}Z}$$
Set $f_p=f(\cdotp,p)\colon X\to {\widetilde}Z$ for $p\in P$. The image $f_p(X)$ is a closed Stein subspace of ${\widetilde}Z$ that is biholomorphic to $X$ (since ${\widetilde}h\circ f$ is the identity on $X$).
When $P=\{p\}$ is a singleton, there is only one section $f=f_p$, and the desired conclusion follows by applying Theorem \[SES:OPS\] to the restricted submersion $\pi\colon Z|_{f(X)} \to f(X)$.
In general we consider the family of restricted submersions $Z|_{f_p(X)}\to f_p(X)$ $(p\in P)$. The proof of the parametric Oka principle [@FPrezelj:OP2 Theorem 1.4] requires certain modifications that we now explain. It suffices to obtain a homotopy $F^t$ of liftings of $f$ over a relatively compact subset $D$ of $X$ with $K\subset D$; the proof is then finished by induction over an exhaustion of $X$. The initial step is provided by the following proposition. (No special assumption is needed on the submersion $\pi\colon Z\to {\widetilde}Z$ for this result.)
\[local-ext\] [(Assumptions as above)]{} Let $D$ be an open relatively compact set in $X$ with $K\subset D$. There exists a homotopy of liftings of $f$ over $D$ from $F=F^0|_{D\times P}$ to a lifting $F'$ such that properties ($\alpha$) and ($\gamma$) hold for $F'$, while $(\beta)$ is replaced by
- $F'_p$ is holomorphic on $D$ for all $p$ in a neighborhood $P'_0\subset P$ of $P_0$.
The existence of such local holomorphic extension $F'$ is used at several subsequent steps. We postpone the proof of the proposition to the end of this section and continue with the proof of Theorem \[SES:lifting-maps\]. Replacing $F$ by $F'$ and $X$ by $D$, we assume from now on that $F_p$ is holomorphic on $X$ for all $p\in P'_0$ (a neighborhood of $P_0$).
Assume for the sake of discussion that $X$ is a Stein manifold, that $X'=\emptyset$, and that $\pi \colon Z\to {\widetilde}Z$ is a subelliptic submersion. (The proof in the fiber bundle case is simpler and will be indicated along the way. The case when $X$ has singularities or $X'\ne \emptyset$ uses the induction scheme from [@FF:Kohn], but the details presented here remain unchanged.) It suffices to explain the following:
Let $K\subset L$ be compact strongly pseudoconvex domains in $X$ that are ${\mathcal{O}}(X)$-convex. Assume that $F^0= \{F^0_p\}_{p\in P}$ is a $\pi$-lifting of $f= \{f_p\}_{p\in P}$ such that $F^0_p$ is holomorphic on $K$ for all $p\in P$, and $F^0_p$ is holomorphic on $X$ when $p\in P'_0$. Find a homotopy of liftings $F^t = \{F^t_p\}_{p\in P}$ $(t\in[0,1])$ that are holomorphic on $K$, uniformly close to $F^0$ on $K\times P$, the homotopy is fixed for all $p$ in a neighborhood of $P_0$, and $F^1_p$ is holomorphic on $L$ for all $p\in P$.
Granted the Main Step, a solution satisfying the conclusion of Theorem \[SES:lifting-maps\] is then obtained by induction over a suitable exhaustion of $X$.
[*Proof of Main Step.*]{} We cover the compact set $\bigcup_{p\in P} f_p(\overline{L\bs K})\subset {\widetilde}Z$ by open sets $U_1,\ldots, U_N\subset {\widetilde}Z$ such that every restricted submersion $\pi\colon Z|_{U_j}\to U_j$ admits a finite dominating family of $\pi$-sprays. In the fiber bundle case we choose the sets $U_j$ such that $Z|_{U_j}$ is isomorphic to the trivial bundle $U_j\times Y \to U_j$ with POP fiber $Y$.
Choose a Cartan string ${\mathcal{A}}=(A_0,A_1,\ldots,A_n)$ in $X$ [@FPrezelj:OP2 Def. 4.2] such that $K=A_0$ and $L=\bigcup_{j=0}^n A_j$. The construction is explained in [@FPrezelj:OP2 Corollary 4.5]: It suffices to choose each of the compact sets $A_k$ to be a small strongly pseudoconvex domain such that $\left(\bigcup_{j=0}^{k-1} A_j,A_k\right)$ is a Cartan pair for all $k=1,\ldots,n$. In addition, we choose the sets $A_1,\ldots, A_n$ small enough such that $f_p(A_j)$ is contained in one of the sets $U_l$ for every $p\in P$ and $j=1,\ldots, n$.
We cover $P$ by compact subsets $P_1,\ldots,P_m$ such that for every $j=1,\ldots,m$ and $k=1,\ldots, n$, there is a neighborhood $P'_j\subset P$ of $P_j$ such that the set $
\bigcup_{p\in P'_j} f_p(A_k)
$ is contained in one of the sets $U_l$.
As in [@FPrezelj:OP2] we denote by ${\mathcal{K}}({\mathcal{A}})$ the [*nerve complex*]{} of ${\mathcal{A}}=(A_0,A_1,\ldots,A_n)$, i.e., a combinatorial simplicial complex consisting of all multiindices $J=(j_0,j_1,\ldots,j_k)$, with $0\le j_0<j_1<\cdots < j_k \le n$, such that $$A_J=A_{j_0}\cap A_{j_1}\cap\cdots\cap A_{j_k} \ne \emptyset.$$ Its [*geometric realization*]{}, $K({\mathcal{A}})$, is a finite polyhedron in which every multiindex $J=(j_0,j_1,\ldots,j_k)\in {\mathcal{K}}({\mathcal{A}})$ of length $k+1$ determines a closed $k$-dimensional face $|J|\subset K({\mathcal{A}})$, homeomorphic to the standard $k$-simplex in ${\mathbb{R}}^k$, and every $k$-dimensional face of $K({\mathcal{A}})$ is of this form. The face $|J|$ is called the [*body*]{} (or [*carrier*]{}) of $J$, and $J$ is the [*vertex scheme*]{} of $|J|$. Given $I,J\in {\mathcal{K}}({\mathcal{A}})$ we have $|I|\cap |J|=|I\cup J|$. The vertices of $K({\mathcal{A}})$ correspond to the individual sets $A_j$ in ${\mathcal{A}}$, i.e., to singletons $(j) \in {\mathcal{K}}({\mathcal{A}})$. (See [@Hurewicz-Wallman] or [@Spanier] for simplicial complexes and polyhedra.)
Given a compact set $A$ in $X$, we denote by $\Gamma_{\mathcal{O}}(A,Z)$ the space of all sections of $h\colon Z\to X$ that are holomorphic over some unspecified open neighborhood $A$ in $Z$, in the sense of germs at $A$.
Recall that a [*holomorphic ${\mathcal{K}}({\mathcal{A}},Z)$-complex*]{} [@FPrezelj:OP2 Def. 3.2] is a continuous family of holomorphic sections $$F_* = \{F_J \colon |J| \to \Gamma_{\mathcal{O}}(A_J,Z),\ \ J \in {\mathcal{K}}({\mathcal{A}}) \}$$ satisfying the following compatibility conditions: $$I,J\in {\mathcal{K}}({\mathcal{A}}),\ I\subset J \Longrightarrow
F_J(t) = F_I(t)|_{A_J} \ (\forall t\in |I|).$$ Note that
- $F_{(k)}$ a holomorphic section over (a neighborhood of) $A_k$,
- $F_{(k_0,k_1)}$ is a homotopy of holomorphic sections over $A_{k_0}\cap A_{k_1}$ connecting $F_{(k_0)}$ and $F_{(k_1)}$,
- $F_{(k_0,k_1,k_2)}$ is a triangle of homotopies with vertices $F_{(k_0)},F_{(k_1)},F_{(k_2)}$ and sides $F_{(k_0,k_1)},F_{(k_0,k_2)},F_{(k_1,k_2)}$, etc.
Similarly one defines a continuous ${\mathcal{K}}({\mathcal{A}},Z)$-complex.
A ${\mathcal{K}}({\mathcal{A}},Z;P)$-complex is defined in an obvious way by adding the parameter $p\in P$. It can be viewed as a ${\mathcal{K}}({\mathcal{A}},Z)$-complex of $P$-sections of $Z\to X$, or as a family of ${\mathcal{K}}({\mathcal{A}},Z)$-complexes depending continuously on the parameter $p\in P$. Similarly, a ${\mathcal{K}}({\mathcal{A}},Z;P,P_0)$-complex is a ${\mathcal{K}}({\mathcal{A}},Z;P)$-complex consisting of holomorphic sections (over the set $L=\bigcup_{j=0}^n A_j$) for the parameter values $p\in P_0$. The terminology of Def. \[P-section\] naturally applies to complexes of sections.
By choosing the sets $A_1,\ldots, A_n$ sufficiently small and by shrinking the neighborhood $P'_0$ (furnished by Proposition \[local-ext\]) around $P_0$ if necessary we can deform $F=F^0$ to a holomorphic ${\mathcal{K}}({\mathcal{A}},Z;P,P'_0)$-complex $F_{*,*} = \{F_{*,p}\}_{p\in P}$ such that
- every section in $F_{*,p}$ projects by $\pi\colon Z\to {\widetilde}Z$ to the section $f_p$ (such $F_{*,*}$ is called a [*lifting*]{} of the holomorphic $P$-section $f=\{f_p\}_{p\in P}$),
- $F_{(0),p}$ is the restriction to $A_0=K$ of the initial section $F^0_p$, and
- for $p\in P'_0$, every section in $F_{*,p}$ is the restriction of $F^0_p$ to the appropriate subdomain (i.e., the deformation from $F^0$ to $F_{*,*}$ is fixed over $P'_0$).
A completely elementary construction of such [*initial holomorphic complex*]{} $F_{*,*}$ can be found in [@FPrezelj:OP2 Proposition 4.7].
We observe that, although the map $h={\widetilde}h\circ \pi \colon Z\to X$ is not necessarily a submersion (since the projection ${\widetilde}h\colon {\widetilde}Z\to X$ may have singular fibers), the construction in [@FPrezelj:OP2] still applies since we only work with the fiber component of $F_p$ (over $f_p$) with respect to the submersion $\pi\colon Z \to {\widetilde}Z$. All lifting problems locally reduce to working with functions.
The rest of the construction amounts to finitely many homotopic modifications of the complex $F_{*,*}$. At every step we collapse one of the cells in the complex and obtain a family (parametrized by $P$) of holomorphic sections over the union of the sets that determine the cell. In finitely many steps we obtain a family of [*constant complexes*]{} $F^1=\{F^1_p\}_{p\in P}$, that is, $F^1_p$ is a holomorphic section of $Z\to X$ over $L$. This procedure is explained in [@FPrezelj:OP2 Sect. 5] (see in particular Proposition 5.1.). The additional lifting condition is easily satisfied at every step of the construction. In the end, the homotopy of complexes from $F^0$ to $F^1$ is replaced by a homotopy of constant complexes, i.e., a homotopy of liftings $F^t$ of $f$ that consist of sections over $L$ (see the conclusion of proof of Theorem 1.5 in [@FPrezelj:OP2 p. 657]).
Let us describe more carefully the main step — collapsing a segment in a holomorphic complex. (All substeps in collapsing a cell reduce to collapsing a segment, each time with an additional parameter set.)
We have a special pair $(A,B)$ of compact sets contained in $L\subset X$, called a [*Cartan pair*]{} [@FPrezelj:OP3 Def. 4], with $B$ contained in one of the sets $A_1,\ldots,A_n$ in our Cartan string ${\mathcal{A}}$. (Indeed, $B$ is the intersection of some of these sets.) Further, we have an additional compact parameter set ${\widetilde}P$ (which appears in the proof) and families of holomorphic sections of $h\colon Z\to X$, $a_{(p,\tilde p)}$ over $A$ and $b_{(p,\tilde p)}$ over $B$, depending continuously on $(p,\tilde p)\in P\times {\widetilde}P$ and projecting by $\pi\colon Z\to {\widetilde}Z$ to the section $f_p$. For $p\in P'_0$ we have $a_{(p,\tilde p)}=b_{(p,\tilde p)}$ over $A\cap B$. These two families are connected over $A\cap B$ by a homotopy of holomorphic sections $b^t_{(p,\tilde p)}$ $(t\in[0,1])$ such that $$b^0_{(p,\tilde p)}=a_{(p,\tilde p)},
\quad
b^1_{(p,\tilde p)}=b_{(p,\tilde p)},
\quad \pi\circ b^t_{(p,\tilde p)} =f_p$$ hold for each $p\in P$ and $t\in [0,1]$, and the homotopy is fixed for $p\in P'_0$. These two families are joined into a family of holomorphic sections ${\widetilde}a_{(p,\tilde p)}$ over $A\cup B$, projecting by $\pi$ to $f_p$. The deformation consists of two substeps:
1. by applying the Oka-Weil theorem [@FPrezelj:OP1 Theorem 4.2] over the pair $A\cap B\subset B$ we approximate the family $a_{(p,\tilde p)}$ sufficiently closely, uniformly on a neighborhood of $A\cap B$, by a family ${\widetilde}b_{(p,\tilde p)}$ of holomorphic sections over $B$;
2. assuming that the approximation in (1) is sufficiently close, we glue the families $a_{(p,\tilde p)}$ and ${\widetilde}b_{(p,\tilde p)}$ into a family of holomorphic sections ${\widetilde}a_{(p,\tilde p)}$ over $A\cup B$ such that $\pi\circ {\widetilde}a_{(p,\tilde p)}=f_p$.
For Substep (2) we can use local holomorphic sprays as in [@FF:CAP Proposition 3.1], or we apply [@FPrezelj:OP1 Theorem 5.5]. The projection condition $\pi\circ {\widetilde}a_{(p,\tilde p)}=f_p$ is a trivial addition.
Substep (1) is somewhat more problematic as it requires a dominating family of $\pi$-sprays on $Z|_U$ over an open set $U\subset {\widetilde}Z$ to which the sections $b^t_{(p,\tilde p)}$ project. (In the fiber bundle case we need triviality of the restricted bundle $Z|_U \to U$ and POP of the fiber.) Recall that $B$ is contained in one of the sets $A_k$, and therefore $$\bigcup_{p\in P'_j} f_p(B) \subset \bigcup_{p\in P'_j} f_p(A_k) \subset U_{l(j,k)}.$$ Since $\pi\circ b^t_{(p,\tilde p)} = f_p$ and $Z$ admits a dominating family of $\pi$-sprays over each set $U_l$, Substep (1) applies separately to each of the $m$ families $$\{b^t_{(p,\tilde p)} \colon p\in P'_j,\ \tilde p\in {\widetilde}P,\ t\in[0,1]\},
\quad j=1,\ldots,m.$$
To conclude the proof of the Main Step we use the [*stepwise extension method*]{}, similar to the one in [@FPrezelj:OP2 pp. 138-139]. In each step we make the lifting holomorphic for the parameter values in one of the sets $P_j$, keeping the homotopy fixed over the union of the previous sets.
We begin with $P_1$. The above shows that the Main Step can be accomplished in finitely many applications of Substeps (1) and (2), using the pair of parameter spaces $P_0 \cap P'_1 \subset P'_1$ (instead of $P_0\subset P$). We obtain a homotopy of liftings $\{F^t_p\colon p\in P'_1, t\in [0,1]\}$ of $f_p$ such that $F^{1}_p$ is holomorphic on $L$ for all $p$ in a neighborhood of $P_1$, and $F^t_p=F^0_p$ for all $t\in[0,1]$ and all $p$ in a relative neighborhood of $P_0\cap P'_1$ in $P'_1$. We extend this homotopy to all values $p\in P$ by replacing $F^t_p$ by $F^{t\chi(p)}_p$, where $\chi\colon P\to[0,1]$ is a continuous function that equals one near $P_1$ and has support contained in $P'_1$. Thus $F^1_p$ is holomorphic on $L$ for all $p$ in a neighborhood $V_1$ of $P_0\cup P_1$, and $F^1_p=F^0_p$ for all $p$ in a neighborhood of $P_0$.
We now repeat the same procedure with $F^1$ as the ‘initial’ lifting of $f$, using the pair of parameter spaces $(P_0\cup P_1) \cap P'_2 \subset P'_2$. We obtain a homotopy of liftings $\{F^t_p\}_{t\in[1,2]}$ of $f_p$ for $p\in P'_2$ such that the homotopy is fixed for all $p$ in a neighborhood of $(P_0\cup P_1) \cap P'_2$ in $P'_2$, and $F^2_p$ is holomorphic on $L$ for all $p$ in a neighborhood of $P_0\cup P_1\cup P_2$ in $P$.
In $m$ steps of this kind we get a homotopy $\{F^t\}_{t\in [0,m]}$ of liftings of $f$ such that $F^m_p$ is holomorphic on $L$ for all $p\in P$, and the homotopy is fixed in a neighborhood of $P_0$ in $P$. It remains to rescale the parameter interval $[0,m]$ back to $[0,1]$.
This concludes the proof in the special case when $X$ is a Stein manifold and $X'=\emptyset$. In the general case we follow the induction scheme in the proof of the parametric Oka principle for stratified fiber bundles with POP fibers in [@FF:Kohn]; Cartan strings are now used inside the smooth strata.
When $\pi \colon Z\to {\widetilde}Z$ is a fiber bundle, we apply the one-step approximation and gluing procedure as in [@FF:CAP], without having to deal with holomorphic complexes. The Oka-Weil approximation theorem in Substep (1) is replaced by POP of the fiber.
We begin by considering the special case when $\pi\colon Z={\widetilde}Z\times {\mathbb{C}}\to{\widetilde}Z$ is a trivial line bundle. We have $F_p=(f_p,g_p)$ where $g_p$ is a holomorphic function on $X$ for $p\in P_0$, and is holomorphic on $K\cup X'$ for all $p\in P$. We replace $X$ by a relatively compact subset containing $\bar D$ and consider it as a closed complex subvariety of a Euclidean space ${\mathbb{C}}^N$. Choose bounded pseudoconvex domains $\Omega\Subset \Omega'$ in ${\mathbb{C}}^N$ such that $\bar D \subset \Omega\cap X$.
By [@FPrezelj:OP3 Lemma 3.1] there exist bounded linear extension operators $$\begin{aligned}
S \colon H^\infty(X\cap \,\Omega')
&\longrightarrow& H^2(\Omega) = L^2(\Omega)\cap {\mathcal{O}}(\Omega), \\
S' \colon H^\infty(X'\cap \,\Omega') &\longrightarrow& H^2(\Omega),\end{aligned}$$ such that $S(g)|_{X\cap\, \Omega}=g|_{X\cap\, \Omega}$, and likewise for $S'$. (In [@FPrezelj:OP3] we obtained an extension operator into $H^\infty(\Omega)$, but the Bergman space appeared as an intermediate step. Unlike the Ohsawa-Takegoshi extension theorem [@OhsawaT], this is a soft result depending on the Cartan extension theorem and some functional analysis; the price is shrinking of the domain.) Set $$h_p = S(g_p|_{X\cap \,\Omega'}) - S'(g_p|_{X'\cap \,\Omega'}) \in H^2(\Omega),
\qquad p\in P_0.$$ Then $h_p$ vanishes on $X'$, and hence it belongs to the closed subspace $H^2_{X'}(\Omega)$ consisting of all functions in $H^2(\Omega)$ that vanish on $X'\cap \Omega$. Since these are Hilbert spaces, the generalized Tietze extension theorem (a special case of Michael’s convex selection theorem; see [@Repovs Part C, Theorem 1.2, p. 232] or [@Dowker; @Lee]) furnishes a continuous extension of the map $P_0 \to H^2_{X'}(\Omega)$, $p \to h_p$, to a map $P\ni p \to {\widetilde}h_p\in H^2_{X'}(\Omega)$. Set $$G_p = {\widetilde}h_p + S'(g_p|_{X'\cap \,\Omega'}) \in H^2(\Omega),\qquad p\in P.$$ Then $$G_p|_{X'\cap\, \Omega} = g_p|_{X'\cap\, \Omega} \ (\forall p\in P), \quad
G_p|_{X\cap\, \Omega} = g_p|_{X\cap\, \Omega} \ (\forall p\in P_0).$$ This solves the problem, except that $G_p$ should approximate $g_p$ uniformly on $K$. Choose holomorphic functions $\phi_1,\ldots,\phi_m$ on ${\mathbb{C}}^N$ that generate the ideal sheaf of the subvariety $X'$ at every point in $\Omega'$. A standard application of Cartan’s Theorem B shows that in a neighborhood of $K$ we have $$g_p = G_p + \sum_{j=1}^m \phi_j \, \xi_{j,p}$$ for some holomorphic functions $\xi_{j,p}$ in a neighborhood of $K$, depending continuously on $p\in P$ and vanishing identically on $X$ for $p\in P_0$. (See e.g. [@FPrezelj:OP2 Lemma 8.1].)
Since the set $K$ is ${\mathcal{O}}(X)$-convex, and hence polynomially convex in ${\mathbb{C}}^N$, an extension of the Oka-Weil approximation theorem (using a bounded linear solution operator for the $\dibar$-equation, given for instance by Hörmander’s $L^2$-methods [@Hor] or by integral kernels) furnishes functions ${\widetilde}\xi_{j,p} \in{\mathcal{O}}(\Omega)$, depending continuously on $p\in P$, such that ${\widetilde}\xi_{j,p}$ approximates $\xi_{j,p}$ as close as desired uniformly on $K$, and it vanishes on $X\cap \Omega$ when $p\in P_0$. Setting $${\widetilde}g_p = G_p + \sum_{j=1}^m \phi_j \, {\widetilde}\xi_{j,p},\quad p\in P$$ gives the solution. This proof also applies to vector-valued maps by applying it componentwise.
The general case reduces to the special case by using that for every $p_0 \in P_0$, the Stein subspace $F_{p_0}(X)$ (resp. $f_{p_0}(X)$) admits an open Stein neighborhood in $Z$ (resp. in ${\widetilde}Z$) according to a theorem of Siu [@Dem; @Siu]. Embedding these neighborhoods in Euclidean spaces and using holomorphic retractions onto fibers of $\pi$ (see [@FF:Kohn Proposition 3.2]), the special case furnishes neighborhoods $U_{p_0}\subset U'_{p_0}$ of $p_0$ in $P$ and a $P$-section $F'\colon \bar D\times P\to Z$, homotopic to $F$ through liftings of $f$, such that
- $\pi\circ F'_p=f_p$ for all $p\in P$,
- $F'_p$ is holomorphic on $\bar D$ when $p\in U_{p_0}$,
- $F'_p=F_p$ for $p\in P_0 \cup (P\bs U'_{p_0})$,
- $F'_p|_{X'\cap D}=F_p|_{X'\cap D}$ for all $p\in P$, and
- $F'$ approximates $F$ on $K\times P$.
(The special case is first used for parameter values $p$ in a neighborhood $U'_{p_0}$ of $p_0$; the resulting family of holomorphic maps $\bar D\times U'_{p_0} \to {\mathbb{C}}^N$ is then patched with $F$ by using a cut-off function $\chi(p)$ with support in $U'_{p_0}$ that equals one on a neighborhood $U_{p_0}$ of $p_0$, and applying holomorphic retractions onto the fibers of $\pi$.) In finitely many steps of this kind we complete the proof.
One might wish to extend Theorem \[SES:lifting-maps\] to the case when $\pi\colon E\to B$ is a [*stratified*]{} subelliptic submersion, or a stratified fiber bundle with POP fibers. The problem is that the induced stratifications on the pull-back submersions $f_p^* E\to X$ may change discontinuously with respect to the parameter $p$. Perhaps one could get a positive result by assuming that the stratification of $E\to B$ is suitably compatible with the variable map $f_p \colon X \to B$.
Recall (Def. \[POPmaps\]) that a holomorphic map $\pi\colon E\to B$ satisfies POP if the conclusion of Theorem \[SES:lifting-maps\] holds. We show that this is a local property.
\[localization\] [**(Localization principle for POP)**]{} A holomorphic submersion $\pi\colon E\to B$ of a complex space $E$ onto a complex space $B$ satisfies [[POP]{}]{} if and only if every point $x\in B$ admits an open neighborhood $U_x\subset B$ such that the restricted submersion $\pi\colon E|_{U_x}\to U_x$ satisfies [[POP]{}]{}.
If $\pi\colon E\to B$ satisfies POP then clearly so does its restriction to any open subset $U$ of $B$.
Conversely, assume that $B$ admits an open covering ${\mathcal{U}}=\{U_\alpha\}$ by open sets such that every restriction $E|_{U_\alpha}\to U_\alpha$ enjoys POP. When proving POP for $\pi\colon E\to B$, a typical step amounts to choosing small compact sets $A_1,\ldots,A_n$ in the source (Stein) space $X$ such that, for a given compact set $A_0\subset X$, ${\mathcal{A}}=(A_0,A_1,\ldots,A_n)$ is a Cartan string. We can choose the sets $A_1,\ldots,A_n$ sufficiently small such that each map $f_p\colon X\to B$ in the given family sends each $A_j$ into one of the sets $U_\alpha\in{\mathcal{U}}$.
To the string ${\mathcal{A}}$ we associate a ${\mathcal{K}}({\mathcal{A}},Z;P,P_0)$-complex $F_{*,*}$ which is then inductively deformed into a holomorphic $P$-map ${\widetilde}F \colon \bigcup_{j=0}^n A_j \times P \to E$ such that $\pi\circ {\widetilde}F=f$. The main step in the inductive procedure amounts to patching a pair of liftings over a Cartan pair $(A',B')$ in $X$, where the set $B'$ is contained in one of the sets $A_1,\ldots, A_n$ in the Cartan string ${\mathcal{A}}$. This is subdivided into substeps (1) and (2) (see the proof of Theorem \[SES:lifting-maps\]). Only the first of these substeps, which requires a Runge-type approximation property, is a nontrivial condition on the submersion $E\to B$. It is immediate from the definitions that this approximation property holds if there is an open set $U\subset B$ containing the image $f_p(B')$ (for a certain set of parameter values $p\in P$) such that the restricted submersion $E|_U \to U$ satisfies POP. In our case this is so since we have insured that $f_p(B')\subset f_p(A_j)\subset U_\alpha$ for some $j\in\{1,\ldots,n\}$ and $U_\alpha\in{\mathcal{U}}$.
Ascent and descent of the parametric Oka property {#AD}
=================================================
In this section we prove Theorem \[SES:ascend-descend\] stated in Sec. \[Sec:Oka\].
[*Proof of (i):*]{} Assume that $B$ enjoys POP (which is equivalent to PCAP). Let $(K,Q)$ be a special convex pair in ${\mathbb{C}}^n$ (Def. \[special\]), and let $F \colon Q \times P\to E$ be a $(P,P_0)$-map that is holomorphic on $K$ (Def. \[P-section\]).
Then $f=\pi\circ F\colon Q\times P\to B$ is a $(P,P_0)$-map that is holomorphic on $K$. Since $B$ enjoys POP, there is a holomorphic $P$-map $g\colon Q\times P\to B$ that agrees with $f$ on $Q\times P_0$ and is uniformly close to $f$ on a neighborhood of $K\times P$ in ${\mathbb{C}}^n\times P$.
If the latter approximation is close enough, there exists a holomorphic $P$-map $G\colon K\times P\to E$ such that $\pi\circ G=g$, $G$ approximates $F$ on $K\times P$, and $G=F$ on $K\times P_0$. To find such lifting of $g$, we consider graphs of these maps (as in the proof of Theorem \[SES:lifting-maps\]) and apply a holomorphic retraction onto the fibers of $\pi$ [@FF:Kohn Proposition 3.2].
Since $G=F$ on $K\times P_0$, we can extend $G$ to $(K\times P) \cup (Q\times P_0)$ by setting $G=F$ on $Q\times P_0$.
Since $\pi\colon E\to B$ is a Serre fibration and $K$ is a strong deformation retract of $Q$ (these sets are convex), $G$ extends to a continuous $(P,P_0)$-map $G\colon Q\times P\to E$ such that $\pi\circ G=g$. The extended map remains holomorphic on $K$.
By Theorem \[SES:lifting-maps\] there is a homotopy of liftings $G^t\colon Q\times P\to E$ of $g$ $(t\in[0,1])$ which is fixed on $Q \times P_0$ and is holomorphic and uniformly close to $G^0=G$ on $K\times P$. The holomorphic $P$-map $G^1\colon Q\times P\to E$ then satisfies the condition in Def. \[def:CAP\] relative to $F$. This proves that $E$ enjoys PCAP and hence POP.
[*Proof of (ii):*]{} Assume that $E$ enjoys POP. Let $(K,Q)$ be a special convex pair, and let $f \colon Q \times P\to B$ be a $(P,P_0)$-map that is holomorphic on $K$. Assuming that $P$ is contractible, the Serre fibration property of $\pi\colon E\to B$ insures the existence of a continuous $P$-map $F\colon Q\times P\to E$ such that $\pi\circ F=f$. (The subset $P_0$ of $P$ does not play any role here.) Theorem \[SES:lifting-maps\] furnishes a homotopy $F^t\colon Q\times P\to E$ $(t\in [0,1])$ such that
- $F^0=F$,
- $\pi\circ F^t=f$ for each $t\in[0,1]$, and
- $F^1$ is a $(P,P_0)$-map that is holomorphic on $K$.
This is accomplished in two steps: We initially apply Theorem \[SES:lifting-maps\] with $Q\times P_0$ to obtain a homotopy $F^t\colon Q\times P_0\to E$ $(t\in [0,\frac{1}{2}])$, satisfying properties (a) and (b) above, such that $F^{1/2}_p$ is holomorphic on $Q$ for all $p\in P_0$. For trivial reasons this homotopy extends continuously to all values $p\in P$. In the second step we apply Theorem \[SES:lifting-maps\] over $K\times P$, with $F^{1/2}$ as the initial lifting of $f$ and keeping the homotopy fixed for $p\in P_0$ (where it is already holomorphic), to get a homotopy $F^t$ $(t\in[\frac{1}{2},1])$ such that $\pi\circ F^t=f$ and $F^1_p$ is holomorphic on $K$ for all $p\in P$.
Since $E$ enjoys POP, $F^1$ can be approximated uniformly on $K\times P$ by holomorphic $P$-maps ${\widetilde}F \colon Q\times P\to E$ such that ${\widetilde}F=F^1$ on $Q\times P_0$. Then $${\widetilde}f=\pi\circ {\widetilde}F\colon Q\times P\to B$$ is a holomorphic $P$-map that agrees with $f$ on $Q\times P_0$ and is close to $f$ on $K\times P$.
This shows that $B$ enjoys PCAP for any contractible (compact, Hausdorff) parameter space $P$ and for any closed subspace $P_0$ of $P$. Since the implication PCAP$\Longrightarrow$POP in Theorem \[CAP\] holds for each specific pair $(P_0,P)$ of parameter spaces, we infer that $B$ also enjoys POP for such parameter pairs. This completes the proof of (ii).
[*Proof of (iii):*]{} Contractibility of $P$ was used in the proof of (ii) to lift the map $f \colon Q \times P\to B$ to a map $F\colon Q \times P\to E$. Such a lift exists for every topological space if $\pi\colon E\to B$ is a weak homotopy equivalence. This is because a Serre fibration between smooth manifolds is also a Hurewicz fibration (by Cauty [@Cauty]), and a weak homotopy equivalence between them is a homotopy equivalence by the Whitehead Lemma.
I express my sincere thanks to Finnur Lárusson for his questions which led to this paper, and for very helpful discussions and remarks. I also thank Petar Pavešić and Dušan Repovš for advice on Tietze extension theorem used in the proof of Proposition \[local-ext\].
[10]{}
Cauty, R.: Sur les ouverts des ${\rm CW}$-complexes et les fibrés de Serre, Colloq. Math. **63**, 1–7 (1992)
Demailly, J.-P., Cohomology of $q$-convex spaces in top degrees. Math. Z. *204*, 283–295 (1990)
Dowker, C. H.: Mapping theorems for non-compact spaces. Amer. J. Math. **69**, 200–242 (1947)
Forstnerič, F.: The Oka principle for sections of subelliptic submersions. Math. Z. **241**, 527–551 (2002)
Forstnerič, F.: The homotopy principle in complex analysis: A survey. *Explorations in Complex and Riemannian Geometry: A Volume dedicated to Robert E. Greene*, pp. 73–99, Contemporary Mathematics, **332**, American Mathematical Society, Providence, 2003.
Forstnerič, F.: The Oka principle for multivalued sections of ramified mappings. Forum Math. **15**, 309–328 (2003)
Forstnerič, F.: Holomorphic flexibility properties of complex manifolds. Amer. J. Math. **128**, 239–270 (2006)
Forstnerič, F.: Runge approximation on convex sets implies Oka’s property. Ann. Math. (2) **163**, 689–707 (2006)
Forstnerič, F.: Extending holomorphic mappings from subvarieties in Stein manifolds. Ann. Inst. Fourier **55**, 733–751 (2005)
Forstnerič, F.: The Oka principle for sections of stratified fiber bundles. Pure and Appl. Math. Quarterly, to appear (2009). arXiv: Math.CV/0705.0591
Forstnerič, F., Prezelj, J.: Oka’s principle for holomorphic fiber bundles with sprays. Math. Ann. **317**, 117–154 (2000)
Forstnerič, F., Prezelj, J.: Oka’s principle for holomorphic submersions with sprays. Math. Ann. **322**, 633–666 (2002)
Forstnerič, F., Prezelj, J.: Extending holomorphic sections from complex subvarieties. Math. Z. **236**, 43–68 (2001)
Grauert, H.: Holomorphe Funktionen mit Werten in komplexen Lieschen Gruppen. Math. Ann. **133**, 450–472 (1957)
Grauert, H.: Analytische Faserungen über holomorph-vollständigen Räumen. Math. Ann. **135**, 263–273 (1958)
Gromov, M.: Oka’s principle for holomorphic sections of elliptic bundles. J. Amer. Math. Soc. **2**, 851-897 (1989)
Hörmander, L.: $L\sp{2}$ estimates and existence theorems for the $\bar \partial$ operator. Acta Math. **113** 89–152 (1965)
Hurewicz, W., Wallman, H.: Dimension Theory. Princeton Mathematical Series [**4**]{}, Princeton University Press, Princeton (1941)
Ivarsson, B., Kutzschebauch, F.: A solution of Gromov’s Vaserstein problem. C. R. Acad. Sci. Paris, Ser. I, **336** (2008)
Ivarsson, B., Kutzschebauch, F.: Holomorphic factorization of maps into $\mathrm{Sl}_n({\mathbb{C}})$. Preprint (2008)
Lárusson, F.: Excision for simplicial sheaves on the Stein site and Gromov’s Oka principle. Internat. J. Math. **14**, 191–209 (2003)
Lárusson, F.: Model structures and the Oka principle. J. Pure Appl. Algebra **192**, 203–223 (2004)
Lárusson, F.: Mapping cylinders and the Oka principle. Indiana Univ. Math. J. **54**, 1145–1159 (2005)
Lárusson, F.: Applications of a parametric Oka principle for liftings. Preprint (2009)
Lee, S. W. On the theory of selections. Honam Math. J. **19**, 125–130 (1997)
Ohsawa, T., Takegoshi, K.: On the extension of $L\sp 2$ holomorphic functions. Math. Z. **195**, 197–204 (1987)
Oka, K.: Sur les fonctions des plusieurs variables. III: Deuxième problème de Cousin. J. Sc. Hiroshima Univ. **9**, 7–19 (1939)
Repovš, D., Semenov, P. V.: Continuous Selections of Multivalued Mappings. Mathematics and its Applications **455**, Kluwer Academic Publishers, Dordrecht (1998)
Siu, J.-T., Every Stein subvariety admits a Stein neighborhood. Invent. Math. **38**, 89–100 (1976)
Spanier, E. H.: Algebraic topology. Corrected reprint. Springer-Verlag, New York-Berlin (1981)
Vaserstein, L.: Reduction of a matrix depending on parameters to a diagonal form by addition operations. Proc. Amer. Math. Soc. **103**, 741–746 (1988)
Whitehead, G. W.: Elements of Homotopy Theory. Graduate Texts in Math. **61**, Springer-Verlag, Berlin (1978)
[^1]: Supported by the research program P1-0291, Republic of Slovenia.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This Letter presents the first distance measurement to the massive, semi-detached, eclipsing binary LMC-SC1-105, located in the LH 81 association of the Large Magellanic Cloud (LMC). Previously determined parameters of the system are combined with new near-infrared photometry and a new temperature analysis to constrain the reddening toward the system, and determine a distance of $50.6\pm1.6$ kpc (corresponding to a distance modulus of $18.52\pm0.07$ mag), in agreement with previous eclipsing binary measurements. This is the sixth distance measurement to an eclipsing binary in the LMC, although the first to an O-type system. We thus demonstrate the suitability of O-type eclipsing binaries (EBs) as distance indicators. We suggest using bright, early-type EBs to measure distances along different sight lines, as an independent way to map the depth of the LMC and resolve the controversy about its three-dimensional structure.'
author:
- 'Alceste Z. Bonanos, Norberto Castro, Lucas M. Macri, Rolf-Peter Kudritzki'
title: |
The Distance to the Massive Eclipsing Binary LMC-SC1-105\
in the Large Magellanic Cloud
---
Introduction {#section:intro}
============
As one of the nearest galaxies to the Milky Way, the Large Magellanic Cloud (LMC) has naturally been an attractive first rung for the Extragalactic Distance Scale. The [*HST*]{} Key Project [@Freedman01] adopted a distance modulus $\mu=18.50\pm0.10$ mag (corresponding to a distance of 50.1$\pm2.4$ kpc) to the LMC, which has since become the consensus in the community. @Schaefer08 pointed out that overestimation of error bars and band-wagon effects are present in the literature, with pre-2001 LMC distance measurements yielding values between 18.1 and 18.8 mag [see @Benedict02], and post-2001 values clustering around the Key Project value. Given that different systematic errors accompany each method, a careful comparison of the distances resulting from different methods is necessary to characterize them. Furthermore, there is increasing evidence for substantial and complex vertical structure in the disk of the LMC [see review by @vanderMarel06] from studies of red clump stars [@Olsen02; @Subramanian10], Cepheid variables [@Nikolaev04] and RR Lyrae stars [@Pejcha09], which demands further exploration.
The only direct, geometrical method available for measuring distances to stars in the LMC is with eclipsing binaries (EBs). In particular, the light curve provides the fractional radii of the components, the radial velocity semi-amplitudes determine the masses and size of the orbit, which together with the effective temperature determination (e.g. by comparison with synthetic spectra), yield luminosities and therefore distances [see reviews by @Andersen91; @Torres10]. The EB distance method has so far been applied to four early-B type systems [@Guinan98; @Ribas02; @Fitzpatrick02; @Fitzpatrick03] and one G-type giant system [@Pietrzynski09] in the LMC, with individual uncertainties ranging from 1.2 to 2.2 kpc. Four of these systems are located within the bar of the LMC and their individual distances are consistent with the quoted uncertainties, yielding an error-weighted mean value of $49.4\pm1.1$ kpc. A fifth system, located several degrees away in the north-east quadrant of the disk of the LMC, gives a $3\sigma$ shorter distance of $43.2\pm1.8$ kpc.
![Spatial distribution of known EBs from OGLE II and MACHO (blue circles) on the [*Spitzer*]{} 3.6$\mu$m image of the LMC. EBs with measured distances are labeled. Yellow circles mark the most suitable detached EBs for distance determination [@Michalska05]; red circles mark the OGLE II binaries we plan to measure distances to next. The kinematic center (white “x”) from @Kim98 and the dynamical center or center of the bar (green “x”) from @vanderMarel02 are labeled; the solid line corresponds to the line of nodes [@vanderMarel02]. Coordinates are given for J2000.[]{data-label="map"}](fig1.eps){width="8.5cm"}
Figure \[map\] shows the spatial distribution of all known EBs from the OGLE II [@Wyrzykowski03] and MACHO [@Derekas07; @Faccioli07] microlensing surveys of the LMC, and the systems with measured distances, overlaid onto the [*Spitzer*]{} SAGE image in the IRAC 3.6 $\mu$m band [@Meixner06]. A magnitude cut ($V<17$ mag) and period cut ($>1.5$ days) were both applied to the EB catalogs to reject foreground systems and faint systems whose immediate follow up is unrealistic or impossible. The detached EBs selected by @Michalska05 among the OGLE II systems as being most suitable for distance determination are also shown. Both the kinematic center [@Kim98] and the dynamical center [or center of the bar; @vanderMarel02] are overplotted, as is the line of nodes [$\Theta=129.\!^{\circ}9\pm6.\!^{\circ}0\deg$; @vanderMarel02].
Motivated by the evidence for vertical structure in the LMC and the one discrepant EB distance, we proceed to compute the distance to LMC-SC1-105[^1]. LMC-SC1-105 is a massive, semi-detached, short period ($P=4.25$ days) O-type system, with component masses of $\rm
M_{1}=30.9\pm1.0\;{\mbox{M$_{\odot}$}}$, $\rm M_{2}=13.0\pm0.7\;{\mbox{M$_{\odot}$}}$, and radii of $\rm R_{1}=15.1\pm0.2\;{\mbox{R$_{\odot}$}}$, $\rm R_{2}=11.9\pm0.2\;{\mbox{R$_{\odot}$}}$ [determined by @Bonanos09]. The very accurate measurement of the radii ($<2\%$) renders the system suitable for a distance determination, given that EB distances are independent of the usual distance ladder and therefore important checks for other methods. However, accurate radii are not sufficient for an accurate distance. Accurate fluxes (i.e. effective temperatures) and extinction estimates are also needed, therefore this Letter sets out to determine these quantities and obtain the distance. Specifically, Section 2 presents new near-infrared photometry of LMC-SC1-105, Section 3 an analysis of the spectra with state-of-the-art model atmospheres, Section 4 the distance determination, and finally, Section 5 a discussion of our results.
Near-Infrared Data
==================
![Phased CPAPIR $JHK_s-$band light curves of LMC-SC1-105.[]{data-label="lcs"}](fig2.eps){width="8.5cm"}
This study makes use of $JHK_s$ observations of LMC-SC1-105 obtained with the CPAPIR camera [@Artigau04] at the CTIO 1.5-m, as part of a synoptic survey of Cepheid variables in the LMC (L. M. Macri et al. 2011, in prep.). The EB was observed at 20 different epochs on 11 nights between 2006 November 5 and 2007 December 2. Time-series PSF photometry was carried out using DAOPHOT and ALLFRAME [@Stetson87; @Stetson94]. Photometric zeropoints were determined using $\sim 2,500$ stars from the 2MASS Point Source Catalog, located within $15\arcmin$ of the system and with $10.5<K_s<13.5$ mag, while color terms were derived using nearly $5\times 10^5$ 2MASS stars across the entire bar of the LMC. Figure \[lcs\] shows the calibrated, phased light curves of LMC-SC1-105. We calculated error-weighted out-of-eclipse mean magnitudes of $J=13.22\pm0.04$, $H=13.27\pm0.04$ and $K_s=13.26\pm0.04$ mag.
Effective Temperature Analysis
==============================
An accurate distance measurement to LMC-SC-105 requires an accurate flux determination for its binary components. We proceed to refine the effective temperatures estimated by @Bonanos09[^2] with the state-of-the-art, NLTE stellar atmosphere code FASTWIND [@Santolaya97; @Puls05], which includes the effects of stellar winds and spherical atmospheric extension.
The analysis involves a direct comparison between the helium lines, which are the main temperature diagnostics at these spectral types, plus H$\alpha$, to constrain the stellar wind, with a complete FASTWIND model grid designed to study O-type stars at the metallicity of the LMC. The grid was developed within the FLAMES-II collaboration [@Evans10] and constructed at the Instituto de Astrofísica de Canarias. Specifically, we derived the set of models that provide the lowest $\chi^2$, using H$\alpha$ and the 10 strongest and lines available[^3]. The synthetic models were downgraded to the instrumental resolution of the observed spectra and the projected rotational velocities $v \sin i$ were refined to 160 km s$^{-1}$ and 120 km s$^{-1}$, for the primary and secondary component, respectively. We fixed the surface gravities to the values determined by @Bonanos09: $\log(g_1)=3.57\pm0.02$ and $\log(g_2)=3.40\pm0.03$[^4]. In practice, we rounded the values to the first decimal point, to match the 0.1 dex step size of the grid. The $\chi^2$ method provides the stellar parameters and their corresponding errors.
The technique was applied to the two highest S/N spectra of LMC-SC1-105 [see @Bonanos09], obtained at phases 0.27 and 0.75, i.e. at the first and second quadratures. Both phases yielded the same temperature for each component, within the errors. Specifically, at first quadrature, we found best fit values of T$_{\rm eff1}=36100\pm1000$ K, T$_{\rm eff2}=33200\pm800$ K, while at the second quadrature T$_{\rm
eff1}=35700\pm1100$ K, T$_{\rm
eff2}=33100\pm900$ K. Figures \[oct23\] and \[oct25\] show the best fit FASTWIND models, plus the effects of the temperature errors in the profiles. The synthetic models, which only include transitions of , and , provide a good match to the observed spectra. Despite not including the Balmer lines in the analysis (except H$\alpha$), the wings of these lines are in good agreement with the models, confirming the accuracy of the $\log(g)$ determination from the EB analysis.
![Best fit FASTWIND model (red) of LMC-SC1-105, at the first quadrature. The blue (green) lines correspond to models with the best fit T$\rm_{eff}$ plus (minus) the 1$\sigma$ error. The set of lines with smaller Doppler shifts corresponds to the primary.[]{data-label="oct23"}](fig3.ps){width="8.5cm"}
![Same as Figure \[oct23\], but for the second quadrature. The set of lines with larger Doppler shifts corresponds to the primary.[]{data-label="oct25"}](fig4.ps){width="8.5cm"}
@Bonanos09 reported changes of the spectral types with phase due to the Struve-Sahade effect [@Stickland97], the largest being from O7V to O8V for the primary, which would have an impact on the temperature of $\sim2000$ K [@Martins05]. Our analysis, however, does not yield any remarkable differences in temperature between the two quadratures. The reason for this is that the classification criteria [@Walborn90] hinge on the lines $\lambda4541$, $\lambda4471$, $\lambda4200$, and +II $\lambda4026$, while the FASTWIND analysis averaged over 10 and lines in the spectrum. The imperfect fits of $\lambda4200$ and $\lambda4471$ by the models (see Figure \[oct23\]), are consistent with a spectral type change.
At phase 0.75, the secondary star shows important deviations in the cores of H$\beta$ and H$\gamma$ from the model (see Figure \[oct25\]), which might be due to excess emission arising from the slow mass transfer or the distorted line profiles of Roche lobe-filling stars [see @Bitner06]. Nonetheless, the rest of the and lines are well modeled within the errors. Some of the lines (e.g. $\lambda$4541) might indicate a slightly higher temperature, however these differences lie within the errors.
Distance
========
The flux $f_{\lambda}$ measured at Earth at a certain wavelength $\lambda$ from a binary at distance $d$ is given by
$$\label{disteq}
f_{\lambda}=\frac{1}{d^2} \left(R_{1}^2 \; F_{1,\lambda}+R_{2}^2 \;
F_{2,\lambda} \right) \times 10^{-0.4 \; A\left(\lambda \right)},$$
where $R_{1}$ and $R_{2}$ are the radii of the two stars and $F_{1,\lambda}$ and $F_{2,\lambda}$ the surface fluxes. The total extinction $A\left(\lambda \right)$ is a function of the reddening $E(B-V)$, the normalized extinction curve $k(\lambda-V) \equiv
E(\lambda-V)/E(B-V)$ and the ratio of total to selective extinction in the $V$ band, $R_V \equiv A(V)/E(B-V)$:
$$A\left(\lambda \right) =
E\left(B-V\right)\left[k\left(\lambda-V\right)+R_V\right].$$
Having measured the temperatures of the stars from the spectra, we computed fluxes and fit to the observed magnitudes, using Equation \[disteq\] and the best-fit FASTWIND model atmospheres for each quadrature determined above. Note that we used the mean radii[^5] of the stars instead of their volume radii as better approximations to compute their projected surface areas.
Following the procedure outlined in @Bonanos06 for the detached EB in M33, we calculated synthetic photometry of the composite spectrum over the appropriate Johnson-Cousins optical filter functions as defined by @Bessell90 and calibrated by @Landolt92, and the 2MASS filter set. Monochromatic fluxes were measured at the isophotal wavelengths [see @Tokunaga05], which best represent the flux in a passband. We used zeropoints from @Bessell98 [Appendix A] and @Cohen03 to convert the fluxes to magnitudes. We reddened the model spectrum using the reddening law parameterization of @Cardelli89, as prescribed in @Schlegel98, and simultaneously fit the optical[^6] and near-infrared $BVIJHK_s$ photometry. Specifically, we computed the intrinsic $(B-V)_0=-0.27$ mag from the model atmospheres at the isophotal wavelengths, thus yielding $E(B-V)=0.11\pm0.01$ mag.
The value of $R_V$ was determined as the value that minimized the error in the SED fit over the six photometric bands. For phase 0.27, we found $R_V=5.8\pm0.4$ and for phase 0.75, $R_V=5.7\pm0.4$. The resulting distance to LMC-SC1-105 and thus the LMC bar is $50.6\pm1.6$ kpc ($\mu=18.52\pm0.07$ mag) for the first quadrature and $50.4\pm1.6$ kpc ($\mu=18.51\pm0.07$ mag) for the second quadrature. The distances are identical within errors. Given the better fit of the FASTWIND models to the spectra at first quadrature, we adopt the distance derived for first quadrature. The fit of the reddened model spectrum to the photometry and the residuals of the fit are shown in the upper and lower panels of Figure \[sed\], respectively. The error in the distances was computed by a bootstrap resampling procedure. We repeated the spectral energy distribution (SED) fitting procedure 1000 times for each quadrature, by randomly selecting (using Gaussian sampling) all the parameters within their errors. We adopt the $\sigma$ of the resulting Gaussian distribution as the uncertainty in the distance.
We tested the robustness of our reddening and distance results, by first fitting the $BVI$ photometry alone, which yielded an identical value for the distance ($50.8\pm1.6$ kpc or $\mu=18.53\pm0.07$ mag, with $R_V=5.7\pm0.4$), thus demonstrating the consistency of the near-infrared with the optical photometry. Next, if we fix $R_V=3.1$, the best fit value for $E(B-V)=0.18$ mag, resulting in a distance of $51.9\pm1.6$ kpc ($\mu=18.58\pm0.07$ mag), i.e. in agreement with our reported result, within errors. If instead we assume $R_V=3.1$ and fix $E(B-V)=0.11\pm0.01$ mag (based on our photometry and the model spectra)[^7], we would derive a much larger distance of 55.2 kpc ($\mu=18.71$ mag), which yields a SED fit error of 0.05 mag (versus 0.01 mag) that is inconsistent with the photometry. The validity and implications of the high value of $R_V$ that we have measured are discussed in the following Section.
![Upper panel: fit of the reddened EB model spectrum (for phase 0.27) to the $BVIJHK_s$ photometry. Lower panel: residuals of the SED fit, in terms of the flux ratio. Error bars correspond to the photometric error for each band in flux units. The best fit values of $E(B-V)=0.11\pm0.01$ mag and $R_{V}=5.8\pm0.4$ yield a distance modulus to the EB and thus the LMC bar of $50.6\pm1.6$ kpc ($\mu=18.52\pm0.07$ mag).[]{data-label="sed"}](fig5.eps){width="8.5cm"}
The error quoted above for $R_V$ was estimated using the Bayesian code CHORIZOS [@Maiz-Apellaniz04]. The available $BVIJHK_s$ photometry was given as input, with Teff in the range 33000$-$36000 K and $\log(g)$ fixed to 3.50, from TLUSTY models. The code yielded best fit mean values (for a single star) of T$_{\rm eff}=34500\pm1100$ K, $R_{\lambda5495}=5.4\pm0.4$ and $E(\lambda4405-\lambda5495)=0.10\pm0.01$ mag, consistent with the values we derived.
Discussion
==========
LMC-SC1-105 is located in the LH 81 association [@Massey00], near the center of the LMC bar. It contains two early O-type stars and three Wolf-Rayet systems, one of which was recently found to be an EB [@Szczygiel10]. Furthermore, this association resides in the superbubble N 154 [@Henize56] = DEM 246 [@Davies76]. We have determined a large value of $R_V=5.8\pm0.4$ toward LMC-SC1-105, however, such high values are not uncommon. @Cardelli89 find $5<R_V\leq5.6$ for 6 out of the 29 OB stars in their sample, while @Fitzpatrick07 find $R_V>5$ for 12 out of the 328 stars in their sample. Large values of $R_V$ simply imply larger dust grain sizes, which are expected to occur in dense regions of the interstellar medium due to accretion and coagulation of grains. We therefore conclude that the environment in which LMC-SC1-105 resides has large dust grains.
In this Letter, we have determined the distance to LMC-SC1-105 and consequently the LMC bar to be $50.6\pm1.6$ kpc ($\mu=18.52\pm0.07$ mag). The agreement we find with previous EB distances to systems in the bar with different spectral types testifies to the robustness of the EB method and its potential as a powerful, independent distance indicator. Furthermore, it confirms that O-type (and semi-detached) EBs are suitable for distance determination, i.e. that the fluxes predicted by FASTWIND are indeed accurate. EB-based distance determinations to M31 [@Ribas05; @Vilardell10] and M33 [@Bonanos06] can therefore provide an independent absolute calibration of the Extragalactic Distance Scale. Future distance determinations to EBs in the LMC (e.g. those marked in Figure \[map\]), will additionally provide $R_V$ values in different environments of the LMC. Finally, we suggest using bright, early-type EBs to measure distances along different sight lines to the LMC, as an independent way to map its depth and resolve the controversy about its vertical structure.
[50]{} natexlab\#1[\#1]{}
, J. 1991, , 3, 91
, E., [Doyon]{}, R., [Vallee]{}, P., [et al.]{} 2004, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 5492, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. [A. F. M. Moorwood & M. Iye]{}, 1479–1486
, G. F., [McArthur]{}, B. E., [Fredrick]{}, L. W., [et al.]{} 2002, , 123, 473
, M. S. 1990, , 102, 1181
, M. S., [Castelli]{}, F., & [Plez]{}, B. 1998, , 333, 231
, M. A. & [Robinson]{}, E. L. 2006, , 131, 1712
, A. Z. 2009, , 691, 407
, A. Z., [Stanek]{}, K. Z., [Kudritzki]{}, R. P., [et al.]{} 2006, , 652, 313
, J. A., [Clayton]{}, G. C., & [Mathis]{}, J. S. 1989, , 345, 245
, M., [Wheaton]{}, W. A., & [Megeath]{}, S. T. 2003, , 126, 1090
, R. D., [Elliott]{}, K. H., & [Meaburn]{}, J. 1976, , 81, 89
, A., [Kiss]{}, L. L., & [Bedding]{}, T. R. 2007, , 663, 249
, C. J., [Bastian]{}, N., [Beletsky]{}, Y., [et al.]{} 2010, in IAU Symposium, Vol. 266, IAU Symposium, ed. [R. de Grijs & J. R. D. L[é]{}pine]{}, 35–40
, L., [Alcock]{}, C., [Cook]{}, K., [et al.]{} 2007, , 134, 1963
, E. L. & [Massa]{}, D. L. 2007, , 663, 320
, E. L., [Ribas]{}, I., [Guinan]{}, E. F., [et al.]{} 2002, , 564, 260
—. 2003, , 587, 685
, W. L., [Madore]{}, B. F., [Gibson]{}, B. K., [et al.]{} 2001, , 553, 47
, E. F., [Fitzpatrick]{}, E. L., [Dewarf]{}, L. E., [et al.]{} 1998, , 509, L21
, K. G. 1956, , 2, 315
, S., [Staveley-Smith]{}, L., [Dopita]{}, M. A., [et al.]{} 1998, , 503, 674
, A. U. 1992, , 104, 340
, T. & [Hubeny]{}, I. 2003, , 146, 417
, J. 2004, , 116, 859
, F., [Schaerer]{}, D., & [Hillier]{}, D. J. 2005, , 436, 1049
, P., [Waterhouse]{}, E., & [DeGioia-Eastwood]{}, K. 2000, , 119, 2214
, M., [Gordon]{}, K. D., [Indebetouw]{}, R., [et al.]{} 2006, , 132, 2268
, G. & [Pigulski]{}, A. 2005, , 434, 89
, S., [Drake]{}, A. J., [Keller]{}, S. C., [et al.]{} 2004, , 601, 260
, K. A. G. & [Salyk]{}, C. 2002, , 124, 2045
, O. & [Stanek]{}, K. Z. 2009, , 704, 1730
, G., [Thompson]{}, I. B., [Graczyk]{}, D., [et al.]{} 2009, , 697, 862
, J., [Urbaneja]{}, M. A., [Venero]{}, R., [et al.]{} 2005, , 435, 669
, I., [Fitzpatrick]{}, E. L., [Maloney]{}, F. P., [et al.]{} 2002, , 574, 771
, I., [Jordi]{}, C., [Vilardell]{}, F., [et al.]{} 2005, , 635, L37
, A. E., [Puls]{}, J., & [Herrero]{}, A. 1997, , 323, 488
, B. E. 2008, , 135, 112
, D. J., [Finkbeiner]{}, D. P., & [Davis]{}, M. 1998, , 500, 525
, P. B. 1987, , 99, 191
—. 1994, , 106, 250
, D. J. 1997, The Observatory, 117, 37
, S. & [Subramaniam]{}, A. 2010, , 520, A24
, D. M., [Stanek]{}, K. Z., [Bonanos]{}, A. Z., [et al.]{} 2010, , 140, 14
, A. T. & [Vacca]{}, W. D. 2005, , 117, 421
, G., [Andersen]{}, J., & [Gim[é]{}nez]{}, A. 2010, , 18, 67
, R. P. 2006, in The Local Group as an Astrophysical Laboratory, ed. [M. Livio & T. M. Brown]{}, 47–71
, R. P., [Alves]{}, D. R., [Hardy]{}, E., [et al.]{} 2002, , 124, 2639
, F., [Ribas]{}, I., [Jordi]{}, C., [et al.]{} 2010, , 509, A70
, N. R. & [Fitzpatrick]{}, E. L. 1990, , 102, 379
, L., [Udalski]{}, A., [Kubiak]{}, M., [et al.]{} 2003, Acta Astronomica, 53, 1
[^1]: Or OGLE J053448.26-694236.4 = MACHO 81.8881.21 = LH 81-72.
[^2]: T$_{\rm
eff1}=35\pm2.5$kK, T$_{\rm eff2}=32.5\pm2.5$kK, for $\log(g)=3.50$ (fixed), from best fit TLUSTY models [@Lanz03].
[^3]: $\lambda\lambda$4026, 4143, 4471, 4713, 4922, 5015, 5875 and $\lambda\lambda$4200, 4541, 5411.
[^4]: Note, the $\log(g)$ error bars given in Table 5 of @Bonanos09 incorrectly correspond to the errors in $g$.
[^5]: $\rm (r_{pole}+r_{side}+r_{back})/3$
[^6]: $B_{\rm max}=12.81\pm0.01$ mag, $V_{\rm max}=12.97\pm0.01$ mag, $I_{\rm max}=13.04\pm0.01$ mag [@Wyrzykowski03].
[^7]: Note, our $E(B-V)$ value is consistent with the range (0.13-0.23 mag) measured by @Massey00 for 34 stars in LH 81.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper describes advances in microwave frequency standards using laser-cooled atoms at BNM-SYRTE. First, recent improvements of the $^{133}$Cs and $^{87}$Rb atomic fountains are described. Thanks to the routine use of a cryogenic sapphire oscillator as an ultra-stable local frequency reference, a fountain frequency instability of $1.6\times 10^{-14}\tau^{-1/2}$ where $\tau $ is the measurement time in seconds is measured. The second advance is a powerful method to control the frequency shift due to cold collisions. These two advances lead to a frequency stability of $2\times 10^{-16}$ at $50~000$ s for the first time for primary standards. In addition, these clocks realize the SI second with an accuracy of $7\times 10^{-16}$, one order of magnitude below that of uncooled devices. In a second part, we describe tests of possible variations of fundamental constants using $^{87}$Rb and $^{133}$Cs fountains. Finally we give an update on the cold atom space clock PHARAO developed in collaboration with CNES. This clock is one of the main instruments of the ACES/ESA mission which is scheduled to fly on board the International Space Station in 2008, enabling a new generation of relativity tests.'
author:
- |
S. Bize, P. Laurent, M. Abgrall,\
H. Marion, I. Maksimovic, L. Cacciapuoti, J. Grünert\
C. Vian, F. Pereira dos Santos, P. Rosenbusch,\
P. Lemonde, G. Santarelli, P. Wolf and A. Clairon\
[*BNM-SYRTE, Observatoire de Paris*]{}\
[*61 Avenue de l’Observatoire 75014 Paris, France.*]{}\
A. Luiten, M. Tobar\
[*The University of Western Australia, School of Physics*]{}\
[*35 Stirling Highway, Crawley, Western Australia.*]{}\
C. Salomon\
[*Laboratoire Kastler Brossel, ENS*]{}\
[*24 rue Lhomond, 75005 Paris, France*]{}
title: Cold atom Clocks and Applications
---
Introduction: Einstein’s legacy in modern clocks {#sec:introduction}
================================================
Modern clocks using laser cooled atoms owe a great deal to the famous 1905 “annus mirabilis” of Einstein. Indeed the three theoretical problems that Einstein beautifully solved in 1905 are key ingredients in current atomic clocks, hundred years after Einstein’s work.
1. First the quanta of light, photons, are routinely used to cool atoms to microkelvin temperatures and to confine them in electromagnetic traps. Atom manipulation is a direct application of energy and momentum exchanges between light and matter. At one microkelvin, cesium atoms which form the basis for the current definition of the SI unit of time, the second, move at an average speed of 7mm.s$^{-1}$, enabling extremely long observation times and thus precision measurements. On Earth, atomic fountains enable unperturbed ballistic flight with duration approaching one second. Furthermore, every experiment in atomic physics routinely uses the Einstein’s photoelectric effect in photodiodes to detect, and control light beams. Furthermore, the concept of photon is intimately connected to the famous Planck relationship $E=h \nu$ between energy, Planck’s constant and frequency of electromagnetic radiation, which is of paramount importance in atomic clocks.
2. Second Einstein’s theory of brownien motion with the famous relationship $k_BT=D/\alpha$ between temperature, diffusion coefficient and friction coefficient not only proved the existence of atoms, but beautifully applies to Doppler and sub-Doppler laser cooling mechanisms at work in every cold atom experiment [@Nobel97]. In optical molasses atoms are viscously confined by the bath of photons, they experience a three-dimensional random walk in position and storage times in excess of 10 seconds have been observed for this brownien motion.
3. Third Einstein’s theory of special (and later general) relativity introduced a new approach relating space and time, and the fundamental concept of relativistic invariance and Lorentz transformation. Einstein predicted that time in a fast moving frame seems to slow down to someone not moving with it, and distances appear shorter. This revolutionary approach had major fundamental as well as practical consequences in the following century. Clocks in different reference frames tick at different rates and the well-known GPS receivers which equip boats, cars, and planes use routinely Einstein’s relativity to determine their position with 10 meter accuracy. Indeed, the atomic clocks onboard the 24 GPS satellites orbiting the Earth at an altitude of 20 000 kms must be corrected for relativistic effects (time dilation and gravitational shift) in order to be synchronized with ground clocks and to reach this positioning accuracy. The correction is about 38 microseconds per day. If each satellite would not apply this compensation the positioning error would reach 11 kilometers per day ! These three papers have had revolutionary consequences in Science and society.
Historically, clocks have played a major role in tests of predictions of relativity theories, from the Hafele-Keating clock transport in jet planes, the Pound-Rebka gravitational shift measurement, the Vessot-Levine GP-A Space hydrogen maser red-shift measurement, and radar ranging Shapiro delay experiment [@Will93]. In addition the current definition of time in the SI unit system, the second, relies on Einstein Equivalence Principle (EEP). This principle is the foundation for all gravitational metric theories that describe gravity as a consequence of curved space-time. The Einstein Equivalence Principle states [@Will93]:
1. if an uncharged test body is placed at an initial event in spacetime and given an initial velocity there, then its subsequent trajectory will be independent of its internal structure and composition
2. In any freely falling frame, the outcome of any local non gravitational test experiments is independent of the velocity of the frame
3. the outcome of any local non gravitational test experiment is independent of where and when in the universe it is performed
An immediate consequence of EEP is that the fundamental constants of physics such as the gravitational constant $G$, or the fine structure constant $\alpha= e^2/4\pi \epsilon_0 \hbar c$ must be independent of time and space.
In this article we first describe recent progress in the realization of the SI second using laser cooled cesium and rubidium clocks. In a second part we use these highly stable devices to perform new tests of Einstein Equivalence Principle, namely the constancy of fundamental constants.
Atomic Fountains
================
The ever increasing control of the motion of atomic samples is at the origin of recent progress in atomic frequency standards and precision measurements [@Gill2001]. Laser cooled and trapped atoms enable long observation times required for high precision measurements. Charged particles confined in Paul or Penning traps offer extremely long storage enabling high precision mass measurements, fundamental tests, and the realization of ultra-stable microwave and optical clocks. The recent NPL frequency measurement of an optical transition in Sr$^+$ ion with an uncertainty of $3\,
10^{-15}$ [@Gill2004] is only a factor three or four worse than the current accuracy of cesium fountains. Precision measurements with neutral atoms on the other hand are usually performed in an atomic fountain where laser cooled atoms ballistically propagate for durations up to one second. In the last decade, atomic clocks and inertial sensors using matter wave interferometry in fountains have become two of the most important applications of cold atoms [@Gill2001; @Clairon1995]. About two dozens of fountain devices are now used for a variety of applications. It has been shown recently that microwave and optical clocks as well as matter-wave inertial sensors belong to the same general class of atom interferometers [@Borde2002]. As an example the current sensitivity in acceleration measurement with atom interferometers is on the order of $3\times 10^{-8}\,$m.s$^{-2}$ in one minute measurement duration and, in a decade, cesium fountain clocks have gained almost two orders of magnitude in accuracy. As we show in this paper the fractional inaccuracy of the BNM-SYRTE fountains at Paris Observatory do not exceed today $7\times 10^{-16}$ which corresponds to less than a single second error over 50 million years, allowing for the realization of SI unit of time, the second, at the same level. About half a dozen fountains throughout the world at metrology institutes including PTB, NIST, IEN, NPL, have now an accuracy near $10^{-15}$, making fountains a major contributor to the accuracy of the TAI (Temps Atomique International). In the future, many applications, such as positioning systems (GPS, GALILEO, GLONASS) as well as scientific applications will benefit from these developments. For instance, deep space satellites have travel durations of several years across the solar system. Precise monitoring of their position requires timescales with very low long term drift. Also, using advanced time and frequency transfer systems (operating at higher carrier frequency and chip rate and/or using two way transfer techniques) may lead to positioning accuracy at the millimeter level for averaging time of a few hundred seconds. This would impact many geodetic applications.
In this paper we show that prospects for further improvements are important. A frequency comparison between two fountains exhibits a stability of $2\, 10^{-16}$ at 50 000 second averaging time, for the first time for atomic standards. This frequency resolution sets the stage for clock accuracy at the $10^{-16}$ level for cesium, almost one order of magnitude potential gain, and even better for rubidium with its far reduced collision shift [@Sortais2000; @Fertig2000]. We begin by recalling the basic operation of fountain atomic clocks and introduce several new techniques which demonstrate frequency measurements with a frequency resolution at the $10^{-16}$ level. The first technique makes use of an ultra-stable cryogenic oscillator to interrogate the clock transition in the fountain. Thanks to its extremely good short term frequency stability and low phase noise, the frequency stability of cesium and rubidium fountains is one order of magnitude below that of fountains using an ultra-stable quartz oscillator as interrogation oscillator. It currently reaches $1.6\times 10^{-14}~\tau^{-1/2}$ where $\tau$ is the averaging time in seconds. The fundamental quantum noise of the clock is now reached with atomic samples of up to $10^7$ atoms. The second advance deals with a new technique to measure and cancel with high precision the collisional shift in the clock. This shift is a major plague in cesium clocks and is much reduced (two orders of magnitude) in rubidium devices [@Sortais2000; @Fertig2000]. The method uses interrupted adiabatic population transfer to prepare precise ratios of atomic densities. We show here that the cesium collisional shift can be measured and cancelled at the $10^{-3}$ level. By comparing rubidium and cesium fountains over a duration of six years, a new upper limit for the drift of fundamental constants has been obtained. Finally we present the development status of the PHARAO cold atom space clock which is under industrial realization. PHARAO will fly onboard the international Space station in the frame of the European ACES mission in 2008-2009 and perform fundamental physics tests such as an improved measurement of Einstein’s red-shift, search for drift of fundamental constants and special relativity tests.
Recent advances in cesium and rubidium fountains {#sec:developments}
================================================
In this section we briefly review recent advances on cesium and rubidium fountains performed in our laboratory, BNM-SYRTE where three laser cooled atomic fountains are in operation. The first one (FO1), in operation since 1994 [@Clairon1995], has been refurbished recently. The second one (FOM), a transportable fountain, is derived from the PHARAO space clock prototype [@Laurent1998]. This fountain was transported on two occasions to the Max Planck Institute in Garching for direct frequency measurement of the hydrogen $1s \rightarrow 2s$ transition [@Fischer2004](See section on stability of fundamental constants). The third one (FO2), a dual fountain operating with $^{133}$Cs or $^{87}$Rb, is described in [@Sortais2000]. Here we only briefly describe the present design and recent improvements of FO1 and FO2. A scheme of the fountain apparatus is shown in Fig.\[fig:fountain\]. An optical bench provides through optical fibers all beams required for manipulating and detecting the atoms. The fountains operate with lin $\perp$ lin optical molasses. Atoms are cooled by six laser beams supplied by preadjusted fiber couplers precisely fixed to the vacuum tank and aligned along the axes of a 3 dimensional coordinate system, where the (111) direction is vertical. In FOM, optical molasses is loaded from a $^{133}$Cs vapor and $3\times 10^7$ atoms are cooled in 400 ms. In FO1 and FO2, optical molasses are loaded from a laser slowed atomic beam which is created by diffusing $^{133}$Cs or $^{87}$Rb vapor through a bundle of capillary tubes. With this setup $3\times 10^8$ $^{133}$Cs atoms can be loaded in 400 ms in FO1. In FO2 an additional transverse cooling of the atomic beam increases the loading rate to $10^9$ atoms in 100 ms for $^{133}$Cs.
![Schematic view of the dual Cs-Rb atomic fountain.[]{data-label="fig:fountain"}](fountain.eps){height="11cm"}
The atoms are launched upwards at 4 m.s$^{-1}$ by using moving optical molasses and cooled to $\sim 1~\mu$K in the moving frame by adiabatically decreasing the laser intensity and increasing the laser detuning. In normal operation atoms in the clock level $|F=3,m_{\mathrm{F}}=0\rangle$ are selected by microwave and light pulses.
![BNM-SYRTE fountain ensemble.[]{data-label="fig:cryogenic_oscillator"}](cryogenic_oscillator.eps){height="12cm"}
About 50 cm above the capture zone, a cylindrical copper cavity (TE$_{011}$ mode) is used to probe the hyperfine transition in a Ramsey interrogation scheme. The cavities have a loaded quality factor of $Q_{\mathrm{FO1}}=10000$ for FO1 and $Q_{\mathrm{FO2}}=6600$ for FO2. Both cavities can be fed with two coupling irises oppositely located on the cavity diameter. Symmetric or asymmetric feedings are used to evaluate and reduce the residual Doppler effect due to imperfections of the standing wave in the cavity and a tilt of the launch direction of the atoms.
![Experimental Ramsey fringes (transition probability as a function of the microwave detuning) measured with $^{133}$Cs in FO2 fountain. The insert shows the central fringe with a FWHM of $\sim
1$ Hz. Each point is a single $1.3$ s measurement. At half maximum of the central fringe, the signal to noise ratio is 5 000, within $20\%$ of the fundamental quantum noise with $\sim 10^{7}$ detected atoms.[]{data-label="fig:ramsey_Fringes"}](fringes.eps){height="10cm"}
The microwaves feeding the cavities are synthesized from the signal of an ultra-stable cryogenic sapphire resonator oscillator (CSO) developed at the University of Western Australia [@Mann2001]. As shown in Fig.\[fig:cryogenic\_oscillator\], the three fountains use the same CSO oscillator to synthesize the microwave signals probing the atomic transition. To reduce its drift, the CSO is weakly phase-locked to a hydrogen maser. This maser contributes to the local timescale and TAI (Temps Atomique International) through various time and frequency transfer systems. With this setup, atomic fountains are used as primary frequency standard to calibrate TAI and can be compared to other remote clocks. Nowadays, atomic fountains are the dominant contributors to the accuracy of TAI.
The 11.932 GHz output signal from the CSO is converted in order to synthesize 11.98 GHz and 100 MHz signals, both phase coherent with the H-maser. FO2 uses the 11.98 GHz signal to generate 9.192 GHz by a home-build low noise synthesizer which achieves a frequency stability of $3\times 10^{-15}$ at 1 s by operating only in the microwave domain. This scheme reduces at the minimum the phase noise and the spurious side-bands induced by the down conversion process. A similar setup is used to synthesize the 6.834 GHz required for the FO2 fountain operation with $^{87}$Rb. The 150 m distance between FO1, FOM and the CSO prevents the direct use of the 11.98 GHz signal. Instead, a 100 MHz signal is synthesized from the CSO and distributed to FO1 and FOM via a high stability RF cable. Finally, a 100 MHz to 9.192 GHz home-made synthesizer generates the interrogation signal. These additional steps degrade the phase noise of the interrogation signal in FO1 and FOM with a frequency stability currently limited to $\sim 2\times10^{-14}$ at 1 s.
Frequency stability {#frequency-stability .unnumbered}
-------------------
![Fractional frequency instability of FO2 against CSO for high density (HD, red squares) and low density (LD, green circles) configurations. It demonstrates a stability of $1.6\times
10^{-14}\tau^{-1/2}$ for a $^{133}$Cs primary standard. Also shown is the fractional frequency instability for the differential measurement between HD and LD (blue triangles). This curve demonstrates an excellent rejection of the CSO fluctuations in the differential measurement, allowing for a fractional frequency resolution of $2.5\times 10^{-16}$ at 20 000 s. In this measurement, the collisional shift at LD is the frequency difference between HD and LD ($\sim 5\times 10^{-14}$). It is obtained with a resolution close to 2 parts in $10^{-16}$ and it is stable at the $0.5\%$ level over 20 000 s.[]{data-label="fig:FO2HDBD"}](FO2HDBD.eps){height="8cm"}
Atoms selected in $|F=3,m_F=0\rangle$ cross the microwave cavity on the way up and on the way down, completing the two Ramsey interactions. After the Ramsey interrogation, the populations $N_e$ and $N_g$ of both clock levels $|e\rangle$ and $|g\rangle$ are measured by fluorescence. The number of detected atoms is typically $0.5\%$ of the initially captured atoms. The signal $p=N_e/(N_e+N_g)$ is equal to the atomic transition probability and is insensitive to atom number fluctuations. A typical Ramsey resonance is presented in Fig.\[fig:ramsey\_Fringes\]. From the transition probability, measured on both sides of the central Ramsey fringe, we compute an error signal to lock the microwave interrogation frequency to the atomic transition using a digital servo loop. At the quantum limit one expects $S/N= 1/\sigma_{\delta p}= 2\sqrt{N}$ for $N$ detected atoms, where $\sigma_{\delta p}$ is the shot to shot standard deviation of the fluctuations of the transition probability. The frequency corrections are applied to a computer controlled high resolution DDS synthesizer in the microwave generator. These corrections are used for the accuracy and frequency stability evaluations of each fountain. The fractional frequency instability of the FO2 fountain operating with $\sim 10^7$ detected atoms and measured against the cryogenic oscillator is plotted Fig.\[fig:FO2vsFO1\]. At the quantum limit one expects a frequency instability, characterized by the fractional Allan standard deviation, given by: $\sigma_{\mathrm{y}}(\tau)= (1/\pi
Q_{at})\sqrt{T_{\mathrm{c}}/N\tau}$, where $Q_{at}\sim 10^{10}$ is the atomic quality factor, $\tau$ and $T_{\mathrm{c}}$ are respectively the averaging time and the cycle duration. Above the servo-loop time constant ($\sim 3$ s) and below 100 s, the fractional instabilities of FO1 and FO2 are $\sigma_{\mathrm{y}}(\tau)=2.9\times 10^{-14}\tau^{-1/2}$ and $1.6\times 10^{-14}\tau^{-1/2}$ respectively, within $\sim 20\%$ of the standard quantum limit. For longer averaging time the frequency instability is dominated by the frequency fluctuations of the CSO and the H-maser. This is the first demonstration of routinely operated primary frequency standards with frequency instabilities in the low $10^{-14}\tau^{-1/2}$ region. We will show below that this excellent short term stability enables an evaluation of systematic frequency shifts and frequency comparisons between clocks at the $10^{-16}$ level in a few days.
Accuracy {#accuracy .unnumbered}
--------
All known systematic frequency shifts are evaluated in our fountains. The accuracy budget for each shift is given in table I for $^{133}$Cs. The overall uncertainty, the quadratic sum of all uncertainties is $7.5\times 10^{-16}$ for FO1, $6.5\times
10^{-16}$ for FO2 and $8\times 10^{-16}$ for FOM. In the following, we only discuss some of the most bothersome effects and the recent improvements in their evaluation. A more complete discussion of systematic effects can be found in [@Vian2004].
---------------------------------------- ---------------------------- ---------------------------- ----------------------------
[FO1 ($\times 10^{16}$)]{} [FO2 ($\times 10^{16}$)]{} [FOM ($\times 10^{16}$)]{}
$1199.7\pm4.5$ $1927.3\pm0.3$ $351.9\pm 2.4$
[Blackbody radiation]{} $-162.8\pm2.5$ $-168.2\pm2.5$ $-191.0\pm2.5$
[Collisions and cavity pulling (HD)]{} $-197.9\pm2.4$ $-357.5\pm2.0$ $-34.0\pm5.8$
[Spectral purity & leakage]{} $0.0\pm3.3$ $0.0\pm4.3$ $0.0\pm2.4$
[First order Doppler effect]{} $<3$ $<3$ $<2$
[Ramsey & Rabi pulling]{} $<1$ $<1$ $<1$
[Microwave recoil]{} $<1.4$ $<1.4$ $<1.4$
[Second order Doppler effect]{} $<0.08$ $<0.08$ $<0.08$
[Background collisions]{} $<1$ $<1$ $<1$
[Total uncertainty]{} $\pm7.5$ $\pm6.5$ $\pm7.7$
---------------------------------------- ---------------------------- ---------------------------- ----------------------------
: Systematic fractional frequency shifts for FO1 and FO2.[]{data-label="tb:accuracy"}
### Cold collisions and cavity pulling {#cold-collisions-and-cavity-pulling .unnumbered}
The cold collision frequency shift is known to be particularly large for $^{133}$Cs [@Gibble1993; @Ghezali1996]. For instance, when FO2 is operated at its best frequency stability the shift is $\sim
10^{-13}$. The linear extrapolation of this effect to zero density is known to be delicate. As pointed out in the first paper observing the cold collision shift in $^{133}$Cs fountains [@Gibble1993], selecting atoms in the clock levels using microwaves may lead to distortions of the position or velocity distribution. Methods to cope with these effects have been proposed [@Fertig2000], yet the linear extrapolations have proved to be valid only at the 5% to 10% level.
To evaluate the collision shift at the $10^{-3}$ level (a requirement for a frequency stability and accuracy at $10^{-16}$), we recently developed a new method based on interrupted adiabatic passage to select atoms in the $|F=3,m_{\mathrm{F}}=0\rangle$ state [@Pereira2002]. Atomic samples are prepared by transferring atoms from $|F=4,m_{\mathrm{F}}=0\rangle$ to $|F=3,m_{\mathrm{F}}=0\rangle$ with an efficiency precisely equal to $100\%$ (high density, HD) or $50\%$ (low density, LD). With this method, the atom number is changed without affecting either the velocity or the position distributions. Therefore, the density ratio LD/HD is equal to the atom number ratio and is $1/2$ at the $10^{-3}$ level. Since the collisional shift is proportional to the atomic density, it can be extrapolated to zero density with this accuracy. In addition, with this method, the cavity frequency pulling [@Sortais2000; @Fertig2000; @Bize2001] is also accounted for.
![Fractional instability of the ratio of the detected atom number in $|F=4,m_F=0\rangle$ between low density and high density configurations as a function of the number of fountain cycles. The measured ratio is $0.5005(2)$. Each cycle lasts $\sim 1.3$ s. The stability (solid line) decreases as the square root of the number of cycles.[]{data-label="fig:ratioHDLDFO1"}](ratioHDLDFO1.eps){height="7cm"}
The collisional shift is measured in real-time with the following differential method. The clock is operated alternately in the HD configuration for $60$ s and in the LD configuration for the next $60$ s. This timing choice minimizes the noise due to frequency instabilities of the CSO oscillator. As seen in Fig.\[fig:FO2HDBD\], at $120$ s the stability of FO2 against CSO is near its minimum. Also, over this $120$ s period, density fluctuations do not exceed $\sim 1\%$. One the other hand, due to slow changes in the clock environment, we observe that the density may fluctuate up to $10-20\%$ over one or several days. Our differential method efficiently cancels these slow daily density variations.
In [@Pereira2002], our calculations predicted that the interrupted adiabatic passage method does provide a LD/HD ratio precisely equal to 1/2 to better than $10^{-3}$. Initially, we were experimentally able to realize this ratio at the $1\%$ level. Improvements in the accuracy of the microwave frequency synthesis for the adiabatic passage enable us to now reach a precision of $2\times 10^{-3}$ for this ratio.
During routine operation of the fountains, the number of detected atoms in each hyperfine state is recorded for both LD and HD configurations. As seen in Fig.\[fig:ratioHDLDFO1\], the Allan standard deviation of the measured LD/HD atom number ratio decreases as the square root of the number of fountain cycles (or time), down to a few parts in $10^4$ for one day of averaging. Despite the $10-20\%$ slow drift in atom number over days, this ratio remains remarkably constant. The LD/HD atom number ratio in $|F=4,m_{\mathrm{F}}=0\rangle$ is found equal to $1/2$ to better than $10^{-3}$. This method relies on fluorescence measurements made in the detection zones for each fountain cycle. Various measurements have been performed to establish that the measurement of this ratio is not biased by more than $10^{-3}$ due optical thickness effects in the detection. On the other hand, the LD/HD atom number ratio in $|F=3,m_{\mathrm{F}}=0\rangle$ is found to slightly differ from 1/2 by $0.3\%$ typically. This deviation originates from atoms in the $|F=3,m_{\mathrm{F}}\neq
0\rangle$ states populated by imperfections in the state preparation. This deviation must be taken into account in the evaluation of the collisional shift. In [@Marion2004], we showed that the frequency shift of the clock transition due to $|F=3,m_{\mathrm{F}}\neq 0\rangle$ atoms is at most $1/3$ of that of collisions between $|F=3,m_{\mathrm{F}}=0\rangle$ and $|F=4,m_{\mathrm{F}}=0\rangle$ clock states. Their contribution to the collisional frequency shift is thus at the $0.1\%$ level. In summary, when this adiabatic passage method is used, we take a $2\times 10^{-3}$ relative uncertainty for the determination of the high density cold collision shift.
### Effect of microwave spectral purity and leakage {#effect-of-microwave-spectral-purity-and-leakage .unnumbered}
Spectral impurities of the interrogation signal and microwave leakage may cause shifts of the clock frequency. In order to evaluate these effects, we make use of their dependance with the microwave power. We alternate the Ramsey interrogation between a configuration of $\pi/2$ and $3\pi/2$ pulses, i.e. a variation of a factor of 9 in microwave power. Within the resolution of the measurement of $3.3\times 10^{-16}$, no frequency shift is observed. In this measurement, four data sets are recorded, LD and HD at $\pi/2$ and LD and HD at $3\pi/2$. In this way, the collisional shift (which may also change with the microwave power) is evaluated and cancelled for both $\pi/2$ and $3\pi/2$ configurations by the differential method described above, allowing for the extraction of a possible influence of microwave spectral purity and leakage alone.
### Residual first order Doppler effect {#residual-first-order-doppler-effect .unnumbered}
A frequency shift due to the first order Doppler effect can occur if the microwave field inside the interrogation cavity exhibits a phase gradient and the atoms pass the cavity with a slight inclination from the cavity axis. We determine the frequency shift due to the linear component of the phase gradient in a differential measurement by coupling the microwave interrogation signal “from the left”, “from the right” or symmetrically into the cavity, providing 3 data sets. The observed shift between the “left” and symmetric configuration is $(-25.3\pm 1.1)\times 10^{-16}$ while the shift between the “right” and symmetric configuration is $(+24.0\pm 1.2)\times
10^{-16}$. The magnitude of this residual first order Doppler effect is consistent with a simple estimate of the residual traveling wave component in the cavity [@Schroder2004] together with a misalignment between the local gravity and the launch direction $\lesssim 1$ mrad. The mean of these two measurements is $(-0.7\pm
0.8)\times 10^{-16}$ and consistent with zero, indicating that the traveling wave component is well cancelled in the symmetric coupling configuration. Using the atoms as a probe, we can indeed ensure that the cavity is fed symmetrically to better than $1\%$ in amplitude and $60$ mrad in phase, which cancels the effect of linear phase gradient to $\sim 1\%$, better than the above measurement resolution. As a consequence, only the quadratic phase dependence of the microwave field remains as a possible source of residual Doppler shift. A worst case estimate based on [@Schroder2004] gives an upper bound for the fractional frequency shift of 3 parts in $10^{16}$, which we conservatively take as the overall uncertainty associated with residual first order Doppler effect. Other contributions to the accuracy budget are listed in Table \[tb:accuracy\]. The total accuracy currently reaches $7.5$ parts in $10^{16}$ for FO1, $6.5$ parts in $10^{16}$ for FO2, and $8.0$ parts in $10^{16}$ for FOM. This represents a one order of magnitude improvement over uncooled cesium devices. In the future, we anticipate that the extensive use of the methods described above will enable us to bring the accuracy of $^{133}$Cs fountains below 2 parts in $10^{16}$ and the accuracy of $^{87}$Rb to a even lower value thanks to its 100-fold lower collision shift [@Sortais2000; @Fertig2000].
### Frequency comparisons between two $^{133}$Cs fountains at $2\,10^{-16}$ {#frequency-comparisons-between-two-133cs-fountains-at-210-16 .unnumbered}
![Fractional frequency instability (Allan deviation) between FO1 and FO2 fountains (blue triangles). After $50~000$ s of averaging, the stability between the two fountains is $2.2$ parts in $10^{16}$. Also plotted is the fractional frequency instability of FO1 (red circles) and FO2 (black squares) against the CSO locked to the hydrogen maser.[]{data-label="fig:FO2vsFO1"}](FO2vsFO1.eps){height="9cm"}
The routine operation of two atomic fountains near the quantum noise limit using the CSO as an interrogation oscillator enables frequency comparisons in the low $10^{-16}$ range, for the first time for primary frequency standards. Figure \[fig:FO2vsFO1\] presents the frequency stability between FO1, FO2 and CSO. Each fountain is operated in differential mode in order to permanently evaluate and cancel the collision shift. Appropriate post-processing of the data thus enables us to construct for each fountain, a clock which is free of the cold collision shift and whose stability is shown in Fig. \[fig:FO2vsFO1\] against the CSO oscillator. Fig. \[fig:FO2vsFO1\] also shows that the combined stability between these two clocks reaches $2.2\times 10^{-16}$ at $50~000$ s, a previously unattained long term stability. From this data, we infer that at least one of the two fountains has a stability below $
(2.2/\sqrt{2})\times 10^{-16}=1.6 \times 10^{-16}$ at the same averaging time. The mean fractional frequency difference between the two fountains is $4\times 10^{-16}$, fully compatible with the accuracy of each of the two clocks as stated in Table \[tb:accuracy\]. This very good stability sets a new challenge for time and frequency transfer systems between remote clocks. As an example, long distance frequency comparisons between PTB and NIST fountains were performed at the level of only $6\times 10^{-16}$ after 2 weeks of averaging with GPS [@Parker2001]. Similarly, comparisons between BNM-SYRTE and PTB recently achieved $2\times
10^{-15}$ for one day of integration with TWSTFT [@Richard2004].
Einstein Equivalence Principle and Stability of fundamental constants {#sec:alpha}
=====================================================================
Highly accurate atomic clocks offer the possibility to perform laboratory tests of possible variations of fundamental constants. Such tests interestingly complement experimental tests of the Local Lorentz Invariance and of the Universality of free-fall to experimentally establish the validity of Einstein’s Equivalence Principle (EEP). They also complement tests of the variability of fundamental constants on different timescales, geological timescale [@Damour1996; @Fujii2003] and cosmological timescale [@Webb2001; @Srianand2004]. Nearly all unification theories (in particular string theories) violate EEP at some level [@Marciano1984; @Damour1994; @Damou2002] which strongly motivates experimental searches for such violations.
Tests described here are based on highly accurate comparisons of atomic energies. In principle, it is possible to express any atomic energy as a function of the elementary particle properties and the coupling constants of fundamental interactions using Quantum Electro-Dynamics (QED) and Quantum Chromo-Dynamics (QCD). As a consequence, it is possible to deduce a constraint to the variation of fundamental constants from a measurement of the stability of the ratio between various atomic frequencies.
Different types of atomic transitions are linked to different fundamental constants. The hyperfine frequency in a given electronic state of alkali-like atoms (involved for instance in $^{133}$Cs, $^{87}$Rb [@Marion2003], $^{199}$Hg$^{+}$ [@Prestage1995; @Berkeland1998], $^{171}$Yb$^{+}$ microwave clocks) can be approximated by: $$\label{eq:hyperfine_energy}
\nu_{\mathrm{hfs}}^{(i)} \simeq R_\infty
c \times\mathcal{A}_{\mathrm{hfs}}^{(i)} \times
g^{(i)}\left(\frac{m_e}{m_p}\right)~\alpha^2~F_{\mathrm{hfs}}^{(i)}(\alpha),$$ where the superscript $(i)$ indicates that the quantity depends on each particular atom. $R_\infty$ is the Rydberg constant, $c$ the speed of light, $g^{(i)}$ the nuclear g-factor, $m_e/m_p$ the electron to proton mass ratio and $\alpha$ the fine structure constant. In this equation, the dimension is given by $R_\infty
c$, the atomic unit of frequency. $\mathcal{A}_{\mathrm{hfs}}^{(i)}$ is a numerical factor which depends on each particular atom. $F_{\mathrm{hfs}}^{(i)}(\alpha)$ is a relativistic correction factor to the motion of the valence electron in the vicinity of the nucleus. This factor strongly depends on the atomic number $Z$ and has a major contribution for heavy nuclei. Similarly, the frequency of an electronic transition (involved in H [@Niering2000], $^{40}$Ca [@Helmcke2003], $^{199}$Hg$^{+}$ [@Bize2003], $^{171}$Yb$^{+}$ [@Stenger2001; @Peik2003] optical clocks) can be approximated by $$\label{eq:electronic_energy}
\nu_{\mathrm{elec}}^{(i)} \simeq R_\infty
c \times \mathcal{A}_{\mathrm{elec}}^{(i)} \times F_{\mathrm{elec}}^{(i)}(\alpha).$$ Again, the dimension is given by $R_\infty c$. $\mathcal{A}_{\mathrm{elec}}^{(i)}$ is a numerical factor. $F_{\mathrm{elec}}^{(i)}(\alpha)$ is a function of $\alpha$ which accounts for relativistic effects, spin-orbit couplings and many-body effects [^1].
According to [@Flambaum2003; @Flambaum2004], the sensitivity to g-factors $g^{(i)}$ and to the proton mass $m_p$ can be related to a sensitivity to fundamental parameters, namely the mass scale of QCD $\Lambda_{\mathrm{QCD}}$ and the quark masses $m_{q}=(m_{u}+m_{d})/2$ and $m_s$. Therefore, any measurement of the ratio between atomic frequencies can be interpreted as testing the stability of 4 dimensionless fundamental constants: $\alpha$, $m_{q}/\Lambda_{\mathrm{QCD}}$, $m_{s}/\Lambda_{\mathrm{QCD}}$ and $m_{e}/\Lambda_{\mathrm{QCD}}$. The sensitivity to $m_{s}/\Lambda_{\mathrm{QCD}}$ is relatively weak compared to the 3 other constants. The sensitivity coefficients have now been calculated for a large number of atomic species used in atomic clocks [@Prestage1995; @Flambaum2003; @Dzuba1999a; @Dzuba1999; @Dzuba2000; @Karshenboim2000b; @Dzuba2003; @Flambaum2004; @Angstmann2004]. Reliable knowledge of these sensitivity coefficient at the 1$\%$ to $10\%$ level is required to deduce limits to a possible variation of each of these fundamental parameters by combining the results of several complementary clock comparisons.
![Measured $^{87}$Rb frequencies referenced to the $^{133}$Cs fountains over 72 months. The 1999 measurement value ($\nu_{\mathrm{Rb}}(1999)=6\,834\,682\,610.904\,333\,$Hz) is conventionally used as reference. A weighted linear fit to the data (solid line) gives $\frac{d}{dt}\ln\left(\frac{\nu_{\mathrm{Rb}}}{\nu_{\mathrm{Cs}}}\right)=(-0.5\pm
5.3)\times 10^{-16}\,\mathrm{yr}^{-1}$. MJD represent Modified Julian Dates.[]{data-label="fig:alphapoint"}](alphapoint.eps){height="8cm"}
Fig.\[fig:alphapoint\] summarizes the comparison between $^{87}$Rb and $^{133}$Cs hyperfine frequencies that have been performed using the above described fountain ensemble over a duration of 6 years. Each point on the graph summarizes the result of one to two months of measurements which include each time an evaluation of all known systematic effects [@Marion2003; @Bize1999; @Bize01]. A weighted linear fit to the data in Fig.\[fig:alphapoint\] determines how these measurements constrain a possible time variation of $ \nu_{\mathrm{Rb}}/\nu_{\mathrm{Cs}}$. We find: $$\label{eq:alpha1}
\frac{d}{dt}\ln\left(\frac{\nu_{\mathrm{Rb}}}{\nu_{\mathrm{Cs}}}\right)=(-0.5\pm
5.3)\times 10^{-16}\,\mathrm{yr}^{-1}$$ which represents a 100-fold improvement over the Hg$^+$-H hyperfine energy comparison [@Prestage1995]. This results implies the following constraint: $$\label{eq:alpha2}
\frac{d}{dt}\ln\left(\frac{g_{\mathrm{Cs}}}{g_{\mathrm{Rb}}}\alpha^{0.49}\right)=(0.5\pm5.3)\times
10^{-16}\,\mathrm{yr}^{-1}.$$ Using the calculated link between g-factors and $m_{q}$, $m_s$ and $\Lambda_{\mathrm{QCD}}$ (ref. [@Flambaum2003; @Flambaum2004]), we find the following constraint to the variation of fundamental constants: $$\label{eq:alpha3}
\frac{d}{dt}\ln\left(\alpha^{0.49}\left[m_{q}/\Lambda_{\mathrm{QCD}}
\right]^{0.174}\left[m_{s}/\Lambda_{\mathrm{QCD}}
\right]^{0.027}\right)=(0.5\pm5.3)\times 10^{-16}\,\mathrm{yr}^{-1}.$$ As pointed out in [@Damour1994; @Calmet2002; @Langacker2002], the hypothetical unification of all interactions implies that a variation of the fine-structure constant $\alpha$ should be accompanied by a variation of the strong interaction constant and of elementary particle masses. Within this framework, current estimates gives [@Damour1994; @Flambaum2004; @Calmet2002; @Langacker2002]: $$\label{eq:alpha_unified_theories}
\frac{\delta(m/\Lambda_{\mathrm{QCD}})}{(m/\Lambda_{\mathrm{QCD}})}\sim
35\times\frac{\delta\alpha}{\alpha}.$$ Within this theoretical framework, the present comparison between Rb and Cs fountains (Eq. \[eq:alpha1\]) constrains a time variation of $\alpha$ at the level of $7\times 10^{-17}$ yr$^{-1}$. In the future, improvement of $^{87}$Rb and $^{133}$Cs fountain to accuracies of few parts in $10^{16}$ and repeated comparisons over several years between these devices will improve the above result by at least one order of magnitude.
The transportable fountain FOM has similarly been used as a primary standard in the measurement of the frequency $\nu_{\mathrm{H}}$ of the hydrogen 1S-2S transition performed at Max-Planck Institut in Garching (Germany) [@Niering2000; @Fischer2004]. Two measurements performed over a 4 year period constrain fractional variations of $\nu_{\mathrm{Cs}}/\nu_{\mathrm{H}}$ at the level of $(3.2\pm
6.3)\times 10^{-15}$.yr$^{-1}$. This constrains fractional variations of $g_{\mathrm{Cs}}(m_e/m_p)\alpha^{2.83}$ at the same level [@Prestage1995; @Dzuba1999a]. Combining these results with other recent comparisons ($^{199}$Hg$^+$ optical clock versus $^{133}$Cs fountain [@Bize2003; @Udem2001], $^{171}$Yb$^+$ optical clock versus $^{133}$Cs fountain [@Stenger2001; @Peik2004]), it is possible to independently set limits on variations of $\alpha$, $g_{\mathrm{Rb}}/g_{\mathrm{Cs}}$ and $g_{\mathrm{Cs}}(m_e/m_p)$. These measurements test the stability of the electroweak interaction ($\alpha$) and of the strong interaction ($g_{\mathrm{Rb}}/g_{\mathrm{Cs}}$, $g_{\mathrm{Cs}}(m_e/m_p)$) separately [@Fischer2004; @Peik2004] and independently of any cosmological model.
The PHARAO cold atom space clock and ACES
=========================================
PHARAO, as “Projet d’Horloge Atomique par Refroidissement d’Atomes en Orbite”, has started in 1993 with the objective of performing fundamental metrology with a space cold atom clock [@Laurent1995]. The combination of laser cooling techniques [@lasercooling89] and microgravity environment indeed allows the development of space clocks with unprecedented performances.
To demonstrate the feasibility of a compact cold atom clock operating in microgravity, BNM-SYRTE and LKB with the support of CNES (the French space agency) have undertaken the construction of a clock prototype in 1994. The prototype was successfully tested in 1997 in jet plane parabolic flights [@Laurent1998]. The same year, ESA, the European Space Agency, selected the ACES proposal (Atomic Clock Ensemble in Space) [@Salomon1996]. ACES will perform fundamental physics tests by using the PHARAO cold atom clock, a H-maser (developed by the Neuchâtel Observatory) and a time and frequency transfer system MWL on a platform developed by ESA. This ensemble will fly on board the international space station in 2008-2009. The station is orbiting at a mean elevation of 400 km with a 90 mn period and an inclination angle of 51.6$^\mathrm{o}$. The planned mission duration is 18 months. During the first 6 months, the performances of the PHARAO cold atom clock in space will be established. Thanks to the microgravity environment the linewidth of the atomic resonance will be varied by two orders of magnitude (from 11 Hz to 110 mHz). The target performance is $7\times
10^{-14}\tau^{-1/2}$ for the frequency stability and $10^{-16}$ for the frequency accuracy. In the second part of the mission, the onboard clocks will be compared to a number of ground based clocks operating both in the microwave and the optical domain.
In 2001, PHARAO entered into industrial development under the management of CNES with the construction of two clock models, an engineering model for test and validation, and a flight model.
### The PHARAO instrument {#the-pharao-instrument .unnumbered}
The clock is composed of four main sub-systems as shown in figure \[fig:pharao\_setup\]. Each sub-system has been subcontracted to different manufacturers and they will be assembled at CNES Toulouse to validate the clock operation.
The laser source provides all the laser tools for cooling, launching and detection of the atoms. Two extended cavity diode lasers [@Allard04] are used as master lasers. One of them injection-locks two slave diode lasers to provide high laser power for capturing $\sim 10^{8}$ in optical molasses. The second laser is used as a repumping laser. The two master laser frequencies are stabilized by servo-loops using absorption signals through cesium cells. The other laser frequencies are synthesized by using 6 acousto-optic modulators (AOM). These AOM also control the laser beam amplitudes. The output laser beams are connected to the cesium tube through polarization maintaining optical fibers. Figure \[fig:pharao\_lasersource\] shows the PHARAO optical bench during the assembly. The total mass is 20 kg, the volume is 26 liters and the power consumption is 40 W.
The cesium tube provides the atomic source, the controlled environment for the atomic manipulation, the interrogation and detection process (figure \[fig:pharao\_lasersource\]). Its design is similar to atomic fountains except for the interrogation zone where a two zone Ramsey cavity is used. The Ramsey cavity (figure \[fig:pharao\_ramseycavity\]) has been specially developed for this application and forms a ring resonator. One coupling system feeds two symmetrical lateral waveguides which meet at the two interaction zones. The advantage of this configuration is to provide very weak phase disturbances of the internal microwave field while enabling large holes ($8\times 9$ mm) for the atom path. The flight model of the microwave cavity is currently mounted (September 2004) inside the atomic fountain FO1 to measure the end to end cavity phase shift before integration in the flight model. These measurements and numerical simulations, should enable us to determine the cavity phase shift effect with an accuracy of a few parts in $10^{17}$.
The cesium tube is designed for a vacuum of $10^{-8}$ Pa in order to minimize the cold atom losses with the background gas collisions. Three layers of magnetic shields and a servo system maintain the magnetic field instability in the interaction zone below 20 pT. Similarly, the interaction zone temperature is regulated to better than 0.2$^\mathrm{o}$C.
The microwave chain synthesizes the two radiofrequency signals for the state selection cavity and the interrogation cavity. A 100 MHz VCXO (Voltage Control Oscillator) is phase-locked to an Ultra Stable Oscillator (USO) for the short term stability and to the Space Hydrogen Maser (SHM) for the medium term. Three USOs have been space qualified for our application. We have compared these quartz oscillators to the BNM-SYRTE CSO. Their frequency stability is on the order of $7\times10^{-14}$ from 1 to 10 s integration time. The engineering model of the chain has been fully tested and the results are in agreement with the performance objectives of the space clock. A further performance verification is currently being made by using the microwave source in the FO2 fountain. All PHARAO sub-systems are driven and controlled by a computer (UGB, On Board Data Processing Unit). The UGB also manages the data flux between the clock and the ACES payload. When assembled, the clock fills a volume of about 200 l for a weight of 91 kg and an electric consumption of 114 W.
The final assembly of the engineering model of the PHARAO clock will start at the end of 2005 at CNES-Toulouse. After the clock functional and performance test are made, the flight model will be assembled and finally tested. For both models, we expect to reach $10^{-15}$ frequency accuracy in the Earth gravity environment and $10^{-16}$ in microgravity environment.
### Scientific objectives of the ACES mission {#scientific-objectives-of-the-aces-mission .unnumbered}
The objectives of PHARAO/ACES are (i) to explore and demonstrate the high performances of the cold atom space clock (ii) to achieve time and frequency transfer with stability better than $10^{-16}$ and (iii) to perform fundamental physics tests. A detailed account can be found in [@Salomon2001].
The combination of PHARAO with SHM will define an on board frequency reference having a long term stability and accuracy provided by PHARAO and a short term stability determined by SHM. The resulting fluctuations of ACES frequency reference are expected to be about 10 ps per day. The orbit of ISS will allow ground users to compare and synchronize their own clock to ACES clocks, leading to a worldwide access to the ultra stable frequency reference of ACES. The results of these comparisons at $10^{-16}$ level will provide new tests in fundamental physics such as an improved measurement of Einstein’s gravitational red-shift, a search for a possible anisotropy of the speed of light and a search for possible space-time variations of fundamental physical constants, similar to that described above in section \[sec:alpha\]. The current most precise measurement of the red-shift was made by the space mission Gravitational Probe A (GPA) with an accuracy of $7\times 10^{-5}$ [@Vessot1980]. PHARAO/ACES will improve this test by a factor 30. By allowing worldwide comparison between distant clocks, operating with different atomic species, ACES will play a major role in establishing new limits for variations of fundamental constants.
Finally, PHARAO/ACES will be a pioneering cold atom experiment in space. The PHARAO technology can be extended for the development of a new generation of high performance inertial sensors and clocks using matter wave interferometry. As for atomic clocks, such sensors may achieve extremely high sensitivity in micro-gravity environment, as pointed out in the ESA HYPER project [@HYPER2000]. These instruments could then be used for a large variety of scientific space missions such as VLBI, gravitational wave detection, and deep space navigation.
Conclusions
===========
With methods described in this paper, we expect to bring the accuracy of $^{133}$Cs fountains at 1 or 2 parts in $10^{16}$. For $^{87}$Rb, a frequency stability of $1\times 10^{-14}\tau^{-1/2}$ i.e. $3\times 10^{-17}$ at one day seems accessible, together with an excellent accuracy. Routine operation of these devices over several years will have a profound impact on ultra-precise time keeping and fundamental physics tests. To take full benefit of this performance, long distance time transfer systems must be upgraded. In particular, the ACES time and frequency transfer system will enable comparisons at the level of $10^{-16}$ per day in 2008-2009. Another route currently under study makes use of telecom optical fibers and over $100$ km distance a stability of $1\times 10^{-14}$ at 1 s and $2\times 10^{-17}$ at one 1 day has already been demonstrated [@Narbonneau2004]. Extension to larger distances is under study.
More generally, clocks operating in the optical domain rather than in the microwave domain are making rapid progress on the ground [@Ekstrom2003]. The frequency of these clocks is four to five orders of magnitude higher than the frequency of microwave standards and with an equivalent linewidth, the quality factor of the resonance exceeds that of cesium clocks by the same factor. Using laser cooled atoms or ions and ultra-stable laser sources [@Young1999], these optical clocks will likely open the $10^{-17}-10^{-18}$ stability range. Using the wide frequency comb generated by femtosecond lasers, it is now possible to connect virtually all frequency standards together throughout the microwave to ultra-violet frequency domain [@Udem2001; @Udem2002]. The attractive proposal of [@Katori2001; @Katori2003] to realize an optical lattice clock is currently receiving a great deal of interest. In this method, neutral atoms are confined in an optical lattice in the Lamb-Dicke regime. Light-shifts of the clock levels induced by the lattice beams are differentially compensated at an appropriate laser detuning. This proposal combines several interesting features such as long observation time, large number of atoms, and recoil-free resonance [@Takamoto2003]. Promising atoms to implement this method are alkaline-earth atoms because of their strongly forbidden inter-combination line. Ca [@Udem2001; @Stenger2004], Sr [@Takamoto2003; @Courtillot2003] and Yb [@Kuwamoto1999; @Park2003; @Porsev2004] are actively studied.
In the frequency stability range of $10^{-17}-10^{-18}$, it is clear that fluctuations of the Earth potential at the clock location induced, for instance, by sea tides will affect comparisons between distant clocks. This limitation could be turned into an advantage if one installs such ultra-stable clock in space where the gravitational potential can present far reduced fluctuations compared to ground. As in the past, clocks with very high stability will have an ever increasing impact on scientific and industrial applications.
Acknowledgements
================
BNM-SYRTE and Laboratoire Kastler Brossel are Unités Associées au CNRS, UMR 8630 and 8552. This work was supported in part by BNM, CNRS, CNES and ESA. P. Wolf is on leave from Bureau International des Poids et Mesures, Pavillon de Breteuil, 92312 Sèvres Cedex, France. J. Grünert and L. Cacciapuoti acknowledge financial support from the CAUAC European Research Training Network.
[88.]{}
See, for instance, the 1997 Nobel lectures by Steven Chu, C. Cohen-Tannoudji and W. Phillips in Reviews of Modern Physics 70, 685 (1998)
See for instance, C. Will, Theory and experiment in gravitational physics, Cambridge University Press, (1993). See for instance Proceedings of the 6$^{\mathrm{th}}$ Symposium on Frequency Standards and Metrology, Gill P. ed., World Scientific, Singapore, 2001. H. Margolis, G. Barwood, G. Huang, H. Klein, S. Lea, K. Szymaniec, P. Gill, Science [**306**]{}, 1355 (2004)
Clairon A. et al., A cesium fountain frequency standard: recent results, IEEE Trans. Instrum. Meas. 44 (1995) 128. Bordé C.J., Atomic Clocks and Inertial Sensors, Metrologia 39 (2002) 435. Sortais Y. et al., Cold Collision Frequency Shifts in a $^{87}$Rb Fountain, Phys. Rev. Lett. 85 (2000) 3117. Fertig C., Gibble K., Measurement and Cancellation of the Cold Collision Frequency Shift in an $^{87}$Rb Fountain Clock, Phys. Rev. Lett. 85 (2000) 1622. Laurent P. et al., A cold atom clock in absence of gravity, Eur. Phys. J. D 3 (1998) 201. Mann A.G., Chang S. and Luiten A.N., Cryogenic Sapphire Oscillator with Exceptionally High Frequency Stability, IEEE Trans. Instrum. Meas. 50 (2001) 519. Vian C. et al., BNM-SYRTE Fountains: Recent Results, submitted to IEEE Trans. Instrum. Meas. (2004). Gibble K., Chu S., A laser cooled Cs Frequency Standard and a Measurement of the Frequency Shift due to Ultra-cold Collisions, Phys. Rev. Lett. 70 (1993) 1771. Ghezali S., Laurent P., Lea S.N., Clairon A., An Experimental Study of the Spin-Exchange Frequency Shift in a Laser Cooled Cesium Fountain Standard, Europhys. Lett. 36 (1996) 25. Pereira Dos Santos F. et al., Controlling the Cold Collision Shift in High Precision Atomic Interferometry, Phys. Rev. Lett. 89 (2002) 233004. Bize S. et al., Cavity Frequency Pulling in Cold Atom Fountains, IEEE Trans. Instrum. Meas. 50 (2001) 503. Marion H. et al., First Observation of Feshbach Resonances at Very Low Magnetic Field in a $^{133}$Cs Fountain, Proc. of the 2004 EFTF (2004) arxiv:physics/0407064. Schröder R., Hübner U., Griebsch D., Design and realization of the microwave cavity in the PTB caesium atomic fountain clock CSF1, IEEE Trans. Ultrason. Ferroelect. Freq. Contr. 49 (2002) 383. Parker T.E. et al., First Comparison of Remote Cesium Fountains, Proc. of the 2001 IEEE Intl. Freq. Cont. Symp. (2001) 63. Richard J.-Y. et al., Comparison of remote cesium fountains using GPS P3 and TWSTFT links, Proc. of the 2004 EFTF (2004). Damour T., Dyson F., The Oklo bound on the time variation of the fine-structure constant revisited, Nucl. Phys. B 480 (1996) 37. Fujii Y., Time-variability of the fine-structure constant expected from the Oklo constraint and the QSO absorption lines, Phys. Lett. B 573 (2003) 39. Webb J.K. et al., Further Evidence for Cosmological Evolution of the Fine Structure Constant, Phys. Rev. Lett. 87 (2001) 091301. Srianand R., Chand H., Petitjean P., Aracil B., Limits on the Time Variation of the Electromagnetic Fine-Structure Constant in the Low Energy Limit from Absorption Lines in the Spectra of Distant Quasars, Phys. Rev. Lett. 92 (2004) 121302. Marciano W.J., Time Variation of the Fundamental “Constants” and Kaluza-Klein Theories, Phys. Rev. Lett. 52 (1984) 489. Damour T., Polyakov A., The String Dilaton and a Least Coupling Principle, Nucl. Phys. B 423 (1994) 532. Damour T., Piazza F., Veneziano G., Runaway Dilaton and Equivalence Principle Violations, Phys. Rev. Lett. 89 (2002) 081601. Marion H. et al., Search for Variations of Fundamental Constants using Atomic Fountain Clocks, Phys. Rev. Lett. 90 (2003) 150801. Prestage J.D., Tjoelker R.L., Maleki L., Atomic Clocks and Variations of the Fine Structure Constant, Phys. Rev. Lett. 74 (1995) 3511. Berkeland D.J. et al., Laser-Cooled Mercury Ion Frequency Standard, Phys. Rev. Lett. 80 (1998) 2089. Niering M. et al., Measurement of the Hydrogen 1S-2S Transition Frequency by Phase Coherent Comparison with a Microwave Cesium Fountain Clock, Phys. Rev. Lett. 84 (2000) 5496. Helmcke J. et al., Optical Frequency Standard Based on Cold Ca Atoms, IEEE Trans. Instrum. Meas. 52 (2003) 250. Bize S. et al., Testing the Stability of Fundamental Constants with the $^{199}$Hg$^+$ Single-Ion Optical Clock, Phys. Rev. Lett. 90 (2003) 150802. Stenger J. et al., Absolute Frequency Measurement of the 435.5 nm $^{171}$Yb$^+$ Clock Transition with a Kerr-lens Mode-locked Femtosecond Laser, Opt. Lett. 26 (2001) 1589. Peik E. et al., in Proc. of the Joint Mtg. IEEE Intl. Freq. Cont. Symp. and EFTF Conf. (2003). Flambaum V.V., (2003) arxiv:physics/0302015. Flambaum V.V., Leinweber D.B., Thomas A.W., Young R.D., Limits on the Temporal Variation of the Fine Structure Constant, Quark Masses and Strong Interaction from Quasar Absorption Spectra and Atomic Clock Experiments, Phys. Rev. D 69 (2004) 115006. Dzuba V.A. , Flambaum V.V.,Webb J.K., Calculations of the Relativistic Effects in Many-electron Atoms and Space-time Variation of Fundamental Constants, Phys. Rev. A 59 (1999) 230. Dzuba V.A. , Flambaum V.V., Webb J.K., Space-Time Variation of Physical Constants and Relativistic Corrections in Atoms, Phys. Rev. Lett. 82 (1999) 888. Dzuba V.A., Flambaum V.V., Webb J.K., Atomic Optical Clocks and Search for Variation of the Fine-Structure Constant, Phys. Rev. A 61 (2000) 034502. Karshenboim S.G. , Some Possibilities for Laboratory Searches for Variations of Fundamental Constants, Can. J. Phys. 78 (2000) 639. Dzuba V.A., Flambaum V.V., Marchenko M.V., Relativistic Effects in Sr, Dy, YbII and YbIII and Search for Variation of the Fine Structure Constant, Phys. Rev. A 68 (2003) 022506. Angstmann E.J., Dzuba V.A., Flambaum V.V., Relativistic Effects in Two Valence-electron Atoms and Ions and the Search for Variation of the Fine-structure Constant, Phys. Rev. A 70 (2004) 014102. Bize S. et al., High-accuracy Measurement of the $^{87}$Rb Ground-tate Hyperfine Splitting in an Atomic Fountain, Europhys. Lett. 45 (1999) 558. Bize S. et al., in Proc. of the 6$^{\mathrm{th}}$ Symposium on Frequency Standards and Metrology, Gill P. ed., World Scientific, Singapore (2001) 53. Calmet X., Fritzsch H., The Cosmological Evolution of the Nucleon Mass and the Electroweak Coupling Constants, Eur. Phys. J. C 24 (2002) 639. Langacker P., Segre G., Strassler M.J., Implications of Gauge Unification for Time Variation of the Fine Structure Constant, Phys. Lett. B 528 (2002) 121. Fischer M. et al., New Limits on the Drift of Fundamental Constants from Laboratory Measurements, Phys. Rev. Lett. 92 (2004) 230802. Udem Th. et al., Absolute Frequency Measurements of Hg$^{+}$ and Ca Optical Clock Transitions with a Femtosecond Laser, Phys. Rev. Lett. 86 (2001) 4996. Peik E. et al., New Limit on the Present Temporal Variation of the Fine Structure Constant, (2004) arxiv:physics/0402132. Laurent P. et al., Cesium fountains and micro-gravity clocks, in Proc. of the 25$^{\mathrm{th}}$ Moriond Conf. on Dark Matter in Cosmology, Clocks and Tests of Fundamental Laws (1995). See for instance J. Opt. Soc. Am. B (special issue) 6 (1989) 2020. Salomon C., Veillet C., ACES: Atomic Clock Ensemble in Space, Proc. of the 1$^{\mathrm{st}}$ ESA symposium on Space Station Utilization, SP385 (1996) 295 . Allard F., Maksimovic I., Abgrall M., Laurent P., Automatic system to control the operation of an extended cavity diode laser, Rev. Sci. Instrum. 75 (2004) 54. Salomon C. et al., Cold Atoms in space and atomic clocks : ACES, C.R. Acad. Sci. Paris, T2 Série 4 (2001) 1313. Vessot R.F.C. et al., Tests of Relativistic Gravitation with a Space-borne Hydrogen Maser, Phys. Rev. Lett. 45 (1980) 2081. HYPER: Hyper-Precision Cold Atom Interferometry in Space, ESA-SCI (2000) 10. Narbonneau F. et al., Proc. of the 2004 EFTF conf. (2004). See for instance Proc. of the 2003 IFCS-EFTF conf. (2003). Young B.C., Cruz F.C., Itano W.M., Bergquist J.C., Visible Lasers with Subhertz Linewidths, Phys. Rev. Lett. 82 (1999) 3799. Udem Th., Holzwarth R., Hänsch T.W., Optical frequency metrology, Nature 416 (2002) 233. Katori H., Spectroscopy of strontium atoms in the Lamb-Dicke confinement, Proc. of the 6$^{th}$ Symposium on Frequency Standards and Metrology, Gill P. ed., World scientific, Singapore (2001) 323. H. Katori, M. Takamoto, V.G. Pal’chikov, V.D. Ovsiannikov, Ultrastable optical clock with neutral atoms in an Engineered ligth shift Trap, Phys. Rev. Lett. 91 (2003) 173005. Takamoto M., Katori H., Spectroscopy of the $^{1}$S$_0$-$^{3}$P$_0$ clock transition of $^{87}$Sr in an optical lattice, Phys. Rev. Lett. 91 (2003) 223001. Stenger J. et al., Phase-coherent frequency measurement of the Ca intercombination line at 657 nm with a Kerr-lens mode-locked femtosecond laser, Phys. Rev. A 63 (2001) 021802. Courtillot I. et al., Clock transition for a future optical frequency standard with trapped atoms, Phys. Rev. A 68 (2003) 030501. Kuwamoto T., Honda K., Takahashi Y., Yabuzaki T., Magneto-optical trapping of Yb atoms using an intercombination transition, Phys. Rev. A 60 (1999) R745. Park C.Y., Yoon T. H., Efficient magneto-optical trapping of Yb atoms with a violet laser diode, Phys. Rev. A 68 (2003) 055401. Porsev S.G., Derevianko A., Fortson E.N., Possibility of an optical clock using the 6 $^{1}$S$_0$-6 $^{3}$P$_0$ transition in $^{171,173}$Yb atoms held in an optical lattice, Phys. Rev. A 69 (2004) 021403.
[^1]: It should be noted that in general the energy of an electronic transition has in fact a contribution from the hyperfine interaction. However, this contribution is a small fraction of the total transition energy and thus carries no significant sensitivity to a variation of fundamental constants. The same applies to higher order terms in the expression of the hyperfine energy (\[eq:hyperfine\_energy\]). A precision of 1 to 10 $\%$ on the sensitivity is sufficient to interpret current experiments.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This review focuses on the X-ray emission processes of extra-galactic jets on scales resolvable by the sub arcsec resolution of the Chandra X-ray Observatory. It is divided into 4 parts. The introductory chapter reviews the classical problems for jets, as well as those associated directly with the X-ray emission. Throughout this section, we deal with the dualisms of low powered radio sources versus high powered radio galaxies and quasars; synchrotron models versus inverse Compton models; and the distinction between the relativistic plasma responsible for the received radiation and the medium responsible for the transport of energy down the jet. The second part collects the observational and inferred parameters for the currently detected X-ray jets and attempts to put their relative sizes and luminosities in perspective. In part 3, we first give the relevant radio and optical jet characteristics, and then examine the details of the X-ray data and how they can be related to various jet attributes. The last section is devoted to a critique of the two non-thermal emission processes and to prospects for progress in our understanding of jets.'
author:
- 'D. E. Harris Henric Krawczynski'
bibliography:
- 'master.bib'
title: 'X-RAY EMISSION FROM EXTRAGALACTIC JETS'
---
epsf.tex epsf.def psfig.sty
relativistic jets, X-ray jets, synchrotron emission, inverse Compton emission
THE PROBLEMS {#sec:intro}
============
Jets are giant collimated plasma outflows associated with some types of Active Galactic Nuclei (AGN). The first jet was discovered in 1918 within the elliptical galaxy M87 in the Virgo cluster: “A curious straight ray lies in a gap in the nebulosity in p.a. 20$^{\circ}$, apparently connected with the nucleus by a thin line of matter. The ray is brightest at its inner end, which is 11$^{\prime\prime}$ from the nucleus.” [@curt18]. At that time, the extended feature was a mere curiosity and its nature was not understood. When radio telescopes with good angular resolution and high sensitivities became available in the sixties, it was found that many galaxies exhibited extended radio emission consisting of a nuclear component, jets, hotspot complexes, and radio lobes. According to the standard picture, jets originate in the vicinity of a super-massive black hole (’SMBH’ with several million to several billion solar masses) located at the center of the AGN; (c.f. the early ideas of @salp64). The jets are most likely powered by these black holes, and the jets themselves transport energy, momentum, and angular momentum over vast distances [@rees71; @blan74; @sche74], from the “tiny” black hole of radius $r\,=$ $10^{-4}$ $M_{\rm BH}/10^9 M_\odot$ pc to radio hotspots, hotspot complexes and lobes which may be a Mpc or more away. Thus the study of jets must address a range of scales covering a factor of 10$^{10}$!
Even now, after thirty years of intensive studies of radio galaxies in the radio regime, no consensus has emerged on their fundamental attributes such as composition, formation, and collimation. With the advent of the Hubble Space Telescope (HST) and the Chandra X-ray Observatory (CXO), the optical and X-ray emission from jets can be studied and new tests can be evaluated which were not possible based on radio data alone. This follows because the radio, optical, and X-ray jet emissions are emitted by electrons with quite different energies (i.e. Lorentz factors, $\gamma$).
This review is focused on what X-ray observations of relativistic jets can contribute to our understanding of the physical processes in jets. Although some jet detections were made with the imaging X-ray observatories [*Einstein*]{} and [*ROSAT*]{}, significant progress blossomed only with the CXO [@weis03] launch in 1999. For this reason, together with the limitations of space, we emphasize results obtained between the years 2000 and 2005.5. We will concentrate on spatially resolved X-ray emission from the kpc-scale jets. Radio observations of the pc-scale jets and broadband observations of the spatially unresolved but highly variable core emission from the sub-pc jets of blazar-type AGN will only be discussed when they have direct implications for the inner workings of kpc-jets. Furthermore, we will not cover Galactic X-ray jets even though they bear many similarities to their extragalactic counterparts. Although there have been reports of thermal X-ray emission associated with jets (mainly in the context of ’jet-cloud interactions’), our main concern is with the non-thermal emissions, already well established as the major process for radio through X-ray frequencies from multiple lines of argument including polarization data, Faraday screen parameters, X-ray spectral fitting, and the absence of emission lines.
Reviews on some aspects of jets include “Theory of extragalactic radio sources“ [@bege84], “Beams and Jets in Astrophysics” [@hugh91], “Parsec-Scale Jets in Extragalactic Radio Sources” [@zens97], and “Relativistic jets in AGNs” [@tave04b]. Among the many jet related meetings in the last ten years are: “Relativistic jets in AGNs” Cracow, 1997, [@ostr97]; “Ringberg Workshop on Relativistic Jets” Ringberg Castle, 2001[^1]; “The Physics of Relativistic Jets in the CHANDRA and XMM Era” Bologna, 2002, [@brun03]; “Triggering Relativistic Jets” Cozumel, 2005 [@lee06]; and “Ultra-Relativistic Jets in Astrophysics: Observations, Theory and Simulations” Banff, 2005[^2]. Conference reviews on the unresolved core emission have been given by @siko01 [@copp99; @kraw04; @kraw05] and @tave05.
We use the conventional definition of spectral index, $\alpha$, for power-law radiation spectra: flux density, S$_{\nu}~\propto~\nu^{-\alpha}$. It is not yet known if electrons alone, or electrons and positrons radiate the observed jet emission. We thus refer in the following to either electrons or electrons and positrons as “electrons”. We use $\gamma$ as the Lorentz factor of particles in the jet-frame of reference, and $\Gamma$ for the bulk Lorentz factor of the jet plasma. As most X-ray emitting jets are detected on only one side of otherwise double radio sources, $\Gamma~\geq$ a few seems likely to be generally applicable.
Jet Composition {#sec:1fluid}
---------------
We take the essence of a jet to be a quasi-lossless transmission line: a conduit containing relativistically moving particles and magnetic field (either of which could dominate the local energy) and/or Poynting flux. We distinguish between two substances: the “medium” which is responsible for delivering the power generated in the nucleus of the host galaxy to the end of the jet and thence to the radio lobes; and the non-thermal plasma responsible for the emission we detect in the radio, optical, and X-ray bands. While these two substances can be one and the same for some jet models, we prefer to think of them as quite distinct. Most models explain the appearance of radio, optical or X-ray bright hotspots in some jets as caused by the transfer of some form of energy (for example, energy associated with the medium’s bulk motion or magnetic field energy), to highly relativistic emitting particles. The reader should note that we use the term “medium” lacking more precise knowledge about the nature of the jet material.
While the basic make-up of jets is still largely unknown, observations of polarized radio and optical emission show that at least some of the continuum jet emission originates as synchrotron emission from relativistic electrons gyrating in a magnetic field. Although we have this direct evidence about the emitting plasma, the jet medium responsible for delivering power to the end of the jets is largely unconstrained. The jet medium cannot entirely consist of the relativistic electrons that produce the observed radiation since unavoidable inverse Compton (IC) losses off the cosmic microwave background (CMB) photons would preclude the flow of high energy electrons all the way to the end of some jets. Positing a minimal magnetic field strength of 3$\mu$G, and ignoring the IC losses associated with starlight or quasar light which would shorten the relevant lifetimes even more, it has been shown that electrons with $\gamma~\geq$ a few thousand cannot survive for the time required to travel from the environs of the SMBH to the end of some jets [e.g. @harr06b].
The main contenders for the underlying jet medium are Poynting flux, electrons with $\gamma~\leq$ 1000, and protons. In addition, ’neutral beams’ have been suggested [e.g. neutrons, @atoy04]. The latter hypothesis requires that the direction into which the jet is launched changes with time to account for large scale bending and discrete deflections such as those in 3C 120 and 3C 390.3. Real jets may be made of several components, or may involve the transition of a jet dominated by one component into a jet dominated by another; e.g. a class of models postulates that an initially electromagnetic jet transforms into a particle dominated jet further downstream.
Jet formation, structure and propagation {#sec:1genesis}
----------------------------------------
### Jet formation {#jet-formation .unnumbered}
Jets are believed to be launched from accreting supermassive black holes and powered by either the gravitational energy of accreting matter that moves toward the black hole or, in the Blandford-Znajek process [@blan77], by the rotational energy of a rotating black hole. In the first case, jets may either be launched purely electromagnetically [@blan76; @love76], or as the result of magnetohydrodynamic processes at the inner regions of the accretion disk [@blan82; @bege84; @koid99]. In the Blandford-Znajek process, the black hole rotating in the magnetic field supported by the accretion disk gives rise to a Poynting flux. Most models of jet formation face the $\sigma$-problem ($\sigma$ is the ratio of electromagnetic energy density to particle energy density), namely that they predict a Poynting flux dominated energy transport by a strongly magnetized or high-$\sigma$ plasma, while pc scale observations indicate that the jets consist of particle dominated, low-$\sigma$ plasma [@celo93; @kraw02; @kino02]. Understanding the launching of jets may thus require the solution of two problems: the launching of a magnetically dominated outflow, and the conversion of such an outflow into a particle dominated jet. The latter transition is poorly understood, and requires more theoretical work.
The process of jet formation will have an impact on the steadiness of the jet-flow, and will affect the amplitudes and time scales of jet luminosity variations. Modulations of the power output are believed to cause the large amplitude brightness variations of the (unresolved) X-ray and $\gamma$-ray emission from blazars [@spad01; @tani03]. Large amplitude variations on time scales of thousands of years may be responsible for the radio, optical and X-ray knots observed in many kpc-scale jets [@staw04a; @staw04b], and the bright X-ray flare of the M87 jet [@harr06a]. Several recent studies have shown that the flaring activity of AGN can be described in the language of noise processes [@uttl05]; i.e. the study of power spectra. Blazar flares show that the noise process that drives flares has a rising amplitude of the power spectrum on the relatively long time scales of a few years. If jet knots reflect nuclear variability, it would require substantial power at much longer time scales (red noise) which of course are not available for direct observation.
### Transverse jet structure {#transverse-jet-structure .unnumbered}
In addition to the obvious uncertainties as to the identity of the jet medium and its bulk velocity, several jet models involve jet structure perpendicular to the jet axis. Radio observations of transversely resolved jets [e.g. @swai98; @lain04; @lara04; @push05] and theoretical models of the core emission of blazars [@chia00] indicate a velocity gradient across the jet. Simple models use a two-zone structure, a fast moving spine that carries most of the jet energy, surrounded by a slower sheath, each with a characteristic value of $\Gamma$ [@chia00]. @lain04 assume a gradual decline of $\Gamma$ from the jet center to the outer parts of the jet: i.e. many layers with different velocities. If the velocity difference between layers is large, the particles in some layers see the relativistically boosted photons from other layers, resulting in an increase of the IC emission [@ghis05]. A wealth of different jet structures has been proposed and studied in the framework of explaining the prompt and afterglow emission from Gamma-Ray Bursts [e.g. @graz06] and some of these may be relevant to kpc scale jets.
The fact that jets may have a complex structure is important for interpreting the observational data. For example, the dominance of the bright jet over the dim counter-jet in a number of sources was previously thought to constrain the bulk Lorentz factor of the jets [e.g. @ward97]. However, the observations may merely show that most of the radio emission comes from a slow moving plasma, and thus may not constrain the bulk Lorentz factor of the jet component that carries most of the jet energy and momentum. The boundaries between jet layers of different velocity may accelerate particles [@staw02] and are of special interest for jet stability considerations.
### Jet propagation and the occurrence of knots {#jet-propagation-and-the-occurrence-of-knots .unnumbered}
The origin of jet knots (localized brightness enhancements) and the mechanism that controls the location, strength, and longevity of the shocks thought to be responsible for the existence of knots, have not yet been identified unambiguously. It is important to remember however, that there is probably more than one type of knot and that there are several suggested methods of producing brightness enhancements in addition to the conventional explanation of particle acceleration at shocks. In the case of the M87 jet (fig. \[fig:m87\]), the inner knots, D, E, and F appear quasi regular in size and spacing suggesting a possible origin associated with standing waves similar to those described by @bere03, or by the elliptical mode Kelvin-Helmholtz instability [@loba03]. Quite different are the knots A and C for which steep, quasi-planar gradients in radio brightness suggest reverse and forward shocks (see comment on shocks in the sidebar). Note however that @bick96 have devised a detailed model of the M87 jet. They argue that all the knots can be explained by oblique shocks, with the apparent differences being ascribed to relativistic effects. Their model requires the angle between the jet axis and the line of sight to be 30$^{\circ}$ to 35$^{\circ}$, a value substantially larger than the 10$^{\circ}$ to 20$^{\circ}$ required by the observation of fast moving blobs downstream from the leading edge of the knot HST-1 (see sec. \[sec:3kpc\]).
One of the alternative explanation of knots is that knots in relativistic jets could be manifestations of a change in the beaming factor. The relativistic beaming factor, $\delta$ depends both on $\Gamma$ and on the viewing angle, $\theta$ (the angle between the jet axis and the line of sight in the observer’s frame):
$$\delta^{-1}~=~\Gamma(1~-~\beta~cos~\theta).
\label{eq:delta1}$$
If the jet medium moves in a straight line so that $\theta$ is fixed, an increase in $\delta$ requires a significant increase in $\Gamma$. While we can imagine plausible ways to lower $\Gamma$, the critical question is, are there ways to increase $\Gamma$ far from the central engine? This would entail a supply of energy such that the total power flow could decrease yet $\Gamma$ could increase (e.g. by converting some power from the flow as in magnetic reconnection). @siko05 discuss this scenario, but deal only with the situation close to the black hole. While there is circumstantial evidence for acceleration of jet features on pc-scales [e.g. @hardee05] and it is generally accepted that both FRI [@lain02b] and quasar [@ward97] jets decelerate on pc to kpc scales there is no indication that significant jet acceleration occurs on kpc scales which may be required for some IC models of X-ray emission.
If the jet medium is allowed to significantly change its direction, modest changes in $\theta$ can produce large changes in $\delta$. On VLBI scales, there has been a long standing debate on ballistic vs. curved trajectories. On the kpc scale the question arises: does the medium move in a straight or gently curved path, or might it follow a helical pattern controlled by a field structure of the same topology? If the latter case holds, the changes in brightness along the jet could be explained by beaming effects and some of the problems for high $\Gamma$ jet models, such as excessive jet length, would be mitigated. @bahc95 remark on the apparent helical morphology of the HST image of 3C 273 (fig. \[fig:hst273\]) and @naka01 argue for a ’torsional Alvèn wave train’ moving out to large distances from the central engine as a method of controlling the large scale structure. There are of course numerous examples of large scale bending [e.g. 3C 120: @walk87] and discrete deflections [e.g. 3C 390.3 @harr99]; but in these cases we would anticipate deceleration only.
\*
### Terminal Hotspots {#sec:1hotspots .unnumbered}
Terminal hotspots, like knots, are thought to be localized volumes of high emissivity which are produced by strong shocks or a system of shocks. The somewhat hazy distinction between hotspots and knots is that downstream from a knot, the jet usually propagates much as before, whereas at the terminal hotspot, the jet itself terminates and the remaining flow is thought to create the radio lobes or tails. Thus the underlying jet medium must suffer severe deceleration and the outward flow from the hotspot is non-relativistic and is not confined to a small angle. This is patently not true for the so called ’primary’ hotspots in double or multiple systems. Instead of a terminal shock, primaries (and also aberrations such as hotspot B in 3C 390.3 North) may have oblique reflectors in essence, although the actual mechanism for bending might be more akin to refraction. For an extensive discussion of the differences between knots and hotspots, see @brid94.
Knots are a common property of FRI jets and generally do not lead to a total disruption of the jet which maintains its identity downstream, be it relativistic or not. What we call knots in quasar jets, may have little in common with FRI knots given their relative physical sizes (fig. \[fig:size\]).
Insofar as the X-ray emission mechanism is concerned, the initial X-ray detection of the Cygnus A hotspots [@harr94] was accompanied with a demonstration that synchrotron self Compton (SSC) emission provided a consistent explanation if the average magnetic field strength was close to the equipartition value under the assumption that the relativistic particle energy density was dominated by electrons, not protons. Essentially all the emission models for jet knots, on the other hand, have shown that SSC emission is completely inadequate to explain X-ray emission unless the magnetic field is orders of magnitude less than the equipartition value.
As the number of hotspot detections increased from CXO observations, many were found to be consistent with SSC predictions but a significant number appeared to have a larger X-ray intensity than predicted. This excess could be attributed to a field strength well below equipartition, IC emission from the decelerating jet ’seeing’ Doppler boosted hotspot emission [@geor03], or an additional synchrotron component [@hard04b]. The last named authors show that the strength of the excess correlates with hotspot luminosity in the sense that the strongest hotspots are consistent with SSC emission whereas the weaker radio hotspots required the extra synchrotron component.
For many distant and/or faint jets, it is often difficult to be certain that a feature is a knot, a hotspot, or even a lobe. In some extreme cases, the true nature of even bright hotspots is ambiguous. An example is the double hotspot system in 3C351 shown in fig. \[fig:351\]. Displaced from the north radio lobe is a double hotspot to the NE of the core. These are bright at radio and X-ray bands. The southern radio lobe has only a weak hotspot with at most 4% of the radio intensity of the NE hotspots at 1.4 GHz. Thus the double hotspot has the hallmarks of relativistic beaming in spite of the commonly held view that hotspot radiation is not beamed [see however @denn97 for a discussion of beamed emission from hotspots]. Given the fact that these bright features are not located at the outer edge of the lobe, perhaps they are knots in a jet very close to our line of sight.
\*
Entrainment and Collimation {#sec:1entrain}
---------------------------
Long-standing problems for low-loss jets include the suppression of mixing with ambient material and the collimation and stability of jets [@hugh91]. The process of entrainment of ambient material is closely related to the process of jet deceleration. Both processes have been studied observationally [e.g. @lain03] and numerically [e.g. @ross04]. Possible mechanisms causing entrainment include velocity shear and Kelvin-Helmholtz instabilities [@bodo03]. @lain03 have studied FRI radio galaxies assuming that the two sides of the jets are intrinsically identical and that the observed differences in radio brightness and polarization are caused by the viewing angle and relativistic beaming effects. They find the velocity of the jet plasma decreases moving away from the jet axis and that this velocity shear decelerates the jet substantially. These arguments purport to demonstrate that there is a clear distinction between FRI and FRII radio galaxies insofar as their jet properties are concerned. Since the powerful jets of FRII radio galaxies and quasars are able to escape the high ambient density of their host galaxies and maintain their collimation out to the prominent hotspots, it is inferred that they suffer less entrainment and deceleration than FRI jets [@bick95].
Collimation of jets has to be addressed both on sub-parsec scale during the process of launching the jet and on kpc-scales to explain the remarkable stability of jets. @tsin02 for example consider the former problem and demonstrate collimation for a relativistic component by a second, non-relativistic less-collimated outflow (wind). Other collimation mechanisms include confinement by magnetic fields [@saut02], ram pressure of the ambient medium [@komi94], and by radiation [@fuku01].
Particle Acceleration and Emission Mechanisms {#sec:1radi}
---------------------------------------------
The Chandra X-ray observatory increased the number of jets with X-ray emission from a handful to $\approx$50 sources. Of these, 60% are classified as high-luminosity sources (quasars and FRII radio galaxies) and the remaining are low-luminosity sources (a mix of FRI’s, BL Lac’s and a Seyfert galaxy). The observations indicate that the radio to X-ray emission from low-luminosity FRI sources can be explained by synchrotron models, while that from the high-luminosity FRII sources requires multi-zone synchrotron models, synchrotron and IC models, or more exotic variants.
### Synchrotron models for FRI galaxies {#synchrotron-models-for-fri-galaxies .unnumbered}
For low-luminosity (FRI) radio sources, there is strong support for the synchrotron process as the dominant emission mechanism for the X-rays, optical, and of course radio emissions. Among the arguments supporting this view are the intensity variability found for knots in the M87 jet [@harr06a]; the fact that in most cases the X-ray spectral index, $\alpha_x$ is $>$ 1 and significantly larger than the radio index, $\alpha_r$; and the relative morphologies in radio, optical, and X-rays. For the sorts of magnetic field strengths generally ascribed to jet knots (10 to 1000 $\mu$G), synchrotron X-ray emission requires the presence of electrons of energies in the range 10$^7~<~\gamma~<~10^8$. As the highest energy electrons cool in equipartition magnetic fields on time scales of years, the observations of single power-law spectral energy distributions extending all the way from the radio to the X-ray regime pose the problem of why there is no sign of radiative cooling. A possible solution may be that electrons escape the high-magnetic field emission region before they cool.
The radio to X-ray observations require the presence of one or more populations of high energy electrons (or protons if proton synchrotron emission is viable, [@ahar02]). A common assumption is that the particles are accelerated at strong magnetohydrodynamic (MHD) shocks by the Fermi I mechanism [@bell78; @blan78][see also the review by @kirk99]. However, there are several uncertainties. First we cannot be sure that the Fermi process is relevant since if the jet is strongly magnetized with a tangled field geometry, shock acceleration is not as effective as for strong shocks which can exist when the field does not dominate. Next, the uncertainty of the bulk Lorentz factor of the jet medium means that we can’t be sure that $\Gamma$ is large enough to allow the possibility of relativistic shocks. Finally, even if the bulk velocity of the jet is relativistic, it is still possible to have non-relativistic shocks in the jet frame. For mildly relativistic shocks, Fermi I shock acceleration is more complicated than is the case for the non-relativistic regime, and it is not yet well understood [@kirk99]).
#### Distributed Acceleration {#sec:1distrib}
For FRI jets such as that in M87 (fig. \[fig:m87\]), X-ray emitting electrons with $\gamma\approx~10^7$ will cool on time scales of a few years, and optical and UV emitting electrons will cool on time scales of a few decades. Thus, interpreting the X-ray and optical emission from these sources as synchrotron emission implies that the emitting regions cannot be much larger than the electron acceleration regions. For bright knots which have traditionally been associated with strong shocks in the jet flow, these “life-time constraints” can easily be accommodated. However, Chandra detected several jets with quasi-continuous emission along the jet [e.g. Cen A, @kata06] suggesting that electron acceleration may be spatially distributed rather than being restricted to a few bright knots. @wang02 finds that plasma turbulent waves can be a mechanism for efficient particle acceleration, producing high energy electrons in the context of blazar jets. @nish05 [@staw02] propose turbulent acceleration in a jet’s ’boundary’ or ’shear’ layer surrounding the jet spine. @staw02 also argue that the resulting electron energy distribution should show an excess near the high energy cutoff, thereby producing a harder X-ray spectrum than would be expected based on the extrapolation of the radio and optical data.
Other explanations for the quasi-continuous emission include low-level IC/CMB emission of low-energy electrons (see the discussion in the next paragraph) and synchrotron emission from electrons accelerated by magnetic reconnection. However, if the knot emission is produced by synchrotron emission from shock-accelerated electrons and the continuous emission has another origin, one might expect that the two jet regions would show markedly different spectral energy distributions. Measurement of X-ray spectral indices of the continuous emission is usually difficult, because of the fewer photons available for analysis. In the case of M87, @perl03 find no change of $\alpha_x$ between the knots and the quasi-continuous emission within a statistical accuracy of $\pm$0.15 in the X-ray spectral index.
#### Departures from power law spectra {#sec:1bowtie}
One of the primary reasons that IC/CMB models are preferred over synchrotron models for most FRII radio galaxies and quasars is the so-called ‘bow-tie problem’. Conventional synchrotron spectral energy distributions call for a concave downward spectral shape, allowing for spectral breaks to steeper spectra at higher frequencies and eventual high frequency cutoffs. Thus we expect $\alpha_x~\geq~\alpha_{ox}$ ($\alpha_{ox}$ is the spectral index between optical/UV and X-ray). When this is not the case, the ’bow-tie’ showing the X-ray flux density and allowed range of $\alpha_x$ does not permit a smooth fit of a concave downward curve and instead requires a flattening of the X-ray spectrum. Examples are provided in fig. \[fig:2sed\] which shows the SED’s for 3 knots in the 3C273 jet. The bow-tie problem is more common for FRII radio galaxies and quasars but is also found for some of the FRI radio galaxies.
In addition to the hypothesis of a second spectral component, there have been two suggestions for accommodating this behavior with synchrotron emission. The first is that mentioned above: boundary layer acceleration [@staw02] producing a flatter spectrum for the high energy part of the electron spectrum. The other suggestion is restricted to the case where IC dominates the E$^2$ losses. @derm02 argue that for the highest energy electrons, IC losses are reduced by the lower Klein-Nishina cross section so that the top end of the electron distribution experiences a reduced loss rate and thus an excess above the expected distribution builds up at high energies producing a hard synchrotron spectrum at X-ray frequencies. Although this is a clever method of solving the bow-tie problem, in order to work, the photon energy density in the jet frame, u$^{\prime}$($\nu$) must be larger than the magnetic field energy density, u(B). To realize this, a $\Gamma^2$ boosting of the CMB is required, and the practical result is that by invoking the necessary $\Gamma$, you will already produce the observed X-ray emission by the IC/CMB process.
### Emission models for FRII galaxies and quasars {#emission-models-for-frii-galaxies-and-quasars .unnumbered}
The most pressing problem for X-ray emission from relativistic jets is the emission mechanism for the powerful jets from FRII radio galaxies and quasars. As mentioned above, the radio to X-ray spectral energy distributions of most of these sources cannot be described by a one-component synchrotron model. Such models predict a spectral energy distribution which softens at high energies. In terms of spectral indices, we expect $\alpha_x\geq\alpha_{ox}$), whereas the Chandra observations showed that $\alpha_x<\alpha_{ox}$ for many quasar jets. The most popular explanation is the IC model put forth by @celo01 and @tave00. The observed large ratios of X-ray to radio luminosities are explained by postulating very fast jets with high bulk Lorentz factors $\Gamma$. Relativistic boosting increases the energy density of the CMB in the jet frame: $$u'(\rm CMB)\,= 4 \times 10^{-13}~(1+z)^4~\Gamma^2~\rm erg~cm^{-3}.$$ In this way, a single population of electrons is able to produce the radio and optical synchrotron emission in a magnetic field close to equipartition (B$_{eq}$ generally less than 100 $\mu$G), and the IC X-ray emission by scattering off the relativistically boosted CMB.
While we evaluate the various difficulties confronting the IC/CMB model in the final section of this review (\[sec:4ic\]), we would like to emphasize here that the IC/CMB model requires two key-ingredients for which there is at present no independent observational verification: enough low-energy electrons and highly relativistic plasma motion on kpc-scales. Analysis of the spectral energy distributions of several FRII sources shows that electrons with Lorentz factors $\gamma'~\approx$ 100 produce the observed X-rays: [e.g. @harr02a eq. B4]:
$$\gamma' \,=\, \sqrt{\frac{6.25\times 10^{-12}\, \nu_{\rm
ic}(obs)}{(1+\mu'_{\rm j}) \delta \Gamma}}.
\label{eq:gammaic}$$
where the prime is used to denote quantities in the jet-frame, $\mu'_{\rm j}$ equals cos($\theta'$) and $\theta'$ is the angle between the jet direction and the line of sight. The uncertainties of extrapolating the electron spectrum to low energies is illustrated in fig. \[fig:extrap\] which shows the spectrum of the knot in the jet of PKS0637-752. The low-energy electrons responsible for the X-ray emission produce synchrotron emission in the 1-30 MHz range, well below frequencies available from the Earth with reasonable angular resolutions. For this example we have used $\Gamma$=10, which is the value required for the IC/CMB model [@tave00; @celo01]. The actual electron spectrum could flatten significantly for $\gamma\leq$3000 or even suffer a low energy cutoff. In that case there would be fewer electrons than calculated, and the required value of $\Gamma$ would have to be increased to compensate. It is even conceivable that the electron spectrum could steepen at low energies and the required $\Gamma$ would be much smaller than estimated. Our ignorance of the low end of the electron spectrum is very general; only in sources with very high values of magnetic field strength do ground based radio data begin to give us the required information.
On the positive side, the IC/CMB model avoids the ’far-from-equipartition’ requirement of models which explain the high X-ray fluxes as synchrotron self-Compton emission from electrons up-scattering long-wavelength synchrotron photons into the X-ray band [e.g. @schw00]. Furthermore, it does not require the ad-hoc introduction of additional particle components, required by multi-zone synchrotron models [e.g. @harr99].
Physical Comparisons of Resolved X-ray Jets {#sec:phys}
===========================================
We are used to looking at images of jets that fit nicely on the page, be they Galactic micro quasars, jets from relatively local FRI radio galaxies, or jets from quasars with substantial redshifts. We are struck by a number of similarities and are tempted to consider them all to share fundamental properties. In an effort to sharpen our perspective, we have devoted this part of the review to presenting the observed and deduced parameters for the X-ray jets known to us. Most of these data exist in the literature, but we have adjusted published values, where necessary, to conform to the currently standard cosmology: H$_0$=71 km s$^{-1}$; $\Omega_m$=0.3; and $\Omega_{\Lambda}$=0.7. Gathering these data will allow us to compare physical sizes and apparent luminosities.
In fig. \[fig:size\] we show the relative sizes of 3 jets: M87, 3C273, and PKS1127-145. Although the indicated sizes (1.6, 56, and 238 kpc, respectively) are projected sizes, it can be seen that the entire jet of M87 would easily fit within a single knot of the 3C273 jet. Note also that a single 0.049$^{\prime\prime}$ pixel in the top panel corresponds to a few pc, the scale of VLBI jets and also comparable to the total size of jets from microquasars in our galaxy. Given the vast range of scales, can we really expect similar physical processes to operate all the way from pc to Mpc scales?
Gross jet properties {#sec:2jets}
--------------------
In Table \[tab:fr1jets\] we list parameters for X-ray jets of low radio power sources from the XJET website[^3] (2005.6), and Table \[tab:fr2jets\] contains the data for the more powerful sources classified as quasars or FRII radio galaxies. We have not included sources if only the terminal hotspots and/or lobes have been detected in X-rays. In almost all cases, the division between the two tables corresponds to how the original investigators interpret the X-ray emission process. All the entries in Table \[tab:fr1jets\] are ascribed to synchrotron emission except for Cen B, and similarly the jets of Table \[tab:fr2jets\] are described on the basis of the IC/CMB model except for PKS2152-69 for which @ly05 suggested thermal emission; Pictor A, PKS 1136-135 and 3C 273 for which both synchrotron and IC/CMB have been suggested; and 1928+738 and 3C403 which have been ascribed to synchrotron emission.
The projected jet length, both in arcsec and kpc should be accurate to about 10%; it is meant to describe the length of the X-ray jet as detected by the CXO and not the total length of the radio jet. The apparent X-ray luminosity is derived from the observed flux or flux density, assuming $\alpha_x$=1. As pointed out by @list03a, such luminosities are not directly useful for correlations since we are dealing with relativistic beaming which alters the jet frame luminosity, depending on $\Gamma$ and $\theta$. The value of $\alpha_x$ given is a published value, either for the whole jet, or from a brighter knot. If a reasonable estimate of the angle of the jet to the line of sight is given in the literature, it is quoted here. For a number of the quasars, $\theta$ is estimated from the IC/CMB calculation, and is thus model dependent; for others, it is estimated from VLBI studies. The resulting deprojected length suffers from similar uncertainties.
[@lllcrllccl@]{} Host name & z & Scale & Length & Length & log L$_x$ & $\alpha_x$ & $\theta$ & Deproj. & Reference\
& & (kpc/”) & (arcsec) & (kpc) & (ergs$^{-1}$) & & (deg) & (kpc) &\
3C15\* & 0.0730 & 1.4 & 4 & 5.6 & 41.03 & 0.7$\pm$0.4 & .. & .. & 1\
NGC 315 & 0.0165 & 0.33 & 13 & 4.3 & 40.54 & 1.5$\pm$0.7 & .. & .. & 2\
3C31\* & 0.0167 & 0.34 & 8 & 2.7 & 40.56 & 1.1$\pm$0.2 & 52 & 3.4 & 3\
B2 0206+35 & 0.0369 & 0.72 & 2 & 1.4 & 41.12 & .. & .. & .. & 4\
3C 66B\* & 0.0215 & 0.43 & 7 & 3.0 & 41.03 & 1.3$\pm$0.1 & .. & .. & 5\
3C 120\* & 0.0330 & 0.65 & 80 & 52 & 41.95 & .. & .. & .. & 6\
3C 129 & 0.0208 & 0.42 & 2.5 & 1.0 & 39.64 & .. & .. & .. & 7\
PKS 0521-365\* & 0.055 & 1.06 & 2 & 2.1 & 41.90 & 1.2$\pm$0.3 & .. & .. & 8\
B2 0755+37\*& 0.0428 & 0.83 & 4 & 3.3 & 41.52 & .. & .. & .. & 4\
3C270 & 0.0074 & 0.15 & 35 & 5.2 & 39.13 & .. & .. & .. & 9, 10\
M84 & ... & 0.082 & 3.9 & 0.3 & 38.71 & 0.8$\pm$0.3 & 50 & 0.4 & 11\
M87\* & 0.0043 & 0.077 & 20 & 1.5 & 41.32 & $>$1 & 20 & 4.5 & 12, 13\
Cen A & ... & 0.017 & 120 & 2.0 & 39.39 & 0.4 to 2.2 & 15 & 7.7 & 14\
Cen B & 0.013 & 0.26 & 8 & 2.1 & 40.13 & .. & ... & ... & 15\
3C296 & 0.0237 & 0.47 & 10 & 4.7 & 40.09 & 1.0$\pm$0.4 & .. & .. & 16\
NGC6251\* & 0.0249 & 0.49 & 410 & 200 & .. & 1.30$\pm$0.14 & .. & .. & 17\
3C 346\* & 0.161 & 2.7 & 2 & 5.4 & 41.96 & 1.0$\pm$0.3 & 20 & 16 & 18\
3C 371\* & 0.051 & 0.98 & 4 & 3.9 & 41.87 & 0.7+0.4,-0.2 & 18 & 12.6 & 19\
3C 465 & 0.0293 & 0.58 & 7.5 & 4.4 & 40.30 & $\approx$ 1.4 & ... & ... & 16\
Notes\
The scale is given in units of kpc per arcsec.\
All sources are classified as FRI radio galaxies except for 3C120, a Seyfert I galaxy, and the two BL Lac objects, PKS0521-365 and 3C371.\
A “\*” after the source name indicates that an optical detection has been reported, see http://home.fnal.gov/$\sim$jester/optjets/.\
References\
1 @kata03; 2 @worr03; 3 @hard02; 4 @worr01; 5 @hard01; 6 @harr04; 7 @harr02c; 8 @birk02; 9 @zeza05; 10 @chia03; 11 @harr02b; 12 @wils02; 13 @mars02; 14 @hard03; 15 @mars05; 16 @hard05a; 17 @evan05; 18 @worr05; 19 @pesc01.
[@lllcrllccrl@]{}
Host name & z & Scale & Length & Length & log L$_x$ & $\alpha_x$ & $\theta$ & Deproj. & $\delta$ & Reference\
& & (kpc/as) & (arcsec) & (kpc) & (ergs$^{-1}$) & & (deg) & (kpc) & &\
3C9 & 2.012 & 8.5 & 6.4 & 54 & 44.34 & .. & .. & .. & .. & 1\
PKS 0208-512 & 0.999 & 8.04 & 5 & 40 & 44.47 & .. & 8 & 262 & 7 & 2,3\
PKS 0413-21 & 0.808 & 7.54 & 2 & 15 & 43.99 & .. & 20 &44 & 3 & 2\
Pictor A & 0.0350 & 0.69 & 114 & 79 & 40.84 & 0.97$\pm$0.07 & $>$23 & $<$201 & $<$3 & 4,5\
PKS 0605-085 & 0.870 & 7.7 & 4 & 31 & 44.58 & 0.4$\pm$0.7 & .. &.. & .. & 6\
PKS 0637-752\* & 0.651 & 6.9 & 12 & 83 & 44.34 & 0.85$\pm$0.08 & 5.7 & 836 & 10 & 7,8\
3C 179 & 0.846 & 7.7 & 4.4 & 34 & 44.45 & .. & .. & .. & .. & 6\
B2 0738+313 & 0.635 & 6.9 & 35 & 241 & 42.93 & 0.5 to 1.4 & 8 & 1730 & 7 & 9\
0827+243 & 0.939 & 7.9 & 6.2 & 49 & 44.14 & 0.4$\pm$0.2 & 2.5 & 1100 & 20 & 10\
3C 207 & 0.68 & 7.1 & 4.6 & 33 & 43.97 & 0.3$\pm$0.3 & 8 & 237 & 7 & 6\
3C 212\* & 1.049 & 8.1 & 4 & 32 & 43.52 & .. & .. & .. & .. & 11\
PKS 0903-57 & 0.695 & 7.1 & 3.5 & 25 & 43.90 & .. & 20 & 73 & 3 & 2\
PKS 0920-39 & 0.591 & 6.6 & 10 & 66 & 43.70 & .. & 7 & 322 & 8 & 2,3\
3C 219 & 0.174 & 2.9 & 20 & 58 & (41.68) & .. & .. & .. & .. & 12\
Q0957+561 & 1.41 & 8.5 & 8 & 68 & 43.69 & 0.9$\pm$0.6 & .. & .. & 1.4 & 13\
PKS 1030-357 & 1.455 & 8.5 & 12 & 102 & 44.99 & .. & 8.6 & 682 & 9 & 2,3\
PKS 1046-40 & 0.620 & 6.8 & 4 & 27 & 43.44 & .. & 17 & 93 & 3 & 2\
PKS 1127-145 & 1.18 & 8.3 & 30 & 249 & 44.62 & 0.5$\pm$0.2 & 24 & 612 & 4 & 14\
PKS 1136-135\*& 0.554 & 6.4 & 6.7 & 43 & 43.92 & 0.4$\pm$0.4 & 6 & 410 & 10 & 6\
4C49.22\* & 0.334 & 4.8 & 5.6 & 27 & 43.62 & .. & 6 & 270 & 14 & 6\
PKS 1202-262 & 0.789 & 7.5 & 5 & 37 & 44.73 & .. & 4.9 & 568 & 12 & 2,3\
3C 273\* & 0.1583 & 2.7 & 21 & 57 & 43.58 & 0.6 to 0.9 & 5 & 654 & 5 & 16,17\
4C19.44\* & 0.720 & 7.2 & 14.4 & 104 & 44.55 & .. & 10 & 616 & 14 & 6\
3C 303\* & 0.141 & 2.5 & 9 & 22 & 41.56 & .. & .. & .. & .. & 17\
GB 1508+5714 & 4.3 & 6.9 & 2.2 & 15 & 44.96 & 0.9$\pm$0.4 & 15 & 58 & 4 & 18,21\
PKS 1510-089 & 0.361 & 5.0 & 5.2 & 26 & 43.85 & 0.5$\pm$0.4 & .. & .. & .. & 6\
3C 345\* & 0.594 & 6.6 & 2.7 & 18 & 43.65 & 0.7$\pm$0.9 & 7 & 138 & 7 & 6\
1642+690 & 0.751 & 7.3 & 2.7 & 20 & 43.67 & .. & .. & .. & .. & 6\
3C 380\* & 0.692 & 7.1 & 1.8 & 13 & 44.68 & .. & 13 & 57 & .. & 2\
1928+738\* & 0.302 & 4.4 & 2.6 & 11 & 43.21 & 0.7$\pm$0.7 & 6 & 105 & 10 & 6\
3C403\* & 0.059 & 1.13 & 45 & 51 & 41.48 & 0.7$\pm$0.4 & .. & .. & .. & 19\
PKS 2101-490 & (1.04) & 8.1 & 6 & 49 & 44.17 & .. & 25 & 116 & .. & 2\
PKS 2152-69 & 0.0283 & 0.56 & 10 & 5.6 & 40.66 & 1.6$\pm$0.4 & .. & .. & .. & 20\
3C 454.3\* & 0.859 & 7.7 & 5.2 & 40 & 44.62 & .. & 18 & 129 & .. & 2\
Notes\
The scale is given in units of kpc per arcsec.\
All sources are classified as quasars except for the 4 FRII radio galaxies: Pictor A, 3C219, 3C403, and PKS2152-69.\
A “\*” after the source name indicates that an optical detection has been reported see http://home.fnal.gov/$\sim$jester/optjets/.\
The redshift for PKS 2101-490 is uncertain (described as ’tentative’ by @mars05).\
PKS2152-69 is odd, having a bright knot close to the core, and a disparity between the radio, optical, and X-ray distributions. @ly05 argue for a thermal interpretation of the X-ray emission.\
References\
1, @fabi03; 2, @mars05; 3, @schw06; 4, @wils01; 5, @hard05b; 6, @samb04; 7, @char00; 8, @schw00; 9, @siem03a; 10, @jors04a;11, @aldc03; 12, @coma03; 13, @char02; 14, @siem02; 15, @mars01; 16, @samb01; 17, @kata03; 18, @cheu04; 19, @kraf05; 20, @ly05; 21, @siem03b.
Evaluation {#sec:2eval}
----------
In fig. \[fig:obs\_phys\] we show a plot of the observed parameters, jet length (projected) and observed (i.e. assuming isotropic emission) X-ray luminosity, L$_x$. This plot conforms to the common perception that quasars have powerful jets and are generally longer than those of FRI galaxies. Perhaps the only surprise is the gap with no jets lying between 10$^{42}$ and 10$^{43}$ ergs s$^{-1}$. The lower right is sparsely populated partly because in a large fraction of FRI jets, only the inner segment is detected in X-rays. The upper left is empty because short jets at typical quasar redshifts will be difficult to resolve from the nuclear emission with arcsec resolutions. A separation of $\approx$2$''$ is required to detect a jet close to a bright quasar, and at a typical redshift of 0.5, this already corresponds to 10kpc.
The FRII radio galaxies have projected sizes comparable to those of the quasars, but are of lower apparent luminosity. The weakest jet (lower left corner) is M84 for which X-ray emission has been detected in the very inner part of the radio jet. 3C129 is quite similar, and joins Cen A, both points lying to the lower left of the main clump of FRI’s.
This sort of plot is useful for comparative purposes, but not for interpretation because L$_x$ is only an apparent luminosity and not the true luminosity in the jet frame and also because the length is a lower limit because of projection.
For a subset of the sources plotted in fig. \[fig:obs\_phys\], some reasonable estimate for the angle between the line of sight and the jet has been published. Since most of these jets are sensibly straight on kpc scales (3C120 being a notable exception), we can obtain a deprojected length. For most of these, an estimate of the beaming factor is also available. For the majority of the quasars, the value of $\delta$ given in Table \[tab:fr2jets\] is model dependent since it is the beaming factor required for the IC/CMB model. For the FRI and FRII radio galaxies, the $\delta$ values are derived from various lines of arguments based on geometry of the lobes, VLBI superluminal motions, and other more or less reliable methods. Assumed $\delta$’s for the FRI jets are: 1.3, 3C31 and M84; 3.5, M87 and 3C371; 4, Cen A; and 3, 3C346. However, all beaming factors are suspect and the corresponding uncertainty will most likely introduce scatter in plots such as fig. \[fig:lxp\_phys\] which plots L$_x^{\prime}$=L$_x(obs)/\delta^4$ against length(obs)/sin$\theta$.
The main purpose of fig. \[fig:lxp\_phys\] is to demonstrate that with the ’current community interpretation’ (i.e. FRI jets come from synchrotron emission whereas quasar jets are dominated by IC/CMB emission), FRI jets and quasar jets are more clearly separated on the basis of size rather than luminosity. Parameters for the smallest quasar jets (the group of 5 around log L$_x^{\prime}$=42, length=70kpc) are less secure since the jet emission is only of order one resolution element from the quasar core emission for these sources.
L$_x^{\prime}$ values are compromised by model dependency. If quasar jets were to come from synchrotron emission instead of IC/CMB emission, the appropriate $\delta$ could well be of order 3 or 4 (similar to that for FRI’s, and adequate to explain the jet one-sidedness) instead of typical values like 10. Thus the luminosity correction when moving to the jet frame would be closer to a factor of 100 than 10,000 and the plot would be closer to a scaled version of fig. \[fig:obs\_phys\].
For both of these figures we need to remember that ’low-power’ and ’high-power’ sources are so divided according to their total radio luminosity. When we plot the jet luminosity we are dealing with a parameter that quantifies the jet loss, not the jet power. Since FRI jets are commonly thought of as being ’lossy’, the underlying assumption is that a larger fraction of the FRI jet power is radiated than is the case for FRII jets. Thus both the characteristic power and the fractional energy lost to radiation for both classes of sources are ’free’ parameters and the resulting luminosities (luminosity = total jet power $\times$ fractional loss to radiation) would not necessarily be expected to be similar as in fig. \[fig:lxp\_phys\].
Observations of Resolved Jets {#sec:data}
=============================
Relevant Radio and Optical Considerations {#sec:3radopt}
-----------------------------------------
In addition to the critical role of radio and optical flux densities, which complement the X-ray intensities in defining the SED’s of jet knots, these longer wavelengths provide 2 critical capabilities for jet observations: higher angular resolution than that of the CXO, and polarization. Moreover, in most cases, we are confident that we can interpret the data on the basis of synchrotron emission rather than being faced with the uncertainty of IC vs. synchrotron emission, as is the case for the X-rays.
For the SED’s, the IR-optical-UV data are usually those which determine if a synchrotron spectrum (broken power law with high energy cutoff) can be used to describe the radio to X-ray data. For example, @jest05 find a spectral flattening from HST data in the 3C273 jet which is the basis for the claim that a simple synchrotron spectrum cannot fit all the data. Often the optical upper limit for a non detection is used to preclude a single synchrotron component whereas if no optical data were available, the radio and X-ray data could have been interpreted as a single (broken) power law.
There are at least two observational problems affecting the construction of SED’s. The first is the uncertainty that our photometry is measuring the same entity in all bands. The CXO resolution is significantly worse than that of the HST so to gather the counts for photometry, one needs at least a circle with radius of 0.5$^{\prime\prime}$. Thus when measuring the SED of the knots in e.g. 3C273 (fig. \[fig:hst273\]), we implicitly assume that the X-rays are coming from the same emitting volume as the optical/radio, and not from some additional volume such as a sheath around knots.
The second uncertainty is the absorption correction which mainly affects the UV and soft X-ray data, and depends not only on the column density to the source, but also on the gas to dust ratio.
### Morphology and Polarization at kpc scales {#sec:3kpc}
One of the more significant advances in understanding radio jets has been achieved by @lain02a [@lain02b]. For the case of FRI jets for which both sides are visible and well resolved, they have been able to use the laws of energy and momentum conservation to solve for all the physical jet parameters assuming that the observed differences in brightness and polarization between the two sides are caused by relativistic effects only [see @koni80 for a general discussion of relativistic effects]. For 3C31 they find $\beta$=0.8 to 0.85 initially, then decelerating to $\beta$=0.2 at a few kpc’s, with loading by entrainment being the cause of the deceleration. The solution requires cross jet velocity structure: the outside has to be going slower than the center. This is consistent with, but does not require, a simple ’spine/sheath’ structure.
Optical and radio polarization have also been used to study the field configuration in relation to the properties of internal shocks in jets. For example, @perl05 find that the peaks in X-ray brightness in the M87 jet coincide with minima of the optical polarization. They conclude that this is consistent with the location of internal shocks which both produce the X-ray emission via particle acceleration and change the magnetic field direction. The observed reduction in polarized signal would then be a result of beam smearing over a region of swiftly changing field direction.
Other notable progress coming from optical data includes @bire99 who found features moving at $\approx$6c downstream from M87/HST-1 (the same knot which later flared by a factor of 50 across the spectrum). This demonstrates that at least mildly relativistic velocities persist to kpc scales. Several investigators [see for example @macc96] have also noted that optical emission away from bright knots requires continuous acceleration processes since the E$^2$ loss times are so short that the electrons responsible for the observed emission cannot travel from the shock locations; the same sort of argument was later deduced from similar morphologies observed at X-ray frequencies.
### Parsec scale structures {#sec:3pc}
The most relevant aspect of VLBI work for X-ray jet physics is the accumulating database containing monitoring of a reasonably large sample of quasar, blazar, and BL Lac jets. The original work was the “2 cm survey” which has now become institutionalized on the web as MOJAVE[^4]. These data provide a wealth of information such as the distribution of beaming factors and jet velocities (if one accepts the notion that observed proper motions of jet features reflect the underlying jet velocity and that the sources are at distances indicated by their redshifts). @kell04 find apparent velocities $\beta$ ranging from zero to 15, with a tail extending up to 30 for individual features. With assumptions about brightness temperatures, this can be translated to $\Gamma$ values covering a similar range. If bulk velocities of this magnitude persist to kpc scales, one of the prerequisites of the IC/CMB model for X-ray jet emission will be satisfied.
Another VLBA monitoring project is described in @jors05. They find a similar range for $\Gamma$ using intensity variability timescales to estimate $\delta$, with most quasar components having $\Gamma$ of order 16 to 18. In both of these works, there is ample evidence of non-ballistic motion: velocity vectors of components mis-aligned with the jet vector.
@gabu04 have used the transverse polarization structure of jets resolved with VLBI to argue for a helical structure for the magnetic field governing the emitting region; circumstantial evidence for non-ballistic motions.
@ward98 argued for jet composition being a pair plasma based on circular polarization inferences and @hiro99 suggested that two components in the jet of 3C279 were dominated by pair plasma on the basis of electron density arguments.
The X-ray Data {#sec:3xray}
--------------
We will not cover inferences from unresolved X-ray behavior of cores, but concentrate on jet features for which we have some confidence that the radio, optical, and X-ray emission comes from the same emitting volume. We make the usual assumptions that all relativistic plasmas will emit both synchrotron and IC radiation and since most/all X-ray jets are one sided, and these sides are the same as the those which have VLBI superluminal jets, $\Gamma >$ 1, but not necessarily $\geq$ 5 (for kpc scales).
### Jet Structure {#sec:3structure}
So far, there is very little transverse structure available from X-ray data. A notable exception is knot 3C120/k25 [@harr04] which is resolved into 3 components. The jet of Cen A is well resolved since it is the nearest jet source [@kraf02]. M87/knots A, B, and C are a bit larger than the point spread function [@perl05] and Marshall (private communication) reports that several features in the jet of 3C273 are also resolved by the CXO.
In general, there is good correspondence between jet knots mapped in the radio, optical, and X-ray bands. For both IC and synchrotron emission models, this is expected in first-order approximations if there is a single relativistic electron distribution responsible for all observed emissions. Relative intensities between bands can vary depending on the relative magnitudes of energy densities in the magnetic field and in the photons, as well as on the form of the electron distribution, N(E), since different bands come from different segments of N(E).
There are, however, a few cases of gross mis-alignment between an X-ray feature and emissions at lower frequencies. In the M87 jet, beyond knot C (upper right of fig. \[fig:m87\]) the radio jet makes a sudden excursion to the north downstream from a sharp gradient in radio brightness (in the opposite sense to that of the leading edge of knot A). Although there is weak radio emission downstream of this radio edge, the X-ray brightens. One interpretation might be that the radio jet encounters an ’obstacle’ causing an internal shock and a jet deflection. The X-ray emission would then come from the obstacle, and not properly be associated with the jet itself. A similar situation occurs just downstream of knot A in the jet of 3C273 where the radio jet deflects to the South before resuming its principal direction, whereas the X-ray, and in this case also the optical emission, continues further North along the main jet vector defined by knot A and the rest of the jet (fig. \[fig:273K\]).
Another example of discrepant correspondence between radio and X-rays is Cen A [@kraf02; @hard03; @hard04a], shown in fig. \[fig:cena\]. While many radio and X-ray features align well, albeit with quite different relative intensities, there are a few X-ray knots that have no obvious corresponding radio enhancements. For PKS1136-135 [@samb02; @samb04], the radio emission associated with the first bright X-ray knot (’A’) is extremely weak; another example of the range of relative intensities between radio and X-ray emissions.
#### Offsets {#sec:3offsets}
Another common, although not universal, effect is the offset between peak X-ray, optical, and radio brightness distributions when mapped with similar angular resolutions. It is generally the case that when this occurs, the higher frequency brightness peaks at the upstream end of the knot, and the underlying cause seems to be a steepening of the spectra moving downstream [@hard03]. A few examples are Cen A, fig. \[fig:cena\] for knot A2 at RA=13h 25m 29s; M87 knots D and F (fig. \[fig:m87\]); and knot B in PKS1127-145 (fig. \[fig:1127\]). For the nearby sources Cen A and M87, the magnitude of the projected offsets are of order tens of parsecs, whereas for PKS1127-145 at z=1.18, the observed offset is of order 10 kpc. Additional examples from the FRI category are listed in @bai03, a paper devoted to the offset effect.
For the simplest synchrotron scenario, if electrons were to be accelerated at a single location (i.e. a shock) and then be advected down the jet, all synchrotron bands should coincide insofar as peak brightness goes, even though downstream, we would expect to lose the highest energy electrons sooner than the lower energy electrons responsible for the radio emission. This statement assumes perfect angular resolution whereas normally, our beam sizes are not adequate to discern these structural differences. Thus even with the same angular resolution, the peak brightness of the X-ray emission, being centered on the shock, can occur upstream of the radio centroid which has been shifted downstream a bit since the downstream plasma will continue to produce radio emission (but not X-rays).
Another possibility to explain the observed offsets which is not fine-tuned to the beam size is an increasing magnetic field strength downstream from the shock, thereby enhancing the radio emissivity. While the physical scales for the nearby sources are reasonably consistent with these models (travel time matching E$^2$ halflives), the 10kpc offset for PKS1127-145 (again assuming synchrotron emission) would most likely rely on the second explanation. It is also the case, as emphasized earlier, that whereas we expect the electrons responsible for the X-ray emission to travel no further than some tens of light years, there is emission between the bright knots in many jets, and this supports the presence of a distributed, quasi continuous acceleration mechanism (see sec. \[sec:1radi\]).
For the IC/CMB model, currently there is no reasonable explanation as to why the X-ray emission should drop off more rapidly than the radio/optical synchrotron emission since the X-ray producing electrons have $\gamma\leq$200, and thus much longer halflives than the electrons responsible for the radio (and optical) emission. There are of course [*ad hoc*]{} possibilities such as a precursor shock system (or some other mechanism) which would accelerate copious numbers of electrons only up to some small value of $\gamma$ like 1000.
#### Progressions {#sec:3progressions}
Another effect which is closely associated with offsets is what we call “progressions”. This term is applied to those jets for which the X-ray intensity is highest at the upstream end, thereafter generally decreasing down the jet, whereas the radio intensity increases along the jet. Progressions are rather common; in @samb04 this effect is shown for 7 quasars. The most striking example is 3C273 and profiles are shown in fig. \[fig:273prof\]. Note that the optical knots are of relatively constant brightness. If this jet were to be observed with a single resolution element, there would be a clear offset between peak brightnesses in the radio and X-ray which would most likely be comparable to the length of the bright part of the jet, $\approx$ 6$^{\prime\prime}$ (15 kpc). Referring to fig. \[fig:size\], we see that the 3C273 jet is about the size of a single ’knot’ in the PKS1127-145 jet. Thus we see that ’progressions’ and ’offsets’ can be considered to be two manifestations of an underlying spectral behavior. Both offsets and progressions are observed in FRI and in quasar jets although the common explanations for the two classes differ. For synchrotron models an increasing magnetic field strength is posited, thereby increasing the synchrotron emissivity. For IC models, a gradual decrease in jet bulk velocity is assumed, leading to a diminishing u’($\nu$) in the jet frame [@geor04].
\*
#### Emission between the knots {#sec:3interknot}
Although the conventional view of synchrotron jets is that electrons responsible for X-ray emission cannot propagate more than a few light years from their acceleration site, lower brightness emission is often detected between the brighter knots. In their survey of quasar jets, @samb04 [their sec. 4.3] remark on this attribute for PKS0605-085 and possibly also for 3C207 and PKS1136-135. Because of the lower brightness levels in both radio and X-ray bands, it has been difficult to obtain the data necessary to perform spectral tests for the emission mechanism. Quasi continuous jet emission is expected for the IC/CMB process whereas a synchrotron hypothesis would require continuous acceleration processes, as outlined in sec. \[sec:1distrib\].
### Relative Intensities and SED’s {#sec:3sed}
Combining radio, optical, and X-ray photometry to create a broad band spectrum is the standard method to discriminate between synchrotron and IC emission. Resting on the common assumptions of a power law distribution for the relativistic electrons and E$^2$ losses affecting the highest energy electrons more severely than the lower energy electrons, the notion that a one zone synchrotron source must have a concave downward spectrum has been generally accepted. This approach can work even if only 3 flux densities are available (i.e. one radio, one optical, and the X-ray measurement), but can become stronger with more data, permitting estimates of the spectral index to be obtained within each band.
There are several variations of this test such as plotting $\alpha_{ro}$ against $\alpha_{ox}$ [e.g. @samb04 fig. 4], or simply demonstrating that an optical upper limit lies below the line connecting the radio and X-ray data [e.g. @schw06]. Of the 34 quasars and FRII radio galaxies listed in Table \[tab:fr2jets\], 9 have been shown by @samb04 to have knots with $\alpha_{ro}>\alpha_{ox}$ and another 8 have optical upper limits which preclude a simple synchrotron fit. Twelve of the sources do not yet have useful optical data available, and the remaining 5 consist of a few FRII radio galaxies and a couple of quasars for which some knots have spectra consistent with a synchrotron fit while others do not.
Although it is thus fairly simple to demonstrate that a simple (i.e. ’single zone’) synchrotron spectrum fails to apply to most knots in quasar jets, once high quality data are available, serious problems arise also for the single zone IC/CMB models. Examples are spine-sheath models devised to benefit both from high and low $\Gamma$ effects and 3C273 for which some knots seem to have $\alpha_x~>~\alpha_r$, contrary to the expectation that electron spectra will most likely flatten at low energies [@jest06].
### Variability {#sec:3varib}
Since the CXO has been providing X-ray photometry of jet components for only 6 years, to detect variability we require a small physical size abetted by a significant value of $\delta$ to compress the elapsed time in our frame.
Thus, clear intensity variability has so far only been found in Cen A [@hard03] and M87 [@harr06a]. This is not meant to preclude the possibility of detection of variability in larger structures which could well contain small scale structure. If one were to ascribe the factor of 100 difference in apparent luminosities between 3C273 and M87 to a factor of 3 difference in $\delta$, then an event such as the flare in knot HST-1 of M87 [@harr06a] would be easily detected. The factor of 50 increase in the X-ray flux from HST-1 means that what was once an inconspicuous X-ray knot for a time outshone the remainder of the jet plus the unresolved core of M87.
Proper motion has been observed for radio features in Cen A [@hard03] and optical features moving at up to 6c downstream of HST-1 in M87 [@bire99]. In both cases, the associated X-ray features align with stationary radio or optical components.
Jet Detection Statistics {#sec:statistics}
------------------------
In the standard picture of AGN unification [@urry95], there are two main classes of AGN, low-power AGN (BL Lac objects and low-power radio galaxies) and high-power AGN (Quasars and high-power radio galaxies). The differences of AGN within each class are explained with a different degree of alignment between the line of sight and the symmetry axis of the AGN (assumed to be parallel with the AGN jets). BL Lac objects are interpreted to be the aligned versions of FRI radio galaxies, and steep spectrum radio quasars (SSRQ - sources with a radio spectral index $\alpha_{\rm r}$, $>$ 0.5 at a few GHz) and flat spectrum radio quasars (FSRQ, $\alpha_{\rm r}$, $<$ 0.5) are increasingly aligned versions of the FRII parent population. @urry95 derive the luminosity functions of the beamed AGN from the luminosity function of the parent populations. For their specific source samples, their analysis indicates that FRII radio galaxies have jets with bulk Lorentz factors of between 5 and 40, and that SSRQs and FSRQs are FRIIs with jets aligned to within $\sim$38$^\circ$ and $\sim$14$^\circ$ to the line of sight, respectively. The bulk Lorentz factors of the FRI jets are less constrained but seem to be somewhat lower than those of the FRII jets, and the jets of their radio selected BL Lacs seem to be aligned to within 12$^\circ$ to the line of sight.
In this context, we now consider the sources with X-ray jets (see Tables 1 and 2). The low-power sources with X-ray jets are mostly FRI radio galaxies, except for 3C120, a Seyfert I galaxy and the two BL Lacs PKS0521-365 and 3C371. The high-power sources with X-ray jets are all classified as quasars except for the four FRII radio galaxies Pictor A, 3C219, 3C403 and PKS2152-69. Remarkably, almost all X-ray jets from the non-aligned FRI and FRII sources can be explained as synchrotron emission from mildly relativistic jets; Lorentz factors of a few are needed to explain the non-detection of counterjets. Most sources for which the simple synchrotron picture does not work are quasars. In this case, explaining the X-ray emission requires Lorentz factors on the order of 10 and viewing angles on the order of 10$^\circ$. The IC/CMB interpretation of the X-ray emission thus indicates that the X-ray jets detected so far are similarly closely aligned to the line of sight as the average radio-selected FSRQs used in the FRII/quasar unification analysis described above.
@samb04 [@mars05] used the CXO to study the fraction of sources with X-ray jet emission for certain source samples. @samb04 studied sources with bright 1.4 GHz radio emission and a radio knot detection more than 3$''$ away from the core. Out of a sample of 17 sources, X-ray jets were detected for 10 sources. @mars05 studied two samples of flat spectrum radio sources. One sample consisted of sources selected for their high 5 GHz flux density. The other sample consisted of sources with one-sided linear radio jet morphology. Out of 19 sources of the first sample, 16 were detected with short CXO observations. The detection probability in the second sample was lower, but this finding was not statistically significant.
The samples used in these “survey-type studies” were biased toward beamed sources. As the alignment of the sources is poorly constrained by the longer wavelength data, the high detection fraction with CXO cannot be used to argue for or against the IC/CMB model for those sources for which simple synchrotron models do not account for the X-ray emission.
As mentioned above, the radio spectral index can be used as an indicator of the jet orientation relative to the line of sight. A similar indicator is the lobe over core dominance at intermediate radio frequencies like 5 GHz. A test of the IC/CMB model is to check that the orientation parameters indicate an aligned jet for all quasars that exhibit the bow-tie problem. Indeed only for one source (PKS 1136-135) do we find at the same time $\alpha_r>>$0.5, lobe over core dominance and a SED which indicates a bow-tie problem. However, this source does not make a strong case against the IC/CMB model. Only knot B exhibits the bow-tie problem and close inspection of the radio-X-ray morphology shows that it may well be a hot-spot rather than a jet knot. We conclude that the IC/CMB scenario is not grossly inconsistent with other orientation indicators.
Discussion & Summary {#sec:conclu}
====================
Critique of the Synchrotron Emission Model {#sec:4sync}
------------------------------------------
There seems to be little doubt that the X-ray emission from most or all jets of FRI sources is dominated by the synchrotron process. When SED’s are available, they are consistent with concave downward fits. There are no problems with the synchrotron parameters such as magnetic field strength or energy requirements. Light curves for variable knots also support the synchrotron model even if the predictions for behavior at lower frequencies still need to be verified. The alternative of IC/CMB emission requires unreasonable beaming parameters such as angles to the line of sight which are too small compared to a host of other estimates.
Perhaps the most important implication to be deduced from FRI jets is the necessity for distributed emission rather than a finite number of shocks. While we don’t doubt the evidence for strong, discrete shocks (e.g. a large gradient in radio brightness, often facing upstream), some additional process is required.
The most likely alternatives for the ’additional process’ are the aforementioned ’distributed acceleration’ and IC/CMB emission. As outlined in sec. \[sec:1fluid\], one of the candidates for the underlying jet ’medium’ is electrons with $\gamma~<$ a few thousand. If that option were to be correct, then it could well be the case that even very modest values of $\Gamma, \delta$, and $\theta$ would suffice for an IC/CMB model of inter-knot emission, and many of the problems of this process for knots would not be present. It is likely that sufficient data have now accumulated in the relevant archives that this test could be performed for a number of the brighter jets with well defined knots.
To explain jet segments devoid of detectable emission, this scenario would indicate that jets are inherently intermittent. Aside from testing this suggestion by careful photometry, spectral analysis, and calculation of beaming parameters, it would be a somewhat unbelievable coincidence if the energy spectral index, p, \[N($\gamma$)$\propto~\gamma^{-p}$\] were to be the same for the postulated low energy electrons responsible for the jet’s energy transport and for the highest energy electrons with $\gamma\approx~10^7$ responsible for the knot X-ray emission. Thus one could reasonably expect to see a marked change in $\alpha_x$ moving from knot to inter-knot regions.
The largest hurdle for the application of synchrotron emission to jets of quasars comes from those cases for which the optical intensity is so low (or undetected) that it precludes a concave downward spectral fit from radio to X-rays. The associated ’bow-tie’ problem (sec. \[sec:1bowtie\]) has been reported also for FRI sources [e.g. 3C120, knot ‘k25’ @harr04]. None of the possible solutions has been accepted by the community and progress on this issue depends on a demonstration that a key ingredient of spectral hardening at high energy is indicated by some independent means. Examples would be the confirmation of a prediction from the two-zone model or finding independent support for the shear layer acceleration model.
Critique of the IC/CMB Emission Model {#sec:4ic}
-------------------------------------
The idea of augmenting u($\nu$) compared to u(B) in the jet frame had been used for jets close to black holes where u($\nu$) was thought to be dominated by UV radiation [e.g. @derm94; @blan95; @siko97]. @celo01 and @tave00 applied this concept to kpc scale jets for which the CMB dominates u($\nu$). By positing that the X-ray knots of PKS0637-752 had a value of $\Gamma\approx$10, similar to the values deduced from superluminal proper motions for the pc scale jet [@ting00], they were able to show that IC/CMB was able to explain the observed X-ray intensities while still maintaining equipartition conditions between u(B) and u(p) (where u(p) is the energy density in relativistic particles). This idea was quickly adopted by the community because it was already realized that the preponderance of one-sided X-ray jets requires $\Gamma\geq$3 or 4 and it provided a solution to the vexing problem of too little optical intensity to provide a reasonable synchrotron fit to the spectrum.
Additional support for this model is supplied by the jets that show the progression, discussed above, with a decreasing ratio of X-ray to radio intensity moving away from the core. Under the IC/CMB model, all that is required is a general deceleration of the jet, thereby reducing u$^{\prime}$($\nu$) in the jet frame [e.g. for 3C273, @samb01]. There is, of course the problem of explaining why the IC/CMB X-ray intensity of 3C273/knot A happens to fall so close to the extrapolation of the radio/optical synchrotron spectrum [@mars01].
There are a number of additional uncertainties and problems for the IC/CMB model although none of these represents a definitive refutation.
#### Offsets and lifetime considerations {#sec:4offsets}
In sec. \[sec:3offsets\] we discussed offsets between X-ray and radio brightness distributions of jet knots. The low energy ($\gamma\approx$100) electrons responsible for the X-ray emission will have E$^2$ lifetimes in excess of 10$^6$ yr; sufficient to travel to the end of even a Mpc jet. Thus when confronted with a knotty X-ray jet, the question arises: once a copious supply of these electrons are generated (e.g. at knot A in 3C273), why does the emission fade to a low level and then rise again for the next knot instead of forming a continuous or cumulatively brightening jet? One might devise a rather contrived scenario by having the beaming factor decrease to end one knot, and then either increase at the location of the next knot, or posit the injection of enough new electrons to produce a bright knot, even if the beaming factor were smaller than that of the first knot. This explanation is unsatisfactory when the radio emission is considered since it should follow the X-ray behavior if $\delta$ is the controlling factor. The same sort of problem affects the observed offsets [@staw04a; @atoy04]; the X-ray brightness should persist further downstream than the optical and radio, the opposite of what is observed.
#### Energetics {#sec:4energy}
@derm04 [@atoy04] have made a comprehensive review of the X-ray emission processes and emphasize that the original formulation of IC/CMB emission worked on the equipartition assumption based on the radio data. When the electron spectrum is extended down to the low energies required by IC/CMB, the particle energy density \[and hence also u(B)\] increase significantly. This leads them to conclude that excessive energies for the jet are required, even under the “optimistic” assumption that the jet is made of an electron/positron plasma without cold protons. For PKS0637-752, they find the kinetic luminosity is $\geq7\times10^{46}$ erg s$^{-1}$ for $\delta$=27, $\theta\leq2^{\circ}$, and increases for more reasonable beaming parameters. The associated total energy is $\geq10^{57}$ ergs for the case of $\delta=\Gamma\approx$10, $\theta\approx$5$^{\circ}$.
#### Uncertainty of extrapolation of the electron spectrum {#sec:4extrap}
One of the implicit assumptions of every IC/CMB calculation (i.e. to determine the required beaming parameters to explain the observed radio and X-ray intensities) is that the electron spectrum extends to very low energies with a slope p=2$\alpha_r$+1. If that were to be the case, then $\alpha_x$ should have the same value as $\alpha_r$. However, as demonstrated in fig. \[fig:extrap\], we currently have no knowledge that this condition holds. If $\alpha_x$ is less than $\alpha_r$, it would indicate a low frequency break to a flatter spectrum and the estimated beaming parameters would be wrong. With fewer low energy electrons than assumed by the extrapolation, $\Gamma$ and $\delta$ would have to be larger and $\theta$ correspondingly smaller, exacerbating some of the problems listed above. It is, of course, conceivable that the electron spectrum takes an upturn at low energies, in which case the error goes in the opposite direction.
Another assumption often, but not always present is that of equipartition. Since every calculation requires a value of the magnetic field in order to move from the observed segment of the synchrotron spectrum to obtain the corresponding segment of the electron spectrum, the usual method is to assume equipartition. When that constraint is removed as in the case of arguing for a field strength well below equipartition [e.g. @kata05], the electron spectrum can be considered to be undefined and one can conjure up whatever number of low energy electrons are needed to explain the X-rays for a given beaming factor. In the case of @kata05, a small value of $\Gamma$ was invoked based on radio asymmetry arguments [@ward97]. The initial analysis of PKS0637-752 [@schw00] also suggested substantial dominance of u(p) since the IC/CMB scenario with beaming was not widely known at that time.
Finally, not only do IC/CMB models require a substantial extrapolation of the electron energy distribution to low energies with a fixed power law, they also require some fine tuning of a strict cutoff in the distribution at some slightly lower $\gamma$ in order not to over-produce the optical emission [e.g. @samb04 Table 7].
#### Small angles to the line of sight and physical length of jets {#sec:4theta}
From recent quasar surveys with the CXO [@samb04; @mars05; @schw06], fitting IC/CMB models yield $\delta$ values which range from 3 to 11. For $\Gamma=\delta$, this means $\theta$ is most commonly between 4$^{\circ}$ and 11$^{\circ}$. Since most X-ray jets are reasonably straight, the physical length of jets sometimes exceeds 1 Mpc (n.b. the jet lengths given in Tables \[tab:fr1jets\] and \[tab:fr2jets\] refer primarily to the X-ray extent; the radio jet is often longer).
Many workers [e.g. @derm04] find Mpc scale quasar jets uncomfortably long, and it is certainly the case that most FRII radio galaxies, the ’face-on’ counterparts of quasars under the unified scheme, are much smaller. However, there are a small number of ’giant radio galaxies’, and even a few quasars with sizes considerably greater than 1 Mpc [e.g. @rile90].
In some cases source morphology inferences are in conflict with small $\theta$. @wils01 argue that if the IC/CMB model with equipartition is applied to the jet in Pictor A, $\Gamma=\delta$=7.2 and $\theta$=8$^{\circ}$. Such an angle to the line of sight would mean that the total extent of the source would be on the order of 3 Mpc and the hotspots at the outer end of each lobe, should be seen projected onto the radio lobes instead of protruding beyond the lobes as they are actually situated. Although @wils01 conclude that an IC/CMB model at a more reasonable $\theta\approx23^{\circ}$ would require B$<$B$_{eq}$, @hard05b subsequently have made a strong case that the X-ray emission from the Pictor A jet is synchrotron emission, not IC/CMB.
#### Expectations for jets with z$>$1 {#sec:4highz}
@schw02 has argued that at higher redshifts there should be more jet detections since the increase in u($\nu$) of the CMB by the factor (1+z)$^4$ will compensate for the usual redshift dimming of surface brightness. In addition to this effect, we might expect to see more of the lower $\Gamma$ jets with larger beaming cones because the (1+z)$^4$ factor already will statistically increase the ratio of u($\nu$)/u(B) regardless of the $\Gamma^2$ factor from the jet’s bulk velocity. So far, these predictions have not been realized [@bass04], and @kata05 [see their fig. 10] have emphasized that the required $\delta$ values for the IC/CMB model generally decrease with redshift. At this stage, the only quasar jet detection with z substantially $\geq$2 is GB 1508+5714 (z=4.3).
Tests to differentiate between synchrotron and IC/CMB models {#sec:4test}
------------------------------------------------------------
The basic tenet of the IC/CMB model for jet knots is that the X-ray emission is sampling the low energy end of the power law electron distribution. Therefore, the IC emission must continue to higher frequencies, unlike the synchrotron spectrum which is already relatively steep, and most likely will show an exponential cutoff at somewhat harder X-ray energies than available with the CXO. If we could measure the X-ray spectrum of quasar knots at much higher frequencies, and found a smooth continuation, it would be a clear confirmation of the IC/CMB model. If on the other hand, we were to find a cutoff in the X-ray spectrum, that would indicate synchrotron emission. Unfortunately, there are no real prospects of convincingly performing this test since it is so difficult to reach the required sensitivity and angular resolution above 10 keV. The CXO band is too narrow to define the expected cutoff which may well be smeared over a wide frequency band by internal source structure.
Another option for discriminating these emission mechanisms will become available as new radio telescopes with unprecedented sensitivity and resolution at low frequencies come on line in the next few years. Both LOFAR in the Netherlands and the LWA (Long Wavelength Array) in the US will have the capability to resolve jet knots and determine the characteristics of the electron distribution at the low energies of interest (fig. \[fig:extrap\]). Each of these instruments will have a reasonably wide frequency coverage so that not only the amplitude, but also the slope of the low frequency emission can be measured. If we find that the low frequency radio data indicates that the spectrum flattens significantly or has a low frequency cutoff, then the IC/CMB model will have serious problems.
Optical and IR telescopes can be used to achieve detections of jet knots which currently have only upper limits. This band plays a crucial role for the IC/CMB model because there is still substantial uncertainty as to the origin of the currently detected optical features: is this emission from the top end of the synchrotron spectrum or the bottom end of the IC spectrum? Robust detections and photometry at several wavelengths should clarify this problem which impacts on the general ’fine tuning’ of the low energy end of the electron spectrum (sec. \[sec:4extrap\]).
Detectability of the Extended Jet Emission By Gamma-Ray Telescopes
------------------------------------------------------------------
The [*EGRET*]{} detector on board the Compton Gamma-Ray Observatory established that blazars, AGNs with their jets aligned with the line of sight, are strong sources of gamma-rays. The EGRET experiment (approximately 20 MeV to 30 GeV, or 5$\times10^{21}$ to 7$\times10^{24}$ Hz) detected a total of 66 blazars with redshifts up to $z$ $\sim$ 2 [@hart99]. A small number of blazars (currently 10) with redshifts between 0.031 and 0.186 have been detected at even higher energies (GeV to TeV, frequencies above 10$^{25}$ Hz) with ground-based Cherenkov telescopes [@kraw05]. Rapid gamma-ray flux variability on time scales between 15 min and a few hours, together with assumptions about infrared to UV emission co-spatially emitted with the gamma-rays, have been used to derive a lower limit on the Doppler factor $\delta ^{>}_\sim$ 10 of the emitting plasma based on gamma-ray opacity arguments [@gaid96; @matt97]. All of these observations refer to very small physical scales, resulting in completely unresolved data from the nuclear regions. If the extended jet emission detected by Chandra indeed originates from the IC/CMB process, the IC component should in principle be detectable in the MeV/GeV energy range with the Gamma-ray Large Area Space Telescope (GLAST) to be launched in 2007 [@mcen04], and possibly also in the GeV/TeV energy regime with ground-based telescopes like H.E.S.S., VERITAS, MAGIC, and CANGAROO III [@ahar04; @week03]. GLAST has a sensitivity for the flux above 100 MeV of 3$\times~10^{-13}$ ergs cm$^{-2}$s$^{-1}$ for 5 yrs of sky-survey observations. Cherenkov telescopes like VERITAS and H.E.S.S. have a 100 GeV sensitivity of 9$\times~10^{-13}$ ergs cm$^{-2}$s$^{-1}$ for 100hrs integration. These estimates have been derived by the instrument teams for photon indices of 2. For harder photon spectra with indices of 1.5, the $\nu\times f_{\nu}$ sensitivities are about a factor of 2 better. IC/CMB models predict gamma-ray fluxes between 10$^{-13}$ and a few times 10$^{-12}$ ergs cm$^{-2}$ s$^{-1}$ [@tave04a; @derm04] so these new observatories should have sufficient sensitivity for detection.
The angular resolution of GLAST for a single photon will be 3.4$^\circ$ at 100 MeV, and 0.1$^\circ$ at 10 GeV; typical source localization accuracies will be tens of arc-minutes near detection threshold and 0.5 arcmin for very strong sources [^5]. For most sources, the angular distance between the core and kpc-scale jet is only a few arcsec, and GLAST will not be able to distinguish between core and jet emission on the basis of the spatial information. Furthermore, variability studies will be limited to rather long time scales and large fractional flux variations.
Cherenkov telescopes have better angular resolutions ($\approx$0.1$^\circ$) and source localization accuracies ($\approx$20$''$). For $\sim$100 GeV photons however, the transparency of the Universe is limited to redshifts on the order of 0.5 owing to the gamma-rays pair-producing on IR background photons (1 to 40 microns) from galaxies. Detection and identification of gamma-rays from kpc-scale jets would thus require very strong sources with very extended X-ray jets at low redshifts; the chances for obtaining unambiguous results are not promising.
Prospects {#sec:prospects}
---------
### Synchrotron Emission
In general, synchrotron emission is a powerful diagnostic of relativistic plasmas, and in the particular case of X-ray frequencies, informs us as to the location of acceleration sites. The major problem is the unknown magnetic field strength which precludes a direct determination of the electron energy distribution.
Since the X-ray emitting electrons have such a high energy and consequently short lifetime, we expect variability in jets will continue to offer new insights. With a multifrequency monitoring, it should be possible to disentangle light travel times from E$^2$ halflives and thus obtain a different estimate of the magnetic field strength and or u’($\nu$) as well as $\delta$ [@harr06a].
As more jets are studied with greater sensitivity, we believe the chances are good that we should find a few objects that display the effects of a high energy cutoff in the CXO band. Though we assume that all synchrotron plasmas have cutoffs, few if any have actually been observed in radio, optical, or X-rays. This result would impact the acceleration scenario by providing an estimate of the extent in energy of the electron distribution.
On the theoretical front, we need additional ideas of how deviations from a power law electron spectrum can occur. The two proposals currently available are rather restricted in applicability and should be further developed.
### IC emission
If the jet X-ray emission from powerful sources is indeed from the IC/CMB process, we can study different attributes of the underlying relativistic plasma than those involved in synchrotron emission. In particular, we can obtain vital information about the low energy part of the electron spectrum. Both the amplitude and slope for $\gamma\leq$1000 are germane to the injection problem for shock acceleration as well as permitting greatly improved estimates of the total particle energy density and hence the energetics of the emitting plasma.
As is well known, estimates of the photon energy density are amenable to direct observational input, and this permits us to pass more confidently from the emission spectrum to the electron spectrum. Once the electron spectrum is known, then the observed synchrotron component will provide the magnetic field strength. The basic physics is understood and IC emission is mandatory in all relativistic plasmas. The only questions are, how much emission is there and what is the frequency range of the emission.
For the beaming IC/CMB model applied to jets, some ’paradigm shifts’ will be in order. If current estimates of beaming parameters are correct, many of the relatively bright X-ray knots are, in their own frame, rather unimpressive: luminosities of order 10$^{38}$ to 10$^{39}$erg s$^{-1}$ would be common and the canonical 10$^{44}$erg s$^{-1}$ would no longer be relevant.
Another effect means that our view of jets close to the line of sight is actually a stretched out version of the time history of a very small fraction of the ’current jet length’ (by which we mean the distance from the outermost knot or hotspot to the core, at the time we observe the jet tip). This can be quickly grasped by reversing time and sending a signal from the earth to the quasar. As the wavefront of our signal passes the jet tip, the jet is moving relativistically towards the quasar. For example, take a 100,000 l.y. jet at 5$^{\circ}$ to our line of sight. If the jet has a bulk velocity of 0.99c, by the time our wavefront reaches the quasar about 98.6% of the jet (as it existed when our wavefront first reached the tip) has now been swallowed by the black hole, and is thus not observable by us. What we see, which appears to be 100,000 l.y. in length, is actually just the 1,400 l.y. long tip of the ’current jet’, as it was at progressively earlier times as we move back from the tip. The most important aspect of this effect is to make the necessary adjustments when comparing quasar jets to those lying closer to the plane of the sky. What might we actually be studying if all we see is 1% of the ’current jet length? The hotspot? If so, what we call knots in the jets would actually be bits of the hotspot brightening and fading over its 100,000 year long journey to its ’present’ location.
Summary {#sec:4summary}
-------
Within a few years, the uncertainty as to the X-ray emission process for quasar jets should be eliminated and then we will either have a method of measuring the low energy end of the relativistic electron distribution (if the IC/CMB model applies) or we will have new insights into the behavior and loss mechanisms affecting the highest energy electrons (if synchrotron models apply). If we are convinced that the IC/CMB model is correct, then a number of conclusions are already clear: detected quasar jets lie close to the line of sight and have large Lorentz factors. That in turn means we can solve for some of the basic jet parameters such as energy flux, and most likely we will improve our understanding of cross-jet velocity structure: many different lines of argument point to the necessity of some sort of ’spine-sheath’ structure.
In part 2 of this review, we have examined the differences between the jets of FRI radio galaxies and those of quasars. Will the distinctions in jet length and luminosity translate to differences in X-ray emission process? If so, why are there so many similarities between low power and high power sources such as offsets and progressions? We may also expect to better understand the underlying reasons for brightness fluctuations along jets, and if the small knots of FRI jets have the same genesis as the kpc scale knots in quasar jets. All of these lines of investigation will hopefully elucidate the dichotomy between the plasma that emits the radiation we observe and the medium which transports the energy and momentum over such vast distances.
ACKNOWLEDGMENTS
===============
We thank C. Cheung, S. Jester, M. Hardcastle, and many other colleagues for useful discussions. C. Cheung, and L. Stawarz kindly gave us helpful comments on the manuscript and the editor, R. Blandford, provided valuable advice. This work has made use of NASA’s Astrophysics Data System Bibliographic Services and the XJET website. Partial support was provided by NASA contract NAS8-03060 and grant GO3-4124A. HK thanks the Department of Energy for support in the framework of the Outstanding Junior Investigator program.
[^1]: http://www.mpa-garching.mpg.de/$\sim$ensslin/Jets/Proceedings/
[^2]: to download talks: http://www.capca.ucalgary.ca/meetings/banff2005/index.html
[^3]: http://hea-www.harvard.edu/XJET/
[^4]: http://www.physics.purdue.edu/astro/MOJAVE/
[^5]: http://www-glast.slac.stanford.edu/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper we compute general-relativistic polytropic models simulating rigidly rotating, pulsating neutron stars. These relativistic compact objects, with a radius of $\sim 10 \, \mathrm{km}$ and mass between $\sim 1.4$ and $3.2$ solar masses, are closely related to pulsars. We emphasize on computing the change in the pulsation eigenfrequencies owing to a rigid rotation, which, in turn, is a decisive issue for studying stability of such objects. In our computations, we keep rotational perturbation terms of up to second order in the angular velocity.\
\
**Keywords:** general-relativistic polytropic models; Hartle’s perturbation method; neutron stars; quasi-radial pulsation modes; rigid rotation
author:
- |
Vassilis Geroyannis$^1$, Eleftheria Tzelati$^2$\
$^{1,2}$Department of Physics, University of Patras, Greece\
$^[email protected], $^[email protected]
title:
---
Introduction
============
According to “Hartle’s perturbation method” (also called “Hartle–Thorne perturbation method”; [@4H67], [@5HT68]), we treat rotating relativistic neutron stars as perturbative solutions of the Einstein’s field equations, which describe a static spherically symmetric relativistic model. Our aim here is to compute, based on the work of Hartle et al. [@13HTC72] and also of Hartle and Friedmann [@6HF75], the zeroth- and second-order eigenfrequencies of the lowest three modes of radial pulsation for general-relativistic polytropic models. The zeroth-order eigenfrequencies, $[\sigma^2]^{(0)}$, are the eigenfrequencies of the nonrotating model, while the second-order ones, $[\sigma^2]^{(2)}$, are the rotation-induced changes in the eigenfrequencies.
In this study, we use “gravitational units” (abbreviated “gu”; see e.g. [@7GK08] and [@8GS11]), also called “geometrized units”. In these units the speed of light, $c$, and the gravitational constant, $G$, are equal to unity, $c_\mathrm{gu} = G_\mathrm{gu} = 1$, and the length is the unique “base unit”, measured in centimeters (abbreviated “cm”). On the other hand, the well-known “centimeter–gram–second units” (abbreviated “cgs”) have three base units: the length measured in cm, the mass measured in grams (abbreviated “g”), and the time measured in seconds (abbreviated “s”).
The Nonrotating Model {#NRM}
=====================
The Schwarzschild metric for a nonrotating spherical object, expressed in polar coordinates, is given by the relation ([@4H67], Equation (25)) $$ds^2=-e^{\nu}dt^2+e^{\lambda}dr^2+r^2(d\theta^2+\sin^2\theta d\phi^2),$$ where $\nu$ and $\lambda$ are metric functions of $r$. The exponential functions $e^{\nu}$ and $e^{\lambda}$ represent the fluctuations in the time rate flow and the divergence from the Euclidean geometry, respectively. The gravitational potential $\Phi$ is usually defined as (see e.g. [@7GK08], Equation (2)) $$\Phi=\frac{\nu}{2},$$ and the metric function $\lambda$ by (see e.g. [@7GK08], Equation (5)) $$e^{\lambda} = \left( 1-\frac{2m}{r} \right)^{-1},$$ where $m(r)$ is the mass-energy of the star.
A relativistic neutron star obeys the Tolman–Oppenheimer–Volkoff (TOV) equations of (i) hydrostatic equilibrium ([@4H67], Eq. (28)) $$\frac{dP}{dr} = - \, \, \frac{(E+P)(m + 4 \pi \, r^3 \, P)}{r(r-2m)},
\label{TOVP}$$ and (ii) mass–energy ([@4H67], Eq. (29a)) $$\frac{dm}{dr}=4\pi r^2 E,
\label{TOVM}$$ fulfilling the initial conditions $$P(0) = P_c = P(E_c), \ \ E(0) = E_c, \ \ m(0) = m_c = 0,
\label{TOVIC}$$ where $P(r)$ is the pressure and $E(r)$ the mass-energy density.
To treat numerically the system (\[TOVP\])–(\[TOVM\]), we need an equation relating $P$ to $E$, $P = P(E)$, that is, an “equation of state” (EOS). In this study, we consider the polytropic EOS (see e.g. [@9CST94], Equation (3); see also [@10T65], Section II) $$P = K \, \varrho^{1 + (1/n)}.
\label{poleos}$$ The parameters $K$ and $n$ are the so-called “polyropic constant” and “polytropic index”, respectively. The so-called “adiabatic index”, $\Gamma$, is defined as (cf. [@13HTC72], Equation (2.3)) $$\Gamma = \, \frac{E+P}{P} \, \left(\frac{dP}{dE} \right)_\mathrm{constant \, entropy},
\label{Gamma}$$ and, in general, is associated with perturbations about equilibrium under constant entropy. On the other hand, the so-called “adiabatic index associated with the equation of state”, $\gamma$, is defined as (cf. [@13HTC72], Equation (3.12)) $$\gamma = \, \frac{E+P}{P} \, \left(\frac{dP}{dE} \right)_\mathrm{equation \, of \, state}.
\label{gamma}$$ In the polytropic EOS (\[poleos\]), $\varrho(r)$ is the rest-mass density, related to the mass-energy density via the equation ([@8GS11], Equation (9)) $$E = \varrho + n \, P.$$
To solve the system (\[TOVP\])–(\[TOVM\]), we write the two equations as ([@8GS11], Equations (8) and (9), respectively) $$\frac{d\varrho}{dr}=-\frac{[\varrho+(1+n)P](m+4\pi r^3 P)}{r(r-2m)(dP/d\varrho)}
=-\frac{[\varrho+K(1+n)\varrho^{\Gamma}](m+4K\pi r^3\varrho^{\Gamma})}{K\Gamma r(r-2m)\varrho^{1/n}},$$ $$\frac{dm}{dr}=4\pi r^2(\varrho+nP)=4\pi r^2(\varrho+Kn\varrho^{\Gamma}).$$ The initial conditions are now $$\varrho(0) = \varrho_c, \ \ m(0)=0.
\label{MODIC}$$
To complete the study of the nonrotating model, we need to solve the differential equation for the gravitational potential $\Phi$ ([@4H67], Equation (29b)) $$\frac{d\Phi}{dr}= - \, \, \frac{1}{E+P} \, \frac{dP}{dr} =
\frac{m + 4 \pi K r^3 \varrho^{\,\Gamma}}{r(r-2m)},$$ obeying the boundary condition at the surface of the star (cf. [@8GS11], Equation (30)) $$\Phi_R = \, \frac{1}{2} \, \ln \left( 1-\frac{2M}{R} \right).$$
Rigid Rotation {#RR}
==============
When a star is rotating, its shape deviates from sphericity. Furthermore, when the star is in equilibrium state, there is a balance between the pressure forces, the gravitational forces, and the centrifugal forces. Assuming that the star is rotating rigidly and slowly (see however the starting remark in Section \[NRD\]), the perturbed metric is given by ([@5HT68], Equation (4)) $$\begin{aligned}
ds^2 &= -e^{\nu} \left[ 1 + 2(h_0+h_2P_2) \right] dt^2 +
\frac{1+2(m_0+m_2P_2)/(r-2m)}{1-2m/r} \, dr^2 \\
&+ r^2 \left[ 1 + 2(\upsilon_2-h_2)P_2 \right]
\left[ d\theta^2+sin^2\theta(d\phi-\omega dt)^2 \right] + \mathcal{O}(\Omega^3),
\end{aligned}
\label{pm}$$ where $P_l=P_l(\cos \theta)$ are Legendre polynomials; the perturbation functions $m_0, \, h_0, \, m_2, \, h_2,$ and $\upsilon_2$ are radial functions proportional to $\Omega^2$; the centrifugal forces depend on the angular velocity $\Omega$ relative to a distant observer, as well as on the angular velocity $\varpi$ relative to the local inertial frame, these two angular velocities connected via the relation ([@5HT68], Equation (6)) $$\varpi = \Omega - \omega,$$ where $\omega$ is the angular velocity of the local inertial frame. To calculate $\varpi$, we solve the so-called “frame-dragging equation” ([@5HT68], Eq. (9)) $$\frac{1}{r^4} \, \frac{d}{dr} \left( r^4 \, j \,
\frac{d\varpi}{dr} \right) + \, \frac{4}{r} \, \frac{dj}{dr}\, \varpi = 0,
\label{fd}$$ where ([@5HT68], Equation (10)) $$j = e^{-\Phi} \, \left( 1 - \frac{2 \, m}{r} \right)^{1/2}.$$ Outside the star, $\varpi$ takes the form ([@5HT68], Equation (13b)) $$\varpi = \Omega - \frac{2 \, J}{r^3},
\label{varpi}$$ where $J$ is the total angular momentum of the star, given by ([@5HT68], Equation (13a)) $$J = \, \frac{1}{6} \, R^4 \, \left( \frac{d\varpi}{dr} \right)_{r=R},
\label{J}$$ where $R$ is the radius of the star.
To solve Equation (\[fd\]), we integrate from the center of the star outwards, imposing the initial conditions $$\varpi = \varpi_c, \ \ \ \ \ \ \frac{d\varpi}{dr} = 0,$$ where the constant $\varpi_c$ is chosen arbitrarily. Since, due to Equation (\[varpi\]), the angular velocity $\Omega_\mathrm{arb}$ corresponding to the arbitrarily chosen initial value $\varpi_c$ is equal to $$\Omega_\mathrm{arb} = \varpi(R) + \, \frac{2J}{R^3},$$ and since, in general, the prescribed angular velocity $\Omega$ is different to $\Omega_\mathrm{arb}$, we must rescale the solution $\varpi(r)$ of Equation (\[fd\]) in order for the prescribed value $\Omega$ to be applied to our model (for details on this matter, see [@7GK08], Section 5.2), $$\varpi_\mathrm{new}(r) = \, \frac{\Omega}{\Omega_\mathrm{arb}} \, \, \varpi(r).$$
Spherical Deformation {#SD}
=====================
The spherical deformations due to rigid rotation are the mass perturbation function, $m_0$, and the pressure perturbation function, $p_0$. They are calculated by integrating the “$l=0$ equations of hydrostatic equilibrium” ([@4H67], Section VII), which become after a long algebra ([@5HT68], Equations (15a) and (15b), respectively) $$\frac{dm_0}{dr}=4\pi r^2 \frac{dE}{dP}(E+P)p_0+\frac{1}{12}j^2r^4
\left( \frac{d\varpi}{dr} \right)^2-\frac{1}{3}r^3 \, \frac{dj^2}{dr} \, \varpi^2,
\label{m0}$$
$$\begin{aligned}
\frac{dp_0}{dr} &= \, - \, \frac{m_0(1+8 \pi \, r^2 \, P)}{(r-2m)^2} \, - \, \frac{4\pi \, (E+P) \, r^2}{r-2m} \, p_0 \\
&+ \, \frac{1}{12} \, \frac{r^4 \, j^2}{r-2m} \left(\frac{d\varpi}{dr} \right)^2 \,
+ \, \frac{1}{3} \, \frac{d}{dr} \left( \frac{r^3 \, j^2 \, \varpi^2}{r-2m} \right),
\end{aligned}$$
with initial conditions $m_0=0$ and $p_0=0$ at $r = 0$.
The perturbation function $h_0$ (to be used in Section 5), involved in the metric (\[pm\]), is defined outside the star as ([@5HT68], Equation (17a)) $$h_0 = - \, \frac{\delta M}{r-2 \, M} + \, \frac{J^2}{r^3(r-2 \, M)}$$ and inside the star as ([@5HT68], Equation (17b)) $$h_0 = -p_0 + \, \frac{1}{3} \, r^2 \, e^{-\nu}\varpi^2 + h_{0\mathrm{c}}.$$ The constant $h_{0\mathrm{c}}$ is determined so that $h_0$ be continuous across the surface (see e.g. [@7GK08], Equation (47)) $$h_{0c}=-\frac{1}{R-2M} \, m_0(R) + p_0(R)- \, \frac{1}{3} \, R^2e^{-2\Phi(R)} \, \varpi^2(R).$$ In these equations, $M$ is the mass-energy of the star (also called gravitational mass); and $\delta M$ is the increase in the gravitational mass due to spherical deformation, given by ([@7GK08], Equation (43)) $$\delta M = m_0(R)+ \, \frac{J^2}{R^3}.$$
Quasi-Radial Pulsation {#RO}
======================
When a star is rotating rigidly with a small angular velocity $\Omega$ (see however the starting remark in Section \[NRD\]), the squared eigenfrequencies $\sigma^2$ of its pulsation modes can be expanded in powers of $\Omega$ ([@6HF75], Equation (2.1)), $$\sigma^2 = [\sigma^2]^{(0)} + [\sigma^2]^{(2)} + \dots
\label{wdistort}$$ where the superscript “$(0)$” denotes terms of zeroth order in $\Omega$, and the superscript “$(2)$” terms of second order in $\Omega$. The aim of the present study is to compute the zeroth- and the second-order eigenfrequencies of the lowest three pulsation modes.
Concerning the zeroth-order eigenfrequencies, we begin with the so-called “Chandrasekhar operator” ([@13HTC72], Equation (4.6b)), $\mathcal{L}[U]$, applied to the so-called “displacement function” ([@13HTC72], Equations (4.2a,b,c,d)), $U$, and set equal to zero, $$\mathcal{L}[U] = \mathcal{W} \, [\sigma^2]^{(0)} \, U +
( \mathcal{A} \, U' )' +
\mathcal{F}_1 (-\mathcal{F}_2 - \mathcal{F}_3 + \mathcal{F}_4)\,U = 0,
\label{LU}$$ where primes denote differentiation with respect to $r$, and $$\mathcal{F}_1 = e^{(3\nu + \lambda)/2},$$ $$\mathcal{F}_2 = \, \frac{4 \, P'}{r^3},$$ $$\mathcal{F}_3 = 8 \, \pi \, e^\lambda \, \frac{P(E+P)}{r^2},$$ $$\mathcal{F}_4 = \, \frac{(P')^2}{(E+P) \, r^2},$$ $$\mathcal{A} = \, \frac{\mathcal{F}_1 \, \Gamma \, P}{r^2},$$ $$\mathcal{W} = \, \frac{e^{(\nu + 3\lambda)/2} \, (E+P)}{r^2}.$$ The second-order differential equation (\[LU\]), subject to the initial condition ([@13HTC72], Equation (4.2e)) $$U = \alpha \, r^3 \ \ \ \mathrm{near} \ \ r = 0
\label{SLIC}$$ (without loss of generality, we can take $\alpha=1$) and to the boundary condition $$\Gamma \, P \, U' = 0 \ \ \ \mathrm{at \, the \, surface} \ \ r = R,
\label{SLBC}$$ establishes a Sturm-Liouville (SL) boundary value problem with eigenvalues the pulsation eigenfrequencies $[\sigma^2]^{(0)}$. Equation (\[LU\]) can be put into the so-called “SL form” (cf. [@14KR01], Equation (14)) $$(\mathcal{A} \, U')' +
\left(\mathcal{Q} + \mathcal{W} \, [\sigma^2]^{(0)}\right) U = 0,
\label{SL}$$ where $$\mathcal{Q} = \mathcal{F}_1 \left(-\mathcal{F}_2 - \mathcal{F}_3 + \mathcal{F}_4\right).$$ The SL form (\[SL\]) can be transformed into a system of two first-order differential equations (cf. [@14KR01], Equations (18) and (19)), $$U' = \upsilon, \ \ \ \ \ \
\upsilon' = \, - \, \frac{\mathcal{A}' \, \upsilon +
\left( \mathcal{Q} + \mathcal{W} \, [\sigma^2]^{(0)} \right) U}{\mathcal{A}},$$ subject to the initial conditions $U = r^3$ and $\upsilon = 3 r^2$ near $r = 0$, and to the boundary condition $\Gamma \, P \, \upsilon = \Gamma \, P \, U' = 0$ at the surface (Equation (\[SLBC\])).
To compute the eigenvalue(s) $[\sigma^2]^{(0)}$, we work as follows. We start the numerical integration for a trial value $\sigma^2$ and initial conditions as above. We integrate towards the surface and then check if the resulting solution $U$ satisfies the boundary condition $\Gamma \, P \, U(R)' = 0$. From the point of view of numerical analysis, this boundary condition can be treated as an algebraic equation of the form $f\left( \sigma^2 \right) = 0$; thus our task, to compute the root(s) $[\sigma^2]^{(0)}$ of this equation, can be achieved by a standard numerical method (e.g. the bisection method).
The second-order eigenfrequencies $[\sigma^2]^{(2)}$ are computed by the relation ([@13HTC72], Equation (4.8)) $$[\sigma^2]^{(2)} = \,
\frac{\int_0^R \left[ e^{\nu+\lambda/2} \, U(r) \, \mathfrak{D}(r) \right] dr}
{\int_0^R \left[ \mathcal{W}(r) \, U^2(r) \right] dr}.$$ The driving term $\mathfrak{D}(r)$ is defined as (cf. [@13HTC72], Table 4) $$\begin{aligned}
\mathfrak{D}= & \, \, U' \times
\Biggl\{ m_0 \, r^{-4} \, e^{2\lambda+\Phi} \, \Gamma \left( E+P \right) + \mathcal{T}_0 + \frac{2}{3} \, \varpi^2 r^{-1} e^{\lambda-\Phi} \, \Gamma \, P \times \mathcal{T}_1 \Biggr\} \ \ + \\
& \, \, U \times \Biggl\{ m_0 \, r^{-5} \, e^{3\lambda+\Phi} \times \mathcal{S}_0 + 2 h_0 \, r^{-2} \, (E+P) \, e^{\lambda-\Phi} \, [\sigma^2]^{(0)} \ \ + \\
&p_0 \, \left( E+P \right) \, r^{-4} \, e^{2\lambda+\Phi} \times \mathcal{S}_1 + 4\varpi\varpi' \, r^{-1} e^{-\Phi} \left( E+P+\frac{1}{3}\,\Gamma P \right) \ \ + \\
&\frac{2}{3} \, \varpi'^2 \, e^{\lambda-\Phi} \times \mathcal{S}_2 + \frac{2}{3} \, \varpi^2 \, r^{-2}\, e^{\lambda-\Phi} \times \mathcal{S}_3 \Biggr\},
\end{aligned}$$ where, in turn, $$\begin{aligned}
\mathcal{T}_0 = & \, \Biggl[ \frac{1}{2} \, p_0 \, \Gamma \, \left( E+P \right) r^{-3} e^{\lambda+\Phi} \ \ \times \Biggr. \\
& \left( \frac{E+P}{\gamma P}-\frac{E}{P} \right) \left( 1-e^{-\lambda} \right) -\frac{2}{3} \, \varpi \, \varpi' \, e^{-\Phi} \, \biggl( \Gamma \left( E+P \right) \ \ + \biggr. \\
& \Biggl. \biggl. \frac{2\,(\Gamma P)^2}{E+P} \biggr)
-\frac{1}{12} \, \varpi'^2 \, re^{-\Phi} \, \Gamma \left( E+P \right) \Biggr]
\end{aligned}$$
$$\begin{aligned}
\mathcal{T}_1 = & \left[ -\frac{1}{2} \left( 3e^{-\lambda}-1 \right) \frac{E+P}{P}+\frac{1}{2} \, \frac{\Gamma P}{E+P} \left( 1-5e^{-\lambda} \right) \ \ + \right. \\
& \frac{1}{2} \, \Gamma \left( 1-e^{-\lambda} \right) \left( 1-\frac{1}{\gamma} \right) +4\pi r^2 \Gamma P \left( 1-\frac{1}{\gamma}+\frac{P}{E+P} \right) \ \ - \\
& \left. re^{-\lambda} \frac{\Gamma'P}{E+P} \right],
\end{aligned}$$
$$\begin{aligned}
\mathcal{S}_0 = & \Biggl[ \Gamma(E+P) \biggl[ -\frac{1}{2} \left( 1-e^{-\lambda} \right) \biggr. \Biggr. \ \ + \\
& \biggl. 4\pi r^2P \left( 1+2e^{-\lambda} \right) +64\pi^2r^4P^2 \biggr] + (E+P+\Gamma P) \biggl[ -1-3e^{-\lambda} \biggr. \ \ - \\
& \biggl. 16\pi r^2P \left( 1 + \frac{1}{2} \ e^{-\lambda} \right) -64\pi^2r^4P^2 \biggr] +re^{-\lambda} \Gamma' P \left( 1+8\pi r^2P \right) \ \ - \\
& \Biggl. 2(E+P)[\sigma^{(0)}]^2r^2e^{-\lambda-\nu} \Biggr],
\end{aligned}$$
$$\begin{aligned}
\mathcal{S}_1 = & \biggl( \frac{E+P}{\gamma P}-\frac{E}{P} \biggr) \biggl[ -[\sigma^{(0)}]^2r^2 e^{\lambda-\nu} \ \ - \biggr. \\
& \biggl. \frac{1}{4} \left( 1-e^{-\lambda} \right) \left( 1+7e^{-\lambda} \right) \biggr] +4\pi \Gamma' Pr^3e^{-\lambda}-2\pi Pr^2 \biggl[ (1+e^{-\lambda})(2+\Gamma) \biggr. \ \ + \\
& \biggl. 8\pi r^2P(1+\Gamma) \biggr] -2\pi(E+P)r^2 \biggl[ \left( 1-e^{-\lambda} \right) \Gamma+\left(1+e^{-\lambda}\right)(2-\Gamma) \, \frac{1}{\gamma} \ \ + \biggr. \\
&\biggl. 8\pi Pr^2(1-\Gamma)\left(1+\frac{1}{\gamma}\right)\biggr],
\end{aligned}$$
$$\begin{aligned}
\mathcal{S}_2 = & \biggl[ \pi r^2\left(1-\frac{1}{2}\,\Gamma\right)P(E+P)+\pi r^2\Gamma P^2 \ \ + \biggr. \\
&\frac{1}{16}\,\Gamma(E+P)\left(1-e^{-\lambda}\right)+\frac{1}{8}\left(E+P+\Gamma P \right) \left(1+7e^{-\lambda} \right) \ \ - \\
& \biggl. \frac{1}{8} \, r \, \Gamma' Pe^{-\lambda} \biggr],
\end{aligned}$$
and $$\begin{aligned}
\mathcal{S}_3 = & \Biggl[ -(E+P-\Gamma P)[\sigma^{(0)}]^2 r^2 e^{-\nu} \ \ + \\
&(E+P) \biggl[ \frac{31}{4} \, e^{-\lambda}-\frac{5}{2}-\frac{1}{4} \, e^{\lambda}+\frac{1}{2}\Gamma\left(e^{-\lambda}-1\right)\biggr] + \Gamma P \biggl[ -\frac{11}{4} \, e^{-\lambda} + \frac{3}{2} \ \ + \Biggr. \\
& \biggl. \frac{1}{4} \, e^{\lambda} \biggr] + 4\pi r^2(E+P)P \left( 3+e^{\lambda} \right) \left( \frac{1}{2}\Gamma -1 \right) + 4\pi r^2\Gamma P^2 \left( 1+e^{\lambda} \right) \ \ + \\
& \Biggl. \Biggl. 16\pi^2 r^4 P^2 e^{\lambda} \biggl[ (\Gamma-1)(E+P)+\Gamma P \biggr] + r \, \Gamma' \, P \, e^{-\lambda} \Biggr].
\end{aligned}$$
The Computations
================
In this study, to compute nonrotating models (Section \[NRM\]), rigid rotations (Section \[RR\]), and spherical deformations (Section \[SD\]), we use the corresponding numerical methods described in very detail in [@7GK08] (Sections 5.1 and 5.2). We then combine these methods with the numerical framework described in Section \[RO\] for computing the zeroth- and second-order eigenfrequencies of pulsation.
To implement all the required methods, and thus to compute the results presented here, we have written and used a $\mathit{Mathematica}^\circledR$ program.
Numerical Results and Discussion {#NRD}
================================
To begin with, it is worth remarking here that, as it has been verified by several authors (see e.g. [@BWMB05], Sections 4 and 7; see also [@PG03], Section 5.3; for differentially rotating neutron stars, see [@7GK08], Section 6), Hartle’s perturbation method gives remarkably accurate results even when applied to rapidly rotating neutron stars, although this method has been developed as a slow-rotation perturbation method.
As discussed in [@6HF75] (Section I), pulsars are identified as rotating neutron stars and, therefore, there is a strong interest in studying the influence of a rigid rotation on the properties of such relativistic objects. In particular, it is of great interest to compute the pulsation frequencies of the quasi-radial modes (i.e. modes which would be radial in the absence of rotation) for several models and thus to have a measure of the effect of general relativity on these frequencies. To that purpose, we have computed, and present in this section, relevant numerical results.
Regarding our computations, we resolve four nonrotating general-relativistic polytropic models with polytropic indices $n=1.0, \, 1.5, \, 2.0, \, \mathrm{and} \, \, 2.5$. Each model is resolved for five central mass-energy densities: $E_\mathrm{c} = 10^{13}$, $3.16 \times 10^{13}$, $10^{14}$, $3.16 \times 10^{14}$, and $10^{15}$ cgs. These values have been chosen to be below and relatively close to the “maximum-mass densities” of the corresponding models, being in fact the more interesting ones when considering neutron stars. It is worth mentioning here that the total mass $M$ of a model, treated as a function of the central density $E_\mathrm{c}$, $M = M(E_\mathrm{c})$, obtains a maximum value $M_\mathrm{max}$ for a specific value $E_\mathrm{c}^\mathrm{max}$; such a model is called “maximum-mass model”, and the central density of this model is called “maximum-mass density”. The maximum-mass densities of our models, computed by a method described in [@8GS11] (Section 4), are $E_\mathrm{c}^\mathrm{max} = 3.793 \times 10^{15}, \, 4.890 \times 10^{15}, \, 4.656 \times 10^{15}, \, \mathrm{and} \, \, 3.489 \times 10^{15}$ cgs, respectively ([@8GS11], Tables 2, 3, 4, and 5, respectively). All models, studied here, have $E_\mathrm{c}^\mathrm{max} < 5 \times 10^{15}$ cgs. Accordingly, in our computations the sequence of central mass-energy densities is terminated at $E_\mathrm{c} = 10^{15}$ cgs.
Next, each density case is resolved for the three lowest modes of pulsation: Mode 0, Mode 1, and Mode 2. For each mode, we compute a rigidly rotating configuration with angular velocity equal to the corresponding Keplerian angular velocity, $\Omega_\mathrm{K}$. Hartle’s perturbation method uses proper expansions in the rotation parameter $\epsilon = \Omega/\Omega_\mathrm{max}$, where $\Omega_\mathrm{max} = \sqrt{G \, M / R^3}$ is the angular velocity for which mass shedding starts occuring at the equator of a star. Thus $\Omega_\mathrm{max}$ describes the Newtonian balance between centrifugal and gravitational forces. However, this Newtonian upper bound appears to be a rather overestimated limit for neutron stars. For such relativistic objects, the appropriate upper bound is $\Omega_\mathrm{K}$. Hence, if the angular velocity is slightly greater than $\Omega_\mathrm{K}$, then mass shedding occurs at the equator of a neutron star. $\Omega_\mathrm{K}$ can be computed by several methods (for a discussion on this matter, see [@PG03]; for a detailed description of such a method, see [@BFGM05]; see also [@SG12], and references therein, for results concerning general-relativistic polytropic models). In this study, the Keplerian angular velocities given in Tables \[ta1\], \[ta3\], \[ta5\], and \[ta7\] have been computed by using the well-known “Rotating Neutron Stars Package” (RNS) [@S92].
Furthermore, we assume that the adiabatic indices $\Gamma$ (Equation (\[Gamma\])) and $\gamma$ (Equation (\[gamma\])) do coincide for the polytropic models under consideration, $$\Gamma = \gamma.$$ In addition, a second step towards simplification is to assume, as several authors do (see e.g. [@6HF75]; for a different view, on the other hand, see e.g. [@14KR01]), that $\Gamma$ is associated with the polytropic index $n$ via the polytropic relation (cf. [@8GS11], Equation (5)) $$\Gamma = 1 + \frac{1}{n}.
\label{GammaConst}$$
In all tables of the present paper, parenthesized (positive or negative) integers following numerical values denote powers of ten. For example, the entry $3.16(13)$ is equal to $3.16 \times 10^{13}$ and the entry $3.51(-15)$ is equal to $3.51 \times 10^{-15}$.
For the five density cases with polytropic index $n=1.5$, we can compare our results in Table \[ta4\] with respective results in Table 3 of [@6HF75]. To the purpose of such comparisons, we have written Table \[ta4\] in exactly the same format with that of Table 3 of [@6HF75]. We find excellent agreement between respective results, except for two eigenfrequencies $[\sigma^2]^{(2)}$ belonging to Mode 2. The first one arises when $E_\mathrm{c} = 10^{14} \, \mathrm{cgs}$ and leads to a difference $\sim 4\%$ (our result is “-15.95”, while the result in [@6HF75] is “-16.6”); and the second one occurs when $E_\mathrm{c} = 10^{15} \, \mathrm{cgs}$ and leads to a difference $\sim 3\%$ (our result is “-14.15”, while that in [@6HF75] is “-13.7”). Since all other results do almost coincide, it seems that these two differences are of rather minor significance.
Our main remarks on the numerical results presented in Tables \[ta4\], \[ta2\], \[ta6\], and \[ta8\] have as follows. First, in all cases examined, the eigenfrequencies $[\sigma^2]^{(0)}$ and $[\sigma^2]^{(2)}$ increase in absolute value with the central density. Equivalently, since (for the cases examined) increasing central density implies increasing gravitational mass, all the eigenfrequencies increase in absolute value with the gravitational mass. In addition, since increasing central density does also imply increasing Keplerian angular velocity for a rotating configuration, all the eigenfrequencies $[\sigma^2]^{(2)}$ increase in absolute value with the Keplerian angular velocity.
Second, all eigenfrequencies $[\sigma^2]^{(0)}$ are positive, thus representing stable nonrotating pulsating configurations. Likewise, the eigenfrequencies $[\sigma^2]^{(2)}$ are positive for Mode 0 and for the soft polytropic EOSs $n=2.5$ and $n=2.0$. Consequently, the effect of rotation is to stabilize such rotationally perturbed configurations. On the other hand, the eigenfrequencies $[\sigma^2]^{(2)}$ are negative for Modes 1 and 2 of the soft EOSs $n=2.5$ and $n=2.0$, as well as for the three modes of the stiff EOSs $n=1.5$ and $n=1.0$, thus turning to destabilize the corresponding rotating configurations. It is worth mentioning here that, among the members of a collection of EOSs, the EOS deriving the larger value $P$ for a given $E$ is the “stiffest” EOS in the collection; while the EOS leading to the smaller value $P$ for the same $E$ is the “softest” EOS in this collection. Note that, for increasing polytropic index $n$, the polytropic EOSs are getting softer; equivalently, for decreasing $n$, the polytropic EOSs become stiffer.
Finally, in all cases examined, the zeroth-order eigenfrequencies $[\sigma^2]^{(0)}$ are $\sim$one order of magnitude greater than the respective squared Keplerian angular velocities $\Omega_\mathrm{K}^2$; this inequality is in fact a necessary condition for the perturbation theory developed in [@13HTC72] and [@6HF75] to be valid, as discussed in detail in [@6HF75] (Section II). Consequently, all the cases examined lie within the domain of applicability of the theory, and, especially, of Equation for computing the second-order eigenfrequencies $[\sigma^2]^{(2)}$ which represent the rotationally induced changes in the pulsation eigenfrequencies.
---------------- ---------------------- --------- --------- ---------------------
$E_\mathrm{c}$ $\varrho_\mathrm{c}$ $M$ $R$ $\Omega_\mathrm{K}$
(cgs) (cgs) (gu) (gu) (gu)
1.00(13) 9.99(12) 3.38(3) 1.52(6) 1.98(-8)
3.16(13) 3.15(13) 1.04(4) 1.52(6) 3.51(-8)
1.00(14) 9.89(13) 3.09(4) 1.49(6) 6.24(-8)
3.16(14) 3.06(14) 8.02(4) 1.41(6) 1.10(-7)
1.00(15) 9.08(14) 1.56(5) 1.23(6) 1.92(-7)
---------------- ---------------------- --------- --------- ---------------------
: Central mass-energy density and rest-mass density, gravitational mass, and radius of a nonrotating general-relativistic polytropic model with polytropic index $n = 1.0$ and polytropic constant $K = 10^5 \, \mathrm{cgs} = 1.499 \times 10^{12} \, \mathrm{gu}$. The Keplerian angular velocities, appearing here, have been computed by RNS.\[ta1\]
---------------- ---------------------- --------- --------- ---------------------
$E_\mathrm{c}$ $\varrho_\mathrm{c}$ $M$ $R$ $\Omega_\mathrm{K}$
(cgs) (cgs) (gu) (gu) (gu)
1.00(13) 9.96(12) 1.58(4) 3.13(6) 1.42(-8)
3.16(13) 3.13(13) 2.73(4) 2.56(6) 2.52(-8)
1.00(14) 9.80(13) 4.54(4) 2.08(6) 4.46(-8)
3.16(14) 3.04(14) 7.08(4) 1.66(6) 7.81(-8)
1.00(15) 9.22(14) 9.84(4) 1.29(6) 1.35(-7)
---------------- ---------------------- --------- --------- ---------------------
: Central mass-energy density and rest-mass density, gravitational mass, and radius of a nonrotating general-relativistic polytropic model with polytropic index $n = 1.5$ and polytropic constant $K = 5.380 \times 10^{15} \, \mathrm{cgs} = 3.389 \times 10^{7} \, \mathrm{gu}$. The Keplerian angular velocities, appearing here, have been computed by RNS.\[ta3\]
---------------- ---------------------- --------- --------- ---------------------
$E_\mathrm{c}$ $\varrho_\mathrm{c}$ $M$ $R$ $\Omega_\mathrm{K}$
(cgs) (cgs) (gu) (gu) (gu)
1.00(13) 9.93(12) 2.61(4) 4.59(6) 9.96(-9)
3.16(13) 3.12(13) 3.38(4) 3.43(6) 1.76(-8)
1.00(14) 9.78(13) 4.29(4) 2.52(6) 3.10(-8)
3.16(14) 3.04(14) 5.27(4) 1.87(6) 5.41(-8)
1.00(15) 9.36(14) 6.13(4) 1.37(6) 9.32(-8)
---------------- ---------------------- --------- --------- ---------------------
: Central mass-energy density and rest-mass density, gravitational mass, and radius of a nonrotating general-relativistic polytropic model with polytropic index $n = 2.0$ and polytropic constant $K = 10^{12} \, \mathrm{cgs} = 1.291 \times 10^{5} \, \mathrm{gu}$. The Keplerian angular velocities, appearing here, have been computed by RNS.\[ta5\]
---------------- ---------------------- --------- --------- ---------------------
$E_\mathrm{c}$ $\varrho_\mathrm{c}$ $M$ $R$ $\Omega_\mathrm{K}$
(cgs) (cgs) (gu) (gu) (gu)
1.00(13) 9.93(12) 1.96(4) 5.31(6) 6.74(-9)
3.16(13) 3.13(13) 2.16(4) 3.76(6) 1.19(-8)
1.00(14) 9.84(13) 2.36(4) 2.55(6) 2.10(-8)
3.16(14) 3.08(14) 2.54(4) 1.82(6) 3.67(-8)
1.00(15) 9.60(14) 2.68(4) 1.29(6) 6.38(-8)
---------------- ---------------------- --------- --------- ---------------------
: Central mass-energy density and rest-mass density, gravitational mass, and radius of a nonrotating general-relativistic polytropic model with polytropic index $n = 2.5$ and polytropic constant $K = 1.500 \times 10^{13} \, \mathrm{cgs} = 2.980 \times 10^{3} \, \mathrm{gu}$. The Keplerian angular velocities, appearing here, have been computed by RNS.\[ta7\]
Mode 0 Mode 1 Mode 2
---------- -------------------- ------------------------------------- -------------------- ------------------------------------- -------------------- -------------------------------------
$E_c$ $[\sigma^2]^{(0)}$ $\frac{[\sigma^2]^{(2)}}{\Omega^2}$ $[\sigma^2]^{(0)}$ $\frac{[\sigma^2]^{(2)}}{\Omega^2}$ $[\sigma^2]^{(0)}$ $\frac{[\sigma^2]^{(2)}}{\Omega^2}$
(cgs) (gu) (gu) (gu)
1.00(13) 1.35(-15) -0.33 6.36(-15) -7.40 1.35(-14) -17.27
3.16(13) 4.07(-15) -0.32 1.96(-14) -7.25 4.17(-14) -17.17
1.00(14) 1.16(-14) -0.31 5.87(-14) -6.96 1.26(-13) -15.95
3.16(14) 2.98(-14) -0.28 1.66(-13) -6.65 3.59(-13) -15.51
1.00(15) 5.88(-14) -0.22 4.27(-13) -5.97 9.38(-13) -14.15
: Eigenfrequencies of the lowest three modes of a nonrotating and a rotating general-relativistic polytropic model with polytropic index $n = 1.5$, and polytropic constant and Keplerian angular velocities as in Table \[ta3\].\[ta4\]
Mode 0 Mode 1 Mode 2
---------- -------------------- -------------------- -------------------- -------------------- -------------------- --------------------
$E_c$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$
(cgs) (gu) (gu) (gu) (gu) (gu) (gu)
1.00(13) 3.51(-15) -3.72(-16) 1.50(-14) -3.38(-15) 3.19(-14) -7.77(-15)
3.16(13) 1.08(-14) -1.12(-15) 4.66(-14) -9.23(-15) 9.94(-14) -1.64(-14)
1.00(14) 3.18(-14) -3.74(-15) 1.41(-13) -3.28(-14) 3.01(-13) -7.43(-14)
3.16(14) 7.99(-14) -1.13(-14) 3.87(-13) -9.62(-14) 8.37(-13) -2.19(-13)
1.00(15) 1.33(-13) -2.70(-14) 8.68(-13) -2.40(-13) 1.92(-12) -5.47(-13)
: Eigenfrequencies of the lowest three modes of a nonrotating and a rotating general-relativistic polytropic model with polytropic index $n = 1.0$, and polytropic constant and Keplerian angular velocities as in Table \[ta1\].\[ta2\]
Mode 0 Mode 1 Mode 2
---------- -------------------- -------------------- -------------------- -------------------- -------------------- --------------------
$E_c$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$
(cgs) (gu) (gu) (gu) (gu) (gu) (gu)
1.00(13) 5.20(-16) 4.07(-18) 2.84(-15) -7.02(-16) 5.98(-15) -1.69(-15)
3.16(13) 1.54(-15) 1.44(-17) 8.72(-15) -2.16(-15) 1.84(-14) -5.21(-15)
1.00(14) 4.35(-15) 6.89(-17) 2.63(-14) -6.02(-15) 5.60(-14) -1.15(-14)
3.16(14) 1.10(-14) 2.64(-16) 7.62(-14) -1.85(-14) 1.63(-13) -4.07(-14)
1.00(15) 2.24(-14) 1.13(-15) 2.09(-13) -5.25(-14) 4.52(-13) -1.24(-13)
: Eigenfrequencies of the lowest three modes of a nonrotating and a rotating general-relativistic polytropic model with polytropic index $n = 2.0$, and polytropic constant and Keplerian angular velocities as in Table \[ta5\].\[ta6\]
Mode 0 Mode 1 Mode 2
---------- -------------------- -------------------- -------------------- -------------------- -------------------- --------------------
$E_c$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$ $[\sigma^2]^{(0)}$ $[\sigma^2]^{(2)}$
(cgs) (gu) (gu) (gu) (gu) (gu) (gu)
1.00(13) 1.62(-16) 1.57(-17) 1.25(-15) -3.24(-16) 2.59(-15) -7.89(-16)
3.16(13) 4.72(-16) 5.03(-17) 3.85(-15) -1.00(-15) 8.02(-15) -2.44(-15)
1.00(14) 1.31(-15) 1.69(-16) 1.19(-14) -2.25(-15) 2.59(-14) -2.76(-15)
3.16(14) 3.23(-15) 5.26(-16) 3.54(-14) -8.09(-15) 7.53(-14) -1.32(-14)
1.00(15) 6.16(-15) 1.68(-15) 1.03(-13) -2.52(-14) 2.18(-13) -4.95(-14)
: Eigenfrequencies of the lowest three modes of a nonrotating and a rotating general-relativistic polytropic model with polytropic index $n = 2.5$, and polytropic constant and Keplerian angular velocities as in Table \[ta7\].\[ta8\]
[01]{}
J. B. Hartle, “Slowly Rotating Relativistic Stars. I. Equations of Structure,” [*The Astrophysical Journal*]{}, Vol. 150, 1967, pp. 1005-1029. doi:10.1086/149400
J. B. Hartle and K. S. Thorne, “Slowly Rotating Relativistic Stars. II. Models for Neutron Stars and Supermassive Stars,” [*The Astrophysical Journal*]{}, Vol. 153, 1968, pp. 807-834. doi:10.1086/149707
J. B. Hartle, K. S. Thorne and S. M. Chitre, “Slowly Rotating Relativistic Stars. VI. Stability of the Quasi-Radial Modes,” [*The Astrophysical Journal*]{}, Vol. 176, 1972, pp. 177-194. doi:10.1086/151620
J. B. Hartle and J. L. Friedman, “Slowly Rotating Relativistic Stars. VIII. Frequencies of the Quasi-Radial Modes of an $n = 3/2$ Polytrope,” [*The Astrophysical Journal*]{}, Vol. 196, 1975, pp. 653-660. doi:10.1086/153451
V. S. Geroyannis and A. G. Katelouzos, “Numerical Treatment of Hartle’s Perturbation Method for Differentially Rotating Neutron Stars Simulated by General-Relativistic Polytropic Models,” [*International Journal of Modern Physics C*]{}, Vol. 19, 2008, pp. 1863-1908. doi:10.1142/S0129183108013370
V. S. Geroyannis and I. E. Sfaelos, “Numerical Treatment of Rotating Neutron Stars Simulated by General-Relativistic Polytropic Models: A Complex-Plane Strategy,” [*International Journal of Modern Physics C*]{}, Vol. 22, 2011, pp. 219-248. doi:10.1142/S0129183111016269
G. B. Cook, S. L. Shapiro and S. A. Teukolsky, “Rapidly Rotating Polytropes in General Relativity,” [*The Astrophysical Journal*]{}, Vol. 422, 1994, pp. 227-242. doi:10.1086/173721
R. F. Tooper, “Adiabatic Fluid Spheres in General Relativity,” [*The Astrophysical Journal*]{}, Vol. 142, 1965, pp. 1541-1562. doi:10.1086/148435
K. D. Kokkotas and J. Ruoff, “Radial Oscillations of Relativistic Stars,” [*Astronomy and Astrophysics*]{}, Vol. 366, 2001, pp. 565-572. doi:10.1051/0004-6361:20000216
E. Berti, F. White, A. Maniopoulou, and M. Bruni, “Rotating Neutron Stars: An Invariant Comparison of Approximate and Numerical Space-Time Models,” [*Mon. Not. R. Astron. Soc.*]{}, Vol. 358, 2005, pp. 923-938. doi:10.1111/j.1365-2966.2005.08812.x
P. J. Papasotiriou and V. S. Geroyannis, “A SCILAB Program for Computing General-Relativistic Models of Rotating Neutron Stars by Implementing Hartle’s Perturbation Method,” [*International Journal of Modern Physics C*]{}, Vol. 14, 2003, pp. 321-350. 10.1142/S0129183103004516
O. Benhar, V. Ferrari, L. Gualtieri and S. Marassi, “Perturbative Approach to the Structure of Rapidly Rotating Neutron Stars,” [*Physical Review D*]{}, Vol. 72, No. 4, 2005, Article ID: 044028. doi:10.1103/PhysRevD.72.044028
I. Sfaelos and V. Geroyannis, “Third-Order Corrections and Mass-Shedding Limit of Rotating Neutron Stars Computed by a Complex-Plane Strategy,” [*International Journal of Astronomy and Astrophysics*]{}, Vol. 2012,2, 2012, pp. 210-217. doi:10.4236/ijaa.2012.24027
N. Stergioulas, “Rotating Neutron Stars (RNS) Package,” 1992. http:// www.gravity.phys.uwm.edu/rns/index.html.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We have studied the emission properties of self-organized InAs quantum dots (QDs) grown in an InGaAs quantum well by metalorganic chemical vapor deposition. Low-temperature photoluminescence spectroscopy shows emission from single QDs around 1300 nm; we clearly observe the formation of neutral and charged exciton and biexciton states, and we obtain a biexciton binding energy of 3.1 meV. The dots exhibit an *s*-*p* shell splitting of approximately 100 meV, indicating strong confinement.'
author:
- 'N. I. Cade'
- 'H. Gotoh'
- 'H. Kamada'
- 'T. Tawara'
- 'T. Sogawa'
- 'H. Nakano'
- 'H. Okamoto'
title: 'Charged exciton emission at 1.3 $\mu$m from single InAs quantum dots grown by metalorganic chemical vapor deposition'
---
Semiconductor self-assembled quantum dots (QDs) are of considerable interest for future telecommunication applications, such as low-threshold lasers and non-classical light sources for quantum key distribution systems. Efficient single-photon emission has recently been demonstrated at visible wavelengths using semiconductor QD structures,[@michler00; @santori02; @zwiller03] and there have been many detailed investigations into the low-temperature optical characteristics of QDs emitting at 1150 nm or less.[@lomascolo02; @moskalenko02; @kaiser02] However, to date there have been only a small number of spectroscopic experiments on single QDs emitting in the important telecommunications window around 1300 nm:[@ward04] biexcitonic features have been identified in low-temperature photoluminescence (PL) from QDs grown by molecular beam epitaxy (MBE),[@alloing05] whereas similar investigations for QDs fabricated by metalorganic chemical vapor deposition (MOCVD) show an unclear power dependence in the emission.[@song05]
Quantum dot structures grown by MOCVD have potentially a large commercial value due to the high growth rates achievable; however, for applications at telecommunication wavelengths the growth is complicated by large strain effects and complex surface dynamics within the dot layers.[@passaseo04] Therefore, there is a strong motivation for studying the optical characteristics of these structures in relation to other fabrication techniques. Here, we report on the emission properties of single QDs in a novel dots-in-well (DWELL) heterostructure grown by MOCVD. We present low-temperature PL spectra from individual QDs with an emission wavelength of 1300 nm; power-dependent measurements clearly reveal the formation of an exciton-biexciton system, with a biexciton binding energy of more than 3 meV. We also identify recombination from charged exciton and biexciton complexes, and we observe a large energy difference between *s*- and *p*-shell states.
The QDs were fabricated using conventional low-pressure MOCVD on a (100) GaAs substrate: an InAs(:Bi) dot layer was deposited in a 5 nm In$_{0.12}$Ga$_{0.88}$As(:Bi) quantum well (QW), and the DWELL heterostructure grown between GaAs barrier layers and InGaP cladding layers. Bismuth doping was found to significantly improve the PL intensity and emission wavelength of the dots. The DWELL structure results in a pronounced red-shift relative to similar InAs/GaAs systems due to effects such as strain relaxation[@nishi99] and alloy decomposition.[@guffarth01] Atomic force microscopy (AFM) measurements on similar samples suggest a dot size of $<$ 15 nm with elongation along the $[011]$ axis; the QD sheet density is estimated as $2\times10^{10}$ cm$^{-2}$. A more detailed description of the growth will be published elsewhere. In order to obtain single dot spectroscopy, mesa structures were fabricated by electron-beam lithography and dry etching, with sizes between 2$\times$2 $\mu$m$^2$ and 200$\times$200 nm$^2$. Micro-PL measurements were taken using an Ar$^{+}$ laser (488 nm) focused to a $\leq$ 2 $\mu$m spot; the luminescence was dispersed in a 0.5 m spectrometer and detected with a nitrogen cooled InGaAs photodiode array ($1024\times1$). Unless otherwise stated, the sample temperature was maintained at 5 K in a continuous-flow He cryostat.
Figure \[pl\](a) shows PL spectra from an unetched region of the sample at 300 K, and from a 200 nm mesa at 5 K. In the former case the full width at half maximum (FWHM) of the QD peak is 30 meV, indicating a good growth uniformity. The shorter wavelength peak is from *p*-shell states, as discussed later. At 5 K, emission is observed below 1.0 $\mu$m (1.27 eV) from the hybrid wetting layer – quantum well (WL–QW) that forms in DWELL structures.[@chang05]
In Fig. \[pl\](b) we show PL spectra from a selection of 200 nm mesas, obtained with a power density of $\sim$ 20 Wcm$^{-2}$. Each spectrum has been shifted in energy to align the main emission peak observed at low power. The other emission lines exhibit very close similarities in energy and intensity between the different spectra; this is strong evidence that each spectrum originates from a single QD, and that the different lines arise from various exciton complexes within the dot.
To investigate the origin of these lines, PL spectra were taken for a range of excitation powers as shown in Fig. \[power\](a). At low powers the spectra are composed of four narrow lines ($<$ 100 $\mu$eV, resolution limited). All of these lines are present at the lowest powers measured ($<$ 30 mWcm$^{-2}$) for many different mesas \[see Fig. \[pl\](b)\] and exhibit an almost identical linear increase in intensity over low excitation powers before saturating at $\sim$ 500 nW \[Fig. \[power\](b)\]. These lines are assigned to recombination from neutral (X) and charged (X$^{-}$/X$^{+}$) exciton states. Emission from the neutral exciton state exhibits a polarization-dependent fine-structure splitting of approximately 300 $\mu$eV. This is attributed to splitting of the $M=\pm1$ bright exciton angular momentum states due to dot asymmetry;[@bayer02b] the origin and nature of this fine-structure will be discussed in detail elsewhere.
With increasing power, we observe the appearance of additional lines below the exciton energy. In particular, the intensities of the lines labeled 2X and 2X$^{-}$ are found to increase superlinearly with excitation power \[Fig. \[power\](b)\], which is consistent with emission from biexciton states. The intensity dependence of these lines can be fit very well over low powers by rate-equation models based on random capture of excitons into a dot.[@bacher99] The other lower energy lines are attributed to charged exciton complexes consisting of two or more electrons bound to a hole; this is consistent with theoretical predictions,[@zunger04] and other experimental observations.[@warburton00] The positive trion X$^{+}$ appears at a higher energy than the exciton as the hole wavefunction $l_{h}$ has a smaller lateral extent than the electron wavefunction $l_{e}$.[@lelong96] The lack of any observable exchange energy splitting in polarization dependent PL is further evidence of the charged nature of these complexes.[@bayer02b]
The assignment of the 2X$^{-}$ (X$^{2-}$) state has been corroborated by measuring the energy difference between this line and the 2X (X$^{-}$) line in a study of 9 dots; this was found to be 1.77 meV (1.51 meV), with a standard deviation of $<$ 100 $\mu$eV. From this survey the mean binding energies of the X$^{-}$ (X$^{+}$) and 2X states are 5.6 meV (-1.1 meV) and 3.1 meV respectively; the first value is consistent with the shifts calculated by Finley *et al*. [@finley01b] for dots with $l_{e}<10$ nm. The biexciton binding energy is also in agreement with the values obtained by Kaiser *et al*. [@kaiser02] for a similar strongly confined DWELL system.
Finally, Fig. \[power2\] shows broad-spectrum power dependence of the same dot shown in Fig. \[power\]. At $\sim$ 300 nW (10 Wcm$^{-2}$) we see the appearance of a new set of lines approximately 100 meV above the ground-state *s*-shell multiplet. These appear concomitantly with the 2X lines and therefore can be attributed to recombination of electron-hole pairs in *p*-shell states. The magnitude of this *s*-*p* shell splitting is consistent with the large confinement potential expected for structures with strain-reducing layers.[@liu05]
In summary, we have observed 1.3 $\mu$m emission at 5 K from individual InAs QDs grown in an InGaAs QW by MOCVD. This suggests that this structure can be utilized as a single-photon source at telecommunication wavelengths. Recombination from neutral and charged biexciton states was clearly observed, the former having a binding energy of 3.1 meV. The negative (positive) trion is found to have a binding energy of 5.6 meV (-1.1 meV). These values, and the large energy separation between *s*- and *p*-shell recombination, indicate strong confinement and small dot size.
This work was partly supported by the National Institute of Information and Communications Technology (NICT).
P. Michler, A. Kiraz, C. Becher, W. V. Schoenfeld, P. M. Petroff, L. Zhang, E. Hu, and A. Imamo$\breve{\textrm{g}}$lu, Science **290**, 2282 (2000).
C. Santori, D. Fattal, J. Vučković, G. S. Solomon, and Y. Yamamoto, Nature (London) **419**, 594 (2002).
V. Zwiller, T. Aichele, W. Seifert, J. Persson, and O. Benson, Appl. Phys. Lett. **82**, 1509 (2003).
M. Lomascolo, A. Vergine, T. K. Johal, R. Rinaldi, A. Passaseo, R. Cingolani, S. Patan$\grave{\textrm{e}}$, M. Labardi, M. Allegrini, F. Troiani, and E. Molinari, Phys. Rev. B **66**, 041302 (2002).
E. S. Moskalenko, K. F. Karlsson, P. O. Holtz, B. Monemar, W. V. Schoenfeld, J. M. Garcia, and P. M. Petroff, J. Appl. Phys. **92**, 6787 (2002).
S. Kaiser, T. Mensing, L. Worschech, F. Klopf, J. P. Reithmaier, and A. Forchel, Appl. Phys. Lett. **81**, 4898 (2002).
M. B. Ward, D. C. Unitt, Z. Yuan, P. See, R. M. Stevenson, K. Cooper, P. Atkinson, I. Farrer, D. A. Ritchie, and A. J. Shields, Physica E **21**, 390 (2004).
B. Alloing, C. Zinoni, V. Zwiller, L. H. Li, C. Monat, M. Gobet, G. Buchs, A. Fiore, E. Pelucchi, and E. Kapon, Appl. Phys. Lett. **86**, 101908 (2005).
H. Z. Song, T. Usuki, S. Hirose, K. Takemoto, Y. Nakata, N. Yokoyama, and Y. Sakuma, Appl. Phys. Lett. **86**, 113118 (2005).
A. Passaseo, V. Tasco, M. De Giorgi, M. T. Todaro, M. De Vittorio, and R. Cingolani, Appl. Phys. Lett. **84**, 1868 (2004).
K. Nishi, H. Saito, S. Sugou, and J.-S. Lee, Appl. Phys. Lett. **74**, 1111 (1999).
F. Guffarth, R. Heitz, A. Schliwa, O. Stier, N. N. Ledentsov, A. R. Kovsh, V. M. Ustinov, and D. Bimberg, Phys. Rev. B **64**, 085305 (2001).
W.-H. Chang, H.-Y. Chen, H.-S. Chang, W.-Y. Chen, T. M. Hsu, T.-P. Hsieh, J.-I. Chyi, and N.-T. Yeh, Appl. Phys. Lett. **86**, 131917 (2005).
M. Bayer, G. Ortner, O. Stern, A. Kuther, A. A. Gorbunov, A. Forchel, P. Hawrylak, S. Fafard, K. Hinzer, T. L. Reinecke, S. N. Walck, J. P. Reithmaier, F. Klopf, and F. Schafer, Phys. Rev. B **65**, 195315 (2002).
G. Bacher, R. Weigand, J. Seufert, V. D. Kulakovskii, N. A. Gippius, A. Forchel, K. Leonardi, and D. Hommel, Phys. Rev. Lett. **83**, 4417 (1999).
A. Zunger and G. Bester, Physica E **21**, 204 (2004).
R. J. Warburton, C. Schaflein, F. Haft, F. Bickel, A. Lorke, K. Karrai, J. M. Garcia, W. Schoenfeld, and P. M. Petroff, Nature (London) **405**, 926 (2000).
Ph. Lelong and G. Bastard, Solid State Commun. **98**, 819 (1996).
J. J. Finley, P. W. Fry, A. D. Ashmore, A. Lemaitre, A. I. Tartakovskii, R. Oulton, D. J. Mowbray, M. S. Skolnick, M. Hopkinson, P. D. Buckle, and P. A. Maksym, Phys. Rev. B **63**, 161305 (2001).
W.-S. Liu and J.-I. Chyi, J. Appl. Phys. **97**, 024312 (2005).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We report the detection of a cross-correlation signal between [*Fermi*]{} Large Area Telescope diffuse [$\gamma$]{}-ray maps and catalogs of clusters. In our analysis, we considered three different catalogs: WHL12, redMaPPer and PlanckSZ. They all show a positive correlation with different amplitudes, related to the average mass of the objects in each catalog, which also sets the catalog bias. The signal [detection]{} is [confirmed]{} by [the results of]{} a stacking analysis. The cross-correlation signal extends to rather large angular scales, around 1 degree, that correspond, at the typical redshift of the clusters in these catalogs, to a few to tens of Mpc, i.e. the typical scale-length of the large scale structures in the Universe. Most likely this signal is contributed by the cumulative emission from AGNs associated to the filamentary structures that converge toward the high peaks of the matter density field in which galaxy clusters reside. In addition, our analysis reveals the presence of a second component, more compact in size and compatible with a point-like emission from within individual clusters. At present, we cannot distinguish between the two most likely interpretations for such a signal, i.e. whether it is produced by AGNs inside clusters or if it is a diffuse [$\gamma$]{}-ray emission from the intra-cluster medium. We argue that this latter, intriguing, hypothesis might be tested by applying this technique to a low redshift large mass cluster sample.'
author:
- 'Enzo Branchini$^{2,3,4}$'
- 'Stefano Camera$^{5}$'
- 'Alessandro Cuoco$^{6}$'
- 'Nicolao Fornengo$^{7,8}$'
- 'Marco Regis$^{7,8}$'
- 'Matteo Viel$^{9,10,11}$'
- 'Jun-Qing Xia$^1$'
bibliography:
- 'references.bib'
title: 'Cross-correlating the $\gamma$-ray sky with catalogs of galaxy clusters'
---
Introduction {#sec:intro}
============
Galaxy clusters are the largest virialized objects in the Universe formed by the gravitational instability-driven hierarchical structure formation process. They are also unique astrophysical laboratories hosting galaxies, highly ionised hot gas in thermal equilibrium, dark matter (DM) and a population of relativistic cosmic rays (CRs). The last two components can provide the conditions in which a diffuse [$\gamma$]{}-ray emission can be produced. CRs can lead to the the emission of [$\gamma$]{}-ray photons via three channels: inverse Compton of relativistic electrons with the cosmic microwave background, non-thermal bremsstrahlung, and decay of $\pi^0$ produced through collision of relativistic protons with thermal protons (see, e.g., [@2010MNRAS.409..449P] and [@2016MNRAS.459...70V] for recent simulations of [$\gamma$]{}-ray emission in galaxy clusters). DM can also directly or indirectly produce $\gamma$ rays through annihilation or decay, and clusters are promising targets in the particle DM search, due to their large DM content.
[Clusters of galaxies are not isolated objects. They are located at the node of a complex cosmic web, surrounded by a network of filamentary structures populated by astrophysical sources, like the AGNs, that can contribute to the [$\gamma$]{}-ray emission also from within the cluster itself.]{}
The discovery and characterization of cluster-wide $\gamma$-ray emission is therefore important in several ways. If the signal is induced by CRs, then it can be used to discriminate between different models for the observed radio emission and clarify the nature of radio halos (see, e.g., [@2008SSRv..134...93F], and [@2014IJMPD..2330007B] for recent reviews on extended radio emissions in galaxy clusters). A detection of a signal produced by DM would be its final discovery. In this case, a detailed estimate and understanding of the contribution of all astrophysical sources to the cluster [$\gamma$]{}-ray emission is fundamental to unambiguously detect this more exotic signal.
For these reasons, clusters of galaxies have been primary targets for $\gamma$-ray observatories. Yet, despite numerous efforts, unambiguous detection of extended $\gamma$-ray signal from the intra-cluster medium is lacking. Upper limits on the emission from individual galaxy clusters have been obtained from the analysis of space-based observations, including the EGRET data [@2003ApJ...588..155R] and, subsequently, the first 18 months of [*Fermi*]{} Large Area Telescope (LAT) data [@2010ApJ...717L..71A], and from ground-based observations in the energy band above 100 GeV (for a complete list of references, see the Introduction in [@Ahnen:2016qkt]). The lack of detection has paved the way for the stacking approach that has been adopted in the analyses of the most recent [*Fermi*]{}-LAT data releases [@Zimmer:2011vy].
In [@Dutson:2012ra], they have stacked $\gamma$-ray data at positions taken from an X-ray flux-limited sample of clusters, further selecting objects with a core-dominated brightest cluster galaxy with high radio flux. [@Huber:2013cia] have performed a stacking considering 53 clusters selected from the HIFLUGCS catalog [@2002ApJ...567..716R]. In [@Griffin:2014bra], they have analyzed 78 richest nearby clusters in the Two Micron All-Sky Survey cluster catalog. These searches found no evidence for a signal in the stacked data.
have analyzed 52.5-month [*Fermi*]{}-LAT data at the positions of the 55 X-ray galaxy clusters from the HIFLUGCS sample. Only the brightest objects have been considered in the analysis to maximise the chance of detecting signatures of the neutral pion decay which should scale with the X-ray flux. An excess has been detected with a statistical significance of 4.3 $\sigma$ within a radius of $\sim 0.25$ deg. However, several arguments suggest that the signal is mainly produced by AGN, with no evidence of a contribution from the intra-cluster material.
[@2014ApJ...787...18A] have similarly searched for a spatially extended $\gamma$-ray emission at the locations of 50 HIFLUGS X-ray galaxy clusters in the 4-year [*Fermi*]{}-LAT data, employing an improved LAT data selection (P7REP). They have detected a 2.7 $\sigma$ significant excess in a joint-likelihood analysis of stacked data. This signal, however, seems to be [produced by]{} three objects, Abell 400, Abell 1367 and Abell 311, and has been conservatively attributed to individual sources (radio galaxies) within the cluster rather than to genuine diffuse emission.
The $\gamma$-ray spectra of galaxy clusters have also been searched for monochromatic $\gamma$-ray features, but this characteristic signal has not been revealed in the joint likelihood analysis of different samples of clusters [@2016JCAP...02..026A; @Adams:2016alz]. These results have been used to place upper limits on the velocity-averaged DM cross section for self-annihilation into [$\gamma$]{} rays.
More specialized analyses have targeted nearby, individual objects such as the Coma and the Virgo clusters [@2016ApJ...819..149A; @2015ApJ...812..159A; @2014MNRAS.440..663Z]. The analysis of the Coma cluster in [@2016ApJ...819..149A] has revealed an excess emission within the cluster virial radius. However, its statistical significance is well below the threshold to claim detection of $\gamma$-ray produced by CRs interactions in the cluster. The analysis of the Virgo cluster in [@2015ApJ...812..159A] was mainly aimed at indirect DM detection. It has revealed an extended emission within a radius of 3 deg, which has been regarded as an artefact due to the incompleteness of the interstellar emission model. These results were used to set an upper limit on the $\gamma$-ray flux produced by both CRs interaction with the intra-cluster medium and DM annihilation.
The goal of this work is to expand the galaxy clusters analyses described above along three directions. Firstly we consider three different, non X-ray selected galaxy cluster samples. Namely: [*i)*]{} the redMaPPer catalog consisting of clusters identified through the red-sequence matched-filter Probabilistic Percolation cluster finder applied to the Sloan Digital Sky Survey (SDSS) DR8 [@2014ApJ...785..104R], [*ii)*]{} the WHL12 catalog [@2012ApJS..199...34W; @2015ApJ...807..178W] obtained from the SDSS DR12 [@2015ApJS..219...12A] and [*iii)*]{} the Planck catalog of Sunyaev-Zeldovich (PlanckSZ) clusters [@2015arXiv150201598P]. Secondly, instead of stacking signals, we cross-correlate cluster positions in the three catalogs with [*Fermi*]{}-LAT data [and compute 2-point statistics ]{} both in configuration and Fourier space. Combining the information from the 2-point angular cross-correlation function (CCF) and the cross angular power spectrum (CAPS) allows one to reduce the impact of systematic errors that may affect the stacking analysis which, in any case, we also perform to corroborate our results. We will follow [@xia11] and [@Xia:2014] for the cross-correlation measurements. Thirdly, we shall interpret the detected signal in the framework of the halo model along the line pursued in [@Regis:2015zka; @2015ApJS..221...29C] (see, e.g., [@Cooray:2002dia] for a review).
The paper is organized as follows. Section \[sec:data\] describes the [*Fermi*]{}-LAT data and the catalogs of clusters employed in the analysis. The measurement of the angular cross-correlation and of the stacked profiles is presented in Section \[sec:meas\]. Section \[sec:models\] introduces the theoretical models which are then compared to data in Section \[sec:res\]. We draw our conclusions in Section \[sec:concl\].
![Sky-maps of the cluster counts per pixel for the WHL12 (upper panel), redMaPPer (central panel), and PlanckSZ (lower panel) catalogs of clusters. The images have been smoothed to a resolution of $1^\circ$ to improve the visualization. The maps are shown in Mollweide projection and in Galactic coordinates. []{data-label="fig:maps"}](WHL12_z_smooth.pdf "fig:"){width="40.00000%"} ![Sky-maps of the cluster counts per pixel for the WHL12 (upper panel), redMaPPer (central panel), and PlanckSZ (lower panel) catalogs of clusters. The images have been smoothed to a resolution of $1^\circ$ to improve the visualization. The maps are shown in Mollweide projection and in Galactic coordinates. []{data-label="fig:maps"}](RedMapper_smooth.pdf "fig:"){width="40.00000%"} ![Sky-maps of the cluster counts per pixel for the WHL12 (upper panel), redMaPPer (central panel), and PlanckSZ (lower panel) catalogs of clusters. The images have been smoothed to a resolution of $1^\circ$ to improve the visualization. The maps are shown in Mollweide projection and in Galactic coordinates. []{data-label="fig:maps"}](Planck_SZ_Clusters.pdf "fig:"){width="40.00000%"}
Data {#sec:data}
====
In this work we use four different datasets: measurements of high-latitude diffuse [$\gamma$]{}-ray emission, and three different galaxy cluster catalogs. All of them are described below.
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
Fermi-LAT {#sec:fermidata}
---------
[*Fermi*]{}-LAT is the primary instrument onboard the [*Fermi*]{} Gamma-ray Space Telescope launched in June 2008 [@2009ApJ...697.1071A]. It is a [$\gamma$]{}-ray pair-conversion telescope covering the energy range between 20 MeV and $\sim$ 1 TeV. Due to its excellent angular resolution ($\sim 0.1^{\circ}$ above 10 GeV), large field of view ($\sim 2.4$ sr), and very efficient rejection of background from charged particles, it is currently the best experiment to investigate the nature of the extra-galactic [$\gamma$]{}-ray background [EGB, @2015IGRB] in the GeV energy range.
For our analysis, we have used 78 months of data from August 4th 2008 to January 31th 2015 ([*Fermi*]{} Mission Elapsed Time 239557418 s - 444441067 s), considering the Pass 8 event selection[^1]. Furthermore, to reduce the contamination from the bright Earth limb emission, we exclude photons detected with measured zenith angle larger than 100$^{\circ}$. In order to generate the final flux maps we have produced the corresponding exposure maps using the standard routines from the LAT *Science Tools*[^2] version 10-01-01, and the Pass 8 CLEAN event class, namely the [`P8R2_CLEAN_V6`]{} instrument response functions (IRFs). We use both back-converting and front-converting events. The GaRDiAn package [@FermiLAT:2012aa; @Ackermann:2009zz] was adopted to pixelize both photon count and exposure maps in HEALPix[^3] format [@2005ApJ...622..759G]. The maps contain $N_{\rm pix} = 12, 582, 912 $ pixels with mean spacing of $\sim 0.06^{\circ}$ corresponding to the HEALPix resolution parameter $N_{\rm side}=1024$. Finally, the flux maps are obtained by dividing the count maps by exposure maps. We perform the bulk of our analysis in the three separate energy intervals: $0.5<E<1$ GeV, $1<E<10$ GeV and $10<E<100$ GeV. For a more accurate study of the spectral dependence we will also use a finer energy binning, with 8 bins [cut at]{} 0.25, 0.5, 1, 2 , 5, 10, 50, 200, 500 GeV.
Our analysis is focused on the unresolved [$\gamma$]{}-ray background (UGRB), i.e., the unresolved EGB emission left after subtracting resolved point sources [@2015IGRB]. [To obtain such maps we mask out]{} the [$\gamma$]{}-ray point sources listed in the 3FGL catalog [@3FGL]. More precisely, we mask the 500 brightest point sources (in terms of the integrated photon flux in the 0.1-100 GeV energy range) with a disc of radius $2^\circ$, and the remaining ones with a disc of $1^\circ$ radius. The Small and Large Magellanic Clouds, which are extended sources, are masked with discs of $3^{\circ}$ and $5^{\circ}$ radius respectively. To reduce the impact of the Galactic emission we apply a Galactic latitude cut masking the region with . In [@xia11] we have experimented with different latitude cuts and found that represents the best compromise between pixels statistics and Galactic contamination. We also exclude the regions associated to the [*Fermi*]{} Bubbles and the Loop I structures [as in [@xia11]]{}.
The Galactic diffuse emission can be still significant at the high Galactic latitude used in our analysis and needs to be removed. For this purpose, we use the model of Galactic diffuse emission `gll_iem_v06.fits`[^4], which we subtract from the observed emission to obtain the [*cleaned*]{} [$\gamma$]{}-ray maps. The model, together with an isotropic template, is convolved with the IRFs and fitted to the photon data in each energy bin and in our region of interest using `GaRDiAn`. The best fit diffuse plus isotropic model is subtracted from the count map and this residual count map is further divided by the exposure to give the final residual flux map to be analyzed for the given energy bin.
As the Galactic diffuse emission model is not exact, cleaning is not perfect and the residual flux maps are still contaminated by spurious signals, especially on large angular scales. [However, and this is the main advantage of our analysis, the cross-correlation analyses are expected to be almost immune to these contaminations ]{} since Galactic foreground emission is not expected to correlate with the extragalactic signal that we want to investigate. Nonetheless, [to minimize the chance of systematic errors we adopt a conservative approach and, following [@xia11] and [@Xia:2014]]{} we apply [a further]{} cleaning procedure that, using HEALPix tools, removes all contributions from multipoles up to $\ell=10$.
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
![Observed CAPS (PSF deconvolved) between the [*Fermi*]{}-LAT [$\gamma$]{}-ray map at $E_\gamma>1$ GeV and the PlanckSZ catalog.[]{data-label="fig:data_Planck"}](CAPS_measurements_PlanckSZ.pdf){width="40.00000%"}
Galaxy cluster catalogs {#sec:catadata}
-----------------------
The three catalogs of galaxy clusters considered in our analysis are: [*1)*]{} SDSS redMaPPer [@2014ApJ...785..104R], [*2)*]{} WHL12 [@2012ApJS..199...34W; @2015ApJ...807..178W] and [*3)*]{} PlanckSZ [@2015arXiv150201598P].
1. [Clusters in the [*redMaPPer*]{} catalog are identified using red-sequence galaxies. Adopting these objects as clusters signposts increases the contrast between cluster and background galaxies in color space, thus enabling accurate and precise photometric redshift estimates. In our analysis we consider all 26,350 clusters detected in the redshift range $0.08 < z < 0.55$ over the $\sim$10,400 sq deg area of the SDSS Data Release 8. Photo-$z$ errors are nearly Gaussian with an amplitude $\sigma_z \sim 0.006$ at $z \sim 0.1$, that increases to $\sigma_z \sim 0.02$ at $z \sim 0.5$. Details on the cluster detection procedure and on the iterative method to estimate photometric redshifts can be found in [@2014ApJ...785..104R].]{}
2. [[*WHL12*]{} updates and refines on the [@2009ApJS..183..197W] galaxy cluster catalog from SDSS III [@2011ApJS..193...29A].]{} The original catalog contained 39,668 clusters with photometric redshifts. The new catalog was obtained by [@2012ApJS..199...34W] by applying an improved cluster detection method [@2011ApJ...734...68W] to SDSS-III galaxies, exploiting $\sim$1.35 millions Large Red Galaxies with spectroscopic redshifts in the SDSS 12th Data Release (DR12). The updated WHL12 catalog [@2015ApJ...807..178W] that we considered in our analysis has 158,103 clusters in the the range $0.05<z<0.8$. Its completeness is larger than 95% for objects with mass $M_{200} > 1.0 \times 10^{14}\, M_\odot$ in the redshift range of $0.05 < z < 0.42$, decreasing at higher redshifts.
3. [The [*PSZ2*]{} Second Planck Sunyaev-Zed’dovich catalog contains SZ-selected clusters. It is based on the full 29 month mission data [@2015arXiv150201598P]. The methodology employed to detect clusters refines the one used to produce the first Planck SZ cluster catalog. PSZ2’s 1,653 clusters, distributed across 83.6% of the sky, are the union of outcomes from three cluster detection codes . It contains 1653 detections, of which 1203 are confirmed to be clusters with identified counterparts in external datasets and with a purity larger than 83%. PSZ2 probes clusters at relatively low redshift with the distribution peaking at $z\sim0.2$ and extending up to $z\lesssim0.5$. The catalog also provides estimates for the mass of the SZ clusters as a function of redshift.]{}
For each of the three cluster catalogs we built a HEALPix skymap with resolution parameter $N_{\rm side}=1024$ specifying the cluster counts per pixel $n(\hat{\Omega}_i)$. The three maps are shown in Fig. \[fig:maps\]. Clusters are counted as single objects, i.e. no statistical weight has been used to account for the cluster mass or selection effects. The cross-correlation analysis is then performed between the normalized count maps $n(\hat{\Omega}_i)/\bar{n}$, [where $\bar{n}$ is the mean cluster density in the unmasked area]{}, and the [*Fermi*]{}-LAT residual flux sky-maps. For the two SDSS-based cluster catalogs we use the standard SDSS mask, i.e., the contour of the sky region covered by RedMaPPer which can be seen in Fig. \[fig:maps\]. Further, we conservatively mask also the disconnected south-galactic region. For the PSZ2 catalog no mask is used since the masked area is included in the one applied to the Fermi maps.
To better understand the catalogs properties, we show in Fig. \[fig:cat\_prop\] the histograms of the distributions of redshift, mass and angular size of the clusters in the three catalogs. We note already here that the vast majority of the considered clusters are effectively point-like for the [*Fermi*]{}-LAT telescope, namely, their angular size is smaller than the point-spread function (PSF) of [*Fermi*]{}-LAT (see right panel).
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
Measuring 2-point cross-correlation statistics. {#sec:meas}
===============================================
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
![Approximate significance of the stacking signal (see text for the definition) as a function of the number of clusters in the stacking for the WHL12 (green), redMaPPer (red) and PlanckSZ (blue) catalogs. This plot is for illustrative purposes only and is not used for any quantitative analysis. []{data-label="fig:stack_sig"}](stack_significances.pdf){width="49.00000%"}
Cross-correlation APS {#sec:xcorr}
---------------------
[In our analysis we estimate both the 2-point CCF and the cross APS. In both cases we used PolSpice,[^5],]{} a publicly available toolkit to estimate the angular $CCF^{(\gamma c)}(\theta)$ and the CAPS $\bar C_\ell^{(\gamma c)}$ of any two datasets pixelized in HEALPix format [@szapudi01; @chon04; @efstathiou04; @challinor05]. [[@Xia:2014] have tested the reliability and robustness of the PolSpice estimator in similar analyses.]{} PolSpice also estimates the covariance matrix $\bar V_{\ell\ell'}$ of the different multipoles taking into account the correlation effect induced by the mask.
In general, other, and possibly more rigorous, statistical techniques (see, e.g., [@Baddeley; @Pewsey]) can be applied. However, for the specific problem under investigation (affected by large uncertainties), they are expected not to affect significantly the results as they could do in the case of a precision test.
The CAPS estimated from PolSpice [include the effects of the instrument PSF and pixelization]{}. We deconvolve the results from these effects as in [@Xia:2014]: Firstly we compute the beam window function $W_\ell^{B}$ associated to the PSF, and the pixel window function $W_\ell^{\rm pixel}$ associated to the pixelization. Then, we derive the deconvolved CAPS $C_\ell^{(\gamma c)}$ from the measured ones $\bar C_\ell^{(\gamma c)}$ as $C_\ell^{(\gamma c)}=(W_\ell)^{-1}\,\bar C_\ell^{(\gamma c)}$, where $W_\ell = W_\ell^{B} W_\ell^{\rm pixel} $ is the global window function. The covariance matrix of the deconvolved $C_\ell^{(\gamma c)}$ is then given by $V_{\ell\ell'} = \bar V_{\ell\ell'} W_\ell^{-2}W_{\ell'}^{-2}$. Finally, since the mask induces a strong correlation in nearby multipoles we bin the CAPS measurements into 12 equally spaced logarithmic intervals in the range $\ell\in[10,2000]$. We choose logarithmic bins to account for the rapid loss of power at high $\ell$ induced by the PSF. In what follows we omit the square bracket in the $[\ell]$ subscript and use $C_\ell^{(\gamma c)}$ also to indicate the binned CAPS. In our analysis we mainly focus on the binned quantity and should be clear from the context when, instead, we consider the un-binned CAPS. The $C_\ell^{(\gamma c)}$ in each bin is given by the simple unweighted average of the $C_\ell^{(\gamma c)}$ within the bin. For these binned $C_\ell^{(\gamma c)}$ it is also possible to build the corresponding block covariance matrix as $\sum_{\ell\ell'} V_{\ell\ell'}/\Delta\ell/\Delta\ell'$, where $\Delta\ell, \Delta\ell'$ are the width of the two multipoles bins, and $\ell, \ell'$ run within the multipoles of the first and second multipole bin.
We verified that the binning is very efficient in removing correlation among nearby multipoles, resulting in a block covariance matrix that is, to a good approximation, diagonal. For this reason we have neglected the off-diagonal terms in our analysis.
The results of the various CAPS measurements are shown in Figs. \[fig:data\] and \[fig:data\_Planck\]. In Fig. \[fig:data\] we show the PSF-deconvolved and binned CAPS of the redMaPPer (red) and WHL12 (green) catalogs in threee energy bins: $500\,{\rm MeV}<E_\gamma<1\,{\rm GeV}$ (left panel), $1\,{\rm GeV}<E_\gamma<10\,{\rm GeV}$ (central) and $10\,{\rm GeV}<E_\gamma<100\,{\rm GeV}$ (right). The bars represent 1 $\sigma$ errors from the diagonal elements of the covariance matrix.
We clearly detect a non-zero cross-correlation signal. We quantify the local statistical significance of the detection in each energy bin in terms of number of sigmas $N_\sigma=\sqrt{\sum_j (C_{\Delta\ell_j}^{(\gamma c)}/\delta C_{\Delta\ell_j}^{(\gamma c)})^2}$ with $C_{\Delta\ell_j}^{(\gamma c)}$ the measured CAPS in the $j$ bin and $\delta C_{\Delta\ell_j}^{(\gamma c)}$ its uncertainty quoted in the figure. For WHL12 we find: $N_\sigma=3.7, \, 4.4$ and 2.9 in the three energy bins. The significance for redMaPPer is very similar: $N_\sigma=3.3, \, 5.0$ and 2.7.
The results for PlanckSZ are shown in Fig. \[fig:data\_Planck\]. Given the limited number of clusters in this sample the statistics is too poor to divide results into energy bins. Therefore, we have considered a single bin $E_\gamma>1$ GeV. Also in this case we find a non-negligible cross correlation signal with $N_\sigma=3.7$.
We note that the amplitude of the measured CAPS is the highest for PlanckSZ and the lowest for WHL12, with redMaPPer being in between. This was somehow expected since the average mass of the clusters in the catalog is the largest in PlanckSZ and the smallest in WHL12 (see Fig. \[fig:cat\_prop\]b). Since the redshift distribution of the clusters is not dramatically different, larger mass (typically) implies higher [$\gamma$]{}-ray emission and higher catalog bias, and therefore higher CAPS amplitude.
CCF and Stacked signals {#sec:stack}
-----------------------
[The CCF statistics provides a complementary information to CAPS. Here we estimate CCF from the CAPS as follows:]{} CCF\^[(c)]{}() = \_|C\_\^[(c)]{} P\_, \[eq:2point\] where $\theta$ is the angular separation in the sky and $P_\ell$ are the Legendre polynomials.
The resulting CCFs from the three cluster catalogs are shown Fig. \[fig:CCF\_stack\]. [We show the $1 < E_\gamma < 10$ GeV case only. Unlike the CAPS case, here the CCFs [*are not de-convolved*]{} for the PSF and angular pixels. The reason for this is that unlike the Fourier space case, in which deconvolution is a simple multiplication, deconvolution in configuration space is more unstable if, like in our case, CAPS has large power at high multipoles $\ell$.]{}
The CCF analysis is quite similar to stacking the [$\gamma$]{}-ray signal at clusters’ locations (as we will show at the end of the Section). Since stacking analyses have been popular in previous studies we decided to also perform this type of analysis. In our procedure we first select a region of [4 degrees]{} of radius around the position of each cluster in the three catalogs. Then, we sum the [$\gamma$]{}-ray flux in the circular areas with no attempt to rescale the signal to the object’s properties (e.g. richness). Each image in the stacking is randomly rotated, in order to better investigate the impact of the region around the centered cluster. A potential issue is that the stacked images might not be independent, since they come from partially overlapping fields. This is most severe in the WHL12 catalog in which the mean angular separation of clusters is $\sim 0.25^\circ \ll 4 ^\circ$ radius size. For this reason we will not attempt to perform a quantitative statistical comparison with the results of the CCF analysis.
The stacked images for [$\gamma$]{}-rays in the energy range $1 < E_\gamma < 10$ GeV are shown in Fig. \[fig:stack\_im\]. We see a clear [$\gamma$]{}-ray excess at the clusters’ center. From each image we have obtained a stacked [$\gamma$]{}-ray emission profile $y^{(\gamma c)}(\theta)$ by averaging the stacked flux in logarithmically-spaced circular annuli ignoring possible small anisotropies that may survive the stacking procedure. The [$\gamma$]{}-ray stacked profiles for the clusters in the three catalogs are shown in Fig. \[fig:CCF\_stack\]. The reported error bars are the image-by-image scatter around the stacked flux in each annulus. With the same definition of errors, we show how the stacking signal builds up as the number of clusters increases in Fig. \[fig:stack\_sig\]. We defined an approximate significance given by $\sqrt{\sum_i\,(y_i/\sigma_i)^2}$, where $y_i$ is the measured stacked emission in the annulus $i$ and $\sigma_i$ is the scatter. On the other hand, these approximations are likely to underestimate the real errors and to overestimate the significance, since rely on the assumption that all stacked fields are independent, while in reality they are not, as already mentioned. Therefore both Fig. \[fig:CCF\_stack\] and the insets of Fig. \[fig:stack\_sig\] are shown only to illustrate the trends, and will not be used for any quantitative analyses.
If the [$\gamma$]{}-ray emission is circularly symmetric, the stacked profile $y^{(\gamma c)}(\theta)$ (derived with the procedure outlined above) is an estimator of $\langle \delta_c(0)\,I_\gamma(\theta)\rangle$, where $\delta_c$ is the cluster fluctuation field. This is actually the definition of the CCF, i.e. $CCF^{(\gamma c)}(\bm \theta)= \langle \delta_c(\bm \theta')\,I_\gamma(\bm \theta'+\bm \theta)\rangle=\langle \delta_c(0)\,I_\gamma(\theta)\rangle$. Therefore, in the small angle limit, where one can assume the dominant contribution to come from a single object and the circular symmetry to hold, the two quantity $CCF^{(\gamma c)}$ and $y^{(\gamma c)}(\theta)$ are perfectly equivalent. [^6]
The similarity between the two quantities is indeed evident from Fig. \[fig:CCF\_stack\]. The nice agreement between CCFs and stacking profiles is an important cross-check for our analysis, since the two measurements have been obtained employing two completely different methods.
Models {#sec:models}
======
In the Limber approximation [@1953ApJ...117..134L], the CAPS between [$\gamma$]{}-ray emitters and galaxy clusters can be written as: C\_\^[(c)]{}= W\_() W\_[c]{}()P\_[c]{}(k=/,), \[eq:clfin\] where $\chi(z)$ denotes the radial comoving distance. $W_c(\chi)$ and $W_{\gamma}(\chi)$ are the window functions that characterize the distribution of clusters and [$\gamma$]{}-ray emitters along the line of sight, respectively. $P_{\gamma c}(k,z)$ is the 3D cross power spectrum (PS) at the redshift $z$, $k$ is the modulus of the wavenumber and $\ell$ denotes the angular multipole. The relation $\chi(z)$ is fully specified by the expansion history of the Universe $H(z)$: $d\chi=c\,dz/H(z)$.
In the following, galactic [$\gamma$]{}-ray astrophysical emitters are denoted by the symbol $\gamma_i$, where $i$ indicates the type of source, and we consider the ones that are known to significantly contribute to the UGRB signal: blazars, misaligned AGN \[mAGN\] and star forming galaxies \[SFGs\]. The symbol $\gamma_c$ refers to [$\gamma$]{}-ray emission from the intra-cluster medium \[ICM\], while $c_j$ denotes cluster catalogs ($j=$ redMaPPer, WHL12 or PlanckSZ).
The ingredients of our model are the cross-PS and the window functions entering in the computation of Eq. (\[eq:clfin\]). They are described in the next two subsections. Predictions of Eq. (\[eq:clfin\]) will be then compared with the measured CAPS shown in Figs. \[fig:data\] and \[fig:data\_Planck\].
Window functions {#sec:wf}
----------------
The window functions of the three galactic [$\gamma$]{}-ray sources considered here (blazars, mAGNs and SFGs) are presented in [@2015ApJS..221...29C] (Appendix). The collective [$\gamma$]{}-ray emission from ICM in the Universe gives raise to the window function: W\_[\_c]{}(z)=\_[M\_[c,min]{}]{} [d]{}M, \[eq:wf\_clug\] where $d^2n/dM\,dV$ is the cluster mass function predicted by the ellipsoidal collapse model [@Sheth:1999mn] and $\mathcal{L}_{\gamma_c}$ is the cluster [$\gamma$]{}-ray luminosity per unit energy range. The integral is over the cluster masses above $10^{13.8}\,h^{-1}\,M_\odot$ . The relation between the halo and the cluster virial mass $M_{500}$ is specified in [@2003ApJ...584..702H] (Appendix A). For the luminosity we adopt the empirical relation of : L\_[\_c]{}(100[MeV]{})=A\_[\_c]{}()\^[5/3]{}[s\^[-1]{}GeV\^[-1]{}]{}, \[eq:lumg\_clu\] where $L_{\gamma_c}=\mathcal{L}_{\gamma_c}/E$ and $A_{\gamma_c} \lesssim10^{21}$ is a normalization parameter constrained by the non-detection of [$\gamma$]{}-rays from nearby clusters . Finally, we assume a power-law ICM energy spectrum with index $\alpha_{\gamma_c}=2.2$.
The window function of cluster counts is: W\_[c\_j]{}(z)=[d]{}M, \[eq:wf\_cluc\] where: N\_[c\_j]{}=[d]{}MdV is the number of object in the $j$-th cluster catalog. To model the cluster mass function we adopted two different approaches.
For the redMaPPer and WHL12 clusters we measure the cluster richness distribution $d^2n_{c_j}/{d\lambda dz}$ from the catalogs, assume a lognormal distribution for the cluster richness $\lambda$, $P(\lambda|M_{500})$, with $\lambda(M_{500})$ taken from [@2015MNRAS.454.2305S] (redMaPPer) and [@2015ApJ...807..178W] (WHL12) and derive the cluster mass function as a function of redshift as: =d P(|M\_[500]{}) . Finally, we obtain the mass function per unit volume $d^2n_{c_j}/{dV dM}$ by specifying the cosmology-dependent relation: = . We checked that the derived cluster mass function is in good agreement with the theoretical model of [@Sheth:1999mn], in the relevant mass and redshift ranges, once selection effects and completeness are taken into account.
For the PlanckSZ clusters we estimate the mass function using directly the masses and the redshifts of the objects in the catalog.
![Measured (points) and predicted (lines) autocorrelation APS for the catalogs redMaPPer (red), WHL12 (green) and PlanckSZ (blue).[]{data-label="fig:APS_catalogs"}](APS_catalogs.pdf){width="49.00000%"}
Three dimensional power spectrum {#sec:ps}
--------------------------------
To model the three dimensional cross power spectrum we use the halo model and write the PS as a sum of the one-halo and two-halo terms $P=P^{1h}+P^{2h}$. Below we provide the expressions for $P^{1h}$ and $P^{2h}$. The interested reader can find a detailed discussion in [@Fornengo:2013rga]. We also notice that in some equation the redshift dependence is not explicitly reported.
For the cross correlation between point-like astrophysical [$\gamma$]{}-ray emitters and clusters we have: & & P\_[c\_j,\_i]{}\^[1h]{}(k,z) = \_[\_[min,i]{}(z)]{}\^[\_[max,i]{}(z)]{} [d]{}\_i(,z)\
& & \[eq:PSBd1\]\
& & P\_[c\_j,\_i]{}\^[2h]{}(k,z) =\
& &P\^[lin]{}(k).\[eq:PSBd2\] Both terms depend on the luminosity function of the emitter, $\Phi_i$ and the mean luminosity density $\langle f_{\gamma_i}\rangle = \int {\mathrm d}\mathcal{L} \, \mathcal{L} \, \Phi_i(\mathcal{L},z)$, whereas the linear mass power spectrum $P^{\rm lin}(k)$ and the bias $b_{\gamma_i}$ only enter the two-halo term. For the bias we adopt a simple linear model and assume that the bias of the emitter is equal to that of its halo host $b_{\gamma_i}(\mathcal{L})=b_h(M(\mathcal{L}))$ modeled according to [@Sheth:1999mn]. For the relation between the mass of the halo host and the luminosity of the emitter, $M(\mathcal{L})$, we adopt the one derived by [@Camera:2014rja]. The effective halo occupation of clusters $\langle N_{c_j}\rangle=(dn_{c_j}/dM)/(dn/dM)$ is obtained from the cluster mass functions $dn_{c_j}/dM$ used in Section \[sec:wf\]. In this way, we account for selection effects and completeness of the catalogs. The average number density of clusters at a given redshift is given by $\bar n_{c_j}(z)=\int dM\, \langle N_{c_j}\rangle\,dn/dM$. Note that Eq. (\[eq:PSBd1\]) does not depend on the wavenumber $k$. It describes the picture of point-like [$\gamma$]{}-ray emitters located at the center of the clusters. Being flat, it acts as a shot-noise-like term.
[[@2015ApJS..221...29C] have shown that this halo model is not sufficient to describe the effect of the [*Fermi*]{}-LAT PSF that creates an additional shot-noise-like term on small-scales, which is not captured by the above equations. Quantifying the amplitude of this effect is not straightforward. However, since we know it is scale-independent, we can model it empirically by adding an extra, shot-noise-like constant term in the fit of the measured $C_{\ell}^{(\gamma c)}$. Therefore, following [@2014JCAP...10..061A] and [@2015ApJS..221...29C] we include one additional free parameter for each combination of cluster catalog and [$\gamma$]{}-ray source.]{}
We note also that the 1-halo term model above assumes that the relation $M(\mathcal{L})$ is deterministic. We argue that ignoring the scatter in the relation does not significantly affect our results since the 1-halo term is small (blazars, mAGN and SFG reside in halos typically smaller than the cluster size) and subdominant with respect to the shot-noise term.
For the cross correlation between [$\gamma$]{}-ray emission from the ICM and clusters we have: & & P\_[c\_j,\_c]{}\^[1h]{}(k,z) = \_[M\_[c,min]{}]{}\^[M\_[max]{}]{} [d]{}M \
& & P\_[c\_j,\_c]{}\^[2h]{}(k,z) = \[eq:PSclcl2\]\
& & P\^[lin]{}(k) ,where now the luminosity density is $\langle f_{\gamma_c}\rangle = \int {\mathrm d}M \,dn/dM\, \mathcal{L}/\bar\rho$, and $\tilde v_\delta(k|M)$ is the Fourier transform of the normalized halo density profile $\rho_h(\bm x|M)/\bar \rho_{DM}$, that we assume to have a NFW shape [@Navarro:1996gj]. The underlying assumption is that the [$\gamma$]{}-ray emission from the ICM has the same profile of the host halo (in practice, this is not a crucial assumption, since in the current analysis we do not probe scales smaller than the typical size of a cluster).
Unlike in the previous case, uncertainties in the 1-halo term cannot be ignored. They stem from the fact that no extended [$\gamma$]{}-ray emission from clusters has been unambiguously detected and, consequently, no observational constraint exists for the $\mathcal{L}_{\gamma_c}(M)$ relation. To account for this potential source of systematic error, we again include an additional constant term when we fit the cross-correlation model to the data.
![CAPS between redMaPPer clusters and [$\gamma$]{}-ray emitters (convolved with beam window function $W_\ell^{B}$) compared to the measurement at $E_\gamma>1$ GeV. Dotted line shows a noise term at the level of $C_\ell^{(\gamma c)}=2\cdot10^{-12}{\rm cm^{-2}s^{-1}}$.[]{data-label="fig:CAPS_redMA"}](CAPS_redMA.pdf){width="49.00000%"}
![CAPS between WHL12 clusters and [$\gamma$]{}-ray emitters (convolved with beam window function $W_\ell^{B}$) compared to the measurement at $E_\gamma>1$ GeV. Dotted line shows a noise term at the level of $C_\ell^{(\gamma c)}=7\cdot10^{-13}{\rm cm^{-2}s^{-1}}$.[]{data-label="fig:CAPS_WHL12"}](CAPS_WHL12.pdf){width="49.00000%"}
Results {#sec:res}
=======
Before comparing model and data we performed a sanity check in which we use our model to predict the angular cluster-cluster power spectra and compare it with the measured auto angular power spectrum for each of the three cluster catalogs. The results are shown in Fig. \[fig:APS\_catalogs\]. The agreement is remarkably good except at very large or very small angular scales where, respectively, selection effects and model uncertainties are larger.
The results of the comparisons between measured and predicted cross-correlation between [$\gamma$]{}-ray emitters (blazars, mAGN, SFG and ICM) and clusters (redMaPPer, WHL12 and PlanckSZ) are shown in Figs. \[fig:CAPS\_redMA\], \[fig:CAPS\_WHL12\] and \[fig:CAPS\_Planck\]. Dots with errorbars refer to the data whereas the different curves represent predictions obtained using the models described in Section \[sec:models\]. The latter have been normalized (without changing neither spectral nor spatial shapes) such that blazars, mAGNs and SFGs individually contribute to 100% of the UGRB above 1 GeV. For the ICM case we consider a 1 % contribution to meet the observational constraints described in Section \[sec:wf\].
At $\ell\lesssim100$, the data are well fitted by a model in which the [$\gamma$]{}-ray emission is produced by blazars, mAGNs or SFGs (or a combination of them), provided that they contribute to 100 % of the UGRB. On the contrary, the ICM contribution to the cross correlation is highly subdominant, largely because of the 1 % UGRB contribution constraint. Note however that the ICM would account for much more than 1 % of the CAPS (while blazars, mAGNs or SFGs providing 100 % of the UGRB contribute to a similar fraction, roughly 100 %, of the CAPS). This is because, when compared to galactic [$\gamma$]{}-ray emitters, the ICM contribution has a relatively larger non-linear term and its window function have a better overlapping in redshift with the catalogs window functions. Though small, the amplitude of the ICM CAPS is not the same in all panels. Instead it correlates with the average mass of the clusters in the catalogs. Indeed, one can notice how it increases going from WHL12 to PlanckSZ (the catalogs with, respectively, the smallest and largest average mass, see Fig. \[fig:cat\_prop\]b), in particular in comparison with the milder increase expected in the amplitudes associated to the other [$\gamma$]{}-ray emitters. In the case of PlanckSZ catalog, that contains massive clusters, the ICM contribution is larger than that of the other astrophysical sources at $\ell \gtrsim100$.
The similarity in the theoretical CAPS of all the above [$\gamma$]{}-ray emitters, except the ICM, guarantees the fact that all of them provide a good fit to the data and reflects the similarities in the [$\gamma$]{}-ray window functions. The small differences originates from two effects: the (relatively small) differences in the shapes of the [$\gamma$]{}-ray window functions (see, e.g., the Supplemental Material in [@Regis:2015zka]) and the different redshift dependences of the bias factors of the various emitters (see, e.g., Appendix of [@2015ApJS..221...29C]).
![CAPS between PlanckSZ clusters and [$\gamma$]{}-ray emitters (convolved with beam window function $W_\ell^{B}$) compared to the measurement at $E_\gamma>1$ GeV. Dotted line shows a noise term at the level of $C_\ell^{(\gamma c)}=5\cdot10^{-12}{\rm cm^{-2}s^{-1}}$.[]{data-label="fig:CAPS_Planck"}](CAPS_PlanckSZ.pdf){width="49.00000%"}
At $\ell\gtrsim100$ the models systematically underestimate the measured CAPS. As we have discussed, the validity of our model is expected to break down at small scale. For this reason we included a shot-noise-like term that restores the missing power at high $\ell$. This term, convolved with the [*Fermi*]{}-LAT beam window function, is shown with black dot-dashed lines in Figs. \[fig:CAPS\_redMA\], \[fig:CAPS\_WHL12\] and \[fig:CAPS\_Planck\]. Its introduction clearly improves the quality of the model that now fits the data all the way out to $\ell=1000$.
In principle, one could account for this term by increasing the normalization of the ICM window function by one order of magnitude. Indeed, the shape of the CAPS ICM is similar to the signal shape over the whole multiple range. On the other hand, this increase would make the ICM contribution to the UGRB to grow to $\sim$10%, something which conflicts with current bounds.
The different [$\gamma$]{}-ray emitters have different energy spectra, and we can have a better insight on the origin of the cross-correlation signal by repeating the correlation analysis in finer energy bins. We considered eight of them: $[0.25,0.5] \; [0.5,1.0] \; [1.0,2.0] \; [2.0,5.0] \; [5.0,10.0] \; [10.0,50.0] $ $\; [50.0,200.0] \; [200.0,500.0] $ GeV. Building upon the results of the previous analysis we now move on from the specific benchmark models adopted above and consider a simpler CAPS model: C\_[,a]{}\^[(c)]{}=C\^[1h]{}\_a+A\^[2h]{}\_a C\_[,a]{}\^[2h]{} , \[eq:simpmodel\] with $a=1,..,8$ running over the energy bins[^7], $C^{1h}_a$ is the parameter of the fit related to the 1-halo contribution, $C_{\ell,a}^{2h}$ is the 2-halo model term that we have modeled according to the blazar case since, as we have seen above, the other [$\gamma$]{}-ray emitters produce a similar cross-correlation signal apart from the ICM that, again according to the previous analysis, is subdominant. $A^{2h}_a$ is the second free parameter of the fit and sets the amplitude of the 2-halo term. We remind the reader that, in the halo model, the CAPS is generically given by the sum of a 1-halo and a 2-halo terms, as described in Section \[sec:models\]; in Eq. (\[eq:simpmodel\]), the 1-halo CAPS is constant (since the source is assumed to be point-like) and the 2-halo CAPS is computed plugging Eq. (\[eq:PSBd2\]) into Eq. (\[eq:clfin\]).
The null hypothesis (no signal) can be discarded at high significance. Indeed, the $\Delta \chi^2$ with respect to this model is 43.2 (WHL12), 53.5 (redMaPPer), and 14.4 (PlanckSZ). Considering the number of parameters of the model (16 for WHL12 and redMaPPer and 12 for PlanckSZ) and following Wilks’ theorem [@Wilks:1938], this leads to $p$-values of $2.6\times10^{-4}$ (WHL12), $6.3\times10^{-6}$ (redMaPPer), and 0.27 (PlanckSZ). Note that these are conservative estimates of the significance of the signals, since one could reduce the degrees of freedom by fixing the energy spectrum of the model (that, as we will see in the following, is in fair agreement with expectations for blazars).
We show the derived fitting parameters in Figs. \[fig:CAPS\_spectrum1h\] and \[fig:CAPS\_spectrum2h\]. In Fig. \[fig:CAPS\_spectrum1h\], we plot the $C^{1h}$ term for the different energy bins. The global statistical evidence (i.e. adding up in quadrature the evidences of single energy bins) for the 1-halo component is $3.9\sigma$ (WHL12), $4.7\sigma$ (redMaPPer), and $2.3\sigma$ (PlanckSZ). Fig. \[fig:CAPS\_spectrum2h\] instead shows the 2-halo term $A^{2h}C_{\ell=80}^{2h}$ whose statistical evidence turns out to be $2.6\sigma$ (WHL12), $2.1\sigma$ (redMaPPer), and $1.8\sigma$ (PlanckSZ). The evidence in each energy bin for the 1-halo (2-halo) term (and in turn the error bars in Fig. \[fig:CAPS\_spectrum1h\] (Fig. \[fig:CAPS\_spectrum2h\])) have been derived by evaluating the likelihood ratio between the model in Eq. (\[eq:simpmodel\]) and the same model but with the 1-halo (2-halo) term set to zero.
The energy dependence of the 1-halo term in the redMaPPer case shows a possible break at $E=10$ GeV which suggests the presence of two contributions: a hard component with spectral index close to $-2$ above 10 GeV and a softer component at lower energies. The analysis of the other two catalogs displays a similar trend. However, the quality of the data does not allow us to draw statistically significant conclusions. A fit to the energy spectrum of the redMapper 1-halo CAPS with the sum of two power-laws is preferred over the single power-law case at 85% C.L., with the the best-fit spectral indexes being $-2.9$ and $-2.0$ (while $-2.7$ is the best-fit spectral index in the case of a single power-law). No break is seen in the energy-dependence of the 2-halo term with a spectral index of $\sim -2$. This is consistent with being produced by the same sources responsible for the 1-halo term at $E > 10$ GeV.
Knowing that BL Lac-type of blazar are characterized by hard energy spectra in contrast with the softer spectra of all other galactic [$\gamma$]{}-ray emitters (mAGN, FSRQ, SFG) considered in our model, the emerging picture is that of a cross correlation signal dominated by BL Lacs on large scale (where the 2-halo term dominates) and on small scales, but possibly only at high energy. The soft component, seen in the 1-halo term only, might indicate that the small-scale cross-correlation signal is also contributed by a different type of [$\gamma$]{}-ray emitters that takes over at $E < 10$ GeV. These can be non-BL Lac AGNs or SFGs hosted in the cluster halo, or the ICM itself (or a combination of them).
![Energy spectrum of a constant 1-halo CAPS. This has been derived by fitting the measurements in eight energy bins from 0.25 to 500 GeV with $C^{1h}+A_{2h}C^{2h}_\ell$, where $C^{1h}$ and $A_{2h}$ are the fitting parameters and $C^{2h}_\ell$ is taken to follow the blazar case.[]{data-label="fig:CAPS_spectrum1h"}](CAPS_spectrum1h_chi2.pdf){width="49.00000%"}
![Energy spectrum of the 2-halo CAPS term at $\ell=80$. This has been derived by fitting the measurements in eight energy bins from 0.25 to 500 GeV with $C^{1h}+A_{2h}C^{2h}_\ell$, where $C^{1h}$ and $A_{2h}$ are the fitting parameters and $C^{2h}_\ell$ is taken to follow the blazar case. The plot shows $A_{2h}C^{2h}_{80}$.[]{data-label="fig:CAPS_spectrum2h"}](CAPS_spectrum2h_chi2.pdf){width="49.00000%"}
Conclusions {#sec:concl}
===========
In this work, we have analyzed the 2-point angular cross-correlation between the unresolved EGB observed by [*Fermi*]{}-LAT and the number of galaxy clusters in three different catalogs: WHL12, redMaPPer and PlanckSZ. The main results are:
- For the first time, we detect a cross-correlation signal in both configuration and Fourier space. The measurement has a strong statistical significance for WHL12 and redMaPPer, while it is at the level of a hint for PlanckSZ.
- An alternative, more conventional, stacking analysis was performed that confirms these results. As expected, the [$\gamma$]{}-ray emission profile that emerges from the stacked images turns out to be consistent with that of the angular CCF.
- The cross-correlation measurement confirms that the unresolved EGB observed by [*Fermi*]{}-LAT correlates with the large scale clustering of matter in the Universe (traced by clusters), as found also in [@Fornengo:2014cya] and [@Xia:2014], employing different tracers.
- We model the cross correlation signal in the framework of the halo model and consider [$\gamma$]{}-ray emission from four types of astrophysical sources: blazars, mAGN, SFG and ICM. Our model provides a good fit to the measured CAPS on large scales. At $\ell\lesssim100$ all these sources but the ICM generate a cross-correlation compatible with the observations. The ICM contribution is subdominant. Its small amplitude is constrained by the fact that no [$\gamma$]{}-ray emission from clusters, including the nearby ones, has been detected so far. Only in the PlanckSZ cluster catalog case the ICM may provide a non negligible contribution at intermediate scales.
On small scales our model underestimates the observed CAPS. In this range the angular spectrum is approximately flat and can be fitted by a constant, shot-noise-like term, that we included in our model. Its physical interpretation is as follows. [*Fermi*]{}-LAT has a rather large PSF ($\gtrsim0.1^\circ$). The recorded arrival direction of photons is “randomized” with respect to the true direction on scales below (comparable to) the PSF. This creates a shot-noise-like term and prevents to fully characterize the correlation among sources beyond the PSF angular scale. Noting that the mean cluster redshift in the considered catalogs is $\bar z\sim 0.2-0.4$, any population of sources with physical sizes below a few Mpc could contribute to the small-scale CAPS.
- The cross correlation signal is significantly detected out to $\sim1$ degree, which is beyond the PSF extension (even though the statistical significance of the 2-halo term is not totally conclusive and amounts to $>2\sigma$). At the typical redshifts of the clusters in the considered catalogs, $\sim1$ degree corresponds to a linear scale of $\sim 10$ Mpc. This means that a large fraction of the correlation signal seems to be not physically associated to the clusters. Instead, it can be produced by AGNs or SFGs residing in the larger scale structures that surrounds the high density peaks where clusters reside.
- Finally, we have investigated the energy dependence of the cross-correlation signal. It turns out that on large scale, where the 2-halo term dominates, the signal is contributed by sources with hard energy spectra, consistent with that of the BL Lacs. On small scales, where the 1-halo term dominates, the correlation signal could be contributed by different types of sources. At high ($E>10$ GeV) energies the dominant sources have hard spectra, i.e. they are probably the same BL Lac population. At smaller energies, the correlation signal shows a hint of contribution by sources with softer spectra. These can be non-BL Lac AGNs, SFGs and/or the ICM.
In conclusion, our measurement combined with theoretical expectations suggests that the detected cross-correlation signal is largely contributed by compact sources like AGNs or SFGs. A possible contribution from the ICM, associated to the cluster itself, is however not excluded at small scales. Since its amplitude is expected to increase with the mass of the clusters, a cross-correlation analysis with a wide-field catalog containing a large number of nearby massive clusters can be therefore a suitable way to attempt the detection of the so far elusive [$\gamma$]{}-ray emission from the ICM.
Acknowledgements {#acknowledgements .unnumbered}
================
The *Fermi* LAT Collaboration acknowledges generous ongoing support from a number of agencies and institutes that have supported both the development and the operation of the LAT as well as scientific data analysis. These include the National Aeronautics and Space Administration and the Department of Energy in the United States, the Commissariat à l’Energie Atomique and the Centre National de la Recherche Scientifique / Institut National de Physique Nucléaire et de Physique des Particules in France, the Agenzia Spaziale Italiana and the Istituto Nazionale di Fisica Nucleare in Italy, the Ministry of Education, Culture, Sports, Science and Technology (MEXT), High Energy Accelerator Research Organization (KEK) and Japan Aerospace Exploration Agency (JAXA) in Japan, and the K. A. Wallenberg Foundation, the Swedish Research Council and the Swedish National Space Board in Sweden.
Additional support for science analysis during the operations phase is gratefully acknowledged from the Istituto Nazionale di Astrofisica in Italy and the Centre National d’Études Spatiales in France.
We would like to thank F. Zandanel for discussions. This work is supported by the research grant [*Theoretical Astroparticle Physics*]{} number 2012CPPYP7 under the program PRIN 2012 funded by the Ministero dell’Istruzione, Università e della Ricerca (MIUR); by the research grants [*TAsP (Theoretical Astroparticle Physics)*]{} and [*Fermi*]{} funded by the Istituto Nazionale di Fisica Nucleare (INFN); by the [*Excellent Young PI Grant: The Particle Dark-matter Quest in the Extragalactic Sky*]{}. MV and EB are supported by PRIN MIUR and IS PD51 INDARK grants. MV is also supported by ERC-StG cosmoIGM, PRIN INAF. JX is supported by the National Youth Thousand Talents Program, the National Science Foundation of China under Grant No. 11422323, and the Strategic Priority Research Program, [*The Emergence of Cosmological Structures*]{} of the Chinese Academy of Sciences, Grant No. XDB09000000. SC is supported by ERC Starting Grant No. 280127. Some of the results in this paper have been derived using the HEALPix[^8] package.
Validation and cross-checks {#sec:app1}
===========================
To assess the robustness of our analyses and results, we performed a few different tests, along the lines described in Section 6 of [@Xia:2014]. No unexpected behavior has been found. In this Appendix, we report the results for two tests and, for the sake of definiteness, we focus on the redMaPPer catalog only. First, we considered a different selection of [$\gamma$]{}-ray events. The analysis in the main text has been conducted using the [`P8R2_CLEAN`]{} class of [*Fermi*]{}-LAT events. This is a common choice for diffuse studies since it provides a good compromise between having a clean event class and sufficiently high statistics. Here we analyze also the [`P8R2_ULTRACLEANVETO`]{} photons, namely the cleanest Pass 8 event class. Moreover, we considered a more conservative zenith angle cut, excluding photons detected with measured zenith angle larger than 90$^{\circ}$ (instead of 100$^{\circ}$, as in the analysis presented in the main text). The comparison is shown in Fig. \[fig:CAPS\_valid\] (red versus violet points). It is clear that the results are fully compatible. Also the significances of detection are not dramatically affected. For the more conservative choice of photon events, the $p$-values of the statistical analysis described in Section \[sec:res\] would become $8.8\times 10^{-3}$ (WHL12), $3.7\times 10^{-5}$ (redMaPPer), and 0.18 (PlanckSZ).
To test the robustness of the detection we built a mock realization of the redMaPPer catalog by performing the transformation on the Galactic latitude $b\rightarrow-b$ for each cluster of the sample. This realization preserves the intrinsic clustering of the catalog (i.e., it provides the same autocorrelation signal), but should remove the cross-correlation (for more details, see Section 6.6 in [@Xia:2014]). Indeed, as clear also from Fig. \[fig:CAPS\_valid\] (orange open points), the derived CAPS is compatible with no signal, with the $p$-value now becoming $0.994$, meaning no preference over the null hypothesis.
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}
Consistency among different samples of clusters {#sec:app2}
===============================================
The WHL12 catalog is, among the three samples we considered, the one with the largest number of clusters. This is due to the fact that it has the lowest richness (and in turn mass) threshold, see also Fig. \[fig:cat\_prop\]b. The redshift range is instead not dramatically different, especially between WHL12 and redMaPPer (that also share the fraction of sky probed).
In this Appendix, we test the dependency of the WHL12 cross-correlation signal from the cluster richness (that can be translated into cluster mass, see discussion in Section \[sec:wf\]). We also check the consistency of our findings for the different catalogs, by comparing the signal of a high-richness subsample of WHL12 with the redMaPPer case. To this aim we split the WHL12 into three samples of richness $\lambda<23$, $23<\lambda<35$ and $\lambda>35$. The last bin contains 24,903 clusters, similarly to the redMaPPer catalog. Results are shown in Fig. \[fig:CAPS\_richness\], where we focus on the [$\gamma$]{}-ray energy bin $1\,{\rm GeV}<E_\gamma<10\,{\rm GeV}$ for definiteness. Even though the statistics is not very high, left panel shows that the amplitude increases with richness, as expected. The right panel compares the CAPS of the high-richness subsample of WHL12 with the redMaPPer one. The two cross-correlation measurements are fully compatible.
{width="49.00000%"} {width="49.00000%"}
[^1]: For a definition of the Pass 8 event selections and their characteristics, see http://www.slac.stanford.edu/exp/glast/groups/canda/\
lat\_Performance.htm.
[^2]: http://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/\
Cicerone/
[^3]: http://healpix.jpl.nasa.gov/
[^4]: http://fermi.gsfc.nasa.gov/ssc/data/access/lat/\
BackgroundModels.html \[foot4\]
[^5]: http://www2.iap.fr/users/hivon/software/PolSpice/
[^6]: In the literature, the relation between the stacked profile and the CAPS is typically reported using the Fourier transform (see, e.g., [@2012PhRvD..85b3007F]): y\^[(c)]{}() = |C\_\^[(c)]{} J\_0() , \[eq:ystack\] where $J_0$ is the is the zeroth order Bessel function. In the small angle limit $\ell \gg1$, $\theta \ll1$ we have $J_0(\ell \theta)\simeq P_\ell[\cos(\theta)]$ and thus Eq. (\[eq:2point\]) becomes the discrete form of Eq. (\[eq:ystack\]).
[^7]: In the correlation with the PlanckSZ catalog, we discard the energy bins 1 and 8 since they are meaningless due to their low statistics.
[^8]: http://healpix.sourceforge.net/downloads.php
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper overviews certain radiation detection, perception, and planning challenges for nuclearized robotics that aim to support the waste management and decommissioning mission. To enable the autonomous monitoring, inspection and multi–modal characterization of nuclear sites, we discuss important problems relevant to the tasks of navigation in degraded visual environments, localizability–aware exploration and mapping without any prior knowledge of the environment, as well as robotic radiation detection. Future contributions will focus on each of the relevant problems, will aim to deliver a comprehensive multi–modal mapping result, and will emphasize on extensive field evaluation and system verification.'
author:
- 'Frank Mascarich$^1$, Taylor Wilson$^2$, Tung Dang$^1$, Shehryar Khattak$^1$, Christos Papachristos$^1$, and Kostas Alexis$^1$[^1] [^2][^3] [^4]'
title: '**Towards Robotically Supported Decommissioning of Nuclear Sites** '
---
INTRODUCTION {#sec:intro}
============
A history of nuclear research, power generation and military developments has left a legacy of nuclear sites, now requiring careful decommissioning. In the U.S., the goal is that of safe cleanup of the Manhattan Project nuclear sites, the ensuing Cold War nuclear arms race, and the early years of federal nuclear science research and technology development. Sub–tasks of this broad mission include a) nuclear facility decommissioning, b) soil and water cleanup, c) liquid radioactive waste processing and disposition, d) solid radioactive waste treatment, storage and disposal, as well as e) nuclear materials and spent nuclear fuel management. In these challenging tasks requiring careful inspection, characterization, decommissioning and maintenance, robotic systems can be of unparalleled value. Specialized, robustly autonomous, robotic systems that can deal with the dirty, dull, dangerous and difficult environments of the nuclear sites are now required.
However, to facilitate the vision of broad and reliable robotic support of the waste management and decommissioning efforts, a set of challenges have to be addressed. Among others this includes the need for pioneering platform designs presenting ultimate mobility, robust autonomy in often visually–degraded and GPS–denied environments, high–resolution mapping and semantic classification, radiation detection and its fusion with multi–modal maps, as well as radiation source localization. Despite very important efforts of the community, such as those described in [@qian2012small; @han2013low; @bogue2011robots; @cortez2009distributed; @dudar1999mobile; @westrom1994radiation; @michal2009red], a variety of complex problems are yet to be addressed so that robots can operate autonomously in the sites relevant to the decommissioning effort and provide comprehensive mapping and characterization. Indeed, the complexity and the degraded conditions in the facilities of interest are unique to the domain. Figure \[fig:facilities\] presents photos of relevant sites.
![Indicative facilities of interest: a) H–Canyon and b) PUREX.[]{data-label="fig:facilities"}](nuclear_sites.eps){width="0.99\columnwidth"}
In this paper we discuss the problem of robotically supported nuclear waste management and decommissioning in the sense of autonomous exploration, inspection and characterization of the nuclear sites. In particular, our goal is to discuss some of the sensing, path planning, control, system design and implementation challenges that are particular to the environments and mission goals of nuclear site characterization.
CHALLENGES FOR NUCLEARIZED ROBOTICS {#sec:challenges}
===================================
Nuclear facilities present a unique set of challenges that make robotic support attractive. The most unique and obvious challenge encountered at nuclear facilities is ionizing radiation, ranging from a few times the natural background, to dose rates exceeding many Sv/hr at sites housing spent nuclear fuel, reprocessed material, and also at nuclear accident sites [@groves1983now; @kelly2009manhattan; @gephart2003hanford; @tsubokura2012internal]. This radiation creates an environment hazardous to worker entry, and is in many cases not well characterized because of the lack of human surveys. Loose or airborne contamination creates an internal exposure hazard and penetrating beta, photon and neutron fields pose external exposure threats. Robotic platforms can accurately map radiation fields in environments where dose–rates make personnel entry impossible, and additionally can map radiation and radionuclide contamination in less hazardous environments more efficiently than traditional surveys, keeping with the health physics principle of ALARA (As Low As Reasonably Achievable). Furthermore, nuclear facilities and especially those relevant to the decommissioning mission often are only documented in historic reports. Although rich documentation is available and useful information can be extracted from it, for most of these sites prior maps are not available and capability for GPS–denied operation is required.
In very high radiation areas, robotics have a unique role in performing tasks that are inaccessible or particularly challenging to humans. Indeed there are operational dose limits on robotics as well. However these limits are much higher and can be mitigated with a combination of radiation hardened electronics, additional shielding, and path planning to minimize time of exposure and dose rates. Radiation damage to semiconductor electronics represents the soft–point for nuclearized robotics, and for very high radiation dose rates, hardened electronics should be chosen, with a wealth of knowledge available from the use and evaluation of electronics packages for space applications. Additionally, work should be conducted for the effect of ionizing radiation on a range of optical systems, including cameras, LiDAR and other proximity/ranging systems. For certain modalities, such as CCD and CMOS optical sensors the effects of ionizing radiation have been demonstrated [@hopkinson2004radiation; @goiffon2008ionizing], yet a comprehensive study on the effects of radiation on many common other robotically–employed sensors is yet to be conducted in order to evaluate the performance of each and identify points of failure. This should be further coupled with the specific task the sensors are used for (e.g. SLAM).
Robotics for facilities with unsealed sources of radiation should make consideration for ease of decontamination, or in some cases design systems or components to be replaceable or disposable. Additionally, some robotics for nuclear applications will enter areas of high radiation that will also contain a strong thermal source, so temperature ratings and cooling scenarios should be addressed for these applications. This challenges aspects of the design and especially the battery system. Furthermore, mapping and localization of lost or orphan sources [@jarman2011bayesian; @towler2012radiation; @morelande2009radiological; @christie2016radiation] represents a real, as demonstrated by recent source recovery operations. Along with lowering personnel doses in such operations, an autonomous system can optimize a more efficient search method for multi–source localization.
Finally, it is noted that the development of nuclearized robotics for inspection operations, decommissioning, and accident response could have further applications outside the scope of this paper. Autonomous robotics with radiation detection and mapping systems could represent an effective safeguard tool for nuclear facilities and international inspectors to accurately inventory nuclear materials and safeguard against diversions. Such autonomous robotics could provide round–the–clock inspection and inventory of facilities with large layouts and quantities of material. Additionally, lessons learned in the field of nuclearized robotics could provide valuable design feedback for developing spacecraft systems for missions outside of Low Earth Orbit (LEO), where such systems will encounter high radiation fields, and derive usable mission data from radiation sensing.
AUTONOMOUS OPERATION IN DEGRADED VISUAL ENVIRONMENTS {#sec:dve}
====================================================
Nuclearized robotics will be requested to operate in all sorts of challenging environments. Going beyond the current state–of–the–art in robotics for the nuclear domain, autonomous operation (as opposed to teleoperation) and mission–execution in GPS–denied environments will become common. Even more challenging, it is noted that many important applications (e.g. decommissioning) often take place in Degraded Visual Environments (DVEs). Iconic examples include the inspection of the PUREX tunnels and H–Canyon.
For the problem of autonomous navigation in DVEs, a robust localization and intelligent path planning strategy has to be facilitated. Recent work of our team aims to address the problem through a) multi–modal sensor fusion [@NIRdepth_ICUAS_2017], and b) localization uncertainty–aware Receding Horizon Exploration and Mapping (RHEM) path planning [@RHEM_ICRA_2017]. In this approach, data from visual cameras synchronized with flashing LEDs (or Near Infrared cameras) are fused with inertial sensor cues and a depth sensor in order to enable robust operation in darkness. As given certain environments, a sensing modality can become ill–conditioned, a multi–modal sensor fusion approach can robustify the overall robot operation and also provide mapping results of higher resolution and fidelity. Towards robotic operation in the complex environments relevant to the nuclear decommissioning effort, our team develops a Multi–Modal Mapping Unit (M3U) a prototype of which is presented in Figure \[fig:m3uphoto\], while Figure \[fig:m3udesign\] provides an overview of its components. Recently published work employs a perception unit that relies on Near Infrared cameras and inertial sensors for localization [@NIRdepth_ICUAS_2017].
![Photo of a prototype of the Multi–Modal Mapping Unit.[]{data-label="fig:m3uphoto"}](m3u_photo.eps){width="0.99\columnwidth"}
![Overview of the design diagram of the Multi–Modal Mapping Unit architecture. The microcontroller unit (MCU) is responsible for the visual–inertial subsystem triggering, while a powerful high-level main processing unit (MPU) handles all the data acquisition and processing tasks.[]{data-label="fig:m3udesign"}](m3u_design.eps){width="0.99\columnwidth"}
With the localization pipeline running onboard the robot, the robot pose and tracked landmarks as well as their covariance matrix are estimated. These estimates are then exploited from the path planning module and propagated along sampled paths in order to account for the robot localizability along different trajectories. Figure \[fig:rhem\_steps\] presents the localizability–aware exploration and mapping planner [@RHEM_ICRA_2017]. At first, in an online computed tree, the algorithm identifies the branch that optimizes the amount of new space expected to be explored. The first viewpoint configuration of this branch is selected, but the path towards it is decided through a second planning step. Within that, a new random tree is sampled, admissible branches arriving at the reference viewpoint are found and the robot belief about its state and the tracked landmakrs of the environment is propagated. As system state the concatenation of the robot states and tracked landmarks (features) is considered. Then, the branch that minimizes the localization uncertainty, as factorized using the D–optimality of the pose and landmarks covariance matrix is selected. The corresponding path is conducted by the robot and the process is iteratively repeated. It is noted that this process goes beyond baseline deterministic exploration [@NBVP_ICRA_16; @yoder2016autonomous]. When some knowledge of the environment is available as a prior map, work on optimized coverage can also be exploited to provide a rough global path [@SIP_AURO_2015; @APST_MSC_2015; @bircher_robotica; @BABOOMS_ICRA_15; @galceran2013survey; @papachristos2016distributed; @oettershagen2016long; @alexis2017realizing].
![2D representation of the two–steps uncertainty–aware exploration and mapping planner. The first planning layer samples the path with the maximum exploration gain. The viewpoint configuration of the first vertex of this path becomes the reference to the second planning layer. Then this step, samples admissible paths that arrive to this configuration, performs belief propagation along the tree edges, and selects the one that provides minimum uncertainty over the robot pose and tracked landmarks. []{data-label="fig:rhem_steps"}](planning_layers_explanation.eps){width="0.99\columnwidth"}
ROBOTIC RADIATION DETECTION {#sec:rad}
===========================
Radiation detection is a well–studied and a continuously–evolving field on its own but robotized sensing brings further and new challenges. First of all, good overview of the types of radiation sensing systems, such as proportional gas–filled detectors, semiconductor diode detectors, germanium gamma-ray detectors and other solid–state solutions, scintillation detectors, and radiation cameras, their features, radiation, thermal, and mechanical hardness is required. Furthermore, the critical role of photomultiplier tubes and photodiodes has to be well–understood to enable the appropriate selection and design of the sensing module. Good theoretical references can be found at [@knoll2010radiation; @tsoulfanidis2013measurement]. Specific to the application, the sensing solution has to be decided according to the interest to detect alpha, beta, gamma or neutron activity, the required energy resolution and the power levels of the site to be surveyed. A critical question is if spectroscopy is required. Figure \[fig:raddetectors\] presents indicative radiation detectors. In addition, limitations of the robotic platform will necessarily shape the final detector selection.
![Indicative radiation detectors: a) gas–filled proportional detector, b) solid–state detector, c) scintillator, d) gamma camera.[]{data-label="fig:raddetectors"}](rad_detectors.eps){width="0.99\columnwidth"}
In the area of gamma radiation detection, and depending on the application, three detection technologies namely a) miniature scintillation detectors (e.g. CeBr3, CsI, NaI) with built–in temperature compensated bias generator and a pre-amplifier often alongside a silicon photomultiplier (SiPm) tube [@quarati2013scintillation; @swiderski2015gamma; @valtonen2009radiation; @beck1972situ; @buzhan2003silicon; @herbert2006first], b) miniature solid–state low voltage gamma detectors [@dearnaley1967solid; @lutz1999semiconductor], and c) gamma cameras [@he2010three; @bolotnikov2012array; @zhang2007prototype] are worth of special attention. The first two solutions can be realized at extremely small sizes and low–weights making them affordable for aerial robotic applications, while scintillation devices can provide the sensitivity and energy resolution characteristics required for precise monitoring and source localization. Radiation cameras are still relatively heavy but provide unique characteristics when it comes to radiation mapping in correlation with the $3\textrm{D}$ structure. Through a multi–modal sensor fusion approach, comprehensive 3D maps annotated with radiation can be derived.
Neutron detection is also a particularly interesting area with high relevance to homeland security and industrial monitoring (e.g. personnel monitoring, water content in soil) applications. Neutron detection refers to the effective detection of neurons entering a well–positioned detector. Neutrons can be produced through multiple processes such as alpha particle induced reactions, spontaneous fission, and induced fission. Gas–filled proportional detectors such as the family of $^3$He–based detectors [@east1969polyethylene], scintillation neutron detectors (e.g. liquid organic, plastic) [@thomas1962boron], as well as solid–state neutron detectors may be used [@petrillo1996solid]. A selection of a neutron detector with the appropriate radiological sensitivity for the application is required.
Alpha detection is key to many applications in contaminated areas but its detection is particularly challenging. As alpha particles are the heaviest and most highly charged of the common nuclear radiations, they quickly give up their energy to any medium through which they pass, rapidly coming to equilibrium with, and disappearing in the medium. Due to this reason special detection techniques must be used to allow the particles to enter the active region of a detector (e.g. ZnS(Ag)–based scintillation devices)). In field instruments it is common to use an extremely thin piece of aluminized Mylar film on the face of the detector probe to cover a thin layer of florescent material. This is due to the fact that energy attenuation of the incident alpha radiation through Mylar is estimated to be less than $10\%$. However, the use of this film makes the detector extremely fragile - to the level that any contact with a hard object, such as a blade of hard grass, may puncture the film [@websiterad].
Beyond the radiation detectors themselves, a set of methods and techniques are critical to achieve the desired final sensing result. First of all, spectroscopy is critical when characterization matters. Dose and dose–rate equivalent count rate monitoring is important especially for safety–related tasks. Facilitation and tuning of detection directionality through a set of techniques (e.g. compton imaging, coded mask apertures, collimation) allows to realize the desired sensing properties. Furthermore, appropriate design of the interfacing (e.g. amplifiers, multi–channel analyzers, analog–to–digital converters) and processing electronics (e.g. DSPs, FPGAs) is critical and has to be considered in order to achieve the desired sensing functionality at an affordable weight and cost.
An additional critical step is that of detector calibration. The process of radiation detector intrinsics calibration involves the use of a pre–calibrated source and logging of counting statistics over different orientations and distances from the source. It is important to be aware of the polarity characteristics of the radiation detector to be used and ongoing experience indicates that an in–house calibration step is critical to take place. Furthermore, extrinsics calibration with the remaining of the sensor modalities is again required if correlation with the $3\textrm{D}$ reconstructed map or direct sensor data (e.g. camera frames) is to take place. Figure \[fig:teviso\_calib\] presents such calibration data for the case of the TEVISO RD3024 solid–state detector with the use of a $300\textrm{mCi}$ $\textrm{Cs}$-127 source.
![Calibration results of a TEVISO RD3024 solid–state detector with the use of a $300\textrm{mCi}$ $\textrm{Cs}$-127 source. Calibration took place over different sensor orientations and varied distances from the source. []{data-label="fig:teviso_calib"}](teviso-calib2.eps){width="0.99\columnwidth"}
Finally, a challenge of robotized nuclear detection especially related to platforms of high mobility (e.g. aerial robots) is that of localization accuracy. Due to the fact that depending on the application radiation detectors may require significant dwell times, localization accuracy and robustness is critical. This is particularly relevant when smaller spaces are considered, when GPS–denied operation is required, and when accurate estimates of the radiation source location are required.
PRELIMINARY RESULTS {#sec:results}
===================
A set of preliminary studies related to exploration in DVEs and radiation detection have been conducted in order to approach the challenge of developing nuclearized robotics especially related with the problem of supporting the waste management decommissioning effort.
Localization and Mapping inside a Dark Tunnel
---------------------------------------------
For this experimental evaluation, the mission took place within a remote city tunnel during night–time. This kind of environment is unique in multiple aspects: a) it exhibits nearly–complete lack of ambient light due to its closed structure especially at night (while even during the day it still is significantly dark and robotic deployment within such a space would still require handling of this aspect), b) it is littered with dust which can lift up into the air and into the sensors’ fields–of–view due to the turbulence created by an aerial robot’s spinning rotors, and c) its internal structure mainly composed of concrete walls is such that contains little discernible texture.
![Aerial robotic localization and mapping within a dark city tunnel.[]{data-label="fig:tunnel_experiment"}](tunnel_experiment.eps){width="0.99\columnwidth"}
Figure \[fig:tunnel\_experiment\] illustrates these conditions based on the data recorded during the experiment, alongside the localization and mapping results as performed in real–time during the experiment. A video of the experimental sequence is also available at <https://youtu.be/HpWlFUNboR4>
Autonomous Exploration in DVEs
------------------------------
This mission scenario refers to the complete concept of autonomous robotic exploration of DVEs. The mock-up space is a dark indoor location, with dimensions $12\times6.5\times2\textrm{m}$, setup to incorporate artificially–created vertical and T–shaped walls, as well as other structural elements by using $300$ boxes with size $0.4\times0.3\times0.3\textrm{m}$.
![Autonomous aerial robotic exploration within an indoor degraded visual environment containing wall-like structures.[]{data-label="fig:arena_experiment"}](arena_experiment.eps){width="0.99\columnwidth"}
Figure \[fig:arena\_experiment\] illustrates the aforementioned conditions, as well as the progress of this experiment, while a video of the sequence is also available at <https://youtu.be/1-nPFBhyTBM>. For this mission where the human is out–of–the–loop, consistent localization and mapping during autonomous exploration are provided by the localizability-aware RHEM planner.
Radiation Detection
-------------------
Two Teviso RD3024 low voltage SMD/SMT nuclear radiation sensors were installed on the aerial robot. As mounting points the two opposite facing arms of its hexacopter structure were selected to create differential measurements and exploit the polarity of the sensor. With the radiation sensors being initially calibrated with the use of a characterized source (see Figure \[fig:teviso\_calib\]), the robot was then commanded to follow an exploration trajectory. Given the sensor data collected, the source location is estimated. Figure \[fig:radexp\] presents the relevant result, while a video of the experiment is available at <https://youtu.be/b9BbKQTfrY8>.
![Aerial robotic radiation detection.[]{data-label="fig:radexp"}](rad_exp1b.eps){width="0.99\columnwidth"}
CONCLUSIONS {#sec:concl}
===========
This paper discussed certain challenges relevant to the problem of robotically supported nuclear waste management and decommissioning with a special focus on nuclear site characterization. Furthermore, preliminary results on GPS–denied operation in degraded visual environments, exploration and mapping, as well as radiation detection were presented. Future work will focus on the challenges of multi–modal characterization, robustly autonomous exploration and mapping, optimized robotic radiation detection, and real–time multi–source localization.
ACKNOWLEDGEMENTS {#sec:concl}
================
The authors would like to thank Myung Chul Jo, radiation safety officer at the University of Nevada, Reno who provided his valuable help in order to enable the proper and safe use of the radiation equipment involved in this study.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
K. Qian, A. Song, J. Bao, and H. Zhang, “Small teleoperated robot for nuclear radiation and chemical leak detection,” *International Journal of Advanced Robotic Systems*, vol. 9, no. 3, p. 70, 2012.
J. Han, Y. Xu, L. Di, and Y. Chen, “Low-cost multi-uav technologies for contour mapping of nuclear radiation field,” *Journal of Intelligent & Robotic Systems*, pp. 1–10, 2013.
R. Bogue, “Robots in the nuclear industry: a review of technologies and applications,” *Industrial Robot: An International Journal*, vol. 38, no. 2, pp. 113–118, 2011.
R. A. Cortez, H. G. Tanner, and R. Lumia, “Distributed robotic radiation mapping,” in *Experimental Robotics*.1em plus 0.5em minus 0.4emSpringer, 2009, pp. 147–156.
A. M. Dudar, C. R. Ward, J. D. Jones, W. R. Mallet, L. J. Harpring, M. X. Collins, and E. K. Anderson, “Mobile autonomous robotic apparatus for radiologic characterization,” 1999, uS Patent 5,936,240.
G. B. Westrom, R. E. Carlton, and L. R. Tripp, “Radiation mapping system,” 1994, uS Patent 5,286,973.
R. Michal, “Red whittaker and the robots that helped clean uptmi-2,” *Nuclear News*, pp. 37–40, 2009.
L. R. Groves, *Now it can be told: The story of the Manhattan Project*.1em plus 0.5em minus 0.4emDa Capo Press, 1983.
C. C. Kelly, *Manhattan Project: The Birth of the Atomic Bomb in the Words of Its Creators, Eyewitnesses, and Historians*.1em plus 0.5em minus 0.4emBlack Dog & Leventhal, 2009.
R. E. Gephart, “Hanford: A conversation about nuclear waste and cleanup,” Pacific Northwest National Laboratory (PNNL), Richland, WA (US), Tech. Rep., 2003.
M. Tsubokura, S. Gilmour, K. Takahashi, T. Oikawa, and Y. Kanazawa, “Internal radiation exposure after the fukushima nuclear power plant disaster,” *Jama*, vol. 308, no. 7, pp. 669–670, 2012.
G. R. Hopkinson, A. Mohammadzadeh, and R. Harboe-Sorensen, “Radiation effects on a radiation-tolerant cmos active pixel sensor,” *IEEE Transactions on Nuclear Science*, vol. 51, no. 5, pp. 2753–2762, 2004.
V. Goiffon, P. Magnan, F. Bernard, G. Rolland, O. Saint-P[é]{}, N. Huger, and F. Corbi[è]{}re, “Ionizing radiation effects on cmos imagers manufactured in deep submicron process,” in *Proceedings of SPIE-The International Society for Optical Engineering*, vol. 6816, 2008.
K. D. Jarman, E. A. Miller, R. S. Wittman, and C. J. Gesh, “Bayesian radiation source localization,” *Nuclear technology*, vol. 175, no. 1, pp. 326–334, 2011.
J. Towler, B. Krawiec, and K. Kochersberger, “Radiation mapping in post-disaster environments using an autonomous helicopter,” *Remote Sensing*, vol. 4, no. 7, pp. 1995–2015, 2012.
M. R. Morelande and B. Ristic, “Radiological source detection and localisation using bayesian techniques,” *IEEE Transactions on Signal Processing*, vol. 57, no. 11, pp. 4220–4231, 2009.
G. Christie, A. Shoemaker, K. Kochersberger, P. Tokekar, L. McLean, and A. Leonessa, “Radiation search operations using scene understanding with autonomous uav and ugv,” *arXiv preprint arXiv:1609.00017*, 2016.
, “Autonomous exploration of visually-degraded environments using aerial robots,” in *IEEE International Conference on Unmanned Aircraft Systems (ICUAS)*, 2017.
S. K. Christos Papachristos and K. Alexis, “Uncertainty–aware receding horizon exploration and mapping using aerial robots,” in *IEEE International Conference on Robotics and Automation (ICRA)*, June 2017.
, “Receding horizon “next-best-view” planner for 3d exploration,” in *IEEE International Conference on Robotics and Automation (ICRA)*, May 2016.
L. Yoder and S. Scherer, “Autonomous exploration for infrastructure modeling with a micro aerial vehicle,” in *Field and Service Robotics*. 1em plus 0.5em minus 0.4emSpringer, 2016, pp. 427–440.
, “,” **, pp. 1–25, 2015.
, “Uniform coverage structural inspection path-planning for micro aerial vehicles,” September 2015.
, “An incremental sampling-based approach to inspection planning: The rapidly-exploring random tree of trees,” 2015.
, “Structural inspection path planning via iterative viewpoint resampling with application to aerial robotics,” in *IEEE International Conference on Robotics and Automation (ICRA)*, May 2015, pp. 6423–6430.
E. Galceran and M. Carreras, “A survey on coverage path planning for robotics,” *Robotics and Autonomous Systems*, vol. 61, no. 12, pp. 1258–1276, 2013.
C. Papachristos, K. Alexis, L. R. G. Carrillo, and A. Tzes, “Distributed infrastructure inspection path planning for aerial robotics subject to time constraints,” in *Unmanned Aircraft Systems (ICUAS), 2016 International Conference on*.1em plus 0.5em minus 0.4emIEEE, 2016, pp. 406–412.
P. Oettershagen, T. Stastny, T. Mantel, A. Melzer, K. Rudin, P. Gohl, G. Agamennoni, K. Alexis, and R. Siegwart, “Long-endurance sensing and mapping using a hand-launchable solar-powered uav,” in *Field and Service Robotics*.1em plus 0.5em minus 0.4emSpringer International Publishing, 2016, pp. 441–454.
K. Alexis, “Realizing the aerial robotic worker for inspection operations,” *arXiv preprint arXiv:1703.02640*, 2017.
G. F. Knoll, *Radiation detection and measurement*.1em plus 0.5em minus 0.4emJohn Wiley & Sons, 2010.
N. Tsoulfanidis, *Measurement and detection of radiation*.1em plus 0.5em minus 0.4emCRC press, 2013.
F. Quarati, P. Dorenbos, J. Van der Biezen, A. Owens, M. Selle, L. Parthier, and P. Schotanus, “Scintillation and detection characteristics of high-sensitivity cebr 3 gamma-ray spectrometers,” *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, vol. 729, pp. 596–604, 2013.
L. Swiderski, P. Schotanus, E. Bodewits, D. Badocco, T. Batsch, D. Cester, M. Corbo, P. Garosi, A. Iovene, J. Iwanowska-Hanke *et al.*, “Gamma spectrometer based on cebr 3 scintillator with compton suppression for identification of trace activities in water,” in *Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), 2015 IEEE*.1em plus 0.5em minus 0.4emIEEE, 2015, pp. 1–3.
E. Valtonen, J. Peltonen, O. Dudnik, A. Kudin, H. Andersson, Y. A. Borodenko, T. Eronen, J. Huovelin, H. Kettunen, E. Kurbatov *et al.*, “Radiation tolerance tests of small-sized csi (tl) scintillators coupled to photodiodes,” *IEEE Transactions on Nuclear Science*, vol. 56, no. 4, pp. 2149–2154, 2009.
H. L. Beck, J. Decampo, and C. Gogolak, “In situ ge (li) and nai (tl) gamma-ray spectrometry.” Health and Safety Lab.,(AEC), New York, Tech. Rep., 1972.
P. Buzhan, B. Dolgoshein, L. Filatov, A. Ilyin, V. Kantzerov, V. Kaplin, A. Karakash, F. Kayumov, S. Klemin, E. Popova *et al.*, “Silicon photomultiplier and its possible applications,” *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, vol. 504, no. 1, pp. 48–52, 2003.
D. J. Herbert, V. Saveliev, N. Belcari, N. D’Ascenzo, A. Del Guerra, and A. Golovin, “First results of scintillator readout with silicon photomultiplier,” *IEEE Transactions on Nuclear Science*, vol. 53, no. 1, pp. 389–394, 2006.
G. Dearnaley, “Solid-state radiation detectors,” *Contemporary Physics*, vol. 8, no. 6, pp. 607–626, 1967.
G. Lutz *et al.*, *Semiconductor radiation detectors*.1em plus 0.5em minus 0.4emSpringer, 1999, vol. 40.
Z. He and F. Zhang, “Three-dimensional, position-sensitive radiation detection,” 2010, uS Patent 7,692,155.
A. Bolotnikov, J. Butcher, G. Camarda, Y. Cui, G. De Geronimo, J. Fried, R. Gul, P. Fochuk, M. Hamade, A. Hossain *et al.*, “Array of virtual frisch-grid czt detectors with common cathode readout for correcting charge signals and rejection of incomplete charge-collection events,” *IEEE Transactions on Nuclear Science*, vol. 59, no. 4, pp. 1544–1551, 2012.
F. Zhang, Z. He, and C. E. Seifert, “A prototype three-dimensional position sensitive cdznte detector array,” *IEEE Transactions on Nuclear Science*, vol. 54, no. 4, pp. 843–848, 2007.
L. East and R. Walton, “Polyethylene moderated 3he neutron detectors,” *Nuclear Instruments and Methods*, vol. 72, no. 2, pp. 161–166, 1969.
G. Thomas, “A boron-loaded liquid scintillation neutron detector using a single photomultiplier,” *Nuclear Instruments and Methods*, vol. 17, no. 2, pp. 137–139, 1962.
C. Petrillo, F. Sacchetti, O. Toker, and N. Rhodes, “Solid state neutron detectors,” *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, vol. 378, no. 3, pp. 541–551, 1996.
, “Radiation detection and measurement.”
[^1]: This material is based upon work supported by the Department of Energy under Award Number \[DE-EM0004478\].
[^2]: $^1$ The authors are with the Autonomous Robots Lab, University of Nevada, Reno, 1664 N. Virginia, 89557, Reno, NV, USA [[email protected]]{}
[^3]: $^2$ The author is with the Wilson Laboratory for Radiaton Physics, University of Nevada, Reno, 1664 N. Virginia, 89557, Reno, NV, USA
[^4]: This paper is presented at the Autonomous Structural Monitoring and Maintenance using Aerial Robots Workshop, IEEE International Conference on Robotics and Automation (ICRA) 2017, Singapore, May 29, 2017
|
{
"pile_set_name": "ArXiv"
}
|
---
bibliography:
- 'neurogenbib.bib'
---
[**Quantifying the behavioural relevance of hippocampal neurogenesis** ]{}\
Stanley E. Lazic$^{1,\ast}$, Johannes Fuss$^{2}$, Peter Gass$^{3}$\
**[1]{} In Silico Lead Discovery, Novartis Institutes for Biomedical Research, 4002 Basel, Switzerland\
**[2]{} Department of Psychiatry and Psychotherapy, Central Institute of Mental Health, University of Heidelberg, Medical Faculty Mannheim, Mannheim, Germany\
**[3]{} Department of Psychiatry and Psychotherapy, Central Institute of Mental Health, University of Heidelberg, Medical Faculty Mannheim, Mannheim, Germany\
$\ast$ E-mail: [email protected]******
Abstract {#abstract .unnumbered}
========
Few studies that examine the neurogenesis–behaviour relationship formally establish covariation between neurogenesis and behaviour or rule out competing explanations. The behavioural relevance of neurogenesis might therefore be overestimated if other mechanisms account for some, or even all, of the experimental effects. A systematic review of the literature was conducted and the data reanalysed using causal mediation analysis, which can estimate the behavioural contribution of new hippocampal neurons separately from other mechanisms that might be operating. Results from eleven eligible individual studies were then combined in a meta-analysis to increase precision (representing data from 215 animals) and showed that neurogenesis made a negligible contribution to behaviour (standarised effect = 0.15; 95% CI = -0.04 to 0.34; p = 0.128); other mechanisms accounted for the majority of experimental effects (standardised effect = 1.06; 95% CI = 0.74 to 1.38; p = 1.7 $\times 10^{-11}$).
Introduction {#introduction .unnumbered}
============
There is a consensus that new neurons in the hippocampus have a causal and biologically significant influence on cognitive and affective behaviours. However, three conditions commonly cited for establishing causation—temporal precedence, covariation, and the elimination of competing explanations—are rarely satisfied by studies that examine the neurogenesis–behaviour relationship. Temporal precedence is the only condition that is usually met because experimental interventions are applied before behavioural assessments, and changes in neurogenesis occur in between ($Experimental~intervention \rightarrow changes~in~neurogenesis \rightarrow behavioural~assessment$). Few studies quantitatively estimate the covariation between neurogenesis and behaviour; most only provide a qualitative description of a *group level* association by demonstrating that the experimental group with greater neurogenesis had better behavioural performance, or vice versa. This is not the same as demonstrating that *individual animals* with higher levels of neurogenesis have better behavioural performance. Group level associations cannot be used to infer individual level associations—this is known as the ecological fallacy [@Robinson1950]. It is even possible for associations at the group level to differ in sign from those at the individual level. The few studies that perform an individual level analysis unfortunately often do the correlation or regression through all of the data, without regard for the experimental groups. This rarely tests a hypothesis of interest and can lead to incorrect inferences [@Bewick2003; @Hassler2003].
There is a deeper issue in that the experimental interventions used to alter levels of neurogenesis (e.g. exercise, stress, irradiation, hormones, etc.) have numerous effects besides altering neurogenesis (see Table 1 in reference [@Lazic2012]), which may influence behaviour and therefore become competing explanations. Indeed, the more thorough studies measure multiple histological or physiological properties of the hippocampus and often find differences between experimental groups. Each of these outcomes becomes a competing explanation of how a particular experimental intervention affects behaviour. Few studies in the neurogenesis literature directly test whether changes in behaviour can be attributed to changes in neurogenesis and conclusions are based on an unjustified inferential leap (Fig. \[fig:models\]A). The typical study demonstrates (1) that the experimental intervention affects neurogenesis, (2) that the intervention affects behaviour, and then inevitably concludes (3) that neurogenesis was responsible for the behavioural change (dashed line). Few recognise that no conclusion can be made about the relevance of neurogenesis with this analysis and reasoning. To see why the form of reasoning is fallacious one simply has to replace neurogenesis with any other variable: “older mice (1) have more grey hair and (2) perform worse on the memory task, therefore we conclude (3) that hair colour is important for memory performance”. Arguing that neurogenesis is a plausible explanation (unlike hair colour) because of its neuroanatomical location, the electrophysiological properties of new neurons, etc. is assuming the point to be proved. The influence of other mechanisms is highly likely and could account for some—or even all—of the observed behavioural effects. Perfunctory caveats in the discussion noting that other mechanisms might also be involved do not adequately address the issue because if the majority of the behavioural effects are due to other mechanisms then the conclusion about neurogenesis’ relevance needs to be reconsidered. Many of these competing explanations can be eliminated by using a mediation analysis, which separates the total effect of an intervention into the part due to neurogenesis and the part due to all other mechanisms (Fig. \[fig:models\]B), and such methods are used routinely in many fields.
Recent reviews on the neurogenesis–behaviour relationship have classified studies as providing either correlational or causal evidence, although the criteria used to make this distinction are often unspecified and differ between authors [@Zhao2008; @DeCarolis2010; @Koehl2011; @Petrik2012; @Vadodaria2014]. We argue that this distinction is both artificial and unhelpful for three reasons. First, both classes of studies (however defined) are of the same design: they either randomise animals to different treatment conditions or use pre-existing groups such as old and young animals. The difference between groups on neurogenesis and behaviour is then examined. The only difference between studies is the probability that the observed behavioural effects are actually related to neurogenesis and not to some other effect of the intervention. A study is not causal because neurogenesis was increased using a highly selective genetic approach and correlational because physical activity was used. Second, even the so-called causal studies still make inferences by qualitatively describing group level associations. Third, the so-called correlational studies can provide causal evidence by separately estimating the effect of neurogenesis from other effects with a mediation analysis.
Previous studies that used a mediation analysis or related methods did not show a significant effect of neurogenesis, and the estimates were close to zero [@Bizon2004; @Lazic2010a; @Castro2010; @Lazic2012]. This suggests that the relevance of neurogenesis on various behavioural tasks is much more modest than commonly believed (or perhaps even non-existent) and that other mechanisms account for the observed effects. We therefore sought to obtain a better understanding of the neurogenesis–behaviour relationship by reanalysing all publicly available data. A Bayesian mediation analysis was used to estimate neurogenesis’ unique contribution and then results combined across studies to increase the precision of the estimates. Neuroscience studies often have small sample sizes and thus low statistical power [@Button2013] and most published studies are underpowered to detect any role that neurogenesis may have when analysed with a mediation model [@Fritz2007; @Thoemmes2010]. However, by integrating data across studies the low power of individual experiments is irrelevant as they are combined to obtain precise estimates. The combined sample size of 215 makes this one of the largest neurogenesis analyses to date. All animal models, methods of manipulating neurogenesis, behavioural outcomes, and study designs were eligible for inclusion. The quality of the literature was also assessed by quantifying the number of studies that used randomisation, blinding, took litter-effects into account, whether data points were dropped without mention or justification, and whether there was selective reporting of neurogenesis–behaviour associations, as these are known to introduce bias in the reported results.
Results and Discussion {#results-and-discussion .unnumbered}
======================
Neurogenesis’ contribution to behaviour is minimal {#neurogenesis-contribution-to-behaviour-is-minimal .unnumbered}
--------------------------------------------------
Figure \[fig:meta\_results\] displays the individual and combined effect estimates from neurogenesis and other mechanisms, and it is clear that the effect of other mechanisms was larger than that of neurogenesis in all cases. The black diamonds are the combined estimates across the eleven studies. Overall, neurogenesis made a small—and some might argue negligible—12% contribution to behaviour that was not significant (standardised effect = 0.15; 95% CI = -0.04 to 0.34; p = 0.128) and non-neurogenesis mechanisms played a much greater role (standardised effect = 1.06; 95% CI = 0.74 to 1.38; p = 1.7 $\times 10^{-11}$). While not part of the literature search, one additional study was available to us [@Fuss2010], which had a greater effect of neurogenesis compared to other mechanisms. Including this study in the analysis increases the overall effect of neurogenesis from 0.15 to 0.18 (95% CI = -0.01 to 0.36; p=0.062). Assuming that neurogenesis has some effect, including additional studies will eventually produce a statistically significant result and the lower confidence limit will exclude zero. Of greater interest however is the upper confidence limit, which places an upper bound on the magnitude of neurogenesis’ contribution. It is clear in Figure \[fig:meta\_results\] that large or even moderate effects of neurogenesis are not supported by the data, properly analysed. The conclusions of these studies therefore needs to be revisited, and it should be noted that the degree to which the authors believed that their study provided support for a neurogenesis–behaviour relationship varied greatly.
An advantage of a meta-analysis is that the consistency of effects across studies can be estimated [@Borenstein2009]. It is clear that the effect of neurogenesis is consistently small across a variety of different species, behavioural outcomes, and methods of manipulating neurogenesis (Fig. \[fig:meta\_results\]), which is confirmed with a test for between-study heterogeneity (empirical Bayes estimate: $\tau^2$ = 0; SE = 0.048; Q$_{(11)}$ = 4.98, p = 0.932). If the estimated contribution of neurogenesis differed greatly between studies then follow-up analyses could attempt to attribute this variation to properties of the studies. For example, are some behavioural tests more sensitive to changes in neurogenesis than others, or is neurogenesis more important in certain species or strains? Despite differences in the characteristics of the experiments, they all consistently showed little or no effect of neurogenesis. This means that studies not included in the meta-analysis would be unlikely to show a dramatic effect of neurogenesis. Even though only a small proportion of published studies could be included in the meta-analysis the low between-study variability strengthens the generalisability of the results.
Quality of the literature {#quality-of-the-literature .unnumbered}
-------------------------
Only 36% (4/11) of studies reported using blinding, 36% (4/11) used randomisation, and 18% (2/11) commented on how the design of the study accounted for potential litter effects. To our knowledge, no studies have examined or reported on litter-to-litter variation in neurogenesis, but litter effects are present for many outcomes (including behaviour; see reference [@Lazic2013] and references therein) and are a source of both bias and noise. Given the large number of factors that affect neurogenesis, it would be surprising if it were stable across litters. In addition, 45% (5/11) of the studies appeared to have incomplete data in that the number of data points in the figures was less than the number of animals mentioned in the methods section, or the reported degrees of freedom did not match the indicated sample size and method of analysis [@Lazic2010]. Finally, 64% (7/11) of studies selectively reported results in that only some associations between behavioural outcomes and neurogenesis were reported. Not surprisingly, all reported associations from these studies were significant. These results are similar to other preclinical studies using animal models [@Bebarta2003; @Macleod2004; @Crossley2008; @Philip2009; @Kilkenny2009; @Sena2010; @Vesterinen2010; @Rooke2011; @Moja2014] and are known to inflate effect sizes (reviewed by Dirnagl [@Dirnagl2006]).
Is neurogenesis involved with forgetting? {#is-neurogenesis-involved-with-forgetting .unnumbered}
-----------------------------------------
A recent paper by Akers et al. published after the previous analysis was completed suggests that high levels of neurogenesis might be associated with increased forgetting, particularly when neurogenesis is increased after learning has taken place [@Akers2014]. This is a different proposed behavioural function for neurogenesis compared to previous studies. Only data for one experiment could be extracted (Figure 4F in the paper), which used a genetic knock-down of neurogenesis. Wild-type (WT) mice were compared with transgenic mice expressing thymidine kinase (TK) with a nestin promoter. Administering ganciclovir to TK mice causes apoptosis of nestin expressing cells, thus reducing the number of new neurons. The primary outcome in this paper was the length of time animals were immobile or frozen when placed into an environment in which they had previously received an electrical shock, and was expressed as the percent of total time in the test environment. Since the outcome is bounded between 0 and 100%, the values were transformed to the logit scale ($log\left( p/(1-p) \right)$ after dividing by 100 to convert percents into proportions. This allows a normal linear model to be used for the reanalysis; otherwise predictions and confidence intervals would include impossible values (outside of the \[0,1\] interval). The data were then back transformed onto the proportion scale for plotting (Fig. \[fig:freezing\]A).
If the experimental groups are ignored, there is an association between the number of Ki67$^+$ cells and the proportion of time spent freezing, which would appear to provide evidence for a causal role for neurogenesis. However, the direction and strength of this association is irrelevant for addressing the question of whether neurogenesis is important for behaviour—a point which has not been appreciated by the majority of the neurogenesis community. In order to explain why, an unrelated example will be used so that prior opinions do not influence the interpretation. Suppose we have the hypothesis that taller people earn higher incomes, and specifically that greater height *causes* greater income. To simplify the analysis we could examine males only and one profession only, say that of barbers. Collecting data on the height and income of a number of barbers would then allow us to address the association between these variables, which would be a partial test of the hypothesis by establishing whether covariation exists. Suppose that the data were collected from two countries: the Netherlands (tall and high income by global standards) and individuals from a pygmy population in the Congo (short [@Bozzola2009] and low income). A graph of height ($x$-axis) versus income ($y$-axis) would show a cluster of points in the top right (Dutch population) and in the bottom left (African population) and a correlation or regression through all of the data (ignoring the country that individuals are from) would show an impressive positive association and a small p-value. This however is irrelevant for our hypothesis: the Dutch barbers earn higher incomes not because they are so tall but because they live in a modern industrialised country. If height was causally involved in determining income, then we would expect this association to hold *within* the two populations—the tallest Dutchman should earn more than the shortest and the tallest Congolese should earn more than the shortest. It is the association within groups that matters (which can be tested in a single analysis with an analysis of covariance for example and does not require a separate analysis for each population [@Lazic2012]). Similarly, it is the association between neurogenesis and behaviour within the WT group and within the TK group that tests the causal hypothesis, not the overall association. This generalises to any experimental or grouping variable.
In the Akers data set, the association within the experimental groups is still negative but not significant (p = 0.12), which might be due to lower power. A mediation analysis was also performed on the standarised variables and the effect of neurogenesis was 0.399 (95% CI = -0.10 to 1.08), which is 40% of the total effect (Fig. \[fig:freezing\]B). The standarised effect attributed to all other mechanisms was 0.623 (95% CI = -0.36 to 1.54). It is clear from the 95% CI that there is a great deal of uncertainty in the estimates and that a larger sample size is required, but it is clear that there is no overwhelming effect of neurogenesis. It might be surprising that genetic knock-down of neurogenesis still has such a large effect attributed to other mechanisms, as this method of manipulating neurogenesis is less likely to have additional off-target effects. Some of this might be attributed to biological or technical variables that are affecting the results such as litter effects, or the sex of the person handling the animals [@Sorge2014].
There are several points to consider. First, these results show that the behavioural effects of experimental interventions are driven mainly by neurogenesis-independent mechanisms. The studies included in the combined analysis are representative of the literature and include several strains of rats and mice, measures of both cognitive and affective behaviour, and many of the standard methods of manipulating neurogenesis (e.g. exercise, environmental enrichment, stress, and age). It is therefore likely that the relevance of neurogenesis in general has been greatly overestimated, bringing into question the usefulness of increasing neurogenesis for therapeutic purposes.
Second, these results should not be at all surprising because it is known that the experimental interventions typically used to alter levels of neurogenesis have many other effects that influence behaviour [@Lazic2012], and thus they are potential alternative explanations that need to be ruled out to make a convincing case for neurogenesis. For example, Lipp and colleagues showed a high within group correlation between the extent of intra- and infrapyramidal mossy fibre projections from the dentate to the CA3 and performance on a hippocampal dependent behavioural task [@Lipp1988]. This was evident in the control group as well, suggesting that natural variation in mossy fibre projections are a good predictor of behaviour, and any experimental intervention that affects this will possibly influence behaviour, regardless of the intervention’s effect on neurogenesis. Behaviour is a “final common pathway” of countless factors and biological processes, making it *a priori* unlikely that any single factor will dominate.
Third, one could argue that the experiments examined here are too simple to establish causation, and more complex experiments that both increase (e.g. exercise) and decrease (e.g. irradiation) neurogenesis are required, thus ruling out other explanations by experimental means. It is still possible however to be misled with these “$2\times 2$” designs when using standard statistical methods [@Lazic2012]. The problem lies in how inferences are made rather than the experimental design. Many studies attempt to address the issue of off-target effects by using multiple methods to alter levels of neurogenesis. While this provides a qualitative way to increase confidence in the relevance of neurogenesis, a series of biased estimates still does not adequately address the issue.
The fourth point relates to limitations of the analysis. The mediation model does not provide proof of neurogenesis’ role in behaviour (even if the effect was significant), as there might be another variable which is mediating the effect of the intervention and neurogenesis just happens to be correlated with it [@Fiedler2011]. This analysis can provide a necessary but not sufficient condition for demonstrating a causal role for neurogenesis. The three-variable model used here does not capture the full biological complexity, and future studies should include other variables in more detailed and realistic models. This would provide a better understanding of how multiple mechanisms interact and their relative importance for behaviour. For example, the effect of neurogenesis might be underestimated if there are compensating mechanisms that counteract the effect of increased or decreased neurogenesis. An assumption of the mediation analysis is that the behavioural training and testing has no—or at least a negligible—effect on neurogenesis. This will likely not be a problem however unless the effect of training and testing on neurogenesis differs between experimental groups. A final point is that the mediation models used here assume that all groups have at least some animals with neurogenesis. The models would not work well if all animals in the experimental group had a cell count of zero for example, as there needs to be some variation in neurogenesis to account for variation in behaviour.
Fifth, the results of the combined analysis are only as good as the studies that go into it. Remarkably few studies reported the covariation between neurogenesis and behaviour (which explains why so few studies were included despite a large number of published papers), but it is an obvious result to mention since it directly addresses the research hypothesis. The results are likely biased in favour of neurogenesis if studies only displayed the association between neurogenesis and behaviour if it was statistically significant and in the expected direction; and even studies that reported these results did so selectively in that not all neurogenesis–behaviour associations were mentioned. Publication bias is therefore a concern and is known to be a problem in the preclinical biomedical literature [@Sena2010; @Tsilidis2013]. The neurogenesis literature also suffers from the problem of multiplicity: multiple behavioural tests are often conducted, with multiple outcomes for each behavioural test, multiple markers of proliferation or new neurons are used (e.g. Ki67, BrdU, BrdU/NeuN, DCX, BrdU/DCX, etc.), and multiple divisions of the hippocampus are possible (the whole dentate gyrus, dorsal versus ventral, or infrapyramidal versus suprapyramidal blade, etc.). There may be good reasons for making such anatomical distinctions and the point is that there are many neurogenesis–behaviour correlations that can be conducted per study. Such multiple testing increases the chance of false positives and is rarely taken into account. To this one can add variations in analysis such as with and without assuming equal variances, log transformations, non-parametric methods, the removal of inconvenient data points, etc. This flexibility in choice of analysis has been referred to as “researcher degrees of freedom” and can also increase the number of false positives [@Simmons2011]. It is also rare for studies to define primary histological and behavioural outcomes before conducting the study, and thus any significant result is taken as evidence that neurogenesis is relevant for behaviour. After some fifteen years of neurogenesis research, surely we are in a position to define primary outcomes before conducting the experiment—in other words, to make a specific prediction about the expected relationships [@Wagenmakers2012a; @Kimmelman2014]. The above factors are recognised as important for the reproducibility of research [@Landis2012], as is the application of appropriate statistical methods to make valid conclusions, rather than inferential leaps unsupported by the data.
Finally, while this article has been mostly critical, it is worth reflecting on some positive aspects. Compared with other fields where causal explanations are sought, experimental biologists have a number of advantages, including the use of randomisation, a high degree of experimental control, the ability to create large effects, access to homogeneous sample material, ease of reproducing the treatment across subjects (i.e. all animals “comply” with the treatment), and dealing with directly observable phenomena such as cell counts and behaviour. These advantages—coupled with statistical models that directly estimate the hypothesis of interest—provide a powerful set of methods for understanding brain–behaviour relationships.
Materials and Methods {#materials-and-methods .unnumbered}
=====================
To find relevant data a PubMed search using the phrase “neurogenesis AND (dentate OR hippocamp\*) AND (behavior OR behaviour) AND English \[LA\] NOT review \[PT\]” was conducted and returned 990 papers, and included publications up to 4 April 2014. The flow diagram can be found in the Supporting Information (Text S1). Not all 990 papers were relevant as some were review articles while others were *in vitro* studies. The papers were examined for *in vivo* studies that either assigned animals to different treatment conditions or used naturally occurring groups such as old versus young. Otherwise, all animal models, methods of manipulating neurogenesis, behavioural outcomes, and study designs were eligible for inclusion. However, only data presented as scatterplots (neurogenesis versus behaviour) could be extracted and therefore used. None of the studies provided the raw data as supplementary material, otherwise they would have been included. In addition, it was necessary for each animal to be distinguished by experimental group; for example, with different colours or shapes for the plotting symbols, or if the groups were plotted in separate panels or figures. Only fifteen studies remained after applying the above criteria and the data were accurately extracted from the figures using g3data software (www.frantz.fi/software/g3data.php). Two studies were subsequently excluded as over-plotting of points made it impossible to extract the data in one, and the control group was omitted from the scatterplot in the second. Two additional papers had a strong group $\times$ neurogenesis interaction on the behavioural outcome. These were not included in the combined analysis but are discussed in the Supporting Information (Text S1). There were therefore eleven studies in the final analysis with 215 animals.
The mediation models were implemented as Bayesian graphical models and were fit in R (version 3.0.3) using the R2jags package (version 0.04-01) and JAGS (Just Another Gibbs Sampler; version 3.4.0). Measures of neurogenesis and behaviour were standardised to have a mean of zero and standard deviation of one, which allowed the effects to be pooled across studies. Note that this only puts the results of the studies on a common scale, but does not otherwise change them. The results are therefore interpreted in standard deviation units; for example, an estimated effect of 0.2 for neurogenesis means that the experimental intervention via its effect on neurogenesis increases behaviour by 0.2 standard deviations relative to the control group. A similar interpretation applies for the neurogenesis-independent effect. For some behavioural outcomes a lower value indicates better performance and therefore the estimated effect of the intervention would be negative. In such cases the sign of the effect was reversed in the combined analysis so that positive estimates are always associated with better performance. The Markov chain Monte Carlo (MCMC) sampling used three chains of 1,000,000 iterations each, a burn-in period of 5000 iterations, and every tenth value was saved. The three chains were well mixed (Gelman-Rubin statistic $\leq$1.01 for all parameters). The main parameters of interest were the effect of the treatment or group on behaviour via neurogenesis-independent mechanisms ($\beta_4$ in Fig. 1B) and the effect via neurogenesis (which is equal to the product of coefficients $\beta_1 \times \beta_3$). Non-informative priors were used and the results were not sensitive to the form of the prior. Further details of this modelling approach applied to neurogenesis data can be found in Lazic [@Lazic2012], and a general introduction to mediation analysis can be found in references [@Baron1986; @Shrout2002; @MacKinnon2007; @MacKinnon2009; @Hayes2009; @Yuan2009; @Hayes2013]. Causal models are discussed in detail by Pearl [@Pearl2009] and Bayesian graphical models by Koller and Friedman [@Koller2009].
After the estimates were obtained from each of the eleven studies they were combined with a random effects meta-analysis using the metafor R package (version 1.8-0; [@Viechtbauer2010]). The studies often contained multiple measures of neurogenesis, behavioural tests, and outcomes for a particular behavioural test (e.g. latency to immobility and total immobility time on the forced swim test). To avoid “double counting” only the most precise estimate from each study was used in the combined analysis, but all estimates are available in the Supporting Information (File S1). The data extracted from the published manuscripts are also provided in the Supporting Information (File S1) along with the R code. A protocol for this analysis has not been published.
The quality of the eleven studies that went into the combined analysis was assessed using five metrics that are known to introduce bias [@Hooijmans2014]. First, the number of studies mentioning random allocation of animals to different treatment groups (if applicable) was determined. Second, the number of studies indicating that the experimenter was blind when quantifying neurogenesis or when assessing behavioural outcomes was determined. The criteria used was generous in that if blinding was mentioned for only one of the above, the study was counted as having been blinded. Third, the number of studies that discussed how potential litter effects were handled was determined. Litter effects occur because littermates tend to be more alike on a variety of outcomes compared with animals from different litters. In other words, the variation between litters is greater than the variation within litters [@Lazic2013]. If experiments are not designed with litter effects in mind, both the false positive and false negative rates can be dramatically increased. Fourth, a check on data consistency was performed by examining whether data points were removed without comment. This can be determined by inconsistencies in numbers of data points across relevant figures, a discrepancy between the number of animals reported in the methods compared with those shown in the figures, or when the reported degrees of freedom do not match the indicated sample size and method of analysis [@Lazic2010]. Finally, selective reporting was defined as the reporting of only a subset of all possible neurogenesis–behaviour associations. For example, if three behavioural outcomes are described, but the result of only one association with neurogenesis is reported. Or, if multiple measures of neurogenesis are described (e.g. Ki67 and DCX) but the result of only one association with behaviour is reported. The quality metrics were not used in any way in the meta-analysis.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous reviewers for constructive comments.
Figure Legends {#figure-legends .unnumbered}
==============
![[**Relating neurogenesis to behaviour.**]{} A typical analysis (A) only demonstrates that a treatment or experimental intervention affects neurogenesis (p-value for $\beta_1$ is $<0.05$) and behaviour (p-value for $\beta_2$ is $<0.05$). An unjustified inferential leap is then made by concluding that changes in neurogenesis are responsible for changes in behaviour (dashed arrow), thus ignoring other potential explanations. The appropriate mediation analysis (B) estimates the effect of the treatment that is mediated by neurogenesis ($\beta_1 \times \beta_3$) and via all other mechanisms ($\beta_4$).[]{data-label="fig:models"}](models)
![[**Study specific and combined estimates from the causal mediation analysis.**]{} The effect of neurogenesis was smaller than the effect of other mechanisms in all studies. The overall effect of neurogenesis was small, non-significant (p = 0.128), and unlikely to account for a large proportion of the observed behavioural effects (upper 95% CI = 0.34). PMID = PubMed ID; DA = Dark agouti; SD = Sprague-Dawley; MWM = Morris water maze; OF = open field; FST = forced swim test; B/W = black/white discrimination task. Error bars are 95% credible intervals.[]{data-label="fig:meta_results"}](combo_results)
![[**Association between neurogenesis and forgetting.**]{} (A) Raw data extracted from Figure 4F in Akers et al. [@Akers2014] and estimated regression lines. While in the predicted direction (negative association), the relationship is not significant within groups (p = 0.12). (B) The causal mediation analysis divides the total effect of the TK transgene into the part that can be attributed to neurogenesis and to all other mechanisms. The effect of neurogenesis accounted for less than half (40%) of the total effect, although there is a great deal of uncertainty in this estimate. Error bars are 95% (thin) and 50% (thick) credible intervals.[]{data-label="fig:freezing"}](freezing)
Supporting Information Legends {#supporting-information-legends .unnumbered}
==============================
[**Text S1. Supplementary information.**]{} Contains the meta-analysis flow diagram and a discussion of two studies with strong group $\times$ neurogenesis interaction effects.
[**File S1. Data files and R code.**]{}
|
{
"pile_set_name": "ArXiv"
}
|
0
@addtoreset
NSF-ITP-96-75\
hep-th/9608104
**$K3$-Fibrations and Softly Broken $N\!=\!4$ Supersymmetric\
**
Gauge Theories.
César Gómez$^a$, Rafael Hernández$^{a,b}$ and Esperanza López$^c$
$^a$[*Instituto de Matemáticas y Física Fundamental, CSIC,\
Serrano 123, 28006 Madrid, Spain*]{}
$^b$[*Departamento de Física Teórica, C-XI, Universidad Autónoma de Madrid,\
Cantoblanco, 28049 Madrid, Spain*]{}
$^c$[*Institute for Theoretical Physics, University of California,\
Santa Barbara, CA 93106-4030*]{}
Global geometry of $K3$-fibration Calabi-Yau threefolds, with Hodge number $h_{2,1}=r+1$, is used to define $N\!=\!4$ softly broken $SU(r+1)$ gauge theories, with the bare coupling constant given by the dual heterotic dilaton, and the mass of the adjoint hypermultiplet given by the heterotic string tension. The $U(r+1)$ Donagi-Witten integrable model is also derived from the $K3$-fibration structure, with the extra $U(1)$ associated to the heterotic dilaton. The case of $SU(2)$ gauge group is analyzed in detail. String physics beyond the heterotic point particle limit is partially described by the $N\!=\!4$ softly broken theory.
Introduction
============
The Seiberg-Witten solution [@SW]-[@8] of $N\!=\!2$ supersymmetric gauge theories turns out to be intimately related to both integrable models [@in]-[@9] and string theory [@KV]-[@KLMVW]. The mass formula for BPS states $$M = \sum_{i=1}^{r} \| n_{i}^{e} a_{i}(\vec{u}) +
n_{i}^{m} a_{i}^{D}(\vec{u}) \|,$$ where $r$ is the rank of the gauge group of the theory, admits a geometrical representation in terms of the periods, $$a_{i}(\vec{u}) = \oint_{\alpha_i} \lambda, \: \: \: \: \: \: \: \:
a_{i}^{D}(\vec{u}) = \oint_{\beta_i} \lambda,
\label{eq:I2}$$ of a meromorphic form $\lambda$ on an hyperelliptic curve of genus r, $\Sigma_{\vec{u}}$, with $\vec{u}=u_{1}, \ldots ,u_{r}$ and $u_i$ the Casimir expectation values.
The solution given by (\[eq:I2\]) defines a family of abelian varieties, i.e. the Jacobian of $\Sigma_{\vec{u}}$, parameterized by the quantum moduli manifold. This is the generic structure underlying algebraic integrable models [@DM]. As it was shown in [@DW], an integrable model describing $N\!=\!2$ gauge theories can be directly derived, through Hitchin’s construction [@H], from a two dimensional Higgs system defined on an elliptic Riemann surface, $E_{\tau}$. This approach leads to the Seiberg-Witten solution for $N\!=\!2$ gauge theories with one massive hypermultiplet in the adjoint representation. The pure gauge theory appears as a double scaling limit in which hypermultiplet is decoupled $(m
\rightarrow \infty )$, and the bare coupling constant is sent to $\infty$.
In addition, the geometrical representation in terms of periods (\[eq:I2\]) of the BPS mass formula, calls for a stringy interpretation of Seiberg-Witten geometry based on some non critical string, with effective string tension $\lambda$, winding around an “internal” space described by the Riemann surface $\Sigma_{\vec{u}}$ [@KLMVW]. This string interpretation nicely combines two ingredients, namely the string derivation of Seiberg-Witten solution, and its integrable model representation. The first one arises from the discovery of heterotic-type II dual pairs [@HT; @svd; @KV; @SQ], with the heterotic string compactified on $K3
\times
T^2$. Higgsing the gauge group leads to different heterotic compactifications having $r+2$ vector excitations and $s$ neutral hypermultiplets, therefore the corresponding type IIA dual should be defined on a Calabi-Yau threefold with $h_{1,1}=r+1$, and $h_{2,1}=s-1$. These Calabi-Yau threefolds provide the string theory extension of Seiberg-Witten quantum moduli for a gauge group of rank $r$. An important step in the construction of dual pairs is the use of threefolds which are $K3$-fibrations [@K3]. The reason for this requirement goes back to the identification of the Calabi-Yau modulus that we could put in correspondence with the heterotic dilaton. This moduli is singled out, for $K3$-fibrations, as the one associated with the size of the $\IP^{1}$ base space [@AL].
On the other hand, the integrable model representation of the Seiberg-Witten curve for pure gauge theories, can be defined as a fibration of the spectral cover set given by the vanishing locus of the Landau-Ginzburg potential associated to the Dynkin diagram of the corresponding gauge group [@MW]. The relation between this $0$-dimensional fibration, and the $K3$-fibration defining the Calabi-Yau threefold can be now obtained by fixing the $K3$-fiber at an orbifold point, described by the corresponding Dynkin diagram, and blowing it up to an ALE space [@KLMVW]. This geometrical manipulation can be formally undertaken by turning gravity off $(S \rightarrow \infty)$ and simultaneously going to the point particle regime by sending the string tension to zero [@KKLMV]. Within this approach, a string interpretation of the periods (\[eq:I2\]) is obtained as the wrapping of a $3$-brane on a $3$-cycle of the Calabi-Yau threefold. To put it more precisely, the meromorphic form $\lambda$ appears as the string tension of the self-dual string obtained when wrapping a $3$-brane on $2$-cycles of the ALE-space.
In this paper, we will try to go one step further in the study of the deep interplay between $K3$-fibered Calabi-Yau threefolds and the integrability underlying the Seiberg-Witten solution. In order to do that we will choose, continuing the line in reference [@GHL], the approach to integrability used by Donagi and Witten [@DW], based on Hitchin’s gauge model on $E_{\tau}$, where one naturally lands onto the Seiberg-Witten solution for $N\!=\!4$ softly broken gauge theories. The physical reason for such a choice is related to the existence, in addition to the Higgs moduli, of two extra parameters, namely, the bare coupling constant $\tau$ and the bare mass for the hypermultiplet in the adjoint representation. Our main goal will be to find a string derivation of this theory. Moreover we will claim that $K3$-fibered threefolds naturally lead to the $N\!=\!4$ softly broken version of $N\!=\!2$ gauge theories, with the bare coupling constant $\tau$ and the hypermultiplet mass mapped in a concrete way into the heterotic dilaton and the string scale $\alpha'$ respectively.
The main difference between our case and the string derivation of pure $N\!=\!2$ gauge theory is that, in order to obtain the $N\!=\!4$ softly broken version, we should work beyond the point particle limit, at generic values of the heterotic dilaton field. Thus, we will be using the global information on the $K3$-fibration structure. The important physical question raised by this analysis is, of course, to unravel the type of global string dynamics that we are capturing with the $N\!=\!4$ softly broken quantum field theory. We will concentrate mainly in the case of gauge group $SU(2)$. What we observe is that the gauge theory that we are associating with the threefold fibration captures the features which are universal for a certain set of $K3$-fibrations, with identical Hodge number $h_{2,1}=2$, but differing in the modular properties of the mirror map and in Hodge number $h_{1,1}$. In particular, following [@AP], we will relate the different modular properties of the mirror map to the Kac-Moody level at which the gauge symmetry is realized in the heterotic dual string. The $N=4$ softly broken theory will then determine the relation between the Kac-Moody level and the genus of a curve of singularities developed by the Calabi-Yau manifolds at the locus corresponding to vanishing heterotic dilaton [@KM; @KMP].
The $N\!=\!4$ softly broken version of the Calabi-Yau threefold singles out a particular singular locus on the moduli of complex structures of the threefold, namely that in correspondence with the field theory locus where some component of the adjoint hypermultiplet becomes massless. If the mass of the hypermultiplet is correctly capturing part of the dynamics of the string scale, $\alpha'$, we should expect the corresponding singular locus of the Calabi-Yau threefold to be somehow related to the existence of the string scale, in very much a similar way as the self dual point for compactifications on $S^{1}$ is given by a radius $R$ equal to $\sqrt{\alpha'}$. We will present some evidence in this direction.
Concerning Donagi-Witten version of integrability as based on Hitchin’s gauge model on an elliptic curve $E_{\tau}$, we will observe that this “reference Riemann surface” can also be recovered from the geometrical data of the $K3$-fibration.
The Calabi-Yau Curve and its Quantum Field Theory Analogue.
===========================================================
$K3$-Fibrations.
----------------
Let us consider the string embedding of $SU(2)$ $N=2$ supersymmetric Yang-Mills theory according to [@KV]. The Higgs mechanism that produces the desired gauge group originates 129 neutral hypermultiplets, therefore we must choose as the type IIA dual a $K3$-fibered threefold with Hodge numbers $h_{1,1}=2$, $h_{2,1}=128$, i.e. $W=\IP_{\{1,1,2,2,6\}}^{4}[12]$. The mirror manifold $W^*$ can be obtained from the orbifold construction [@GP], and has defining polynomial $$W^{\ast} = \frac{1}{12} x_{1}^{12} + \frac{1}{12} x_{2}^{12} +
\frac{1}{6} x_{3}^6 + \frac{1}{6} x_{4}^6 + \frac{1}{2} x_{5}^2
- \psi x_1 x_2 x_3 x_4 x_5 - \frac{1}{6} \phi (x_1 x_2)^6.
\label{1000}$$ The moduli space of complex deformations of $W^*$ is parameterized by $(\psi,\phi)$, subject to the global symmetry $${\cal A}:(\psi,\phi) \rightarrow (\beta \psi, - \phi)~, \hspace{1cm}
\beta^{12}=1.
\label{555}$$ This symmetry forces to introduce invariant quantities; we will use $b=1/\phi^2$ and $c=-\phi / \psi^6$. The $K3$-fibration structure of (\[1000\]) becomes manifest by the change of variables $x_1/x_2 \equiv z^{1/6}b^{-1/12}$, $x_{1}^2 \equiv x_0 z^{1/6}$ [@KLMVW]: $$W^{\ast}= \frac{1}{12}(z+\frac{b}{z}+2) x_{0}^6 + \frac{1}{6} x_{3}^6
+
\frac{1}{6} x_{4}^6 + \frac{1}{2} x_{5}^2 + c^{-1/6} x_0 x_3 x_4
x_5,
\label{2}$$ with the variable $z$ acting as coordinate on the $\IP^1$ base space. It is convenient to define $$d(z;b)=\frac{1}{2}(z+\frac{b}{z}+2)~, \hspace{1cm}
\hat{c}(z;b,c)=c \: d(z;b).
\label{557}$$ Substituting this into (\[2\]) and rescaling $x_0$, $W^{\ast}$ acquires the explicit form of a $K3$-surface $$W^{\ast} = \frac{1}{6} x_{0}^6 + \frac{1}{6} x_{3}^6 +
\frac{1}{6} x_{4}^6 + \frac{1}{2} x_{5}^2 +
\hat{c}(z;b,c)^{-1/6} x_0 x_3 x_4 x_5.
\label{1001}$$ As we move in $\IP^1$, the $K3$-fiber can become singular. From (\[1001\]) it is easy to deduce that this occurs for the $K3$ modulus values $\hat{c}(z;b,c)=0,1$. These values of $\hat{c}$ are acquired at the following $\IP^1$ points, $z=e_{i}^{\pm}$: $$\begin{aligned}
\label{3}
\hat{c}=0 & \rightarrow & e^{\pm}_0 = -1 \pm \sqrt{1-b}, \nonumber
\\
\hat{c}=1 & \rightarrow & e^{\pm}_1 = \frac{1 - c \pm
\sqrt{(1-c)^2 - b c^2}}{c}.\end{aligned}$$ The discriminant of (\[1001\]) is therefore given by $\Delta(z;b,c)=
\prod_{i=0}^{1} (z-e^{+}_i(b,c))(z-e^{-}_i(b,c))$. There is an additional singularity at $\hat{c}(z;b,c)=\infty$, which is originated in the quotient by discrete reparameterizations of (\[1001\]) inherited from the orbifold construction of $W^{\ast}$. It corresponds to the points $$\begin{aligned}
\hat{c}=\infty & \rightarrow & z=0,\infty \hspace{1cm}
(b \neq \infty).
\label{560}\end{aligned}$$
The Calabi-Yau manifold becomes singular when some of the points (\[3\])-(\[560\]) coalesce. We will now analyze the regions in moduli space where this situation happens [@can; @CYY] (we will follow notation in [@can]). The loci $$\begin{aligned}
\label{l1}
{\cal C}_1 & = & \{ b=1 \}, \nonumber \\
{\cal C}_C & = & \{(1-c)^2-bc^2=0\} ,
\label{eq:c4}\end{aligned}$$ are respectively obtained from the identifications $e_0^+=e_0^-$ and $e_1^+=e_1^-$. ${\cal C}_C$ is the conifold locus, where $3$-cycles of the threefold degenerate to points, while ${\cal C}_1$ corresponds to the appearance of a genus two curve of $A_1$ singularities. We can also consider $$\begin{aligned}
{\cal C}_0 & = & \{ c = \infty \}, \nonumber \\
{\cal C}_{\infty} & = & \{ b = 0 \},
\label{eq:c5}\end{aligned}$$ which are defined, respectively, by the identifications $e_1^{\pm}=e_0^{\pm}$ and $e_0^+=e_1^+=0$. ${\cal C}_0$ is an orbifold locus, given by the fixed points under ${\cal A}^2$. ${\cal C}_{\infty}$ corresponds to the weak coupling limit locus, once we identify the heterotic dilaton with the size of the base space [@AL]. In addition we have $${\cal D}_{(0,-1)} = \{ c = 0 \},
\label{l2}$$ implying $e_{1}^+=0$, $e_{1}^-=\infty$. ${\cal D}_{(0,-1)}$ is an exceptional divisor in moduli space whose intersection with ${\cal C}_{\infty}$ identifies the large complex structures limit of (\[1000\]). Finally, let us notice that at $b=\infty$ the points (\[560\]) are ill defined, giving raise to a very degenerate situation. We can put in correspondence $${\cal D}_{(-1,0)} = \{ b = \infty \},
\label{556}$$ with the exceptional divisor introduced to resolve a conical singularity in the moduli space generated from quotienting by the discrete transformation ${\cal A}$[^1]. In section 2.3 we will return again to this point, in relation with a double covering of the moduli space[^2].
More in general, we can consider $K3$-fibered Calabi-Yau threefolds whose mirror $W^*$ is also a $K3$-fibration that can be written as [@KLMVW] $$W^{\ast}=\frac{1}{2n}\left( z + \frac{b}{z} +2 \right) x_{0}^n +
W_{K3}^{\ast} (x_0,x_3,x_4,x_5;c_i).
\label{999}$$ The discriminant of the fiber will be given by a polynomial $\Delta(z;b,c_i)$, depending on the point $z$ in the base and the moduli parameters, with zeroes at $z=e_i^{\pm}$ where the $K3$-fiber becomes singular. The number of singular points is given by $2 h_{2,1}=2 (r+1)$.
The Calabi-Yau Curve.
---------------------
From the above $K3$-fibration structure it is possible to recover [@KLMVW], in the heterotic point particle limit $b \rightarrow 0$, $\alpha'
\rightarrow 0$ [@KKLMV], the Seiberg-Witten curves for $N\!=\!2$ $SU(r+1)$ Yang-Mills theory [@KLTY; @AF; @MW]. Namely for the $SU(2)$ case, using the map $$\begin{aligned}
b & = & \alpha'^{2} \Lambda^{4}, \nonumber\\
c & = & 1 + \alpha' u,
\label{eq:cc1}\end{aligned}$$ with $\Lambda$ the $SU(2)$ dynamical scale, and rescaling $z \rightarrow \alpha' z$, the points (\[3\]) become the branch points of the associated Seiberg-Witten curve [@MW] $$\begin{aligned}
e_0^{\pm} & = & 0, \infty , \nonumber\\
e_1^{\pm} & = & - u \pm \sqrt{u^2-\Lambda^4}.
\label{888}\end{aligned}$$ In general, the geometrical meaning of the point particle limit amounts to replacing the $K3$-fiber by an ALE space that blows up an $A_r$ orbifold singular $K3$. Let us denote by $G=SU(r+1)$ the gauge group whose Dynkin diagram describes the orbifold singularity. In this situation, the branched cover of the $\IP^1$ base space that defines the Seiberg-Witten curve for the pure gauge theory in the integrable model formulation of [@MW] (whose branch points are at the generalization of (\[888\])), ${\cal C}_G$, can be directly derived from the ALE space homology.
A different approach consists in using the global structure of the $K3$-fibration in order to build an hyperelliptic curve such that its moduli space is isomorphic to the Calabi-Yau moduli. We can define $${\cal C}^{CY}: \hspace{5mm} y^2=\Delta(z;b,c_i)=\prod_{i=0}^{r}
(z-e^{+}_i(b,c_i))(z-e^{-}_i(b,c_i)).
\label{eq:c13}$$ By construction, this curve becomes singular at the moduli values where the threefold $W^*$ acquires a singularity, i.e. those where two roots coalesce. Performing the point particle limit [@KKLMV], ${\cal C}^{CY}$ reduces to the Seiberg-Witten curve ${\cal C}_G$. However, it is important to stress the following. In order to define ${\cal C}^{CY}$ we have interpreted the singular points of the $K3$-fibration as branch points of an associated hyperelliptic curve. But contrary to the point particle limit, in which the $K3$-fiber is substituted by an ALE space, we are not using the $K3$ periods for the direct construction of the curve.
Our first aim in analyzing (\[eq:c13\]) will be, following previous work in reference [@GHL], to identify the $N=2$ supersymmetric gauge theory represented by ${\cal C}^{CY}$. Being defined for arbitrary values of the moduli parameters, the first difference between ${\cal C}_G$ and ${\cal C}^{CY}$ is that the second depends on the “dilaton modulus” $b$. In the heterotic string framework, the expectation value of the dilaton field determines the bare gauge coupling constant. Therefore we should look for a gauge theory in which the coupling constant behaves as a modulus, i.e. an ultraviolet finite theory.
The simplest candidate for ${\cal C}^{CY}$ is a theory with the field content of $N=4$, namely Yang-Mills plus a matter hypermultiplet in the adjoint representation of $G$, where in general the adjoint hypermultiplet can be massive. Let us consider the case where the adjoint hypermultiplet has a bare mass $m$. It was shown in [@SW2] that in a double scaling limit that sends $m$ and the $N=4$ coupling constant to $\infty$, we can recover the pure Yang-Mills theory. In this process the mass plays a role formally analogous to that of the string coupling constant in the double limit that takes ${\cal C}^{CY}$ to ${\cal C}_G$. Therefore, in the proposed interpretation of string notions in terms of gauge theory quantities, we should identify $(\alpha')^{-1} \sim m^2$.
Let us now review briefly some results in reference [@GHL] for the case $G=SU(2)$. The curve for $SU(2)$ $N=2$ gauge theory with one massive adjoint hypermultiplet [@SW2], is given by $$y^2= (x - a_1 \hat{u} + a_{2}^2)(x + a_2 (\hat{u} - a_1))
(x - a_2 (\hat{u} - a_1)),
\label{11}$$ with $u= \mbox {Tr} \phi^2$ the quadratic Casimir, and $\hat u$ defined by the convenient normalization $\frac{1}{4}m^2 {\hat u}= u$. The quantities $a_1$ and $a_2$ depend on the asymptotic value of the gauge coupling constant of the theory, $\tau=\frac{\theta}{2\pi}+
i \frac{4 \pi}{g^2}$, according to[^3] $$a_1= \frac{3}{2}e_1(\tau)~~, \hspace{1cm} a_2= \frac{1}{2} (e_3(\tau)
- e_2(\tau)).$$ In the moduli space $(\hat{u}, \tau)$, we can differentiate the following loci where (\[11\]) becomes singular $$\begin{array}{lcl}
\hat{{\cal C}}_{0} & = & \{ \hat{u}(\tau)= a_1(\tau) \}, \\
\hat{{\cal C}}_{C}^{(1)} & =
& \{ \hat{u}(\tau)= a_2 (\tau) \}, \\
\hat{{\cal C}}_{C}^{(2)} & = & \{ \hat{u}(\tau)= -a_2(\tau) \}, \\
\hat{{\cal D}} & = & \{ \hat{u} = \infty \}.
\end{array}
\hspace{1cm}
\begin{array}{lcl}
\hat{{\cal C}}_{\infty} & = & \{ \tau= i \infty \}, \\
\hat{{\cal C}}_{1}^+ & = & \{\tau = 0 \}, \\
\hat{{\cal C}}_{1}^- & = & \{\tau = 1 \}, \\
&&
\end{array}
\label{12}$$
We can now try to put in correspondence the moduli space of Kähler deformations of $\IP^{4}_{\{1,1,2,2,6\}}[12]$ with the moduli space of the $N=4$ softly broken theory (\[11\]). The basic idea will be to map the singular loci described by the fibration structure of the mirror $W^{\ast}$, (\[l1\])-(\[l2\]), with the set (\[12\]). This is achieved by the map [@GHL] $$c=\frac{a_1(\tau)}{a_1(\tau)-\hat{u}}~~, \hspace{1cm}
b=\left( \frac{a_2 (\tau)}{a_1(\tau)} \right)^2.
\label{14}$$ It is important to notice, using the modular properties of $a_i(\tau)$[^4], that (\[14\]) is effectively quotienting by the $(\hat{u},\tau)$-plane transformation $$T:(\hat{u},\tau) \rightarrow (\hat{u},\tau +1).
\label{eq:c11}$$ Indeed the proposed map sends $\hat{{\cal C}}_{C}^{(1,2)} \rightarrow
{\cal C}_{C}$ and $\hat{{\cal C}}^{\pm}_1 \rightarrow {\cal C}_1$, while $\hat{{\cal C}}_{\infty}$, $\hat{{\cal C}}_{0}$ and $\hat{{\cal
D}}$, which are fixed under (\[eq:c11\]), are mapped respectively into ${\cal C}_{\infty}$, ${\cal C}_0$ and ${\cal D}_{(0,-1)}$[^5]. Therefore we observe that the $(\hat{u},\tau)$-plane behaves as a double cover of the Calabi-Yau moduli space.
In the weak coupling limit $\tau \rightarrow i \infty$, the map (\[14\]) becomes $$c=\frac{1}{1-\hat{u}}~~, \hspace{1cm}
b=64 e^{2 \pi i\tau}.
\label{15}$$ From the heterotic-type II identification at leading order $b=e^{-S}$ [@KV], with $S$ the heterotic dilaton, the above expression explicitly shows that we are associating $\tau$ with $S$. Setting $(\alpha')^{-1}=\frac{1}{4}m^2$, as we proposed, (\[15\]) reproduces the relation $\frac{c^2 b}{(1-c)^2}=\frac{\Lambda^4}{u^2}$ used in [@KKLMV] for defining the point particle limit of the string (of which (\[eq:cc1\]) is the first order).
We would like now to prove that, by the map (\[14\]), the curve ${\cal C}^{CY}$ is in fact the Seiberg-Witten curve for $N\!=\!2$ $SU(2)$ Yang-Mills theory with one massive hypermultiplet in the adjoint representation. In order to do so, we rewrite the Calabi-Yau curve (\[eq:c13\]) for the $SU(2)$ case as $$y^2= \prod_{i=0,1} (z-e_{i}^+(\hat{u},\tau))(z-e_{i}^-
(\hat{u},\tau)),
\label{16}$$ with $$\begin{aligned}
\label{510}
e_{0}^{\pm} & = & \frac{-a_1(\tau) \pm \sqrt{ a_1(\tau)^2 -
a_2(\tau)^2}}{ a_1(\tau)}, \nonumber \\
e_{1}^{\pm} & = & \frac{- \hat{u} \pm \sqrt{ {\hat u}^2 -
a_2(\tau)^2}}{ a_1(\tau)}.\end{aligned}$$
It is convenient now to transform the Calabi-Yau curve (\[16\]) into the standard cubic form $y^2=z(z-1)(z-\lambda)$, where $$\lambda=\frac{(e_{1}^{-} - e_{0}^+)(e_{0}^{-} -
e_{1}^+)}{(e_{1}^{+} - e_{0}^+)(e_{0}^{-} - e_{1}^-)}.
\label{17}$$ Using results in Appendix A, we observe that the field theory curve (\[11\]) is isogenic to the quartic $$y^2=(x^2 + a_1 \hat{u} - a_{2}^2)^2 - (a_2(\hat{u}-a_1))^2.
\label{20}$$ The $\lambda$ parameter that characterizes the standard cubic form of (\[20\]) is given by $$\lambda'= \frac{a_1 \hat{u}-a_{2}^2 + \sqrt{(a_{1}^2 -
a_{2}^2)(\hat{u}^2 - a_{2}^2)}}{a_1 \hat{u} - a_{2}^2 -
\sqrt{(a_{1}^2 -
a_{2}^2)(\hat{u}^2 - a_{2}^2)}}.
\label{511}$$ Substituting (\[510\]), we can see that $\lambda$ and $\lambda'$ precisely coincide.
This concludes the proof that ${\cal C}^{CY}$ for the Calabi-Yau mirror of $\IP_{\{1,1,2,2,6\}}^{4}[12]$ is, by the map (\[14\]), the curve describing $SU(2)$ $N=2$ supersymmetric Yang-Mills with one massive hypermultiplet in the adjoint representation. The necessity of introducing an isogeny transformation is originated in different conventions for the Higgs field normalization. The curve (\[11\]) follows the convention in [@SW], which is appropriate when only integer charges for BPS states can appear. However the Seiberg-Witten curve ${\cal C}_G$, obtained from the point particle limit of the string, adopts a normalization adequated to gauge theories that can include fundamental matter [@SW2]. Since ${\cal C}^{CY}$ flows to ${\cal C}_G$ in the point particle limit, it shares with it the same normalization of the Higgs field, differing, up to an isogenic transformation (see (\[A1\])), of that in (\[11\]).
Double Covering and $K3$-Fibrations.
------------------------------------
We have pointed out that the $(\hat{u},\tau)$-moduli space of the $SU(2)$ theory with adjoint matter, using the map (\[14\]), acts as a double covering of the Calabi-Yau $\IP_{\{1,1,2,2,6\}}[12]$ moduli space $(b,c)$. At the same time, we have seen that the Seiberg-Witten curve for this gauge theory can be explicitly constructed from string theory compactification. This raises the following puzzle: if the string moduli space is correctly labeled by $(b,c)$, how can the double covering variables $(\hat{u},\tau)$ be naturally derived from the Calabi-Yau moduli?
The answer comes from the way we define the $K_3$-fibration. In (\[2\]) the choice $\sqrt{b}=- \frac {1}{\phi}$ has been implicitly done. If we instead choose $\sqrt{b}= \frac {1}{\phi}$, the fibration structure (\[557\]) changes into $$\hat{c}~'= \frac{1}{2}\left(-z-\frac{b}{z} +2 \right)c.
\label{eq:c20}$$ We see that the choice of one or another branch of $\sqrt{b}$ amounts to changing $z \rightarrow -z$. This is an effect that can be absorbed in a trivial redefinition of $z$, and therefore does not, in essence, affect ${\cal C}^{CY}$. However it shows how the $K_3$-fibration structure, and any notion based on it, leads to a double covering of the moduli space determined by the two branches of $\sqrt{b}$. Moreover in terms of the map (\[14\]), written as $$\sqrt{b}=\frac{a_2(\tau)}{a_1(\tau)},
\label{200}$$ the change of branch becomes equivalent to the transformation $T: \tau \rightarrow \tau +1$.
This problem is related to a subtlety in the derivation of the Seiberg-Witten solution for $SU(2)$ Yang-Mills in reference [@KKLMV], accomplished by performing the blow up of the tangency point $(b=0,c=1)$ between the conifold locus ${\cal C}_C$ and the weak coupling locus ${\cal C}_{\infty}$. The variable parameterizing the second exceptional divisor arising from the blow up of the tangency, $\frac {bc^2}{(1-c)^2}$, is identified with the field theory [**Z**]{}$_2$ invariant quantity $\frac
{\Lambda^4}{u^2}$. The reason why the blow up approach recovers $(\Lambda^2/u)^2$ is that, from (\[eq:cc1\]), the change of branch $\sqrt{b} \rightarrow -\sqrt{b}$ implies $\Lambda^2 \rightarrow
-\Lambda^2$, and the Calabi-Yau moduli space naturally quotients by this transformation. On the contrary, from the $K3$-fibration in the ALE limit we get in a direct way the Seiberg-Witten curve, parameterized by $u$, and not its [**Z**]{}$_2$ quotient.
The transformation $T$ has indeed an string analogue. Using the map (\[14\]), $T$ is equivalent to the discrete symmetry (\[555\]) of the Calabi-Yau moduli space, ${\cal A}$. After performing the ${\cal A}$-quotient, the Calabi-Yau moduli space becomes isomorphic [@can] to the space $\IP^{2}_{\{1,1,2\}}$, which has a conical singularity at the origin (see footnote 1). This is precisely the geometry of the $T$-quotiented $(\hat{u},\tau)$-plane[^6]. We stress that the $(\hat{u},\tau)$ variables, although undo the ${\cal A}$-quotient, preserve an ${\cal A}^2$-quotient. This can be seen from the fact that ${\cal A}^2$ fixes the locus ${\cal C}_0$, which has a counterpart in the $(\hat{u},\tau)$ moduli space.
Finally we notice that $(\hat{u},\tau)$ points are not in a one-to-one correspondence with $(\sqrt{b},c)$. The divisor ${\cal D}_{(-1,0)}=\{ b=\infty \}$, which appears in the blow up of the conical singularity created by the ${\cal
A}$-quotient, presents non-trivial monodromy. However in the $(\hat{u},\tau)$-plane, having undone the quotient by ${\cal A}$, the value $b=\infty$ should not imply an additional monodromy locus. This is in fact the case. Since $b=\infty$ corresponds to $a_1=0$, from (\[14\]) we see that ${\cal D}_{(-1,0)}$ is blown down to the point $\hat{u}=0$. The fibration structure is also reflecting this remark through its dependence on the $K3$-fiber modulus $\hat{c}=cd$. The value $b=\infty$ implies $d=\infty$, so that the combination $\hat{c}$ blows down the line $(b=\infty,c)$ to a point.
Integrability.
==============
In the previous section we have compared ${\cal C}^{CY}$ with the curve for $SU(2)$ $N=2$ Yang-Mills with one massive hypermultiplet in the adjoint representation. In this section we will recover, from the Calabi-Yau manifold, the basic building elements used in the Donagi-Witten formulation [@DW] of the integrability of gauge theories with adjoint matter.
To start with, let us just briefly recall the construction in [@DW] for the simple case of $SU(2)$ gauge theory. According to Hitchin’s construction, we start with the elliptic curve $$E_{\tau}: \: \: y^2 = (x-e_1(\tau))(x-e_2(\tau))(x-e_3(\tau)),
\label{eq:i1}$$ on which a two dimensional Higgs field $\phi$ is defined as an holomorphic 1-form transforming in the adjoint representation of the $SU(2)$ gauge group. In terms of $\phi$, an spectral cover curve is defined through (\[eq:i1\]) and $$0=t^2 -x + A_2,
\label{eq:i2}$$ with $A_2$ related to the Higgs expectation value of the $N\!=\!2$ theory by $A_2 = u - \frac {1}{2}e_1(\tau)$. Equations (\[eq:i1\]) and (\[eq:i2\]) define a genus $2$ curve, symmetric under the [**Z**]{}$_2 \times$[**Z**]{}$_2$ transformation $t \rightarrow
-t$, $y \rightarrow -y$. The curve (\[11\]) we have employed in the previous section corresponds to the [**Z**]{}$_2 \times$[**Z**]{}$_2$ invariant part.
The integrable model version of pure gauge theory curves [@in; @MW], which is recovered from string theory in the point particle limit [@KLMVW], is given by $$z + \frac {\Lambda^4}{z} + 2 P_{A_1}(t,u)=0.
\label{eq:curve}$$ This can be derived from the “classical” expression (\[eq:i2\])[^7] by the “quantization” procedure $x \rightarrow -\frac{1}{2}(z + \frac {\Lambda^4}{z})$. Instead, in [@DW] the quantization of (\[eq:i2\]) is implemented by forcing $x$ to live on the elliptic curve (\[eq:i1\]), parameterized by the $N=4$ coupling constant $\tau$.
In order to show that the Donagi-Witten integrability construction is the natural extension of (\[eq:curve\]) when we move off the point particle limit, we have still to determine the elliptic curve $E_{\tau}$ from Calabi-Yau data. Associated to a $K3$-fibration threefold $$W^{\ast}=\frac{1}{n} d(z;b)~ x_{0}^n + W_{K3}^{\ast}
(x_0,x_3,x_4,x_5;c_i)
\label{100}$$ with $d(z;b)$ given by (\[557\]), we can always consider the following four points on the $\IP^1$ base space where the $K3$-fiber becomes singular $$\begin{array}{ccc}
d(z;b)=0 & \rightarrow & e_{0}^{\pm} = -1 \pm \sqrt{1-b}, \\
d(z;b)=\infty & \rightarrow & z=0,~\infty.
\label{eq:i6}
\end{array}$$ In the spirit of the previous section, we can define the elliptic curve associated to these data $$y^2=z(z-e_{0}^+)(z-e_{0}^-).
\label{105}$$ This curve is independent of the particular Calabi-Yau we are working with, as far as the $K3$-fibration structure is defined by $d=\frac{1}{2}(z+\frac{b}{z}+2)$. The curve (\[105\]) is characterized by the $\lambda$ parameter $$\lambda=\frac{e_{0}^-}{e_{0}^+},
\label{110}$$ which only depends on the dilaton modulus $b$.
We will follow the same path as in section 2.2 for comparing (\[105\]) with $E_{\tau}$. Namely, after shifting $x \rightarrow
x+\frac{1}{2}e_1$, we apply the isogeny described in Appendix A. This transforms $E_{\tau}$ into the quartic $$y^2=(x^2 +a_1(\tau))^2 - a_{2}(\tau)^2.$$ Using the map between $b$ and $\tau$ proposed in (\[14\]), the $\lambda$ parameter of this curve can be easily seen to coincide with (\[110\]), implying that (\[105\]) is isogenic to $E_{\tau}$. In this sense we notice that, according to (\[A1\]), the elliptic parameter of the curve (\[105\]) is $2\tau$.
Let us stress the meaning of the extra singularity at $d = \infty$ used in (\[eq:i6\]), which corresponds to $z=0,\infty$. The Seiberg-Witten differential for ${\cal C}^{CY}$, which we will describe in the begining of the next section (see (\[41\])), has poles with residue at the points $z=0,\infty$. The residue at these poles is defining the mass of the hypermultiplet in the adjoint [@SW2]. (In the point particle limit [@KKLMV] the structure (\[105\]) disappears, as $e_0^{\pm} \rightarrow 0, \infty$). In the context of the Donagi-Witten formulation, the singularities at $z=0,\infty$ have a similar role to the point $x=\infty$ in $E_{\tau}$ which, by equation (\[eq:i2\]), corresponds to a degenerate spectral set $t = \pm \infty$. Recalling the underlying Hitchin model, we observe that $x=\infty$ is the pole of the associated two dimensional Higgs field, whose residue also defines the mass of the adjoint hypermultiplet [@DW].
In summary, from the Calabi-Yau geometry we get the integrability structure of the Seiberg-Witten model, as it is described in the Donagi-Witten construction. This integrability structure only shows up when we keep alive both the string scale, $\alpha'$, and the gravitational effects due to the existence of the dilaton. However, we can not expect the Donagi-Witten model to be equivalent to full fledged string theory, an issue that will be addressed in the next section.
Picard-Fuchs Equations.
=======================
It was shown in reference [@KLMVW] that in the point particle limit, where the $K3$ degenerates to an ALE space, we can effectively map the second homology group of $K3$ into the homology group of 0-cycles defined by the spectral set $\{ t~~ |~~ P_{A_1}(t;u_i)=0
\}$. Alternatively, the integration of the holomorphic top form of the Calabi-Yau manifold on a 2-cycle of the $K3$, in its ALE limit, defines the meromorphic Seiberg-Witten form $\lambda$ for the curve (\[eq:curve\]). This form is [@MW] $$\lambda= \sqrt{u + \frac{1}{2} (z + \frac{\Lambda^4}{z} )}~
\frac{dz}{z}.
\label{40}$$ Using the Calabi-Yau curve defined in the previous sections we can propose a generalization of (\[40\]) to the case gravity is turned on. Namely, we can consider the meromorphic form $\lambda$ derived from the ${\cal C}^{CY}$ by means of the map (\[14\]). The result is $$\tilde{\lambda} = \sqrt{1-\frac{1}{\hat{c}}}~ \frac{dz}{z},
\label{41}$$ with $\hat c$ defined in (\[557\]). Our aim in this section is to understand the meaning of (\[41\]), and equivalently ${\cal C}^{CY}$, in the string context. In order to do so, we will analyze in what way $\tilde{\lambda}$ is related with the periods of $K3$ and with the associated Picard-Fuchs equation.
The information used to construct ${\cal C}^{CY}$ reduces to the discriminant of the $K3$-fiber $\Delta(z;b,c)$ and, derived from it, the discriminant of the Calabi-Yau threefold. However these data do not determine in an unique way the threefold. There exist different Calabi-Yau spaces whose moduli of Kähler deformations share common features. As an example the manifold $\IP_{\{1,1,2,2,6\}}^{4}[12]$ we have been considering up to now, exhibits the same singular loci and Yukawa couplings structure that the hypersurface $\IP_{\{1,1,2,2,2\}}^{4}[8]$ and the complete intersections $\IP_{\{1,1,2,2,2,2\}}^{5}[4,6]$ and $\IP_{\{1,1,2,2,2,2,2\}}^{6}[4,4,4]$ [@can; @CYY; @Y2]. These four manifolds are $K3$-fibrations and from any of them, in the point particle limit (\[eq:cc1\]), can be derived the exact physics of $SU(2)$ $N=2$ Yang-Mills theory. For simplicity we will denote them as $A:\IP_{\{1,1,2,2,6\}}^{4}[12]$, $B:\IP_{\{1,1,2,2,2\}}^{4}[8]$, $C:\IP_{\{1,1,2,2,2,2\}}^{5}[4,6]$ and $D: \IP_{\{1,1,2,2,2,2,2\}}^{6}[4,4,4]$ (see Appendix C for a brief description of these spaces).
The Picard-Fuchs equation for their $K3$-fiber is [@LY] $$L= \theta_{\hat c}^3 - {\hat c} \left(\theta_{\hat c}^3 +
\frac{3}{2}\theta_{\hat c}^2 + \frac{1}{2}\theta_{\hat c} +
2r ( \theta_{\hat c} + \frac{1}{2}) \right),
\label{43}$$ with $r=\frac{N-1}{2N^2}$ and $N=6,4,3,2$ respectively. Since by construction ${\cal C}^{CY}$ can not distinguish between the mentioned four spaces, it is natural to expect that (\[41\]) is related to the common part of their Picard-Fuchs equation. Indeed, it is easy to see that (\[41\]) is solution of $$L_1= \theta_{\hat c}^3 - {\hat c} \left(\theta_{\hat c}^3 +
\frac{3}{2}\theta_{\hat c}^2 + \frac{1}{2}\theta_{\hat c} \right)
- \frac{1}{4}\theta_{\hat c}.
\label{44}$$ The operators $L$ and $L_1$ differ in the first order differential operator $$L_2= L-L_1= \frac{1}{4}\theta_{\hat c} - 2r {\hat c}
\left(\theta_{\hat c} + \frac{1}{2} \right).$$
The singular points of the $K3$ Picard-Fuchs equation (\[43\]) are at $\hat{c}=0,1,\infty$. At $\hat{c}=1$ the $K3$ develops an $A_1$ singularity, and its associated monodromy reproduces the Weyl group for $SU(2)$. However the point $\hat{c}=0$ presents logarithmic monodromy, as can be deduced from the corresponding indicial equations. This fact does not allow to use the solutions of (\[43\]), in particular the period carrying Weyl monodromy around $\hat{c}=1$, for defining the Seiberg-Witten differential of a Riemann surface. Definition (\[44\]) is the necessary modification of $L$ in order to achieve this. The singular points of $L_1$ are still at $\hat{c}=0,1,\infty$, but the asymptotic behavior for both $0,\infty$ have been modified. The indicial equations of $L$ and $L_1$ at $\hat{c}=1$ coincide, therefore the leading behavior of the solutions at this singularity is not affected.
We want now to analyze how working with $L_1$ instead of $L$ affects the physics that we obtain. Namely, we will compare the physics associated to ${\cal C}^{CY}$ with that of a type IIA string compactified in manifolds $A$, $B$, $C$ or $D$, and their corresponding heterotic dual strings.
Let us begin considering the conifold locus of these Calabi-Yau manifolds. This locus corresponds to the melting of the $K3$-fiber singular points $e_{1}^{\pm}$, given in (\[3\]). At these points $\hat{c}=1$ and $L$ and $L_1$ coincide. Therefore, in a neighborhood of the conifold locus, ${\cal C}^{CY}$ (equivalently an $N=4$ softly broken gauge theory) should describe essentially the same physics as the threefold. Indeed in both cases the singularity is interpreted as due to BPS dyons becoming massless [@SW; @SW2], [@S].
The singular locus $\hat{{\cal C}}_0$ of the $N=4$ softly broken theory corresponds to an electric hypermultiplet acquiring zero mass [@SW2]. This multiplet derives from components of the initial massive adjoint hypermultiplet, and the singularity occurs for Higgs expectation values of order $m$, with $m$ the adjoint hypermultiplet bare mass. The gravity counterpart of this locus is ${\cal C}_0$ in (\[eq:c5\]), which we observed can be represented by the melting of the singular points $e_{0}^{\pm}=
e_{1}^{\pm}$. Since over $e_{0}^{\pm}$ the $K3$ develops the $\hat{c}=0$ singularity, at which the operators $L$ and $L_1$ strongly differ, the interpretation of field theory and string loci, $\hat{{\cal
C}}_0$ and ${\cal C}_0$, can naturally be different. However we will propose that both share a common origin in the presence of an additional (non-moduli) scale in the theory. For the $N=4$ softly broken theory, this scale is of course provided by the mass of the adjoint hypermultiplet. In the case of the Calabi-Yau space we want to argue that the scale behind ${\cal C}_0$ is the string tension $\alpha'$. Let us stress that the identification between $m^2$ and $(\alpha')^{-1}$ was one of the main consequences of the map (\[14\]) between gauge theory variables $(\hat{u},\tau)$ and string moduli $(b,c)$.
We will concentrate in the Calabi-Yau model $A:~
\IP_{\{1,1,2,2,6\}}^{4}[12]$. The locus ${\cal C}_0$ was defined as the set of points invariant under ${\cal A}^2$, with ${\cal A}$ given by the moduli space symmetry (\[eq:c11\]). This transformation satisfies ${\cal A}^6=-1$, discarding its origin in an additional massless particle. The mirror map relates complex structures of $A^{\ast}$, parameterized by $(b,c)$, to Kähler structures of $A$, parameterized by special coordinates $(t_1,t_2)$ ($t_1 \sim \mbox{log}~c$, $t_2 \sim
\mbox{log}~b$, for $c,b \rightarrow 0$ [@MM]). Heterotic-type II duality further relates the Kähler coordinates $(t_1,t_2)$ to the heterotic moduli [@KV; @KKLMV] $$t_1=T~~, \hspace{2cm} t_2=-\frac{S}{2 \pi i},
\label{800}$$ with $T$ an space-time modulus and $S$ the heterotic dilaton. In [@can; @Y2; @KKLMV] the monodromy group for the manifold $A$ was worked out, from where it can be deduced $${\cal A}^2 : T \rightarrow -\frac{T+1}{T}~, \hspace{1cm}
S \rightarrow S.
\label{501}$$ We see that the monodromy around ${\cal C}_0$ produces a $T$-duality transformation for the heterotic dual string[^8].
More in general, and without doing reference to string-string dualities, we can relate ${\cal C}_0$-type loci, i.e. set of fixed points under global symmetries in moduli space, to self-dual regions with respect to the Calabi-Yau generalization of the target-space duality $R\rightarrow \frac{1}{R}$, where $R$ stands for a Kähler modulus. This can can already be observed in the simpler case of the quintic [@COGP]. In that case we have a single moduli, $\psi$; the conifold singularity is located at $\psi=1$. The symmetry transformation ${\cal A}:\psi \rightarrow \beta
\psi$, with $\beta^5=1$, lets fixed the point $\psi=0$, which is the analog in this simple example of the ${\cal C}_0$ locus. At this point the Zamolodchikov’s metric is regular, while the Kähler potential becomes singular. However, this singularity in the Kähler potential is not producing any singular contribution to the partition function, as can be seen from the topological analysis of reference [@BCOV]. Thus, we should not expect new massless particles at $\psi=0$[^9]. Alternatively, the ${\cal
A}$ transformation acts on the Calabi-Yau radius in a similar (but more complicated) way to $R \rightarrow \frac {1}{R}$ [@COGP].
$T$-duality transformations always imply the introduction of an additional scale. Having identified ${\cal C}_0$ as the Calabi-Yau generalization of a self-dual locus, we should determine the scale that is associated to it. Expression (\[501\]), together with the fact that the Calabi-Yau weak coupling limit $b \rightarrow 0$ gets mapped by string-string duality to heterotic perturbative effects [@KV], indicate that this scale is given by the heterotic-dual string tension $\alpha'$. Restoring unities, the self-dual point of (\[501\]) is $T=\rho~ \sqrt{\alpha'}$, with $\rho^3
=1$. Therefore perturbative string excitations will have, at that point, a typical mass ${\alpha'}^{-1/2}$. This agrees with the field theory interpretation ${\alpha'}^{-1/2}\sim m$, for $m$ the adjoint hypermultiplet bare mass. Values $b>0$ will imply corrections of the characteristic mass of string excitations at ${\cal C}_0$. In this case, the natural candidate for comparing square masses should be the Casimir expectation value $u\sim m^2 e_1(\tau)$, that determines the field theory locus $\hat{{\cal C}}_0$.
From the type II perspective, we should obtain the same result for the typical mass of excitations in a neighborhood of ${\cal C}_0$, by looking at BPS states associated to Ramond-Ramond charged branes. Notice that, although ${\cal C}_0$ is equally a self-dual locus on the type IIA side, the mass of these excitations will not be governed by the type II string tension, since Ramond-Ramond charged branes are non-perturbative objects [@P].
The same considerations apply to models $B$ and $C$. The case $D$ is however different, as the monodromy around ${\cal C}_0$ involves logarithms[^10] and therefore it does not admit to be interpreted as a self-dual locus. In this case we could think in the appearance of a massless particle to explain the singularity. If underlying scale is still $\alpha'$, it would indicate an electrically charged particle. Physically, the $D$ model can be the one more closely related to ${\cal C}^{CY}$. It is worth noticing that the singular points of the first order differential operator $L_2$ are at $\hat{c}= 0, 8r,\infty$. Only for the model $D$ we have $8r=1$, and the singular points of $L_2$ coincide with those of $L$ and $L_1$. This fact could be directly connected with passing from a logarithmic monodromy transformation around the hypermultiplet locus $\hat{{\cal C}}_0$ in the field theory approach, to an smoothed out (non-logarithmic) monodromy ${\cal A}^2$ for the models $A$, $B$ and $C$, while not for the Calabi-Yau $D$.
Finally we consider the locus ${\cal D}_{(0,-1)}= \{ c=0 \}$, which correspond to the degenerate situation $e_{1}^+=0$ and $e_{1}^-=\infty$. The operators $L$ and $L_1$ also differ at $z=0,\infty$, implying that close to this locus ${\cal C}^{CY}$ will not reproduce the string dynamics. This is indeed as expected, since the mirror map [@can] fixes $t_1=i \infty$ at ${\cal D}_{(0,-1)}$, therefore representing a decompactification limit. Using (\[14\]), this locus corresponds to the ultraviolet regime $u=\infty$ for the $N=4$ softly broken theory, where string and field theory should strongly differ.
Summarizing, we have seen that ${\cal C}_{CY}$ can provide a good description of string phenomena only in a neighborhood of the conifold locus. However, ${\cal C}_{CY}$ proved useful in interpreting the locus ${\cal C}_0$. In the next section, following this path, we will use ${\cal C}_{CY}$ as a tool for understanding further differences in the coupling to gravity of $SU(2)$ $N=2$ Yang-Mills provided by models $A$, $B$, $C$ and $D$.
Higher Kac-Moody Level String Models.
=====================================
The differences between the four Calabi-Yau manifolds we are considering can be resumed in the properties of the mirror map for their $K3$-fiber, or equivalently, the mirror map between $c$ and $t_1$ for the dilaton modulus value $b=0$. Using the string-string identification $t_1=T$, we have $$c=\frac{h_k(T_{0}^{(k)})}{h_k(T)},
\label{801}$$ with $h_k$ the Hauptmodul function of $\Gamma_0 (k)_+$ (shifted by a constant), and $k=1,2,3,4$ for spaces $A$, $B$, $C$ and $D$ respectively [@can; @LY]. The value $T_{0}^{(k)}$ is given by the self-dual point of one distinguished element of the group $\Gamma_0
(k)_+$, the Atkin-Lehner involution $$T \rightarrow -\frac{1}{kT}.$$ From (\[801\]) we see that the $\Gamma_0 (k)_+$ determines the $T$-duality group of the heterotic dual string. The dual for the $A$ model is a rank three heterotic string compactified in $K3 \times
T^2$ [@KV], where $T$ corresponds to a $T^2$ modulus. Since $\Gamma_0 (1)_+=PSL(2,Z)$ and $h_1$ coincides with the $j$-invariant, we obtain the usual $T$-duality group for this model.
In [@AP] the case $B:\IP^{4}_{\{1,1,2,2,2\} }[8]$ was considered. It was proposed to be dual to an heterotic string compactification with $SU(2)$ perturbative enhancement of symmetry realized at Kac-Moody level 2. The mirror map (\[801\]) fixes the heterotic enhancement of symmetry point at $T=i/\sqrt{2}$. The right and left moving momenta for the rank three models are given [@N] by $$\begin{aligned}
\label{50}
p_L & = & \frac{i \sqrt{2}}{T-\bar{T}}(n_1 + n_2 {\bar T}^2 + 2m
{\bar T}), \nonumber \\
p_R & = & \frac{i \sqrt{2}}{T-\bar{T}}(n_1 + n_2 T \bar{T} +
m (T+\bar{T}) ).\end{aligned}$$ The condition for new massless states, $p_L=0$, $|p_{R}|^2 \leq 2$, is satisfied at $T=i/\sqrt{2}$ for $n_1=2 \: n_2=\pm 1/2$, $m=0$. This implies the value $p_{R}^2 =1$. If we denote the Kac-Moody level at which the enhanced gauge group is realized by $k_G$, the following relation holds $$\label{900}
k_G~ |p_{R}|^2 = 2 .$$ Therefore $k_G=2$ for the heterotic dual associated to the manifold $B$. In an analogous way we can analyze the heterotic duals that could correspond to models $C$, $D$. The mirror map fixes the perturbative enhancement of symmetry at $T=i/\sqrt{k}$, with $k=3,4$ respectively. At this value, using (\[50\]), the conditions for two additional massless states are again verified, satisfying (\[900\]) for $k_G=k$. Although further checks along the lines of [@KT; @K3; @AP] have to be done for the cases $C$, $D$, and the explicit construction of the heterotic dual [@KV] accomplished for models $B$, $C$ and $D$, we will assume that a main difference between the coupling to gravity of $N=2$ $SU(2)$ Yang-Mills that the four manifolds can provide is given by the Kac-Moody level at which they realize the gauge symmetry, being $k_G=1,2,3,4$ respectively.
The Kac-Moody level also affects the relation between gauge group bare coupling constant and heterotic dilaton [@G]. With a convenient normalization we can set $\tau_G=- \frac{k S^{inv}}{2 \pi i}$, where now it is important to distinguish between the special coordinate dilaton $S$, with modular properties under $T$-duality transformations, and the invariant dilaton $S^{inv}$ [@Sinv; @Sinv2]. Since ${\cal C}^{CY}$ captures the Calabi-Yau information in a neighborhood of the conifold locus, and in the limit $b\rightarrow 0$ this information reproduces the heterotic enhancement of symmetry, we can, in this limit, identify $\tau$ and $\tau_G$. Using (\[15\]), the relation between $\tau$ and $S^{inv}$ translates into $$t_2 = - \frac{kS}{2 \pi i}
\label{51}$$ for the special geometry coordinates.
We have analyzed properties of the models $A$, $B$, $C$ and $D$ which correspond to the heterotic-dual weak coupling regime. We can now use both pieces of information, namely ${\cal C}^{CY}$ and the level $k$, in trying to reproduce properties associated to the heterotic strong coupling limit. We will concentrate therefore on the locus ${\cal C}_1=\{ b=1 \}$, which by the mirror map implies $t_2=0$ [@KM]. On the type IIA side this locus is associated to the appearance of a curve of [**Z**]{}$_2$ quotient singularities, ${\cal C}_{Z_2}$, corresponding to blowing down the exceptional divisor whose size controls the dilaton modulus $b$. In [@KMP; @KM] this singularity was interpreted as a non-perturbative enhancement of symmetry in the $U(1)$ vector field associated to the dilaton, $U(1) \rightarrow SU(2)$, together with the appearance of massless hypermultiplets in the adjoint representation of $SU(2)$. The number of massless adjoint hypermultiplets is given by the genus $g$ of the curve of singularities ${\cal C}_{Z_2}$. The monodromy around ${\cal C}_1$ can be resumed in the $2\times 2$ matrix $$\left(
\begin{array}{cc}
-1 & 2(g-1) \\
0 & -1
\end{array}
\right),
\label{52}$$ acting on a column vector whose first entry we denote by $t_{2}^D$, and the second is given by the special coordinate $t_2$.
Since $t_2$ behaves as the special coordinate version of the $N=4$ coupling constant $\tau$, and $\tau$ is an intrinsic parameter of ${\cal C}^{CY}$, it should be possible to describe the different monodromy matrices (\[52\]) for the spaces $A$, $B$, $C$ and $D$ in an unified way. In order to do this, we will allow the freedom to change the normalization of $t_{2}^D$, while preserving that of $t_2$. Inspired by (\[51\]) we consider the change $t_{2}^D \rightarrow t_{2}^D/k$, transforming (\[52\]) into $$\left(
\begin{array}{cc}
-1 & 2(g-1)/k \\
0 & -1
\end{array}
\right).
\label{eq:pf8}$$ For the four models we are considering the following relation between the genus $g$ of the curve of singularities at ${\cal C}_1$, and the level $k$ dictated by the modular properties of the mirror map, is verified (see Appendix C) $$\frac{g-1}{k}= 1.$$ Substituting this into (\[eq:pf8\]) we obtain a single representation of the ${\cal C}_1$ monodromy for the four Calabi-Yau spaces. It is given by $P T^{-2} $, where now $T$ denote the $Sl(2;Z)$ generator and $P=-1$.
It must be noticed that a $T^{-2}$ monodromy is the one we will get for the Donagi-Witten curve at the singular locus $\tau=0$[^11], which by the map between field theory and string variables (\[14\]) corresponds to ${\cal C}_1$. Namely, encoding the monodromy around $\tau=0$ through its action on a column vector with entries $1$ and $\tau$, the matrix $T^{-2}$ implies $\tau^D
\rightarrow
\tau^D + 2$, with $\tau^D =-1/\tau$. We observe that, although the ${\cal C}^{CY}$ curve can not describe the physics at ${\cal C}_1$ because is missing the parameter $g$ and the Weyl generator $P$, it can however explain its underlying structure. Indeed, using ${\cal C}^{CY}$ and the Kac-Moody level information, it is possible to reproduce in a natural way (\[52\]). We can also put in correspondence the monodromies for the Calabi-Yau and field theory loci ${\cal C}_{\infty}$ and $\tau=
i\infty$. Taking into account the double covering of the string moduli space that the field theory is doing, the monodromy $S T^{-2} S^{-1}$ around $\tau =
i \infty$, producing $\tau \rightarrow \tau +2$, becomes the ${\cal C}_{\infty}$ monodromy [@can; @Y2] $t_2 \rightarrow t_2 +
1$.
Final Comments.
===============
Ultraviolet finite $N=2$ supersymmetric gauge theories are described [@SW2] in terms of the Higgs vacuum expectation values, the bare coupling constant $\tau$ and the masses of their matter content. In order to relate these theories with a string compactification on a Calabi-Yau threefold, we can not follow the standard path of performing the heterotic point particle limit. In this paper we have shown that $N=2$ gauge theories with $N=4$ matter content are associated with the global structure of $K3$-fibration threefolds. Moreover the global geometrical information on $K3$-fibrations reproduces all the ingredients used in the characterization [@DW] of the integrability of these ultraviolet finite $N=2$ theories. For the case of $SU(2)$, our philosophy has been to consider the gauge theory with $N=4$ matter as encoding the common geometrical structure of $K3$-fibrations sharing the same $h_{2,1}$ Hodge number, and structure of the complex deformations moduli space.
The special role of a ultraviolet finite $N=2$ $SU(2)$ gauge theory, in particular the case $2N_c = N_f$, appears in a different context [@Sen] as proves [@BDS] of F-theory [@V] backgrounds. The prove approach is related to duality properties of type II strings, while our approach should be related to strong-weak coupling duality of the heterotic string.
[**Acknowledgments**]{}
We thank conversations with K. Landsteiner, W. Lerche, A. Sen, C. Vafa and N. Warner. This work is partially supported under grant OFES contract number 930083 by European Community grant ERBCHRXCT920069 and by PB92-1092. The work of E. L. is supported by C.A.P.V. fellowship. The work of R. H. is supported by U.A.M. fellowship.
[**Appendices.**]{}
Isogeny Between Cubic and Quartic Curves.
=========================================
We consider an elliptic curve, given by the quartic $$y^2= (x^2 + a)^2 - b^2,
\label{a1}$$ for certain constants $a$ and $b$. By defining $x'=x^2 + a$, $y'=yx$ we can convert (\[a1\]) into an isogenic cubic $${y'}^2=(x'-a)({x'}^2 - b^2).
\label{a2}$$ This process amounts to quotienting by the symmetry $x \rightarrow -x$ of (\[a1\]), and blowing down the line $(x=0,y)$ in order to get again an elliptic curve.
It is immediate to see that the abelian differential for quartic and cubic are related by $$\frac{dx'}{y'}=2\frac{dx}{y}.$$ Let us choose a basis of cycles of (\[a2\]) in the following form: $\gamma_1'$ surrounding the branch points $\pm b$, and $\gamma_2'$ surrounding the branch points $b$ and $-a$. With this choice the cycles $\gamma_1$ and $\gamma_2$, transformed of $\gamma_1'$ and $\gamma_2'$ respectively, satisfy $$\gamma_1=\gamma_1'~~, \hspace{2cm} \gamma_2=2\gamma_2'.$$ Therefore the difference between quartic and cubic, from the point of view of the periods, reduces to a change in normalization $$\int_{\gamma_1} \omega = \frac{1}{2} \int_{\gamma_1'} \omega'~~,
\hspace{1cm} \int_{\gamma_2} \omega= \int_{\gamma_2'} \omega',
\label{A1}$$ for $\omega$, $\omega'$ the respective abelian differentials. This is precisely the difference between the two curves considered in [@SW], [@SW2] for solving $SU(2)$ $N=2$ Yang-Mills.
The quartic form (\[a1\]) reproduces the structure, for the particular case $SU(2)$, of the hyperelliptic curves for higher rank gauge groups proposed in [@KLTY; @AF] $$y^2=P_{SU(N)}(x;u_i)^2 - \Lambda^{2N},$$ with $u_i$ the $SU(N)$ Casimirs.
We should also note that the Jacobi invariants for (\[a1\]) and (\[a2\]) differ. This can be seen by transforming both curves to the standard form $$y^2=x(x-1)(x-\lambda),$$ and comparing the $\lambda$-parameters obtained. We have $$\lambda=\frac{a+\sqrt{a^2 - b^2}}{a -\sqrt{a^2 - b^2}}~~,
\hspace{1cm} \lambda'=\frac{a+b}{a-b}.$$
The Map Between $SU(3)$ Moduli Spaces.
======================================
In [@KV] the Calabi-Yau $\IP_{\{1,1,2,8,12\}}^{4}[24]$ was proposed as the type II dual for the string embedding of $N=2$ $SU(3)$ Yang-Mills theory. In this Appendix, extending part of the analysis done in previous sections for $SU(2)$, we will build a map between the moduli space of $N\!=\!2$ $SU(3)$ Yang-Mills theory with adjoint matter and the moduli space of the hypersurface $\IP_{\{1,1,2,8,12\}}^{4}[24]$. The analysis will be restricted to the weak coupling sections $\tau =i \infty$ and $b=0$ of the respective moduli spaces.
The discriminant locus of the genus two curve describing $SU(3)$ with adjoint matter [@DW], in the limit $\tau=i \infty$, is given by $$\Delta_{SU(3)} = (4\hat{u}^3-27\hat{v}^2)(4(\hat{u}-1)(\hat{u}-4)^2-
27\hat{v}^2),
\label{eq:b1}$$ where $\hat{u}$ and $\hat{v}$ are the quadratic and cubic Casimir expectation values respectively, normalized by the adjoint hypermultiplet mass. The first factor corresponds to a classical enhancement of $SU(2)$ gauge symmetry. We will denote its two branches by $${\cal C}^{\pm} = \{3\sqrt{3} \hat{v} = \pm 2 \hat{u}^{3/2} \}.
\label{eq:b2}$$ The second factor goes to zero when some component of the $N\!=\!4$ hypermultiplet become massless; we will also divide it into $${\cal C}_{h}^{\pm} = \{3\sqrt{3} \hat{v} = \pm
2(\hat{u}-4)\sqrt{\hat{u}-1} \}.
\label{eq:b3}$$
If we now concentrate on the $K3$-fibration structure of the $\IP_{\{1,1,2,8,12\}}^{4}[24]$ Calabi-Yau manifold [@KLMVW] $$\begin{aligned}
W & = & \frac{1}{24} (z + \frac {b}{z} + 2) +
\frac{1}{12} x_{3}^{12} + \frac{1}{3} x_{4}^3 + \frac{1}{2} x_{5}^2 +
\nonumber \\
&& + \frac {1}{6\sqrt{c}}(x_0 x_3)^6 + \left( \frac {a}{\sqrt{c}}
\right)^{1/6} x_0 x_3 x_4 x_5 = 0,\end{aligned}$$ we notice that the $K3$-fiber becomes singular at six points $(z=e_i^{\pm},
\: \: i=0,1,2)$ of the $\IP^1$ base space $$\begin{aligned}
e_0^{\pm} & = & -1 \pm \sqrt{1-b}, \nonumber\\
e_1^{\pm} & = & \frac {1-c \pm \sqrt{(1-c)^2-bc^2}}{c}, \nonumber\\
e_2^{\pm} & = & \frac {(1-a)^2-c \pm \sqrt{((1-a)^2-c)^2-bc^2}}{c}.
\label{eq:b5}\end{aligned}$$ Merging of these points leads to two conifold-type loci, when $e_i^{+}=
e_i^{-}$ for $i=1,2$, and a strong dilaton locus ${\cal C}_1$, when $e_0^{+}=e_0^{-}$: $$\begin{aligned}
{\cal C}_C^{(1)} & = & \{ (1-c)^2-bc^2 = 0 \}, \nonumber\\
{\cal C}_C^{(2)} & = & \{ ((1-a)^2-c)^2-bc^2 = 0 \}, \nonumber\\
{\cal C}_1 & = & \{ b-1 = 0 \}.
\label{eq:b6}\end{aligned}$$ The weak gravity locus $${\cal C}_{\infty} = \{ b = 0 \},$$ corresponds to $0=e_0^+=e_1^+=e_2^+$. We can also consider the merging $e_0^{\pm}=e_i^{\pm}, \: \: i=1,2$, which happens respectively at $$\begin{aligned}
{\cal C}_0^{+} & = & \{ c = \infty ~~~ (\forall~ a/\sqrt{c} ) \},
\nonumber \\
{\cal C}_0^{-} & = & \{ a -1 = 0 \},
\label{eq:b7}\end{aligned}$$
We will restrict now to the $(a,c)$ section of the Calabi-Yau moduli space at $b=0$. In the spirit of (\[14\]), we can propose the following map between the $(u,v)$-plane at $\tau= i \infty$, and the Calabi-Yau moduli space at $b=0$ $$1-c = \frac {-2\hat{u}^{3/2}+3\sqrt{3}\hat{v}}{2(\hat{u}-4)
\sqrt{\hat{u}-1}+3\sqrt{3}\hat{v}}, \: \: \: \:
1- \frac {c}{(1-a)^2} = \frac {2\hat{u}^{3/2}+3\sqrt{3}
\hat{v}}{2(\hat{u}-4)\sqrt{\hat{u}-1}-3\sqrt{3}\hat{v}}.
\label{eq:b8}$$ This map is built by requiring that the loci ${\cal C}^{\pm}$ get mapped into the two conifold ${\cal C}_{C}^{(1,2)}$, and the hypermultiplet loci ${\cal C}_{h}^{\pm}$ go into ${\cal C}_0^{\pm}$.
The point $(a=0,c=1)$, one of the intersections between ${\cal C}_C^{(1)}$ and ${\cal C}_C^{(2)}$ at $b=0$, corresponds by heterotic-type II duality to a perturbative $SU(3)$ enhancement of symmetry point [@KV]. A first check of (\[eq:b8\]) is that, sending the mass of the adjoint hypermultiplet $m \rightarrow \infty$ and identifying again $m^2$ with the string tension $(\alpha')^{-1}$, it reproduces the point particle limit map for $SU(3)$ [@KKLMV; @KLMVW] $$\begin{aligned}
a & = & -2 (\alpha u)^{3/2}, \nonumber\\
c & = & 1 - \alpha^{3/2}(-2 u^{3/2} +3\sqrt{3} v).\end{aligned}$$ The scale governing the strong coupling effects of the $N=2$ $SU(3)$ Yang-Mills theory with adjoint matter is given by $\Lambda^{6} \sim e^{2 \pi i \tau} m^{6}$ [@DW]. Assuming the relation (\[15\]) between the dilaton modulus $b$ and the $N=4$ coupling constant $\tau$, this implies the double scaling identification required to perform the point particle limit [@KKLMV] $$b = \alpha^{3} \Lambda^{6}.$$
It would be interesting to see if the map (\[eq:b8\]) can help to locate, by comparison with the field theory, Argyres-Douglas points [@AD] in string theory.
Summary of $A$, $B$, $C$ and $D$ models.
========================================
The Picard-Fuchs equations that govern the Kähler structure deformations of the Calabi-Yau manifolds $A:\IP_{\{1,1,2,2,6\}}^{4}[12]$, $B:\IP_{\{1,1,2,2,2\}}^{4}[8]$, $C:\IP_{\{1,1,2,2,2,2\}}^{5}[4,6]$ and $D: \IP_{\{1,1,2,2,2,2,2\}}^{6}[4,4,4]$, consist of two differential operators of second and third order [@CYY; @Y2]. The second order operator is common to the four manifolds $$L^{(1)}= 4\theta_b^2 - b (2\theta_b - \theta_c + 1)
(2\theta_b - \theta_c),
\label{c1}$$ while they differ in the third order one $$\begin{aligned}
\label{C1}
A: && L^{(2)}= \theta^2_c \: (\theta_c - 2\theta_b) - c(\theta_c
+ 5/6)(\theta_c + 1/2)(\theta_c + 1/6), \nonumber \\
B: && L^{(2)}= \theta^2_c \: (\theta_c - 2\theta_b) - c(\theta_c
+ 3/4)(\theta_c + 1/2)(\theta_c + 1/4), \nonumber \\
C: && L^{(2)}= \theta^2_c \: (\theta_c - 2\theta_b) - c(\theta_c
+ 2/3)(\theta_c + 1/2)(\theta_c + 1/3), \\
D: && L^{(2)}= \theta^2_c \: (\theta_c - 2\theta_b) - c(\theta_c
+ 1/2)^3. \nonumber\end{aligned}$$ We are using the normalization conventions for the moduli parameters $(b,c)$ implied in (\[1000\]) and (\[2\]). The second order operator (\[c1\]), in the point particle limit defined by (\[eq:cc1\]), becomes precisely the Picard-Fuchs operator for $N=2$ $SU(2)$ Yang-Mills theory [@KKLMV].
These four spaces develop a curve of [**Z**]{}$_2$ quotient singularities, ${\cal C}_{Z_2}$, at ${\cal C}_1=\{ b=1 \}$. This singularity comes from an ambient singularity in the weighted projective spaces in which they are immersed. Namely, in our case points $(x_1,x_2,x_3,x_4,..)$ and $(\lambda x_1, \lambda x_2, \lambda^2 x_3,
\lambda^2 x_4,..)$ should be identified, fixing, for $\lambda=-1$, the hyperplane of orbifold points $x_1=x_2=0$. The intersection of this hyperplane with the defining hypersurface, or complete intersection, of the Calabi-Yau manifolds determines ${\cal C}_{Z_2}$ in each case. The genus of ${\cal C}_{Z_2}$ is $g=2,3,4,5$ for the spaces $A$, $B$, $C$ and $D$ respectively [@can; @KMP].
The exceptional divisor whose size controls the dilaton modulus $b$, is introduced in order to resolve the mentioned singularity. However for the value $b=1$, the manifolds develop the original curve of [**Z**]{}$_2$ quotient singularities. This structure can be read from the following piece in the defining expression of their mirror manifolds $$W^{\ast}= \frac{1}{2n} x_1^{2n} + \frac{1}{2n} x_2^{2n} +
\frac{1}{\sqrt{b}~n} (x_1 x_2)^n +...~.$$ Changing variables to $x_1/x_2= z^{1/n} b^{-1/2n}$ and $x_1^2= x_0
z^{1/n}$ [@KLMVW], the $K3$-fibration structure (\[999\]) for manifolds $A$, $B$, $C$ and $D$ becomes manifest. The $K3$-fiber of their mirror, which for model $A$ is given in (\[1001\]), is given by $$\begin{aligned}
B: && x_{0}^{4}/4 + x_{3}^4/4 +
x_{4}^4/4 + x_{5}^4/4
+ \hat{c}^{-1/4} x_0 x_3 x_4 x_5 =0, \frac{}{} \nonumber \\
C: && x_{0}^2/2 + x_{3}^2/2
+ x_{4}^2/2 + \hat{c}^{-1/5} x_5 x_6 =0 \nonumber \\
&& x_{5}^3/3 + x_{6}^3/3 +\hat{c}^{-1/5} x_0 x_3
x_4 =0 \frac{}{}, \nonumber \\
D: && x_{0}^2/2 + x_{3}^2/2
+ \hat{c}^{-1/6} x_4 x_5 =0 \nonumber \\
&& x_{4}^2/2 + x_{5}^2/2 + \hat{c}^{-1/6}
x_6 x_7=0 \nonumber \\
&& x_{6}^2/2 + x_{7}^2/2 + \hat{c}^{-1/6} x_0 x_3=0,\end{aligned}$$ with $\hat{c}(z;b,c)$ defined in (\[557\]). The Picard-Fuchs equation governing the above $K3$ surfaces can be obtained from the third order differential operator of the threefold (\[C1\]), by sending $b \rightarrow 0$.
------- ---------------------------------------- ------------- ----------- -----------------
Model Manifold $h_{2,1}$ Genus $g$ Mirror
modular group
$A$ $\IP_{\{1,1,2,2,6\}}^{4}[12]$ $128$ $2$ $\Gamma$
$B$ $\IP_{\{1,1,2,2,2\}}^{4}[8]$ $86$ $3$ $\Gamma_0(2)_+$
$C$ $\IP_{\{1,1,2,2,2,2\}}^{5}[4,6]$ $68$ $4$ $\Gamma_0(3)_+$
$D$ $\IP_{\{1,1,2,2,2,2,2\}}^{6}[4,4,4]$ $58$ $5$ $\Gamma_0(4)_+$
------- ---------------------------------------- ------------- ----------- -----------------
The four manifolds we are considering have common Yukawa couplings [@CYY; @Y2] $$K_{ccc}= \frac{1}{c^3 \Delta_C},~ K_{ccb}=\frac{1-c}{2 c^2 b
\Delta_C},~
K_{cbb}=\frac{2c-1}{4cb\Delta_C \Delta_1},~ K_{bbb}= \frac{1-c+b-
3cb}{8b^2 \Delta_C \Delta_1^2},$$ where $\Delta_C=(1-c)^2 - c^2 ~ b$ is the conifold factor in the discriminant, and $\Delta_1=1-b$. To finish, let us notice that the Yukawa couplings of Calabi-Yau three-folds are determined by the Picard-Fuchs equations coefficients $f_{l}^{k_1,..,k_n}$ $$L^{(l)} = \sum_{k_1,..,k_n} f_{l}^{k_1,..,k_n} \partial_{c_1}^{k_1}..
\partial_{c_n}^{k_n}~~,$$ with $\sum k_i \geq 2$ [@CYY] ($c_i$, $i=1,..,n$ includes all Calabi-Yau moduli parameters and $l$ labels the set of Picard-Fuchs differential operators). This information is retained in the modified Picard-Fuchs operator (\[44\]) that governs the Seiberg-Witten differential for ${\cal C}^{CY}$.
[99]{}
N. Seiberg and E. Witten, Nucl. Phys. B426 (1994) 19, hep-th/9407087.
N. Seiberg and E. Witten, Nucl. Phys. B431 (1994) 484, hep-th/9408099.
A. Klemm, W. Lerche, S. Theisen and S. Yankielowicz, Phys. Lett. B344 (1995) 169, hep-th/9411048.
P. C. Argyres and A. E. Faraggi, Phys. Rev. Lett. 74 (1995) 3931, hep-th/9411057.
U. H. Danielsson and B. Sundborg, Phys. Lett. B358 (1995) 273, hep-th/9504102.
A. Hanany and Y. Oz, Nucl. Phys. B452 (1995) 283, hep-th/9505075.
P. Argyres, R. Plesser and A. Shapere, Phys. Rev. Lett. 75 (1995) 1699, hep-th/9505100.
A. Brandhuber and K. Landsteiner, Phys. Lett. B358 (1995) 73, hep-th/9507008.
J. A. Minahan and D. Nemeschansky, hep-th/9507032.
P. Argyres and A. Shapere, hep-th/9509175.
A. Hanany, hep-th/9509176.
A. Gorskii, I. Krichever, A. Marshakov, A. Mironov and A. Morozov, Phys. Lett. B355 (1995) 466, hep-th/9505035.
E. Martinec and N. P. Warner, Nucl. Phys. B459 (1996) 97, hep-th/9509161.
T. Nakatsu and K. Takasaki, Mod. Phys. Lett. A11 (1996) 157, hep-th/9509162.
R. Donagi and E. Witten, Nucl. Phys. B460 (1996) 299, hep-th/9510101.
E. Martinec, Phys. Lett. B367 (1996) 91, hep-th/9510204.
E. Martinec and N. P. Warner, hep-th/9511052.
H. Itoyama and A. Morozov, hep-th/9511126 and hep-th/9512161.
C. Ahn and S. Nam, hep-th/9603028.
A. Gorsky, A. Marshakov, A. Mironov and A. Morozov, hep-th/9603140.
C. Gómez, R. Hernández and E. López, hep-th/9604057.
I. M. Krichever and D. H. Phong, hep-th/9604199.
S. Kachru and C. Vafa, Nucl. Phys. B450 (1995) 69, hep-th/9505105.
S. Ferrara, J. A. Harvey, A. Strominger and C. Vafa, hep-th/9505162.
S. Kachru, A. Klemm, W. Lerche, P. Mayr and C. Vafa, Nucl. Phys. B459 (1996) 537, hep-th/9508155.
A. Klemm, W. Lerche, P. Mayr, C. Vafa and N. Warner, hep-th/9604034.
R. Donagi and E. Markman, alg-geom/9408004.
N. Hitchin, Duke Math. J. 54 (1987) 91.
C. Hull and P. Townsend, Nucl. Phys. B438 (1995) 109, hep-yh/9410167.
E. Witten, Nucl. Phys. B443 (1995) 85, hep-th/9503124.
A. Klemm, W. Lerche and P. Mayr, Phys. Lett. B357 (1995) 313, hep-th/9506112.
P. S. Aspinwall and J. Louis, Phys. Lett. B369 (1996) 233, hep-th/9510234.
C. Gómez, R. Hernández and E. López, hep-th/9512017.
I. Antoniadis and H. Partouche, Nucl. Phys. B460 (1996) 470, hep-th/9509009.
A. Klemm and P. Mayr, hep-th/9601014.
S. Katz, D. R. Morrison and M. R. Plesser, hep-th/9601108.
B. R. Greene and M. R. Plesser, Nucl. Phys. B338 (1990) 15.
P. Candelas, X. de la Ossa, A. Font, S. Katz and D. Morrison, Nucl. Phys. B416 (1994) 481, hep-th/9308083.
S. Hosono, A. Klemm, S. Theisen and S. T. Yau, Comm. Math. Phys. 167 (1995) 301, hep-th/9308122.
S. Hosono, A. Klemm, S. Theisen and S. T. Yau, Nucl. Phys. B433 (1995) 501-554, hep-th/940605.
B. H. Lian and S. T. Yau, hep-th/9411234.
A. Strominger, Nucl. Phys. B451 (1995) 96, hep-th/9504090.
P. S. Aspinwall, B. R. Greene and D. R. Morrison, Nucl. Phys. B420 (1994) 184-242, hep-th/9311042.
P. Candelas, X. de la Ossa, P. Green and L. Parkes, Nucl. Phys. B359 (1991), 21.
M. Bershadsky, S. Cecotti, H. Ooguri and C. Vafa, Commun. Math. Phys. 165 (1994) 311, hep-th/9309140.
J. Polchinski, Phys. Rev. Lett. 75 (1995) 4724, hep-th/9510017.
I. Antoniadis, E. Gava, K. S. Narain and T. R. Taylor, hep-th/9507115.
V. Kaplunovsky, J. Louis and S. Theisen, Phys. Lett. B357 (1995) 71-75, hep-th/9506110.
P. Ginsparg, Phys. Lett. B197 (1987) 139.
B. de Wit, V. Kaplunovsky, J. Louis and D. Luest, Nucl. Phys. B451 (1995) 53-95, hep-th/9504006.
I. Antoniadis, S. Ferrara, E. Gava, K. S. Narain and T. R. Taylor, Nucl. Phys. B447 (1995) 35-61, hep-th/9504034.
A. Sen, hep-th/9605150.
T. Banks, M. R. Douglas and N. Seiberg, hep-th/9605199.
C. Vafa, hep-th/9602022.
P. C. Argyres and M. R. Douglas, Nucl. Phys. B448 (1995) 93, hep-th/9505062.
[^1]: Defining the ${\cal A}$-invariant quantities $\xi \equiv \psi^8$, $\eta \equiv \psi^4 \phi$ and $\zeta
\equiv \phi^2$, the quotiented moduli space is given by the projective cone $\xi \zeta = \eta^2$.
[^2]: This will explain why the analysis of singular loci provided by the fibration structure is missing the ${\cal D}_{(-1,-1)}$ divisor, which, together with ${\cal
D}_{(0,-1)}$, is associated with the resolution of a tangency point between ${\cal C}_{\infty}$ and ${\cal C}_1$ originated in the ${\cal A}$ quotient. This completes the set of toric divisors in the compactification of the moduli space worked out in [@can].
[^3]: The Weierstrass invariants $e_{i}$ can be defined in the terms of Jacobi theta functions: $e_{1}=\frac
{1}{3}(\theta_{2}^{4}(0,\tau)+\theta_{3}^{4}(0,\tau)),~~
e_{2}=- \frac
{1}{3}(\theta_{1}^{4}(0,\tau)+\theta_{3}^{4}(0,\tau)),~~
e_{3}=\frac {1}{3}(\theta_{1}^{4}(0,\tau)-\theta_{2}^{4}(0,\tau))$.
[^4]: We have $a_1(\tau+1)=a_1(\tau)$ and $a_2(\tau+1)=
-a_2(\tau)$.
[^5]: A similar map between the singular loci of the hyperelliptic curve describing $SU(3)$ $N\!=\!2$ supersymmetric Yang-Mills theory with adjoint matter and the moduli space of the $SU(3)$ gauge group Calabi-Yau manifold [@KV], $\IP_{\{1,1,2,8,12\}}^{4}[24]$, is proposed in Appendix B.
[^6]: Defining the invariants with respect to (\[eq:c11\]) $\xi \equiv \tilde{u}^2$, $\eta \equiv \frac {\tilde{u}}{\epsilon}$, $\zeta \equiv
\frac {1}{\epsilon^2}$, where $\epsilon \equiv 8 e^{\pi i \tau}$, $\tilde{u} \equiv \frac {\hat{u}} {\epsilon}$, the quotiented $(\hat{u},\tau)$-moduli space is also given by the projective cone $\xi \zeta = \eta^2$.
[^7]: The Landau-Ginzburg potential $P_{A_1}(t,u)=t^2+u$ is being used.
[^8]: The monodromy around the locus ${\cal D}_{(0,-1)}= \{ c=0 \}$ corresponds to $T
\rightarrow
T+1$. Combining this and (\[501\]), or equivalently surrounding the two conifold branches, we can obtain the standard $T \rightarrow
-1/T$.
[^9]: It must be recalled that the Kähler potential enters always into the special geometry equations in the form $e^{-K} \| f \|^2$, with $f$ a meromorphic section of the special geometry Hodge line bundle. A zero of $f$ at the origin ($\psi=0$ for the quintic) regularizes the singularity of $K$. The meromorphic section $f$ contributes to the definition of the topological propagator from which the topological version of the partition function is built up. At the conifold locus, this is the propagator reproducing the logarithmic singularity at one loop. At the ${\cal C}_0$ locus, this propagator is smoothed out with the help of special geometry structure, namely, the freedom to normalize the vacuum section of the Hodge line bundle.
[^10]: The indicial equations derived from the third order Picard-Fuchs operator (\[C1\]) of model $D$, around $c=\infty$, posses a triple solution.
[^11]: The dependence on $\tau$ of the Donagi-Witten construction is built using the Weierstrass invariants $e_{i}(\tau)$ (see footnote 3), which are modular functions of $\Gamma(2)$. Therefore the monodromy matrices for the singular loci $\tau=0,1,i
\infty$ in (\[12\]), have to be conjugate to $T^2$ or $T^{-2}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a catalogue of H–band spectra for 85 stars of approximately solar abundance observed at a resolving power of 3000 with the KPNO Mayall 4m FTS. The atlas covers spectral types O7–M5 and luminosity classes I-V as defined on the MK system. We identify both atomic and molecular indices and line–ratios which are temperature and luminosity sensitive allowing spectral classification to be carried out in the H–band. The line ratios permit spectral classification in the presence of continuum excess emission, which is commonly found in pre–main sequence and evolved stars. We demonstrate that with spectra of $R = 1000$ obtained at $SNR > 50$ it is possible to derive spectral types within $\pm 2$ subclasses for late–type stars. These data are available electronically through the Astronomical Data Center in addition to being served on the World–Wide–Web.'
author:
- 'Michael R. Meyer'
- Suzan Edwards
- 'Kenneth H. Hinkle'
- 'Stephen E. Strom'
title: |
Near–Infrared Classification Spectroscopy:\
H–band Spectra of Fundamental MK Standards
---
\#1\#2\#3\#4\#5\#6\#7[ to ]{}
=-0.2 truein
Introduction
============
With the recent development of large–format infrared array detectors, high quality photometric surveys are routinely conducted at wavelengths between 1–2.5 $\mu$m. Soon the completion of the 2 Micron All–Sky Survey (2MASS; Skrutskie 1997) and DENIS (Epchtein et al. 1997) will provide comprehensive catalogues of near–infrared sources with detection limits sensitive to a wide variety of stellar and non–stellar objects. Infrared spectra will be required for appropriate identification of many of these sources, and for further study of their astrophysical properties.
The pioneering study of Johnson and Mendez (1970) was the first to explore the spectra of a large sample of normal stars in the near–infrared. However many years passed before improvements in instrumentation made possible similar observations of large numbers of targets of astrophysical interest. The majority of the work done in near–infrared spectroscopy to date has been focused on the K–band, in large part because intrinsically cool or heavily obscured objects are typically brighter at K–band than in the J– or H–bands. In 1986, Kleinmann and Hall (1986; KH86) provided the first comprehensive medium resolution atlas ($R=3000$) of stellar spectra in the K-band, covering all luminosity classes, but restricted to spectral types between F8-M7. More recently, Wallace and Hinkle (1997; WH97) have extended the KH96 K–band atlas, using the same FTS spectrograph on the KPNO 4m with $R=3000$, but including stellar spectra spanning spectral types O-M and luminosity classes I-V. They also summarize the considerable body of work directed toward K–band spectroscopy in the last decade.
While in many situations, the K–band will be the wavelength selection of choice for spectroscopic studies of highly obscured or very cool objects, the presence of circumstellar dust ($T_{vap} < 2000 K$; Pollack et al. 1994) often results in significant excess continuum emission longward of $2~\mu$m. This continuum excess is commonly found in two important classes of objects: young stars with circumstellar disks (e.g. Meyer, Calvet, & Hillenbrand, 1997) and evolved stars with extensive envelopes from mass-loss (e.g. Le Bertre 1997). Near–infrared excess due to warm dust can also complicate spectroscopic studies of composite stellar systems aimed at discerning the stellar populations of other galaxies (e.g. Schinnerer et al. 1997). Continuum excess longward of $2~\mu$m will weaken or even render invisible the photospheric features in the K–band, while the photosphere will dominate at shorter wavelengths. In such a situation, near infrared spectra shortward of 2 $\mu$m will be required to see the stellar photosphere too obscured to be detected optically. To date, there has been relatively little work in the H–band (1.55–1.75 $\mu$m). Recent publications include: i) observations of 40 G, K, and M stars of luminosity class I and III at $R=1500$ by Origlia, Moorwood, and Oliva (1993); ii) the library of 56 spectra O–M of luminosity class I, II, and V at $R=500$ (Lancon & Rocca–Volmerange, 1992); iii) a library of 37 stars of luminosity classes I, III, V at $R=1500-2000$ (Dallier, Boisson, & Joly 1996) over a limited portion of the H–band; and iv) a study of 9 OB Stars at $R=570$ (Blum et al. 1997).
Here we present an H–band spectral atlas for 85 stars of nearly solar abundance with spectral types on the MK classification system ranging from 07-M5 and luminosity classes I-V. These $R=3000$ spectra were collected with the same FTS at the KPNO 4m as the K–band atlases of KH86 and WH97. In Section 2, we describe the sample selection and in Section 3 we describe the observations and calibration of the data. In Section 4 we discuss the dependence of the spectral features on temperature and luminosity and suggest a two–dimensional classification appropriate for late-type stars. In Section 5 we discuss near–IR spectral classification with regard to wavelength range/spectral resolution, and conclude with a summary of our results.
Defining the Sample
===================
In our sample selection, we chose optically visible stars which had previously been identified on the temperature and luminosity scales of the revised MK system (Keenan 1987). [^1] The majority of the stars were drawn from the following fundamental lists: i) Morgan, Abt, and Tapscott (1978) for 29 stars O6–G0; ii) Keenan and McNeil (1989) for 45 stars G0–K5; and iii) Kirkpatrick, Henry, and McCarthy (1991) for 5 late–type dwarfs K5–M3. We supplemented these primary standards with an additional 5 secondary standards from the compilation of Jaschek, Conde, and de Sierra (1964) and one late–type dwarf classified by Henry, Kirkpatrick, and Simons (1994).
In order to cover as complete a range of stellar temperature and luminosity as possible, we defined a two-dimensional grid with 26 bins of spectral type and three bins of luminosity class. Our temperature grid is binned $\times 2$ more coarsely in spectral subclass than the revised MK system, so that we typically sample only every other MK subclass. The three luminosity bins are divided into supergiants (I–II), giants (III), and subgiants/dwarf stars (IV–V). A full sampling of this grid would have resulted in 78 distinct temperature/luminosity pairs. Our atlas includes a total of 85 sources with 53 of the bins filled. Grid coverage was finer among the later spectral types, where for stars GO and later, we filled 26 of the 27 bins (9 spectral types $\times$ 3 luminosity classes). In contrast, for stars earlier than GO, only 27 of the 51 bins were covered (17 spectral types $\times$ 3 luminosity classes).
The 85 individual stars in our H–band survey are listed in Table \[sample\] along with relevent stellar properties taken from the Bright Star Catalogue (Hoffleit & Jaschek 1982). Additional restrictions on the sample selection included: i) v$sini$ $<$ 250 km s$^{-1}$ with exception of HR3982 (B7 V); ii) near-solar metallicity (avoiding those MK standards which exhibit spectral peculiarities due to enhanced or deficient metal abundance); and iii) no visual companions within the beam (separations $1-5$”). Our program was begun in advance of the K–band FTS atlas of WH97, and their sources are drawn in large measure from our sample. We note in Table \[sample\] the stars for which K–band spectra can be found in the WH97 digital atlas.
Table \[temp\] and Figure \[fig1\] provide additional insight into the temperature and luminosity coverage of our sample. In Table \[temp\], we list each of the spectral type and luminosity bins we have “filled”. For each bin in which there is at least one spectrum, we give the corresponding effective temperature. For most stars, we adopted the temperature scale of Tokunaga (1996), except for giants earlier than G0 where we adopted Schmidt–Kaler (1982) [^2]. Figure \[fig1\] provides a schematic illustrating the temperature and luminosity coverage for the 85 stars in our sample. In this illustration we have applied the same main sequence bolometric corrections to both dwarf (27) and subgiant (11) stars; as such they are indistinguishable in this diagram.
Observations and Data Calibration
=================================
Observations of our 85 sample stars were obtained at the Mayall 4m telescope at Kitt Peak National Observatory during four separate observing runs from 1993–1994 (Table \[ftslog\]). We used the Fourier Transform Spectrometer (FTS) dual–output interferometer (Hall et al. 1979). The FTS was ideal for this program for several reasons. First, the wavelength coverage of the FTS is limited only by the bandpass of the blocking filters, independent of the spectral resolution. This gave us complete coverage in the J– and H–bands which would have been difficult to obtain with available grating spectrographs. For example, our H–band spectral range is a factor of two greater than the spectra of Dallier et al. (1996). Secondly, the spectral resolution is fixed by the path difference scanned with the interferometer so we were able to chose the highest resolution possible and achieve S/N in excess of 75 for the majority of our sources. [^3] Finally, because of the novel background subtraction algorithm of the 4m FTS described below, we were able to observe the brightest stars in our sample ($H < 3.0^m$) during good daytime conditions (typically mornings). Combining daytime observations with targeted nighttime observations of key faint sources, the FTS provided a uniform set of high quality spectra for a large sample of spectral standards.
Our observing program included simultaneous spectral coverage in both the J–band and the H–band. However, the J–band data presented difficulties which made it expeditious to focus our initial effort on the H–band. The primary problems with the J-band spectra were; i) the inherent difficulty in data reduction due to rapid temporal variations in telluric water vapor absorption; and ii) and the relative paucity of strong features which would allow spectral classification over the full range of stellar temperature and luminosity. We defer discussion of the J–band spectra to a future contribution.
Spectra were collected simultaneously in the J– and H–bands with the use of a dichroic beamsplitter to separate the wavelengths longward and shortward of 1.5 $\mu$m. Each star was centered within an input aperture of 3.8 arcsec while sky background was measured through an identical aperture 50.0 arcsec away. The interferogram was scanned at a rate of 1 kHz as the path difference was varied continuously from 0.0–0.75 cm providing an unapodized resolution of 0.8 cm$^{-1}$. Data were obtained as separate scan pairs, with the path difference varied first in one direction and then the other. A forward–backward scan pair was treated as an “observation” and observations were repeated in beam–switching mode (A–B–B–A). Because the sky background from each aperture produces an interferogram shifted in phase by 180$^{\circ}$ at each set of detectors, source spectra are background subtracted in fourier space as they are collected. This permits observations of bright stars to be obtained during good daytime conditions. These beam–switched observations were repeated and scans were averaged until adequate signal–to–noise ratio (SNR) was achieved. The interferograms were transformed at Kitt Peak National Observatory yielding spectra in units of relative flux versus wavenumber ($\sigma$ in cm$^{-1}$). The transformed spectra were converted into fits format images and all data reduction was performed using the IRAF software package [^4]. The spectra were then convolved with a gaussian filter of half–width $\delta = 1.2$ cm$^{-1}$. This procedure, commonly referred to as apodization, eliminates “ringing” observed in the FTS spectra due to the finite scan path of the interferometer. The resulting apodized resolution (Rayleigh criterion) was $\delta \sigma = 2.1$ cm$^{-1}$ giving a mean resolving power of $R=3000$ in the H–band. At this stage the J– and H–band spectra were separated for ease of reduction. The slope of the continuum was normalized to 1.0 using a four–segment spline fitting function. Care was taken to keep the residuals from this fit to within 1 %.
Next we corrected the spectra for telluric absorption features present in the spectra which varied with zenith angle. We attempted to construct an opacity map for the earth’s atmosphere by dividing normalized spectra obtained of the A0 star standards at different airmass. Because of the simplicity of the A0 star spectra, showing primarily hydrogen lines in absorption, it was relatively easy to monitor the degree to which this procedure was successful. In dividing two normalized spectra of the same star taken at different airmass, all stellar photospheric absorption features should directly cancel, leaving only those absorption features due to the earth’s atmosphere. If we assume that the opacity of the telluric absorption is directly proportional to airmass we derive: $$\tau(\sigma,X = 1.0) = \frac{1}{(X_{high}-X_{low})} \times ln[ \frac{I(\sigma,X_{low})}{I(\sigma,X_{high})}]$$ where $\tau$ is the atmospheric opacity, $X_{low}$ is the low airmass value, and $X_{high}$ is the high airmass value. A typical opacity map derived in this way for the H–band is shown in Figure \[fig2\]. Several of the features in this map identified with known constituents of the earth’s atmosphere such as water vapor, methane, and carbon dioxide, are denoted in Figure \[fig2\]. Again if the atmospheric opacity varies linearly with airmass we can simply scale the opacity for each star so that $\tau(\sigma,X) = X \times \tau(\sigma,X=1.0)$. Using this technique we corrected the spectra to zero–airmass; $$I(\sigma,X=0.0) = I(\sigma,X) \times e^{\tau(\sigma,X)}$$ We used the highest signal–to–noise A0 standard star spectra ($SNR > 100$) with the largest $\Delta X$ to define the opacity. We found some residual telluric absorptions, possibly due to water vapor which do not vary strictly with airmass. Such features severely complicate the reduction of the J–band spectra.
Finally, the forward and backward scans of each star were averaged and residuals of the differenced spectra were calculated in order to evaluate the average SNR. The observations were obtained with the goal of achieving SNR of 75 or greater. In most cases this was achieved with the highest quality spectra reaching values of several hundreds. The average SNR for each stellar spectrum is included in Table \[ftshewsup\] –Table \[ftshewdw\] below.
Line Identification and Dependence on Temperature and Luminosity
================================================================
Representative H–band spectra are shown in Figures \[fig3\]– \[fig6\] for luminosity classes I–II, III, IV, and V, with prominent atomic and molecular features identified. Line identifications were made for the strongest lines from comparison with the solar photospheric and umbral near infrared atlases (Livingston & Wallace 1991 (LW91); Wallace & Livingston 1992 (WL92)). However, at our moderate spectral resolution many features are blended and we found the model atmosphere calculations of Oliva, Moorwood, and Origlia (1993) to be useful in identifying the dominant contributors to a blend in late–type stars.
Visual inspection of the features in Figures \[fig3\]– \[fig6\] reveals that $R=3000$ H–band spectra contain sufficent temperature and luminosity sensitive features to enable spectral classes to be distinguished. Beginning with the early type stars, the dominant spectral features are HeI 5882 cm$^{-1}$ (1.700 $\mu$m) and the Brackett series of hydrogen from lines 4-10 (1.736 $\mu$m) to 4-16 (1.556 $\mu$m). The He I line exceeds the strength of the Brackett lines in the very earliest stars (06 to B0), with a maximum equivalent width of $\sim 0.83$ cm$^{-1}$ (HR1903; B0 Ia), and recedes to undetectable levels ($\sim 0.10$ cm$^{-1}$) by spectral type B8. From the late B to early F stars, the Brackett series dominates the spectrum, after which lines of neutral atomic metals begin to take prominence. The strongest metallic lines include MgI, SiI, CaI, AlI, and FeI, which increase in strength toward the K stars. Finally molecular features of OH and CO dominate the spectra of the latest–type stars from K5–M5. The most striking luminosity-sensitive feature is the second–overtone CO bandhead $[v, v^{'} = 6, 3]$ at 6177 cm$^{-1}$ (1.619 $\mu$m), which is found in the spectra of the K and M stars. This feature is signficantly stronger in stars of lower surface gravity at equivalent spectral type.
To further enable spectral classification in the H–band, we have identified a set of 9 features which are prominent in stars of spectral type A-M. These include a relatively isolated Brackett line (H4-11), 5 neutral metals, and 3 molecular bands. In Table \[ftshband\], we define 9 narrow band indices with bandpasses ranging from 10 to 50 cm$^{-1}$, which include each of these features. The variable widths of the bandpasses were selected to minimize line blending, contamination from residual telluric absorption, and sensitivity to radial velocity shifts. Table \[ftshband\] also identifies the wavenumber of the dominant contributor and the lower state energy level, the central wavenumber and passband of the index, and additional species which may contribute to the index strength. The equivalent widths of these 9 indices were evaluated from the normalized spectra of our 85 survey stars, and are tabulated in Tables \[ftshewsup\] to \[ftshewdw\] in units of cm$^{-1}$ [^5]. Uncertainties in these equivalent widths depend on the SNR of the spectrum in question and the bandpass/strength of the feature. Errors range from $\sigma_{EW} = 0.02-0.1$ cm$^{-1}$ exceeding this upper limit in very few cases. Multiple observations of several sources are listed for comparison.
The temperature and luminosity dependence for four representative indices is illustrated in Figure \[fig7\]. The 4–11 Brackett line (HI5950) behaves as expected, with a rapid rise to a maximum (at a peak equivalent width of $\sim$ 3 cm$^{-1}$) as $T_{eff}$ approaches 10000 K, and a slower decline toward higher temperatures. The behavior of the index is similar in both the dwarfs and the giants, although the luminosity class I/II sources show a larger scatter, presumably due to intrinsic variability (e.g. Kaufer et al. 1996). The general behavior of the neutral atomic features is illustrated by the Mg6345 index. In luminosity classes IV–V, this index reaches a maximum strength between 5000-6000 K, with a peak equivalent width of $\sim$ 2.5 cm$^{-1}$. In contrast, the maximum strength of this index in the lower surface gravity objects (also $\sim$ 2.5 cm$^{-1}$) is found in the coolest stars in our sample, monotonically decreasing toward higher temperatures; as expected given the behavior of ionization state as a function of surface gravity. The two SiI indices exhibit similar behavior, but the AlI index, (not shown) turns over at much lower temperatures in our dwarf stars because of its lower ionization potential. Note that we have chosen not to form an index based on the strongest SiI line at 6292 cm$^{-1}$ (1.5892 $\mu$m) because it is coincident with the 4–14 Brackett line of HI at 6297 cm$^{-1}$ (1.5881 $\mu$m).
The behavior of the molecular features is illustrated for both the second–overtone $^{12}$CO (6,3) and the OH ($\Delta v = 2$) indices. Both indices exhibit a similar behavior with temperature and luminosity, becoming detectable around $T_{eff}$ = 5000, with a strength in the giants approximately twice that in the dwarfs. Similar behavior was noted by KH86 in the first–overtone CO features in the K–band. In dwarf stars the second–overtone CO index reaches a maximum before M5, and displays a turnover toward the coolest stars. This may be due in part to features of CaI and FeI which contaminate the index for intermediate spectral types (F5–K3). Ali et al. (1995) find that the relationship between T$_{eff}$ and equivalent widths of the first–overtone CO bandheads flatten out between 3500–5000 K in dwarf stars. From high resolution ($R > 45,000$) FTS spectra, Wallace & Hinkle (1996) observed that the 2 $\mu$m continuum in late–type dwarf stars is suppressed by numerous water vapor features which are blended at intermediate to low resolution. Predicting the equivalent widths of features where the apparent continuum is subject to temperature and luminosity effects is not straight–forward. In contrast, both the CO and OH indices continue to rise at the coolest temperatures for stars of higher luminosity. However, the magnitude (and temperature) of the maximum in the dwarf stars differs between the CO and OH indices, which we use in the next section to define a two-dimensional classification scheme for late–type stars. We note that the OH index begins to include a contribution from the stark–broadened 4-11 Brackett line at 5949 cm$^{-1}$ creating a secondary maxiumum in the strength of this index around 10,000 K in the dwarf stars.
While the temperature and luminosity dependence of the atomic features is readily understood through application of the Saha and Boltzman equations governing the population of the ionization states and energy levels respectively, the explanation behind the behavior of the molecular features is more subtle. Two possibilities for the factor of two enhancement in the molecular bands in the giants over the dwarfs have been explored in the literature. One attributes the luminosity dependence in the molecular features to differing microturbulence in the atmospheres of dwarfs and giants. The expectation is that larger microturbulence in the lower surface gravity giants effectively broadens the opacity of the feature over a larger frequency interval in these saturated features, thereby enhancing the equivalent width (McWilliams & Lambert 1984). Another possible contributor is the differing depth of the line formation region in the dwarfs versus the giants, which is fixed by the H$^{-}$ opacity. As described by Gray (1992) higher surface gravity results in a higher electron pressure (and thus H$^-$ column density). This brings the CO line formation region closer to the stellar surface reducing $N_{CO}$ according to the following proportionality: $$P_e \sim g^{1/3} \sim N_{H^-} \sim 1/N_{CO}$$ In any case, this luminosity dependence of the band strength gives an excellent empirical discriminant between giants and dwarfs, which we exploit below to develop a two dimensional spectral index.
To discern surface gravity effects between the super–giants and giants or between sub–giants and dwarf stars requires more careful study. A detailed examination of line strengths as a function of surface gravity at a fixed temperature reveal the expected trends. However, this behavior does not reveal itself in the coarse analysis afforded by our narrow–band indices.
While the temperature and luminosity sensitivities outlined above can provide good spectral classification in many instances, discriminants which do not rely on absolute line strength are required when a star is subject to near-infrared continuum veiling. In this case line ratio diagnostics are to be preferred, since absorption features will appear shallower in the presence of continuum excess but line ratios will be preserved as long as the excess is not strongly wavelength dependent. We have identified one diagnostic based on line ratios which can be used to evaluate both temperature and luminosity for stars from K3-M5 in the presence of continuum veiling. This two–dimensional spectral index is defined as:
$$\frac{EW[OH5920]}{EW[Mg6345]} \ vs.
\ \frac{EW[CO6018+CO6170]}{EW[Mg6345]}$$
where EW is the equivalent width in cm$^{-1}$ for the indices identified in Table \[ftshband\] and listed in Tables \[ftshewsup\]– \[ftshewdw\]. The temperature and luminosity dependence of this diagnostic is illustrated in Figure \[fig8\]. In this diagnostic, the ratio of the OH5920 to Mg6345 indices is temperature sensitive, with distinct temperature dependences for dwarfs and giants. Specifically we find
$$T_{eff} (V) = 4640 \pm 250 - (2610 \pm 110) \frac{EW[OH5920]}{EW[Mg6345]}$$
and
$$T_{eff} (III) = 5100 \pm 180 - (2730 \pm 80)\times \frac{EW[OH5920]}{EW[Mg6345]}$$
The comparison of this temperature sensitive ratio with the sum of the two $^{12}$CO indices, also normalized to Mg6345, then provides an excellent means of identifying both the temperature and luminosity class of late–type stars. Formal errors in the equivalent width suggest that spectral types can be evaluated to within $\pm 2$ subclasses ($\pm$ 300 K) from K3–M5 using spectra with $SNR > 50$ based on these indices alone.
Discussion and Summary
======================
Spectral classification in the near-infrared will become increasingly important in the next decade, as the 2MASS and DENIS near-infrared sky surveys reveal unprecented numbers of stars which are optically–invisible. Because the 1–2.5 $\mu$m region is on the Rayleigh–Jeans tail of most stellar SEDs, it is not an ideal wavelength regime to pursue spectral classification. Yet, there are sufficient features in both the H– and the K–bands to allow most stellar photospheres to be classified. For heavily reddened sources, the K–band will be the wavelength of choice. However, continuum emission from circumstellar dust with temperatures less than 2000 K can heavily veil stellar photospheres at wavelengths greater than 2.0 $\mu$m. In this case, shorter wavelength spectra are required in order to identify the underlying star.
The early-type stars are probably the most challenging for spectral classification in the near–infrared. We find that a rough classification from 07-B8 can be made in the H–band by the relative strengths of HeI 5882 cm$^{-1}$ and the Brackett series (see also Blum et al. 1997). Hanson, Conti, and Rieke (1995) have established a classification scheme in the K–band for O–B stars relying on lines of helium as well as higher ionization species obtained at $R > 1000$. For stars A through early K, the H–band may be superior to the K band in providing a large number of intermediate ionization potential species with strong features such as MgI, SiI, and FeI in addition to the numerous Brackett series features. Stars K3–M5 are probably best classified in the K–band (KH86; Ali et al. 1995; WH97) using atomic features of MgI, CaI, and NaI as well as the first–overtone CO bandheads observed at $R \sim 1000$. However we have found that these stars also have strong temperature and luminosity sensitive features in the H–band such as MgI, AlI, OH, and the second–overtone CO bandheads. The very latest–type stars ($> M5$) have very strong, broad, molecular features which can be identified at resolutions as low as $R \sim 300$. Kirkpatrick et al. 1993 (see also Jones et al. 1996) have classified stars in the I– or the J–band employing features due to VO, TiO, and FeH. In addition, broad water vapor bands observed throughout the 1–2.5 $\mu$m region (Jones et al. 1995) are an important opacity source in the atmospheres of the coolest stars as well as brown dwarfs (Allard & Hauschildt 1995). While the I– or J–bands are probably the best spectral regions to classify extremely cool stars (as they lie on the Wien side of the Planck function for these objects) more heavily obscured objects can still be profitably observed at low resolution in the J– and H–bands (e.g. NICMOS on HST) or in the K–band (Wilking, Greene, & Meyer 1998) in search of these water vapor absorptions.
The H–band spectral atlas we have presented is comprised of moderate resolution spectra with $R \sim 3000$. In contrast, most spectral classification is typically carried out with $R \sim 500-1000$. The strongest and broadest features in the H–band are the CO(6-3) bands and the Brackett lines. These features could be identified with much lower spectral resolution than our survey, at $R \sim 500$. The most crowded region in the H–band spectra is that in the vicinity of the HI line at 5948.50 cm$^{-1}$ (1.68110 $\mu$m), the AlI triplet at 5964–5980 cm$^{-1}$ (1.677–1.672 $\mu$m), and the SiI line at 5993.29 cm$^{-1}$ (1.66853 $\mu$m). In order to properly separate these important features from each other, a resolving power of $R \sim 1000$ is required. At this resolution, one can also obtain measurements of the HeI line at 5882 cm$^{-1}$ (1.700 $\mu$m). At $R = 3000$ one can resolve individual components of the AlI triplet, the Mg I doublet, and the CO bandheads, as well as the stark–broadened Brackett lines in the early–type dwarf stars (Table \[ftshband\]). An additional issue in the near–infrared is the significant contribution to shot-noise from air-glow lines. In the H–band air–glow from OH is sufficently bright and variable that they compromise $R=1000$ H–band spectral classification for very faint sources. Spectral resolution as high as $R \sim 5000$ will be required to resolve the bulk of these air–glow features and to obtain adequate SNR spectra of faint objects [^6].
In summary, we present an H–band spectral atlas at a resolving power of $R = 3000$ that spans a wide range in stellar temperature (O7-M5) and luminosity class (I-V). This spectral region contains a number of temperature and/or luminosity sensitive atomic and molecular features which will allow spectral classification to be carried out in the H-band. As an example of the efficacy of this spectral range for distinguishing stellar spectral types, we define a set of narrow–band indices which, with $ SNR \sim 50$, permit classification of late–type stars on the MK system within $\pm 2$ subclasses. It appears however, that for most applications obtaining H–band spectra at $R \sim 1000$ will be sufficient for classification.
Appendix A: Electronic Availability of the Data
===============================================
The final reduced averaged spectra as well as the difference of the forward and backward scan pairs (see Section 3 for description of the reduction procedure) are available through the Astronomical Data Center (ADC) for each observation listed in this paper. The ADC can be contacted directly: i) by post at Astronomical Data Center, NASA Goddard Space Flight Center, Code 631, Greenbelt, MD 20771; or ii) by telephone at (301) 286–8310; or iii) by fax at (301) 286–1771; or iv) via the internet at http://adc.gsfc.nasa.gov. The data are in fits format with pertinent header information included for each image. These fits format files, useful plotting routines, and other relevent information are also available on the World Wide Web at http://donald.phast.umass.edu. The raw FTS data are also available directly from NOAO (contact KHH for details).
We would like to thank Lori Allen, Ed Chang, Lynne Hillenbrand, Susan Kleinmann, Michael Skrutskie, and Lloyd Wallace for helpful discussions. Special thanks to John Carpenter for assisting in the initial compilation of the standard star lists, and to Karen Strom and Stephen Friedman for their assistance in making the data available electronically. Antonella Romano provided assistance in preparing the tables and figures for publication. Support for MRM during the final stages of this work was provided by NASA through Hubble Fellowship grant \# HF–01098.01–97A awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA under contract NAS 5–26555. SE acknowledges support from the National Science Foundation’s Faculty Award for Women Program. This work was supported in part through a grant from the National Science Foundation (\# AST–9114863) to SES.
Ali, B., Carr, J.S., DePoy, D.L., Frogel, J.A., & Sellgren, K. 1995, AJ, 110, 2415
Allard, F., & Hausschildt, P.H. 1995, ApJ, 445, 433
Bell, J.R. 1974, Fourier Transform Spectroscopy, (John Willey and Sons: New York)
Bessell, M.S., Castelli, F., & Plez, B. 1998, A&A, 333, 231
Blum, R.D., Ramond, T.M., Conti, P.S., Figer, D.F., & Sellgren, K. 1997, AJ, 113, 1855
Coxon, G. 1965, Ark. Fys., 28, 381
Dallier, R., Boisson, C., & Joly, M. 1996, A&ASS, 116, 239
Epchtein, N. et al. 1997, Impact of Large Scale Near-IR Sky Surveys, eds. Gorzon, F., Epchtein, N., Omont, A., Burton, B., & Persi, P. (Kluwer: Amsterdam, Netherlands)
Eriksson, K.B.S., & Isberg, H.B.S. 1963, Ark. Fys., 23, 527
Garcia, J.D., & Mack, J.E. 1965, JOSA, 55, 654
George, T., Urban, W., & LeFloch, A. 1994, J. Mol. Spec., 165, 50
Gray, D.F. 1992, The Observation and Analysis of Stellar Photospheres, (Cambridge University Press: Cambridge), p. 374
Hall, D.N.B, Ridgeway, S., Bell, E., & Yarborough, J.M. 1979, Proc. SPIE, 248, 898
Hanson, M., Conti, P., & Rieke, M. 1996, ApJS, 107, 281
Henry, T.J., Kirkpatrick, J.D., & Simons, D. 1994, AJ, 108, 1437
Herbst, T.M. 1994, PASP, 106, 1298
Hoffleit, D., & Jaschek, C. 1982, The Bright Star Catalogue, Yale Observatory
Jaschek, C., Conde, H., & de Sierra, A.C. 1964, La Plata Observatory Bulletin, 28
Johnson, H.L., & Mendez, M.E. 1970, AJ, 75, 785.
Jones, H.R.A., Longmore, A.J., Allard, F., Hauschildt, P.H., Miller, S., & Tennyson, J. 1995, MNRAS, 277, 767
Jones, H.R.A., Longmore, A.J., Allard, F., & Hauschildt, P.H. 1996, MNRAS, 280, 77
Kaufer, A. et al. 1997, A&A, 320, 273
Keenan, P.C., 1985, IAU, 111, 123
Keenan, P.C., 1987, PASP, 99, 713
Keenan, P.C., & McNeil, R. 1989, APJS, 71, 245
Kirkpatrick, J.D., Henry, T.J., & McCarthy, D.W. Jr. 1991, APJS, 77, 417
Kirkpatrick, J.D., Kelly, D.M., Rieke, G.H., Liebert, J., Allard, F., Wehrse, R. 1993, ApJ, 402, 643
Kleinmann, S.G., & Hall, D.N.B. 1986, APJS, 62, 501 (KH86)
Koornneef, J. 1983, A&A, 128, 84
Lancon, A., & Rocca-Volmerange, B. 1992, A&AS, 96, 593
Le Bertre, T. 1997, A&A, 324, 1059
Litzen, U. 1964, Ark. Fys., 28, 239
Livingston, W., & Wallace, L. 1991, Solar Atlas 1–5 $\mu$m, Kitt Peak Observatory Bulletin (LW91)
McWilliam, A., & Lamber, D.L. 1984, PASP, 96, 882
Meyer, M.R., Calvet, N., & Hillenbrand, L.A. 1997, AJ, 114, 288
Morgan, W., Abt, H., & Tabscott, J. 1978, Revised MK Spectral Atlas for Stars Ealier than the Sun, Yerkes and Kitt Peak Observatories
Origlia, L., Moorwood, A., & Oliva, E. 1993, A&A, 280, 536
Pollack, J.B., Hollenbach, D., Beckwith, S.V.W., Damon, P., Rousch, T., & Fong, W. 1994, ApJ, 421, 615.
Risberg, G. 1965, Ark. Fys., 28, 381
Schinnerer, E., Eckart, A., Quirrenbach, A., Boker, T., Tacconi–Garman, L.E., Downes, D. 1997, ApJ, 488, 174
Skrutskie, M.F. 1997, Impact of Large Scale Near-IR Sky Surveys, eds. Gorzon, F., Epchtein, N., Omont, A., Burton, B., & Persi, P. (Kluwer: Amsterdam, Netherlands)
Schmidt–Kaler, T.H. 1982, Physical Parameters of Stars, Landolt-Bornstein New Series, Vol. 2b, Astronomy and Astrophysics, Stars and Star Clusters, ed. K. Shaifers & H. H. Voigt, (Springer–Verlag: New York)
Tokunaga, A., 1996, Astrophysical Quantities, submitted
Wallace, L., & Livingston, W. 1992, Atlas of Dark Sunspot Spectrum 1–5 $\mu$m, Kitt Peak Observatory Bulletin (WL92)
Wallace, L. & Hinkle, K. 1996, APJS, 107, 312
Wallace, L. & Hinkle, K. 1997, APJS, 111, 445 (WH97)
Wilking, B.A., Greene, T.P., & Meyer, M.R. 1998, AJ, submitted
[lllllll]{} & & & & & &\
1903 & \*46 $\epsilon$ Ori & 2.4 & B0 & Ia & 26SB & 87\
1203 & 44 $\zeta$ Per & 3.4 & B1 & Ib & 20SB & 59\
2827 & 31 $\eta$ CMa & 2.5 & B5 & Ia & 41V & 45\
1713 & \*19 $\beta$ Ori & 0.1 & B8 & Ia & 21SB & 33\
3975 & \*30 $\eta$ Leo & 3.3 & A0 & Ib & 3V & 20\
7924 & \*50 $\alpha$ Cyg & 1.0 & A2 & Ia & -5SBO & 21\
1865 & 11 $\alpha$ Lep & 2.0 & F0 & Ib & 24 & 13\
1017 & \*33 $\alpha$ Per & 0.9 & F5 & Ib & -2V & 18\
7796 & \*37 $\gamma$ Cyg & 1.1 & F8 & Ib & -8V & 20\
8232 & \*22 $\beta$ Aqr & 1.5 & G0 & Ib & 7V? & 18\
7479 & 5 $\alpha$ Sge & 2.4 & G1 & II & 2V? & 0\
7063 & $\beta$ Sct & 2.2 & G4 & IIa & -22SB10 & 10\
8752 & \*- & 3.6 & G4v & $>I$ & -58V? & 35\
7314 & \*21 $\theta$ Lyr & 2.1 & K0 & +II & -31V & $<$19\
6713 & 93 Her & 2.4 & K0.5 & IIb & -24 & $<$17\
8465 & \*21 $\zeta$ Cep & 1.1 & K1.5 & Ib & -18SB & $<$17\
6498 & 49 $\sigma$ Oph & 1.9 & K2 & II & -27 & $<$19\
603 & 57 $\gamma^1$ And & -0.5 & K3 & -IIb & -12SB & $<$17\
8089 & \*63 Cyg & 1.5 & K4 & Ib-IIa & -26V & –\
8079 & \*62 $\xi$ Cyg & 0.5 & K4.5 & Ib-II & -20SB & $<$17\
2061 & 58 $\alpha$ Ori & -2.? & M1-2 & Ia-Iab & 21SB & –\
1155 & \*– & 0.5 & M2 & +IIab & -3V & –\
921 & 25 $\rho$ Per & -1.7 & M4 & II & 28 & –\
7009 & \*– & 0.6 & M4.5-M5 & +II & -19 & –\
6406 & \*64 $\alpha^1$ Her & -2.4 & M5 & Ib-II & -33V & 21\
1899 & 44 $\iota$ Ori & 3.5 & O9 & III & 22SB2O & 130\
1552 & 3 $\pi^4$ Ori & 4.1 & B2+B2 & III & 23SBO & 40\
5291 & \*11 $\alpha$ Dra & 3.5 & A0 & III & -13SBO & 18\
403 & \*37 $\delta$ Cas & 2.3 & A5 & III-IV & 7SB & 113\
1412 & \*78 $\theta^2$ Tau & 2.9 & A7 & III & 40SB1O & 78\
4031 & \*36 $\zeta$ Leo & 2.8 & F0 & III & -16SB & 84\
21 & \* 11 $\beta$ Cas & 1.6 & F2 & III–IV & 12SB & 70\
5017 & \*20 CVn & 3.9 & F3 & III & 8V? & 17\
2706 & 48 Gem & 5.0 & F5 & III-IV & 13V & 74\
8905 & \*68 $\upsilon$ Peg & 3.3 & F8 & III & -11 & 79\
4883 & \*31 Com & 3.0 & G0 & III & -1V? & 77\
4716 & 5 CVn & 2.8 & G6 & III & -12SB & $<$17\
7328 & 1 $\kappa$ Cyg & 1.6 & G9 & III & -29SB & $<$17\
7949 & 53 $\epsilon$ Cyg & 0.2 & K0 & -III & -11SB? & $<$17\
8317 & \*11 Cep & 2.3 & K1 & III & -37 & $<$17\
6299 & \*27 $\kappa$ Oph & 0.8 & K2 & III & -56V & $<$17\
165 & 31 $\delta$ And & 0.5 & K3 & III & -7SB1O & $<$17\
6705 & \*33 $\gamma$ Dra & -1.2 & K5 & III & -28 & $<$17\
152 & – & 1.7 & K5-M0 & III & -33V & $<$17\
4517 & \*3 $\nu$ Vir & 0.3 & M1 & IIIab & 51V? & –\
6242 & – & 0.4 & M4 & +III-IIIa & -7V? & –\
7886 & \*– & -0.6 & M6 & III & -66V? & –\
\[sample\]
[lllllll]{} & & & & & &\
6588 & 85 $\iota$ Her & 4.3 & B3 & IV & -20SB1O & 11\
4033 & \*33 $\lambda$ UMa & 3.3 & A2 & IV & 18V & 48\
1351 & 57 Tau & 4.9 & F0 & IV & 42SB1? & 109\
5235 & 8 $\eta$ Boo & 1.5 & G0 & IV & 0SB1O & 13\
5409 & 105 $\phi$ Vir & 2.8 & G2 & IV & -10SB & 0\
6623 & 86 $\mu$ Her & 1.4 & G5 & IV & -16V & 20\
995 & 59 Ari & 3.9 & G6 & IV & 0V & –\
7602 & 60 $\beta$ Aql & 1.7 & G8 & IV & -40V & $<$ 16\
7957 & \*3 $\eta$ Cep & 1.2 & K0 & IV & -87 & $<$17\
5901 & 11 $\kappa$ CrB & 2.5 & K1 & IVa & -24 & $<$17\
6014 & – & 3.6 & K1.5 & IV & -4V & –\
2456 & \*15 Mon & 5.5 & O7 & V(e) & 33SB & 63\
5191 & \*85 $\eta$ UMa & 2.4 & B3 & V & -11SB? & 205\
3982 & \*32 $\alpha$ Leo & 1.6 & B7 & V & 6SB & 329\
7001 & \* $\alpha$ Lyr & 0.0 & A0 & V & -14V & 15\
2491 & \*9 $\alpha$ CMa & -1.5 & A1 & Vm & -8SBO & 13\
4534 & \*94 $\beta$ Leo & 2.0 & A3 & V & 0V & 121\
4357 & \*68 $\delta$ Leo & 2.3 & A4 & V & -20V & 181\
4931 & 78 UMa & 4.1 & F2 & V & -10V? & 92\
1279 & – & 5.1 & F3 & V & 36SB1? & 25\
2943 & \*10 $\alpha$ CMi & -0.6 & F5 & IV-V & -3SBO & 6\
1538 & 59 Eri & 2.3 & F6 & V & 35 & –\
4375 & \*53 $\xi$ UMa & 3.0 & G0 & V & -16SB1O & 1\
4983 & 43 $\beta$ Com & 3.1 & F9.5 & V & 6SB? & 6\
483 & \*– & 3.7 & G1.5 & V & 4V? & 2\
4374 & 53 $\xi$ UMa & 3.5 & G0 & V & -16SB1O & 3\
5072 & 70 Vir & 3.6 & G4 & V & 5V & 1\
4496 & \*61 UMa & 3.8 & G8 & V & -5V & $<$17\
7462 & \*61 $\sigma$ Dra & 3.0 & K0 & V & 27V & $<$17\
1084 & \*18 $\epsilon$ Eri & 1.6 & K2 & V & 15V? & $<$17\
8832 & \*– & 3.2 & K3 & V & -18V & –\
– & GL570A & 3.0 & K4 & V & - & -\
8085 & \*61 Cyg & 2.4 & K5 & V & -64V & $<$17\
8086 & 61 Cyg & 3.1 & K7 & V & -64V? & =$<$25\
– & GL338A & 4.5 & M0 & V & - & -\
– & GL526 & 4.5 & M1.5 & V & - & -\
– & \*GL411 & 3.6 & M2 & V & - & -\
– & \*GL725A & 4.7 & M3 & V & - & -\
[rlll]{} & & &\
O6-O8– & & 37000 & 38000\
O9– & 32500 & 32000 & 33200\
O9.5– & & & 31450\
B0– & 26000 & 29000 & 29700\
B1– & 20700 & 24000 & 25600\
B2– & 17800 & 20300 & 22300\
B3– & 15600 & 17100 & 19000\
B4 & 13900 & & 17200\
B5– & 13400 & 15000 & 15400\
B6 & 12700 & 14100 & 14100\
B7– & 12000 & 13200 & 13000\
B8– & 11200 & 12400 & 11800\
B9– & 10500 & 11000 & 10700\
A0– & 9730 & 10100 & 9480\
A1– & 9230 & 9480 &\
A2 & 9080 & 9000 & 8810\
A5– & 8510 & 8100 & 8160\
A7– & & 7650 & 7930\
F0– & 7700 & 7150 & 7020\
F2– & 7170 & 6870 & 6750\
F5– & 6640 & 6470 & 6530\
F7– & & & 6240\
F8– & 6100 & 6150 &\
G0– & 5510 & 5910 & 5930\
G2 & & & 5830\
G3 & 4980 & &\
G4 & & 5190 & 5740\
G6 & & 5050 & 5620\
G8– & 4590 & 4960 &\
K0– & 4420 & 4810 & 5240\
K1– & 4330 & 4610 &\
K2– & 4260 & 4500 & 5010\
K3– & 4130 & 4320 &\
K4– & & 4080 & 4560\
K5– & 3850 & 3980 & 4340\
K7 & & & 4040\
M0– & 3650 & 3820 & 3800\
M1– & 3550 & 3780 & 3680\
M2– & 3450 & 3710 & 3530\
M3– & 3200 & 3630 & 3380\
M4– & 2980 & 3560 & 3180\
M5– & & 3420 & 3030\
M6– & & 3250 & 2850\
\[temp\]
[lll]{} & &\
March 9–10, 1993 & Day & 17\
April 1–3, 1993 & Day/Night & 30\
May 18–19, 1993 & Day & 10\
January 30–31, 1994 & Day/Night & 42\
\[ftslog\]
[lllllll]{} & & & & & &\
MgI(4s–4p) & 5.39 & 5843.41 & 1.71133 & 5844 & 10 & CO, Fe, Ni, OH\
OH($\Delta v = 2$) & 0.76 & 5920: & 1.689 & 5920 & 20 & C, CO, Fe, Ni\
HI(4-11) & 12.75 & 5948.50 & 1.68110 & 5950 & 20 & CO, Fe, Ni, Si\
AlI(4p–4d tr) & 4.09 & 5963.76 & 1.67679 & 5972.5 & 25 & CO, Fe, Ni, OH\
& & 5968.31 & 1.67552 & & &\
& & 5979.60 & 1.67235 & & &\
SiI(4p–3d) & 5.98 & 5993.29 & 1.66853 & 5993 & 10 & CO, Fe, Ni, OH\
$^{12}$CO(8,5)bh & 1.55 & 6018 & 1.662 & 6017.5 & 15 & Fe, OH, S\
$^{12}$CO(6,3)bh & 1.05 & 6177 & 1.619 & 6170 & 50 & Ca, Fe, Ni, OH, Si\
SiI(4p–5s) & 5.98 & 6263.92 & 1.59644 & 6264 & 10 & Fe, Mg, Ni, OH\
MgI(4s–4p tr) & 5.93 & 6341.10 & 1.57701 & 6345 & 20 & CN, CO, Fe, H$_2$O, Ni, OH\
& & 6347.88 & 1.57533 & & &\
& & 6351.22 & 1.57450 & & &\
\[ftshband\]
[llllllllllll]{} & & & & & & & & & & &\
HR1903 & 26000 & 203 & -0.03 & 0.09 & 0.75 & 0.17 & -0.02 & 0.11 & 0.27 & 0.11 & 0.47\
HR1903 & 26000 & 167 & 0.01 & 0.07 & 0.73 & 0.11 & -0.00 & 0.09 & 0.19 & 0.12 & 0.32\
HR1203 & 20700 & 160 & -0.01 & 0.12 & 0.80 & 0.18 & 0.01 & 0.11 & 0.23 & 0.11 & 0.56\
HR2827 & 13400 & 045 & -0.05 & -0.02 & 1.06 & -0.04 & 0.20 & -0.04 & 0.01 & -0.05 & 0.08\
HR1713 & 11200 & 218 & 0.04 & 0.06 & 1.27 & 0.23 & 0.05 & 0.09 & 0.19 & 0.17 & 0.50\
HR3975 & 9730 & 068 & -0.08 & 0.11 & 2.12 & 0.23 & 0.12 & -0.09 & 0.33 & 0.20 & 0.58\
HR7924 & 9080 & 225 & -0.05 & 0.00 & 1.40 & 0.20 & -0.02 & 0.03 & -0.14 & 0.08 & 0.08\
HR1865 & 7700 & 196 & -0.01 & 0.22 & 2.21 & 0.32 & 0.14 & -0.13 & 0.30 & 0.25 & 0.52\
HR1017 & 6640 & 248 & 0.08 & 0.25 & 1.88 & 0.43 & 0.06 & -0.02 & 0.53 & 0.41 & 1.05\
HR7796 & 6100 & 324 & 0.10 & 0.29 & 1.58 & 0.69 & 0.14 & 0.03 & 0.79 & 0.60 & 0.94\
HR8232 & 5510 & 290 & 0.09 & 0.37 & 1.28 & 0.67 & 0.17 & 0.03 & 1.18 & 0.67 & 1.19\
HR8752 & 5510 & 084 & -0.13 & 0.34 & 1.16 & 0.14 & 0.07 & 0.02 & 0.16 & 0.16 & 0.41\
HR7479 & 5333 & 096 & 0.15 & 0.13 & 1.13 & 0.70 & 0.10 & 0.00 & 0.92 & 0.41 & 0.69\
HR7479 & 5333 & 109 & 0.10 & 0.22 & 1.07 & 0.49 & 0.11 & 0.02 & 0.76 & 0.54 & 1.04\
HR7479 & 5333 & 182 & 0.13 & 0.12 & 1.08 & 0.56 & 0.05 & 0.07 & 0.94 & 0.49 & 1.06\
HR7479 & 5333 & 093 & 0.06 & -0.02 & 1.20 & 0.53 & 0.03 & 0.14 & 1.05 & 0.54 & 0.93\
HR7063 & 4902 & 260 & 0.17 & 0.12 & 0.94 & 0.84 & 0.14 & 0.25 & 1.66 & 0.76 & 1.06\
HR7314 & 4420 & 193 & 0.23 & 0.15 & 0.84 & 0.88 & 0.15 & 0.46 & 2.25 & 1.00 & 1.60\
HR7314 & 4420 & 257 & 0.15 & 0.17 & 0.67 & 0.59 & 0.10 & 0.37 & 1.91 & 0.86 & 1.36\
HR6713 & 4375 & 265 & 0.22 & 0.19 & 0.81 & 0.80 & 0.16 & 0.40 & 1.86 & 0.79 & 1.25\
HR8465 & 4295 & 308 & 0.13 & 0.28 & 0.68 & 0.70 & 0.19 & 0.63 & 2.35 & 1.02 & 1.79\
HR6498 & 4260 & 247 & 0.24 & 0.30 & 0.84 & 1.00 & 0.21 & 0.83 & 2.75 & 1.11 & 1.84\
HR603 & 4130 & 453 & 0.24 & 0.34 & 0.69 & 0.61 & 0.19 & 0.49 & 2.09 & 0.87 & 1.46\
HR8089 & 3990 & 202 & 0.21 & 0.45 & 0.50 & 0.67 & 0.14 & 0.68 & 2.26 & 0.94 & 1.59\
HR8079 & 3920 & 445 & 0.24 & 0.67 & 0.62 & 0.80 & 0.24 & 1.05 & 2.70 & 1.05 & 1.72\
HR2061 & 3550 & 327 & 0.54 & 1.08 & 0.36 & 1.39 & 0.23 & 1.46 & 3.92 & 0.89 & 2.22\
HR1155 & 3450 & 237 & 0.46 & 1.31 & 0.58 & 1.12 & 0.38 & 1.37 & 3.35 & 1.12 & 2.03\
HR921 & 2980 & 225 & 0.77 & 1.41 & 0.29 & 1.27 & 0.37 & 1.10 & 4.08 & 0.67 & 1.52\
HR7009 & 2925 & 292 & 0.63 & 1.43 & 0.79 & 1.55 & 0.54 & 1.79 & 4.68 & 1.26 & 2.21\
HR6406 & 2800 & 319 & 0.59 & 1.41 & 0.80 & 1.66 & 0.54 & 1.85 & 4.98 & 1.31 & 2.31\
\[ftshewsup\]
[llllllllllll]{} & & & & & & & & & & &\
HR1899 & 32000 & 164 & -0.03 & 0.17 & 0.54 & 0.15 & -0.04 & 0.03 & 0.18 & 0.10 & 0.49\
HR1899 & 32000 & 182 & -0.04 & 0.20 & 0.62 & 0.11 & 0.05 & -0.03 & 0.07 & -0.00 & 0.07\
HR1552 & 20300 & 115 & -0.07 & 0.13 & 1.29 & 0.31 & 0.10 & -0.09 & 0.12 & -0.01 & 0.09\
HR5291 & 10100 & 159 & -0.07 & 0.19 & 2.65 & 0.78 & 0.02 & 0.05 & 0.28 & 0.11 & 0.41\
HR403 & 8100 & 202 & -0.02 & 0.39 & 2.79 & 0.77 & 0.09 & -0.13 & 0.25 & 0.08 & 0.54\
HR1412 & 7650 & 186 & -0.02 & 0.43 & 2.70 & 0.71 & 0.05 & -0.10 & 0.60 & 0.25 & 0.91\
HR4031 & 7150 & 193 & 0.02 & 0.29 & 2.30 & 0.54 & 0.12 & -0.09 & 0.25 & 0.13 & 0.48\
HR21 & 6870 & 245 & 0.03 & 0.32 & 1.99 & 0.55 & 0.04 & -0.11 & 0.20 & 0.22 & 0.82\
HR21 & 6870 & 179 & 0.01 & 0.25 & 2.21 & 0.63 & 0.13 & -0.14 & 0.46 & 0.24 & 0.69\
HR5017 & 6700 & 174 & 0.04 & 0.28 & 2.58 & 0.83 & 0.07 & -0.03 & 0.43 & 0.36 & 0.96\
HR2706 & 6470 & 086 & 0.01 & 0.39 & 2.08 & 0.47 & 0.16 & -0.10 & 0.13 & 0.20 & 0.64\
HR8905 & 6270 & 059 & 0.06 & 0.13 & 1.06 & 0.45 & -0.01 & -0.16 & 0.30 & 0.28 & 1.02\
HR4883 & 5910 & 280 & 0.15 & 0.20 & 1.07 & 0.52 & 0.09 & 0.04 & 0.90 & 0.43 & 1.02\
HR4716 & 5050 & 133 & 0.17 & 0.10 & 0.97 & 0.47 & 0.13 & 0.07 & 1.29 & 0.54 & 0.92\
HR7328 & 4885 & 281 & 0.19 & 0.20 & 0.78 & 0.57 & 0.11 & 0.11 & 1.33 & 0.67 & 1.35\
HR7949 & 4810 & 369 & 0.21 & 0.24 & 0.77 & 0.59 & 0.13 & 0.18 & 1.39 & 0.67 & 1.38\
HR8317 & 4710 & 191 & 0.29 & 0.14 & 0.63 & 0.58 & 0.08 & 0.26 & 1.79 & 0.83 & 1.64\
HR6299 & 4500 & 532 & 0.26 & 0.09 & 0.77 & 0.81 & 0.10 & 0.44 & 1.92 & 0.76 & 1.40\
HR165 & 4320 & 266 & 0.34 & 0.24 & 0.64 & 0.58 & 0.18 & 0.41 & 1.98 & 0.89 & 1.45\
HR6705 & 3990 & 298 & 0.39 & 0.60 & 0.77 & 1.12 & 0.26 & 0.96 & 3.05 & 1.12 & 1.75\
HR6705 & 3990 & 694 & 0.41 & 0.61 & 0.76 & 1.07 & 0.25 & 0.89 & 2.87 & 1.06 & 1.66\
HR152 & 3956 & 270 & 0.37 & 0.86 & 0.58 & 0.92 & 0.34 & 0.87 & 2.59 & 0.78 & 1.40\
HR4517 & 3780 & 673 & 0.66 & 1.06 & 0.33 & 1.07 & 0.21 & 0.89 & 3.14 & 0.70 & 1.93\
HR6242 & 3560 & 458 & 0.52 & 1.12 & 0.57 & 1.18 & 0.39 & 1.45 & 3.93 & 1.16 & 1.96\
HR7886 & 3250 & 574 & 0.61 & 1.54 & 0.89 & 1.71 & 0.64 & 2.25 & 5.69 & 1.41 & 2.42\
\[ftshewgn\]
[llllllllllll]{} & & & & & & & & & & &\
HR6588 & 19000 & 162 & -0.03 & 0.10 & 1.65 & 0.48 & -0.03 & 0.02 & 0.17 & 0.07 & 0.40\
HR4033 & 8810 & 146 & -0.00 & 0.33 & 2.62 & 0.73 & 0.06 & 0.02 & 0.27 & 0.13 & 0.35\
HR1351 & 7020 & 094 & -0.06 & 0.40 & 2.18 & 0.62 & 0.19 & -0.19 & -0.01 & 0.10 & 0.24\
HR5235 & 5930 & 341 & 0.16 & 0.12 & 1.21 & 0.83 & 0.11 & 0.03 & 1.30 & 0.64 & 1.14\
HR5235 & 5930 & 263 & 0.17 & 0.26 & 1.26 & 0.68 & 0.18 & -0.01 & 0.97 & 0.50 & 1.03\
HR5409 & 5830 & 158 & 0.16 & 0.16 & 1.02 & 0.51 & 0.12 & -0.03 & 0.91 & 0.44 & 0.95\
HR6623 & 5680 & 211 & 0.33 & 0.12 & 0.87 & 0.87 & 0.09 & 0.09 & 1.21 & 0.69 & 1.38\
HR995 & 5620 & 143 & 0.22 & 0.08 & 0.80 & 0.61 & 0.23 & 0.03 & 1.09 & 0.48 & 1.14\
HR7602 & 5430 & 056 & 0.46 & 0.15 & 0.71 & 0.86 & 0.01 & 0.17 & 1.40 & 0.57 & 1.49\
HR7957 & 5240 & 189 & 0.22 & 0.09 & 0.68 & 0.70 & 0.07 & 0.15 & 1.04 & 0.46 & 0.99\
HR7957 & 5240 & 192 & 0.09 & 0.18 & 0.56 & 0.52 & -0.01 & 0.09 & 0.75 & 0.40 & 1.17\
HR5901 & 5125 & 395 & 0.37 & 0.13 & 0.55 & 0.62 & 0.08 & 0.21 & 1.61 & 0.74 & 1.60\
HR5901 & 5125 & 153 & 0.37 & 0.09 & 0.75 & 0.93 & 0.15 & 0.30 & 1.54 & 0.71 & 1.13\
HR6014 & 5068 & 072 & 0.31 & -0.12 & 0.39 & 0.59 & 0.14 & 0.05 & 1.10 & 0.70 & 1.40\
HR6014 & 5068 & 133 & 0.42 & 0.11 & 0.58 & 0.55 & 0.18 & 0.09 & 1.48 & 0.65 & 1.24\
\[ftshewsub\]
[llllllllllll]{} & & & & & & & & & & &\
HR2456 & 38000 & 053 & -0.03 & 0.36 & 0.39 & 0.26 & 0.16 & 0.12 & -0.32 & 0.11 & 0.31\
HR2456 & 38000 & 073 & -0.12 & 0.18 & 0.48 & 0.10 & 0.06 & -0.15 & 0.15 & 0.03 & 0.10\
HR5191 & 19000 & 282 & -0.05 & 0.19 & 1.71 & 0.57 & 0.08 & -0.07 & 0.22 & 0.07 & 0.24\
HR3982 & 13000 & 241 & -0.08 & 0.13 & 2.22 & 0.47 & 0.04 & -0.09 & 0.28 & 0.03 & 0.15\
HR3982 & 13000 & 241 & -0.07 & 0.09 & 2.14 & 0.51 & 0.06 & -0.09 & 0.32 & 0.12 & 0.26\
HR7001 & 9480 & 678 & -0.05 & 0.36 & 2.83 & 1.06 & 0.03 & -0.03 & 0.41 & 0.15 & 0.52\
HR7001 & 9480 & 146 & -0.10 & 0.38 & 2.81 & 1.02 & 0.10 & -0.05 & 0.43 & 0.17 & 0.41\
HR2491 & 9145 & 111 & -0.07 & 0.36 & 2.71 & 0.83 & 0.16 & -0.14 & 0.37 & -0.08 & 0.10\
HR2491 & 9145 & 148 & -0.14 & 0.40 & 2.67 & 0.93 & 0.12 & -0.13 & 0.39 & 0.05 & 0.20\
HR4534 & 8593 & 205 & -0.04 & 0.45 & 2.67 & 1.01 & 0.11 & -0.12 & 0.30 & 0.07 & 0.34\
HR4357 & 8377 & 192 & -0.05 & 0.40 & 2.83 & 0.96 & 0.09 & -0.12 & 0.24 & 0.06 & 0.38\
HR4931 & 6750 & 119 & 0.04 & 0.24 & 1.47 & 0.62 & -0.01 & -0.06 & 0.09 & 0.09 & 0.86\
HR1279 & 6677 & 080 & 0.10 & 0.32 & 1.62 & 0.67 & 0.15 & -0.09 & 0.26 & 0.12 & 0.64\
HR2943 & 6530 & 327 & 0.04 & 0.33 & 1.60 & 0.57 & 0.15 & -0.14 & 0.54 & 0.27 & 0.74\
HR1538 & 6385 & 281 & 0.06 & 0.34 & 1.26 & 0.53 & 0.18 & -0.11 & 0.51 & 0.22 & 0.90\
HR4375 & 6085 & 254 & 0.26 & 0.15 & 0.63 & 0.56 & 0.15 & -0.06 & 0.68 & 0.37 & 1.09\
HR4983 & 5930 & 123 & 0.21 & 0.16 & 0.75 & 0.66 & 0.09 & 0.04 & 0.71 & 0.45 & 1.20\
HR4983 & 5930 & 173 & 0.19 & 0.21 & 0.90 & 0.52 & 0.19 & -0.06 & 0.87 & 0.47 & 0.96\
HR483 & 5855 & 059 & 0.24 & 0.02 & 0.68 & 0.29 & 0.04 & -0.17 & 0.18 & 0.37 & 1.09\
HR483 & 5855 & 079 & 0.24 & 0.19 & 0.90 & 0.46 & 0.19 & -0.11 & 0.67 & 0.48 & 1.30\
HR4374 & 5830 & 250 & 0.28 & 0.20 & 0.62 & 0.63 & 0.10 & -0.01 & 0.90 & 0.47 & 1.35\
HR5072 & 5740 & 197 & 0.26 & 0.15 & 0.74 & 0.53 & 0.16 & -0.04 & 0.99 & 0.53 & 1.28\
HR4496 & 5430 & 105 & 0.35 & 0.17 & 0.59 & 0.59 & 0.21 & -0.08 & 1.12 & 0.61 & 1.45\
HR4496 & 5430 & 090 & 0.31 & 0.16 & 0.39 & 0.50 & 0.13 & -0.04 & 0.69 & 0.43 & 1.08\
HR7462 & 5240 & 131 & 0.52 & 0.04 & 0.38 & 0.72 & 0.09 & 0.09 & 1.20 & 0.62 & 1.89\
HR7462 & 5240 & 095 & 0.51 & 0.07 & 0.37 & 0.69 & 0.13 & 0.06 & 1.16 & 0.72 & 2.05\
HR1084 & 5010 & 223 & 0.61 & 0.19 & 0.44 & 0.74 & 0.24 & -0.01 & 1.59 & 0.76 & 2.00\
HR8832 & 4785 & 111 & 0.88 & 0.14 & 0.29 & 1.06 & 0.17 & 0.18 & 1.88 & 0.90 & 2.50\
GL570A & 4560 & 120 & 0.90 & 0.22 & 0.37 & 1.13 & 0.28 & 0.04 & 2.02 & 0.90 & 2.29\
HR8085 & 4340 & 125 & 0.65 & 0.18 & 0.01 & 1.15 & 0.11 & 0.15 & 1.59 & 0.60 & 2.38\
HR8086 & 4040 & 170 & 0.72 & 0.39 & 0.06 & 1.46 & 0.14 & 0.11 & 1.78 & 0.50 & 2.07\
HR8086 & 4040 & 181 & 0.72 & 0.32 & 0.07 & 1.48 & 0.15 & 0.12 & 1.86 & 0.52 & 2.18\
GL338A & 3800 & 072 & 0.98 & 0.37 & 0.08 & 1.36 & 0.13 & 0.02 & 1.51 & 0.34 & 1.80\
GL526 & 3605 & 068 & 0.62 & 0.42 & -0.13 & 1.32 & 0.16 & 0.10 & 1.04 & 0.04 & 0.92\
GL411 & 3530 & 202 & 0.48 & 0.53 & -0.04 & 1.29 & 0.16 & 0.01 & 1.17 & 0.10 & 1.19\
GL411 & 3530 & 189 & 0.47 & 0.55 & -0.07 & 1.29 & 0.18 & 0.06 & 1.21 & 0.12 & 1.23\
GL725A & 3380 & 083 & 0.52 & 0.57 & -0.10 & 1.15 & 0.06 & 0.03 & 0.93 & -0.00 & 0.99\
GL725A & 3380 & 107 & 0.54 & 0.56 & -0.12 & 1.25 & 0.02 & 0.08 & 1.20 & -0.02 & 1.09\
\[ftshewdw\]
[^1]: For a detailed listing of spectral types and luminosity classes in the revised MK system see Keenan (1985).
[^2]: Recent work by Bessell, Castelli, and Plez (1998) provides updated temperatures, colors, and bolometric corrections for a wide range of spectral types and luminosity classes.
[^3]: For details concerning the advantages and disadvantages of fourier transform spectroscopy, see Bell (1974).
[^4]: IRAF is distributed by the National Optical Astronomy Observatories, which is operated by the Association of Universities for Research in Astronomy, Inc., under contract to the National Science Foundation.
[^5]: The conversion to angstroms is $EW(\AA) = [EW(cm^{-1})/ \sigma^2] \times 10^8$
[^6]: See Herbst (1994) for a comprehensive discussion of OH airglow background supression strategies.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this article we introduce the notion of projective isomonodromy, which is a special type of monodromy evolving deformation of linear differential equations, based on the example of the Darboux-Halphen equation. We give an algebraic condition for a paramaterized linear differential equation to be projectively isomonodromic, in terms of the derived group of its parameterized Picard-Vessiot group.'
address:
- 'Institut de Recherche Mathématique Avancée, Université de Strasbourg et CNRS, 7 rue René Descartes, 67084 Strasbourg Cedex, France'
- 'Department of Mathematics, North Carolina State University, Box 8205, Raleigh, North Carolina 27695-8205'
author:
- Claude Mitschi
- 'Michael F. Singer'
title: Projective Isomonodromy and Galois Groups
---
[^1]
[^2]
Introduction
============
Classical, monodromy preserving deformations of Fuchsian systems have been investigated by many authors who described them in terms of the Schlesinger equation and its links to Painlevé equations. In [@Landesman], Landesmann developed a new Galois theory for parameterized differential equations. A special case was developed in [@CaSi] where the authors consider paramterized [*linear*]{} differential equations and discuss various properties of the parameterized Picard-Vessiot group, the PPV-groups for short. This is a linear differential algebraic group in the sense of Cassidy [@cassidy1]. As is well known, the differential Galois group of a system [ with regular singularities]{} is, as a linear algebraic group, Zariski topologically generated by the monodromy matrices with respect to a fundamental solution. Cassidy and Singer have shown that a parameterized family of such systems is isomonodromic if and only if its PPV-group is conjugate to a (constant) linear algebraic group.
Analogous to the Schlesinger and Painlevé equations’ relation to isomonodromic deformations of Fuchsian systems, the Darboux-Halphen V equation accounts for a special type of monodromy evolving deformation of Fuchsian systems, as was shown by Chakravarty and Ablowitz in [@ChAb]. In this article we first describe the Darboux-Halphen system, then define the general notion of projective isomonodromy illustrated by this example. We characterize projective isomonodromy in different ways, by a condition on the residue matrices for families of Fuchsian systems, and by the condition that the derived group $(G,G)$ of the PPV-group $G$ be conjugate to a constant linear algebraic group when the given equation is absolutely irreducible.
We wish to thank Stephane Malek for making us aware of [@ChAb].
Classical isomonodromy
======================
In the classical study of isomonodromic deformations, only parameterized systems are considered. Furthermore, these systems are assumed to be parameterized in a very special way, that is, the systems are written as
$$\label{isocl}
\frac{dY}{dx}=\sum_{i=1}^{m}\frac{A_i(a)}{x-a_i}, \ \ \ \sum_{i=1}^{m}{A_i(a)}=0$$
where the $n\times n$ matrices $A_i(a)$ depend holomorphically on the multi-parameter $a=(a_1,\ldots,a_n)$ in some open polydisk $D(a^0)$ and the condition on the residue matrices guarantees, for simplicity, that $\infty$ is not singular.
The polydisk $D(a^0)=
D_1\times\ldots\times D_m$ has center at the initial location $a^0=(a^0_1,\ldots,a^0_m)\in \CX^m$ of the poles, with $D_i\subset \CX$ a disk with center $a^0_i$ and $D_i\cap D_j\ne \emptyset$ for all $i\ne j$. Let $x_0\in\calD=\pp\setminus\bigcup_i D_i$.
For fixed $a\in D(a^0)$ and local fundamental solution $Y_a$ of (\[isocl\]) at $a$, analytic continuation along a loop $\gamma$ in $\calD_a=\pp\setminus\{a_1,\ldots,a_m \}$ yields a solution $Y_a^{\gamma}$. The [*monodromy representation*]{} with respect to $Y_a$ is $$\begin{aligned}
\label{chi}
\chi_a : \pi_1(\calD_a;x_0)\rightarrow \GL_n(\CX)\end{aligned}$$ defined by $$Y_a^{\gamma}=Y_a.\chi_a(\gamma),$$ for all $[\gamma]\in \pi_1(\calD_a;x_0)$.
Equation is [*isomonodromic*]{}, or an [*isomonodromic deformation*]{}, if for all $a\in D(a^0)$ there are matrices $C(a)\in \GL_n(\CX)$ such that $$\chi_a=C(a)\ \chi_{a^0}\ C(a)^{-1}.$$
Bolibrukh ([@Bol_iso_def], [@Bol_cfl]) has characterized isomonodromic deformations as follows.
\[bol\] Equation is isomonodromic if and only if the following equivalent conditions hold
1. There is a differential $1$-form $\omega$ on $\pp\times {D(a^0)\setminus \bigcup_{i=1}^m
\{(x,a)\ | \ x-a_i=0 \}}$ such that
- for each fixed $a\in D(a^0),$ $$\omega=\sum_{i=1}^{m}\frac{A_i(a)}{x-a_i} dx$$
- $d\omega=\omega\wedge\omega.$
2. For each $a\in D(a^0)$ there is a fundamental solution $Y_a$ of (\[isocl\]) such that $Y_a(x)$ is analytic in $x$ and $a$, and the corresponding monodromy representation $\chi_a$ does not depend on $a$, that is, $\chi_a=\chi_{a^0}$.
A special type of isomonodromic deformation is given by the Schlesinger differential form $$\omega_s=\sum_{i=1}^m \frac{A_i(a)}{x-a_i}d(x-a_i)$$ whose integrability condition is known as the [*Schlesinger equation*]{} $$\label{schl} dA_i(a)=-\sum_{j=1, j\ne i}^m \frac{[A_i(a),A_j(a)]}{a_i-a_j}d(a_i-a_j), \ \ i=1,\ldots m.$$
Bolibrukh gave examples [@Bol_iso_def] of isomonodromic deformations that are not of the Schlesinger type and he described the general differential forms that occur in Theorem .
In the special case of order two Fuchsian systems with four singularities one can, generically, reduce each system to an order two linear scalar differential equation satisfied by the first component of the dependent variable $Y$, namely a Fuchsian scalar equation with an additional apparent singularity $\lambda$. It is well known that the Schlesinger isomonodromy condition then translates into a non-linear equation of Painlevé VI type satisfied by $\lambda$. For basic results about Painlevé equations and isomonodromic deformations, we refer to [@japanese] and [@AbCla] .
An example of a monodromy evolving deformation
==============================================
In [@ChAb], Chakravarty and Ablowitz describe the Darboux-Halphen system $$\begin{aligned}
\label{DaH}
\left\{ \begin{array}{cccccccc}
\omega_1' &= &&\omega_2\omega_3 &-&\omega_1(\omega_2+\omega_3)&+& \phi^2 \\
\omega_2' &=&& \omega_3\omega_1 &-&\omega_2(\omega_3+\omega_1)&+ &\theta^2 \\
\omega_3' &=&& \omega_1\omega_2 &-&\omega_3(\omega_1+\omega_2)&- &\theta\phi \\
\phi' &=&&\omega_1(\theta - \phi) &-&\omega_3(\theta + \phi) &&\\
\theta' &=& - &\omega_2(\theta- \phi) &-&\omega_3(\theta + \phi) &&\\
\end{array}
\right. \end{aligned}$$ as a prototype of a class of non-linear systems arising as the integrability conditions of an associated Lax pair in the same way as the Painlevé and Schlesinger equations do. This system occurs in the Bianchi IX cosmological models and arises from a special reduction of the self-dual Yang-Mills (SDYM) equation ([*cf*]{}. [@AbCla], [@ChAb], [@ChAb2]). It is also related to the Chazy and Painlevé VI equations (see [@AbCla] for a complete study of such equations and reductions of the SDYM equation). We will follow Ohyama [@ohyama] who studied this equation in more details, and refer to as the Darboux-Halphen V equation or DH-V for short.
Originally ([*cf*]{}. [@ChAb]) the DH-V system with the special condition $\phi=\omega_3=0$ arose from a geometrical problem studied by Darboux, who in 1878 obtained it as the integrability condition for the existence in Euclidean space of a one-parameter family of surfaces of second degree orthogonal to two arbitrary given independent families of parallel surfaces. Halphen solved this system in 1881.
Ohyama ([@ohyama], [@ohyama2]) shows how DH-V is, in the generic case, equivalent to Halphen’s second equation H-II $$x_i'=Q(x_i), \ \ i=1,2,3$$ where $$\begin{aligned}
\label{quadr}
Q(x)=x^2+a(x_1-x_2)^2+b(x_2-x_3)^2+c(x_3-x_1)^2\end{aligned}$$ with constants $a,b,c$ such that $a+b=c+b=-{1}/{4}$ (all derivatives are with respect to the complex variable $t$).
As pointed out in [@ohyama], these equations do not satisfy the Painlevé property (for their movable singularities) and may therefore not be expected to be monodromy-preserving conditions. Nevertheless Chakravarty and Ablowitz [@ChAb] showed how these non-linear equations actually express a special type of monodromy evolving deformation, in the same way as the Schlesinger and Painlevé VI equations rule the isomonodromic deformations of the Schlesinger type.
Using the connection relating the self-dual Yang-Mills equation and the conformally self-dual Bianchi equations, Chakravarty and Ablowitz [@ChAb], followed by Ohyama [@ohyama], showed that DH-V, and hence H-II, actually is the compatibility condition of a Lax pair $$\begin{aligned}
\label{eqn1}
\frac{\partial Y}{\partial x} & = & \left( {\frac{\mu}{P}}I+\sum_{i=1}^3{\frac{\lambda_iS}{x-x_i}} \right)Y\end{aligned}$$ $$\begin{aligned}
\label{eqn2}
\frac{\partial Y}{\partial t} & = & \left( \nu I+\sum_{i=1}^3 \lambda_ix_iS\right)Y-Q(x) {\frac{\partial Y}{\partial x} }\end{aligned}$$ of $2\times 2$ matrix equations, where $x_1,x_2,x_3$ depend on $t$ and $P(x)=(x-x_1)(x-x_2)(x-x_3)$, and $S$ is a traceless constant matrix (the diagonal entries are equal to zero), and $\mu$ and the $\lambda_i$’s are constants with $\mu\ne 0$, $\lambda_1+\lambda_2+\lambda_3=0$, and $\nu(x,t)$ satisfies the auxiliary equation $$\begin{aligned}
\label{eqn7a}
\frac{\partial \nu}{\partial x}=-{\frac{x+x_1+x_2+x_3}{P}}\mu.\end{aligned}$$
Under these asumptions, (\[eqn1\]) Êis for fixed $t$ a Fuchsian system with three singular points $x_1,x_2,x_3$, and Equation (\[eqn7a\]) implies that $\nu$ is not a rational function of $x$. Therefore the Lax pair (\[eqn1\]), (\[eqn2\]) does not describe an isomonodromic deformation, since otherwise the coefficients of (\[eqn2\]) would be rational ([*cf.*]{} [@Sibuya], Remark A.5.2.5).
Let us fix $t_0\in \CX$, and small open disjoint disks $D_i$ with center at $x_i(t_0)$, $i=1,2,3$. Let [ $U(t_0)$]{} be a neighborhood of $t_0$ in $\CX$ such that $x_i(t)\in D_i$ for each $i$ and all $t\in U(t_0)$, and let $x_0\in \CX$ be a fixed base-point, $x_0\notin \bigcup_i D_i$.
Let $Y(t,x)$, for $t\in { U(t_0)}$, denote a fundamental solution, in a neighborhood of $x_0$, of the Lax pair (, ). It is therefore analytic in both $t$ and $x$. For fixed $t\in { U(t_0)}$, we can write an analytic continuation of the fundamental solution $Y(t,x)$ to a punctured neighborhood of $x_i$ as $$\begin{aligned}
\label{eqn4}
{Y(t,x)=Y_i(t,x-x_i(t)).(x-x_i(t))^{L_i(t)}}\end{aligned}$$ where $Y_i(t, x-x_i(t))$ is single-valued, and the matrix $L_i(t)$ does not depend on $x$. Note that $Y_i(t,x-x_i(t))$ is analytic in $t$ and $x$ and $L_i(t)$ is analytic in $t$. Indeed, for fixed $t\in { U(t_0)}$, analytic continuation of $Y$ along an elementary loop around $x_i(t)$ yields a fundamental solution ${\tilde Y}(t,x)$ of which is again analytic in both $t$ and $x$, by the theorem about analytic dependence on initial conditions ([*cf.*]{} [@cartan]). The monodromy matrix $M_i(t)$ is therefore analytic in $t$, as well as $L_i=(1/2\pi i)\log M_i(t)$, and hence $Y_i(t,x-x_i(t))=Y(t,x)\ (x-x_i(t))^{-L_i(t)}$ is analytic in $t$ and $x$ in $U(t_0)\times D_i\setminus \{(t,x)\ | \ x-x_i(t)=0 \}$.
\[prop0\] With notation as above, let $M_i(t)$ for any fixed $t\in U(t_0)$ denote the monodromy matrix of (\[eqn1\]) with respect to $Y$, defined by analytic continuation along an elementary loop around $x_i(t)$. Then $$\begin{aligned}
M_i(t)=c_i(t)G_i\end{aligned}$$ where $G_i$ is a constant matrix and $c_i(t)=e^{-2\pi\mu\sqrt{-1}\int_{t_0}^t \alpha_i(t)dt}$ and the $\alpha_i$ are the residues of $$\begin{aligned}
\label{projfact}
{\frac{x+x_1+x_2+x_3}{P}}=\sum_{i=1}^{3}\frac{\alpha_i}{x-x_i}.\end{aligned}$$
Let us show that $$\begin{aligned}
\label{exp}
\frac{d L_i}{d t}=-\alpha_i \mu I
\end{aligned}$$ where $\alpha_i$ is the $x_i$-residue of $(x+x_1+x_2+x_3)/P$, that is, $$\alpha_i=\frac{x_i+\sigma}{\prod_{j\ne i}(x_i-x_j)}$$ with $\sigma=x_1+x_2+x_3$. Note that $Y_i$ is a function of the local coordinate $x-x_i$, hence $${\d Y_i}/{\d x_i}=-{\d Y_i}/{\d x}.$$ We have $$\frac{\d Y}{\d x_i}=- \frac{\d Y_i}{\d x}(x-x_i)^{L_i} + Y_i (x-x_i)^{L_i}\left(\frac{\d L_i}{\d x_i}\log (x-x_i)-\frac{L_i}{x-x_i}\right)$$ $$=\left(-\frac{\d Y_i}{\d x}(x-x_i)^{L_i}-Y_i(x-x_i)^{L_i}\frac{L_i}{x-x_i} \right) + Y_i(x-x_i)^{L_i}\left(\frac{\d L_i}{\d x_i}\log (x-x_i)\right)$$ $$= -\frac{\d Y}{\d x}+Y_i(x-x_i)^{L_i}\left(\frac{\d L_i}{\d x_i}\log (x-x_i)\right),$$ and $$\frac{\d Y}{\d t}=\frac{\d Y}{\d x_i}\frac{\d x_i}{\d t}=Q(x_i)\frac{\d Y}{\d x_i}=-\frac{\d Y}{\d x}Q(x_i)+Y\left( \frac{\d L_i}{\d x_i}\log (x-x_i)\right)Q(x_i).$$ For any fixed $t$, this is also equal to (see Equation of the Lax pair) $$\frac{\d Y}{\d t}=-\frac{\d Y}{\d x}Q(x) +\left(\nu I +\sum_{i=1}^{3}c_ix_iS \right)Y$$ and as $$\frac{d L_i}{d t}=Q(x_i)\frac{d L_i}{d x_i},$$ (we abusively use the same notation for $L_i$ as a function of $x_i$ and $L_i$ as a function of $t$ via $x_i(t)$), comparing the two expressions we get $$-\frac{\d Y}{\d x}Q(x_i)+ Y(\frac{d L_i}{d t}\log (x-x_i))=-\frac{\d Y}{\d x}Q(x) +\left(\nu I +\sum_{i=1}^{3}c_ix_iS \right)Y.$$ From Equation (\[eqn7a\]) we have that $$\nu=\mu\log \prod_{i=1}^3(x-x_i)^{-\alpha_i}+\phi(t)$$ for some function $\phi(t)$, and hence as $x$ tends to $x_i$ (simplifying and then comparing the leading terms on each side) we get that $$\frac{d L_i}{d t}\log (x-x_i)\sim -\alpha_i \mu \log (x-x_i),$$ that is, $$\frac{d L_i}{d t}=-\alpha_i \mu I .$$ The monodromy matrix of (\[eqn1\]) with respect to $x_0$ and $Y$ around $x_i$ is $M_i=e^{2\pi i L_i}$, which in view of (\[exp\]) is of the form $$M_i(t)=c_i(t)G_i$$ where $G_i$ is the initial monodromy matrix around $x_i(t_0)$, and $c_i(t)=e^{-2\pi\mu\sqrt{-1}\int_{t_0}^t \alpha_i(t)dt}$.
This is an example of what we will call projectively isomonodromic deformations and study from an algebraic point of view.
Projective isomonodromy
========================
Let $\calD$ be an open connected subset of $\PX^1(\CX)$, $\calP$ be an open connected subset of $\CX^r$, and $x_0 \in \calD$. Assume that $\pi_1(\calD,x_0)$ is finitely generated by $\gamma_1, \ldots ,\gamma_m$. Let $A(x,\tbar) \in \gl_n(\calO)$, the ring of $n \times n$ matrices whose entries are functions analytic on $\calD \times \calP$. We will consider the behavior of solutions of the differential equation $$\begin{aligned}
\label{eqn5}
\frac{dY}{dx} &= &A(x,\tbar)Y.\end{aligned}$$ In the following we let $\Scal_n$ be the group of nonzero $n\times n$ scalar matrices.
Equation (\[eqn5\]) is [*projectively isomonodromic*]{} if there exist $m$ analytic functions $c_i:\calP \rightarrow \Scal_n(\CX)$ and fixed matrices $G_1, \ldots , G_m \in \GL_n(\CX)$ such that for each $\tbar \in \calP$ there is a local solution $Y_\tbar(x)$ of (\[eqn5\]) at $x_0$ such that analytic continuation of $Y_\tbar(x)$ along $\gamma_i$ yields $Y_\tbar(x)\cdot G_i \cdot c_i(\tbar)$, for each $i$.
Let $\bar{Y}(x, \tbar)$ be any solution of (\[eqn5\]) analytic in $\calD_0\times\calP$, where $\calD_0$ is a neighborhood of $x_0$ in $\calD$ and let $G_i(\tbar)$ denote the monodromy matrix corresponding to analytic continuation of this solution around $\gamma_i$. Note that $G_i(\tbar)$ depends analytically on $\tbar$. If (\[eqn5\]) is projectively isomonodromic then there exists a function $C(\tbar):\calP \rightarrow \GL_n(\CX)$ such that $$G_i(\tbar) = C(\tbar)^{-1} G_i c_i(\tbar)C(\tbar).$$ Since there may be many ways of selecting $C(\tbar)$, this function need not depend analytically on $\tbar$. However, we will show that one can find a function ${C}(\tbar)$ satisfying the above [*and analytic in $\tbar$*]{}. This fact can be deduced easily from the following result of Andrey Bolibruch whose proof is contained in the proof of Proposition 1 of [@Bol_iso_def].
\[prop2\] For each $ i=1, \ldots , m$, let $H_i(\tbar): \calP \rightarrow \GL_n(\CX)$ be analytic on $\calP$ and let $G_i \in \GL_n(\CX)$. Assume that there is a function $C(\tbar):\calP \rightarrow \GL_n(\CX)$ such that $$H_i(\tbar)=C(t)^{-1}G_iC(t)$$ for all $\tbar \in \calP$ and $i=1,\ldots,m$. Then there exists an analytic function $C(t)$ with the same property.
We can now prove the following
\[prop3\] If (\[eqn5\]) is projectively isomonodromic, then there exists a solution $Y(x, \tbar)$ of (\[eqn5\]) analytic in $\calD_0\times\calP$, where $\calD_0$ is a neighborhood of $x_0$ in $\calD$ such that for all $\tbar \in \calP$ the monodromy matrix of $Y(x, \tbar)$ along $\gamma$ $ G_i \cdot c_i(\tbar)$.
Let $\bar{Y}(x, \tbar)$ be any solution of (\[eqn5\]) analytic in $\calD_0\times\calP$, where $\calD_0$ is a neighborhood of $x_0$ in $\calD$ and let $G_i(\tbar)$ denote the monodromy matrix corresponding to analytic continuation of this solution around $\gamma_i$. Since (\[eqn5\]) is projectively isomonodromic, there is a function $C(\tbar):\calP \rightarrow \GL_n(\CX)$ such that $G_i(\tbar)=C(t)^{-1}G_ic_i(t)C(t)$. Applying Proposition \[prop2\] to $H_i(t)=G_i(t)c_i^{-1}(t)$ and $G_i$, we may assume that $C(t)$ is analytic and thus $Y(x,\tbar) = \bar{Y}(x,\tbar)\cdot {C}(\tbar)$ satisfies the conclusion of this Proposition.
Isomonodromy versus projective isomonodromy
===========================================
We now turn to the relation between Fuchsian isomonodromic equations and Fuchsian projectively isomonodromic equations. Consider the equation $$\begin{aligned}
\label{eqn6}
\frac{dY}{dx} & = & \sum_{i=1}^m \frac{A_i(\tbar)}{x-x_i(\tbar)}Y
\end{aligned}$$ together with
1. $\calP$, a simply connected open subset of $\CX^r$ and
2. $\calD$, an open subset in $\pp$ and $x_0 \in \calD$
such that
1. the functions $A_i(\tbar): \calP \rightarrow \gl_n(\CX)$ and the $x_i(\tbar): \calP \rightarrow \CX$ are analytic functions,
2. $\pp\backslash\calD$ is the union of $m$ disjoint closed disks and
3. for $\tbar\in \calP$ we have $x_i(\tbar) \in D_i$.
Let $x_0 \in \calD$ and $\gamma_i$, $i = 1, \ldots , m,$ be the obvious loops generating $\pi_1(\calD, x_0)$. We then have that Equation (\[eqn6\]) is analytic in $\calD \times \calP$ and we can speak of monodromy matrices $G_i(\tbar)$ corresponding to analytic continuation of a fundamental solution matrix along $\gamma_i$. We can now state
\[descr\] Let $\calD$ and $\calP$ be as above. Equation (\[eqn6\]) is projectively isomonodromic if and only if for each $ i=1, \ldots , m$, there exist functions $b_i:\calP \rightarrow \CX$ and $B_i:\calP \rightarrow \gl_n(\CX)$, analytic on $\calP$ such that
1. $A_i(\tbar) = B_i(\tbar) + b_i(\tbar)I_n$ for $i = 1, \ldots , m$ and
2. $$\begin{aligned}
\label{eqn7}
\frac{dY}{dx} & = & \left(\sum_{i=1}^m\frac{B_i(\tbar)}{x-x_i(\tbar)}\right)Y
\end{aligned}$$
is isomonodromic.
Assume that Equation (\[eqn6\]) is projectively isomonodromic and let $Y(x,\tbar), G_i$ and $c_i(\tbar)$ be as in the conclusion of Proposition \[prop3\]. Since $\calP$ is simply connected and the $c_i(\tbar)$ are nonzero, there exist analytic $b_i(\tbar):\calP \rightarrow \CX$ such that $e^{2\pi\sqrt{-1}b_i(\tbar)} = c_i(\tbar)$ for each $i$. Let $$Z(x,\tbar) = Y(x,\tbar)\cdot \prod_{i=1}^m(x-x_i(\tbar))^{-b_i(\tbar)}I_n$$ One sees that the monodromy of $Z$ along $\gamma_i$ is given by $G_i$ and so is independent of $\tbar$. Therefore, letting $B_i(\tbar) = A_i(\tbar) - b_i(\tbar)I_n$, we have that
$$\begin{aligned}
\frac{dY}{dx} & = & \left(\sum_{i=1}^s \frac{B_i(\tbar)}{x-x_i(t)}\right)Y
\end{aligned}$$
is isomonodromic.
Now assume that $A_i,B_i,b_i$ are as in items 1. and 2. of the proposition and that Equation (\[eqn7\]) is isomonodromic. If $Y(x,\tbar)$ is a local solution of (\[eqn7\]) with constant monodromy matrices $G_i$ along $\gamma_i$, then $Z(x,\tbar) = Y(x,\tbar) \cdot\prod_{i=1}^m (x-x_i(\tbar))^{b_i(\tbar)}I_n$ will have monodromy $c_i(\tbar)G_i$ along $\gamma_i$, with $c_i(\tbar) = e^{2\pi\sqrt{-1}b_i(\tbar)}\ $. Thus Equation (\[eqn6\]) is projectively isomonodromic.
Proposition \[descr\] applies to the DH-$\mathrm V$ example of Chakravarty and Ablowitz, since we can rewrite Equation (\[eqn1\]) of the Lax pair as $$\begin{aligned}
\frac{\partial Y}{\partial x} = \left(\sum_{i=1}^3 \frac{A_i(t)}{(x-x_i)}\right) Y\end{aligned}$$ where $$\begin{aligned}
A_i(t) & = & B_i(t) + b_i(t)I_n\end{aligned}$$ with $$\begin{aligned}
B_i(t) & = & \lambda_iS\\
b_i(t) &=& {\frac{\mu }{\prod_{i\neq j}(x_i-x_j)}}.\end{aligned}$$ An easy computation shows that since $x'_i-x'_j=Q(x_i)-Q(x_j)=x_i^2-x_j^2$ for all $i,j$, we have $$b'_i=\frac{db_i}{dt}= -{\frac{x_i+\sigma }{\prod_{i\neq j}(x_i-x_j)}}\mu==-\alpha_i \mu.$$ and we recover the result of Proposition \[prop0\], that the monodromy of this equation is evolving ‘projectively’ and equal to $$M_i(t)=e^{2\pi\sqrt{-1}b_i(\tbar)}G_i= e^{-2\pi\mu\sqrt{-1}\int_{t_0}^t \alpha_i(t)dt}G_i.$$
Parameterized differential Galois groups
========================================
In this section we examine the parameterized differential Galois groups of projectively isomonodromic equations. Parameterized differential Galois groups ([*cf.*]{} [@CaSi], [@Landesman]) generalize the concept of differential Galois groups of the classical Picard-Vessiot theory and we begin this section by briefly describing the underlying theory.
Let $$\begin{aligned}
\label{pveqn}\frac{d Y}{d x} &=& A(x) Y\end{aligned}$$ be a differential equation where $A(x)$ is an $n\times n$ matrix with entries in $\CX(x)$. The usual existence theorems for differential equations imply that if $x = x_0$ is a point in $\CX$ such that the entries of $A(x)$ are analytic at $x_0$, then there exists a nonsingular matrix $Z = (z_{i,j})$ of functions analytic in a neighborhood of $x_0$ such that $\frac{d Z}{d x} = A(x)Z$. Note that the field $K = \CX(z_{1,1}, \ldots , z_{n,n})$ is closed with respect to taking the derivation $\frac{d}{d x}$ and this is an example of a [*Picard-Vessiot extension*]{}[^3]. The set of field-theoretic isomorphisms of $K$ that leave $\CX(x)$ elementwise-fixed and commute with $\frac{d}{d x}$ forms a group $G$ called the [*Picard-Vessiot group*]{} or [*differential Galois group*]{} of (\[pveqn\]). One can show that for any $\sigma \in G$, there exists a matrix $M_\sigma \in \GL_n(\CX)$ such that $\sigma(Z) =(\sigma(z_{i,j})) = ZM_\sigma$. The map $\sigma \mapsto M_\sigma$ is an isomorphism whose image is furthermore a [*linear algebraic group*]{}, that is, a group of invertible matrices whose entries satisfy some fixed set of polynomial equations in $n^2$ variables. There is a well developed Galois theory for these groups that identifies certain subgroups of $G$ with certain subfields of $K$ and associates properties of the equation (\[pveqn\]) with properties of the groups $G$. The elements of the monodromy group of (\[pveqn\]) may be identified with elements of this group and, when (\[pveqn\]) has only regular singular points, it is known that $G$ is the smallest linear algebraic group containing these elements ([*cf.*]{} [@PuSi2003], Theorem 5.8). Further facts about this Galois theory can be found in [@DAAG] and [@PuSi2003].
Now let $$\begin{aligned}
\label{ppveqn}
\frac{d Y}{d x} & = & A(x,\tbar)Y\end{aligned}$$ be a parameterized system of linear differential equations where $A(x,\tbar)$ is an $n\times n$ matrix whose entries are rational functions of $x$ with coefficients that are functions of $\tbar = (t_1, \ldots, t_r)$, analytic in some domain in $\CX^r$. A differential Galois theory for such equations was developed in [@CaSi] and in greater generality in [@Landesman]. Let $k_0$ be a suitably large field[^4] containing $\CX(t_1, \ldots, t_r)$ and the functions of $\tbar$ appearing as coefficients in the entries of $A$ and such that $k_0$ is closed under the derivations $\Pi = \{\d_1, \ldots, \d_r\}$ where each $\d_i$ restricts to $\frac{\d}{\d t_i}$ on $\CX(t_1, \ldots, t_r)$ and the intersection of the kernels of the $\d_i$ is $\CX$. As before, existence theorems for solutions of differential equations guarantee the existence of a nonsingular matrix $Z(x,\tbar) = (z_{i,j}(x,\tbar))$ of functions, analytic in some suitable domain in $\CX \times \CX^r$, such that $\frac{d Z}{d x} = AZ$. We will let $k = k_0(x)$ be the differential field with derivations $\Delta = \{\d_x, \d_1, \ldots , \d_r\}$ where $\d_x(x) = 1, \ \d_x(z) = 0 \mbox{ for all $z \in k_0$ and the } \d_i \mbox{ extend the previous } \d_i \mbox{ with } \d_i(x) = 0.$ Finally we will denote by $K$ the smallest field containing $k$ and the $z_{i,j}$ that is closed under the derivations of $\Delta$. This field is called the [*parameterized Picard-Vessiot field*]{} or PPV-field of (\[ppveqn\]). The set of field theoretic automorphisms of $K$ that leave $k$ elementwise-fixed and commute with the elements of $\Delta$ forms a group $G$ called the [*parameterized Picard-Vessiot group* ]{} (PPV-group) or [*parameterized differential Galois group*]{} of (\[ppveqn\]). One can show that for any $\sigma \in G$, there exists a matrix $M_\sigma \in \GL_n(k_0)$ such that $\sigma(Z) =(\sigma(z_{i,j})) = ZM_\sigma$. Note that $\d_x$ applied to an entry of such an $ M_\sigma$ is $0$ since these entries are elements of $k_0$ but that such an entry need not be constant with respect to the elements of $\Pi$. One may think of these entries as functions of $\tbar$. In [@CaSi], the authors show that the map $\sigma \mapsto M_\sigma$ is an isomorphism whose image is furthermore a [*linear differential algebraic group*]{}, that is, a group of invertible matrices whose entries satisfies some fixed set of polynomial [*differential*]{} equations (with respect to the derivations $\Pi = \{\d_1, \ldots, \d_r\}$) in $n^2$ variables. We say that a set $X \subset \GL_n(k_0)$ is [*Kolchin closed*]{} if it is the zero set of such a set of polynomial differential equations. One can show that the Kolchin closed sets form the closed sets of a topology, called the [*Kolchin topology*]{} on $\GL_n(k_0)$ ([*cf.*]{} [@cassidy1; @cassidy6; @CaSi; @kolchin_groups]). In [@CaSi], the authors showed that for parameterized systems of linear differential equations with regular singular points, the parameterized monodromy is Kolchin dense in the PPV-group.
The following result shows how the PPV-group can be used to characterize isomonodromy. As in Section 4, let $\calP$ be a simply connected subset of $\CX^r$ and $\calD$ an open subset of $\pp$ with $x_0 \in \calD$. We assume that $A(\tbar,x)$ in Equation (\[ppveqn\]) is analytic in $\calD\times \calP$. Assume that $\pp\backslash \calD$ is the union of $m$ disjoint disks $D_i$ and that for each $\tbar \in \calP$, Equation (\[ppveqn\]) has a unique singular point in each $D_i$ and that this singular point is a regular singular point. Let $\gamma_i, i = 1, \ldots , m$ be the obvious loops generating $\pi_1(\calD, x_0)$. We then have that Equation (\[ppveqn\]) is analytic in $\calD \times \calP$ and we can speak of monodromy matrices $G_i(\tbar)$ corresponding to analytic continuation of a fundamental solution matrix along $\gamma_i$.
([[*cf.*]{} [@CaSi], Proposition 5.4]{}) Assume that $\calD$, $\calP$ and Equation (\[ppveqn\]) are as above. Then this equation is isomonodromic in $\calD\times \calP'$ for some subset $\calP' \subset \calP$ if and only if the PPV-group of this equation over $k$ is conjugate to $G(\CX)$ for some linear algebraic group $G$ defined over $\CX$.
An algebraic condition for projective isomonodromy
==================================================
We now relate the property of projective isomonodromy to properties of the PPV-group.
\[prop3.1\]Let $k,K, A,$ and $G$ be as above. Equation (\[ppveqn\]) is projectively isomonodromic if and only if $G$ is conjugate to a subgroup of $ \GL_n(\CX) \cdot \Scal_n(k_0)\subset \GL_n(k_0)$.
As noted above, for parameterized systems of linear differential equations with regular singular points, the parameterized monodromy is Kolchin dense in the PPV-group. The group $ \GL_n(\CX)\cdot \Scal_n(k_0) \subset \GL_n(k_0)$ is the homomorphic image of the linear differential group $\GL_n(\CX)\times \Scal_n(k_0)\subset \GL_n(k_0)\times \GL_n(k_0)$ and so by Proposition 7 of [@cassidy1], it is also Kolchin closed. Therefore, if the monodromy matrices are in $ \GL_n(\CX)\cdot \Scal_n(k_0)$, then $G \subset \Scal_n(k_0)\cdot \GL_n(\CX)$. The converse is clear.
One easy consequence of Propostion \[prop3.1\] is
\[cor3.1\] Let $k,K, A,$ and $G$ be as above. If (\[ppveqn\]) is projectively isomonodromic then $(G,G)$ is conjugate to a subgroup of $\GL_n(\CX)$.
This corollary yields a simple test to show that (\[ppveqn\]) is not projectively isomonodromic: If the eigenvalues of the commutators of the monodromy matrices (with respect to any fundamental solution matrix) are not constant, then (\[ppveqn\]) is not projectively isomonodromic. In particular, if the determinant or trace of any of these matrices is not constant then (\[ppveqn\]) is not projectively isomonodromic. The converse of the corollary is not true in general (see Remark \[remark3.1\] below) but it is true if Equation (\[ppveqn\]) is absolutely irreducible, that is, when (\[ppveqn\]) does not factor over $\overline{k}$, the algebraic closure of $k$. Before we prove this, we will discuss some group theoretic facts.
In the following, we say that a subgroup $H \subset \GL_n(k_0)$ is [*irreducible*]{} if the only $H$-invariant subspaces of $k_0^n$ are $\{0\}$ and $k_0^n$.
\[lem1\] Let $H$ be an irreducible subgroup of $\GL_n(\CX)$ and let $g\in \GL_n(k_0)$ normalize $H$. Then $g \in \GL_n(\CX)\cdot \Scal_n(k_0)$.
For any $h \in H$ and $g \in \GL_n(k_0)$ normalizing $H$, we have that $$0= \d_i(g^{-1}hg) = -g^{-1}\d_i(g)g^{-1}hg + g^{-1}h\d_i(g)$$ for all $\d_i \in \Pi$. Therefore, $$\d_i(g)g^{-1} h = h \d_i(g)g^{-1}.$$ Since $H$ is irreducible, Schur’s Lemma implies that $\d_i(g)g^{-1}\in \Scal_n(k_0)$. This means that if $g = (g_{r,s})$, then there exists a $z_i\in k_0$ such that $\d_i g_{r,s} = z_i g_{r,s}$ for all $r,s$. One can check that the $z_i$ satisfy the integrability conditions so there exists a nonzero $u \in k_0$ such that $\d_i u = z_i u$ for all $i$. Therefore $g_{r,s} = h_{r,s} u$ for some $h_{r,s} \in \CX$ and so $g = uI_n\cdot h$ for some $h \in \GL_n(\CX)$.
It is well known that if $G$ and $H$ are linear algebraic groups with $H$ normal in $ G$, then $G/H$ is also a linear algebraic group. For $\Scal_n(k_0) \lhd \GL_n(k_0)$, we will denote by $\rho$ the canonical map $\rho: \GL_n(k_0) \rightarrow \GL_n(k_0)/\Scal_n(k_0)$.
\[lem2\] Let $H \subset \GL_n(k_0)$ be a Kolchin connected linear differential algebraic group and let $\Hbar$ be its Zarski closure in $\GL_n(k_0)$. Assume that that $\Hbar$ is irreducible. Then $$H \subset (H,H)_\Pi \cdot \Scal_n(k_0),$$ where $(H,H)_\Pi$ is the Kolchin closure of $(H,H) $.
Since $\Hbar$ is irreducible, it must be reductive ([@springer], p. 37). Since $H$ is Kolchin connected, $\Hbar$ is Zariski connected so we can write $\Hbar = Z(\Hbar)\cdot (\Hbar,\Hbar)$ where $Z(\Hbar)$ is the center of $\Hbar$ ([@humphreys], Ch. 27.5). Using the irreducibility again, Schur’s Lemma implies that $Z(\Hbar) \subset \Scal_n(k_0)$. Using the map $\rho$ above, we have that $\rho(\Hbar)$ is isomorphic to $ (\Hbar,\Hbar)/(Z(\Hbar)\cap (\Hbar,\Hbar))$ and so is a connected semisimple linear algebraic group. Furthermore, $\rho(H)$ is a Zariski dense, Kolchin connected, subgroup of $\rho(\Hbar)$. Propositions 11 and 13 and of [@cassidy6] imply that $\rho(H)$ equals $(\rho(H),\rho(H))_\Pi$, the Kolchin closure of its commutator subgroup $(\rho(H),\rho(H))$. Since $(H,H)_\Pi$ is a linear differential algebraic group, we have that $\rho((H,H)_\Pi)$ is a linear differential algebraic group containing $(\rho(H),\rho(H))$ and therefore contains $(\rho(H), \rho(H))_\Pi$. Since $(H,H)_\Pi \subset H$ we that $\rho((H,H)_\Pi) = \rho(H)$. Therefore $H \subset (H,H)_\Pi \cdot \Scal_n(k_0)$.
\[lem3\] Let $G \subset \GL_n(k_0)$ be a linear differential group and assume that
1. $(G,G) \subset \GL_n(\CX)$ and
2. the identity component $\Gbar^0$ of $\Gbar$, the Zariski closure of $G$ in $\GL_n(k_0)$, is irreducible.
Then $G \subset \GL_n(\CX) \cdot \Scal_n(k_0)$.
We first note that the Zariski closure of $G^0$, the Kolchin component of the identity of $G$ is Zariski connected and of finite index in $\Gbar$. Therefore $\Gbar^0$ is the Zariski closure $\overline{G^0}$ of $G^0$. We now apply Lemma \[lem2\] to $H = G^0$ and conclude that $G^0 \subset (G^0,G^0)_\Pi \cdot \Scal_n(k_0)$. Since $(G,G) \subset \GL_n(\CX)$ we have that $(G^0,G^0)_\Pi \subset \GL_n(\CX)$. Furthermore, since $\Gbar^0$ is irreducible and is the Zariski closure of $G^0$, we have that $G^0$ is irreducible. Therefore $(G^0,G^0)_\Pi$ is an irreducible subgroup of $\GL_n(\CX)$. Any $g \in G$ normalizes $G^0$ and therefore normalizes $(G^0,G^0)_\Pi$. Applying Lemma \[lem1\] to $H = (G^0,G^0)_\Pi$, we have that $G \subset \GL_n(\CX) \cdot \Scal_n(k_0)$.
\[remark3.1\] [ Simple examples ([*e.g.*]{}, $G = \diag_n(k_0)$, the group of diagonal matrices) show that condition $(G,G) \subset \GL_n(\CX)$ does not imply $G \subset \GL_n(\CX) \cdot \Scal_n(k_0)$ without some additional hypotheses. ]{}
\[prop3.2\] Let $k,K,A,G$ be as in Proposition \[prop3.1\]. If Equation (\[ppveqn\]) is absolutely irreducible and $(G,G)$ is conjugate to a subgroup of $\GL_n(\CX)$, then (\[ppveqn\]) is projectively isomonodromic.
As noted above, $\Gbar$ is the usual Picard-Vessiot group of (\[ppveqn\]) over $k$. If (\[ppveqn\]) is absolutely irreducible, then $\Gbar^0$ is an irreducible subgroup of $\GL_n(k_0)$. Lemma \[lem3\] implies that $G \subset \GL_n(\CX) \cdot \Scal_n(k_0)$
[10]{}
M. J. Ablowitz, P. A. Clarkson Solitons, Nonlinear Evolution Equations and Inverse Scattering , 149, Cambridge University Press, 1991
D. V. Anosov, A. A. Bolibruch The Riemann-Hilbert Problem Aspects of Mathematics, Vieweg, 1994
A. A. Bolibruch On Isomonodromic Deformations of Fuchsian Systems , 3(4):589-604, 1997.
A. A. Bolibruch On Isomonodromic Confluences of of Fuchsian Singularities , vol. 221, 117-132, 1998.
H. Cartan Théorie élémentaire des fonctions analytiques d’une ou plusieurs variables complexes Hermann, Paris, 1961.
P. J. Cassidy Differential algebraic groups , 94:891-954, 1972.
P. J. Cassidy The classification of the semisimple differential algebraic groups and the linear semisimple differential algebraic [L]{}ie algebras , 121(1):169-238, 1989.
P. J. Cassidy, M. F. Singer Galois Theory of Parameterized Differential Equations and Linear Differential Algebraic Groups , D. Bertrand et. al., eds., IRMA Lectures in Mathematics and Theoretical Physics, 9:113-157,2006.
S. Chakravarty, M. J. Ablowitz Integrability, monodromy evolving deformations, and self-dual Bianchi IX systems , 76(6):857-860,1996.
S. Chakravarty, M. J. Ablowitz, L.A. Takhtajan, in ed. by M. Boiti, L. Martina and F. Pempinelli (World Scientific, Singapore, 1992.
J. Humphreys Linear Algebraic Groups , Springer-Verlag, New York, 1975
K. Iwasaki, H. Kimura, S. Shimomura, M. Yoshida From Gauss to Painlevé A Modern Theory of Special Functions , Vieweg, 1991
E. R. Kolchin Differential Algebra and Algebraic Groups Academic Press, New York, 1976
E. R. Kolchin Differential algebraic groups Academic Press, New York, 1985
P. Landesman Generalized differential Galois theory . 360(8):4441–4495, 2008.
Y. Ohyama Quadratic equations and monodromy evolving deformations :0709.4587v1 \[math.CA\] 28 Sep 2007.
Y. Ohyama Monodromy evolving deformations and Halphen’s equation in Groups and Symmetries, [*CRM Proc. Lecture Notes*]{}, 47, Amer.Math.Soc., Providence, RI, 2009.
M. van der Put and M. F. Singer Galois Theory of Linear Differential Equations , 328, Springer-Verlag, 2003
Y. Sibuya Linear Differential Equations in the Complex Domain: Problems of Analytic Continuation American Mathematical Society, 1990
T. A. Springer Linear Algebraic Groups, Second Edition , 9, Birkhäuser, Boston, 1998
[^1]:
[^2]: The second author was partially supported by NSF Grant CCR-0634123. He would also like to thank the Institut de Recherche Mathématique Avancée, Université de Strasbourg et C.N.R.S., for its hospitality and support during the preparation of this paper.
[^3]: Picard-Vessiot extensions and the related Picard-Vessiot theory is developed in a fuller generality in [@DAAG] and [@PuSi2003] but we shall restrict ourselves to the above context to be concrete.
[^4]: To be precise, we need $k_0$ to be [*differentially closed*]{} with respect to $\Pi$, that is, any system of polynomial differential equations in arbitrary unknowns having a solution in an extension field already has a solution in $k_0$. See [@CaSi] for a discussion of differentially closed fields in the context of this Galois theory.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Motivated by quite recent research involving the relationship between the dimension of a poset and graph theoretic properties of its cover graph, we show that for every $d\ge 1$, if $P$ is a poset and the dimension of a subposet $B$ of $P$ is at most $d$ whenever the cover graph of $B$ is a block of the cover graph of $P$, then the dimension of $P$ is at most $d+2$. We also construct examples which show that this inequality is best possible. We consider the proof of the upper bound to be fairly elegant and relatively compact. However, we know of no simple proof for the lower bound, and our argument requires a powerful tool known as the Product Ramsey Theorem. As a consequence, our constructions involve posets of enormous size.'
address:
- 'School of Mathematics, Georgia Institute of Technology, Atlanta, GA, USA'
- 'Theoretical Computer Science Department, Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland; School of Mathematics, Georgia Institute of Technology, Atlanta, GA, USA'
- 'School of Mathematics, Georgia Institute of Technology, Atlanta, GA, USA'
author:
- 'William T. Trotter'
- Bartosz Walczak
- Ruidong Wang
title: |
Dimension and Cut Vertices:\
An Application of Ramsey Theory
---
Introduction
============
We assume that the reader is familiar with basic notation and terminology for partially ordered sets (here we use the short term *posets*), including chains and antichains, minimal and maximal elements, linear extensions, order diagrams and cover graphs. Extensive background information on the combinatorics of posets can be found in [@bib:Trot-Book; @bib:Trot-Handbook].
We will also assume that the reader is familiar with basic concepts of graph theory, including the following terms: connected and disconnected graphs; components; cut vertices; and $k$-connected graphs for an integer $k\ge2$. Recall that when $G$ is a connected graph, a connected induced subgraph $H$ of $G$ is called a *block* of $G$ when $H$ is $2$-connected and there is no subgraph $H'$ of $G$ which contains $H$ as a proper subgraph and is also $2$-connected.
Here are the analogous concepts for posets. A poset $P$ is said to be *connected* if its cover graph is connected. A subposet $B$ of $P$ is said to be *convex* if $y\in B$ whenever $x,z\in B$ and $x<y<z$ in $P$. Note that when $B$ is a convex subposet of $P$, the cover graph of $B$ is an induced subgraph of the cover graph of $P$. A convex subposet $B$ of $P$ is called a *component* of $P$ when the cover graph of $B$ is a component of the cover graph of $P$. A convex subposet $B$ of $P$ will be called a *block* of $P$, when the cover graph of $B$ is a block in the cover graph of $P$.
Motivated by questions raised in recent papers exploring connections between the dimension of a poset $P$ and graph theoretic properties of the cover graph of $P$, our main theorem will be the following result.
\[thm:main\] For every $d\ge1$, if $P$ is a poset and every block in $P$ has dimension at most $d$, the the dimension of $P$ is at most $d+2$. Furthermore, this inequality is best possible.
The remainder of this paper is organized as follows. In the next section, we present a brief discussion of background material which serves to motivate this line of research and puts our theorem in historical perspective. Section \[sec:dimension\] includes a compact summary of essential material from dimension theory. Section \[sec:upper-bound\] contains the proof of the upper bound in our main theorem, and in Section \[sec:best-possible\], we give a construction which shows that our upper bound is best possible. This construction uses the Product Ramsey Theorem and produces posets of enormous size. We close in Section \[sec:closing\] with some brief remarks about challenges that remain.
Background Motivation {#sec:background}
=====================
A family ${\mathcal{F}}=\{L_1,L_2,\dots,L_d\}$ of linear extensions of a poset $P$ is called a *realizer* of $P$ when $x\le y$ in $P$ if and only if $x\le y$ in $L_i$ for each $i=1,2,\dots,d$. The *dimension* of $P$, denoted $\dim(P)$, is the least positive integer $d$ for which $P$ has a realizer of size $d$. For simplifying the details of arguments to follow, we consider families of linear extensions with repetition allowed. So if $\dim(P)=d'$, then $P$ has a realizer of size $d$ for every $d\ge d'$. For an integer $d\ge2$, a poset $P$ is said to be *$d$-irreducible* if $\dim(P)=d$ and $\dim(B)<d$ for every proper subposet $B$ of $P$.
As is well known, the dimension of a poset $P$ is just the maximum of the dimension of the components of $P$ *except* when $P$ is the disjoint sum of two or more chains. In this case, $\dim(P)=2$ while all components of $P$ have dimension $1$. Accordingly, when $d\ge3$, a poset $P$ with $\dim(P)=d$ has a component $Q$ with $\dim(Q)=d$.
It is easy to see that if the chromatic number of a connected graph $G$ is $r$ and $r\ge2$, then there is a block $H$ of $G$ so that that chromatic number of $H$ is also $r$. The analogous statement for posets is not true. We show in Figure \[fig:cut-vertices\] representatives of two infinite families of posets. When $n\ge 2$, the poset $P_n$ shown on the left side is $3$-irreducible (see [@bib:TroMoo-1] or [@bib:Kell] for the full list of all $3$-irreducible posets). For each $m\ge3$, the poset $Q_m$ shown on the right is $m+1$-irreducible. This second example is part of an exercise given on page $20$ in [@bib:Trot-Book]. Together, these examples show that for every $d\ge2$, there are posets of dimension $d+1$ where every block has dimension at most $d$.
=\[circle,minimum size=4pt,inner sep=0pt,draw\] =\[rectangle,label distance=3pt,draw=none\] (a) at (0,0) ; (b1) at (-2,1) ; (b2) at (-1,1) ; (b3) at (0,1) ; (b4) at (1,1) ; at (0.5,1) [$\vphantom{1}\cdots$]{}; (b5) at (2,1) ; (e) at (3,1.2) ; (c1) at (-2.5,2) ; (c2) at (-1.5,2) ; (c3) at (-0.5,2) ; (c4) at (0.5,2) ; (c5) at (1.5,2) ; (d) at (0,3) ; (a) edge (b1) edge (b2) edge (b3) edge (b4) edge\[bend right=10\] (e); (d) edge (c2) edge (c3) edge (c4) edge (c5) edge\[bend left=10\] (e); (b1) edge (c1) edge (c2); (b2) edge (c2) edge (c3); (b3) edge (c3) edge (c4); (b4) edge (c4) edge (c5); (b5) edge (c5); at (0,-0.7) [$P_n$; $n\geq 2$]{};
=\[circle,minimum size=4pt,inner sep=0pt,draw\] =\[rectangle,label distance=3pt,draw=none\] (a) at (-0.5,0) ; (b1) at (-2,1) ; (b2) at (-1,1) ; (b3) at (0,1) ; (b4) at (1,1) ; (b5) at (2,1) ; (c1) at (-2,2) ; (c2) at (-1,2) ; (c3) at (0,2) ; (c4) at (1,2) ; (c5) at (2,2) ; (d1) at (-2,3) ; (d2) at (-1,3) ; (d3) at (0,3) ; at (0.5,3) ; (d4) at (1,3) ; (a) edge (b1) edge (b2) edge (b3) edge (b4); (c1) edge (b1) edge (d2) edge (d3) edge (d4); (c2) edge (d1) edge (b2) edge (d3) edge (d4); (c3) edge (d1) edge (d2) edge (b3) edge (d4); (c4) edge (d1) edge (d2) edge (d3) edge (b4); (c5) edge (b1) edge (b2) edge (b3) edge (b4) edge (b5); at (-0.25,-0.7) [$R_n$; $n\geq 3$]{};
We call a poset $P$ a *tree* when the cover graph of $P$ is a tree. The following theorem is proved in [@bib:TroMoo-2].
\[thm:tree\] If $P$ is a poset and the cover graph of $P$ is a tree, then $\dim(P)\le 3$.
In Figure \[fig:3dim-trees\], we show two posets whose cover graphs are trees. These examples appear in [@bib:TroMoo-2], and we leave it as an exercise to verify that each of them has dimension $3$. Accordingly, the inequality in Theorem \[thm:tree\] is best possible. In the language of this paper, we note that when the cover graph of $P$ is a tree and $|P|\ge2$, then every block of $P$ is a $2$-element chain and has dimension $1$. Accordingly, in the case $d=1$, our main theorem reduces to a result which has been known for nearly $40$ years. However, we emphasize that the proof we give in Section \[sec:upper-bound\] of the upper bound in Theorem \[thm:main\] is not inductive and works for all $d\ge1$ simultaneously. For this reason, it provides a new proof of Theorem \[thm:tree\] as a special case.
=\[circle,minimum size=4pt,inner sep=0pt,draw\] (a) at (0,0) ; (b1) at (-2,1) ; (b2) at (-1,1) ; (b3) at (0,1) ; (b4) at (1,1) ; (c1) at (-1.5,2) ; (c2) at (-0.5,2) ; (a) edge (b2) edge (b4); (c1) edge (b1) edge (b2); (c2) edge (b2) edge (b3); (x1) at (3,0.5) ; (x2) at (4,0.5) ; (x3) at (5,0.5) ; (x4) at (6,0.5) ; (y1) at (3,1.5) ; (y2) at (4,1.5) ; (y3) at (5,1.5) ; (y1) edge (x1) edge (x2); (y2) edge (x1) edge (x3); (y3) edge (x1) edge (x4);
A second paper in which trees and cut vertices are discussed is [@bib:Trot-3], but the results of this paper are considerably stronger. Here is a more recent result [@bib:FeTrWi], and only recently has the connection with blocks and cut vertices become clear.
\[thm:outerplanar\] If $P$ is a poset and the cover graph of $P$ is outerplanar, then $\dim(P)\le 4$. Furthermore, this inequality is best possible.
In [@bib:FeTrWi], the poset shown in Figure \[fig:diamonds\] is given, and it is shown that when $n\ge17$, the poset has dimension $4$. Note that the cover graph of this poset is outerplanar. As a consequence, the inequality in Theorem \[thm:outerplanar\] is best possible. Moreover, every block of $P_n$ is a four element subposet having dimension $2$ (these subposets are called “diamonds”). We may then conclude that when $d=2$, the inequality in our main theorem is best possible. However, the construction we present in Section \[sec:best-possible\] to show that our upper bound is best possible will again handle all values of $d$ with $d\ge2$ at the same time, so it will not use this result either.
=\[circle,minimum size=4pt,inner sep=0pt,draw\] =\[rectangle,label distance=3pt,draw=none\] (x) at (-2.5,0) ; (a1) at (0,-1) ; (a2) at (0,-2.5) ; (a3) at (0,-4) ; (a4) at (0,-5.5) ; (b1) at (1,0) ; (b2) at (2.5,0) ; (b3) at (4,0) ; (b4) at (5.5,0) ; (c1) at (0,1) ; (c2) at (0,2.5) ; (c3) at (0,4) ; (c4) at (0,5.5) ; (x) edge (a1) edge (a2) edge (a3) edge (a4) edge (c1) edge (c2) edge (c3) edge (c4); (b1) edge (a1) edge (c1); (b2) edge (a2) edge (c2); (b3) edge (a3) edge (c3); (b4) edge (a4) edge (c4);
The results of this paper are part of a more comprehensive series of papers exploring connections between dimension of a poset and the graph theoretic properties of its cover graphs. Recent related papers include [@bib:BiKeYo; @bib:JMMTWW; @bib:JMTWW; @bib:MicWie; @bib:StrTro; @bib:TroWan; @bib:Walc]. However, many of these modern research themes have their roots in results, such as Theorem \[thm:tree\], obtained in the 1970s or even earlier.
Here is one such example and again, Theorem \[thm:tree\] was the starting point. The result is given in [@bib:JMMTWW].
\[thm:tree-width\] For any positive integers $t$ and $h$, there is a least positive $d=d(t,h)$ so that if $P$ is a poset of height $h$ and the cover graph of $P$ has tree-width $t$, then $\dim(P)\le d$.
As discussed in greater detail in [@bib:JMMTWW], the function $d(t,h)$ must go to infinity with $h$ when $t\ge3$, and in view of Theorem \[thm:tree\], it is bounded for all $h$ when $t=1$. These observations left open the question as to whether $d(2,h)$ is bounded or goes to infinity with $h$. It is now known that $d(2,h)$ is bounded [@bib:BiKeYo; @bib:JMTWW]. However, in attacking this problem, the fact that one can restrict their attention to posets with $2$-connected cover graphs was a useful detail.
Also, the role of cut vertices in cover graphs surfaced in [@bib:Walc], where the following result, which is considerably stronger than Theorem \[thm:tree-width\], is proved.
\[thm:minors\] For any positive integers $t$ and $h$, there is a least positive integer $d=d(n,h)$ so that if $P$ is a poset of height at most $h$ and the cover graph of $P$ does not contain the complete graph $K_n$ as a minor, then $\dim(P)\le d$.
The proof given in [@bib:Walc] uses the machinery of structural graph theory. Subsequently, an alternative proof, using only elementary methods, was given in [@bib:MicWie].
Dimension Theory Essentials {#sec:dimension}
===========================
Let $P$ be a poset with ground set $X$. Then let ${\operatorname{Inc}}(P)$ denote the set of all ordered pairs $(x,y)\in X\times X$ where $x$ is incomparable to $y$ in $P$. The binary relation ${\operatorname{Inc}}(P)$ is of course symmetric, and it is empty when $P$ is a total order, and in this case, $\dim(P)=1$.
A subset $R\subseteq{\operatorname{Inc}}(P)$ is *reversible* when there is a linear extension $L$ of $P$ so that $x>y$ in $L$ for all $(x,y)\in R$. When ${\operatorname{Inc}}(P)\neq
\emptyset$, the dimension of $P$ is then the least positive integer $d$ for which there is a covering $${\operatorname{Inc}}(P)=R_1\cup R_2\cup\dots\cup R_d$$ such that $R_j$ is reversible for each $j=1,2,\dots,d$.
In the proof of the upper bound in our main theorem, we will apply these observations to show that a poset $P$ has dimension at most $d+2$ by first constructing a family ${\mathcal{F}}=\{L_1,L_2,\dots,L_d\}$ of linear extensions of $P$ and then setting $R=\{(x,y)\in P: x>y$ in $L_j$ for all $j=1,2,\dots,d\}.$ If $R=\emptyset$, then $\dim(P)\le d$, and when $R\neq\emptyset$, we will find a covering $R=R_{d+1}\cup R_{d+1}$, where both $R_{d+1}$ and $R_{d+2}$ are reversible. If $L_{d+1}$ and $L_{d+2}$ are linear extensions of $P$ so that for each $j=d+1,d+2$, $x>y$ in $L_j$ whenever $(x,y)\in R_j$, then ${\mathcal{R}}=\{L_1,L_2,\dots,L_{d},L_{d+1},L_{d+2}\}$ is a realizer of $P$, which shows $\dim(P)\le d+2$.
An indexed subset $\{(x_i,y_i)\colon 1\le i\le k\}\subseteq{\operatorname{Inc}}(P)$ is called an *alternating cycle* of length $n$ when $x_i\le y_{i+1}$ in $P$, for all $i=1,2,\dots,k$ (here, subscripts are interpreted cyclically so that $x_n\le y_1$ in $P$). In [@bib:TroMoo-2], the following elementary result is proved.
\[lem:alt-cycle\] Let $P$ be a poset and let $R\subseteq{\operatorname{Inc}}(P)$. Then $R$ is reversible if and only if $R$ does not contain an alternating cycle.
The following construction was given in [@bib:DusMil], where the concept of dimension was introduced. For an integer $d\ge2$, let $S_d$ be the following height $2$ poset: $S_d$ has $d$ minimal elements $\{a_1,a_2,\dots,a_d\}$ and $d$ maximal elements $\{b_1,b_2,\dots,b_d\}$. The partial ordering on $S_d$ is defined by setting $a_i<b_j$ in $S_d$ if and only if $i\neq j$. The poset $S_d$ is called the *standard example* (of dimension $d$).
In dimension theory, standard examples play a role which in many ways parallels the role of complete graphs in the study of chromatic number, and we refer the reader to [@bib:BHPT] for additional details on extremal problems for which results for graphs and results for posets have a similar flavor. In this paper, we will only need the following basic information about standard examples, which was noted in [@bib:DusMil].
\[pro:Sd\] For every $d\ge2$, we have $\dim(S_d)=d$. In fact, if $S_d$ is a standard example, and $a_i>b_i$ in a linear extension $L$ of $S_d$ then $a_j<b_j$ in $L$ whenever $i\neq j$.
Proof of the Upper Bound {#sec:upper-bound}
========================
Before launching into the main body of the proof, we pause to present an important proposition, which will be very useful in the argument to follow. When $M$ is a linear extension of a poset $P$ and $w\in P$, we will write $M=[A<w<B]$ when the elements of $P$ can be labeled so that $M=[u_1<u_2<\dots<u_m]$, $A=[u_1<u_2<\dots<u_{k-1}]$, $w=u_k$, and $B=[u_{k+1}<u_{k+2}<\dots<u_m]$. The generalization of this notation to an expression such as $M=[A<C<w<D<B]$ should be clear. Given this notation, the following is nearly self-evident, but it is stated for emphasis.
\[pro:merge\] Let $P$ be a poset, and let $w$ be a cut vertex in $P$. Let $P'$ and $P''$ be subposets of $P$ whose only common element is $w$. If $M'=[A<w<B]$ and $M''=[C<w<D]$ are linear extensions of $P'$ and $P''$, respectively, then $M=[A<C<w<D<B]$ is a linear extension of $P$. Furthermore, the restriction of $M$ to $P'$ is $M'$ and the restriction of $M$ to $P''$ is $M''$.
The rule $M=[A<C<w<D<B]$ will be called the *merge rule*. When it is applied, we will consider $M'=[A<w<B]$ as a linear extension of an “old” poset $P'$ which shares a cut vertex $w$ with a “new” poset $P''$ for which $M''=[C<w<D]$ is a linear extension. We apply the merge rule to form a linear extension $M=[A<C<w<D<B]$ of the union $P'\cup P''$ and note that $M$ forces old points in $A\cup B$ to the outside while concentrating new points from $C\cup D$ close to $w$.
Now on to the proof. We fix a positive integer $d\ge1$ and let $P$ be a poset for which $\dim(B)\le d$ for every block $B$ of $P$. The remainder of the proof is directed towards proving that $\dim(P)\le d+2$. Let $G$ be the cover graph of $P$. Since $d+2\ge3$, we may assume that $G$ is connected.
Let ${\mathcal{B}}$ be the family of blocks in $P$, and let $t=|{\mathcal{B}}|$. Then, let ${\mathcal{B}}=\{B_1,B_2,\dots,B_t\}$ be any labeling of the blocks of $P$ such that for every $i=2,3,\dots,t$, one of the vertices of $B_i$ belongs to some of the blocks $B_1,B_2,\dots,B_{i-1}$. Such a vertex of $B_i$ is unique and is a cut vertex of $P$—we call it the *root* of $B_i$ and denote it by $\rho(B_i)$.
For every block $B_i\in{\mathcal{B}}$ and every element $u\in B_i$, we define the *tail* of $u$ relative to $B_i$, denoted by $T(u,B_i)$, to be the subposet of $P$ consisting of all elements $v\in\{u\}\cup
B_{i+1}\cup B_{i+2}\cup\dots\cup B_t$ for which every path from $v$ to any vertex in $B_i$ passes through $u$. Note that $T(u,B_i)=\{u\}$ if $u$ is not a cut vertex. Also, if $u\in B_i$, $v\in B_{i'}$, and $(u,i)\neq(v,i')$, then either $T(u,B_i)\cap T(v,B_{i'})=\emptyset$ or one of $T(u,B_i)$ and $T(v,B_{i'})$ is a proper subset of the other.
For every block $B_i\in{\mathcal{B}}$, using the fact that $\dim(B_i)\le d$, we may choose a realizer ${\mathcal{R}}_i=\{L_j(B_i)\colon 1\le j\le d\}$ of size $d$ for $B_i$. For each $i=1,2,\dots,t$, set $P_i=
B_1\cup B_2\cup\dots\cup B_i$. Note that when $2\le i\le t$, we have $\rho(B_i)\in P_{i-1}$.
Fix an integer $j$ with $1\le j\le d$ and set $M_j(1)=L_j(B_1)$. Then, repeat the following for $i=2,3,\dots,t$. Suppose that we have a linear extension $M_j(i-1)$ of $P_{i-1}$. Let $w=\rho(B_i)$. Since $w\in P_{i-1}$, we can write $M_j(i-1)=[A<w<B]$. If $L_j(B_i)=
[C<w<D]$, we then use the merge rule to set $M_j(i)=[A<C<w<D<B]$. When the procedure halts, take $L_j=M_j(t)$. This construction is performed for all $j=1,2,\dots,d$ to detemine a family ${\mathcal{F}}=\{L_1,L_2,\dots,L_d\}$ of linear extensions of $P$. The family ${\mathcal{F}}$ is a realizer for a poset $P^*$ which is an extension of $P$. As outlined in the preceding section, we set $R=\{(x,y)\in{\operatorname{Inc}}(P)\colon x<y$ in $L_j$ for every $j=1,2,\dots,d\}$. We will show that there is a covering $R=R_{d+1}\cup R_{d+2}$ of $R$ by two reversible sets. This is enough to prove $\dim(P)\leq d+2$.
Repeated application of Proposition \[pro:merge\] immediately yields the following.
For each $j=1,2,\dots,d$ and each block $B_i\in{\mathcal{B}}$, the restriction of $L_j$ to $B_i$ is $L_j(B_i)$.
When $L$ is a linear order on a set $X$ and $S\subseteq X$, we say $S$ is an *interval* in $L$ if $y\in S$ whenever $x,z\in S$ and $x<y<z$ in $L$. The next property follows easily from the observation that the merge rule concentrates new points close around the cut vertex $w$ while pushing old points to the outside.
For every $j=1,2,\dots,d$, and every pair $(u,i)$ with $u\in B_i$, the tail $T(u,B_i)$ of $u$ relative to $B_i$ is an interval in $L_j$.
Let $(x,y)\in R$. Then, let $i$ be the least positive integer for which every path from $x$ to $y$ in the cover graph of $P$ contains at least two elements of the block $B_i$. We then define elements $u,v\in B_i$ by the following rules:
1. $u$ is the unique first common element of $B_i$ with every path from $x$ to $y$;
2. $v$ is the unique last common element of $B_i$ with every path from $x$ to $y$.
Note that $u\neq v$, $u=x$ when $x\in B_i$, and $v=y$ when $y\in B_i$.
\[cla:1\] The following two statements hold:
1. $x\in T(u,B_i)$, $y\notin T(u,B_i)$, $y\in T(v,B_i)$, and $x\notin T(v,B_i)$;
2. $u<v$ in $P$.
The first statement is an immediate consequence of the definition of tails. For the proof of the second statement, suppose to the contrary that $u\not<v$ in $P$. Since $u,v\in B_i$ and $u\neq v$, there is some $j$ with $1\le j\le d$ so that $u>v$ in $L_j(B_i)$. Therefore, $u>v$ in $L_j$. Since, $T(u,B_i)$ and $T(v,B_i)$ are disjoint intervals in $L_j$, we conclude that $x>y$ in $L_j$. This contradiction shows $u<v$ in $P$, as claimed.
\[cla:2\] At least one of the following two statements holds:
1. for all $y'$ with $y'\ge x$ in $P$, we have $y'\in T(u,B_i)$ and $y'<y$ in $P^*$;
2. for all $x'$ with $x'\le y$ in $P$, we have $x'\in T(v,B_i)$ and $x<x'$ in $P^*$.
Suppose to the contrary that neither of the two statements holds. Since $T(u,B_i)$ is an interval in $L_j$, $x\in T(u,B_i)$, $y\notin T(u,B_i)$, and $x<y$ in $L_j$ for each $j=1,2,\dots,d$, there must exist some $y'$ with $y'>x$ in $P$ and $y'\notin T(u,B_i)$. Then a path in $G$ from $x$ to $y'$ witnessing the inequality $x<y'$ in $P$ must include the point $u$. In particular, this implies that $x\le u$ in $P$. Similarly, we have $v\le y$ in $P$. This and Claim \[cla:1\] (2) yield $x\le u<v\le y$ in $P$, which is a contradiction.
Now, it is clear how to define the covering $R=R_{d+1}\cup R_{d+2}$. We assign $(x,y)$ to $R_{d+1}$ when the first statement in Claim \[cla:2\] applies, and we assign it to $R_{d+2}$ when the second statement applies. We show that $R_{d+1}$ is reversible. The argument for $R_{d+2}$ is symmetric. Suppose to the contrary that $R_{d+1}$ is not reversible. Then there is an integer $n\ge2$ and an alternating cycle $\{(x_i,
y_i)\colon 1\le i\le n\}$ contained in $R_{d+1}$. Then $x_i\le y_{i+1}$ for each $i=1,2,\dots,n$. However, since $(x_i,y_i)\in R_{d+1}$, we know that $y_{i+1}<y_i$ in $P^*$ for every $i=1,2,\dots,n$. This is impossible, because $P^*$ is a partial order. The proof of the upper bound in Theorem \[thm:main\] is now complete.
Proof that the Upper Bound is Best Possible {#sec:best-possible}
===========================================
As we noted previously, the examples shown in Figure \[fig:3dim-trees\] show that our upper bound is best possible when $d=1$. So in this section, we will fix an integer $d\ge2$ and show that there is a poset $P$ so that $\dim(P)=d+2$ while $\dim(B)\le d$ for every block $B$ of $P$.
For a positive integer $n$, we denote by ${\mathbf{n}}$ the $n$-element chain $\{0<1<\dots<n-1\}$. Also, we let ${\mathbf{n}}^d$ denote the cartesian product of $d$ copies of ${\mathbf{n}}$, that is, the elements of ${\mathbf{n}}^d$ are $d$-tuples of the form $u=(u_1,u_2,\dots,u_d)$ where each coordinate $u_i$ is an integer with $0\le u_i<n$. Also the partial order on ${\mathbf{n}}^d$ is defined by setting $u=(u_1,u_2,\dots,u_d)\le(v_1,v_2,\dots,
v_d)=v$ in ${\mathbf{n}}^d$ if and only if $u_i\le v_i$ in ${\mathbf{n}}$ for all $i=1,2,\dots,d$. As is well known, $\dim({\mathbf{n}}^d)=d$ for all $n\ge2$.
For each $n\ge2$, we then construct a poset $P=P(n)$ as follows. We start with a base poset $W$ which is a copy of ${\mathbf{n}}^d$. The base poset $W$ will be a block in $P$, and $W$ will also be the set of cut vertices in $P$. All other blocks in $P$ will be “diamonds”, i.e., copies of the $2$-dimensional poset on $4$ points discussed in conjunction with Figure \[fig:diamonds\]. Namely, for each element $w\in W$ we attach a three-element chain $x_w<y_w<z_w$ so that $x_w<w<z_w$ while $w$ is incomparable to $y_w$. In this way, the four-element subposet $\{w,x_w,y_w,z_w\}$ is a diamond.
We will now prove the following claim.
If $n$ is sufficiently large, then $\dim(P)\ge d+2$.
We must first gather some necessary tools from Ramsey theory. In particular we need a special case of a result which has become known as the “Product Ramsey Theorem,” and appears in the classic text [@bib:GrRoSp] as Theorem 5 on page 113. However, we will use slightly different notation in discussing this result.
Given a finite set $X$ and an integer $k$ with $0\le k\le|X|$, we denote the set of all $k$-element subsets of $X$ by $\binom{X}{k}$. When $T_1,T_2,\dots,T_d$ are $k$-element subsets of $X_1,X_2,\dots,X_d$, respectively, we refer to the product $g=T_1\times T_2\times\dots\times T_d$ as a *${\mathbf{k}}^d$-grid* in $X_1\times X_2\times\dots\times X_d$.
Here is a formal statement of the version of the Product Ramsey Theorem we require for our proof.
\[thm:prod-ramsey\] For every $4$-tuple $(r,d,k,m)$ of positive integers with $m\ge k$, there is an integer $n_0\ge k$ such that if $|X_i|\ge n_0$ for every $i=1,2,\dots,d$, then whenever we have a coloring $\phi$ which assigns to each ${\mathbf{k}}^t$-grid $g$ in $X_1\times X_2\times\dots\times X_d$ a color $\phi(g)$ from a set $R$ of $r$ colors, then there is a color $\alpha\in R$ so that for each $j=1,2,\dots,d$, there is an $m$-element subset $H_j\subseteq X_j$ such that $\phi(g)=\alpha$ for every ${\mathbf{k}}^t$-grid $g$ in $H_1\times H_2\times\dots\times H_d$.
We will apply this theorem with $k=2$, and since $k$ and $d$ are now both fixed, we will just refer to a ${\mathbf{k}}^d$-grid as a grid. When $g=T_1\times T_2\times\dots\times T_d$ is a grid, we consider the elements $w\in W$ with $w_j\in T_j$ for each $j=1,2,\dots,d$. Clearly, there are $2^d$ such points. Counting the diamonds attached to these points, there are $4\cdot 2^d$ points in $P$ associated with the grid $g$. But we want to focus on $4d+8$ of them: First, we consider an antichain $A=\{a_1,a_2,\dots,a_d\}$ defined as follows: For each $i,j=1,2,\dots,d$, coordinate $j$ of $a_i$ is: $\max(T_j)$ when $i=j$ and $\min(T_j)$ when $i\neq j$.
Dually, the antichain $B=\{b_1,b_2,\dots,b_d\}$ is defined as follows: For each $i,j=1,2,\dots,d$, coordinate $j$ of $b_i$ is: $\min(T_j)$ when $i=j$ and $\max(T_j)$ when $i\neq j$. We then note that when $1\le i,j\le d$, we have $a_i<b_j$ in $P$ if and only if $i\neq j$. As a consequence, the subposet of $P$ determined by $A\cup B$ is the standard example $S_d$ discussed previously. Note further that the points in the two antichains $\{x_{a_j}\colon 1\le j\le d\}$ and $\{z_{b_j}\colon 1\le j\le d\}$ also form a copy of $S_d$. Furthermore, if $x_{a_j}>z_{b_j}$ in some linear extension $L$ of $P$, then $a_j>b_j$ in $L$.
Now we consider two special points $c$ and $d$ in $W$ associated with the grid $g$ together with the points in the diamonds attached at $c$ and $d$. First, we take $c$ with $c_j=\min(T_j)$ for each $j=1,2,\dots,d$, and then we take $d$ with $d_j=\max(T_j)$ for each $j=1,2,\dots,d$. We note that $x_c<c<d<z_d$ in $P$. However, we also note that both $(x_c,y_d)$ and $(y_c,z_d)$ are in ${\operatorname{Inc}}(P)$.
Now suppose that $\dim(P)\le d+1$ and that ${\mathcal{R}}=\{L_1,L_2,\dots,L_{d+1}\}$ is a realizer of $P$. We will argue to a contradiction provided $n$ is sufficiently large. To accomplish this, we define a coloring $\phi$ of the grids in ${\mathbf{n}}^d$ using $(d+1)^{d+2}$ colors. Let $g$ be a grid and consider the $2d+4$ points discussed above. The color $\phi(g)$ will be a vector $(\alpha_1,\alpha_2,\dots,\alpha_{d+2})$ of length $d+2$ defined as follows:
1. for $j=1,2,\dots,d$, $\alpha_j$ is the least index $\alpha$ for which $x_{a_j}>z_{b_j}$ in $L_\alpha$;
2. $\alpha_{d+1}$ is the least index $\alpha$ for which $x_c>y_d$ in $L_\alpha$;
3. $\alpha_{d+2}$ is the least index $\alpha$ for which $y_c>z_d$ in $L_\alpha$.
We apply Theorem \[thm:prod-ramsey\] with $r=(d+1)^{d+2}$ and $m=3$. It follows that there is some fixed color $(\beta_1,
\beta_2,\dots,\beta_{d+2})$ such that for each $j=1,2,\dots,d$, there is a $3$-element subset $H_j=\{u_{1,j}<u_{2,j}<u_{3,j}\}
\subseteq\{0,1,\dots,n-1\}$ so that $\phi(g)=(\beta_1,\beta_2,\dots,
\beta_{d+2})$ for all grids $g$ in $H_1\times H_2\times\dots\times H_d$.
First, we claim that $\beta_{d+1}\neq\beta_{d+2}$. To see this, suppose that $\beta=\beta_{d+1}=\beta_{d+2}$. Then let $c,c',d,d'$ be vectors with $c_j=u_{1,j}$, $d_j=c'_j=u_{2,j}$ and $d'_j=u_{3,j}$ for each $j=1,2,\dots,d$. Then $$c>x_c>y_d=y_{c'}>z_{d'}>d'>c\quad\text{in $L_\beta$.}$$ Clearly, this is impossible.
In view of our earlier remarks concerning standard examples, we may relabel the linear extensions in the realizer so that $\beta_j=j$ for each $j=1,2,\dots,d$. Given the fact that $\beta_{d+1}\neq\beta_{d+2}$, (at least) one of $\beta_{d+1}$ and $\beta_{d+2}$ is in $\{1,2,\dots,d\}$. We complete the argument assuming that $1\le j=\beta_{d+1}\le d$, noting that the other case is symmetric.
Now let $c,d,e$ be points in $W$ defined as follows: First, set $c_j=u_{1,j}$, $e_j=u_{2,j}$ and $d_j=u_{3,j}$. Then for each $i=1,2,\dots,d$ with $i\neq j$, set $c_i=u_{1,i}$, $d_i=u_{2,i}$ and $e_i=u_{3,i}$. Note that $c<d$ and $c<e$ in $P$.
It follows then that $$c>x_c>y_d>x_d>z_e>e>c\quad\text{in $L_j$},$$ where the inequality $x_d>z_e$ follows from the fact that $\phi(g)=(\beta_1,\beta_2,\dots,\beta_{d+2})$ for the grid $g=\{u_{2,1},u_{3,1}\}\times\{u_{2,2},u_{3,2}\}\times\dots\times\{u_{2,d},u_{3,d}\}$. The contradiction completes the proof that the upper bound in our main theorem is best possible.
Closing Comments {#sec:closing}
================
There are two other instances where the Product Ramsey Theorem has been applied to combinatorial problems for posets, although, to be completely accurate, the first only uses it implicitly. The following two inequalities are proved in [@bib:Trot-2].
\[thm:max(P)\] Let $P$ be a poset which is not an antichain, and let $w$ be the width of the subposet $P-{\operatorname{Max}}(P)$. Then $\dim(P)\le w+1$.
\[thm:P-A\] Let $A$ be an antichain in a poset $P$. If $P-A\neq\emptyset$ and $w$ is the width of $P-A$, then $\dim(P)\le 2w+1$.
Both results admit quite simple proofs, and the only real challenge is to show that they are best possible. An explicit construction is given in [@bib:Trot-2] for a family of posets showing that Theorem \[thm:max(P)\] is tight, but the construction for the second is far more complicated and deferred to a separate paper [@bib:Trot-1]. Readers who are familiar with the details of this construction will recognize that it is an implicit application of the Product Ramsey Theorem.
The second application appears in [@bib:FeFiTr] where it is shown that there is a *finite* $3$-dimensional poset which cannot be represented as a family of spheres in Euclidean space—of any dimension—ordered by inclusion. While it may in fact be the case that such posets exist with only a few hundred points, the proof produces an example which is extraordinarily large. This results from the fact that a further strengthening of the Product Ramsey Theorem to a lexicographic version (see [@bib:FisGra]) is required.
We consider it a major challenge to construct examples of modest size for each of these three problems to replace the enormous posets resulting from the application of the Product Ramsey Theorem.
[99]{}
C. Biró, M. Keller and S. Young, Posets with cover graphs of path-width two have bounded dimension, *Order*, to appear, arXiv:1308.4877.
C. Biró, P. Hamburger, A. Pór and W. T. Trotter, Forcing posets with large dimension to contain large standard examples, submitted.
B. Dushnik and E. W. Miller, Partially ordered sets, *Amer. J. Math.* **63** (1941), 600–610.
S. Felsner, P. C. Fishburn and W. T. Trotter, Finite three dimensional partial orders which are not sphere orders, *Discrete Math.* **201** (1999), 101–132.
S. Felsner, W. T. Trotter and V. Wiechert, The dimension of posets with planar cover graphs, *Graphs Combin.*, in press, doi:10.1007/s00373-014-1430-4.
P. C. Fishburn and R. L. Graham, Lexicographic Ramsey theory, *J. Combin. Theory Ser. A* **62** (1993), 280–298.
R. L. Graham, B. L. Rothschild and J. H. Spencer, *Ramsey Theory*, 2nd Edition, J. H. Wiley, New York, 1990.
G. Joret, P. Micek, K. G. Milans, W. T. Trotter, B. Walczak and R. Wang, Tree-width and dimension, *Combinatorica*, in press, doi:10.1007/s00493-014-3081-8.
G. Joret, P. Micek, W. T. Trotter, R. Wang and V. Wiechert, On the dimension of posets with cover graphs of tree-width $2$, submitted, arXiv:1406.3397.
D. Kelly, The $3$-irreducible partially ordered sets, *Canad. J. Math.* **29** (1977), 367–383.
P. Micek and V. Wiechert, Topological minors of cover graphs and dimension, submitted.
N. Streib and W. T. Trotter, Dimension and height for posets with planar cover graphs, *European J. Combin.* **3** (2014), 474–489.
W. T. Trotter, Irreducible posets with arbitrarily large height exist, *J. Combin. Theory Ser. A* **17** (1974), 337–344.
W. T. Trotter, Inequalities in dimension theory for posets, *Proc. Amer. Math. Soc.* **47** (1975), 311–316.
W. T. Trotter, Combinatorial problems in dimension theory for partially ordered Sets, in *Problemes Combinatoires et Theorie des Graphes*, Colloque Internationeaux C.N.R.S. **260** (1978), 403–406.
W. T. Trotter, *Combinatorics and Partially Ordered Sets: Dimension Theory*, The Johns Hopkins University Press, Baltimore, 1992.
W. T. Trotter, Partially ordered sets, in *Handbook of Combinatorics*, R. L. Graham, M. Grötschel and L. Lovász, eds., Elsevier, Amsterdam, 1995, 433–480.
W. T. Trotter and J. I. Moore, Characterization problems for graphs, partially ordered sets, lattices, and families of sets, *Discrete Math.* **16** (1976), 361–381.
W. T. Trotter and J. I. Moore, The dimension of planar posets, *J. Combin. Theory Ser. B* **21** (1977), 51–67.
W. T. Trotter and R. Wang, Dimension and matchings in comparability and incomparability graphs, submitted.
B. Walczak, Minors and dimension, in *Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2015)*, pages 1698–1707, 2015, arXiv:1407.4066.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The local magnetism induced by vacancies in the presence of the spin-orbital interaction is investigated based on the half-filled Kane-Mele-Hubbard model on the honeycomb lattice. Using the self-consistent mean-field theory, we find that the spin-orbital coupling will enhance the localization of the spin moments near a single vacancy. We further study the magnetic structures along the zigzag edges formed by a chain of vacancies. We find that the spin-orbital coupling tends to suppress the counter-polarized ferrimagnetic order on the upper and lower edges, because of the open of the spin-orbital gap. As a result, in the case of the balance number of sublattices, it will suppress completely this kind of ferrimagnetic order. But, for the imbalance case, a ferrimagnetic order along both edges exists because additional zero modes will not be affected by the spin-orbital coupling.'
author:
- 'Weng-Hang Leong'
- 'Shun-Li Yu'
- 'Jian-Xin Li'
title: 'Effects of the spin-orbital coupling on the vacancy-induced magnetism on the honeycomb lattice'
---
INTRODUCTION
============
Graphene and related nanostructured materials have attracted much interest in solid state physics recently due to their bidimensional character and a host of peculiar properties [@Neto]. Among them, the investigation of the magnetic properties in graphene is one of the fascinating topics, as no $d$ and $f$ elements are necessary in the induction of magnetism in comparison with the usual magnetic materials. Theoretical predictions and experimental investigations have revealed that a nonmagnetic defect such as an impurity or a vacancy can induce the non-trivial localized magnetism [@Yazyev; @Hirashima; @Yazyev2; @Ugeda; @Nair]. Similarly, a random arrangement of a large number of vacancies which are generated by the high-dose exposure of graphene to strong electron irradiation [@Meyer] can also induce magnetism theoretically [@Fern2]. These studies not only have the fundamental importance, but also open a door for the possibility of application in new technologies for designing nanoscale magnetic and spin electronic devices.
On the other hand, the topological insulating electronic phases driven by the spin-orbital (SO) interaction have also attracted much interest recently. The Kane-Mele model for the topological band insulator is defined on the honeycomb lattice [@Kane; @Kane2] which is the same lattice structure as graphene. Possible realization of an appreciable SO coupling in the honeycomb lattice includes the cold fermionic atoms trapped in an extraordinary optical lattice [@Lee], the transition-metal oxide Na$_{2}$IrO$_{3}$ [@Shitade] and the ternaty compounds such as LiAuSe and KHgSb [@Hai]. Topological band insulator has a nontrivial topological order and exhibits a bulk energy gap with gapless, helical states at the edge [@Zhang; @Konig; @Hasan]. These edge states are protected by the time reversal symmetry and are robust with respect to the time-reversal symmetric perturbations, such as non-magnetic impurities. It is shown that a vacancy, acting as a minimal circular inner edge, will induce novel time-reversal invariant bound states in the band gap of the topological insulator [@Shen; @Gonz; @Liang]. Theoretically, it is also shown that the SO coupling suppresses the edge magnetism induced in the zigzag ribbon of the honeycomb lattice in the presence of electron-electron interactions [@Fern]. Thus, it is expected that the SO coupling would also affect the local magnetism in the bulk induced by vacancies.
In this paper, we study theoretically the effects of the SO coupling on the local magnetism induced by a single and a multi-site vacancy on the honeycomb lattice, based on the Kane-Mele-Hubbard model where both the SO coupling and the Hubbard interaction between electrons are taken into consideration. This model has been extensively studied to explore the effect of the strong correlation on the topological insulators [@SRachel; @MHohenadler; @SLYu; @DZheng; @WWu; @CGriset; @JWen]. Making use of the self-consistent mean field approximation, we calculate the local spin moments and their distribution around the vacancies. For a single vacancy, we find that the main effect of the SO coupling is to localize the spin moments to be near the vacancy, so that it will enhance the local spin moments. For a large stripe vacancy by taking out a chain of sites from the lattice, we find that the SO coupling tends to suppress the counter-polarized ferrimagnetic order induced along the zigzag edges, because of the open of the SO gap. As a result, in the case of the balance number of sublattices (with even number of vacancies), the SO coupling will suppress completely the counter-polarized ferrimagnetic order along the upper and lower edges. While, in the case of the imbalance number of sublattices (with odd number of vacancies), a ferrimagnetic order along both edges exists because additional zero modes will not be affected by the SO coupling.
We will introduce the model and the method of the self-consistent mean-field approximation in Sec.II. In Sec.III and IV, we present the results for a single vacancy and a multi-site vacancy, respectively. Finally, a brief summary will be given in Sec.V.
MODELS and COMPUTATIONAL METHODS
================================
We start from the Kane-Mele model [@Kane], in which the intrinsic SO coupling with a coupling constant $\lambda$ is included. $$H_0=-t\sum_{\langle ij\rangle,\sigma}c^\dag_{i\sigma}
c_{j\sigma}+i\lambda\sum_{\langle\langle ij\rangle\rangle\sigma\sigma'}v_{ij}\sigma^z_{\sigma\sigma'}
c^\dag_{i\sigma}c_{j\sigma'},$$ where $c^{\dag}_{i\sigma}$ ($c_{j\sigma}$) is the creation(annihilation) operator of the electron with spin $\sigma$ on the lattice site $i$, $\langle ij\rangle$ represents the pairs of the nearest neighbor sites (the hopping is $t$) and $\langle\langle ij\rangle\rangle$ those of the next-nearest neighbors. $v_{ij}=+1(-1)$ if the electron makes a left(right) turn to get to the second bond. The size of our system is considered to be finite with periodic boundary condition. So, the position of each lattice site can be described specifically by $i=\Gamma(m,n)$, representing that the lattice site $i$ is in the $m$th column and the $n$th row, and $\Gamma=A,B$ the sublattice labels. The number of the unit cells is denoted by $N_c=L^2$, therefore the total number of the lattice sites is $N_l=2L^2$. To consider the correlation between electrons, we will include the Hubbard term in the Hamiltonian, which is given by $H_I$, $$H_I=U\sum_{i}\hat{n}_{i\uparrow}\hat{n}_{i\downarrow},$$ where $\hat{n}_{i\sigma}=c^\dag_{i\sigma}c_{i\sigma}$. When vacancies are introduced, the hoppings between the vacancy and the nearest neighbors and the on-site interaction on that vacancy are subtracted from the overall Hamiltonian. Hence the corresponding number of the lattice sites is $N_l=2L^2-N_v$, where $N_v$ is the number of vacancies. The total number of electrons $N_e$ is fixed to be at the half-filling ($N_e=N_l$). The Hubbard interaction term is treated with the self-consistent mean field approximation, so that we will obtain an effective single-particle Hamiltonian where the electrons interact with a spin-dependent potential, $$H_I\simeq U\sum_{i,\sigma}
\langle\hat{n}_{i-\sigma}\rangle\hat{n}_{i\sigma}-U\sum_{i}\langle\hat{n}_{i\uparrow}\rangle
\langle\hat{n}_{i\downarrow}\rangle.$$ And the overall mean field Hamiltonian $H_{mf}$ is then given by, $$H_{mf}=U\sum_{i\sigma}\langle\hat{n}_{i-\sigma}\rangle\hat{n}_{i\sigma}+H_0.$$ After diagonalizing the Hamiltonian $H_{mf}$, we can determine the occupation number $\langle\hat{n}_{i-\sigma}\rangle$ at each site with different spins using the eigenvectors of $H_{mf}$, and this process is carried out iteratively until a required accuracy is reached. Then the magnetic moment of each site $m_{i}=\langle\hat{n}_{i\uparrow}-\hat{n}_{i\downarrow}\rangle$ can be calculated. We note that a collinear magnetic texture is assumed in our system, as used before for the investigations of Kane-Mele-Hubbard model [@Fern; @SRachel]. We have checked the results with the non-collinear magnetic texture and found that the collinear magnetic texture is favored.
MAGNETISM WITH ONE VACANCY
==========================
![(color online). (a) and (b): Distribution of the spin moments $m_{i}$ on lattice sites around a single vacancy at $A(7,7)$ with $U=1.0t$, in which (a) corresponds to the SO coupling constant $\lambda=0.0$ and (b)$\lambda=0.1t$. The area and color of the hollow circles represent the magnitude of the spin moments. (c)$m_{i}$ on the $B$ sublattice as a function of the distance $r$ away from the vacancy. The unit $a$ is the distance between the nearest sites.[]{data-label="loc1"}](loc_moment_1.eps){width="46.00000%"}
The calculation is carried out on the lattice with $N_c=14\times14$ unit cells in which a single vacancy is introduced on the site $A(7,7)$. Figure \[loc1\] displays the distribution of the magnetic moment when the Hubbard interaction is taken to be $U=1.0t$, in which the size and the color of the circle on each lattice site denote the magnitude of the local spin moment. From Fig.1(a) where the SO coupling is turned off, one can see that localized magnetic moments are induced around the vacancy in the presence of a finite Hubbard interaction $U$. This is in agreement with the prediction of the Lieb theorem [@Lieb] regarding the total spin $S$ of the exact ground state of the Hubbard model on bipartite lattices. It states that the total spin $S$ is given by the sublattice imbalance $2S=|N_{A}-N_{B}|$, with $N_{A}$ and $N_{B}$ the number of atoms belonging to each sublattice. With the introducing of a single vacancy on the $A$ sublattice, an imbalance $N_{B}-N_{A}=1$ appears and a magnetic structure near the vacancy with the total spin $S=1/2$ will form. Similar results have also been obtained in recent studies in graphene [@Yazyev; @Hirashima; @Yazyev2].
In the presence of the SO coupling, the magnitude of the magnetic moments around the vacancy increases, as shown in Fig. \[loc1\](b) for $\lambda=0.1t$. At the meantime, if we check the distribution of the magnetic moments, as shown in Fig.1(c) where the magnitude of the magnetic moments on sublattice $B$ as a function of the distance $r$ away from the vacancy is presented, one will find that the magnetic moments are more localized with the increase of the SO coupling. These features demonstrate that the SO coupling will enhance the magnetic moments near the vacancy notably.
![(color online). LDOS for $\lambda=0.0$ \[left column, including (a),(c),(e),(g)\] and for $\lambda=0.1t$ \[right column, including (b),(d),(f),(h)\], where the Hubbard interaction $U=0.0$ for (a) and (b), $U=1.6t$ for (c) and (d), $U=2.6t$ for (e) and (f), and $U=3.6t$ for (g) and (h), respectively. LDOS for different spins is resolved, those with the spin up are denoted by the blue lines and the spin down the red lines. The grey dash lines represent the LDOS on the lattice site away from the vacancy. []{data-label="loc2"}](loc_moment_2.eps){width="46.00000%"}
In order to show the emergence of the magnetism induced by the vacancy in more detail, we calculate the spin resolved local density of state(LDOS) as defined by, $$D_{\sigma}(\epsilon)=\Sigma_{n,i}|u^{n}_{i,\sigma}|^2\delta(\epsilon-\epsilon_n),$$ where $i$ runs over the lattice sites surrounding the vacancy up to the third-nearest neighbors, as those linked by the green line in Fig.\[loc1\](a) and (b). $u^{n}_{i,\sigma}$ is the single-particle amplitude on the $i$th site with spin $\sigma$ and the corresponding eigenvalue is $\epsilon_n$. The Delta function in Eq.(2) is replaced by the Lorentzian function for plotting. The results for the LDOS are presented in Fig. \[loc2\](a)-(h) for different Hubbard interaction $U$ and SO interaction $\lambda$. The red and blue lines represent the LDOS for the spin up and spin down components respectively, and the dash lines show the LDOS away from the vacancy for a comparison. In the case of $U=\lambda=0.0$ as shown in Fig. \[loc2\](a), the LDOS shows a V-shape linear behavior near the Fermi level for those lattice sites far away from the vacancy (denoted by the dashed line) which is the consequence of the linear dispersion relation of the electrons, the so-called Dirac fermions. For those around the vacancy, a peak at the Fermi level emerges as shown by the solid line, which corresponds to the localized states induced by the vacancy [@Castro]. After turning on the SO coupling, such as that for $\lambda=0.1t$\[see Fig.2(b)\], we can see that an energy gap opens for those lattice sites far away from the vacancy [@Kane; @Kane2], so that now a U-shape LDOS near the Fermi level occurs. In this way, the mid-gap peak is enhanced noticeably because the decay rate of the localized states into the continuum is reduced largely due to the open of the energy gap. This will lead to the increase in the spectral weight of the localized states around the vacancy. However, for both cases, one will find that the LDOS for the spin up and spin down components degenerates, so that the system will not show magnetism as a whole without the Hubbard interaction.
The effect of a finite Hubbard interaction $U$ is to split the spin degenerate LDOS, so that two peaks occur corresponding to different spins, as shown in Fig. \[loc2\](c)-(f). Consequently, the localized spin up and down moments will not cancel out in this case, and a net magnetism around the vacancy is induced.
![(color online). Local moments $M_{loc}$ (see text) are shown as a function of the SO coupling $\lambda$ for different Hubbard interaction $U$(a) and of $U$ for different $\lambda$(b).[]{data-label="loc3"}](loc_moment_3.eps){width="44.00000%"}
The magnetism may be quantified by the local moment $M_{loc}=\sum_{i}m_{i}$, where the sum runs over the lattice sites surrounding the vacancy up to the third-nearest neighbors as used above in the calculation for the LDOS. The results are presented in Fig. \[loc3\](a) and (b) for different $U$ and $\lambda$, respectively. The local moment $M_{loc}$ shows a monotonic increase with the SO coupling $\lambda$, so it reinforces our observation that the local magnetism is enhanced by the SO coupling as shown in Fig.1. On the other hand, $M_{loc}$ shows a nonmonotonic dependence on the Hubbard interaction $U$, namely it increases with $U$ firstly and then decreases with a further increase of $U$ after a critical value $U_{c}$. As discussed above, the local magnetism is determined by the spin-split localized states induced by the vacancy, and it is the Hubbard interaction $U$ to split the spin-degenerate states. Because the open of the gap due to the SO coupling will decrease the decay rate of the localized states into the continuum, so it will enhance the spectral weight of the localized states\[see also Fig. \[loc2\]\], consequently the localized magnetism. The splitting between the two localized states with different spins is proportional to $U$, so the two split localized states will situate in the SO gap for a small $U$\[Fig. \[loc2\](c)-(f)\]. However, when $U>U_{c}$ the splitting will be larger than the SO gap, and it pushes the localized states to merge into the continuum\[Fig.2(g) and (h)\], so the local magnetism will decrease.
THE CASE OF MULTI-SITE VACANCY
==============================
![(color online). Distribution of the spin moments $m_{i}$ on the lattice sites surrounding the vacancies for $U=1.0t$ and $L=14$. A cluster of vacancies is formed with the number of missing sites for (a), (b) $N_{v}=8$ and (c), (d) $N_{v}=7$. SO coupling is set to be $\lambda=0.0$ for (a), (c) and $\lambda=0.1t$ for (b), (d). The area and color of hollow circles represent the magnitude of the moments. []{data-label="loc4"}](loc_moment_4.eps){width="46.00000%"}
The multi-site vacancy can be formed by removing the sites continuously. Here, we consider a large stripe vacancy by taking out a chain of sites from the lattice as illuminated in Fig. \[loc4\]. In this way, the stripe vacancy consists of one upper and one lower zigzag edges. As clarified by the Lieb theorem [@Lieb], the sublattice imbalance between the number of atoms belonging to different sublattices will have significant effect on the magnetism. For the stripe vacancy considered here, the imbalance is expressed by the parity of the number of vacancies, where the number is even ($N_{A}=N_{B}$) in Fig. \[loc4\](a) and (b), and odd ($N_{A}\neq N_{B}$) in Fig. \[loc4\](c) and (d), thus the total spin of the system is $S=0$ and $1/2$ respectively.
In the case of even number of vacancies, a ferrimagnetic spin order emerges on both the upper and lower zigzag edges around the stripe vacancies when there is no SO coupling, as shown in Fig. \[loc4\](a). The ferrimagnetic arrangement and the magnitude of the spin moments on these two edges are symmetric, but they are counter-polarized, so they cancel out exactly and the whole system will not show magnetism. This is consistent with the Lieb theorem [@Lieb]. The ferrimagnetic order on a sufficiently long zigzag edge around the stripe vacancies here is similar to the spin order formed at the outer edge of the zigzag ribbon [@Fujita; @Wakabayashi; @Young; @Fern3] and the graphene nanoisland [@Pala]. In the case of odd number of vacancies, a similar ferrimagnetic spin order is also induced with a slightly large magnitude \[Fig. \[loc4\](c)\]. Interestingly, this ferrimagnetic order occurs only on the upper zigzag edge, not on the lower edge. This phenomenon is ascribed to the presence of an extra spin when a sublattice imbalance $N_{A}\neq N_{B}$ exists, as described by the Lieb theorem [@Lieb].
![(color online). (a)Local moments $M_{loc}$ are plotted as a function of $\lambda$ while $U=1.0t$ and $L=14$. The function in different size of vacancy is distinguished by different color and shape of points. The cases of even $N_{v}$ are not plotted as local moments are always zero obeying Lieb theorem [@Lieb]. (b) The function of edge moments $M_{e}$ versus $\lambda$ are given in different $N_{v}$. []{data-label="loc5"}](loc_moment_5.eps){width="48.00000%"}
After turning on the SO coupling, such as for $\lambda=0.1t$, the ferrimagnetic spin order on both the upper and lower zigzag edges around the stripe vacancies disappears completely in the case of even number of vacancies\[Fig. \[loc4\](b)\]. However, the effect of the SO coupling on local magnetism is quite different for the case of an odd number of vacancies. Here, a ferrimagnetic spin order similar to that on the upper edge emerges on the lower edge, though the magnitude of the individual spin moment is reduced\[Fig. \[loc4\](d)\]. To show variation of the total magnetism, we plot the quantity $M_{loc}$ as a function of the SO coupling $\lambda$ in Fig. \[loc5\](a), here $M_{loc}$ is the sum of the spin moments on the sites which are on the zigzag edges around the vacancies. Since $M_{loc}$ is always zero in the case of even $N_{v}$, it is not plotted here. With an odd $N_{v}$, the local moment $M_{loc}$ increases with the increase of $\lambda$, which shows a similar behavior as that in the case of a single vacancy. This indicates that the total local magnetism shown in Fig. \[loc4\](d) is in fact enhanced with the introduction of the SO coupling and approaches the saturation value 1 finally. From Fig. \[loc5\](a), one can also find that $M_{loc}$ increases with the increase of the number of vacancies $N_{v}$. This suggests that the SO coupling will localize the induced spin moments to those lattice sites which are neighboring the vacancies.
To quantify the variation of the spin moments with $\lambda$ on the upper zigzag edge, we also present $M_{e}$ as a function of $\lambda$ in Fig. \[loc5\](b), here $M_{e}$ is the sum of the spin moments only on the sites on the upper zigzag edge. Let us consider firstly the case of even number of $N_{v}$, for a small number of even vacancies, $M_{e}$ is always zero. Up to $N_{v}\geq8$, a finite $M_{e}$ occurs and it increases with $N_{v}$ by the formation of the zigzag edges. However, $M_{e}$ drops rapidly to zero after turning on the SO coupling. These results quantify the physical picture derived from Fig. \[loc4\](a) and (b). Now let us turn to the case of odd number of $N_{v}$. Without the SO coupling, $M_{e}$ also shows an increase with $N_{v}$. With the introduction of the SO coupling, $M_{e}$ shows a decrease with $\lambda$ and saturates to near one half of ${M_{loc}}$.
![(color online). The single-particle energy levels labeled with $m$ (see text) near the Fermi level of the non-interacting systems for (a)$N_v=8$ and (b) $N_v=7$. SO coupling is set to be $\lambda=0.0$ for the red circles and $\lambda=0.1t$ for the blue squares.[]{data-label="loc6"}](loc_moment_6.eps){width="48.00000%"}
In fact, we can make an analogy between the stripe vacancy and the graphene ribbon with zigzag edges. A remarkable feature of the graphene ribbon with zigzag edges is that it has a flat band localized on the zigzag edge[@Neto]. An important effect of this flat band is that a counter-polarized ferromagnetic order along the upper and lower edges will be induced when the Hubbard interaction between electrons is included [@Fujita; @Wakabayashi; @Young; @Fern3]. In view of this, we plot the single-particle spectra for the systems with the stripe vacancy without the Hubbard interaction in Fig.\[loc6\](a) and (b) for $N_{\nu}=8$ and $N_{\nu}=7$, respectively. Each energy level is labeled with $m=n-N_{e}-1/2$ in order to indicate that the energy level with $m<0$ is occupied by electron. For the systems without SO coupling, we find that there are four near-degeneracy localized states \[red circles indicated by arrows in Fig.\[loc6\](a) and (b)\] which is near the Fermi level for both $N_{\nu}=8$ and $N_{\nu}=7$. These states will have the same effect as the flat band in the zigzag ribbon when a suitable Hubbard $U$ is turned on. So, a counter-polarized ferrimagnetic order as shown in Fig.\[loc4\](a) will emerge. However, we note that there are two additional zero modes for $N_{\nu}=7$ relative to $N_{\nu}=8$, due to the imbalance between the sublattices ($N_{A}>N_{B}$). These zero modes will induce extra spin moments on both edges, which counteract the antiparallel moments on the lower edge. Thus, in the case of $N_{\nu}=7$, only the ferrimagnetic order on the upper edge appears. After turning on the SO coupling, those localized states \[blue squares indicated by arrows in Fig.\[loc6\](a) and (b)\] are pushed away from the Fermi level due to the open of the SO gap. Thus, as shown in Fig.\[loc4\](b), a small Hubbard U is not enough to induce the counter-polarized ferrimagnetic order on the upper and lower edges. However, the zero modes originating from the imbalance of sublattices are not affected by the SO coupling \[Fig.\[loc6\](b)\]. So, the additional ferrimagnetic order on both edges induced by these zero modes will remain for $N_{\nu}=7$.
CONCLUSION
==========
In a summary, we have studied the local magnetism induced by vacancies on the honeycomb lattice based on the Kane-Mele-Hubbard model. It is shown that the SO coupling tends to localize and consequently enhances the local magnetic moments near a single vacancy. Furthermore, along the zigzag edges formed by a chain of vacancies, the SO coupling will suppress completely the counter-polarized ferrimagnetic order along the edges. Therefore, the system will not show any local magnetism in the case of even number of vacancies. For an odd number of vacancies, a ferrimagnetic order along both edges exists and the total magnetic moments along both edges will increase.
This work was supported by the National Natural Science Foundation of China (Grant Nos. 91021001, 11190023 and 11204125) and the Ministry of Science and Technology of China (973 Project grant numbers 2011CB922101 and 2011CB605902).
[99]{} A. H. C. Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, and A. K. Geim, Rev. Mod. Phys. [**81**]{}, 109 (2009). O. V. Yazyev and L. Helm, Phys. Rev. B [**75**]{}, 125408 (2007). H. Kumazaki and D. S. Hirashima, J. Phys. Soc. Jpn. [**76**]{}, 064713 (2007). O. V. Yazyev, Phys. Rev. Lett. [**101**]{}, 037203 (2008). M. M. Ugeda, I. Brihuega, F. Guinea, and J. M. Gómez-Rodríguez, Phys. Rev. Lett. [**104**]{}, 096804 (2010). R. R. Nair, M. Sepioni, I-L. Tsai, O. Lehtinen, J. Keinonen, A. V. Krasheninnikov, T. Thomson, A. K. Geim, and I. V. Grigorieva, Nature Phys. [**8**]{}, 199-202 (2012). J. Kotakoski, A. V. Krasheninnikov, U. Kaiser, and J. C. Meyer, Phys. Rev. Lett. [**106**]{},105505 (2011). J. J. Palacios, J. Fernández-Rossier, and L. Brey, Phys. Rev. B [**77**]{}, 195428 (2008). C. L. Kane and E. J. Mele, Phys. Rev. Lett. [**95**]{}, 146802 (2005). C. L. Kane and E. J. Mele, Phys. Rev. Lett. [**95**]{}, 226801 (2005). K. L. Lee, B. Gremaud, R. Han, B. G. Englert, and C. Miniatura, Phys. Rev. A [**80**]{}, 043411 (2009). A. Shitade, H. Katsura, J. Kunes, X. -L. Qi, S. -C. Zhang, and N. Nagaosa, Phys. Rev. Lett. [**102**]{}, 256403 (2009). H. -J. Zhang, S. Chadov, L. Müchler, B. Yan, X. -L. Qi, J. Kübler, S. -C. Zhang, and C. Felser, Phys. Rev. Lett. [**106**]{}, 156402 (2011). B. A. Bernevig, T. A. Hughes, and S. C. Zhang, Science [**314**]{}, 1757 (2006). M. König, Wiedmann, C. Brüne, A. Roth, H. Buhmann, L. W. Molenkamp, X. L. Qi, and S. C. Zhang, Science [**318**]{}, 766 (2007). M. Z. Hasan and C. L. Kane, Rev. Mod. Phys. [**82**]{}, 3045 (2010). W. -Y. Shan, J. Lu, H. -Z. Lu, and S. -Q. Shen, Phys. Rev. B [**84**]{}, 035307 (2011). J. W. González and J. Fernández-Rossier, Phys. Rev. B [**86**]{}, 115327 (2012). J. He, Y. -X. Zhu, Y. -J. Wu, L. -F. Liu, Y. Liang, and S. -P. Kou, Phys. Rev. B [**87**]{}, 075126 (2013). D. Soriano and J. Fernández-Rossier, Phys. Rev. B [**82**]{}, 161302(R) (2010). S. Rachel and K. Le Hur, Phys. Rev. B [**82**]{}, 075106 (2010). M. Hohenadler, T. C. Lang, and F. F. Assaad, Phys. Rev. Lett. [**106**]{}, 100403 (2011). S. L. Yu, X. C. Xie, and J. X. Li, Phys. Rev. Lett. [**107**]{}, 010401 (2011). D. Zheng, G. M. Zhang, and C. Wu, Phys. Rev. B [**84**]{}, 205121 (2011). W. Wu, S. Rachel, W. M. Liu, and K. Le Hur, Phys. Rev. B [**85**]{}, 205102 (2012). C. Griset and C. Xu, Phys. Rev. B [**85**]{}, 045123 (2012). J. Wen, M. Kargarian, A. Vaezi, and G. A. Fiete, Phys. Rev. B [**84**]{}, 235149 (2011). E. H. Lieb, Phys. Rev. Lett. [**62**]{}, 1201 (1989). V. M. Pereira, F. Guinea, J. M. B. Lopes dos Santos, N. M. R. Peres, and A. H. Castro Neto, Phys. Rev. Lett. [**96**]{}, 036801 (2006). M. Fujita, K. Wakabayashi, K. Nakada, and K. Kusakabe, J. Phys. Soc. Jpn. [**65**]{}, 1920 (1996). K. Wakabayashi, M. Sigrist, and M. Fujita, J. Phys. Soc. Jpn. [**67**]{}, 2089 (1998). Y. W. Son, M. L. Cohen, and S. G. Louie, Phys. Rev. Lett. [**97**]{}, 216803 (2006). J. Fernández-Rossier, Phys. Rev. B [**77**]{}, 075430 (2008). J. Fernández-Rossier and J. J. Palacios, Phys. Rev. Lett. [**99**]{}, 177204 (2007).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we introduce a discrete variant of the meta-learning framework. Meta-learning aims at exploiting prior experience and data to improve performance on future tasks. By now, there exist numerous formulations for meta-learning in the continuous domain. Notably, the Model-Agnostic Meta-Learning (MAML) formulation views each task as a continuous optimization problem and based on prior data learns a suitable initialization that can be adapted to new, unseen tasks after a few simple gradient updates. Motivated by this terminology, we propose a novel meta-learning framework in the discrete domain where each task is equivalent to maximizing a set function under a cardinality constraint. Our approach aims at using prior data, i.e., previously visited tasks, to train a proper initial solution set that can be quickly adapted to a new task at a relatively low computational cost. This approach leads to (i) a personalized solution for each individual task, and (ii) significantly reduced computational cost at test time compared to the case where the solution is fully optimized once the new task is revealed. The training procedure is performed by solving a challenging discrete optimization problem for which we present deterministic and randomized algorithms. In the case where the tasks are monotone and submodular, we show strong theoretical guarantees for our proposed methods even though the training objective may not be submodular. We also demonstrate the effectiveness of our framework on two real-world problem instances where we observe that our methods lead to a significant reduction in computational complexity in solving the new tasks while incurring a small performance loss compared to when the tasks are fully optimized.'
author:
- 'Arman Adibi[^1] ,Aryan Mokhtari[^2] ,Hamed Hassani[^3]'
bibliography:
- 'ref.bib'
- 'ref\_2.bib'
title: ' Submodular Meta-Learning '
---
Introduction
============
Many applications in artificial intelligence necessitate exploiting prior data and experience to enhance quality and efficiency on new tasks. This is often manifested through a set of tasks given in the training phase from which we can learn a model or representation that can be later used for new unseen tasks in the test phase. In this regard, meta-learning aims at exploiting the data from the available tasks to learn model parameters or representation that can be later used to perform well on new unseen tasks, in particular, when we have access to limited data and computational power at the test time [@thrun2012learning; @schmidhuber1992learning; @bengio1990learning; @vilalta2002perspective]. By now, there are several formulations for meta-learning, but perhaps one of the most successful ones is the Model-Agnostic Meta-Learning (MAML) framework proposed in [@finn2017model]. In MAML, we aim to train the model parameters such that applying a few steps of gradient-based updates with a small number of samples from a new task would perform well on that task. MAML can also be viewed as a way to provide a proper initialization, from which performance on a new task can be optimized after a few gradient-based updates. Alas, this scheme only applies to settings in which the decision variable belongs to a *continuous* domain and can be adjusted using gradient-based methods at the test time.
Our goal is to extend the methodology of MAML to the *discrete* setting. We consider a setting in which our decision variable is a discrete set, and our goal is to come up with a good initial set that can be quickly adjusted to perform well over a wide range of new tasks. In particular, we focus on submodular maximization to represent the tasks which is an essential class of discrete optimization.
There are numerous applications where the submodular meta-learning framework can be applied to find a personalized solution for each task while significantly reducing the computation load. In general, most recommendation tasks can be cast as an instance of this setting[@gabillon2013adaptive; @el2009turning; @yue2011linear]. Consider the task of recommending a set of items, e.g., products, locations, ads, to a set of users. One approach for solving such a problem is to find the subset of items that have the highest score over all the previously-visited users and recommend that subset to a new user. Indeed, this approach leads to a reasonable performance at test time; however, it does not provide a user-specific solution for a new user. Another approach is to find the whole subset at the test time when the new user arrives. In contrast to the previous approach, this scheme leads to a user-specific solution, but at the cost of running a computationally expensive algorithm to select all the elements at the test time.
In our meta-learning framework, the process of selecting set items to be recommended to a new user is done in two parts: In the first part, a set of items is selected offline according to prior experience. These items are the most popular items to the previously-visited users (depending on the context). In the second part, which happens at the test time, a set of items that is *personalized* to the coming user is selected. These are items that are computed specifically according to the features of the coming user. In this manner, the computation for each coming user would be reduced to the selection of the second part, which typically constitutes a small portion of the final set of recommended items. The first part can be done offline with a lower frequency. For instance, in a real recommender system, the first part can be computed once every hour, and the second part can be computed specifically for each coming user (or for a class of similar users). While we have mentioned recommendation (or more generally facility location) as a specific example, it is easy to see that this framework can be easily used to reduce computation in other notable applications of submodular optimization. **Contributions.** Our contributions are threefold:
- We propose a novel discrete meta-learning framework where each task is equivalent to maximizing a set function under some cardinality constraint. Our framework aims at using prior data, i.e., previously visited tasks, to train a proper initial solution set that can be quickly adapted to a new task at a low computational cost to obtain a task-specific solution.
- We present computationally efficient deterministic and randomized meta-greedy algorithms to solve the resulting meta-learning problem. When the tasks are monotone and submodular, we prove that the solution obtained by the deterministic algorithm is at least $0.53$-optimal, and the solution of the randomized algorithm is $(1-1/e-o(1))$-optimal in expectation, where the $o(1)$ term vanishes by the size of the solution. These guarantees are obtained by introducing new techniques, despite that the meta-learning objective is *not* submodular.
- We study the performance of our proposed meta-learning framework and algorithms for movie recommendation and ride-sharing problems. Our experiments illustrate that the solution of our proposed meta-learning scheme, which chooses a large portion of the solution in the training phase and a small portion adaptively at test time, is very close to the solution obtained by choosing the entire solution at the test time when a new task is revealed.
Related work
------------
**Continuous Meta-Learning.** Meta-learning has gained considerable attention recently mainly due to its success in few shot learning [@vinyals2016matching; @DBLP:conf/iclr/RaviL17; @snell2017prototypical; @wang2019few] as well as reinforcement learning [@DBLP:journals/corr/DuanSCBSA16; @wang2016learning; @DBLP:conf/iclr/SongGYCPT20; @fallah2020provably]. One of the most successful forms of meta-learning is the gradient-based *Model Agnostic Meta-learning* (MAML) approach[@finn2017model]. MAML aims at learning an initialization that can be adapted to a new task after performing one (or a few) gradient-based update(s); see, e.g., [@fallah2019convergence]. This problem can be written as $$\min_{w \in W} \mathbb{E}_{a\sim P}[f_a(w-\nabla f_a(w))],$$ where $W \subseteq \mathbb{R}^d$ is the feasible set and $l$ is the probability distribution over tasks. The previous works on MAML including [@nichol2018first; @finn2018probabilistic; @DBLP:conf/iclr/GrantFLDG18; @yoon2018bayesian; @DBLP:conf/iclr/AntoniouES19; @rajeswaran2019meta; @fallah2019convergence; @collins2020distribution] consider the case where $W$ is a continuous space. In fact none of these works can be applied to the case where the feasible parameter space is discrete. In this paper, we aim to close this gap and extend the terminology of MAML to discrete settings.
**Submodular Maximization.** Submodular functions have become key concepts in numerous applications such as data summarization [@lin2011class; @wei2013using; @kirchhoff2014submodularity; @mirzasoleiman2016distributed], viral marketing [@kempe2003maximizing], sensor placement [@krause2008near], dictionary learning [@das2011submodular], and influence maximization [@kempe2003maximizing]. It is well-known that for maximizing a monotone and submodular function under the cardinality constraint, the greedy algorithm provides a $(1-1/e)$-optimal solution [@krause2014submodular; @nemhauser1978best; @wolsey1982analysis]. There has been significant effort to improve the scalability and efficiency of the greedy algorithm using lazy, stochastic, and distributed methods [@mirzasoleiman2015lazier; @karimi2017stochastic; @barbosa2015power; @mirrokni2015randomized; @kumar2015fast; @hassani2017gradient; @mokhtari2020stochastic; @karbasi2019stochastic; @balkanski2019exponential]. However, our framework is fundamentally different and complementary to these approaches as it proposes a new approach to use data at training time to improve performance at new tasks. Indeed, all the aforementioned techniques can be readily used to further speed-up our algorithms. Optimization of related submodular tasks has been a well-studied problem with works on structured prediction [@lin2012learning], submodular bandits [@yue2011linear; @zhang2019online], online submodular optimization [@jegelka2011online; @streeter2009online; @golovin2014online; @chen2018projection], and public-private data summarization [@mirzasoleiman2016fast]. However, unlike our work, these approaches are not concerned with train-test phases for optimization. Another recently-developed methodology to reduce computation is the two-stage submodular optimization framework [@balkanski2016learning; @mitrovic2018data; @stan2017probabilistic], which aims at summarizing the ground set to a reasonably small set that can be used at test time. The main difference of our framework with the two-stage approaches is that we allow for *personalization*: A small subset of items that can be found at test time specific to the task at hand. This leads to a completely new problem formulation, and consequently, new algorithms.
Problem Statement: Discrete Meta-Learning {#sect:prob-statement}
=========================================
**Setup.** We consider a family of tasks $\mathcal{T}=\{\mathcal{T}_i\}_{i\in \mathcal{I}}$, where the set $\mathcal{I}$ could be of infinite size. Each task $\mathcal{T}_i$ is represented via a set function $f_i:2^{V} \to {\mathbb{R}}_+$ that measures the reward of a set $S \subseteq V $ for the $i$-th task, and performing the task $\mathcal{T}_i$ would mean to maximize the function $f_i$ subject to a given constraint. For instance, in a recommender system where we aim to recommend a subset of the items to the users, the set $\mathcal{I}$ denotes the set of all the possible users and selecting which items to recommend to a user $i \in \mathcal{I}$ is viewed as the task $\mathcal{T}_i$. Moreover, the function $f_i$ encodes the users satisfaction, i.e., $f_i(S)$ quantifies how suitable the set of items $S$ is for user $i$. Taking a statistical perspective, we assume that the tasks $\mathcal{T}_i$ occur according to a possibly unknown probability distribution $i \sim p$.
In this paper, we focus on the case where the functions $f_i$ are monotone and submodular set functions and each task $\mathcal{T}_i$ amounts to maximizing $f_i$ under the $k$-cardinality constraint. That is, the task $\mathcal{T}_i$ is to select a subset $S \subseteq V$ of size $k$ such that the value of $f_i(S)$ is maximized. Submodularity of $f_i$ means that for any $A,B \subseteq V$ the following inequality holds $f_i(A) + f_i(B) \geq f_i(A \cup B) + f_i(A \cap B)$. Furthermore, $f_i$ is called monotone if for any $A \subseteq B$ we have $f_i(A) \leq f_i(B)$.
**Training and test tasks.** We assume access to a collection of *training* tasks $\{\mathcal{T}_i\}_{i=1}^m$. These are the tasks that we have already experienced, i.e., they correspond to the users that we have already seen. Formally, this means that for each training task $\mathcal{T}_i$, we assume knowledge of the corresponding function $f_i$. In our formulation, each of the training tasks is assumed to be generated i.i.d. according to the distribution $p$. Indeed, eventually we aim to optimize performance at *test* time, i.e., obtain the best performance for new and unseen tasks generated independently from the distribution $p$. For instance, in our recommendation setting, test tasks correspond to new users that will arrive in the future. Our goal is to use the training tasks to reduce the computation load at test time.
**Two extremes of computation.** Let us use $\mathcal{T}_{\rm test}$ (and $f_{\rm{test}}$) to denote the task (and its corresponding set function) that we aim to learn at test time. Ideally, if we have sufficient computational power, then we should directly optimize $f_{\rm{test}}$ by solving the following problem $$\label{test_problem}
\max_{S\in V, |S|\leq k}\ f_{\rm{test}}(S).$$ We denote the optimal solution of by $S_{\rm test}^*$. For instance, we can use the greedy procedure to solve which leads to a $(1-1/e)$-optimal solution using $\mathcal{O}(kn)$ evaluations of $f_{\rm test}$, and through $k$ passes over the ground set. However, the available computational power and time in the test phase is often limited, either because we need to make quick decisions to respond to new users or since we need to save energy. For instance, in real-world advertising or recommendation systems, both these requirements are crucial: many users arrive within each hour which means fast optimization is crucial (especially if $n,k$ are large), and also, reducing computation load would lead to huge energy savings in the long run. In such cases, Problem should be solved approximately with less computation.
An alternative to reduce computation at test time is to solve the problem associated with the expected reward over all possible tasks in the training phase (when we have sufficient computation time), i.e., $$\label{expected_max}
\max_{S\in V, |S|\leq k} \ \mathbb{E}_{i \sim p}\; [f_i(S)].$$ We denote the optimal solution of by $S_{\rm exp}^*$. The rationale behind this approach is that the optimal solution to this problem would generalize well over an unseen task if the new task is also drawn according to the probability distribution $p$. In other words, the solution of should perform well for the problem in that we aim to solve at the test time, assuming that $f_{\rm test}$ is sampled according to $p$. In this way, we do not need any extra computation at the test time. However, in this case, the solution that we obtain would not be the best possible solution for the task that we observe at the test time, i.e., $S_{\rm test}^*$ is not equal to $S_{\rm exp}^*$. Note that we often do not have access to the underlying probability distribution $p$, and we only have access to a large set of realizations of tasks in the training phase. As a result, instead of solving , we settle for maximizing the sample average function $$\label{empirical_max}
\max_{S\in V, |S|\leq k} \ \frac{1}{m}\sum_{i=1}^m f_i(S),$$ where $m$ is the number of available tasks in the training phase.
Problems and can be considered as two different extreme cases. In the first option, by solving , we avoid any pre-processing in the training phase, and we obtain the best possible guarantee for the new task, but at the cost of performing computationally expensive operations (e.g., full greedy) at the test time. In the second approach, by solving in the training phase, we obtain a solution that possibly performs reasonably without any computation at the test phase, but the quality of the solution may not be as good as the first option. In summary, there exists a trade-off between the required computational cost at the test time and the performance guarantee on the unseen task. Hence, a fundamental question that arises is what would be the best scheme at the training phase assuming that at test time we have some limited computational power. For instance, in the monotone submodular case, assume that instead of running the greedy algorithm for $k$ rounds, which has a complexity of $\mathcal{O}(kn)$, we can only afford to run $\alpha k$ rounds of greedy at test time, which has a complexity of $\mathcal{O}(\alpha nk)$, where $\alpha\in(0,1)$ is small. In this case, a natural solution would be to find an appropriate set of $(1-\alpha)k$ elements in the training phase, and add the remaining $\alpha k$ elements at test time when a new task arrives. This discussion also applies to any other greedy method (e.g., lazy or stochastic greedy). We now formally state this problem.
[0.21]{} {width="1.0\linewidth" height="1\linewidth"}
[0.21]{} {width="1.0\linewidth" height="1\linewidth"}
[0.21]{} {width="1.0\linewidth" height="1\linewidth"}
[0.21]{} {width="1.0\linewidth" height="1\linewidth"}
**Discrete Meta-Learning.** As we discussed so far, when computational power is limited at test time, it makes sense to divide the process of choosing the best decision between training and test phases. To be more specific, in the training phase, we choose a subset of elements from the ground set that would perform over the training tasks, and then select (or optimize) the remaining elements at the test time *specifically* with respect to the task at hand. To state this problem, consider ${S_{\rm{tr}}}\subseteq V$ with cardinality $|{S_{\rm{tr}}}|=l$, where $l<k$, as the initial set that we aim to find at the training phase, and the set $S_i$ that we add to the initial set ${S_{\rm{tr}}}$ at test time (See Figure \[fig:toy\] for an illustration). Hence, the problem of interest can be written as $$\label{eq:ML_submodular}
\max_{{S_{\rm{tr}}}\in V, |{S_{\rm{tr}}}|\leq l} \ \mathbb{E}_{i \sim p} \Big[\max_{S_i\in V, |S_i|\leq k-l} f_i({S_{\rm{tr}}}\cup S_i)\Big],$$
Note that the critical decision variable that we need to find is ${S_{\rm{tr}}}$ which is the best initial subset of size $l$ overall all possible choices of task when a best subset of size $k-l$ is added to that. In fact, if we define $f_i'({S_{\rm{tr}}}):=\max_{S_i\in V, |S_i|\leq k-l} f_i({S_{\rm{tr}}}\cup S_i)$, then we can rewrite the problem in as $$\max_{{S_{\rm{tr}}}\in V, |{S_{\rm{tr}}}|\leq l} \ \mathbb{E}_{i \sim p}\; \left[ f_i'({S_{\rm{tr}}}) \right].$$
As described previously, we often do not have access to the underlying probability distribution $p$ of the tasks, and we instead have access to a large number of sampled tasked that are drawn independently according to $p$. Hence, instead of solving , we solve its sample average approximation given by $$\label{eq:ML_submodular_sample_avg}
\max_{{S_{\rm{tr}}}\in V, |{S_{\rm{tr}}}|\leq l} \ \frac{1}{m}\sum_{i=1}^m\; \left[\max_{S_i\in V, |S_i|\leq k-l} f_i({S_{\rm{tr}}}\cup S_i)\right]\ = \ \max_{{S_{\rm{tr}}}\in V, |{S_{\rm{tr}}}|\leq l} \ \frac{1}{m}\sum_{i=1}^m\; \left[ f_i'({S_{\rm{tr}}}) \right],$$ where $m$ is the number of tasks in the training set which are sampled according to $p$. Even though the functions $f_i$ are submodular, $f'_i$ *is not submodular* or $k$-submodular [@ohsaka2015monotone] (see Appendix \[sec:counter\_example\] for specific counter examples). Hence, Problem is not a submodular maximization problem. In the next section, we present algorithms for solving Problem with provable guarantees.
We finally note that Problem will be solved at *training* time to find the solution ${S_{\rm{tr}}}$ of size $l$. This solution is then *completed at test time*, by, e.g., running $k-l$ further rounds of greedy on the new task, to obtain a task-specific solution of size $k$.
Submodular Cross-Learning
-------------------------
\[pr: cross-learning submodular\] Let $f_i:2^{\mathcal{X}}\xrightarrow{}{\mathbb{R}}_+$ for $i\in [m-1]$, be a monotone submodular function over ground set ${\mathcal{X}}$. We want to find $S_m$ with size at most $l$ whose subests with average size of at most $k$ maximize the sum of $f_i$ for $i\in [m-1]$, more formally:
$$\max_{S_m\subseteq {\mathcal{X}}, \mid S_m \mid\leq l}\;\sum_{i=1}^{m-1}\; \max_{S_i\subseteq S_m,\; \frac{1}{m-1}\sum\limits_{i=1}^{m-1} \mid S_i\mid \leq k}\; f_i(S_i)$$
\[eq:cross-learning submodular\] $$\begin{aligned}
{2}
&\operatorname*{maximize}&&\sum_{i=1}^{m-1}\; \; f_i(S_i) \\
&\operatorname*{subject\,\, to \quad}&& {S_m\subseteq {\mathcal{X}}, \mid S_m \mid\leq l} \\
& && {S_i\subseteq S_m,\; \frac{1}{m-1}\sum\limits_{i=1}^{m-1} \mid S_i\mid \leq k}\end{aligned}$$ \[eq:basic-formulation\]
Algorithms for Discrete Submodular Meta-Learning {#sec:algorithms_ML}
================================================
Solving Problem requires finding a set ${S_{\rm{tr}}}$ for the outer maximization and sets $\{S_i\}_{i=1}^m$ for the inner maximization. In this section, we describe our proposed greedy-type algorithms to select the elements ${S_{\rm{tr}}}$ and $\{S_i\}_{i=1}^m$. As we deal with $m+1$ sets, the order in which the sets ${S_{\rm{tr}}}$ and $\{S_i\}_{i=1}^m$ are updated becomes crucial, i.e., it is not clear which of the sets ${S_{\rm{tr}}}$ or $S_i$’s should be preferably updated in each round and how can the functions $f_i$ be incorporated in finding the right order, which is the main challenge in designing greedy methods to solve . We design greedy procedures with both deterministic and randomized orders and provide strong guarantees for their solutions.
Deterministic Algorithms
------------------------
In this section, we first describe Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\] which use specific orderings to solve Problem . Based on these two, we then design Algorithm \[alg: main discrete meta-Greedy\] as our *main deterministic* algorithm. Throughout this section, we use $\Delta_i(e|S)=f_i(S\cup\{e\})-f_i(S)$ to denote the marginal gain of adding an element $e$ to set $S$ for function $f_i$. In brief, Algorithm \[alg: discrete meta-Greedy\] first fills ${S_{\rm{tr}}}$ greedily up to completion and then it constructs each of the $S_i$’s greedily on the top of ${S_{\rm{tr}}}$. Specifically, starting from the empty set initialization for ${S_{\rm{tr}}}$ and $S_i$’s, Algorithm \[alg: discrete meta-Greedy\] constructs in its first phase the set ${S_{\rm{tr}}}$ in $l$ rounds, by adding one element per round, where the next element in each round is chosen according to $ e^* = \operatorname*{arg\,max}_{e \in V} \sum_{i=1}^m f_i({S_{\rm{tr}}}\cup \{e\}) - f_{i}({S_{\rm{tr}}})$. Once ${S_{\rm{tr}}}$ is completed, in the second phase, each of the sets $S_i$ is constructed in parallel by running the greedy algorithm on $f_i$. That is, each $S_i$ is updated in $k-l$ rounds where in each round an element with maximum marginal on $f_i$ is added to $S_i$ based on $ e^*_i = \operatorname*{arg\,max}_{e \in V} f_i({S_{\rm{tr}}}\cup S_i \cup\{e\}) - f_i({S_{\rm{tr}}}\cup S_i)$.
**Initialize** ${S_{\rm{tr}}}=\{S_i\}_{i=1}^m=\emptyset$ Find $e^{*}\!\!\! =\!\! \operatorname*{arg\,max}_{ e\in V}\sum\limits_{i=1}^{m}\Delta_i(e|{S_{\rm{tr}}}) $ ${S_{\rm{tr}}}\xleftarrow{} {S_{\rm{tr}}}\cup\{e^{*}\}$ **end for** Find $e_i^{*} \!=\!\operatorname*{arg\,max}_{ e\in V} \Delta_i(e|{S_{\rm{tr}}}\!\cup\! S_i)$ $S_i\xleftarrow{} S_i\cup\{e_i^{*}\}$ **end for** **end for** Return ${S_{\rm{tr}}}$ and $S_i$
**Initialize** ${S_{\rm{tr}}}=\{S_i\}_{i=1}^m=\emptyset$ Find $e_i^{*} = \operatorname*{arg\,max}_{ e\in V} \Delta_i(e|S_i)$ $S_i\xleftarrow{} S_i\cup\{e_i^{*}\}$ **end for** **end for** Find $ e^{*} =\!\operatorname*{arg\,max}_{ e\in V} \!\sum\limits_{i=1}^{m}\!\Delta_{i}(e|{S_{\rm{tr}}}\cup S_i)$ ${S_{\rm{tr}}}\xleftarrow{} {S_{\rm{tr}}}\cup\{e^{*}\}$ **end for** Return ${S_{\rm{tr}}}$ and $S_i$
**Initialize** the sets ${S_{\rm{tr}}}$ and $\{S_i\}_{i=1}^m$ to the empty set. Find $e^{*} = \operatorname*{arg\,max}_{ e\in V} \sum_{i=1}^{m}f_i({S_{\rm{tr}}}\cup\{e\})-f_i({S_{\rm{tr}}})$ ${S_{\rm{tr}}}\xleftarrow{} {S_{\rm{tr}}}\cup\{e^{*}\}$ **end for** Find $e_i^{*} = \operatorname*{arg\,max}_{ e\in V} f_i({S_{\rm{tr}}}\cup S_i\cup\{e\})-f_i({S_{\rm{tr}}}\cup S_i)$ $S_i\xleftarrow{} S_i\cup\{e_i^{*}\}$ **end for** **end for** return ${S_{\rm{tr}}}$ and $S_i$
Algorithm \[alg: Reverse discrete meta-Greedy\] uses the opposite ordering of Algorithm \[alg: discrete meta-Greedy\]. Initializing with all sets to be empty, in the first phase it constructs the sets $S_i$ using the greedy procedure on $f_i$, i.e., each $S_i$ is updated in parallel in $k-l$ rounds, where in each round the element $e_i^{*}$ defined as $e_i^{*} = \operatorname*{arg\,max}_{ e\in V} f_i({S_{\rm{tr}}}\cup S_i\cup\{e\})-f_i({S_{\rm{tr}}}\cup S_i)$ is added to $S_i$. In the second phase, the set ${S_{\rm{tr}}}$ is formed greedily in $l$ rounds, and in each round the element $e^{*} $ defined as $e^{*} = \operatorname*{arg\,max}_{ e\in V} \sum_{i=1}^{m}f_i({S_{\rm{tr}}}\cup\{e\}\cup S_i)-f_i({S_{\rm{tr}}}\cup S_i)$ is added.
**Initialize** the sets ${S_{\rm{tr}}}$ and $\{S_i\}_{i=1}^m$ to the empty set. Find $e_i^{*} = \operatorname*{arg\,max}_{ e\in V} f_i(S_i\cup\{e\})-f_i(S_i)$ $S_i\xleftarrow{} S_i\cup\{e_i^{*}\}$ **end for** **end for** Find $e^{*} = \operatorname*{arg\,max}_{ e\in V} \sum_{i=1}^{m}f_i({S_{\rm{tr}}}\cup\{e\}\cup S_i)-f_i({S_{\rm{tr}}}\cup S_i)$ ${S_{\rm{tr}}}\xleftarrow{} {S_{\rm{tr}}}\cup\{e^{*}\}$ **end for**
return ${S_{\rm{tr}}}$ and $S_i$
While the solutions obtained by Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\] are guaranteed to be near-optimal, it turns out that they can be complementary with respect to each other. Our *main* deterministic algorithm, called `Meta-Greedy`, runs both Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\] and chooses as output the solution, among the two, that leads to a higher objective value in . To be more specific, if we consider ${S_{\rm{tr}}}^{(1)}, \{S_i^{(1)}\}_{i=1}^m$ as the outputs of Algorithm \[alg: discrete meta-Greedy\] and ${S_{\rm{tr}}}^{(2)}, \{S_i^{(2)}\}_{i=1}^m$ as the outputs of Algorithm \[alg: Reverse discrete meta-Greedy\], then `Meta-Greedy` compares the values of $ \sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{(1)}\cup S_i^{(1)})$ and $ \sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{(2)}\cup S_i^{(2)})$ and chooses the solution set that has the higher objective function value. Note that as we described earlier, the main output of this procedure should be the set ${S_{\rm{tr}}}$ of size $l$. Hence, the output of `Meta-Greedy` is either ${S_{\rm{tr}}}^{(1)}$ or ${S_{\rm{tr}}}^{(2)}$ and the sets $ \{S_i^{(1)}\}_{i=1}^m$ and $ \{S_i^{(2)}\}_{i=1}^m$ are only evaluated for the purpose of comparing objective function values.
Next, we explain why our $\texttt{Meta-Greedy}$ method can outperform both Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\]. This will be done by providing the theoretical guarantees for these methods and consequently explaining why Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\] are complementary.
Run Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\] and obtain respective solution sets ${S_{\rm{tr}}}^{(1)}, \{S_i^{(1)}\}_{i=1}^m$ and ${S_{\rm{tr}}}^{(2)}, \{S_i^{(2)}\}_{i=1}^m$.
Compute the objective value $ \sum_{i=1}^{m}f_i({S_{\rm{tr}}}\cup S_i)$ for both solution sets.
**Theoretical guarantees.** We begin with the analysis of Algorithm \[alg: discrete meta-Greedy\]. The following proposition relates the overall performance of Algorithm \[alg: discrete meta-Greedy\] to its performance after phase 1 and shows that the output of the algorithm is at least $1/2$-optimal. We use OPT for the optimal value of Problem .
\[alg1lemma\] Let ${S_{\rm{tr}}}^{(1)}, \{S_i^{(1)}\}_{i=1}^m$ be the output of Algorithm \[alg: discrete meta-Greedy\], and define $\beta$ as $\beta :=\frac{1}{m} \sum_{i=1}^m f_i({S_{\rm{tr}}}^{(1)})$. If the set functions $f_i$ are monotone and submodular, then $$\frac{1}{m} \sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(1)} \cup S_i^{(1)})
\geq \max \Bigl\{ \beta \,, \, (1-1/e)({\rm{OPT}} - 2 \beta) + \beta \Bigr \}.$$ Consequently, the solution obtained by Algorithm \[alg: discrete meta-Greedy\] is at least $1/2$-optimal for any value of $\beta$.
Check Appendix \[sec:proof\_prop\_1\].
The proof of this proposition is relegated to the appendix. The key step in the proof is to relate the progress made in phase 1 to the gap to OPT. This is indeed challenging as phase 1 only involves updates on the outer maximization of . In this regard, we prove a novel technical lemma that can be generally applicable to any mini-max submodular problem. The guarantee given in Proposition \[alg1lemma\] is minimized when $\beta = \text{OPT}/2$. If $\beta$ is small (e.g., $\beta = 0$) or if $\beta$ is large (e.g. if $\beta = (1-1/e)\text{OPT}$) then the guarantee becomes tight (e.g. $(1-1/e)\text{OPT})$. This is indeed expected from the greedy nature of the two phases of Algorithm \[alg: discrete meta-Greedy\]. What is non-trivial about the result of Proposition \[alg1lemma\] is that it provides a strong guarantee for any value of $\beta$, and not just cases that $\beta$ is small or large. Similarly, we can provide near-optimality guarantees for Algorithm \[alg: Reverse discrete meta-Greedy\].
\[alg2lemma\] Let ${S_{\rm{tr}}}^{(2)}, \{S_i^{(2)}\}_{i=1}^m$ be the output of Algorithm \[alg: Reverse discrete meta-Greedy\], and define $\gamma$ as $\gamma := \frac{1}{m}\sum_{i=1}^m f_i(S_i^{(2)})$. If the set functions $f_i$ are monotone and submodular, then $$\frac{1}{m} \sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(2)} \cup S_i^{(2)}) \geq \max \Bigl\{ \gamma \,, \, (1-1/e)({\rm{OPT}} - 2\gamma) + \gamma \Bigr \}.$$ Consequently, the solution obtained by Algorithm \[alg: Reverse discrete meta-Greedy\] is at least $1/2$-optimal for any value of $\gamma$.
Check Appendix \[sec:proof\_prop\_2\].
Similarly, we can show that $\gamma=\text{OPT}/2$ leads to (the worst) guarantee $1/2$-OPT, while for large and small values of $\gamma$ the bound in Proposition \[alg2lemma\] approaches the optimal approximation $(1-1/e)\text{OPT}$.
We note that the values $\beta$ in Proposition \[alg1lemma\] (Algorithm \[alg: discrete meta-Greedy\]) and $\gamma$ in Proposition \[alg2lemma\] (Algorithm \[alg: Reverse discrete meta-Greedy\]) represent two different extremes. The value $\beta$ represents the significance of the role of ${S_{\rm{tr}}}$ in solving Problem , and $\gamma$ represents how significant the role of the sets $\{S_i\}_{i=1}^m$ can be. Even though the worst-case guarantees of Propositions \[alg1lemma\] and \[alg2lemma\] are obtained when $\beta, \gamma = \text{OPT}/2$, a coupled analysis of the algorithms show that in this case at least one of the algorithms should output a solution which is strictly better than $1/2$-optimal. In other words, the outcomes of Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\] are dependent to one another, and the best performance is achieved when the maximum of the two is considered. This justifies why our main algorithm $\texttt{Meta-Greedy}$ can perform strictly better than each of the Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\]. Using a coupled analysis of the outcome of Algorithms \[alg: discrete meta-Greedy\] and \[alg: Reverse discrete meta-Greedy\], we can bound the performance of $\texttt{Meta-Greedy}$ for different values of $\beta$ and $\gamma$ (see the proof of Theorem \[thm:meta-greedy\] in the appendix). In particular, we can show that the output of $\texttt{Meta-Greedy}$ is at least $0.53$-optimal. The proof of the following theorem carefully analyzes the interplay between the role of the inner and outer maximization problems in . We emphasize that the proof introduces new techniques applicable to other types of minimax submodular problems.
\[thm:meta-greedy\] Consider the $\texttt{Meta-Greedy}$ algorithm outlined in Algorithm \[alg: main discrete meta-Greedy\]. If the functions $f_i$ are monotone and submodular, then we have $$\label{bound-meta-greedy}
\max \Bigl\{ \frac{1}{m}\sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(1)} \cup S_i^{(1)}) \, , \, \frac{1}{m} \sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(2)} \cup S_i^{(2)}) \Bigr\} \geq 0.53 \times {\rm{OPT}}.$$
Check Appendix \[sec:proof\_of\_main\_thm\].
Note that for all the results in Propositions \[alg1lemma\] and \[alg2lemma\] as well as Theorem \[thm:meta-greedy\], for given output sets ${S_{\rm{tr}}}$ or $\{S_i\}_{i=1}^m$, the value of $\frac{1}{m} \sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i)$ is a lower bound for the objective function value of Problem evaluated at the output set ${S_{\rm{tr}}}$. To be more precise, the accurate measure for evaluating the quality of the output set ${S_{\rm{tr}}}$ is $\frac{1}{m}\sum_{i=1}^m\; \left[\max_{S_i\in V, |S_i|\leq k-l} f_i({S_{\rm{tr}}}\cup S_i)\right]$ which is indeed larger than $\frac{1}{m} \sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i)$. Hence, all the guarantees that have obtained in the statements above (as well as Theorem \[thm: randomized meta\] below) would directly translate into the same guarantees when we evaluate the objective in on the set ${S_{\rm{tr}}}$.
Randomized Algorithm
--------------------
In this section, we consider greedy procedures in which the decision to alternate between the set ${S_{\rm{tr}}}$ (the outer maximization) and the sets $\{S_i\}_{i=1}^m$ (the inner maximization) is done based on a randomized scheme. The `Randomized meta-Greedy` procedure, outlined in Algorithm \[alg: Randomized meta-Greedy\], provides a specific randomized order. In each round, with probability $l/k$ we choose to perform a greedy update on ${S_{\rm{tr}}}$, and with probability $1-l/k$ we choose to perform a greedy update on *all* the $S_i$’s, $i=1, \cdots,m$. This procedure continues until either ${S_{\rm{tr}}}$ or $\{S_i\}_{i=1}^m$ hit their corresponding carnality constraint, in which case we continue to update the other set(s) greedily until they also become full.
**Initialize** the sets ${S_{\rm{tr}}}$ and $\{S_i\}_{i=1}^m$ to the empty set. $e^{*}_i \xleftarrow{} \operatorname*{arg\,max}_{ e\in V} f_i({S_{\rm{tr}}}\cup S_i\cup\{e\})-f_i({S_{\rm{tr}}}\cup S_i)$ $e^{*}_{tr} \xleftarrow{} \operatorname*{arg\,max}_{ e\in V} \sum_{i=1}^{m} f_i({S_{\rm{tr}}}\cup S_i\cup\{e\})-f_i({S_{\rm{tr}}}\cup S_i)$ **w.p. $\frac{l}{k}$:** ${S_{\rm{tr}}}={S_{\rm{tr}}}\cup\{e^{*}_{tr}\}$ **w.p. $\frac{k-l}{k}$:** $S_i=S_i\cup\{e_i^{*}\}$, $\forall i = 1, \cdots, m$ **end** If ${S_{\rm{tr}}}$ or $S_i$’s have not yet reached their cardinality limit then fill them greedily until their limit is reached Return ${S_{\rm{tr}}}$ and $S_i$
The randomized update of Algorithm \[alg: Randomized meta-Greedy\] is designed to optimally connect the expected increase the objective value at each round with the gap to OPT (as shown in the proof of Theorem \[thm: randomized meta\]). Hence, the `Randomized meta-Greedy` procedure is able to achieve in expectation a guarantee close to the tight value $(1-1/e)\text{OPT}$. However, due to the randomized nature of the algorithm, the sets ${S_{\rm{tr}}}$ or $S_i$ might hit their carnality constraint earlier than expected. Analyzing the function value at this “stopping time” is another technical challenge that we resolve in the following theorem to obtain a guarantee that becomes slightly worse than $(1-1/e)\text{OPT}$ depending on the values of $k-l$ and $l$.
\[thm: randomized meta\] Let the (random) sets ${S_{\rm{tr}}}$, $\{S_i\}_{i=1}^m$ be the output of Algorithm \[alg: Randomized meta-Greedy\]. If the functions $f_i$ are monotone and submodular, then $$\mathbb{E}\Bigl[\frac{1}{m}\sum_{i=1}^m f_i ({S_{\rm{tr}}}\cup S_i) \Bigr]
\geq \left(1 - \frac{1}{e} - b \right) \rm{OPT},$$ where $b \to 0$ as $k-l$ and $l$ grow. More precisely, letting $c = \max\{\frac{1}{k-l}, \frac{1}{l}\},$ we have $b = c + (\exp(3\sqrt{c\log1/c}) - 1)/e = {\mathcal{O}}(\sqrt{c \log 1/c}).$
Check Appendix \[sec:proof\_of\_main\_random\_theorem\].
All presented algorithms are designed for the training phase and their output is the set ${S_{\rm{tr}}}$ with size $l$. The sets $\{S_i\}_{i=1}^m$ are only computed for algorithmic purposes. Given a new task at the test phase, the remaining $k-l$ task-specific elements will be added to ${S_{\rm{tr}}}$ using for instance greedy updates that require a total complexity of $O((k-l)n)$ in function evaluations. Also, the training complexity of the proposed algorithms is $O(kmn)$, however, certain phases can be implemented in parallel.
Simulation Results {#sec:Simulation}
==================
We provide two experimental setups to evaluate the performance of our proposed algorithms and compare with other baselines. Each setup involves a different set of tasks which are represented as submodular maximization problems subject to the $k$-cardinality constraint. In our experiments, we consider the following algorithms: **Meta-Greedy** (Algorithm \[alg: main discrete meta-Greedy\]), **Randomized Meta-Greedy** (Algorithm \[alg: Randomized meta-Greedy\]), **Greedy-Train** (which chooses all the $k$ elements during the training phase–see and the discussion therein), **Greedy-Test** (which chooses all the $k$ elements during the test phase–see and the discussion therein), and **Random** (which chooses a random set of $k$ elements). In the following, we briefly explain the data and tasks and refer the reader to the supplementary materials for more details.
We described the Meta submodular learning problem and algorithms in previous sections. In this section, we provide two applications:
- Ride Share Optimization
- Movie Recommendation
We compare the performance of proposed algorithms on them by showing the comparison in figure\[fig:simulation-ride\].
[0.48]{} {width="1\linewidth"}
[0.48]{} {width="1\linewidth"}
[0.48]{} {width="1\linewidth"}
[0.48]{} {width="1\linewidth"}
[0.48]{} {width="1.1\linewidth"}
[0.48]{} {width="1.1\linewidth"}
**Ride Share Optimization.** We will formalize and solve a facility location problem on the Uber dataset [@uber_2019]. Our experiments were run on the portion of data corresponding to Uber pick-ups in Manhattan in the period of September 2014. This portion consists of $\sim 10^6$ data points each represented as a triplet $(latitude, longitude, DateTime)$. A customer and a driver are specified through their locations on the map. We use $u=(x_u,y_u)$ for a customer a and $r=(x_r, y_r)$ for a driver. We define the “convenience score” of a (customer, driver) pair as $c(u,r)=2-\frac{2}{1+e^{-200 d(u,r)}}$, where $d(u,r)$ denotes the Manhattan distance [@mitrovic2018data]. Given a specific time $a$, we define a time slot $T_a$ and picking inside the data set 10 points in half an hour prior to time $a$, and for each point we further pick 10 points in its 1 km neighborhood, which makes a total of 100 points (locations) on the map. A task ${\mathcal{T}}_i$ takes place at a corresponding time $a_i$, and by defining the set of locations $T_{a_i}$ as above, we let $f_i$ be a monotone submodular function defined over a set $S$ of driver locations as $f_i(S)=\sum_{u\in T_{a_i}}\max_{r\in S}c(u,r)$. We pick 100,000 locations at random from the September 2014 Uber pick-up locations as a ground set. For training we form $m=50$ tasks by picking for each task a random time in the *first* week of Sept. 2014. We test on $m=50$ new tasks formed similarly from the *second* week of Sept. 2014 and report in the figures the average performance obtained at test tasks. Figures \[fig:simulation-1-rideshare\] and \[fig:simulation-2-rideshare\] show the performance of our proposed algorithms against the baselines mentioned above. Figure shows the performance of all algorithms when we fix $k=20$, and vary $l$ from 5 to 18. Larger $l$ means less computation at test time (since we need to further choose $k-l$ elements at test). However, we see that even for large values of $l$ (e.g. $l=16$), the performance of Meta-Greedy is still quite close to the ideal performance of Greedy-Test. Putting this together with the fact that the performance of Greedy-Train is not so good, we can conclude that adding a few personalized elements at test time significantly boosts performance to be even close to the ideal. In Figure , we compare the performance of all the algorithms when $k$ changes from 5 to 30, and $l$ is $80\%$ of $k$ ($l = \lfloor 0.8 k \rfloor$). As we can see, even when we just learn $20\%$ of the set in test time, the performance of Meta-greedy is close to Test-Greedy. Also, when $k-l$ increases, Random-Meta-Greedy performs better than Meta-Greedy. This is in compliance with the results of Theorems \[thm:meta-greedy\], \[thm: randomized meta\].
**Movie Recommendation.** In this application, we use the Movielens dataset [@harper2015movielens] which consists of $10^6$ ratings (from 1 to 5) by $6041$ users for $4000$ movies. We pick the 2000 most rated movies, and 200 users who rated the highest number of movies (similar to [@stan2017probabilistic]). We partitioned the 200 users into 100 users for the training phase and 100 other users for the test phase. Each movie can belong to one of 18 genres. For each genre $t$ we let $G_t$ be the set of all movies with in genre $t$. For each user $i$, we let $R_i$ be the set of all movie rated by the user, and for each movie $v \in R_i$ the corresponding rating is denoted by $r_i(v)$. Furthermore, for user $i$ we define $ f_i(S)=\sum_{t=1}^{18} w_{i,t}.\max_{v\in R_i\cap G_t \cap S} r_i(v)$ which is the weighted average over maximum rate that user $i$ gives to movies from each genre and $w_{i,t}$ is proportion of movies in genre $t$ which is rated by user $i$ out of all the rating he provides. A task ${\mathcal{T}}_i$ involves 5 users $i_1, \cdots, i_5$ and the function assigned to the task is the average of $f_{i_1}, \cdots, f_{i_5}$. We formed $m=50$ training tasks from the users in the training phase, and $m=50$ test tasks from the users in the test phase. Figure (resp. \[fig:simulation-2-movrecom\]) has been obtained in a similar format as Figure \[fig:simulation-1-rideshare\] (resp. Figure \[fig:simulation-2-rideshare\]). We observe a very similar pattern as in the ride share experiments.
![Comparison of two-stage framework and submodular meta-learning framework []{data-label="fig:twostage"}](rideshare_twostage.eps){width="8cm"}
Comparison with Two-stage Submodular Optimization {#sec:two_stage}
=================================================
Two-stage submodular optimization is another way to deal with limited computational power in test time. In this framework, at training time, a reduced ground set is learned which will be used as a ground set at test time. This procedure will reduce the computational time in test time. More formally, the two-stage submodular optimization framework aims to solve the following problem. Let $f_i:2^{\mathcal{X}}\xrightarrow{}{\mathbb{R}}_+$ for $i\in [m]$, be a monotone submodular function over ground set $V$. The goal is to find $S$ with size at most $q$ whose subests of size $k$ maximize the sum of $f_i$ for $i\in [m]$: $$\label{eq:two-stage submodular} \max_{S\subseteq {\mathcal{X}}, \mid S \mid\leq q}\; \frac{1}{m}\sum_{i=1}^{m}\; \max_{S_i\subseteq S, \mid S_i\mid \leq k}\; f_i(S_i)$$ Once the set $S$ is found, it will be used in the test phase (e.g., by running full greedy on $S$ as the reduced ground set) to find $k$ elements for a new task. This framework uses ${\mathcal{O}}(qk)$ function evaluations for each new test task; however, it poorly personalizes to a test task because the set $S$ has been optimized only for the tasks at the training time. This intuition is indeed consistent with our experimental findings reported below. We further remark that the two-stage framework requires very high computational power in training. For example, the Replacement-Greedy algorithm [@stan2017probabilistic] requires computational complexity ${\mathcal{O}}(qkmn)$ (which is a factor $q$ larger than the complexity of the algorithms in this paper). As a result of this issue, we were not able to run the state-of-the-art two-stage algorithms to solve in the setting considered in our main simulation results (presented in Section \[sec:Simulation\]). e.g., for ground set of size $n = 10^5$ our two-stage implementation would take a very long time.
We have considered the ride-sharing application discussed in Section \[sec:Simulation\] and let $n=500$ (ground set size), $m=50$ (number of tasks), and $k$ changing from 5 to 30 (cardinality constraint) while $l=80\%k$ (portion that will fill in the submodular meta-learning during training), and $q=100$ (size of reduced ground set for two-stage framework). For solving the two-stage problem we have used the Replacement-Greedy algorithm introduced in [@stan2017probabilistic]. We choose these parameters based on the following two facts:
1. Because of the high computational cost of the Replacement Greedy algorithm in training for the ride-sharing application, we chose $n$ to be 500.
2. We provide a fair comparison in terms of computational power at test time, which means both Meta-Greedy (our algorithm) and Replacement-Greedy have exactly *the same computational cost* at test time. Formally, $n(k-l)=qk$.
we report the result for the above setting in the Figure \[fig:twostage\]. A few comments are in order: (i) The two stage implementation reduces the ground set of size $n = 500$ to $q = 100$. When $k$ is small, some of the popular elements found at training time would be good enough to warrant a good performance at test time. However, when $k$ increases, the role of personalizing becomes more apparent. As we see, the performance of Replacement-Greedy does not improve much when we increase $k$ and it is close to the performance of Greedy-Train (which chooses all the $k$ elements during the training phase–see and the discussion therein). However, since Meta-Greedy does (a small) task-specific optimization at test time, its performance becomes much better. We emphasize again that, in order to be fair, the comparison in Figure \[fig:twostage\] has been obtained using *the same* computational power allowed at test time for both meta-learning and two-stage approaches.
Conclusion and Future Work
==========================
In this paper, we extended the notion of Model-Agnostic Meta-Learning (MAML) to discrete optimization and in particular to submodular maximization. We proposed a novel formulation in which we aim to find an initial solution set that can be quickly adapted to a new task at a relatively low computational cost. In our meta-learning framework, the process of selecting set items is done in two parts: In the first part, a set of items are selected offline according to prior experience and data. In the second part, which happens at test time, a set of elements that is personalized to the new revealed task is selected. For the proposed problem, we introduced a deterministic variant of the greedy algorithm which obtains a solution that is at least $0.53$-optimal, when the tasks are monotone and submodular. We further presented a randomized algorithm that improves this result and obtains $(1-1/e-o(1))$-approximation in expectation. We also studied the performance of our proposed meta-learning framework and algorithms for two real-world applications: movie recommendation and ride-sharing problems. Our numerical results indicate the advantage of our proposed scheme with respect to traditional learning procedures as well as methods based on two-stage submodular optimization.
There are numerous open directions that can be investigated along the lines of discrete meta-learning and user-specific adaptation for discrete problems (indeed, this work can be considered as a first step). Examples include extending the results to a more general setting when the tasks are (approximately) submodular but non-monotone, considering the case that the tasks at training and test times are drawn according to two different probability distributions (possibly with bounded distance), and desinig
Further, exploring meta-learning continuous extensions and
Acknowledgements {#acknowledgements .unnumbered}
================
The research of Arman Adibi and Hamed Hassani is supported by NSF award CPS-1837253, NSF CAREER award CIF 1943064, and Air Force Office of Scientific Research Young Investigator Program (AFOSR-YIP) under award FA9550-20-1-0111.
Appendix {#appendix .unnumbered}
========
Proof of Proposition \[alg1lemma\] {#sec:proof_prop_1}
==================================
First, we prove lemma which will help us to show .
\[lemma:greedy procedure\] Consider the monotone submodular function $f$ where and the greedy procedure for maximizing it starting from $S^{'}$ and in $t^{th}$ step adding the element with maximum marginal gain $e^{(t)}$ to $S^{(t-1)}$ which means $e^{(t)}=\operatorname*{arg\,max}\limits_e f(S^{(t-1)}\cup e \cup S^{'})-f(S^{(t-1)}\cup S^{'})$, then the following holds for every set $T$ and $S^{''}$ where $S^{'}\subseteq T$ and $\mid T\mid=k$. $$\label{eq: eq1 pf lemma:greedy}
f(S{''}\cup T ) -f( S^{(k)}\cup S{''})\leq f( S^{(k)}\cup S^{'})- f(S^{'})$$
To show define $J^{(t)}$ iteratively as follows. Start from $J^{(0)}=T$ and let $D^{(t)}=J^{(t-1)}\setminus S^{(t-1)}$ and define $o^{(t)}$ in the following way:
1. If $e^{(t)}\in D^{(t)}$, then let $o^{(t)}=e^{(t)}$.
2. Otherwise, if $e^{(t)}\notin D^{(t)}$, let $o^{(t)}$ be one of the elements of $D^{t}$ chosen uniformly at random.
Define $J^{(t)}:=J^{(t-1)}\cup e^{(t)}\setminus o^{(t)}$.\
(TS\^[”]{}) (TJ\^[(1)]{}) … (TJ\^[(t)]{})\
(S\^[(0)]{}S\^[’]{}) (S\^[(1)]{} S\^[’]{}) … (S\^[(t)]{}S\^[’]{})
\
then we can write the following inequalities: $$\begin{aligned}
f(S^{(t)}\cup S^{'})-f(S^{(t-1)}\cup S^{'})
&= f(S^{(t-1)}\cup e^{(t)}\cup S^{'})-f(S^{(t-1)}\cup S^{'})
\\\label{eq: chain1 l1:greedy}&\geq f(S^{(t-1)}\cup S^{'}\cup o_i^{(t)})-f(S^{(t-1)}\cup S^{'})
\\\label{eq: chain2 l1:greedy}&\geq f(J^{(t-1)}\cup S^{'})-f(J^{(t-1)}\cup S^{'}\setminus o^{(t)})\\&\geq f(J^{(t-1)}\cup S^{''})-f(J^{(t-1)}\cup S^{''}\setminus o^{(t)})
\\&\geq f(J^{(t-1)}\cup S^{''})-f(J^{(t-1)}\cup S^{''}\setminus o^{(t)}) \nonumber\\
&\quad-f(S^{''}\cup J^{(t-1)}\cup e^{(t)}\setminus o^{(t)})+f(S^{''}\cup J^{(t-1)}\setminus o^{(t)})\label{eq: chain3 l1:greedy}\\&= f(J^{(t-1)}\cup S^{''})-f(J^{(t-1)}\cup S^{''}\setminus o^{(t)}) \nonumber\\
&\quad-f(S^{''}\cup J^{(t)})+f(S^{''}\cup J^{(t-1)}\setminus o^{(t)})
\\&=f(J^{(t-1)}\cup S^{''})-f(S^{''}\cup J^{(t)})\end{aligned}$$ where follows from $f(S^{(t-1)}\cup e^{(t)} \cup S^{'})-f(S^{(t-1)}\cup S^{'})=\max\limits_e f(S^{(t-1)}\cup e \cup S^{'})-f(S^{(t-1)}\cup S^{'})\geq f(S^{(t-1)}\cup o^{(t)} \cup S^{'})-f(S^{(t-1)}\cup S^{'})$ and follows from submodularity since in each step $S^{(t-1)}\subseteq J^{(t-1)}$ and $o^{(t)}\not\in S^{(t-1)}$. Finally, equation follows from the fact that $-f(J^{(t-1)}\cup e^{(t)}\cup S^{''}\setminus o^{(t)})+f(J^{(t-1)}\cup S^{''}\setminus o^{(t)})\leq 0$. Then, by summing over the above inequality we get the following inequality: $$\begin{aligned}
f(S^{(t)}\cup S^{'})-f(S^{(0)}\cup S^{'})&=\sum_{i=0}^{t}f(S^{(i)}\cup S^{'})-f(S^{(i-1)}\cup S^{'})
\\&\geq \sum_{i=0}^{t} f(S^{''}\cup J^{(i-1)})-f(S^{''}\cup J^{(i)})
\\&=f(S^{''}\cup J^{(0)})-f(S^{''}\cup J^{(t)})
\\&\geq f(S^{''}\cup T)-f(S^{''}\cup J^{(t)})\end{aligned}$$ Note because of the definition of $J^{(t)}$ after k step $J^{(k)}=S^{(k)}$; therefore, we can conclude that: $$\begin{aligned}
f(S^{(k)}\cup S^{'})-f( S^{'})&\geq f(S^{''}\cup T)-f(S^{''}\cup J^{(t)})\notag\\&=f(S^{''}\cup T)-f(S^{''}\cup S^{(k)})\end{aligned}$$ .
Let ${S_{\rm{tr}}}$, $\{S_i\}_{i=1}^{m}$ be the output of Algorithm \[alg: discrete meta-Greedy\] and ${S_{\rm{tr}}}^{*}$, $\{S_i^{*}\}_{i=1}^{m}$ be the optimal solution for problem . We first show that the output of Algorithm \[alg: discrete meta-Greedy\] in phase 1 satisfies the following inequality: $$\label{eq: eq1 pf l1}
\sum_{i=1}^m f_i({S_{\rm{tr}}}^{*} \cup S_i^{*}) - \sum_{i=1}^m f_i( {S_{\rm{tr}}})\leq \sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i^{*})$$
To show let $e^{(t)}$ be the ${t}^{th}$ element of greedy procedure in phase 1, and ${S_{\rm{tr}}}^{(t)}$ be the $t^{th}$ set in this procedure, where $e^{(t)}=\operatorname*{arg\,max}\limits_e \sum\limits_{i=1}^{m} f_i({S_{\rm{tr}}}^{(t-1)}\cup e)-f_i({S_{\rm{tr}}}^{(t-1)})$. let $J^{(0)}={S_{\rm{tr}}}^{*}$ and define $J^{(t)}$ iteratively as follows. Let $D^{(t)}=J^{(t-1)}\setminus {S_{\rm{tr}}}^{(t-1)}$ and define $o^{(t)}$ in the following way:
1. If $e^{(t)}\in D^{(t)}$, then let $o^{(t)}=e^{(t)}$.
2. Otherwise, if $e^{(t)}\notin D^{(t)}$, let $o^{(t)}$ be one of the elements of $D^{t}$ chosen uniformly at random.
Define $J^{(t)}:=J^{(t-1)}\cup e^{(t)}\setminus o^{(t)}$. We show this procedure in the following chain.\
([S\_[[tr]{}]{}]{}\^[\*]{},{S\_i\^[\*]{}}\_[i=1]{}\^[m]{}) (J\^[(1)]{},{S\_i\^[\*]{}}\_[i=1]{}\^[m]{}) … (J\^[(l)]{},{S\_i\^[\*]{}}\_[i=1]{}\^[m]{})\
([S\_[[tr]{}]{}]{}=,{S\_i\^0}\_[i=1]{}\^[m]{}=) ([S\_[[tr]{}]{}]{}\^[(1)]{},{}\_[i=1]{}\^[m]{}) … ([S\_[[tr]{}]{}]{}\^[(l)]{},{}\_[i=1]{}\^[m]{})
\
then we can write the following inequalities: $$\begin{aligned}
\sum_{i=1}^m f_i({S_{\rm{tr}}}^{(t)})-f_i({S_{\rm{tr}}}^{(t-1)})
&= \sum_{i=1}^m f_i({S_{\rm{tr}}}^{(t-1)}\cup e^{(t)})-f_i({S_{\rm{tr}}}^{(t-1)})
\\\label{eq: chain1 l1}&\geq \sum_{i=1}^m f_i({S_{\rm{tr}}}^{(t-1)}\cup o_i^{(t)})-f_i({S_{\rm{tr}}}^{(t-1)})
\\\label{eq: chain2 l1}&\geq \sum_{i=1}^m f_i(S_{i}^{*}\cup J^{(t-1)})-f_i(S_{i}^{*}\cup J^{(t-1)}\setminus o^{(t)})
\\&\geq \sum_{i=1}^m f_i(S_i^{*}\cup J^{(t-1)})-f_i(S_i^{*}\cup J^{(t-1)}\setminus o_i^{(t)}) \nonumber\\
&\quad+\sum_{i=1}^m -f_i(S_{i}^{*}\cup J^{(t)})+f_i(S_{i}^{*}\cup J^{(t-1)}\setminus o_i^{(t)})\label{eq: chain3 l1}
\\&=\sum_{i=1}^m f_i(S_{i}^{*}\cup J^{(t-1)})-f_i(S_{i}^{*}\cup J^{(t)})\end{aligned}$$ where follows from definition of $e^{(t)}$ and the greedy procedure and follows from submodularity since in each step ${S_{\rm{tr}}}^{(t-1)}\subseteq J^{(t-1)}$ and $o^{(t)}\not\in {S_{\rm{tr}}}^{(t-1)}$ and finally, equation follows from the fact that $-f_i(J^{(t)}\cup S_i^{*})+f_i(J^{(t-1)}\cup S_i^{*}\setminus o^{(t)})\leq 0$. Then, by summing over $t$ from 0 to $l$ we get the following inequality: $$\begin{aligned}
\sum_{i=1}^m f_i({S_{\rm{tr}}})=\sum_{i=1}^m f_i({S_{\rm{tr}}}^{(l)})-f_i({S_{\rm{tr}}}^{(0)})&=\sum_{i=1}^m\sum_{t=0}^{l}f_i({S_{\rm{tr}}}^{(t)})-f_i({S_{\rm{tr}}}^{(t-1)})
\\&\geq \sum_{i=1}^m\sum_{t=0}^{l}f_i(S_{i}^{*}\cup J^{(t-1)})-f_i(S_{i}^{*}\cup J^{(t)})
\\&=\sum_{i=1}^m f_i(S_{i}^{*}\cup J^{(0)})-f_i(S_{i}^{*}\cup J^{(l)})
\\&=\sum_{i=1}^m f_i(S_{i}^{*}\cup {S_{\rm{tr}}}^{*})-f_i(S_{i}^{*}\cup {S_{\rm{tr}}})\end{aligned}$$ where the last equality comes from the process of defining $J$. Because, we only change one element by adding element found in greedy process and removing one element from the optimal set in each step and the size of $J^{(t)}$ is $l$ in each step; therefore, after $l$ step $J^{(l)}={S_{\rm{tr}}}$. By rearranging the terms and summing over $i$ the claim in follows.\
\
Second, for the phase 2 of the algorithm \[alg: discrete meta-Greedy\] we can use the usual analysis of greedy[@krause2014submodular] for set $S_i$: $$\begin{aligned}
\sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i)-f_i( {S_{\rm{tr}}})&\geq (1-\frac{1}{e})(\sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i^{opt})-f_i({S_{\rm{tr}}}))
\label{eq: eq3 pf l1 1} \\&\geq (1-\frac{1}{e})(\sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i^{*})-f_i({S_{\rm{tr}}}))
\label{eq: eq3 pf l1 2}\\&\geq (1-\frac{1}{e})(\sum_{i=1}^m f_i({S_{\rm{tr}}}^{*}\cup S_i^{*})-2f_i({S_{\rm{tr}}}))
\label{eq: eq3 pf l1 3}\end{aligned}$$ where $S_i^{opt}=\operatorname*{arg\,max}\limits_{|S_i|\leq k-l}f_i({S_{\rm{tr}}}\cup S_i)$ in the equation . Equation follows from usual greedy analysis, equation follows from definition of ${S_{\rm{tr}}}^{opt}$, and equation follows from equation .
Now divide both sides of by $1/m$ and regroup the terms to obtain $$\label{eq:result_alg_1_part_1}
\frac{1}{m} \sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i) \geq \left(1-\frac{1}{e}\right)({\rm{OPT}}-2\beta)+\beta,$$ where $\beta := \frac{1}{m}\sum_{i=1}^m f_i({S_{\rm{tr}}}).$
Finally, since $S_i\subseteq S_i\cup {S_{\rm{tr}}}$ by monotonicity [[$f_i(S_i\cup {S_{\rm{tr}}})\geq f_i({S_{\rm{tr}}})$]{}]{}. Then, combing this observation with the result in implies $$\frac{1}{m}\sum_{i=1}^m f_i ({S_{\rm{tr}}}\cup S_i) \geq \max \biggl\{ \beta \,, \, (1-1/e)({\rm{OPT}} - 2\beta) + \beta \biggr \}.$$
Proof of Proposition \[alg2lemma\] {#sec:proof_prop_2}
==================================
Let ${S_{\rm{tr}}}$, $\{S_i\}_{i=1}^{m}$ be the output of Algorithm \[alg: Reverse discrete meta-Greedy\] and ${S_{\rm{tr}}}^{*}$, $\{S_i^{*}\}_{i=1}^{m}$ be the optimal solution for problem . We first show the following about the output of algorithm \[alg: Reverse discrete meta-Greedy\], phase 1.
$$\label{eq: eq1 pf l2}
\sum_{i=1}^m f_i({S_{\rm{tr}}}^{*} \cup S_i^{*}) - \sum_{i=1}^m f_i( S_i)\leq \sum_{i=1}^m f_i({S_{\rm{tr}}}^{*} \cup S_i)$$
to show consider the following: let $e_i^{(t)}=\operatorname*{arg\,max}\limits_e f_i(S_i^{(t-1)}\cup e)-f_i(S_i^{(t-1)})$. let $J^{(0)}_i=S_i^{*}$ and define $J^{(t)}_i$ iteratively as follows. Let $D^{t}_{i}=J^{(t-1)}_i\setminus S^{(t-1)}_i$ and define $o_i^{(t)}$ in the following way:
1. If $e_i^{(t)}\in D^{t}_{i}$, then $o^{(t)}=e_i^{(t)}$;
2. Otherwise, if $e_i^{(t)}\notin D^{t}_{i}$, let $o_i^{(t)}$ be one of the elements of $D^{t}_{i}$ chosen uniformly at random;
Define $J^{(t)}_i:=J^{(t-1)}_i\cup e_i^{(t)}\setminus o_i^{(t)}$.\
([S\_[[tr]{}]{}]{}\^[\*]{},{S\_i\^[\*]{}}\_[i=1]{}\^[m]{}) ([S\_[[tr]{}]{}]{}\^[\*]{},{J\_i\^[(1)]{}}\_[i=1]{}\^[m]{}) … ([S\_[[tr]{}]{}]{}\^[\*]{},{J\_i\^[(k-l)]{}}\_[i=1]{}\^[m]{})\
([S\_[[tr]{}]{}]{}=,{S\_i\^0}\_[i=1]{}\^[m]{}=) (,{S\_i\^[(1)]{}}\_[i=1]{}\^[m]{}) … (,{S\_i\^[(k-l)]{}}\_[i=1]{}\^[m]{})
\
then we can write the following inequalities: $$\begin{aligned}
f_i(S_i^{(t)})-f_i(S_i^{(t-1)})
&= f_i(S_i^{(t-1)}\cup e_i^{(t)})-f_i(S_i^{(t-1)})
\\\label{eq: chain1}&\geq f_i(S_i^{(t-1)}\cup o_i^{(t)})-f_i(S_i^{(t-1)})
\\\label{eq: chain2}&\geq f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)})-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)}\setminus o_i^{(t)})
\\&\geq f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)})-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)}\setminus o_i^{(t)}) \nonumber\\
&\quad-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t)})+f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)}\setminus o_i^{(t)})\label{eq: chain3}
\\&=f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)})-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t)})\end{aligned}$$ where follows from definition of $e_i^{(t)}$ and the greedy procedure and follows from the submodularity since in each step $S_i^{(t-1)}\subseteq J_i^{(t-1)}$ and $o_i^{(t)}\not\in S_i^{(t-1)}$ and finally, equation follows from the fact that $-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t)})+f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)}\setminus o_i^{(t)})\leq 0$ because of monotonicity. Then, by summing over $t$ from 0 to $k-l$ we get the following inequality: $$\begin{aligned}
f_i(S_i)=f_i(S_i^{(k-l)})-f_i(S_i^{(0)})&=\sum_{t=0}^{k-l}f_i(S_i^{(t)})-f_i(S_i^{(t-1)})
\\&\geq \sum_{t=0}^{k-l}f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t-1)})-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(t)})
\\&=f_i({S_{\rm{tr}}}^{*}\cup J_i^{(0)})-f_i({S_{\rm{tr}}}^{*}\cup J_i^{(k-l)})
\\&=f_i({S_{\rm{tr}}}^{*}\cup S_i^{*})-f_i({S_{\rm{tr}}}^{*}\cup S_i)\end{aligned}$$ where the last equality comes from the process of defining $J_i^{(k-l)}$; since, the size of $J_i^{(t)}$ is $k-l$ in each step and after $k-l$ step $J_i^{(k-l)}=S_i$. Then, by rearranging and summing over $i$ we can obtain .\
\
Second, for phase 2 of algorithm \[alg: Reverse discrete meta-Greedy\] we can use the usual analysis of greedy[@krause2014submodular] for set ${S_{\rm{tr}}}$ : $$\begin{aligned}
\sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i)-f_i( S_i)&\geq (1-\frac{1}{e})(\sum_{i=1}^m f_i({S_{\rm{tr}}}^{opt}\cup S_i)-f_i(S_i))
\label{eq: eq3 pf l2 1} \\&\geq (1-\frac{1}{e})(\sum_{i=1}^m f_i({S_{\rm{tr}}}^{*}\cup S_i)-f_i(S_i))
\label{eq: eq3 pf l2 2}\\&\geq (1-\frac{1}{e})(\sum_{i=1}^m f_i({S_{\rm{tr}}}^{*}\cup S_i^{*})-2f_i(S_i))
\label{eq: eq3 pf l2 3}\end{aligned}$$ where ${S_{\rm{tr}}}^{opt}=\operatorname*{arg\,max}\limits_{|{S_{\rm{tr}}}|\leq l}\sum\limits_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i)$ in equation . Equation follows from the usual greedy analysis, equation follows from the definition of ${S_{\rm{tr}}}^{opt}$, and equation follows the from equation .
Now divide both sides of by $1/m$ and regroup the terms to obtain $$\label{eq:result_alg_2_part_1}
\frac{1}{m} \sum_{i=1}^m f_i({S_{\rm{tr}}}\cup S_i) \geq \left(1-\frac{1}{e}\right)({\rm{OPT}}-2\gamma)+\gamma,$$ where $\gamma := \frac{1}{m}\sum_{i=1}^m f_i(S_i).$
Finally, since $S_i\subseteq S_i\cup {S_{\rm{tr}}}$ by monotonicity $f_i(S_i\cup {S_{\rm{tr}}})\geq f_i(S_i)$. Then, by combing this result wit we obtain $$\frac{1}{m}\sum_{i=1}^m f_i ({S_{\rm{tr}}}\cup S_i) \geq \max \biggl\{ \gamma \,, \, (1-1/e)({\rm{OPT}} - 2\gamma) + \gamma \biggr \}.$$
The following shows the ratio of lower bound to optimum (a similar plot can be obtained for the lower bound of Proposition \[alg1lemma\] when $\gamma$ is replaced with $\beta$.). As we observe, in the worst case, the approximation factor is $0.5$.
![y-axis: The lower bound of Proposition \[alg2lemma\] divided by OPT, x-axis: $\gamma/\text{OPT}$.[]{data-label="fig:bound"}](boundplot.eps){width="8cm"}
Proof of Theorem \[thm:meta-greedy\] {#sec:proof_of_main_thm}
====================================
Let $\theta_2=\frac{1}{m}\sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(2)} \cup S_i^{(2)})$. Since ${S_{\rm{tr}}}^{(2)}$ found greedily given $\{S_i\}_{i=1}^{m}$ we can write:
$$\label{eq: thm1 pf eq1}
\theta_2-\gamma\geq ({\rm{OPT}}-\gamma)(1-\frac{1}{e})\geq( \frac{1}{m} \sum_{i=1}^{m} f_i(S^{'} \cup S_i^{(2)})-\gamma)(1-\frac{1}{e})$$
for every $\mid S^{'}\mid\leq l$. Also, we can write $$\begin{aligned}
{\rm{OPT}}-\gamma&= \frac{1}{m}\sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{*})-f_i(S_i^{(2)})
\label{eq:tm1 pf align 1} \\&\leq \frac{1}{m}\sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{(2)} \cup S_i^{*})-f_i(S_i^{(2)})
\label{eq:tm1 pf align 2} \\&= \frac{1}{m}\sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{(2)} \cup S_i^{*})+f_i({S_{\rm{tr}}}^{*} \cup S_i^{(2)})-f_i({S_{\rm{tr}}}^{*} \cup S_i^{(2)})-f_i(S_i^{(2)})
\label{eq:tm1 pf align 3} \\&\leq \frac{1}{m}\sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{(2)} \cup S_i^{*})-f_i({S_{\rm{tr}}}^{*} \cup S_i^{(2)})+\frac{\theta_2-\gamma}{1-1/e}
\label{eq:tm1 pf align 4}\\&\leq \frac{1}{m}\sum_{i=1}^{m}f_i( S_i^{(2)} \cup S_i^{*})-f_i( S_i^{(2)})+\frac{\theta_2-\gamma}{1-1/e}
\label{eq:tm1 pf align 5}\end{aligned}$$ where comes from , and comes from submodularity. We thus obtain $$\label{eq: thm l1 1}
{\rm{OPT}} -\frac{\theta_2-\gamma}{1-1/e}-\gamma \leq \frac{1}{m}\sum_{i=1}^{m}f_i( S_i^{(2)} \cup S_i^{*})-f_i( S_i^{(2)})$$ Also we can write for any set $S'$ such that $\mid S^{'}\mid\leq l$: $$\begin{aligned}
\label{eq:tm1 pf align2 1}
\frac{1}{m}\sum_{i=1}^{m}f_i(S^{'}\cup S_i^{*})-f_i(S^{'})
&\geq \frac{1}{m}\sum_{i=1}^{m}f_i(S^{'}\cup S_i^{*}\cup S_i)-f_i(S^{'}\cup S_i)\\
\label{eq:tm1 pf align2 2}
&\geq \frac{1}{m}\sum_{i=1}^{m}f_i(S^{'}\cup S_i^{*}\cup S_i)-f_i(S_i)+f_i(S_i)-f_i(S^{'}\cup S_i)\\ \label{eq:tm1 pf align2 3}
&\geq \frac{1}{m} \sum_{i=1}^{m}f_i( S_i^{*}\cup S_i)-f_i(S_i)+f_i(S_i)-f_i(S^{'}\cup S_i)\\
\label{eq:tm1 pf align2 4}
&\geq {\rm{OPT}} -\frac{\theta_2-\gamma}{1-1/e}-\gamma+ \frac{1}{m} \sum_{i=1}^{m}f_i(S_i)-f_i(S^{'}\cup S_i)\\
\label{eq:tm1 pf align2 5}
&\geq {\rm{OPT}} -2\frac{\theta_2-\gamma}{1-1/e}-\gamma\end{aligned}$$ where follows from submodularity, follows from monotonicity, and follows from , and follows from . This results the following for any set $S'$ such that $|S^{'}|\leq l$: $$\label{eq: thm l1 2}
\frac{1}{m} \sum_{i=1}^{m}f_i(S^{'}\cup S_i^{*})-f_i(S^{'}) \geq {\rm{OPT}} -2\frac{\theta_2-\gamma}{1-1/e}-\gamma$$ Now, from we can find a new bound for the performance of algorithm \[alg: main discrete meta-Greedy\]. From we can write:
$$\label{eq: thm1 pf eq2}
\frac{1}{m} \sum_{i=1}^{m}f_i({S_{\rm{tr}}}^{(1)}\cup S_i^{*})-f_i({S_{\rm{tr}}}^{(1)}) \geq {\rm{OPT}} -2\frac{\theta_2-\gamma}{1-1/e}-\gamma$$
Also, since in Algorithm 1 the set $S_i^{(1)}$ is constructed greedily on the top of ${S_{\rm{tr}}}^{(1)}$, we have: $$\begin{aligned}
\label{eq: thm1 pf eq3}
\frac{1}{m} \sum_{i=1}^{m} f_i({S_{\rm{tr}}}^{(1)} \cup S_i^{(1)})-\beta
&\geq( \frac{1}{m} \sum_{i=1}^{m} f_i({S_{\rm{tr}}}^{(1)} \cup S_i^{*})-\beta)(1-\frac{1}{e})\\
&\geq ( {\rm{OPT}} -2\frac{\theta_2-\gamma}{1-1/e}-\gamma)(1-\frac{1}{e}),\label{eq: thm1 pf eq3}\end{aligned}$$ where follows from . We thus obtain:
$$\begin{aligned}
\label{eq: thm1 pf eq4}
\frac{1}{m} \sum_{i=1}^{m} f_i({S_{\rm{tr}}}^{(1)} \cup S_i^{(1)})
\geq ( {\rm{OPT}} -2\frac{\theta_2-\gamma}{1-1/e}-\gamma)(1-\frac{1}{e})+\beta\end{aligned}$$
Using the same procedure as above, by defining $\theta_1= \frac{1}{m} \sum_{i=1}^{m} f_i({S_{\rm{tr}}}^{(1)} \cup S_i^{(1)})$, we can prove: $$\begin{aligned}
\label{eq: thm1 pf eq5}
\frac{1}{m} \sum_{i=1}^{m} f_i({S_{\rm{tr}}}^{(2)} \cup S_i^{(2)})
\geq ( {\rm{OPT}} -2\frac{\theta_1-\gamma}{1-1/e}-\gamma)(1-\frac{1}{e})+\beta\end{aligned}$$ which results in the following lower bound: $$\begin{aligned}
&\max \biggl\{ \frac{1}{m} \sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(1)} \cup S_i^{(1)}) \, , \, \frac{1}{m}\sum_{i=1}^m f_i ({S_{\rm{tr}}}^{(2)} \cup S_i^{(2)}) \biggr\} \nonumber\\
& \geq \max \biggl\{\theta_1, \theta_2, (1-1/e)({\rm{OPT}} \!-\!\gamma) + \beta -2 (\theta_2-\gamma) ,(1-1/e)({\rm{OPT}} \!-\!\beta) + \gamma -2 (\theta_1-\beta) \biggr\}.\label{eq:thm p6}\end{aligned}$$
Finally, given and , the factor $0.53$ is obtained as a result of the following procedure. Let $\beta$ and $\gamma$ given as $\beta := \frac{1}{m} \sum_{i=1}^m f_i({S_{\rm{tr}}}^{(1)})$ and $\gamma := \frac{1}{m} \sum_{i=1}^m f_i(S_i^{(2)})$. Then the left-hand-side term in is lower bounded by:
\[eq: thm1 opt\] $$\begin{aligned}
{4}
\min\limits_{\theta_1, \theta_2} & \,\, \max \biggl\{\theta_1, \theta_2, (1-1/e)({\rm{OPT}} \!-\!\gamma) + \beta -2 (\theta_2-\gamma) , (1-1/e)({\rm{OPT}} \!-\!\beta) + \gamma -2 (\theta_1-\beta) \biggr\}
\\
& \operatorname*{subject\,\, to \quad}\theta_1 \geq \max\{\beta, (1-1/e)({\rm{OPT}} - 2\beta) + \beta\} \\
& \!\qquad \quad \quad \quad \quad \theta_2 \geq \max\{\gamma, (1-1/e)({\rm{OPT}} - 2\gamma) +\gamma\} \end{aligned}$$
Note that the constraints hold due to the results of Proposition 1 and 2. In particular, the above bound is always larger than $ 0.53 \times \rm{OPT} $ for any value of $\beta$ and $\gamma$.
Proof of Theorem \[thm: randomized meta\] {#sec:proof_of_main_random_theorem}
=========================================
Consider round $t$ in which $\mid {S_{\rm{tr}}}\mid<l$ and $\mid S_i \mid<k-l$ the expected gain of the algorithm with probability $\frac{l}{k}$ is the maximum gain from adding an element $e^{*}=\operatorname*{arg\,max}\limits_e \sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{t}\cup{e}\cup S_i^{t})-f_i({S_{\rm{tr}}}^t\cup S_i^t)$ or with probability $\frac{k-l}{k}$ the gain is $\sum\limits_{i=1}^{m}\max_{e_i}f_i({S_{\rm{tr}}}^{t}\cup{e_i} \cup S_i^{t})-f_i({S_{\rm{tr}}}^t\cup S_i^t)$ which can be written as follows.
$$\begin{aligned}
\mathbb{E}[&\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{t+1}\cup S_i^{t+1})-f_i({S_{\rm{tr}}}^t\cup S_i^t)|{S_{\rm{tr}}}^{t},S_i^t]\nonumber\\=&\frac{l}{k}\max_e\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{t}\cup{e}\cup S_i^{t})-f_i({S_{\rm{tr}}}^t\cup S_i^t)+\frac{k-l}{k}\sum\limits_{i=1}^{m}\max_{e_i}f_i({S_{\rm{tr}}}^{t}\cup{e_i} \cup S_i^{t})-f_i({S_{\rm{tr}}}^t\cup S_i^t)\label{eq: rand alg pf eq begin}\end{aligned}$$
assuming ${S_{\rm{tr}}}^* , S_i^{*}$ is optimal solution, we can also write: $$\begin{aligned}
\frac{1}{k}\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{*})-f_i({S_{\rm{tr}}}^t\cup S_i^t)
&\leq \frac{1}{k}\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{*}\cup {S_{\rm{tr}}}^t\cup S_i^t)-f_i({S_{\rm{tr}}}^t\cup S_i^t)
\label{eq: rand alg pf eq 1} \\&\leq \frac{1}{k}\sum_{e\in {S_{\rm{tr}}}^{*}\setminus {S_{\rm{tr}}}^t}\sum\limits_{i=1}^{m}f_i({e} \cup {S_{\rm{tr}}}^t\cup S_i^t)-f_i( {S_{\rm{tr}}}^t\cup S_i^t)
\nonumber\\&+\frac{1}{k}\sum\limits_{i=1}^{m}\sum_{e\in S_i^{*}\setminus S_i^t}f_i({e} \cup {S_{\rm{tr}}}^t\cup S_i^t)-f_i( {S_{\rm{tr}}}^t\cup S_i^t)
\label{eq: rand alg pf eq 2} \\&\leq \frac{l}{k}\max_e\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{t}\cup{e}\cup S_i^{t})-f_i({S_{\rm{tr}}}^t\cup S_i^t)
\nonumber\\&+\frac{k-l}{k}\sum\limits_{i=1}^{m}\max_{e_i}f_i({S_{\rm{tr}}}^{t}\cup{e_i} \cup S_i^{t})-f_i({S_{\rm{tr}}}^t\cup S_i^t) \label{eq: rand alg pf eq 3}\end{aligned}$$ where follows from monotonicity, and follows from submodularity. Then, from and we conclude that: $$\begin{aligned}
\mathbb{E}[&\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{t+1}\cup S_i^{t+1})-f_i({S_{\rm{tr}}}^t\cup S_i^t)|{S_{\rm{tr}}}^{t},S_i^t]\leq\frac{1}{k}\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{*}\cup S_i^{*})-f_i({S_{\rm{tr}}}^t\cup S_i^t)
\label{eq: rand alg pf eq final}\end{aligned}$$ In other words, the expected improvement in the objective (left-hand side of ) is at least $1/k$ times the gap of the current objective value to OPT (i.e. right-hand side of ). Note that is only valid when $\mid {S_{\rm{tr}}}\mid<l$ and $\mid S_i \mid<k-l$. Hence, by defining the stopping time $\tau$ as first time that either $\mid {S_{\rm{tr}}}\mid=l$ or $\mid S_i \mid=k-l$, and a telescopic usages of the bounds in , we obtain the following bound: $$\begin{aligned}
\mathbb{E} \left[ \frac{1}{m}\sum\limits_{i=1}^{m}f_i({S_{\rm{tr}}}^{\tau}\cup S_i^{\tau}) \right]\nonumber\geq \text{OPT} \times \mathbb{E} \left[\left(1-\left(1-\frac{1}{k}\right)^\tau\right)\right]\end{aligned}$$ The following theorem finds an upper bound on $\mathbb{E}[(1-\frac{1}{k})^\tau]$ which finishes the proof.
\[thm: tau analyze randomized meta convergence\] If stopping time $\tau$ is first time that either $\mid {S_{\rm{tr}}}\mid=l$ or $\mid S_i \mid=k-l$ then $\mathbb{E}[(1-\frac{1}{k})^\tau]\leq c+\exp(-1+\sqrt{3c.log(\frac{k}{c}}))$ where $c=\frac{1}{\min\{l,k-l\}}$.
let $u_1, u_2, \cdots$ be $i.i.d$ random variables with distribution $u_i \sim \text{Bernoulli}(1 - l/k)$, i.e. $p(u_i = 1) = (k-l)/k$. The stopping time $\tau$ is the first time that $\sum_{i=1}^{\tau}u_i=k-l$ or $\tau - \sum_{i=1}^{\tau}u_i = l$. Let us define $X_r=\sum_{i=1}^{r}u_i$. As a result, $\tau=r$ if and only if either $X_r=k-l$ or $X_r=r-l$. Assume without loss of generality $k-l \leq l$ which means $\max\{\sqrt{1/(k-l)},\sqrt{1/l}\}=\sqrt{1/(k-l)}$. Then, we can write: $$\begin{aligned}
\nonumber
\mathbb{E}[(1-\frac{1}{k})^\tau]&\leq \sum_{r=1}^{k-1}(1-\frac{1}{k})^{r}[p(X_r=k-l)+p(X_r=r-l)]\\&\leq\sum_{r=1}^{k-1}\text{exp}(\frac{-r}{k})[p(X_r=k-l)+p(X_r=r-l)] \label{eq_partition}\end{aligned}$$ We focus on the first term in and show that it is less than $e^{-1 + c}/2$. The second term can be shown similarly to have the same upper bound. By using Sterling and Chernoff bounds we can find upper bound on $p(X_r=k-l)$ as follows: $$\begin{aligned}
p(X_r=k-l)&\leq \frac{1}{\sqrt{r}}\text{exp}(-\frac{(k-r)^2(k-l)}{3kr})\\&\leq
\frac{1}{\sqrt{r}}\text{exp}(-\frac{(k-r)^2(k-l)}{3k^2})\end{aligned}$$
As a result: $$\begin{aligned}
\sum_{r=1}^{k-1} \text{exp}(\frac{-r}{k})[p(X_r=k-l)]&\leq \sum_{r=1}^{k-1} \frac{1}{\sqrt{r}}\text{exp}(\frac{-r}{k})\text{exp}(-\frac{(k-r)^2(k-l)}{3k^2})\\&\leq \sum_{i=0}^{ 2\sqrt{k-l} -1 }\frac{1}{\sqrt{k(1-\frac{i}{2\sqrt{k-l}})}}\text{exp}(\frac{-i^2}{12})\text{exp}(-1+\frac{i}{2\sqrt{k-l}})\\&\leq \text{exp}(-1+\frac{1}{\sqrt{k-l}})\sum_{i=0}^{2\sqrt{k-l}-1}\frac{1}{\sqrt{k(1-\frac{i}{2\sqrt{k-l}})}}\text{exp}(\frac{-i^2}{12})\text{exp}(\frac{i-1}{2\sqrt{k-l}})\end{aligned}$$ where one can show using Gaussian integrals that $\sum\limits_{i=0}^{2\sqrt{k-l}-1}\frac{1}{\sqrt{k(1-\frac{i}{2\sqrt{k-l}})}}\text{exp}(\frac{-i^2}{12})\text{exp}(\frac{i-1}{2\sqrt{k-l}})$ is always less than 0.5[@bullen1998dictionary]. Therefore, $\sum\limits_{r=1}^{k-1} \text{exp}(\frac{-r}{k})[p(X_r=k-l)]\leq \frac 12 \text{exp}(-1+\frac{1}{\sqrt{k-l}})$ Similarly, we can use the same argument to prove $\sum\limits_{r=1}^{k-1} \text{exp}(\frac{-r}{k})[p(X_r=r-l)]\leq \frac12 \text{exp}(-1+\frac{1}{\sqrt{k-l}})$, which finishes the proof.
Furthermore, we define $\tau^{'}=r$ when $r$ is the first time that $X_r=r-l$ and $\tau^{''}=r$ when $r$ is the first time that $X_r=k-l$. Also, let $c=\frac{1}{\min\{l,k-l\}}$ as it was defined in the lemma. By this definition, $\tau=\min\{\tau^{''},\tau^{'}\}$ and we can write the following about the probabilities of $\tau^{'}$ and $\tau^{''}$: $$p(\tau^{'}=r)={r-1 \choose l-1}(\frac{k-l}{k})^
{r-l}(\frac{l}{k})^{l}$$ $$p(\tau^{''}=r)={r-1 \choose k-l-1}(\frac{l}{k})^
{r-k+l}(\frac{k-l}{k})^{k-l}$$
then, based on the definition of $\tau^{'}$ and $\tau^{''}$ we have the following properties for $\tau^{'}$ and $\tau^{''}$:
- if $r<k-l$ then $p(\tau^{''}=r)=0$.
- if $r<l$ then $p(\tau^{'}=r)=0$.
- if $r>k$ then $p(\tau^{'}\leq \tau^{''}|\tau^{'}=r)=0$.
- if $r<k$ then $p(\tau^{'}\leq \tau^{''}|\tau^{'}=r)=1$.
- if $r<k$ then $p(\tau^{'}\geq \tau^{''}|\tau^{''}=r)=1$
- if $r>k$ then $p(\tau^{'}\geq \tau^{''}|\tau^{''}=r)=0$.
- $p(\tau^{''}=r|\tau^{'}\geq \tau^{''})=p(\tau=r|\tau^{'}\geq \tau^{''})$.
- $p(\tau^{'}=r|\tau^{'}\leq \tau^{''})=p(\tau=r|\tau^{'}\leq \tau^{''})$.
Moreover using Bayes rule we can write:
- $$p(\tau^{'}=r|\tau^{'}\leq \tau^{''})=\frac{p(\tau^{'}\leq \tau^{''}|\tau^{'}=r)p(\tau^{'}=r)}{p(\tau^{'}\leq \tau^{''})}=\frac{\mathbbm{1}(r\leq k)p(\tau^{'}=r)}{p(\tau^{'}\leq \tau^{''})}.$$
- $$p(\tau^{''}=r|\tau^{'}\geq \tau^{''})=\frac{\mathbbm{1}(r\leq k)p(\tau^{''}=r)}{p(\tau^{''}\leq \tau^{'})}.$$
Let $\bar{X_r}=r-X_r$ we can write $\bar{X_r}=\sum_{i=1}^{r}v_i$ where $v_1,v_2,v_3,\dots$ are $i.i.d$ random variable with distribution $v_i\sim \text{Bernoulli}(l/k)$. Then, we can write the following using Chernoff bound: $$\begin{aligned}
p(\tau^{'}=r)
&\leq p(X_r= r-l)
\\&\leq p(\bar{X_r}\geq l)
\\&\leq p(\bar{X_r}\geq r(\frac{l}{k})-(k-r)\frac{l}{k})
\\&\leq \exp{\left(-\frac{(k-r)^2(\frac{l}{k})^2}{3r(\frac{l}{k})}\right)}
\\&=\exp{\left(-\frac{(k-r)^2(l)}{3rk}\right)}\end{aligned}$$ Similarly: $$\begin{aligned}
p(\tau^{''}=r)
&\leq p(X_r= k-l)
\\&\leq p({X_r}\geq k-l)
\\&\leq p({X_r}\geq r(1-\frac{l}{k})-(k-r)(1-\frac{l}{k}))
\\&\leq \exp{\left(-\frac{(k-r)^2(1-\frac{l}{k})^2}{3r(1-\frac{l}{k})}\right)}
\\&\leq \exp{\left(-\frac{(k-r)^2(k-l)}{3rk}\right)}
\\&\leq \exp{\left(-\frac{(k-r)^2}{3rkc}\right)}\end{aligned}$$ then we can write the $\mathbb{E}[(1-\frac{1}{k})^\tau]$ as follows:
$$\begin{aligned}
\label{eq:pf expectation}
\mathbb{E}[(1-\frac{1}{k})^\tau]&=\sum_{r=1}^{k}(1-\frac{1}{k})^rp(\tau=r)\leq(1-\frac{1}{k})^{k-\alpha \sqrt{c}}+\sum_{r=1}^{k-\alpha \sqrt{c}}(1-\frac{1}{k})^rp(\tau=r)\end{aligned}$$
Our goal is to find proper bound for . we focus on the second term in - and try to find proper bound for it.
$$\begin{aligned}
\sum_{r=1}^{k-\alpha \sqrt{c}}&(1-\frac{1}{k})^rp(\tau=r)
\label{eq:pf thm2 2nd term eq1}\\&=\sum_{r=1}^{k-\alpha \sqrt{c}}(1-\frac{1}{k})^r(p(\tau^{'}=r|\tau^{'}< \tau^{''})p(\tau^{'}<\tau^{''})+ p(\tau^{''}=r|\tau^{'}\geq \tau^{''})p(\tau^{'}\geq\tau^{''}))
\label{eq:pf thm2 2nd term eq2}\\&=\sum_{r=1}^{k-\alpha \sqrt{c}}(1-\frac{1}{k})^r(p(\tau^{'}=r)+ p(\tau^{''}=r))
\label{eq:pf thm2 2nd term eq3}\\&=\sum_{r=l}^{k-\alpha \sqrt{c}}(1-\frac{1}{k})^rp(\tau^{'}=r)+\sum_{r=k-l}^{k-\alpha \sqrt{c}}(1-\frac{1}{k})^r p(\tau^{''}=r)
\label{eq:pf thm2 2nd term eq4}\\& \leq \sum_{r=l}^{k-\alpha \sqrt{c}}\exp{\left(-\frac{(k-r)^2}{3rkc}\right)} + \sum_{r=k-l}^{k-\alpha \sqrt{c}} \exp{\left(-\frac{(k-r)^2}{3rkc}\right)}
\label{eq:pf thm2 2nd term eq5}\\& \leq (k-l)\exp{\left(-\frac{(k-(k-\alpha \sqrt{c}))^2}{3k^2c}\right)} + l \exp{\left(-\frac{(k-(k-\alpha \sqrt{c}))^2l}{3k^2}\right)}
\label{eq:pf thm2 2nd term eq6}\\& \leq (k-l)\exp{\left(-\frac{(\alpha \sqrt{c})^2}{3k^2c}\right)} + l \exp{\left(-\frac{(\alpha \sqrt{c})^2l}{3k^2}\right)}
\label{eq:pf thm2 2nd term eq7}\end{aligned}$$
where follows from law of total probability, follows from bayes rule, follows from Chernoff bound, follows from the fact that $r<k$. Let $\alpha=3\sqrt{\log(\frac{1}{c})}.k$. As result, we have: $$\begin{aligned}
\sum_{r=1}^{k-\alpha \sqrt{c}}&(1-\frac{1}{k})^rp(\tau=r)\leq (k-l)c^{3} + l c^{3cl} \end{aligned}$$ Assume without loss of generality $k-l\leq l$ and $k-l\geq 2$. As a result, $c=\frac{1}{k-l}$. we want to show that $(k-l)c^{3} + l c^{3cl}=c^2+lc^{3cl}\leq c$. To show this, we show the following equivalent inequality : $$\begin{aligned}
l (k-l)^{-3cl}\leq c(1-c)=\frac{k-l-1}{(k-l)^2}\end{aligned}$$ This holds since $k-l\geq 2$ we have $ \frac{l}{(k-l)^3} (k-l)^{-3(cl-1)}\leq \frac{l}{(k-l)^3} 2^{-3(\frac{l}{k-l}-1)}\leq \frac{l}{(k-l)^3\frac{l}{k-l}}=\frac{1}{(k-l)^2}\leq \frac{k-l-1}{(k-l)^2}$. Moreover, we can bound the first term in as follows: $$\begin{aligned}
(1-\frac{1}{k})^{k-\alpha \sqrt{c}}\leq \exp(-1+3\sqrt{c.\log(\frac{1}{c}}))\end{aligned}$$ summing up we can find the following bound for $\mathbb{E}[(1-\frac{1}{k})^\tau]$ which finishes the proof.
$$\begin{aligned}
\mathbb{E}[(1-\frac{1}{k})^\tau]&\leq c+\exp(-1+3\sqrt{c.log(\frac{1}{c}}))\end{aligned}$$
Continuous Algorithm for Submodular Meta-Learning
=================================================
In this section, we provide method that accomplish $1-1/e$ approximation ratio for problem \[eq:ML\_submodular\_sample\_avg\] . In high level, we use the continuous optimization method and dependent rounding technique in [@balkanski2016learning] to obtain a solution.\
\
**New Ground Set:** Similar to [@balkanski2016learning], we define new ground set of size $nm+n$ which has the orginal ground set elements and element for every $(element, function)$ pair ${\mathcal{X}}^{'}=V \cup \{a_{i,j}\}_{i\in[n],j\in[m]}$.
$$\label{eq:cont reformulation}
g(S)=\sum_{j=1}^{m} f_j(\{a_i: a_{i,j}\in S\})$$
\
\
**Continuous problem:** let us associate with each element $a_{i}$ a variable $x_i \in [0,1]$ and for each element $a_{i,j}$ a variable $x_{i,j} \in [0,1]$; then, we define the $G(x)$ as in : $$\label{eq:cont obj1}
G(x)= \mathbb{E}_{s \sim D(x)}g(S)$$ where $D(x)$ is the following distribution:
1. $a_i \in S \sim D(x)$ for each i independently with probability $x_i$
2. $a_{i,j} \in S \sim D(x)$ for each i and for each j independently with probability $\frac{x_{i,j}}{x_i}$ if $a_i \in S$ and with probability 0 otherwise.
we can write the continuous version of the problem as note that the difference between this problem and two-stage submodular problem shows itself in .
\[eq: cont\] $$\begin{aligned}
{4}
& \max_{\mathcal{S}}&& G(x)
\\
&\operatorname*{subject\,\, to \quad}&& x_i\in [0,1]&&&\forall i\in[n]\\
& && x_{i,j}\in [0,1]&&&\forall i\in[n]\forall j\in[m]\\
& && x_{i} \leq x_{i,j} \;\;\;&&&\forall j\in[m]\label{eq: dif two-stage meta sub}\\
& && \sum_{j} x_{i,j} \leq k \;\;\;\\
& && \sum_{i} x_{i} \leq l \;\;\;\end{aligned}$$
\[lemma: cont\] $G_{ME}=G$ for every x where $$\label{eq:cont obj}
G_{ME}(x)= \mathbb{E}_{s \sim x}g(S)$$
[@balkanski2016learning] then we can reformulate the problem as:
\[eq: cont2\] $$\begin{aligned}
{4}
& \max_{\mathcal{S}}&& G_{ME}(x)
\\
&\operatorname*{subject\,\, to \quad}&& x_i\in [0,1]&&&\forall i\in[n]\\
& && x_{i,j}\in [0,1]&&&\forall i\in[n]\forall j\in[m]\\
& && x_{i} \leq x_{i,j} \;\;\;&&&\forall j\in[m]\\
& && \sum_{j} x_{i,j} \leq k \;\;\;\\
& && \sum_{i} x_{i} \leq l \;\;\;\end{aligned}$$
Lemma \[lemma: cont\] reduces the problem to submodular maximization over solvable polytope [@calinescu2011maximizing] which can be solve using continuous greedy method and the solution achieves $(1- \frac{1}{e})$ approximation ratio. But, this algorithm is very slow and can not be use in applications with large amount of data.
Generalization of Submodular Meta-Learning
------------------------------------------
Generalization of submodular maximization has been studied before as $k$-submodularity . The $k$-submodularity studies function of $k$ subset of ground set which are disjoint. And function is submodular in each set[@ohsaka2015monotone]. However, in submodular meta-learning framework, sets can overlap and we don’t have disjoint restriction on the sets. Therefore, we need to come up with more general notion that can be generalize submodular meta-learning. We introduce the notion of orthant-submodularity and propose two algorithm to solve this problem.
### General Submodular Meta-Learning Formulation
A inherent structure of submodular meta-learning involves the m subsets of ground set which lead us to generalize to more general case in which we are dealing with general function of m subsets of ground set. In order to pursue this goal we need to define following definitions.
**Set of m-tuples.** Let ${\mathcal{X}}$ be finite set of elements of size n, and ${\mathcal{Y}}:=2^{\mathcal{X}}$ be the all the subset of ${\mathcal{X}}$, we call ${\mathcal{Y}}^m:=\{(S_1,S_2,\dots,S_m)|\;S_i\in{\mathcal{Y}}\;\}$ set of all m-tuples.
**Partial Order on m-tuples.** For two element ${\mathcal{S}}\, , \,{\mathcal{S}}^{'}\in{\mathcal{Y}}^m$ we call ${\mathcal{S}}\leq {\mathcal{S}}^{'}$ iff for every $i$ in $\{1,2,\dots, m\}$, $S_i\subseteq S_i^{'}$.
**Addition on m-tuples.** For ${\mathcal{S}},{\mathcal{S}}^{'}\in {\mathcal{Y}}^m$ and $e\in {\mathcal{X}}$ we denote ${\mathcal{S}}\xrightarrow{\{j,e\}}{\mathcal{S}}^{'}$ whenever for every $i$ in $\{1,2,\dots, m\}$ except $j$, formally ${S^{'}_i}=S_i$ and $S^{'}_j=S_j\cup\{e\}$.
**Addition and Subtraction on m-tuples.** For ${\mathcal{S}},{\mathcal{S}}^{'}\in {\mathcal{Y}}^m$ and $e\in {\mathcal{X}}$ we denote ${\mathcal{S}}\xrightarrow[\{k,d\}]{\{j,e\}}{\mathcal{S}}^{'}$ whenever for every $i$ in $\{1,2,\dots, m\}$ except $j,k$ ${S^{'}_i}=S_i$ and $S^{'}_j=S_j\cup\{e\}$, $S^{'}_k=S_j\setminus\{d\}$.
**Discrete Derivative of Function on m-tuples.** function $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ is function that map every element of ${\mathcal{Y}}^m=\{(S_1,S_2,\dots,S_m)|\;S_i\in{\mathcal{Y}}\;\}$ to a positive number. Define the discrete derivative of $f$ at ${\mathcal{S}}\in {\mathcal{Y}}^m , e\in{\mathcal{X}}$ as follows: $$\Delta_{i,e}f(S):=f({\mathcal{S}}^{'})-f({\mathcal{S}})$$ where ${\mathcal{S}}\xrightarrow{\{j,e\}}{\mathcal{S}}^{'}$.
**Orthant-Submodularity.** Let ${\mathcal{X}}$ be finite set of elements, and ${\mathcal{Y}}=2^{\mathcal{X}}$ be the all the subset of ${\mathcal{X}}$, function $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ is function that map every element of ${\mathcal{Y}}^m=\{(S_1,S_2,\dots,S_m)|\;S_i\in{\mathcal{Y}}\;\}$ to a positive number. $f$ is **Orthant-submodular** if for every ${\mathcal{S}}$,${\mathcal{S}}^{'}\in {\mathcal{Y}}^m $ ,where ${\mathcal{S}}\leq {\mathcal{S}}^{'}$:
$$\Delta_{i,e}f({\mathcal{S}})\geq \Delta_{i,e}f({\mathcal{S}}^{'})$$
**Monotonicity.** Let ${\mathcal{X}}$ be finite set of elements, and ${\mathcal{Y}}=2^{\mathcal{X}}$ be the all the subset of ${\mathcal{X}}$, function $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ is function that map every element of ${\mathcal{Y}}^m=\{(S_1,S_2,\dots,S_m)|\;S_i\in{\mathcal{Y}}\;\}$ to a positive number. $f$ is **monotone** iff for every ${\mathcal{S}}$,${\mathcal{S}}^{'}\in {\mathcal{Y}}^m $ ,where ${\mathcal{S}}\leq {\mathcal{S}}^{'}$:
$$f({\mathcal{S}}^{'})\geq f({\mathcal{S}})$$
\[pr: monotone orthant submodular\] Consider the function $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ which is monotone and orthant-submodular. **Monotone orthant-submodular maximization with individual size constraints** is the problem of maximizing $f({\mathcal{S}})$ over m-tuples, ${\mathcal{S}}\in {\mathcal{Y}}^m$ subject to individual size constraint on every $S_i$ , formally:
\[eq:monotone orthant submodular\] $$\begin{aligned}
{3}
& \max_{\mathcal{S}}&& f({\mathcal{S}})
\label{eq:monotone orthant submodular obj}\\
&\operatorname*{subject\,\, to \quad}&& {\mathcal{S}}\in {\mathcal{Y}}^m\\
& && |S_i| \leq k_i \;\;\;\forall i\in[m]\end{aligned}$$
\[co: equivalent of two-stage submodular\] let $f_i:{\mathcal{Y}}\xrightarrow{}{\mathbb{R}}_+$ be monotone submodular function for $i\in[m-1]$ let ${\mathcal{S}}\in {\mathcal{Y}}^m$ be ${\mathcal{S}}=(S_1,S_2,...,S_m)$ define $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ as written in \[eq: equivalent of two-stage submodular function\]: $$\label{eq: equivalent of two-stage submodular function}
f({\mathcal{S}}):=\sum_{i=1}^{m-1} f_i(S_i\cap S_m)$$ then the function $f$ is monotone orthant-submodular. Then consider the problem which is monotone orthant-submodular maximization with individual size constraints as in \[eq: equivalent of few shot submodular optimization\] this problem has same optimal solution as problem \[pr: two-stage submodular\].
\[eq: equivalent of two-stage submodular optimization\] $$\begin{aligned}
{4}
& \max_{\mathcal{S}}&& f({\mathcal{S}})
\label{eq: equivalent of two-stage submodular optimization obj}\\
&\operatorname*{subject\,\, to \quad}&& {\mathcal{S}}\in {\mathcal{Y}}^m\\
& && |S_i| \leq k \;\;\;\forall i\in[m-1]\\
& && |S_m| \leq l \;\;\;\end{aligned}$$
As we saw both problem \[pr: Few-Shot Submodular optimization\] and \[pr: two-stage submodular\] are special case of monotone orthant-submodular maximization with individual size constraints. Therefore, the generalization of every meta submodular task can be formulated as maximization of monotone orthant-submodular maximization with individual size constraints.
### Algorithms for General Submodular Meta-Learning {#sec:algorithms}
### Deterministic Algorithms
In this section we will describe two greedy based algorithms that solve the problem \[pr: monotone orthant submodular\]. Note that in each algorithm we use the similar technique as in [@ohsaka2015monotone].\
In algorithm 1, we begin with $\emptyset$ for every set $S_i$ and in each step $t$ we add the pair $e$ to set $S_i$ in which $\{i,e\}$ maximize the marginal gain of $f$ and choosing $\{i,e\}$ does not violate cardinality constraints.
**Input:** a monotone orthant-submodular $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ and size constraints $k_i$ $\forall i \in [m]$\
**Output:** ${\mathcal{S}}\in {\mathcal{Y}}^m$, where $|S_i|=k_i$ $\forall i \in [m]$
Initialize $K=\sum^{m}_{i=1}k_i$ , ${\mathcal{S}}\xleftarrow[]{}\emptyset$
${\mathcal{I}}\xleftarrow[]{}\{i \in [m]\;|\;|S_i|< k_i\}$
$V_i\xleftarrow{}{\mathcal{X}}\setminus S_i$ for $i \in {\mathcal{I}}$
$\{i,e\}\xleftarrow{} \operatorname*{arg\,max}_{ i \in {\mathcal{I}}, e\in V_i} \Delta_{i,e} f({\mathcal{S}})$ ${\mathcal{S}}\xrightarrow{\{i,e\}}{\mathcal{S}}$
**end for**
return ${\mathcal{S}}$
\[thm: main M meta convergence\] Algorithm \[alg: M meta-Greedy\] returns $\frac{1}{2}$-approximation in ${\mathcal{O}}(Kmn)$ evaluation of $f$.
### Fast Randomized Lazy Algorithm
Algorithm \[alg: M meta-Greedy\] has a high computational complexity; therefore, inspired by the idea of [@lazygreedy] we construct randomized algorithm which has lower computational complexity. The only difference is instead of choosing any element in $V_i$ we substitute a random subset of it $R_i$. This change significantly reduce the computational cost.
\[thm: Fast M meta-Greedy\] Algorithm \[alg: Fast M meta-Greedy\] returns $\frac{1}{2}$-approximation with probablity $1-\delta$ in ${\mathcal{O}}(n\,m\;logK\;log\frac{K}{\delta})$ evaluation of $f$.
**Input:** a monotone orthant-submodular $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ and size constraints $k_i$ $\forall i \in [m]$\
**Output:** ${\mathcal{S}}\in {\mathcal{Y}}^m$, where $|S_i|=k_i$ $\forall i \in [m]$
Initialize $K=\sum^{m}_{i=1}k_i$ , ${\mathcal{S}}\xleftarrow[]{}\emptyset$
${\mathcal{I}}\xleftarrow[]{}\{i \in [m]\;|\;|S_i|< k_i\}$
$V_i\xleftarrow{}{\mathcal{X}}\setminus S_i$ for $i \in {\mathcal{I}}$
$R_i\xleftarrow{}$a random subset of size $ \min\{\frac{n-|S_i|}{k_i-|S_i|}log\frac{K}{\delta},n\}$ sampled uniformly from $ V_i $ for $ i \in {\mathcal{I}}$.
$\{i,e\} \xleftarrow{} \operatorname*{arg\,max}_{ i \in {\mathcal{I}}, e\in R_i} \Delta_{i,e} f({\mathcal{S}})$ ${\mathcal{S}}\xrightarrow{\{i,e\}}{\mathcal{S}}$
**end for**
return ${\mathcal{S}}$
### Special Example
\[co: equivalent of few shot submodular\] let $f_i:{\mathcal{Y}}\xrightarrow{}{\mathbb{R}}_+$ be monotone submodular function for $i\in[m-1]$ let ${\mathcal{S}}\in {\mathcal{Y}}^m$ be ${\mathcal{S}}=(S_1,S_2,...,S_m)$ define $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ as written in : $$\label{eq: equivalent of few shot submodular function}
f({\mathcal{S}}):=\sum_{i=1}^{m-1} f_i(S_i\cup S_m)$$ then the function $f$ is monotone orthant-submodular. Consider the problem which is a monotone orthant-submodular maximization with individual size constraints for $f$ as in , this problem has same optimal solution as problem .
\[eq: equivalent of few shot submodular optimization\] $$\begin{aligned}
{4}
& \max_{\mathcal{S}}&& f({\mathcal{S}})
\label{eq: equivalent of few shot submodular optimization obj}\\
&\operatorname*{subject\,\, to \quad}&& {\mathcal{S}}\in {\mathcal{Y}}^m\\
& && |S_i| \leq k-l \;\;\;\forall i\in[m-1]\\
& && |S_m| \leq l \;\;\;\end{aligned}$$
Using Corollary \[co: equivalent of few shot submodular\] we can use algorithms for maximization of general orthant-submodular with individual size constraint, algorithm and , for discrete meta learning problem, problem . In order to do that first note that for objective function $f({\mathcal{S}}):=\sum_{i=1}^{m-1} f_i(S_i\cup S_m)$, the step 5 of algorithm can be simplified as follows. For every i and e we have: $$\Delta_{m,e} f({\mathcal{S}})=\sum_{i=1}^{m-1} f_i(S_i\cup S_m\cup{e})-f_i(S_i\cup S_m)\geq \Delta_{i,e} f({\mathcal{S}})=\ f_i(S_i\cup{e}\cup S_m)-f_i(
S_i\cup{e}\cup S_m)$$ which means in algorithm step 5, until the $S_m$ is not hit the cardinality constraint non of the other sets have chosen. This results simplified algorithm for discrete meta learning problem. Since, algorithm .
### Convergence Analysis
### Proof of corollary \[co: equivalent of few shot submodular\]
We want to proof first that the function in equation \[eq: equivalent of few shot submodular function\] is monotone. Suppose ${\mathcal{S}}\leq {\mathcal{S}}^{'}$ we want to show $f({\mathcal{S}})\leq f({\mathcal{S}}^{'})$. we Know that every $f_i$ is monotone which means for every $j\in[m-1]$ we have $S_j\cup S_m\subseteq S_j^{'}\cup S_m^{'}$; therefore, we can conclude $f_j(S_j\cup S_m)\leq f_j(S_j^{'}\cup S_m^{'})$ which results $f({\mathcal{S}})\leq f({\mathcal{S}}^{'})$. Secondly, we want to show $\Delta_{i,e}f(S^{'})\leq \Delta_{i,e}f(S)$. let us first consider the case that $i\not = m$ then $\Delta_{i,e}f({\mathcal{S}})=f_i((\{e\}\cup S_i)\cup S_m)-f_i( S_i\cup S_m)\geq f_i((\{e\}\cup {S^{'}_i})\cup S_m^{'})-f_i( {S^{'}_i}\cup S_m^{'})=\Delta_{i,e}f({\mathcal{S}}^{'})$ follows from submodularity of $f_i$. If $i=m$ then $\Delta_{i,e}f({\mathcal{S}})=\sum_{t=1}^{m-1}f_t( S_t\cup S_m\cup \{e\})-\sum_{t=1}^{m-1}f_t( S_t\cup S_m)\geq\sum_{t=1}^{m-1}f_t( S_t^{'}\cup S_m^{'}\cup \{e\})-\sum_{t=1}^{m-1}f_t( S_t^{'}\cup S_m^{'})=\Delta_{i,e}f({\mathcal{S}}^{'})$ follows from submodularity of $f_i$.\
For proving the equivalence of two problem. First note that every feasible solution for problem \[eq: equivalent of few shot submodular optimization\] is also feasible for the maximization problem in \[eq:ML\_submodular\_sample\_avg\]. let the ${\mathcal{S}}^{*}=(S^{*}_1,S^{*}_2,\dots,S^{*}_m)$ be a maximizer of function in \[eq: equivalent of few shot submodular function\] with maximum value $OPT_1=\sum_{i=1}^{m-1} f_i(S^{*}_i\cup S^{*}_m)$, and ${\mathcal{S}}^{'}=(S^{'}_1,S^{'}_2,\dots,S^{'}_m)$ be a maximizer of problem in \[eq:ML\_submodular\_sample\_avg\] with maximum value $OPT_2=\sum_{i=1}^{m-1} f_i({S^{'}_i}\cup S^{'}_m)$, then we can conclude $f({\mathcal{S}}^*)\geq f({\mathcal{S}}^{'})$ which means $OPT_2=\sum_{i=1}^{m-1} f_i(S^{*}_i\cup S^{*}_m)\geq \sum_{i=1}^{m-1} f_i({S^{'}_i}\cup S^{'}_m)=OPT_1$. By same argument $OPT_1\geq OPT_2$, which results $OPT_1=OPT_2$.
### Proof of corollary \[co: equivalent of two-stage submodular\]
Let the ${\mathcal{S}}^{*}=(S^{*}_1,S^{*}_2,\dots,S^{*}_m)$ be a maximizer of function in \[eq: equivalent of two-stage submodular function\] with maximum value $OPT_1=\sum_{i=1}^{m-1} f_i(S^{*}_i\cap S^{*}_m)$, and ${\mathcal{S}}^{'}=(S^{'}_1,S^{'}_2,\dots,S^{'}_m)$ be a maximizer of problem \[eq:ML\_submodular\_sample\_avg\] with maximum value $OPT_2=\sum_{i=1}^{m-1} f_i({S^{'}_i})$ . We want to show ${\mathcal{S}}^{''}:=(S^{*}_1\cap S^{*}_m ,\, S^{*}_2\cap S^{*}_m ,\,\dots,\,S^{*}_{m-1}\cap S^{*}_m,S^{*}_m)$ is a optimal solution for problem \[eq:ML\_submodular\_sample\_avg\]. Firstly, ${\mathcal{S}}^{''}$ is feasible solution for problem \[eq:ML\_submodular\_sample\_avg\]. Secondly, $OPT_1=f({\mathcal{S}}^{*})=\sum_{i=1}^{m-1} f_i(S^{*}_i\cap S^{*}_m)=\sum_{i=1}^{m-1} f_i(S^{''}_i)$ , from optimality of $S^{'}$ for problem \[eq:ML\_submodular\_sample\_avg\], $OPT_2=\sum_{i=1}^{m-1} f_i({S^{'}_i})\geq \sum_{i=1}^{m-1} f_i(S^{''}_i)=OPT_1$. Finally, observe that ${\mathcal{S}}^{'}$ is a feasible solution for \[eq: equivalent of two-stage submodular optimization\] which means $OPT_1=\sum_{i=1}^{m-1} f_i(S^{''}_i)\geq \sum_{i=1}^{m-1} f_i({S^{'}_i})=OPT_2$ results in $OPT_1=OPT_2$ .
### Proof of theorem \[thm: main M meta convergence\] {#proof: thm1}
**Incidence Matrix for m-tuples.** for m-tuple ${\mathcal{S}}$ is matrix $L({\mathcal{S}})\in {\mathbb{R}}^{n\times m}$ which define as: $$L({\mathcal{S}}):=\begin{bmatrix}
\mathbbm{1}_{S_1} & \mathbbm{1}_{S_2} &\dots& \mathbbm{1}_{S_m}
\end{bmatrix}$$
Let ${\mathcal{S}}^{(0)}=\emptyset,{\mathcal{S}}^{(1)},\dots,{\mathcal{S}}^{(t)},\dots,{\mathcal{S}}^{(K)}$ be a chain obtained by algorithm \[alg: M meta-Greedy\] and let $S^*$ be the optimal solution. Define the sequence ${\mathcal{J}}^{(0)}={\mathcal{S}}^{*},{\mathcal{J}}^{(1)},\dots,{\mathcal{J}}^{(K)}={\mathcal{S}}^{(K)}$ as follows. In step 5 of algorithm \[alg: M meta-Greedy\], we have defined $i^{(t)},e^{(t)}$ to be ..... We then define $D^{t}_{i}=J^{(t-1)}_i\setminus S^{(t-1)}_i$ and also define $o^{(t)}$ in the following way: (i) If $e^{(t)}\in D^{t}_{i^{(t)}}$, then $o^{(t)}=e^{(t)}$; (ii) Otherwise, if $e^{(t)}\notin D^{t}_{i^{(t)}}$, let $o^{(t)}$ be one of the elements of $D^{t}_{i^{(t)}}$ chosen uniformly at random. We also, define the set ${\mathcal{J}}^{(t-\frac{1}{2})}$ by deleting $o^{(t)}$ from ${\mathcal{J}}^{(t-1)}$, i.e. .... and ${\mathcal{J}}^{(t)}$ by adding $e^{(t)}$ to ${\mathcal{J}}^{(t-\frac{1}{2})}$, i.e. ...... Figure, explain how the sets.... are evolved when $t$ increases from .. to ...
\^[(0)]{}= \^[(1)]{} … \^[(t)]{} … \^[(K)]{}\
\^[(0)]{}=\^\* \^[(1)]{} … \^[(t)]{} … \^[(K)]{}
Note that by definition ${\mathcal{S}}^{(t-1)}\leq {\mathcal{J}}^{t-\frac{1}{2}}$.
\[lemma: in proof thm1\] For every $t\in [K]$ the following inequality holds: $$\label{eq:lemma_arman}
f({\mathcal{S}}^{(t)})-f({\mathcal{S}}^{(t-1)})\geq f({\mathcal{J}}^{(t-1)})-f({\mathcal{J}}^{(t)}).$$
Define $a^{(t)}=\Delta_{e^{(t)},i^{(t)}}f({\mathcal{J}}^{(t-\frac{1}{2})})$, and $a^{(t-\frac{1}{2})}=\Delta_{o^{(t)},i^{(t)}}f({\mathcal{J}}^{(t-\frac{1}{2})})$, and $b^{(t)}=\Delta_{e^{(t)},i^{(t)}}f({\mathcal{S}}^{(t-1)})$. We observe that $f({\mathcal{J}}^{(t-1)})-f({\mathcal{J}}^{(t)})=a^{(t-\frac{1}{2})}-a^{(t)}$. Using this notation, becomes $b^{(t)}\geq a^{(t-\frac{1}{2})}-a^{(t)}$. By monotonicity of $f$ it suffices to show $b^{(t)}\geq a^{(t-\frac{1}{2})}$. By step 5 of algorithm \[alg: M meta-Greedy\] (greedily choosing elements), we have $b^{(t)}=\Delta_{e^{(t)},i^{(t)}} f({\mathcal{S}}^{(t-1)})\geq \Delta_{o^{(t)},i^{(t)}} f({\mathcal{S}}^{(t-1)})$. Finally, by orthant-submodularity of ..., we obtain $\Delta_{o^{(t)},i^{(t)}} f({\mathcal{S}}^{(t-1)})\geq \Delta_{o^{(t)},i^{(t)}}f({\mathcal{J}}^{(t-\frac{1}{2})})=a^{(t-\frac{1}{2})}$ since ${\mathcal{S}}^{(t-1)}\leq {\mathcal{J}}^{t-\frac{1}{2}}$.
Finally, by summing over the relations , we obtain: $$\begin{aligned}
f_{final}-0= f({\mathcal{S}}^{(K)})-f({\mathcal{S}}^{(0)})&=\sum_{t=1}^{t=K} f({\mathcal{S}}^{(t)})-f({\mathcal{S}}^{(t-1)})\geq \sum_{t=1}^{t=K}f({\mathcal{J}}^{(t-1)})-f({\mathcal{J}}^{(t)})
\\&=f({\mathcal{J}}^{(K)})-f({\mathcal{J}}^{(0)})= f_{final}-OPT\end{aligned}$$
### Proof of theorem \[thm: Fast M meta-Greedy\]
We reruse the notation $i^{(t)},e^{(t)},{\mathcal{J}}^{(t)},{\mathcal{S}}^{(t)},D_i^t$ in section \[proof: thm1\]. Assume $R_i^{(t)}$ be the $R_i$ in $t^{th}$ iteration. Let us iteratively define ${\mathcal{J}}^{(0)}={\mathcal{S}}^{*},{\mathcal{J}}^{(1)},\dots,{\mathcal{J}}^{(m)}={\mathcal{S}}^{(K)}$ as follows. If $R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t}$ is empty, then algorithm fails if not set $o^{(t)}=e^{(t)}$ if $e^{(t)}\in R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t} $ or if $e^{(t)}\not\in R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t} $ set $o^{(t)}$ a random element from $R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t}$. Every other part of the proof follows as in section \[proof: thm1\]. We just need to find the probability that the algorithm fails.
With probability $1-\delta$ , $R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t}\not = \emptyset$ for every $t\in [K]$.
If for any $t\in [K]$ $|R_{i^{(t)}}^{t}|=n$ then $Pr[R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t}=\emptyset]=0$ $$\begin{aligned}
Pr[R_{i^{(t)}}^{t}\cap D_{i^{(t)}}^{t}=\emptyset]&=(1-\frac{|S_{i^{(t)}}^{t}|}{|V_{i^{(t)}}^t|})^{(|R_{i^{(t)}}^{t}|)}=(1-\frac{k_{i^{(t)}}-|S_{i^{(t)}}^{t-1}|}{n-|S_{i^{(t)}}^{t-1}|})^{(|R_{i^{(t)}}^{t}|)}\\&\leq exp({-(\frac{k_{i^{(t)}}-|S_{i^{(t)}}^{t-1}|}{n-|S_{i^{(t)}}^{t-1}|}) (\frac{n-|S_{i^{(t)}}^{t-1}|}{k_{i^{(t)}}-|S_{i^{(t)}}^{t-1}|})log(\frac{K}{\delta})})=\frac{K}{\delta}\end{aligned}$$ union bound over $t\in [K]$ prove the lemma.
Algorithm for Worst Case Setting {#sec: Distribution Agnostic Algorithm}
--------------------------------
In the first section, we discussed the proper formulation for the case that we aim to choose an initial subset of the desired set at the training time and the remaining elements at the test time. In particular, we introduced the formulation that finds the best initial set that in expectation or on average would perform well when we later add the best subset to it. Indeed, this formulation leads to a good generalization result since we assume all tasks at the training time are drawn from a common distribution and the task that we observe at the test time is also drawn from that distribution.
However, in some cases, either the tasks in the training phase are not necessarily drawn from the same distribution, or instead of average performance we care about the worst-case performance. In either case, instead of measuring the performance of the initial set in terms of the maximum value of the expected objective function, we aim to maximize the minimum objective function value over all given tasks. In other words, we aim to solve $$\label{eq:ML_submodular_minmax}
\max_{{S_{\rm{tr}}}\in V, |{S_{\rm{tr}}}|\leq l} \ \min_{i=1,\dots, m}\; \max_{S_i\in V, |S_i|\leq k-l} f_i({S_{\rm{tr}}}\cup S_i)
\ = \ \max_{{S_{\rm{tr}}}\in V, |{S_{\rm{tr}}}|\leq l} \ \min_{i=1,\dots, m}\; f_i'({S_{\rm{tr}}}).$$ Note that both problems in are equivalent since $f_i'({S_{\rm{tr}}}):=\max_{S_i\in V, |S_i|\leq k-l} f_i({S_{\rm{tr}}}\cup S_i)$. Later, in Section \[sec:algorithms\_ML\] we formally present efficient methods for solving this problem and characterize their theoretical guarantees.
### Equivalent of Worst Case Setting
\[co: equivalent of MASML\] let $f_i:{\mathcal{Y}}\xrightarrow{}{\mathbb{R}}_+$ be monotone submodular function for $i\in[m]$ let ${\mathcal{S}}\in {\mathcal{Y}}^{m+1}$ be ${\mathcal{S}}=(S_1,S_2,...,{S_{\rm{tr}}})$ define $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$ as written in : $$\label{eq: equivalent of MASML function}
f({\mathcal{S}}):=\min_{i\in [m]} f_i(S_i\cup {S_{\rm{tr}}})$$ Then consider the problem which is a maximization of $f$ with individual size constraints as in \[eq: equivalent of MASML\], this problem has same optimal solution as problem .
\[eq: equivalent of MASML\] $$\begin{aligned}
{4}
& \max_{\mathcal{S}}&& f({\mathcal{S}})
\label{eq: equivalent of MASML obj}\\
&\operatorname*{subject\,\, to \quad}&& {\mathcal{S}}\in {\mathcal{Y}}^{m+1}\\
& && |S_i| \leq k-l \;\;\;\forall i\in[m]\\
& && |{S_{\rm{tr}}}| \leq l \;\;\;\end{aligned}$$
### Proof of corollary \[co: equivalent of MASML\]
For proving the equivalence of two problem. First note that every feasible solution for problem in \[eq: equivalent of MASML\] is also feasible for the maximization problem in . let the ${\mathcal{S}}^{*}=(S^{*}_1,S^{*}_2,\dots,S^{*}_{tr})$ be a maximizer of function in \[eq: equivalent of MASML\] with maximum value $OPT_1=\min\limits_{i\in [m]} f_i(S^{*}_i\cup S^{*}_{tr})$, and ${\mathcal{S}}^{'}=(S^{'}_1,S^{'}_2,\dots,S^{'}_{tr})$ be a maximizer of problem with maximum value $OPT_2=\min\limits_{i\in [m]} f_i({S^{'}_i}\cup S^{'}_{tr})$, then we can conclude $OPT_1\geq \min\limits_{i\in [m]} f_i(S_i\cup S_m)$ for every ${\mathcal{S}}$ in feasible set specially ${\mathcal{S}}={\mathcal{S}}^{'}$ which results $OPT_1\geq OPT_2$. Similarly, $OPT_2=\min\limits_{i\in [m]}\max\limits_{S_i}f_i({S_{\rm{tr}}}^{*}\cup S_i)\geq\min\limits_{i\in [m]}\max\limits_{S_i}f_i({S_{\rm{tr}}}^{'}\cup S_i)\geq\min\limits_{i\in [m]}f_i({S_{\rm{tr}}}^{'}\cup {S^{'}_i})=OPT_1$.
\[co: Hardness of MASML\] Finding any constant factor of problem in is NP-Hard. Because, even in the case of $m=1$ the problem is equivalent to robust submodular optimization which is NP-Hard to find any constant approximation factor.
### Bi-criteria Approximation Algorithm
It is computationally hard to find the approximation solution for problem in \[eq: equivalent of MASML\]. Therefore, we try to solve the problem with Bi-criteria Approximation method. First, Using slack variable $c$ we can reformulate the \[eq: equivalent of MASML\] as \[eq MASML equal slack variable\].
\[eq MASML equal slack variable\]$$\begin{aligned}
{4}
& \max_{{\mathcal{S}},c} && c
\label{eq: MASML equal slack variable obj}\\
&\operatorname*{subject\,\, to \quad}&& f_i(S_i\cup {S_{\rm{tr}}})\geq c\\
& && {\mathcal{S}}\in {\mathcal{Y}}^{m+1}\\
& && |S_i| \leq k-l &&&\forall i\in[m]\\
& && |{S_{\rm{tr}}}| \leq l \;\;\;\end{aligned}$$
Then, inspired by [@krause2008robust] we relaxed the cardinality constraints in \[eq MASML equal slack variable\] which results \[eq: Relaxed MASML2\].
\[eq: Relaxed MASML2\] $$\begin{aligned}
{4}
& \max_{{\mathcal{S}},c} && c
\label{eq: Relaxed MASML2 obj}\\
&\operatorname*{subject\,\, to \quad}&& f_i(S_i\cup {S_{\rm{tr}}})\geq c \;\;\;&&&\forall i\in[m]\label{eq: Relaxed MASML2 const1}\\
& && {\mathcal{S}}\in {\mathcal{Y}}^{m+1}\label{eq: Relaxed MASML2 const2}\\
& && |S_i| \leq \alpha (k-l) &&&\forall i\in[m]\label{eq: Relaxed MASML2 const3}\\
& && |{S_{\rm{tr}}}| \leq \alpha l\label{eq: Relaxed MASML2 const4} \;\;\;\end{aligned}$$
Our goal is approximately solve \[eq: Relaxed MASML2\]. In order to do that we will use binary search procedure over c, by itertaively solving \[eq: Relaxation Binary Search Procedure\] and make the search interval smaller in each step. We can reformulate and relax the constraints in \[eq: Relaxed MASML2 const1\]-\[eq: Relaxed MASML2 const4\] as \[eq: Relaxation Binary Search Procedure const1\]-\[eq: Relaxation Binary Search Procedure const3\].
\[eq: Relaxation Binary Search Procedure\] $$\begin{aligned}
{4}
& {\mathcal{S}}_c^{*}=\operatorname*{arg\,min}_{{\mathcal{S}}} && \mid {\mathcal{S}}\mid = \sum_{i=1}^{m} \mid S_i \mid+\mid {S_{\rm{tr}}}\mid
\label{eq: Relaxation Binary Search Procedure obj}\\
&\operatorname*{subject\,\, to \quad}&& F_c({\mathcal{S}}):=\frac{1}{m}\sum_{i=1}^{m}\min\{f_i(S_i\cup {S_{\rm{tr}}}),c\}= F_c({\mathcal{X}}^{m+1}) \;\;\label{eq: Relaxation Binary Search Procedure const1}\\
& && {\mathcal{S}}\in {\mathcal{Y}}^{m+1}\\
& && \frac{\mid S_i \mid}{\mid {\mathcal{S}}\mid} \leq \frac{(k-l)}{(k-l)(m)+l} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall i\in[m]\label{eq: Relaxation Binary Search Procedure const2}\\
& && \frac{\mid {S_{\rm{tr}}}\mid}{\mid {\mathcal{S}}\mid} \leq \frac{l}{(k-l)(m)+l} \label{eq: Relaxation Binary Search Procedure const3}\end{aligned}$$
We named the optimal value of problem $Z=\min \mid{\mathcal{S}}^{*}_c\mid$. The bi-criteria approximation procedure start with $c_{min}=0$ , $c_{max}=\min_if_i({\mathcal{X}})$ and In each step solves optimization problem in \[eq: Relaxation Binary Search Procedure\] by algorithm \[alg: meta-Saturation Submodular algorithm \]and give us the solution ${\mathcal{S}}^{*}_{c}$ and update $c_{min}$ , $c_{max}$ as follows.
- if $ \mid {\mathcal{S}}^{*}_{c} \mid :=\sum\limits_{i=1}^{m} \mid S_i^{c} \mid +\mid {S_{\rm{tr}}}^{c} \mid\; \geq \; \alpha (l+(m)(k-l))$ or the problem was infeasible then $c_{max}=c$
- if $ \mid {\mathcal{S}}^{*}_{c} \mid =\sum\limits_{i=1}^{m} \mid S_i^{c} \mid+\mid {S_{\rm{tr}}}^{c} \mid \; \leq \; \alpha (l+(m)(k-l))$ then $c_{min}=c$
\[eq: Relaxation Binary Search Procedure\] $$\begin{aligned}
{4}
& {\mathcal{S}}_c^{*}=\operatorname*{arg\,min}_{{\mathcal{S}}} && \mid {\mathcal{S}}\mid = \sum_{i=1}^{m} \mid S_i \mid +\mid {S_{\rm{tr}}}^{c} \mid
\label{eq: Relaxation Binary Search Procedure obj}\\
&\operatorname*{subject\,\, to \quad}&& f_i(S_i\cup {S_{\rm{tr}}})\geq c \;\;&&&\;\; \forall i\in[m]\\
& && {\mathcal{S}}\in {\mathcal{Y}}^m\\
& && \frac{\mid S_i \mid}{\mid {\mathcal{S}}\mid} \leq \frac{k}{k(m)+l} &&&\;\; \forall i\in[m]\\
& && \frac{\mid {S_{\rm{tr}}}\mid}{\mid {\mathcal{S}}\mid} \leq \frac{l}{k(m)+l} &&&\;\; \end{aligned}$$
**Input:** a function $f:{\mathcal{Y}}^{m+1}\xrightarrow{}{\mathbb{R}}_+$ ,c , ${\mathcal{S}}^0=\emptyset$, $t=0$\
**Output:** ${\mathcal{S}}\in {\mathcal{Y}}^m$
${\mathcal{I}}\xleftarrow[]{}\{i \in [m]\;|\; i\; $satisfy \[eq: Relaxation Binary Search Procedure const2\] and \[eq: Relaxation Binary Search Procedure const3\]$ \}$ ${\mathcal{I}}=\{tr\}$$V_i\xleftarrow{}{\mathcal{X}}\setminus S_i$ for $i \in {\mathcal{I}}$
$\{i^t,j^t\} \xleftarrow{} \operatorname*{arg\,max}_{ i \in {\mathcal{I}}, j\in V_i} \Delta_{i,j} f({\mathcal{S}}^t)$ ${\mathcal{S}}^t \xrightarrow{\{i^t,j^t\}}{\mathcal{S}}^{t+1}$, $\theta^{t}:=\;\frac{1}{\Delta_{i^{t},j^{t}}F_c({\mathcal{S}}^{t})}$ , $t \xleftarrow{}t+1$ **end While**
return ${\mathcal{S}}$
In the following lemma we analyze the quality of solution for each run of the greedy algorithm.
\[eq: model agnostic main lemma \] For integral valued $f_i$ and feasible c Algorithm \[alg: M meta-Greedy submodular partial cover (MGPC)\] finds m tuple ${\mathcal{S}}^{G}$ which $f_i({\mathcal{S}}^{G})\geq c$ for all $i$ and $\mid S_i^{G}\mid\leq \alpha \mid S_i^{*} \mid$ where ${\mathcal{S}}^{*}$ is optimal solution and $\alpha=1 +log(max_{e\in{\mathcal{X}},i\in[m]}\Delta_{i,e}f(\emptyset))$.
**Sketch of the proof:**\
Our proof is similar to the proof in [@wolsey1982analysis]. We want to find upper bound on the greedy solution provided by Algorithm \[alg: M meta-Greedy submodular partial cover (MGPC)\] for optimization problem . To do that, we propose some related problems( , and ) that leads us to find a upper bound. First, we propose problem which is equivalent to . This equivalency will be proven in corollary . Then, we relaxed two constraints , and in and provide a upper bound for solution in . In particular, if $Z^L$ is the optimal value of ,then $Z^L\leq Z$. Afterward, using duality we argue that if $D^L$ is the optimal value of dual problem of we have $Z^L\geq D^L$. Finally, we show that if $Z^G$ is the value of for solution of algorithm we have $D^L\geq Z^G(\alpha)^{-1}$ in lemma \[lemma: final in proof of MASMAl1\] and \[lemma: final in proof of MASMAl2\]. In nutshell, we proved lemma \[eq: model agnostic main lemma \] by showing: $$Z\geq Z^L\geq D^L\geq Z^G(\alpha)^{-1}$$ where $Z$ is the optimal value of , $Z^G$ is the value of for solution of algorithm , $Z^L$ is the optimal value of , and $D^L$ is the optimal value of dual problem of .
\[col: MASML1\] We will propose the problem which is equivalent to and have same optimal value. For any m-tuple ${\mathcal{S}}$ which is feasible for let it is indicator matrix $z$ be a matrix in which for every $i\in [m]$ and $j \in [n]$, $z_{i,j}={\mathbf{1}}\{j\in S_i\}$ and let $f_i^c(S)=\min\{c,f_i(S)\}$. Then, $z$ is feasible for and the optimal value of two problem is same. (the argument is similar to proposition 2 in [@wolsey1982analysis] ).
\[eq: Relaxation LP\] $$\begin{aligned}
{4}
& Z=\min_{z} && \sum_{i=1}^{m}\sum_{j=1}^{n} z_{i,j}
\label{eq: LPRelaxation}\\
&\operatorname*{subject\,\, to \quad}&& \sum_{i=l,m}\sum_{j=1}^{n} \Delta_{i,j} f_{i}^{c} (S_{l}\cup {S_{\rm{tr}}})z_{i,j} \geq -f_i^c(S_{l}\cup {S_{\rm{tr}}})+f_i^c({\mathcal{X}})\notag \label{eq:Relaxation LP eq1}\\
& && \;\;\;\;\;\;\quad\;\forall t=0,\dots,T-1 \;,\; l\in [m] \;,\;\forall {\mathcal{S}}\in{\mathcal{Y}}^m\\
& && z_{i,j} \in \{0,1\}\\
& && \frac{\sum_{j=1}^{n} z_{i,j} }{\sum_{i=1}^{m}\sum_{j=1}^{n} z_{i,j} } \leq \frac{k}{k(m)+l} &&&\;\; \forall i\in[m]\label{eq:Relaxation LP eq3}\\
& && \frac{\sum_{j=1}^{n} z_{m,j} }{\sum_{i=1}^{m}\sum_{j=1}^{n} z_{i,j} } \leq \frac{l}{k(m)+l} &&&\label{eq:Relaxation LP eq4}\;\; \end{aligned}$$
The equivalency of and is direct result of the proposition 2 in [@wolsey1982analysis].
let $\theta^{t}$ be the output of algorithm \[alg: M meta-Greedy submodular partial cover (MGPC)\] for $t=0,1,\dots T$ where $T$ is number of iteration of algorithm \[alg: M meta-Greedy submodular partial cover (MGPC)\]. we will prove $\mid S_i^{G}\mid\leq \alpha \frac{l}{k(m)+l} \mid {\mathcal{S}}^{*} \mid=\alpha \mid {\mathcal{S}}_i^{*} \mid$. To show that let us define $Z^l$ as follows:
\[eq: Relaxation LP2\] $$\begin{aligned}
{4}
& Z^{L}=\min_{z} && \sum_{i=1}^{m}\sum_{j=1}^{n} z_{i,j}
\label{eq: LPRelaxation2}\\
&\operatorname*{subject\,\, to \quad}&& \sum_{i=l,m}\sum_{j=1}^{n} \Delta_{i,j}f_i^c({S_{\rm{tr}}}^t\cup {S_{\rm{tr}}}^t)z_{i,j} \geq -f_i^c({S_{\rm{tr}}}^t\cup {S_{\rm{tr}}}^t)+f_i^c({\mathcal{Y}}) \notag\\ \; \; & &&\;\; \forall t=0,\dots,T-1 \;,\; l\in [m]\\
& && z_{i,j} \geq 0\end{aligned}$$
This problem relaxed two constraints , and in and as consequence $Z^L\leq Z$. Then, Let $D^L$ be the optimal value of dual problem of . Then, by weak duality $D^L\leq Z^L$.
\[lemma: final in proof of MASMAl1\] Let us call the final value of Algorithm \[alg: M meta-Greedy submodular partial cover (MGPC)\] $Z^G$, and let $\bar f_{i^t}^c(S):=\sum\limits_{i=1}^m f_{i}^c(S){\mathbf{1}}\{i=i^t\; or\; i^t=tr\} $ .Then, we can write $Z^G$ as follows. $$Z^G =\sum_{t=1}^{T}\theta^{t} (F_c({\mathcal{S}}^{t})-F_c({\mathcal{S}}^{t-1}))=\theta^1(\bar f_{i^1}^c({\mathcal{X}}))+\sum_{t=1}^{T}(\theta^{t}-\theta^{t-1})(\bar f_{i^t}^c({\mathcal{X}})-F_c({\mathcal{S}}^{t-1}))$$ First Inequality is a direct consequence of definition of $\theta^t $ and by rearranging and add and subtract $\bar f^c_i({\mathcal{X}})$ we get the second equality.
\[lemma: final in proof of MASMAl2\] Let $D^L$ be optimal value of dual problem of . We can find following lower bound for $D^L$. $$D^L\geq(\theta^1(f_{i^1}^c({\mathcal{X}}))+ \sum_{t=1}^{T}(\theta^{t}-\theta^{t-1})(\bar f_{i^t}^c({\mathcal{X}})-F_c({\mathcal{S}}^{t-1})))(\alpha)^{-1}$$ First for every $j\in {\mathcal{X}}$ there exist $r\leq T$ that $(\Delta_{i^t,j}F_c({\mathcal{S}}^{r-1}))>0$ and $(\Delta_{i^t,j}F_c({\mathcal{S}}^{r}))=0$ . We want to prove $\alpha^{-1}\theta^*$ is dual feasible where $\theta^*=(\theta^1{\mathbf{1}}_{i^0},(\theta^2 -\theta^1){\mathbf{1}}_{i^1},\dots (\theta^T-\theta^{T-1}){\mathbf{1}}_{i^{T-1}})\in {\mathbb{R}}^{(m)\times T }$ (${\mathbf{1}}_{i}$ is vector of size $m$ which is 1 in the $i^{th}$ place or if $i=tr$ it is all ones) for every $j$ we have the following. $$(\theta^1(f_{i^1}^c({\mathcal{X}}))+ \sum_{t=1}^{T}(\theta^{t}-\theta^{t-1})(\Delta_{i^t,j}F_c({\mathcal{S}}^t))\leq
(\max_{t=1,...,r} (\theta^{t})(\Delta_{i^t,j}F_c({\mathcal{S}}^t))) (\alpha)\leq \alpha$$ The first inequality follows from proposition 3 in [@wolsey1982analysis] and $(\theta^{t})(\Delta_{i^t,j}F_c({\mathcal{S}}^t))\leq1$ is consequence of definition of $(\theta^{t})$ and the fact that $i^t$ chosen greedily. Therefore, $\alpha^{-1}\theta^*$ is dual feasible and we can write the . $$\label{eq: dual LPR}
(\theta^1(f_{i^1}^c({\mathcal{X}}))+ \sum_{t=1}^{T}(\theta^{t}-\theta^{t-1})(\bar f_{i^t}^c({\mathcal{X}})-F_c({\mathcal{S}}^{t-1}))) (\alpha^{-1})\leq D^L$$
Then, combing all we have:
$$Z\geq Z^L\geq D^L\geq(\theta^1(f_{i^1}^c({\mathcal{X}}))+ \sum_{t=1}^{T}(\theta^{t}-\theta^{t-1})(\bar f_{i^t}^c({\mathcal{X}})-F_c({\mathcal{S}}^{t-1})))(\alpha)^{-1}=Z^G(\alpha)^{-1}$$
which resualts $\mid S_i^{G}\mid\leq \alpha \frac{l}{k(m)+l} Z=\alpha \mid {\mathcal{S}}_i^{*} \mid$.
In algorithm \[alg: meta-Saturation Submodular algorithm \], we present the whole procedure for finding solution for worst case setting.
**Input:** a function $f:{\mathcal{Y}}^m\xrightarrow{}{\mathbb{R}}_+$\
**Output:** ${\mathcal{S}}\in {\mathcal{Y}}^m$
$c\xleftarrow[]{}\frac{c_{max}+c_{min}}{2}$
$\Tilde{{\mathcal{S}}}=MGPC(F_c,c)$
$c_{max}\xleftarrow[]{}c$ $c_{min}\xleftarrow[]{}c$ , ${\mathcal{S}}_{best}\xleftarrow[]{}\Tilde{{\mathcal{S}}}$ **end While**
return ${\mathcal{S}}_{best}$
The following theorem prove that the optimal value of worst case setting can be achieve by algorithm \[alg: meta-Saturation Submodular algorithm \] if we relax the conditions as follows.
for any $k,l$ algorithm \[alg: meta-Saturation Submodular algorithm \] find m-tuple ${\mathcal{S}}^{'}$ where $$\min_i f_i(S_i^{'}\cup {S_{\rm{tr}}}^{'})\geq \max_{{\mathcal{S}}}\min_i f_i(S_i\cup {S_{\rm{tr}}})$$ and $\mid S_i^{'}\mid\leq \alpha (k-l)$ for $i\in[m]$ and $\mid {S_{\rm{tr}}}^{'}\mid\leq \alpha l$.
In each step $\mid {\mathcal{S}}_{best}\mid \leq \alpha( l+ (m)(k-l))$ and $c^{*}\in [c_{min}, c_{max})$. Since, $f_i$ are integer, and $c_{max}-c_{min}< \frac{1}{m}$; we can conclude $\min\limits_i f_i(S_i^{'}\cup {S_{\rm{tr}}}^{'})\geq \max\limits_{{\mathcal{S}}}\min\limits_i f_i(S_i\cup {S_{\rm{tr}}})$.
Simulation supplementary
------------------------
### Ride Share Optimization
In this application, we try to solve facility location problem on Uber dataset [@uber_2019]. One part of this dataset, that we work with consists of the September 2014 Uber pick-ups in Manhattan, which is around 1,000,000 data point, each one in the form of $(latitude, longitude, datatime)$. Our goal is to use the data from the past to find the better waiting location for drivers in the future. In particular, We want to find subset of location that minimizes the defined metric between each customers to his closest driver. To be more precise, given a customer $p=(x_p,y_p)$ and driver $r=(x_r, y_r)$ define a manhatan distance between these two point as $d(p,r)=|x_p-x_r|+|y_p-y_r|$. We use the same metric as [@mitrovic2018data], the “convenience score”, which is $c(p,r)=2-\frac{2}{1+e^{-200 d(p,r)}}$.Next, let us define the tasks and learning procedure more clearly. We use September 1 of 2014 as a training set and September 2 of 2014 as a test set. September 1 of 2014 consists of around 700 points and September 1 of 2014 consists of around 1200 points. We define a Time slot $T_a$ by selecting point $a$ and pick 10 data points in half an hour before point $a$ and for each one pick 10 points in 1 km neighborhood of that which makes a total of 100 points. Each task is ${\mathcal{T}}_a=\{\ f_a,T_a\}$ where $f_a$ is monotone submodular function defined for every driver waiting locations set $S$ as $f_a(S)=\sum_{p\in T_a}\max_{r\in S}c(p,r)$, and $T_a$ is corresponding time slot set. We pick 100,000 points at random from the September 2014 Uber pick-ups locations as ground set. For training we pick $m$ points at random from the first week $a_1,a_2,\dots, a_m$ and consider the tasks ${\mathcal{T}}_1, {\mathcal{T}}_2,\dots,{\mathcal{T}}_m$. We run the **Greedy Train**(running greedy at training on sum),**meta-Greedy**, and **Random meta-Greedy** algorithms on them which means we try to optimize $\sum_{i=1}^{m}f_{a_{i}}({S_{\rm{tr}}}\cup S_i)$ and find set $S$ with size $l$ which can be used in test time as initial set. In test time we test the performance of algorithms on the series of tasks ${\mathcal{T}}^{'}_{1}, {\mathcal{T}}^{'}_{2},\dots,{\mathcal{T}}^{'}_{h}$ and report the comparison of performance on average. First, we pick $m$ points at random from second week $a^{'}_{1}, a^{'}_{2},\dots,a^{'}_{m}$ and look at the performance of the **Greedy Train**,**Greedy Test**,**meta Greedy**, and **random meta Greedy** algorithms on this tasks. We report the normalized performance of algorithms in figure \[fig:simulation-ride\]. We can see if we run greedy for each task in test time independently, we achieve better performance, but in the price of computation time.
### Movie Recommendation
In this application we use the data set of movies from movielens dataset [@fivethirtyeight_2019]. The data set consists of movie information and movie ratings between 1 to 5. We pick the subset of 2000 most rated movies and 200 user that rate the highest to these movies same as [@stan2017probabilistic]. Each movies can belong to 18 genres(Horror, Thriller, and etc.). We define a set of all movies in genre $t$ as $G_t$ and define $R_i$ be the set of all movie rated by user $i$, in which for each movie $v \in R_i$, $r_i(v)$ is a corresponding rating. Here we define a user based task which involves 5 users ratings and the related objective functions; formally ${\mathcal{T}}_j=\{ \{f_{j_i}\}_{i=1}^{5},\{R_{j_i}\}_{i=1}^{5}\}$ which $f_{j_i}$ is monotone submodular function and $R_{j_i}$ is set of movie ratings by user $j_i$. In particular, we define $f_i$ as in equation . $$\label{eq: mov recom obj2}
f_i(S)=\sum_{t=1}^{18} w_{i,t}.\max_{v\in R_i\cap G_t \cap S} r_i(v)$$ which is the weighted average over maximum rate that user $i$ gives to movies from each genre and $w_{i,t}$ is proportion of movies in genre $t$ which is rated by user $i$ out of all the rating he provides. We divide 200 users to two parts the first 100 users for training, and the second 100 users for test. We show the results in figure \[fig:simulation-movrecom\]. As we can see the performance of meta-Greedy is really close to Greedy test while it is faster. For example, in simulation of Figure the average time for Greedy test algorithm is 4.33 sec per running, and meta-Greedy is 0.91 sec per running while the performance of meta-greedy is close to Greedy test.
Counter-example for Submodularity of the Objective in {#sec:counter_example}
======================================================
In this section, we provide a counterexample for submodularity of the objective function in the equation . We consider a maximum coverage problem in which the function value is an area covered by a set of elements. We define the ground set $V=\{ABIJ,BCDI,ACDJ,IDEH\\,HEFG,BCEH\}$ which has shown in Figure \[fig:Counter Example\]. Each element is a rectangle, and a function value of that element is an area covered by that element. We refer to each element (rectangle) by it’s vertices.
Let $AC=CD=DE=EF=1$, and $BC=0.75$. Also in we let $m=1$ and $k-l=1$ which means that we are considering a single set function $f$ defined as: $f(S)=\max_{e \in V} A(S\cup e)$, where $A(T)$ is a area of set $T$. Note that the area function $A$ is monotone and submodular, however as we will show below, the function $f$ is not submodular. To do so, we consider two sets $T_1=\emptyset$ and $T_2=\{ACDJ\}$ and add the element $IDEH$ to both sets and observe that $f$ does not satisfy the diminishing returns property. Let us first compute the function value at $T_1$ and $T_2$ as follows: $$f(T_1)=\max_{e \in V} A(e)=A(\{BCEH\})=1.5,$$ and $$f(T_2)=\max_{e \in V} A(T_2\cup e)=A(\{ACDJ,IDEH\})=1.75.$$ Similarly, we compute the function value at $T_1^{'}=T_1\cup \{IDEH\} $, and $T_2^{'}=T_2\cup \{IDEH\} $: $$f(T_1^{'})=\max_{e \in V} A(T_1^{'}\cup e)=A(\{IDEH,ACDJ\})=1.75,$$ and $$f(T_2^{'})=\max_{e \in V} A(T_2^{'}\cup e)=A(\{IDEH,ACDJ,EFGH\})=2.5.$$ We can now see that $T_1\subseteq T_2$, but $f(T_2^{'})-f(T_2)\not \leq f(T_1^{'})-f(T_1)$. Therefore, $f$ is not submodular.
Also let us make a remark about $k$-submodularity which studies functions of $k$ subsets of the ground set that are disjoint sets. This class of functions is submodular in each orthant [@ohsaka2015monotone]. However, in the submodular meta-learning framework, sets can have overlap, and there is no restriction on the sets to be disjoint. Therefore, our framework is different from $k$-submodular maximization.
Objective Value
------------------------ ----------------- --
Greedy-Train 0.896822
Greedy-Test 0.963639
Random 0.839599
Meta-Greedy 0.949555
Randomized Meta-Greedy 0.945020
Replacement Greedy 0.910977
: Comparison of two-stage framework and submodular meta-learning framework
### Two-Stage Submodular Meta-Learning
Although, our submodular meta-learning has the ability of personalizing the solution in test time but it is still use the same ${S_{\rm{tr}}}$ for every task. But learning all the set ${S_{\rm{tr}}}\cup S_i$ is computationally expensive in the test time. Therefore, we need to find the way to personalize ${S_{\rm{tr}}}$ too without increasing the computational power. In order to that, we come up with following approach: Let assume similar to two-stage framework we learn reduced ground set in. training but instead of choosing the. whole set as a subset of this set instead we just choose ${S_{\rm{tr}}}$ as subset of this set and we let $S_i$ be personalized at the test time. Formally, we try to solve the following problem:
\[eq:two-stage sub meta-learning\] $$\begin{aligned}
{4}
& \max_{S,\tilde{S_i}} && \sum_{i=1}^{m}\; \max_{\mid S_i\mid \leq k-l}\; f_i(S_i\cup \tilde{S_i})
\label{eq: LPRelaxation2}\\
&\operatorname*{subject\,\, to \quad}&&\tilde{S_i}\subseteq S\\ \; \;
& &&\mid \tilde{S_i}\mid \leq l\;\;\\
& &&\mid S\mid \leq q\;\;\\ \end{aligned}$$
Note that the optimal value of this problem is higher than related two-stage submodular maximization and related meta-submodular maximization problem; since, if $q=l$ this problem reduced to meta-submodular problem and if $k=l$ this problem reduced to two-stage submodular maximization.
### Two-Stage Submodular Meta-Learning Algorithm
For solving the problem \[eq:two-stage sub meta-learning\] the natural way that come to mind is to some how combining the replacement greedy algorithm and meta-Greedy algorithm but how we can do that? One way to doing that is to fill $S_i$ greedily and then using the replacement greedy to obtain $S,\tilde{S_i}$ then we can use $S$ as reduced ground set in test time.
**Initialize** ${S_{\rm{tr}}}=\{S_i\}_{i=1}^m=\{\tilde{S}_i\}_{i=1}^m=\emptyset$ Find $e_i^{*} = \operatorname*{arg\,max}_{ e\in V} \Delta_i(e|S_i)$ $S_i\xleftarrow{} S_i\cup\{e_i^{*}\}$ **end for** **end for** Find $ e^{*} =\!\operatorname*{arg\,max}_{ e\in V} \!\sum\limits_{i=1}^{m}\!\nabla_{i}(e|\tilde{S_i} \cup S_i)$ $S\xleftarrow{} S\cup\{e^{*}\}$ $\tilde{S_i}\xleftarrow{} \tilde{S_i}\cup\{e^{*}\}\setminus \text{REP}_i(e^{*},\tilde{S_i})$ **end for**
Return $\tilde{S_i}$, $S$ and $S_i$
In the above algorithm $\nabla_{i}(e|S)=\Delta_i(e|S)$ if $\mid S\mid\not=l $, and $\nabla_{i}(e|S)=\max\{0,\max_{x \in S}{\nabla_i{(e,x,S)}}\}$, and $\text{REP}_i(e,S)=\operatorname*{arg\,max}_{x \in S}{\nabla_i{(e,x,S)}}$ if $\mid S\mid=l $ and is $\emptyset$ if $\mid S\mid\not=l $. Finally, $\nabla_i{(e,x,S)}=f_i(e\cup S\setminus x)-f_i(S)$.
### Analysis of Two-Stage submodular Meta-Learning Algorithm
First suppose $S_i$ and $\tilde{S_i}$ and $S$ is a output of algorithm \[alg: two-stage meta-Greedy\]; because, the first phase of algorithm \[alg: two-stage meta-Greedy\] is same as first phase of algorithm \[alg: Reverse discrete meta-Greedy\] we can have the same analysis as the one for which results:
$$\label{eq: two-stage eq1 pf}
\sum_{i=1}^m f_i(\tilde{S_i}^{*} \cup S_i^{*}) - \sum_{i=1}^m f_i( S_i)\leq \sum_{i=1}^m f_i(\tilde{S_i}^{*} \cup S_i)$$
Also by analysis of replacement greedy in [@stan2017probabilistic], we can obtain the following: $$\begin{aligned}
\sum_{i=1}^m f_i(\tilde{S_i}\cup S_i)-f_i( S_i)&\geq \frac{1}{2}(1-\frac{1}{e^2})(\sum_{i=1}^m f_i(\tilde{S_i}^{opt}\cup S_i)-f_i(S_i))
\label{eq: eq3 pf two-stage} \\&\geq \frac{1}{2}(1-\frac{1}{e^2})(\sum_{i=1}^m f_i(\tilde{S_i}^{*}\cup S_i)-f_i(S_i))
\label{eq: eq3 pf two-stage }\\&\geq \frac{1}{2}(1-\frac{1}{e^2})(\sum_{i=1}^m f_i(\tilde{S_i}^{*}\cup S_i^{*})-2f_i(S_i))
\label{eq: eq3 pf two-stage}\end{aligned}$$ Also $S_i\subseteq S_i\cup \tilde{S_i}$ and monotonicity $\sum_{i=1}^m f_i(\tilde{S_i}\cup S_i)\geq \sum_{i=1}^m f_i( S_i)$. therefore, if we let $\theta=\sum_{i=1}^m f_i( S_i)$: $$\sum_{i=1}^m f_i(\tilde{S_i}\cup S_i)\geq \max\{ \theta,\frac{1}{2}(1-\frac{1}{e^2})(\text{OPT}-2\theta)+\theta\}$$ then we can write the as follows:
$$\begin{aligned}
\nonumber
\mathbb{E}[(1-\frac{1}{k})^\tau]&=\mathbb{E}[(1-\frac{1}{k})^\tau|\tau^{''}>\tau^{'}]p(\tau^{''}>\tau^{'})+\mathbb{E}[(1-\frac{1}{k})^\tau|\tau^{''}<\tau^{'}]p(\tau^{''}<\tau^{'})
\\&=\mathbb{E}[(1-\frac{1}{k})^{\tau^{'}}|\tau^{''}>\tau^{'}]p(\tau^{''}>\tau^{'})+\mathbb{E}[(1-\frac{1}{k})^{\tau^{''}}|\tau^{''}<\tau^{'}]p(\tau^{''}<\tau^{'})
\\&=\sum_{r=1}^{k}(1-\frac{1}{k})^r(p(\tau^{'}=r)+p(\tau^{''}=r))
\\&=\sum_{r=l}^{k}(1-\frac{1}{k})^r({r-1 \choose l-1}(\frac{k-l}{k})^
{r-l}(\frac{l}{k})^{l}\nonumber\\&\quad+\sum_{r=k-l}^{k}{r-1 \choose k-l-1}(\frac{l}{k})^
{r-k+l}(\frac{k-l}{k})^{k-l})\\&\leq\sum_{r=l}^{k}\exp(\frac{-r}{k}){r-1 \choose l-1}(\frac{k-l}{k})^
{r-l}(\frac{l}{k})^{l}\nonumber\\&\quad+\sum_{r=k-l}^{k}\exp(\frac{-r}{k}){r-1 \choose k-l-1}(\frac{l}{k})^
{r-k+l}(\frac{k-l}{k})^{k-l}\end{aligned}$$
Then using moment generating function(MGF) of negative binomial(NB) distribution we have :
$$\begin{aligned}
\sum_{r=l}^{k}\exp(\frac{-r}{k}){r-1 \choose l-1}(\frac{k-l}{k})^
{r-l}(\frac{l}{k})^{l}&=\left(\frac{1-\frac{l}{k}}{1-\frac{l}{k}\exp({-\frac{1}{k}})}\right)^l \exp(-\frac{l}{k})\nonumber\\&- \sum_{r=k+1}^{\infty}\exp(\frac{-r}{k}){r-1 \choose l-1}(\frac{k-l}{k})^
{r-l}(\frac{l}{k})^{l}\nonumber\\&\leq \frac{k-l}{k}\exp\left(-\frac{l^2}{k}(1-\exp(-\frac{1}{k}))\right)\nonumber\\&\quad-\exp({-{I_{\frac{l}{k}}}(k+1-l,l+1)})
\\&\leq \frac{k-l}{k}\exp\left(\frac{l^2}{k}((\exp(-1)-1)\frac{1}{k})\right)\nonumber\\&\quad-\exp({-{I_{\frac{l}{k}}}(k+1-l,l+1)})\end{aligned}$$
where the first equality comes from definition of MGF and the first inequality comes from Jensen inequality and definition of CDF of NB distribution and $I_x$ is regularized incomplete beta function. The second inequality comes from the fact that $-1+\exp(-1/k)\leq (\exp(-1)-1)1/k$.\
similarly:
$$\begin{aligned}
\sum_{r=k-l}^{k}\exp(\frac{-r}{k}){r-1 \choose k-l-1}&(\frac{l}{k})^
{r-k+l}(\frac{k-l}{k})^{k-l}=\left(\frac{1-\frac{k-l}{k}}{1-\frac{k-l}{k}\exp({-\frac{1}{k}})}\right)^{k-l} \exp(-\frac{k-l}{k})\nonumber
\\&- \sum_{r=k+1}^{\infty}\exp(\frac{-r}{k}){r-1 \choose k-l-1}(\frac{l}{k})^
{r-k+l}(\frac{k-l}{k})^{k-l}\nonumber\\&\leq \frac{l}{k}\exp\left(\frac{(k-l)^2}{k}(1-\exp(-\frac{1}{k}))\right)
\\&-\exp({-{I_{\frac{k-l}{k}}}(l+1,k+1-l)})
\\&\leq \frac{l}{k}\exp\left(\frac{(k-l)^2}{k}((\exp(-1)-1)\frac{1}{k})\right)-\exp({-{I_{\frac{k-l}{k}}}(l+1,k+1-l)})\end{aligned}$$
summing up we can write: $$\begin{aligned}
\mathbb{E}[(1-\frac{1}{k})^\tau]&\leq\frac{l}{k}\exp\left((\frac{k-l}{k})^2((\exp(-1)-1))\right)\notag\\&+(\frac{k-l}{k})\exp\left((\frac{l}{k})^2((\exp(-1)-1)\right)\notag\\&-\exp({-{I_{\frac{k-l}{k}}}(l+1,k+1-l)})\notag\\&-\exp({-{I_{\frac{l}{k}}}(k+1-l,l+1)})\leq \exp\left(((\exp(-1)-1))(1-c)^2\right).\end{aligned}$$ where c is $\max\{k-l/k,l/k\}$.
[^1]: Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA, USA. {[email protected]}.
[^2]: Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, USA. {[email protected]}.
[^3]: Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA, USA.{[email protected]}.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the dynamical evolution of a large amplitude $r$-mode by numerical simulations. $R$-modes in neutron stars are unstable growing modes, driven by gravitational radiation reaction. In these simulations, $r$-modes of amplitude unity or above are destroyed by a catastrophic decay: A large amplitude $r$-mode gradually leaks energy into other fluid modes, which in turn act nonlinearly with the $r$-mode, leading to the onset of the rapid decay. As a result the $r$-mode suddenly breaks down into a differentially rotating configuration. The catastrophic decay does not appear to be related to shock waves at the star’s surface. The limit it imposes on the $r$-mode amplitude is significantly smaller than that suggested by previous fully nonlinear numerical simulations.'
author:
- Philip Gressman
- 'Lap-Ming Lin[^1]'
- 'Wai-Mo Suen[^2]'
- 'N. Stergioulas'
- 'John L. Friedman'
title: 'Nonlinear $r$-Modes in Neutron Stars: Instability of an unstable mode'
---
The $r$-modes of rotating neutron stars are unstable growing modes driven by gravitational radiation reaction [@Andersson:1; @Friedman:1]. If the $l=m=2$ $r$-mode of a young, rapidly rotating star can grow to an amplitude of order unity, the gravitational radiation it emits would carry away most of the star’s angular momentum and rotational kinetic energy; and the radiation might be detectable by the Laser Interferometric Gravitational Wave Observatory (LIGO II) [@Owen] (for recent reviews see [@Andersson:2; @Friedman:2]). Even if its amplitude were smaller, the $r$-mode instability would limit the periods of hot, young neutron stars (and, possibly, of old stars spun up by accretion). A number of mechanisms to damp the mode have been examined, including shear viscosity enhanced by crust-core coupling and by nonstandard cooling [@crust]; bulk viscosity enhanced by a hyperon-rich core [@hyperon]; and energy loss to a magnetic field driven by differential rotation [@diff_rot]. But none of these definitively eliminates the instability.
The significance of the $r$-mode instability depends strongly on its maximum possible amplitude. In two recent Letters, Stergioulas and Font [@Nick:1] and Lindblom, Tohline, and Vallisneri [@LB:1] performed numerical simulations of nonlinear $r$-modes, both finding that large-amplitude nonlinear $r$-modes can exist for some long period of time. In addition, in [@LB:1; @LB:2], Lindblom [*et al.*]{} carried out numerical simulations of the growth of the $r$-modes driven by the current quadrupole post-Newtonian radiation reaction force in Newtonian hydrodynamics. In order to achieve a significant growth of the $r$-mode amplitude in a reasonably short computational time, they artificially multiplied the radiation reaction force by a factor of 4500. This decreases the growth time of the $r$-mode from about 40 s to 10 ms. The (dimensionless) $r$-mode amplitude $\alpha$ grew to $\approx 3.3$ before shock waves appeared on the surface of the star and the $r$-mode amplitude collapsed. Lindblom [*et al.*]{} suggested that the nonlinear saturation amplitude of the $r$-modes may be set by dissipation of energy in the production of shock waves. Here, we show that a hydrodynamical effect will restrict the $r$-mode amplitude to a value significantly below that reported in [@LB:1; @LB:2].
We note that, with the artificially large radiation reaction, the results in [@LB:1; @LB:2] assume that no hydrodynamic process takes energy from the $r$-mode in a time scale between 10 ms and 40 s (the artificial growth time and the actual physical growth time, respectively). In this paper we investigate the evolution of large amplitude $r$-modes in these time scales (10 ms-40 s). We find that (i) a catastrophic decay of the mode, occuring within these time scales, significantly reduces the amplitude to which an $r$-mode can grow, and (ii) in a large amplitude $r$-mode this catastrophic decay leads to a differentially rotating configuration. (As in Refs. [@LB:1; @LB:2], we assume a perfect fluid with no magnetic fields.)
We solve the Newtonian hydrodynamics equations for a non-viscous fluid flow in the presence of gravitational radiation reaction: $${\partial \rho\over \partial t}+\nabla\cdot\left(\rho \vec{v}\right)=0 \ ,
\label{eq:rho}$$ $$\rho\left({\partial \vec{v}\over \partial t}+ \vec{v}\cdot\nabla\vec{v}\right)
=-\nabla P - \rho\nabla \Phi + \rho \vec{F}_{\rm GR}\ ,
\label{eq:Euler}$$ where $\rho$ is the density, $P$ is the pressure, $\vec{v}$ is the velocity, $\vec{F}_{\rm GR}$ is the radiation reaction force per unit mass, and $\Phi$ is the Newtonian potential, satisfying $$\nabla^2\Phi = 4 \pi G\rho \ .
\label{eq:possion}$$
A high resolution shock capturing scheme (Roe solver) is used to solve the hydrodynamic equations. In addition, as in [@Nick:1], we applied the 3rd order piecewise parabolic method (PPM) [@Collela84] for the cell-reconstruction process in order to simulate rapidly rotating stars accurately for a large number of rotational periods.
As $r$-modes couple to the gravitational radiation mainly through the current multipoles $J_{lm}$, we assume that the contribution to the reaction force $\vec{F}_{\rm GR}$ comes solely from the dominant current multipole $J_{22}$, an approximation used also in [@LB:1; @LB:2]. The resulting expression for $\vec{F}_{\rm GR}$ is given by (see [@Luc:1; @Rez; @LB:1]) $$F_{\rm GR}^x - iF_{\rm GR}^y = -\kappa i\left(x+iy\right)
\left[ 3v^z J_{22}^{(5)} + zJ_{22}^{(6)} \right] \ ,$$ $$F_{\rm GR}^z = -\kappa \ {\rm Im}\left\{\left(x+iy\right)^2\left[
3{v^x+iv^y\over x+iy}J_{22}^{(5)} + J_{22}^{(6)} \right]\right\} \ ,$$ where $J_{22}^{(n)}$ is the nth time derivative of $J_{22}$ and $\kappa =\kappa_{0}\equiv 32\sqrt{\pi}G/(45\sqrt{5} c^7)$. A technical difficulty in evaluating $\vec{F}_{\rm GR}$ is that it depends on high-order time derivatives of $J_{22}$. To circumvent this problem, we assume that $J_{22}^{(n)}=(i\omega)^{n}J_{22}$, with the nonlinear $r$-mode frequency $\omega$ defined by $\omega = - |J_{22}^{(1)}| / |J_{22}|$, as in [@LB:1; @LB:2]. Notice that $J_{22}^{(1)}$ can be expressed as an integral over the fluid variables and is thus calculated on each time slice [@LB:1; @LB:2; @Rez].
To investigate the dynamical properties of a large amplitude $r$-mode numerically, one must first generate initial data. However, analytically we know only the expression of an $r$-mode in the linear regime, under the assumption of $\alpha \ll 1$ and $\Omega_0
\rightarrow 0$. In particular, the $l=m=2$ $r$-mode perturbation (at the first order of $\Omega_0$) is given by [@LB:3] $$\delta \vec{v}= \alpha_{0}R_0 \Omega_{0}\left(r\over R_0 \right)^2
\vec{Y}_{22}^{B} e^{i\omega t} \ ,
\label{eq:delta_v}$$ where $\alpha_0$ is a dimensionless amplitude, $R_0$ and $\Omega_0$ are the radius and angular velocity of the unperturbed rotating star model, and $\vec{Y}_{22}^{B}$ is the magnetic-type vector spherical harmonic defined by $\vec{Y}_{lm}^{B}=[l(l+1)]^{-1/2} \vec r \times \nabla Y_{lm}$. Using Eq. (\[eq:delta\_v\]) with a large $\alpha_0$ introduces other modes in the initial data [@LB:2]; that is, the resulting configuration is not the same as an $r$-mode growing from small amplitude driven by gravitational radiation reaction. In the following, we will show how we obtain a large amplitude $r$-mode in our study.
Beyond the linear regime, there is no unique mode decomposition, and hence no unique definition of an $r$-mode. In this paper, by a large amplitude $r$-mode we mean a configuration resulting from the growth of an infinitesimal $r$-mode due to gravitational radiation reaction. We define the amplitude $\alpha$ of the nonlinear $l=m=2$ mode in terms of its contribution to $J_{22}=\int{\rho r^2 \vec{v}\cdot
\vec{Y}_{22}^{B*} d^3x}$: $$\begin{aligned}
&&\alpha(t) \equiv
\left|{{1\over R^3} \int{ \tilde{\alpha}(x) e^{i\phi (x)} d^3x }}\right|
\ , \cr
&&\cr
&&\tilde{\alpha}(x) e^{i\phi (x)} \equiv
{8\pi R^4 \left(\rho r^2 \vec{v}\cdot \vec{Y}_{22}^{B*}\right)\over
\bar{\Omega}(t)\int{\rho r^4 d^3x} } \ .
\label{eq:alpha} \end{aligned}$$ Here $R$ is the radius of the corresponding nonrotating star model and $\tilde{\alpha}(x)$ is the amplitude density. The phase factor $\phi(x)$ is defined so that $\tilde{\alpha}(x)$ is real. The definition of the amplitude is similar to that of [@LB:1], except that we normalize $\alpha(t)$ by the average angular velocity of the star $\bar{\Omega}(t)$ instead of a fixed initial $\Omega_0$.
![Evolution of $\alpha$ in a slowly rotating star with the correct ($\kappa_0$) and artificial ($9\times 10^7 \kappa_0$) radiation reaction. The solid lines represent the numerical results ($257^3$ resolution), while the dashed lines are the predictions from linear theory.[]{data-label="fig:257_slowtest"}](fig1.ps){width="5.0cm"}
In Fig. \[fig:257\_slowtest\], we show the time evolution of $\alpha$ for two cases in a slowly rotating star with $T=4.42\ {\rm ms}$. In the first case $\kappa$ is set to the correct Post-Newtonian value $\kappa_0$ (represented by the solid line labeled “$\kappa_0$”). The evolution begins with a small (linear) $r$-mode perturbation given by Eq. (\[eq:delta\_v\]) with $\alpha_0=0.1$. The simulation is carried out up to $t=0.6\ {\rm ms}$. To compare, we plotted as dashed line the evolution of $\alpha$ as predicted by the linear theory [@LB:3]: $\alpha=\alpha_0 e^{t/ \tau_{\rm GR} }$, where the gravitational radiation time scale is given by $${1\over \tau_{\rm GR} } = {G\Omega_{0}^6\over 2 c^7}
\left({256\over 405}\right)^2 \left({\kappa\over\kappa_0}\right)
\int{\rho r^4 d^3x} \ .$$ We note that this formula is correct only to linear order in $\alpha$ and in $\Omega$, and is thus accurate only for (i) $\alpha<<1$ and (ii) $\Omega_0/\sqrt{G\bar\rho} <<1$. While assumption (i) is reasonably good, (ii) is actually not accurate for the model used: rotation period 4.42 ms, corresponding to $\Omega / \Omega_{\rm max} \approx0.25$ ($\Omega_{\rm max}\approx
2\sqrt{\pi G\bar{\rho} }/3$ is the Kepler limit, where $\bar{\rho}$ represents the average density of the star). Nevertheless, the two lines nearly coincide.
With radiation reaction coefficient $\kappa_0$, an evolution time $t\approx 3\times 10^{8}\ {\rm ms}$ is needed to reach $\alpha=1$. For a $129^3$ grid-point simulation, this would take $O(10^{11})$ time steps, requiring a clearly impractical $O(10^8)$ hours on a 128 CPU Origin 2000 (MIPS R12000).
To arrive at a large amplitude $r$-mode we use an artificially large $\kappa$ as in [@LB:1; @LB:2]. In Fig. \[fig:257\_slowtest\], the solid line labeled “$9\times 10^7 \kappa_0$” represents the evolution of $\alpha$ with $\kappa = 9\times 10^7 \kappa_0$. For comparison, we also plot the evolution of $\alpha$ as predicted by the linear theory with this $\kappa$ as the dashed line. The slight offset between the dashed and solid lines at the initial time is due to the fact that $\alpha(t=0)$ as calculated from Eq. (\[eq:alpha\]) equals to $\alpha_0$ as defined in Eq. (\[eq:delta\_v\]) only in the limit $\alpha_0 \ll 1$ and $\Omega_0 \rightarrow 0$.
In Fig. \[fig:4500\_2.2\] (left), we show the growth of an $r$-mode (with $\alpha_0=0.5$) to a large amplitude with an artificial $\kappa$ of $4500 \kappa_0$ in our fast rotating star model: The star has a mass of $M=1.64M_{\odot}$ with a polytropic equation of state $P=k\rho^2$. The equatorial radius $R_{e}=14.5\ {\rm km}$. The ratio of the polar to equatorial radii is 0.76. The rotation period is $T=1.24$ ms. This model is used for the rest of the simulations discussed in this paper. Unless otherwise noted, we use $129^3$ Cartesian grid points with $\Delta x=\Delta
y=\Delta z=0.42\ {\rm km}$.
![Left: Growth of the $r$-mode amplitude $\alpha$ during artificial “pumping” with $\kappa=4500\kappa_0$. Right: The velocity profile $v^y$ along the $x$ axis at $t=0$ and at the point when $\alpha=2.0$.[]{data-label="fig:4500_2.2"}](fig2a.ps "fig:"){width="3.8cm"} ![Left: Growth of the $r$-mode amplitude $\alpha$ during artificial “pumping” with $\kappa=4500\kappa_0$. Right: The velocity profile $v^y$ along the $x$ axis at $t=0$ and at the point when $\alpha=2.0$.[]{data-label="fig:4500_2.2"}](fig2b.ps "fig:"){width="4.2cm"}
In Fig. \[fig:4500\_2.2\] (left), the amplitude $\alpha$ rises from 0.5 to 2.2 in 19 ms. An indicator of the accuracy of the simulation is that the total mass of the system is constant to 0.08% by 20 ms in our $129^3$ runs. Also, the actual numerical evolution of the total angular momentum $J=|\int{\rho \vec{r} \times \vec{v} d^3x}|$ agrees with the theoretical prediction \[see Eq. (11) of Ref. [@LB:1]\] to about 1%.
While the discussion above shows the accuracy of our numerical treatment, one still must ask whether the large amplitude $r$-mode obtained with the large artificial pumping is physical or not, in the sense that whether the rapid pumping excites modes that would not be excited with $\kappa = \kappa_0$. We compare the large amplitude $r$-modes obtained with different pumping rates and conclude that the resulting fluid flow pattern does not depend sensitively on the pump rate (as long as the pump rate is large enough so that the large amplitude $r$-mode can be arrived at). In Fig. \[fig:4500\_2.2\] (right), we show the rotational velocity profile $v^y$ along the $x$ axis for $\kappa = 9000,\ 4500 \kappa_0$ at the point when $\alpha =2.0$, starting with the same initial model. For comparison, we also plot the initial profile in the same figure. We see that the two lines, $\kappa=9000,\ 4500\kappa_0$, agree with each other to better than 3%, with smaller discrepancy away from the surface. In the rest of this paper, we will take the “pumped up” configurations given in Fig. \[fig:4500\_2.2\] (left) as the initial state in our investigation of the hydrodynamical behavior of the large amplitude $r$-mode. To the extent that different pumping rates are not affecting the initial state we use, the artificial pumping is not affecting the results we report below.
Here we focus on one question: What is the fate of a large-amplitude $r$-mode in a rapidly rotating neutron star for a sufficiently long time evolution, modeled as the 1.24 ms polytrope described above. That is, what evolution is implied by Eqs. (\[eq:rho\])-(\[eq:possion\]), with the correct amount of radiation reaction?
![Left: Evolution of the “pumped up” $\alpha$ at different values with $\kappa=\kappa_0$. Right: Fourier spectra of $v^z$ along the $x$ axis (at $x=6$ km) at two time slots in the evolution starting out with $\alpha=1.6$. The early (later) time slot is denoted by the dashed (solid) line.[]{data-label="fig:4500_alpha"}](fig3a.ps "fig:"){width="4.0cm"} ![Left: Evolution of the “pumped up” $\alpha$ at different values with $\kappa=\kappa_0$. Right: Fourier spectra of $v^z$ along the $x$ axis (at $x=6$ km) at two time slots in the evolution starting out with $\alpha=1.6$. The early (later) time slot is denoted by the dashed (solid) line.[]{data-label="fig:4500_alpha"}](fig3b.ps "fig:"){width="4.0cm"}
In Fig. \[fig:4500\_alpha\] (left) we show the evolution of $\alpha$ vs time for various large amplitude $r$-modes starting off with $\alpha=2.2,\ 2.0,\ 1.8,\ 1.6$. We see that the mode amplitudes start off slowly decaying, leaking energy to other modes. The decay rate is small, until a certain time. We plot in Fig. \[fig:4500\_alpha\] (right) the Fourier transform of the velocity component $v^z$ along the $x$ axis at a typical point inside the star at two different time slots in the evolution of the $\alpha=1.6$ case \[the lowest line in Fig. \[fig:4500\_alpha\] (left)\]. The dashed line is the profile at the beginning of the evolution (13 ms-35 ms), we see that there is only one large peak at the $r$-mode frequency (0.93 kHz). This is compared to the solid line representing the spectrum at a later time slot (35 ms - 50 ms). We see that various smaller peaks appear in the spectrum, especially the one at twice the $r$-mode frequency (1.86 kHz). Spectra at different points inside the star give similar structure, with those in the core region showing more peaks at different frequencies.
The most interesting feature of Fig. \[fig:4500\_alpha\] (left) is that after some slow leaking of energy into other fluid modes, the $r$-mode amplitude drops catastrophically to a value much smaller than 1. This abrupt drop occurs through nonlinear couplings with other fluid modes: In the slow leaking phase, these other fluid modes are growing linearly until a certain unstable point. The time it takes to reach the unstable point depends sensitively on the $r$-mode amplitude. It shortens from approximately 45 ms to 8 ms when the initial value of $\alpha$ changes from 1.6 to 2.2.
To further investigate this catastrophic decay we perform a set of numerical experiments in which we pump energy into the $r$-mode by turning on the artificial radiation reaction force with coefficient $\kappa =4500\kappa_0$ whenever its amplitude drops down below its initial value. The resulting evolution of $\alpha$ vs time is given in Fig. \[fig:65\_alpha\]. The evolution tracks of $\alpha$ starting off with values $\alpha=2.1,\ 1.9,\ 1.7,\ 1.5,\ 1.4,\ 1.0$ are given. With the large artificial pumping $\alpha$ remains constant despite energy leaking to other modes. In all cases (except $\alpha=1.0$ where the simulation is not evolved long enough), however, the hydrodynamical nonlinear interaction eventually overwhelms the artificial pumping, and the $r$-mode amplitude falls catastrophically.
![Evolution of $\alpha$ ($65^3$ resolution) with artificial pumping of $\kappa = 4500 \kappa_0$ whenever $\alpha$ drops below its initial value.[]{data-label="fig:65_alpha"}](fig4.ps){width="4.8cm"}
![The amplitude density $\tilde{\alpha}(x)$ on the equatorial plane before (left) and after (right) the breakdown for the case where the initial “pumped up” $\alpha$ is 2.0 in Fig. \[fig:4500\_alpha\].[]{data-label="fig:2.0alpha_den"}](fig5a.ps "fig:"){width="4.0cm"} ![The amplitude density $\tilde{\alpha}(x)$ on the equatorial plane before (left) and after (right) the breakdown for the case where the initial “pumped up” $\alpha$ is 2.0 in Fig. \[fig:4500\_alpha\].[]{data-label="fig:2.0alpha_den"}](fig5b.ps "fig:"){width="4.0cm"}
In Fig. \[fig:2.0alpha\_den\] we compare the distribution of the amplitude density $\tilde{\alpha}(x)$ as defined in Eq. (\[eq:alpha\]) on the equatorial plane before (left) and after (right) the breakdown respectively for the case where the initial $\alpha$ is 2.0 (the $\alpha =2.0$ line in Fig. \[fig:4500\_alpha\]). In the figure, the brighter region represents higher amplitude density. During the catastrophic decay, the $r$-mode pattern changes rapidly from a 4-fold “regular” shape (left) to a whirlpool-like spiral (right). We also see in our simulations that strong differential rotation is developed during the breakdown, a potentially important fact regarding whether subsequent re-growth of the $r$-mode is possible or not [@Karino].
In Fig. \[fig:vy\_diffrot\] we plot the rotational velocity profile $v^y$ along the $x$ axis for the case where the initial $\alpha$ is 2.0. The solid line is the initial profile, while the dashed line is the profile after the decay. To further quantify the amount of differential rotation, we define the kinetic energy associated to differential rotation by $I= {1\over 2}\int{ \rho \left( v_{\phi} - \bar{v}_{\phi}\right)^2
d^3x }$, where $\bar{v}_{\phi}= \bar{\Omega}\sqrt{x^2+y^2}$ with $\bar{\Omega}$ being the average angular velocity of the star. We plot the evolution of $\alpha$ and $I$ together in Fig. \[fig:diff\_rot\] (left). It is seen that the amount of differential rotation ($I$) rises rapidly during the breakdown of $\alpha$. We also see in our simulations that the star has a relatively large amplitude pulsation during the breakdown. Fig. \[fig:diff\_rot\] (right) shows the quadrupole-moment component $Q_{xy}$ against time for the same case as in Fig. \[fig:diff\_rot\] (left). We see that $Q_{xy}$ is basically zero until the breakdown and it then oscillates rapidly afterward. However, based on an order-of-magnitude estimation, the gravitational radiation amplitude due to the changing quadrupole moment is only about 1% of that due to the $r$-mode.
![The rotational velocity profile $v^y$ along the $x$ axis for the case where the initial “pumped up” $\alpha$ is 2.0. The solid line is the initial profile at $\alpha=2.0$, while the dashed line is the profile after the breakdown.[]{data-label="fig:vy_diffrot"}](fig6.ps){width="4.5cm"}
In contrast to the study of Refs. [@LB:1; @LB:2], we do not see evidence that this catastrophic decay is due to the generation of shock waves on the surface of the star.
![Left: Evolution of $\alpha$ (dashed) and $I$ (solid) for the case where the initial “pumped up” $\alpha$ is 2.0. Note that $I$ has been rescaled for comparison. Right: Evolution of $Q_{xy}$ (in $G=c=M_{\odot}=1$ units) for the same case.[]{data-label="fig:diff_rot"}](fig7a.ps "fig:"){width="3.9cm"} ![Left: Evolution of $\alpha$ (dashed) and $I$ (solid) for the case where the initial “pumped up” $\alpha$ is 2.0. Note that $I$ has been rescaled for comparison. Right: Evolution of $Q_{xy}$ (in $G=c=M_{\odot}=1$ units) for the same case.[]{data-label="fig:diff_rot"}](fig7b.ps "fig:"){width="4.1cm"}
We found in this paper that a large amplitude $r$-mode will lose energy to other fluid modes whose growth in turn trigger a catastrophic decay of the $r$-mode. For an $r$-mode with an amplitude $\alpha$ of order one, the onset of catastrophic decay requires a time much shorter than the growth time of the $r$-mode due to radiation reaction; and the decay thus limits the $r$-mode amplitude to a value less than that found by [@LB:1; @LB:2]. Further work is in progress to determine the nature of this catastrophic decay, whether it is related to any known hydrodynamical instability of nonlinear flow, and how large an $r$-mode amplitude can be with this effect taken into account.
[*Note added.*]{} Towards the end of the preparation of this paper we learned the results of Arras [*et al.*]{} [@Arras]. We note the following differences between the hydrodynamical phenomenon studied in this paper and that studied by them: The work of Arras [*et al.*]{} points to a slow leakage of the $r$-mode energy into some short wavelength oscillation modes, leading to an equilibrium distribution of mode amplitudes. This in turn, through viscosity dissipation, limits the $r$-mode amplitude to a small value. In this paper we find a sudden and complete breakdown of the $r$-mode that operates independent of viscosity. Further numerical investigation will be carried out to investigate the interactions of $r$-modes with other oscillation modes. Such investigation is beyond the resolution power of our present simulations. Note that, if the conclusions of [@Arras] are correct, there may be no astrophysical situation in which $r$-modes grow to amplitudes large enough to exhibit the sudden decay seen in our simulations.
We thank N. Andersson, G. Comer, E. Evans, S. Iyer, L. Lindblom, M. Miller, Y. Mino, B. Owen and C. Will for useful discussions. The simulations in this paper made use of the following code components: Newton\_evolve (Newtonian gravity and evolution) by P. Gressman, PPM (PPM reconstruction) by T. Font, Newton\_anal (modules for various Newtonian analyses) by L.-M. Lin, and the Cactus computational toolkit by T. Goodale [*et al.*]{} The research is supported by NSF Phy 00-71044, 00-96522, 99-79985 (KDI Astrophysics Simulation Collaboratory Project), NRAC MCS93S025, the EU Programme “Improving the Human Research Potential and the Socio-Economic Knowledge Base” (Research Training Network Contract HPRN-CT-2000-00137), KBN-5P03D01721, and support from the NASA AMES NAS.
[1]{}
N. Andersson, Astrophys. J. [**502**]{}, 708 (1998).
J. L. Friedman and S. M. Morsink, Astrophys. J. [**502**]{}, 714 (1998).
B. J. Owen [*et al.*]{}, Phys. Rev. D [**58**]{}, 084020 (1998).
N. Andersson and K. D. Kokkotas, Int. J. Mod. Phys. D [**10**]{}, 381 (2001).
J. L. Friedman and K. H. Lockitch, gr-qc/0102114.
N. Andersson, D. I. Jones, K. D. Kokkotas, and N. Stergioulas, Astrophys. J. [**534**]{}, L75 (2000); M. Rieutord, astro-ph/0003171; L. Lindblom, B. J. Owen and G. Ushomirsky, Phys. Rev. D [**62**]{}, 084030 (2000); Y. Levin and G. Ushomirsky, Mon. Not. R. Astron. Soc. [**324**]{}, 917 (2001); Y. Wu, C. D. Matzner, and P. Arras, Astrophys. J. [**549**]{}, 1011 (2001); G. Mendell, Phys. Rev. D [**64**]{}, 044009 (2001).
P. B. Jones, Phys. Rev. Lett. [**86**]{}, 1384 (2001); P. B. Jones, Phys. Rev. D [**64**]{}, 084003 (2001); L. Lindblom and B. J. Owen, Phys. Rev. D [**65**]{}, 063006 (2002); P. Haensel, K. P. Levenfish, and D. G. Yakovlev, astro-ph/0110575.
H. C. Spruit, Astron. Astrophys. [**341**]{}, L1 (1999); L. Rezzolla, F. K. Lamb, and S. L. Shapiro, Astrophys. J. [**531**]{}, L139 (2000); L. Rezzolla, F. K. Lamb, D. Markovic,and S. L. Shapiro, Phys. Rev. D [**64**]{}, 104013 and 104014 (2001).
N. Stergioulas and J. A. Font, Phys. Rev. Lett. [**86**]{}, 1148 (2001).
L. Lindblom, J. E. Tohline, and M. Vallisneri, Phys. Rev. Lett. [**86**]{}, 1152 (2001).
L. Lindblom, J. E. Tohline, and M. Vallisneri, Phys. Rev. D [**65**]{}, 084039 (2002).
P. Collela and P. R. Woodward, J. Comput. Phys. [**54**]{}, 174 (1984).
L. Blanchet, Phys. Rev. D [**55**]{}, 714 (1997).
L. Rezzolla [*et al.*]{}, Astrophys. J. [**525**]{}, 935 (1999).
L. Lindblom, B. J. Owen, and S. M. Morsink, Phys. Rev. Lett. [**80**]{}, 4843 (1998).
S. Karino, S. Yoshida, and Y. Eriguchi, Phys. Rev. D [**64**]{}, 024003 (2001).
P. Arras [*et al.*]{}, astro-ph/0202345.
[^1]: Corresponding author.
[^2]: Also at Department of Physics, The Chinese University of Hong Kong, Hong Kong.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'From integral field data we extract the optical spectra of 20 shocked clouds in the supernova remnant N132D in the Large Magellanic Cloud (LMC). Using self-consistent shock modelling, we derive the shock velocity, pre-shock cloud density and shock ram pressure in these clouds. We show that the \[Fe X\] and \[Fe XIV\] emission arises in faster, partially radiative shocks moving through the lower density gas near the periphery of the clouds. In these shocks dust has been effectively destroyed, while in the slower cloud shocks the dust destruction is incomplete until the recombination zone of the shock has been reached. These dense interstellar clouds provide a sampling of the general interstellar medium (ISM) of the LMC. Our shock analysis allows us to make a new determination of the ISM chemical composition in N, O, Ne, S, Cl and Ar, and to obtain accurate estimates of the fraction of refractory grains destroyed. [From the derived cloud shock parameters, we estimate cloud masses and show that the clouds previously existed as typical self-gravitating Bonnor-Ebert spheres into which converging cloud shocks are now being driven.]{}'
author:
- |
Michael A. Dopita , Frédéric P.A. Vogt , Ralph S. Sutherland ,\
Ivo R. Seitenzahl , Ashley J. Ruiter & Parviz Ghavamian
title: |
Shocked Interstellar clouds and dust grain destruction\
in the LMC Supernova Remnant N132D
---
Introduction {#intro}
============
The supernova remnant (SNR) N132D in the Large Magellanic Cloud (LMC) is the brightest X-ray emitting SNR in this galaxy [@Hwang93; @Favata97; @Hughes98; @Borkowski07; @Xiao08]. Located within the stellar bar of the LMC, N132D was first identified as a SNR by @Westerlund66 on the basis of the association of a non-thermal radio source with a \[\]– bright optical structure. @Danziger76 discovered high-velocity \[\] and \[\] material ejected [ in the supernova explosion]{}, which established N132D as a relatively young SNR [ in which the reverse shock has not yet reached the centre of the remnant, marking the formal transition to the Sedov phase of evolution. This would imply that the swept-up interstellar medium (ISM) is less than a few times the mass ejected in the supernova event, a result which is in apparent contradiction to the swept-up mass derived from X-ray observations [@Hughes98].]{} From the dynamics of the fast-moving O and Ne material, @Danziger76 derived a maximum age of 3440yr for the SNR, and a probable age of $\sim 1350$yr. Subsequent analyses gave similar ages; 2350yr [@Lasker80], and 2500yr [@Vogt11].
Although most optical studies have concentrated on the fast-moving O- and Ne-rich ejecta [@Lasker80; @Blair00; @Vogt11], there are a number of luminous, dense shock clouds with apparently ‘normal’ interstellar composition and $\sim200$km/s velocity dispersion. The most prominent of these is the so-called *Lasker’s Bowl* structure in the northern part of the remnant, but HST imaging [@Blair00] reveals many other small complexes of shocked cloudlets. These clouds do not show appreciable enhancements in the strength of the \[\] lines, and are thought to be simply ISM clouds over which the SNR blast wave has recently swept. In this paper we follow the study of N49 by @Dopita16 in investigating the degree of dust destruction in these cloud shocks. We also derive the physical parameters for some 20 different shocked ISM cloudlets and estimate their chemical abundances.
Observations & Data Reduction {#sec:obs}
=============================
The integral field spectra of N132D were obtained between 16 Nov 2017 and 24 Nov 2017 using the Wide Field Spectrograph (WiFeS) [@Dopita07; @Dopita10], an integral field spectrograph mounted on the 2.3-m ANU telescope at Siding Spring Observatory. This instrument delivers a field of view of 25$\times$ 38at a spatial resolution of either 1.0$\times$ 0.5 or 1.0$\times$ 1.0, depending on the binning on the CCD. In these observations, we operated in the binned 1.0x 1.0 mode. The data were obtained in the low resolution mode $R \sim 3000$ (FWHM of $\sim 100$ km/s) using the B3000 & R3000 gratings in each arm of the spectrograph, with the RT560 dichroic which provides a transition between the two arms at around 560nm. For details on the various instrument observing modes, see @Dopita07.
All observations are made at PA=0, giving a long axis in the N-S direction. The basic grid consists of 15 pointings in an overlapping $3\times5$ grid (E-W:N-S), followed by 7 pointings centered on the overlap regions of the base $3\times5$ grid, with an additional pointing in the SE to probe the extent of the photoionised precursor region around N132D. The eighth overlap position was not observed due to deteriorating weather conditions. The typical seeing over the course of the observations was 1.5 arc sec. and ranged from 0.5 to 2.5 arc sec. over the individual exposures. Each region was observed with $2\times1000$s exposure time giving a total integration time on target of 44000s. We used blind offsets from a reference star to move to each field of the mosaic consistently over the course of the observing run, and avoid any gaps in the resulting mosaic. A fault in the offset guide head gave rise to errors of $\lesssim 5$arc sec. in the pointing of the fields; small enough not to result in any gaps within the mosaic.
The wavelength scale is calibrated using a series of Ne-Ar arc lamp exposures, taken throughout the night. Arc exposure times are 50s for the B3000 grating and 1s for the R3000 grating. Flux calibration was performed using the STIS spectrophotometric standard stars HD009051, HD031128 and HD075000 [^1]. In addition, a B-type telluric standard HIP8352 was observed to better correct for the OH and H$_2$O telluric absorption features in the red. The separation of these features by molecular species allows for a more accurate telluric correction by accounting for night to night variations in the column density of these two species. All data cubes were reduced using the PyWiFeS [^2] data reduction pipeline [@Childress14; @Childress14b].
All the individual, reduced cubes were median-averaged into a red and blue mosaic using a custom <span style="font-variant:small-caps;">python</span> script. The WCS solution of both mosaics is set by cross-matching their individual (collapsed) white-light images with all entries in the *Gaia* [@GaiaCollaboration16a] DR1 [@GaiaCollaboration16]; see Fig \[fig1\]). For simplicity, and given the spaxel size of the WiFeS datacubes, we restrict ourselves to integer shifts when combining fields. We estimate that the resulting absolute and relative (field-to-field) alignment accuracy are both $\pm$1arc sec.
HST imaging of the cloudlets
----------------------------
We have downloaded all the HST ACS F475W and F658N images from the *Barbara A. Mikulski archive for space telescopes* (MAST). The former filter encompasses the full velocity range of the \[\]$\lambda\lambda$5959,5007Å emission, whereas the latter covers the rest-frame H$\alpha$ emission. The dataset comprises four exposures of 380s each with the F475W filter, and four exposures of 360s each using the F658N filter, all acquired in January 2004 under program \#12001 (PI Green).
We used the individual calibrated and CTE-corrected frames (<span style="font-variant:small-caps;">\*\_flc.fits</span>) retrieved from MAST to construct a combined, drizzled image in both filters using the following steps. We first correct the WCS solutions of each frame using the <span style="font-variant:small-caps;">tweakreg</span> routine from the <span style="font-variant:small-caps;">drizzlepac 2.1.13</span> package via a dedicated <span style="font-variant:small-caps;">python</span> script. We anchor the WCS solution to the *Gaia* [@GaiaCollaboration16a] DR1 [@GaiaCollaboration16] sources present within the images, retrieved automatically via the <span style="font-variant:small-caps;">astroquery</span> module. The WCS-corrected images are then merged using the <span style="font-variant:small-caps;">astrodrizzle</span> routine, with a pixel scale set to 0.04arcsec for both filters (for simplicity).
{width="\textwidth"}
Spectral Extraction
-------------------
The spectra of 20 prominent shocked ISM clouds were extracted from the global WiFeS mosaic datacubes using using [QFitsView v3.1 rev.741]{}[^3]. The positions of the clouds were chosen so as to largely avoid contamination by high-velocity \[
\] – emitting material, [ see Fig \[fig2\]]{}. The exception is P07, but the velocity shift of the high velocity material here is sufficient as to allow accurate determinations of the narrow-line fluxes. We used a circular extraction aperture with a radius of either 2 or 3 arc sec. to best match the size of the bright region. To remove the residual night sky emission and (approximately) the faint stellar contribution, we subtracted a mean sky reference annulus 1 arc sec. wide surrounding the extraction aperture. The extraction regions were optimised by peaking up the signal in H$\alpha$ in the red data cube, and in H$\gamma$ in the blue data cube, respectively. This procedure may lead to some contamination by the faint extended cloud emission in the case of P07, P08, P14, P15, P18 and P20 (see Figure \[fig3\]). However, the core region strongly dominates such fainter contributions. The positions of the extracted spectra, the extraction radius used, the measured mean radial velocity, and the H$\alpha$ velocity width (FWHM) (after correction for the instrumental resolution) of each cloud are listed in Table \[table1\], and the corresponding positions are shown on the HST ACS H$\alpha$ + \[O III\] image in Figure \[fig2\] [ and each cloud with its spectral extraction region is shown in an $8\times8$arc sec. postage stamp image in Figure \[fig3\]]{}. A spectrum of the bright cloudlet P14 is shown in Figure \[fig4\] to indicate the quality of the data. Note that the typical velocity FWHM of these clouds in only $\sim 180$km/s, which is fairly indicative of the mean shock velocity in those clouds [Fig.6 in @Dopita12].
The mean Heliocentric radial velocity of all 20 clouds is +246km/s, and the mean Heliocentric radial velocity of the photoionised halo of N132D is found to be +261km/s, a difference which lies within the expected statistical error caused by sampling of the cloud shock velocities projected along the lines of sight within the random cloud positions inside the SNR. These figures can be compared to the +264.6km/s [@Smith71], $+262\pm16$km/s [@Danziger76], and +259km/s [@Feast64] for N127, a nearby HII region.
Since most of the H$\alpha$ emission arises in the recombination zone of the cloud shocks, we can estimate a lower limit for the cloud shock velocities, $v_s$, from the most extreme differences in the measured radial velocity of the cloud from the mean Heliocentric radial velocity of the photoionised halo of N132D. These are +173km/s for cloud P04, and -143km/s for cloud P19. Thus we have $v_s \gtrsim 160$km/s, in good agreement with the estimate based on the FWHM of the H$\alpha$ line profiles.
Measuring Emission Line Fluxes
------------------------------
For each extracted spectrum, the spectra were first reduced to rest wavelength based on their measured radial velocities listed in Table \[table1\], and then emission-line fluxes [ in units of erg/cm$^2$/s]{}, their uncertainties, the emission line FWHMs [ (in Å)]{} and the continuum levels were measured using the interactive routines in [Graf]{} [^4]and in [Lines]{} [^5]. The measured line fluxes [ relative to H$\beta$]{} and their uncertainties [ along with the absolute H$\beta$ fluxes in units of erg/cm$^2$/s]{}, are given for each of the clouds in Tables \[tableA1\] – \[tableA4\] in the Appendix.
Emission Line Diagnostics {#diagnostics}
=========================
Self-Consistent Shock Modelling
-------------------------------
To analyse the spectrophotometry we have built a family of radiative shocks with self-consistent pre-ionisation using the MAPPINGS 5.12 code, following the methodology described in @Sutherland17, and applied to the study of Herbig-Haro objects by @Dopita17b. The pre-shock density is set by the ram pressure of the shock, taken to be independent of shock velocity; $P_{\mathrm{ram}} = 1.5\times 10^{-7}$dynes cm$^{-2}$, to ensure that the measured \[\] densities (which give the electron density near the recombination zone of the shock) approximately match those produced by the models. For each abundance set, a set of models was run for $100 \leqslant v_s \leqslant 475$km/s in steps of 25km/s. The magnetic field pressure in the un-shocked cloud ahead of the photo-ionised precursor is assumed to be in equipartition with the gas pressure, $P_{\mathrm{mag}} = P_{\mathrm{gas}}$, and the temperature of the gas entering the shock is given by the self-consistent pre-ionisation computation; see @Sutherland17 for details.
The abundance set was initially taken as being 0.5 times the Local Galactic Concordance (LGC) values [@Nicholls17], but was then iterated manually to achieve a better fit of theory to the spectra of the clouds on the standard @BPT81 and @Veilleux87 diagnostic diagrams. The depletion factors of the heavy elements caused by the condensation of these elements onto dust are defined as the ratio of the gas phase abundance to the total element abundance. The depletion factors are derived from the formulae of @Jenkins09, extended to the other elements on the basis of their condensation temperatures and/or their position on the periodic table. In these shock models, following @Dopita16 we have investigated the effect of changing the logarithmic Fe depletion, $\log D_{\rm Fe}$ in the range $0.0 > \log \left [ D_{\rm Fe} \right ] > -1.0$.
The abundance set adopted for the theoretical grid at the various values of $\log D_{\rm Fe}$ is given in Table \[table2\]. The abundances of C, Mg, and Si are not constrained by our observations and for these we fix the abundances at half of the LGC values. We produce a more refined estimate of the LMC chemical abundances when we build detailed models for the brightest clouds, below.
\
The line intensities and line ratios given in this paper comprise the sum of the radiative shock and its photoionised precursor. In the case of the fastest shock, the extent of the photoionised precursor region and/or the cooling length of the shock may exceed the physical extent of the pre-shocked cloud. In Figure \[fig5\] we show the cooling length of the shock to 1000K, the depth of the photoionised precursor to the point where hydrogen is only 1% ionised, and the fraction of the shock H$\beta$ emission which arises from the precursor. The cooling length and precursor length are given for a ram-pressure of $P_{\mathrm{ram}} = 1.0\times 10^{-7}$dynes cm$^{-2}$. The shock cooling length scales inversely as the ram pressure, and the precursor length remains approximately constant. At the distance of the LMC 0.1pc corresponds to 0.45arc sec.
![The computed cooling length of the shock to $T_e = 1000$K (thick line), the thickness of the photoionised precursor to the point where hydrogen is only 1% ionised (thin line), and the fraction of the shock H$\beta$ emission which arises from the precursor (dotted line). These are all given for a ram pressure of $P_{\mathrm{ram}} = 1.0\times 10^{-7}$dynes cm$^{-2}$. At the densities of those models, the shock cooling length scales inversely as the ram pressure, while the precursor length is almost unaffected by changes in the ram pressure. []{data-label="fig5"}](fig5.pdf)
Shock Velocity Diagnostics
--------------------------
Here we investigate potentially useful diagnostics of the shock velocity using emission line ratios. To avoid issues of chemical abundance difference, the line ratio used should involve different levels of the same ion, or else different ions of the same element. In @Dopita16, we already noted that the \[\] $\lambda\lambda 4363/5007$ ratio was a useful indicator of shock velocity. This is because the \[\] temperature in the cooling zone of the shock is high at low shock velocities (faster than $\sim85$km/s, when this ion is first produced in the cooling zone of the shock), reaching a maximum at $v_s \sim 140$km/s when starts to become ionised to higher ionisation stages. As the velocity increases further, is ionised to , and the \[\] emission becomes confined to the cooler region nearer the recombination zone of the shock. At still higher velocities, a zone photoionised by EUV photons generated in the cooling zone of the shock develops adjacent to the recombination zone. This has still lower mean temperature; $T_e\lesssim 10^4$K. For these physical reasons, the \[\] temperature is more or less a decreasing function of shock velocity, and is relatively insensitive to $\log D_{\rm Fe}$. This behaviour is shown in Figure \[fig6\].
A second diagnostic is the excitation of He, as measured by the / $\lambda\lambda 4686/5876$ ratio. This is shown in the second panel of Figure \[fig6\]. Below $v_s \sim 200$km/s, this ratio is determined by the pre-ionisation of the shock, and the temperature structure of the post-shock region, and is multi-valued. However, above this velocity, it provides a useful diagnostic, albeit somewhat sensitive to $\log D_{\rm Fe}$. This sensitivity is caused by changes in the cooling length with gas-phase Fe abundance, which changes the ionisation structure of He in the shock. In what follows, we will use both the He and \[\] diagnostic ratios as spectroscopic indicators of the shock velocity.
BPT diagnostic diagrams
-----------------------
For shocks, the position of the observations on the standard @BPT81 and @Veilleux87 diagnostic diagrams (hereafter BPT Diagrams), is indicative of the mode of excitation. However, they are not really useful as diagnostics of the detailed shock conditions. This is demonstrated in Figure \[fig7\]. It should be noted that the range in observed line ratio and in the theoretical grid is very small on these diagrams, compared with the range covered by either HII regions or active galaxies and Narrow Line Regions. All these diagrams serve to show is that the theoretical models and the observations show a satisfactory overlap.
This said, a few points on these diagrams are worthy of further mention. First P13 is the only region which shows an excess in the line intensity of \[\], which may be indicative of pre-supernova mass-loss enrichment. However, this seems a little unlikey as the region is close to the boundary of the SNR shell. Second, P20 shows extremely strong \[\] emission. This is evidenced by the blue appearance of this region in Figure \[fig3\]. This region will be discussed in more detail below, where we show that it is due to a finite-age shock in which the gas has not had time to become fully radiative. Lastly, P08 shows particularly weak \[\] emission. This region, also discussed below, lies in a complex of Balmer dominated or non-radiative filaments in which a high velocity non-radiative shock is passing through an un-ionised or partially-ionised precursor medium.
\[\], \[\] and \[\] diagnostics
--------------------------------
As can be seen from Tables \[tableA1\] - \[tableA4\] the \[\] spectrum is very rich. For the purposes of analysis, we have selected the brightest of these lines, \[\]${\lambda 5158}$ and \[\]${\lambda 8617}$. In Figure \[fig8\] we present the observed points superimposed on our theoretical shock model grid. From this, it is immediately apparent that nowhere does the measured iron depletion factor; $\log D_{\rm Fe}$ exceed -0.5. Indeed many of the points are consistent with no depletion at all.
Furthermore, it is clear that the indicated shock velocities in general exceed 200km/s. Thus for most of these the \[\] line ratio can be used to estimate the shock velocity. However, a number of the points lie in the ambiguous region of the / line ratio velocity diagnostic. To the extent to which it is possible, both shock velocities and $\log D_{\rm Fe}$ have been estimated for each cloud from these diagnostic diagrams. The results are given in Tables \[table3\] and \[table4\]. We find a mean shock velocity $\left< v_s \right >$ of 240 km/s, and $\left < \log D_{\rm Fe} \right > = -0.15$. This result agrees very closely with the $\left < \log D_{\rm Fe} \right > = -0.16\pm0.07$ found by @Dopita16 in N49, another bright LMC SNR, and shows that dust has been largely destroyed by the time the recombination zone of the shock has been reached.
A similar analysis was applied to the \[\]$\lambda 5270$/H$\beta$ and \[\]$\lambda 4881$/H$\beta$ line ratios. This is shown in Figure \[fig9\]. The two lines agree less well with each other, the \[\]$\lambda 4881$/H$\beta$ ratio tending to produce a larger depletion factor. In each line ratio, a much wider range of depletion factors is observed, from -0.25 to $\sim -1.25$. Averaging all measurements, we obtain $\left < \log D_{\rm Fe} \right > = -0.74\pm0.25$. This result is also broadly consistent with what was derived in the case of N49; $\left < \log D_{\rm Fe} \right > = -0.956\pm0.15$ [@Dopita16].
Finally, the analysis for \[\]$\lambda 6087$/H$\alpha$ is shown in Figure \[fig10\]. Note that this line requires shock velocities of $v_s > 180$km/s in order to be produced with appreciable strength in the radiative shock. Here, the inferred depletion factors range from -0.0 to $\sim -0.75$ with $\left < \log D_{\rm Fe} \right > = -0.4$.
Given the intrinsic uncertainties in the atomic data for the forbidden lines of Fe, we can conclude that the \[\] and the \[\] lines imply a mean depletion factor of order $\left < \log D_{\rm Fe} \right > = -0.6\pm0.25$, while the \[\] lines give a depletion factor $\left < \log D_{\rm Fe} \right > = -0.15\pm0.15$. Thus, in both of these N132D and in N49, we find that Fe-bearing dust grains are not fully destroyed until they reach the recombination zone of the shock. The detail of the grain destruction process were discussed in @Dopita16, where it was concluded that this type of depletion pattern supports the grain destruction model of @Seab83 and @Borkowski95.
\[Ni II\] diagnostics
---------------------
Since we also observe strong \[\]$\lambda 7378$, it is worth asking whether the Ni-containing grains suffer the same fate as the Fe-containing grains. The diagnostic plots for \[\]/H$\alpha$ are given in Figure \[fig11\]. We find a depletion factor of $\sim -0.4$, intermediate between the \[\] and \[\] results, but again consistent with a large fraction of dust grains being destroyed in the shock.
The highly ionised species of Fe
--------------------------------
To produce the highly ionised species of Fe; \[\]${\lambda 8235}$, \[\]${\lambda 6374}$ and \[\]${\lambda 5303}$ requires shock velocities $v_s \gtrsim 200$km/s. Indeed, the self-consistent radiative shock models do not produce strong \[\]${\lambda 6374}$ until $v_s \sim 250$km/s and \[\]${\lambda 5303}$ becomes bright only for $v_s > 350$km/s. The diagnostics of expansion velocity, velocity dispersion, and the spectral diagnostics presented above all point to shock velocities for typical clouds in the range $160$km/s$ < v_s < 300$km/s. Thus it is clear that the relatively strong \[\] and \[\] lines observed must arise in another region with higher shock velocity.
This is not unexpected, given the appearance of the clouds on the HST images, which often show a bright core in H$\alpha$, surrounded by fainter filamentary structures. We can identify these regions with the bright radiative shocks being driven into the denser cores of these clouds, and a faster, but only partially-radiative shock driven by a similar external ram-pressure sweeping through the less dense outer regions of the same clouds. Indeed in the case of another SNR in the LMC; N49, both the spatial and dynamical distinction between the \[\]${\lambda 5303}$ and the \[\]${\lambda 5159}$ emission are very marked [cf. Fig.3 in @Dopita16].
The detailed diagnostic diagrams for these two Fe species are shown in Figure \[fig12\]. For the standard grid, both fail to reproduce the observed strength of the lines, but the \[\] is particularly bad. Introducing a partially-radiative shock leads to a dramatic improvement. The derivation of the plausible shock parameters [ for these shocks]{} is as follows.
First, we note that the mean cloud shock ram pressure derived from the observations (see Section \[cloudshocks\], below) is $P_{\mathrm{ram}} = 3.3\times10^{-7}$dynes/cm$^2$. At the mean pre-shock particle density implied for these shocks, $n_0 \sim 240$cm$^{-3}$, the mean cooling age to $T_e = 1000$K is 400yr. This provides an initial estimate of the shock age. Another way is to take the age of the SNR as estimated from the \[\] dynamics of the fast-moving material and estimate from the position of the cloud within the SNR how long it has been since the blast wave overran the cloud. In Section \[intro\] we saw that estimates of the age of N132D vary from 1350 – 3440yr for the SNR, with a probable age of $\sim 2500$yr [@Danziger76; @Lasker80; @Vogt11]. The detailed position of the clouds within the SNR blast wave cannot be reliably estimated from their projected distances from the boundary of the shell, but we will take this as being $\sim 25$%, as an upper limit, giving a mean shock age of $\lesssim 600$yr.
With the shock ram pressure given above, the mean pre-shock density at a given shock velocity is fully determined; for shock velocities of 300, 350 and 400km/s the pre-shock particle densities are 160, 117 and 90cm$^{-3}$, respectively. None of these shocks can become fully radiative within this timescale. In the 600yr available, the 300km/s shock cools from $T_e=1.27\times 10^6$K to $T_e=6.06\times 10^5$K; the 350km/s shock from $T_e=1.72\times 10^6$K to $T_e=1.43\times 10^6$K and the 400km/s shock from $T_e=2.25\times 10^6$K to $T_e=2.07\times 10^6$K.
The optical spectra of these shocks only contain emission lines of hydrogen and helium, as well as lines of coronal species of S, Si, Ar, Fe and Ni. The 300km/s shock is too slow to reproduce the observed \[\]/\[\] ratio, while the 400km/s overproduces \[\] relative to \[\], and is somewhat too faint relative to the cloud shock and its precursor. The curves on Figure \[fig12\] are presented for the 350km/s shock. In this figure, we added the contribution of the fast shock to that of the cloud shock, assuming that the fast shock covers an area of either 3 times, or 10 times that of the cloud shock. We assume that dust has been fully destroyed by shattering and sputtering in these fast shocks. Adding this fast shock contribution, we now achieve a satisfactory fit with the observational diagnostics for both \[\] and \[\].
The WiFeS images of N132D support the hypothesis that the faster \[\] and \[\] - emitting shocks cover a greater area than the cloud shocks which are bright in H$\alpha$. This is shown in Figure \[fig13\], in which we have constructed an image in H$\alpha$ (red), \[\] (green) and \[\] (blue). In this image the clouds show up as red or mauve blobs, while the \[\] emission forms diffuse halos about these clouds, and is much more extensive. In general, the \[\] emission is much more closely spatially correlated with the *Chandra X-Ray Observatory* images [@Borkowski07], an effect which is also seen in both N49 in the LMC [@Dopita16] and in the SMC SNR; 1E 0102.2-7219 [@Vogt17].
Detailed Cloud Shock Modelling {#cloudshocks}
==============================
Cloud Shock Parameters {#cloudparms}
----------------------
From the generalised shock diagnostic diagrams, we now turn to the derivation of the shock parameters of the individual shocked ISM clouds in N132D. First, using the shock velocities estimated from both the \[\] and He II/He I diagnostic diagrams, the measured \[\] electron densities given by the \[\] $\lambda\lambda 6731/6717$ ratios, and the mean \[\] electron densities given from the model grid at the same shock velocity, we can estimate the pre-shock particle density, $n_{\mathrm{0}}$, and the ram pressure driving the shock; $P_{\mathrm{ram}}$. These are listed in Table \[table5\]. For the specific regions P08 and P09, the errors in the estimated shock velocities are great, as are the errors in $n_{\mathrm{0}}$. The region P20 is an incompletely radiative shock so therefore the measured \[\] electron density represents an under-estimate of the true density in the \[\] recombination zone.
In the final column of Table \[table5\] we give the time required for the post-shock plasma to cool to 1000K, $ t_{\mathrm{1000K}}$ measured in years. At this point, the optical forbidden lines are no longer emitted in the plasma, and hydrogen is less than 2% ionised, so the shock is essentially fully radiative. We may conclude that these cloud shocks are somewhat older than $ \left<t_{\mathrm{1000K}}\right> =410$yr (excluding P20).
[ The mean ram pressure for the clouds, $P_{\mathrm{ram}} = 3.1\times10^{-7}$dynes/cm$^2$ is much higher than the mean pressure in the X-ray plasma. This may be estimated using the proton densities and electron temperatures listed in @Williams06. This gives $7.6 \times 10^{-8}$ dynes cm$^{-2}$ in the NW region, and $5.3 \times 10^{-8}$ dynes cm$^{-2}$ in the S, for an average of $6.4 \times
10^{-8}$ dynes cm$^{-2}$. Thus the ram pressure in the cloud shocks is typically $\sim5$ times higher than in the surrounding X-ray plasma. The reason for this difference is discussed in Section \[conc\], below. ]{}
The mean pre-shock density inferred for these clouds is surprisingly high; $\left<n_{\mathrm{0}} \right> = 243$cm$^{-3}$ (excluding P08, P09 and P20). However, the pressure in the ISM of the LMC has been estimated using the excited CI emission along a number of sight lines by @Welty16. This analysis gave $\left<\log P/k \right> =4.02$ cm$^{-3}$K. If the N132D clouds were characterised by the same ISM pressure before they were engulfed by the supernova shock wave (which is not at all a certain assumption), then their equilibrium temperature would have been $\sim 40$K, and they can be therefore considered to be part of the cold neutral medium phase of the ISM.
Chemical Abundances & Depletion Factors
---------------------------------------
We selected the four brightest clouds; P12, P14, P15 and P18 for detailed study, with the aim of deriving more precise shock parameters, and to derive the chemical abundances of the elements in the ISM of the LMC through high-precision shock modelling.
In order to measure the goodness of fit of any particular model, we measure the degree to which it reproduces the density-sensitive \[SII\] $\lambda\lambda 6731/6717$ ratio, and we also seek to minimise the L1-norm for the fit. That is to say that we measure the modulus of the mean logarithmic difference in flux (relative to H$\beta$) between the model and the observations *viz.*; $${\rm L1} =\frac{1}{m}{\displaystyle\sum_{n=1}^{m}} \left | \log \left[ \frac{F_n({\rm model})} {F_n({\rm obs.)}} \right] \right |. \label{L1}$$
This procedure weights fainter lines equally with stronger lines, and is therefore more sensitive to the values of the input parameters. We simultaneously fit to those emission lines which are most sensitive to the controlling parameters of shock velocity and ram pressure; effectively HeI, HeII, \[OI\], \[OII\] and \[OIII\] as well as the \[NI\], \[NII\], and \[SII\] lines. In addition, we estimate the gas-phase heavy element abundances using the bright \[FeII\] and \[FeIII\] lines, \[FeV\] 4227Å \[FeVII\] 6087Å \[NiII\] 7378Å MgI\] 4561Å, CaII 3933Å and \[CaII\] 7291Å lines. We assume that the more excited species of Fe are depleted in the gas phase by the same amount as FeIII. For the reason given above we do not attempt to fit either the \[FeX\] 6374Å or the \[FeXIV\] 5303 Å lines, given that they arise in a separate phase of the ISM. Other elements such as Ne, Ar and Cl are fit using the \[NeIII\] 3969Å \[ArIII\] 7136Å and \[ClII\] 8579Å lines.
Our models simultaneously [ optimise the fit to]{} 40 emission lines, and we manually iterate shock velocity, ram pressure and chemical abundances until the fit is optimised. The results for these four clouds are shown in Table \[table6\]. The L1-norm provides a good estimate for the error on the abundance of those elements fitted with only one or two emission lines. For O, N and S the accuracy is higher. In Table 2, we have adopted the abundances of Mg, Ca, Fe and Ni from @Russell92 in order to estimate the depletion factors of these elements.
The best fit was achieved without the inclusion of the precursor emission, which in principle could account for as much as 25% of the total emission. There are two possible reasons for this. First, the finite extent of the cloud may limit the length of the precursor, which is maximised in our plane-parallel models. Second, the local curvature of the shock front may allow significant escape of the upstream ionising UV photons. This is suggested by the fact that our models of these four clouds systematically overestimate the $\lambda 4686$/ $\lambda 5876$ ratio, which suggests that the incoming plasma is over-ionised by the computed precursor radiation field given by the model.
[ The derived depletion factors for Mg given in Table \[table6\] are likely to have very large, and unquantifiable errors attached to them. There is only one Mg line detected, and it is produced by neutral Mg in the tail of the recombination zone of the shock. Given the low ionisation potential of this species (7.64eV), the MgI\] line strength will be highly dependent on the unknown stellar UV radiation field, which by ionising this species would depress the line strength and lead to a larger depletion factor being inferred.]{}
For the remaining elements we derive the following mean LMC interstellar abundances, $12+\log(\mathrm{[X/H]})$, of He: $ 10.86\pm 0.05$, N: $ 7.12\pm 0.07$, O: $ 8.31\pm 0.04$, Ne: $ 7.44 \pm 0.08$, S: $ 7.01\pm 0.06$, Cl: $ 4.66\pm 0.11$ and Ar: $ 5.78\pm 0.11$. These abundances are generally similar to those obtained in our detailed shock fitting of the supernova remnant N49 [@Dopita16]; He: 10.92, N: 7.10, O: 8.46, Ne: 7.79, S: 7.05, Cl: 5.2 and Ar: 6.1.
These abundances should also be compared with those from @Russell92: He: $ 10.94\pm 0.03$, N: $ 7.14\pm 0.15$, O: $ 8.35\pm 0.06$, Ne: $ 7.61\pm 0.05$, S: $ 6.70\pm 0.09$, Cl: $ 4.76\pm 0.08$ and Ar: $ 6.29\pm 0.25$. Given the differences in the methodology used to derive these, and the changes in key atomic input parameters over the years, the agreement between these two is as good as could be expected. No systematic difference is seen for the elements most important in the modelling; N and O, but the derived He, Ne, S, Cl and Ar abundances are only marginally consistent with each other. What is now required is a re-derivation of the abundances using HII regions and the same modelling code to compare with the SNR-derived abundances.
The derived abundances can also be compared to more recent determinations from stellar observations. @Korn05 gave an new analysis of a slow-rotating B star in NGC2002 which gave N: $ 6.99\pm 0.2$, O: $ 8.29\pm 0.2$, and Mg: $ 7.44\pm 0.2$. These values differed very little from his earlier work [@Korn02].
With respect to the refractory elements, @Korn00 derived appreciably lower Mg and Fe abundances than @Russell92; Mg: $ 6.96\pm 0.22$ and Fe: $ 7.09\pm 0.15$. Using these values, our results are consistent with full destruction of the Fe-bearing grains in the \[\]–emitting zone, and the estimated Mg depletion factor is lowered to only $\log D_{\mathrm{MgII}} = -0.38$. However, the later @Korn02 paper gave Mg: $7.37\pm 0.06$ and Fe: $ 7.33\pm 0.0.03$, which is in much closer agreement with the @Russell92 estimates.
In conclusion, we can be fairly confident of the LMC He, N and O abundances; He: $ 10.92\pm 0.04$, N: $ 7.12\pm 0.09$ and O: $ 8.32\pm 0.06$. For the other non-refractory elements the recommended values are Ne: $ 7.52\pm 0.09$, S: $ 7.00\pm 0.15$, Cl: $ 4.70\pm 0.08$ and Ar: $ 5.78\pm 0.25$.
Our results are consistent with nearly full destruction of the Fe–containing grains in these 210–260km/s shocks, but the Mg–containing grains appear to be only partially destroyed. This argues for a different carrier for the two metals, possibly iron(II) oxide (FeO) for iron and the magnesium silicates, forsterite (Mg$_2$SiO$_4$) and enstatite (MgSiO$_3$) in the case of Mg.
The destruction of Fe–containing grains appears to have proceeded to a greater extent in N132D than in N49 [@Dopita16]. Although the estimated shock velocities are similar (210–260km/s in N132D and 200–250km/s in N49), the higher ram pressures and higher pre-shock densities in N132D (160–410cm$^{-3}$) compared with N49 ($\sim 80$cm$^{-3}$) has facilitated this greater fractional dust destruction.
In both N49 and in N132D, the grain destruction in the faster, partially-radiative shocks producing \[\] and \[\] emission is much more advanced. In both these SNR, the shock velocity in these highly-ionised zones is $\sim 350$km/s. Thus, thermal sputtering and non-thermal sputtering due to the high relative velocity of the dust grains relative to the hot-post shock plasma [@Barlow78; @McKee87; @Seab87; @Jones94] dominates the grain destruction in these fast shocks, while gyro-acceleration (aka betatron acceleration) caused by the compression of the magnetic field in the cooling zone of the shock resulting in grain shattering and vaporisation [@Spitzer76; @Borkowski95] appears to dominate as the grain destruction process in the slower, denser shocks.
The partially-radiative region P20
----------------------------------
We already noted in Section \[diagnostics\] that region P20 has anomalously strong \[\] emission – see Figure \[fig7\]. Furthermore, the \[\] lines indicate a low density compared with the other regions which produces the low value of the ram pressure given in Table \[table5\]. All of these point to a shock which has not yet become fully radiative. In this case, it appears that the photoionised precursor is brighter in the optical lines than the shock itself. The extremely strong \[\] would arise in this precursor.
The observed spectrum of this region proved very difficult to model. Our “best" model has a relatively poor fit, with an L1-norm of 0.21. This model has a $\sim 400$yr old shock with $v_s = 300$km/s moving into a medium with a pre-shock H density $n_0 = 235$cm$^{-3}$. In this shock the plasma cools from an initial temperature of $T_e =1.2\times10^6$K to a final temperature of $T_e =8.2\times10^5$K. The precursor material is assumed to have no dust destruction with $\log D_{Fe} = -1.00$, and the density of the pre-shock gas $n_0 = 235$cm$^{-3}$, and is illuminated by the UV radiation field of the partially-radiative shock in plane-parallel one-sided geometry. This gives a precursor thickness of $\sim 0.12$pc (at the point where H is 50% ionised). For this configuration, the shock produces only 4.5% of the total H$\beta$ luminosity. Table \[table7\] provides comparison of the key emission lines predicted by the model, compared with the observations.
Note the extreme discrepancy in the predicted strength of the \[\] $\lambda 4363$ line. From the model, the predicted temperature in the zone is $T_e = 13600$K, while the data suggest an electron temperature of $T_e \sim 49000$K in this zone. Similar, but smaller, temperature discrepancies are suggested for the and zones. This requires another form of heating in the precursor such as electron conduction, cosmic ray heating or else ionisation by the general EUV radiation field of the nebula. Apart from these temperature discrepancies, the \[\] $\lambda\lambda 6731/6717$ ratio also requires a higher electron density; $n_{\mathrm{[SII]}} \sim 760$cm$^{-3}$.
The Balmer-Dominated Shocks in P08
----------------------------------
We noted above that the region P08 also displays an anomalous spectrum, deficient in the forbidden lines with respect to the Balmer lines, with the \[\] lines being particularly weak – see Figure \[fig7\]. The cause of this appears to be “non-radiative” or Balmer-dominated shocks. These occur when a fast shock runs into a cold, essentially un-ionised ISM. A narrow component of the Balmer lines arises from direct collisional excitation of Hydrogen by the fast electrons, and a broad Balmer component is produced by charge exchange with fast protons behind the shock and subsequent collisional excitation of the fast neutral hydrogen resulting by the hot electrons [@Chevalier78; @Chevalier80] – see the reviews by @Heng10 and @Ghavamian13, and references therein. In Figure \[fig14\], we show the HST view of the P08 region, and the spectrum extracted from the WiFeS data cube. It is most likely that the broad component of H$\alpha$ seen in the spectroscopy arises from the filaments which are visible in the H$\alpha$ image, but not in the \[\] image. A broad component is also weakly detected in H$\beta$. This would add N132D to the select group of SNR in which this has been detected – which includes Tycho, SN1006, RCW86, the Cygnus Loop and N103B and SNR 0519-69.0 (DEM N71) in the LMC [@Tuohy82; @Ghavamian01; @Ghavamian02; @Ghavamian17]. [ Most, but not all, of these remnants are from Type Ia supernova explosions, since these do not produce a burst of strong EUV radiation to pre-ionise the pre-shock medium. However, if the ISM is shielded from the EUV flash in some way, or if it is dense enough to recombine in the period between the explosion and the arrival of the shock, then nothing precludes the formation of a Balmer-dominated shock. Given the inferred age of the SNR, the recombination condition would imply that the shock is passing through a medium with a density $ \gtrsim 20$cm$^{-3}$. The width of the H$\alpha$ line suggests that the shock velocity in the Balmer-dominated shock is of order $\sim900$km/s. This is similar to the blast wave velocity derived from the X-ray data [@Favata97; @Hughes98; @Borkowski07], so it seems extremely likely that the broad Balmer lines arise in the blast wave itself. ]{}
Conclusions {#conc}
===========
The environment of N132D provides an excellent sampling of shocked clouds in the Bar of the LMC. The typical physical size of these clouds is $\sim 1.0$pc. Given that they have a typical pre-shock density of $\sim240$cm$^{-3}$ (from Table \[table5\]), we can infer cloud masses of few solar masses. Using the images from Figure \[fig3\] and the pre-shock densities from Table \[table5\], we obtain masses in the range $0.1- 20$M$_{\odot}$ with a mean of $\sim 4$M$_{\odot}$. Thus we infer these clouds initially represented typical ISM self-gravitating Bonnor-Ebert spheres such as those recently investigated on a theoretical basis by @Sipila11, @Fischera14 and @Sipila17. Given that the shock which moves into them following the passage of the supernova blast-wave is strongly compressive, and that such Bonnor-Ebert spheres can be marginally stable against collapse, it is tempting to imagine that the supernova shock may later induce formation of $\sim 1.0 M_{\odot}$ stars within the cores of those clouds.
[ Now, let us consider the discrepancy between the thermal pressure measured by the X-rays, and the ram pressure driving the cloud shocks, pointed out in Section \[cloudparms\]. The X-ray plasma has a thermal pressure of $P_{\mathrm{therm}} = 6.4 \times 10^{-8}$ dynes cm$^{-2}$, while the ram pressure in the clouds is $P_{\mathrm{ram}} = 3.1 \times 10^{-7}$ dynes cm$^{-2}$. However, the pressure driving the cloud shocks is provided by the stagnation pressure behind the cloud bow shock or bow wave produced in the expanding thermal plasma which fills the SNR. This is always greater than the pressure in the pre-shock hot thermal plasma. The geometry of this interaction is pictured in @McKee75 [@Hester86]]{} and @Farage10. In the framework of a plane-parallel strong shock engulfing the cloud the stagnation pressure is about twice as large as the thermal pressure [@McKee75; @Hester86] and this is only very weakly dependent to the Mach number of the primary blast wave (see @Hester86, Table 2).
The X-ray data enable us to estimate the stagnation pressure in two ways. First, we may use the Sedov theory to estimate the blast wave velocity. Using the figures given in @Hughes98 for the explosion energy, age and pre-shock density in the SNR, we derive the blast-wave velocity, $v_{\mathrm B} = 830$km/s. Alternatively, we can use the measured thermal plasma temperatures to derive the blast wave velocity, using the relation $T_{\mathrm B} =(3 \mu m_H/16k)v_{\mathrm B}^2$, where $m_H$ is the pre-shock hydrogen density and the molecular weight $\mu$ is appropriate to a fully-ionised plasma. Estimates of the thermal temperature are in the range $0.68 < kT < 0.8$keV [@Favata97; @Hughes98; @Borkowski07], which implies a blast wave velocity of $\sim 760$km/s. Therefore, the ram pressure associated with expansion is $\sim 4 \times 10^{-8}$ dynes cm$^{-2}$. Adding this to the thermal pressure gives an estimate of the stagnation pressure, $\sim 1.0 \times 10^{-7}$ dynes cm$^{-2}$.
The remaining difference between the ram pressure of the cloud shocks and the estimated stagnation pressure is a factor of three. This difference may be accounted for by the fact that the cloud shocks are convergent towards the centre of mass, a possibility raised in the context of the Cygnus Loop SNR by @Hester86 as a promising mechanism for increasing pressure in shocked dense clouds. For shocks in self-gravitating isothermal spheres such as those considered here, families of self-similar solutions have been obtained by @Lou14. In such shocks, the energy density increases as the shock moves toward the center of the cloud, increasing the ram pressure. This effect can easily account for the difference between the measured ram pressure of the cloud shocks and the estimated stagnation pressure. Additionally, convergent cloud shocks are unstable in the presence of small deviations from sphericity [@Kimura90], and elongated clouds may break up into separate globules [@Kimura91]. Such instabilities are the likely reason for the complex morphologies of the shocks in individual clouds seen in Figure \[fig3\].
Because the shocked clouds of N132D represent gravitationally-confined samples of the ISM as they existed before the supernova event, and because these clouds are dense enough that they would not be appreciably affected in their chemical composition by any pre-supernova mass-loss, they present ideal samples of pristine ISM in the LMC. Our radiative shock analysis has enabled us to estimate accurate gas-phase chemical abundances for a number of elements.
From our model grid, we have demonstrated that the \[\] $\lambda\lambda 4363/5007$ ratio is a good indicator of shock velocity, and that the / $\lambda\lambda 4686/5876$ ratio may also be used for this purpose, although it is rather less reliable. Typical shock velocities are $\sim240$km/s, in agreement with the shock velocities inferred from the kinematics and the measured emission line widths. Using these emission line diagnostics, we have analysed the depletion onto dust in the various ionic stages of Fe, and in and . In common with the SNR N49, we find that dust has been mostly destroyed in the region emitting the \[\] lines, while a smaller fraction has been destroyed in the and zones, consistent with the grain destruction models of @Seab83 and @Borkowski95. However, shows an appreciably higher depletion factor, suggesting that the Mg silicates are more resistant to destruction than the Fe-bearing grains.
It is clear that the highly ionised species of iron; and originate in faster, partially radiative, filamentary and spatially extensive shocks surrounding the dense clouds. Some of these shocks may well be part of the primary blast wave of the SNR. In these, most of the refractory elements have been destroyed by thermal sputtering.
A detailed shock analysis of the four brightest clouds has allowed us to determine the chemical abundances of a number of elements. Comparing these with values given earlier, and with stellar abundance determinations we can now provide a “recommended” set of LMC abundances, which we present here in Table \[table8\]. Here, for completeness, the abundances of C, Mg, Si and Fe have been taken from the work of @Korn02 using Magellanic Cloud B-stars in NGC2004.
Finally we have identified two anomalous regions, P08 and P20. The former contains a contribution from fast Balmer-dominated shocks, while the latter represents an unusual partially-radiative shock, dominated by precursor emission which seems to be heated by an unknown source to very high electron temperatures.
Acknowledgments {#acknowledgments .unnumbered}
===============
MD and RS acknowledge the support of the Australian Research Council (ARC) through Discovery project DP16010363. Parts of this research were conducted by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO 3D), through project number CE170100013. FPAV and IRS thank the CAASTRO AI travel grant for generous support. IRS was supported by the ARC through the Future Fellowship grant FT1601000028. AJR is supported by the Australian Research Council through Future Fellowship grant FT170100243.
This research has made use of <span style="font-variant:small-caps;">matplotlib</span> [@Hunter07], <span style="font-variant:small-caps;">astropy</span>, a community-developed core <span style="font-variant:small-caps;">python</span> package for Astronomy [@AstropyCollaboration13], <span style="font-variant:small-caps;">APLpy</span>, an open-source plotting package for <span style="font-variant:small-caps;">python</span>[@Robitaille12], and <span style="font-variant:small-caps;">montage</span>, funded by the National Science Foundation under Grant Number ACI-1440620 and previously funded by the National Aeronautics and Space Administration’s Earth Science Technology Office, Computation Technologies Project, under Cooperative Agreement Number NCC5-626 between NASA and the California Institute of Technology.
This research has also made use of <span style="font-variant:small-caps;">drizzlepac</span>, a product of the Space Telescope Science Institute, which is operated by AURA for NASA, of the <span style="font-variant:small-caps;">aladin</span> interactive sky atlas [@Bonnarel00], of <span style="font-variant:small-caps;">saoimage ds9</span> [@Joye03] developed by Smithsonian Astrophysical Observatory, of NASA’s Astrophysics Data System, and of the NASA/IPAC Extragalactic Database [@Helou91] which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NNX09AF08G and by other grants and contracts. This research has also made use of NASA’s Astrophysics Data System.
This work has made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<https://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement. Some of the data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NNX09AF08G and by other grants and contracts.
natexlab\#1[\#1]{}\[1\][[\#1](#1)]{}
, [Robitaille]{}, T. P., [Tollerud]{}, E. J., [et al.]{} 2013, , 558, A33
, J. A., [Phillips]{}, M. M., & [Terlevich]{}, R. 1981, , 93, 5
, M. J. 1978, , 183, 367
, W. P., [Morse]{}, J. A., [Raymond]{}, J. C., [et al.]{} 2000, , 537, 667
, F., [Fernique]{}, P., [Bienaym[é]{}]{}, O., [et al.]{} 2000, , 143, 33
, K. J., & [Dwek]{}, E. 1995, , 454, 254
, K. J., [Hendrick]{}, S. P., & [Reynolds]{}, S. P. 2007, , 671, L45
, R. A., [Kirshner]{}, R. P., & [Raymond]{}, J. C. 1980, , 235, 186
, R. A., & [Raymond]{}, J. C. 1978, , 225, L27
, M., [Vogt]{}, F., [Nielsen]{}, J., & [Sharp]{}, R. 2014, [PyWiFeS: Wide Field Spectrograph data reduction pipeline]{}, Astrophysics Source Code Library, , , ascl:1402.034
, M. J., [Vogt]{}, F. P. A., [Nielsen]{}, J., & [Sharp]{}, R. G. 2014, , 349, 617
, I. J., & [Dennefeld]{}, M. 1976, , 207, 394
, M., [Hart]{}, J., [McGregor]{}, P., [et al.]{} 2007, , 310, 255
, M., [Rhee]{}, J., [Farage]{}, C., [et al.]{} 2010, , 327, 245
, M. A., [Payne]{}, J. L., [Filipovi[ć]{}]{}, M. D., & [Pannuti]{}, T. G. 2012, , 427, 956
, M. A., [Seitenzahl]{}, I. R., [Sutherland]{}, R. S., [et al.]{} 2016, , 826, 150
, M. A., & [Sutherland]{}, R. S. 2017, , 229, 35
, C. L., [McGregor]{}, P. J., [Dopita]{}, M. A., & [Bicknell]{}, G. V. 2010, , 724, 267
, F., [Vink]{}, J., [Parmar]{}, A. N., [Kaastra]{}, J. S., & [Mineo]{}, T. 1997, , 324, L45
, M. W. 1964, , 127, 195
, J. 2014, , 571, A95
, E. L. 1986, , 92, 1068
, [Prusti]{}, T., [de Bruijne]{}, J. H. J., [et al.]{} 2016, , 595, A1
, [Brown]{}, A. G. A., [Vallenari]{}, A., [et al.]{} 2016, , 595, A2
, P., [Raymond]{}, J., [Smith]{}, R. C., & [Hartigan]{}, P. 2001, , 547, 995
, P., [Schwartz]{}, S. J., [Mitchell]{}, J., [Masters]{}, A., & [Laming]{}, J. M. 2013, , 178, 633
, P., [Seitenzahl]{}, I. R., [Vogt]{}, F. P. A., [et al.]{} 2017, , 847, 122
, P., [Winkler]{}, P. F., [Raymond]{}, J. C., & [Long]{}, K. S. 2002, , 572, 888
, G., [Madore]{}, B. F., [Schmitz]{}, M., [et al.]{} 1991, in Astrophysics and Space Science Library, Vol. 171, Databases and On-line Data in Astronomy, ed. M. A. [Albrecht]{} & D. [Egret]{}, 89–106
, K. 2010, , 27, 23
Hester, J. J., & Cox, D. P. 1986, , 300, 675
, J. P., [Hayashi]{}, I., & [Koyama]{}, K. 1998, , 505, 732
, J. D. 2007, Computing in Science and Engineering, 9, 90
, U., [Hughes]{}, J. P., [Canizares]{}, C. R., & [Markert]{}, T. H. 1993, , 414, 219
, E. B. 2009, , 700, 1299
, A. P., [Tielens]{}, A. G. G. M., [Hollenbach]{}, D. J., & [McKee]{}, C. F. 1994, , 433, 797
, W. A., & [Mandel]{}, E. 2003, in Astronomical Society of the Pacific Conference Series, Vol. 295, Astronomical Data Analysis Software and Systems XII, ed. H. E. [Payne]{}, R. I. [Jedrzejewski]{}, & R. N. [Hook]{}, 489
, T., & [Tosa]{}, M. 1990, , 245, 365
—. 1991, , 251, 664
, A. J., [Becker]{}, S. R., [Gummersbach]{}, C. A., & [Wolf]{}, B. 2000, , 353, 655
, A. J., [Keller]{}, S. C., [Kaufer]{}, A., [et al.]{} 2002, , 385, 143
, A. J., [Nieva]{}, M. F., [Daflon]{}, S., & [Cunha]{}, K. 2005, , 633, 899
, B. M. 1980, , 237, 765
, Y.-Q., & [Shi]{}, C.-H. 2014, , 442, 741
McKee, C. F., & Cowie, L. L. 1975, , 195, 715
, C. F., [Hollenbach]{}, D. J., [Seab]{}, G. C., & [Tielens]{}, A. G. G. M. 1987, , 318, 674
, D. C., [Sutherland]{}, R. S., [Dopita]{}, M. A., [Kewley]{}, L. J., & [Groves]{}, B. A. 2017, , 466, 4403
, T., & [Bressert]{}, E. 2012, [APLpy: Astronomical Plotting Library in Python]{}, Astrophysics Source Code Library, , , 1208.017
, S. C., & [Dopita]{}, M. A. 1992, , 384, 508
, C. G. 1987, in Astrophysics and Space Science Library, Vol. 134, Interstellar Processes, ed. D. J. [Hollenbach]{} & H. A. [Thronson]{}, Jr., 491–512
, C. G., & [Shull]{}, J. M. 1983, , 275, 652
, O., [Caselli]{}, P., & [Juvela]{}, M. 2017, , 601, A113
, O., [Harju]{}, J., & [Juvela]{}, M. 2011, , 535, A49
, M. G., & [Weedman]{}, D. W. 1971, , 169, 271
, L. 1976, Comments on Astrophysics, 6, 177
, R. S., & [Dopita]{}, M. A. 2017, , 229, 34
, I. R., [Dopita]{}, M. A., [Mathewson]{}, D. S., [Long]{}, K. S., & [Helfand]{}, D. J. 1982, , 261, 473
, S., & [Osterbrock]{}, D. E. 1987, , 63, 295
, F., & [Dopita]{}, M. A. 2011, , 331, 521
, F. P. A., [Seitenzahl]{}, I. R., [Dopita]{}, M. A., & [Ghavamian]{}, P. 2017, , 602, L4
, D. E., [Lauroesch]{}, J. T., [Wong]{}, T., & [York]{}, D. G. 2016, , 821, 118
, B. E., & [Mathewson]{}, D. S. 1966, , 131, 371
, B. J., [Borkowski]{}, K. J., [Reynolds]{}, S. P., [et al.]{} 2006, , 652, L33
, X., & [Chen]{}, Y. 2008, Advances in Space Research, 41, 416
Appendix {#appendix .unnumbered}
========
In the following four Tables, \[tableA1\] through \[tableA4\], we present the measured line fluxes of the 20 shocked ISM clouds in N132D. No reddening correction has been applied, since the mean Balmer Decrement as measured; H$\alpha$/H$\beta$/H$\gamma$/H$\delta$ = 298.7/100.0/44.7/23.9 is indistinguishable (within the errors) with the theoretical Balmer Decrement for $\log n_e =10^4$cm$^{-3}$, and $T_e = 5000$K; 300/100/46.0/25.3. On the basis of the soft X-ray absorption @Borkowski07 determined a best-fit hydrogen column density of $1.4\times10^{20}$ cm$^{-2}$ in the western part of N132D and $1.4-4.1\times10^{21}$ cm$^{-2}$ in the south. Assuming dust is not destroyed in the bar of the LMC, and using the @Fitzpatrick86 $N_{\mathrm H}/E_{\mathrm{B - V}}$ conversion factor, these figures would imply a reddening of $E_{\mathrm{B - V}} = 0.005$mag. and 0.06–0.17mag., respectively.
[^1]: Available at :
[^2]: <http://www.mso.anu.edu.au/pywifes/doku.php.>
[^3]: [QFitsView v3.1]{} is a FITS file viewer using the QT widget library and was developed at the Max Planck Institute for Extraterrestrial Physics by Thomas Ott.
[^4]: Graf is written by R. S. Sutherland and is available at: [<https://miocene.anu.edu.au/graf>]{}
[^5]: Lines is written by R. S. Sutherland and is available at: [<https://miocene.anu.edu.au/lines>]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'For a numerical semigroup $S \subseteq {\mathbb N}$ with embedding dimension $e$, conductor $c$ and left part $L=S \cap [0,c-1]$, set $W(S) = e|L|-c$. In 1978 Wilf asked, in equivalent terms, whether $W(S) \ge 0$ always holds, a question known since as Wilf’s conjecture. Using a closely related lower bound $W_0(S) \le W(S)$, we show that if $|L| \le 12$ then $W_0(S) \ge 0$, thereby settling Wilf’s conjecture in this case. This is best possible, since cases are known where $|L|=13$ and $W_0(S)=-1$. Wilf’s conjecture remains open for $|L| \ge 13$.'
author:
- 'S. Eliahou and D. Marín-Aragón'
title: On numerical semigroups with at most 12 left elements
---
Introduction
============
Let ${\mathbb N}=\{0,1,2,\ldots \}$ denote as usual the set of nonnegative integers. Given integers $a \le b$, we denote by $[a,b[=[a,b-1]=\{z \in {\mathbb Z}\mid a \le z < b\}$, and by $[a,\infty[=\{z \in {\mathbb Z}\mid a \le z\}$, the integer intervals they span. A *numerical semigroup* is a submonoid $S$ of $({\mathbb N},+)$ such that $|{\mathbb N}\setminus S|$ is finite. Equivalently, it is a subset $S$ of ${\mathbb N}$ of the form $S = {\langle a_1,\dots,a_n \rangle}={\mathbb N}a_1+\dots+{\mathbb N}a_n$ where $\gcd(a_1,\dots,a_n)=1$. The least such $n$ is called the *embedding dimension* of $S$ and is often denoted $e$. The *multiplicity* of $S$ is $m = \min S^*$, where $S^*=S \setminus \{0\}$. The *conductor* of $S$ is $c=\max ({\mathbb Z}\setminus S)+1$, or equivalently, the least $c \in {\mathbb N}$ such $[c, \infty[ \, \subseteq S$. The *genus* of $S$ is $g = |{\mathbb N}\setminus S|$. The *left part* of $S$ is $$L=\{s\in S \mid s<c\} = S \cap [0,c[.$$ The *left elements of $S$* are the elements of $L$. Finally, as in [@WilfMacaulay], we denote $$\label{W}
W(S) = e|L|-c.$$ In 1978 Wilf asked, in equivalent terms, whether the inequality $$W(S) \ge 0$$ holds for every numerical semigroup $S$ [@Wilf]. This open question is now known as Wilf’s conjecture. Various particular cases have been settled, including the six independent cases $e \le 3$, $|L| \le 6$, $m \le 18$, $g \le 60$, $c \le 3m$ and $e \ge m/3$. See e.g. [@BGOW; @D; @cota4; @WilfMacaulay; @E2; @FGH; @FH; @K; @Mosca; @Mio; @Sammartano; @Sy]. See also [@D2] for a recent extensive survey on this topic.
The authors of [@cota4] settled Wilf’s conjecture in case $|L|\leq 4$. This was later extended in [@WilfMacaulay], where a certain lower bound $W_0(S) \le W(S)$ was introduced and shown to satisfy $W_0(S) \ge 0$ whenever $|L|\leq 6$. Here we further extend this result by showing that $W_0(S)\geq 0$ holds whenever $|L|\leq 12$. This is best possible since, as shown in [@Jean], there are numerical semigroups $S$ such that $|L| = 13$ and $W_0(S) < 0$. See also Section \[sub W0\].
This paper is organized as follows. In Section \[section notation\] we recall some notation and background, including the definition of $W_0(S)$. In Section \[section profile\] we establish $W_0(S) \ge 0$ in some special circumstances. Our main result, namely $W_0(S) \ge 0$ if $|L| \le 12$, and hence Wilf’s conjecture in that case, is proved in Section \[main\].
For extensive information on numerical semigroups, see [@libro].
Background and notation {#section notation}
=======================
In this section, we recall some notation and terminology introduced in [@WilfMacaulay]. Let $S$ be a numerical semigroup. We denote by $P \subset S^*$ the unique minimal generating set of $S$, so that $S = {\langle P \rangle}$ and $|P|=e$, the embedding dimension. It coincides with the set of *primitive elements* of $S$, i.e. those $x \in S^*$ which are not the sum of two smaller elements of $S^*$. Let $m,c$ be the multiplicity and conductor of $S$, respectively. The *depth* of $S$ is $q = \lceil c/m \rceil$ and its *offset* is $\rho = qm-c$. Thus $$\label{rho}
c=qm-\rho, \ \ \rho \in [0,m[.$$ The set of *decomposable elements* of $S$ is $$D=S^* \setminus P=S^*+S^*.$$ Note that $D$ contains $[c+m,\infty[$. Indeed, if $z \ge c+m$, then $z=m+(z-m)$, so that $z \in S^*+S^*$ since both $m \in S^*$ and $z-m \in S^*$ as $z-m \ge c$. It follows that $$\label{P confined}
P \subseteq [m,c+m[.$$
Throughout Section \[section notation\], the symbols $m,c,q$ and $\rho$ will denote, often tacitly so, the multiplicity, conductor, depth and offset of the numerical semigroup $S$ under consideration, respectively.
The level function $\lambda$
----------------------------
Let $S \subseteq {\mathbb N}$ be a numerical semigroup. We shall further use the following notation, as in [@WilfMacaulay].
\[nota\] For all $j \in {\mathbb Z}$, we denote $$\begin{aligned}
I_j & = & [jm-\rho,(j+1)m-\rho[, \\
S_j & = & S\cap I_j,\,\, P_j=P\cap I_j,\,\, D_j=D\cap I_j.\end{aligned}$$
In particular, we have $$I_q=[qm-\rho, (q+1)m-\rho[ \ = [c,c+m[.$$
The following set addition rules are shown in [@WilfMacaulay]. The proof is straightforward and left to the reader.
\[addition\] Let $S$ be a numerical semigroup. For all $i,j \ge 1$, we have $$\begin{aligned}
S_1 + S_j & \subseteq & S_{1+j} \cup S_{2+j}, \\
S_i + S_j & \subseteq & S_{i+j-1} \cup S_{i+j} \cup S_{i+j+1} \ \textrm{ if } i,j \ge 2. \hspace{1cm}\Box\end{aligned}$$
The *level function* $\lambda_S \colon {\mathbb N}\to {\mathbb N}$ associated to $S$ is defined by $$\lambda_S(x) = j \iff x \in I_j \iff jm-\rho \le x \le (j+1)m - \rho -1$$ for all $x \in {\mathbb N}$. In particular, if $x \in S$ then $\lambda_S(x) = j$ if and only if $x \in S_j$.
In the sequel, for simplicity, we shall write $\lambda$ for $\lambda_S$. Using this function, the above proposition translates as follows.
\[cor level\] Let $x,y \in S^*$. If $\lambda(x), \lambda(y) \ge 2$ then $$\label{level}
\lambda(x)+\lambda(y)-1 \le \lambda(x+y) \le \lambda(x)+\lambda(y)+1.$$ If $\lambda(x) =1$ or $\lambda(y) =1$, then $\lambda(x)+\lambda(y) \le \lambda(x+y) \le \lambda(x)+\lambda(y)+1$. $\Box$
Here are some more consequences.
Let $a,x,y \in S^*$. Then $\lambda(x+y) > \max\{\lambda(x),\lambda(y)\}$. If $\lambda(a+x)=\lambda(a+y)$, then $|\lambda(y)-\lambda(x)| \le 1$.
The first statement directly follows from Corollary \[cor level\]. As for the second one, let $i=\lambda(x), j=\lambda(y)$. We may assume $i \le j$. Let $k=\lambda(a+x)$. Then $a+x, a+y \in S_k$. Hence $|y-x|=|(a+y)-(a+x)| \le m-1$. It follows that $j \le i+1$, since if $j \ge i+2$ then $\min S_j - \max S_i \ge m+1$.
The number $W_0(S)$ {#sub W0}
-------------------
For a numerical semigroup $S$, we denote $$\label{W0}
W_0(S) = |P\cap L||L|-q|D_q|+\rho.$$
Introduced in [@WilfMacaulay], this number bounds $W(S)$ from below and is sometimes easier to evaluate. See also [@D], where $W_0(S)$ is denoted $E(S)$. The following result is Proposition 3.11 in [@WilfMacaulay]. For convenience, we recall the short proof.
Let $S$ be a numerical semigroup. Then $W(S)\geq W_0(S)$.
We have $W(S)=|P||L|-c=|P||L|-qm+\rho$. We have $m=|P_q|+|D_q|$, since $m=|S_q|=|P_q \sqcup D_q|$, and $|P|=|P \cap L|+|P_q|$. It follows that $$W(S)=W_0(S)+|P_q|(|L|-q).$$ Now $|L| \ge q$, since $L$ contains the $q$-subset $\{0,1,\dots,q-1\}m$.
\[Wilf with W0\] Let $S$ be a numerical semigroup such that $W_0(S) \ge 0$. Then $S$ satisfies Wilf’s conjecture.
We have $W(S) \ge W_0(S) \ge 0$.
This corollary is the basis of our approach in this paper, whose main result is that $W_0(S) \ge 0$ whenever $|L| \le 12$. Note that in contrast to Wilf’s conjecture, the number $W_0(S)$ can be negative, but such cases are extremely rare. For instance, among the more than $10^{13}$ numerical semigroups of genus $g \le 60$, only five of them satisfy $W_0(S) < 0$. See [@D; @Jean; @FH]. More specifically, these five exceptions all satisfy $W_0(S)=-1$, $|L|=13$ and $c=4m$, and they occur at genus $43, 51, 55, 55$ and $59$, respectively. The first one, of genus $g=43$, is $S={\langle 14,22,23 \rangle} \cup [56, \infty[$.
The following result has been established in [@WilfMacaulay].
\[q le 3\] Let $S$ be a numerical semigroup of depth $q \le 3$. Then $W_0(S) \ge 0$. In particular, $S$ satisfies Wilf’s conjecture.
Consequently, in proving here that $W_0(S) \ge 0$ if $|L| \le 12$, we only need to consider the case of depth $q \ge 4$. The next three sections focus on the Apéry set of $S$ with respect to $m$ and provide tools to evaluate $W_0(S)$ and prove our main result.
The Apéry profile of $S$ {#sub apery}
------------------------
Let $S \subseteq {\mathbb N}$ be a numerical semigroup. We denote by $$A ={{\rm Ap}}(S,m)=\{x\in S \mid x-m\not\in S\}$$ the *Apéry set* of $S$ with respect to $m$. Equivalently, $A = S \setminus (m+S)$. Each Apéry element $x \in A$ is the smallest element in $S$ of its class mod $m$, since $x-m \notin S$. Hence $|A|=m$. We have $$A \subseteq [0, c+m[.$$ Indeed, this follows from the inclusion $[c+m,\infty[ \: = m + [c,\infty[ \: \subseteq m+S$. We now introduce the Apéry profile of $S$.
\[Ai\] Let $S$ be a numerical semigroup of depth $q \ge 1$. For all $0\leq i\leq q$, we set $A_i = A\cap I_i$ and $${\alpha}_i = |A_i|.$$
We have $A_0=\{0\}$, so ${\alpha}_0=1$. Moreover, $P_1=\{m\} \sqcup A_1$ and $P_i \subseteq A_i$ for all $i \ge 2$.
We call *Apéry profile* of $S$ the $(q-1)$-tuple $${\alpha}(S) = ({\alpha}_1,\dots,{\alpha}_{q-1}) \in {\mathbb N}^{q-1}.$$
As noted above, we have $$\label{alpha i}
{\alpha}_1=|P_1|-1 \textrm{ and }\, {\alpha}_i\geq |P_i| \, \textrm{ for all } 2\leq i< q.$$ Moreover, since $A \subseteq [0, c+m[ = I_0 \sqcup I_1 \sqcup \dots \sqcup I_q$, we have $$\label{part A}
A = A_0 \sqcup A_1 \sqcup \dots \sqcup A_q.$$ Therefore $$\label{eq sum a_i}
\sum_{i=0}^q {\alpha}_i = |A| = m.$$ This justifies why ${\alpha}_0,{\alpha}_q$ are not included in the profile ${\alpha}(S)$, as ${\alpha}_0=1$ and ${\alpha}_q$ may be recovered from ${\alpha}(S)$ and $m$ by the above formula.
Primitive and decomposable Apéry elements
-----------------------------------------
Let $S$ be a numerical semigroup of multiplicity $m$ and Apéry set $A=S \setminus (m+S)$. A key point in the sequel is to distinguish, in $A^* = A \setminus \{0\}$, the primitive elements from the decomposable ones. Indeed, the partition $$A^* = (A \cap P) \sqcup (A \cap D)$$ plays an important role and motivates the following notation.
\[prim and dec\] Let $S$ be a numerical semigroup of depth $q \ge 1$. For all $1 \le i \le q$, we set $${\alpha}_i' = |A_i \cap P|, \ \ {\alpha}_i'' = |A_i \cap D|.$$ Thus $\alpha_i = \alpha_i'+\alpha_i''$ for all $i$.
Since $P \subseteq [m,c+m[$ as seen above, and since $[m,c+m[ \ \subseteq I_1 \cup \dots \cup I_q$, we have $$\label{card of P}
|P| = 1+{\alpha}'_1 + \dots + {\alpha}'_{q}.$$ In particular, $$\begin{aligned}
{\alpha}'_{q} & = & |P \setminus L| \, =\, \big|P \cap [c,c+m[\big|, \\
|D_q| & = & |D \cap [c,c+m[| \, =\, \big|[c,c+m[ \setminus P\big |, \\
m & = & |P_q| +|D_q|.\end{aligned}$$
The following properties of the Apéry set $A$ will be widely used below, often tacitly so.
\[downset\] Let $z \in A \cap D$. If $z =x+y$ with $x,y \in S^*$, then $x,y \in A^*$.
If $x \notin A$, then $x=m+s$ for some $s \in S$, implying $z=x+y=m+(s+y)$. Since $s+y \in S$, it follows that $z\notin A$, contrary to the hypothesis.
For all $k \ge 2$, we have $\displaystyle
A_k \cap D \subseteq \cup_{i,j} (A_i + A_j)
$ where $1 \le i \le j$ and $k-1 \le i+j \le k+1$.
Directly follows from Proposition \[addition\] and Lemma \[downset\].
If $A_k \not= \emptyset$ and $A_i = \emptyset$ for all $1 \le i \le k-1$ for some $k \ge 2$, then $A_k = P_k$.
Directly follows from the above proposition.
Compressed Apéry elements
-------------------------
Throughout, let $S$ denote a numerical semigroup with multiplicity $m$, conductor $c$ and Apéry set $A=S \setminus (m+S)$.
Let $s \in S^*$. We say that $s$ is *compressed* if there exist $x,y \in S^*$ such that $s=x+y$ and $\lambda(s) < \lambda(x)+\lambda(y)$.
By Corollary \[cor level\], we have $\lambda(x+y) \ge \lambda(x)+\lambda(y)-1$ for all $x,y \in S^*$. Thus, the inequality $\lambda(x+y) < \lambda(x)+\lambda(y)$ is equivalent to $\lambda(x+y) = \lambda(x)+\lambda(y)-1$. Estimating the number of compressed elements in $A$ is important in the sequel. This motivates the following notation.
\[C\] $C = C(S)=\{z \in A\cap D \mid z \textrm{ is compressed}\}$.
Recall from Lemma \[downset\] that if $z \in A\cap D$ and $z=x+y$ with $x,y \in S^*$, then in fact $x,y \in A^*$. Consequently, for all $i,j \ge 2$, we have $$\label{on C}
(S_i+S_j)\cap A_{i+j-1} = (A_i+A_j)\cap A_{i+j-1} \ \subseteq \ C.$$ More generally, even if the description below will not be needed here, we have $$C = \bigcup_{k=3}^q A_k \cap (\cup_{i=1}^{k-1} (A_i+A_{k+1-i})).$$ The next result provides a key lower bound on $\rho$, where $\rho$ is the offset as defined in . See also Proposition 3.20 in [@E2].
\[deficit\] Let $S$ be a numerical semigroup. Then $\rho \ge |C|$.
Let $z \in C$, and assume $z = x+y$ with $x,y \in A^*$ such that $\lambda(z)=\lambda(x)+\lambda(y)-1$. Say $\lambda(x)=i$, $\lambda(y)=j$ and $\lambda(z)=i+j-1$. By the definition of $S_i$, we have $$(S_i+S_j) \cap S_{i+j-1} \subseteq [(i+j)m-2\rho,(i+j)m-\rho[.$$ Thus $z \in [(i+j)m-2\rho,(i+j)m-\rho[$. Now, the only classes mod $m$ occurring in the latter interval are those in $[-2\rho,-\rho[$, a set of cardinality $\rho$. Since there is only one element in $A$ per class mod $m$, and since $C \subset A$, the statement follows.
In particular, we shall invoke the following simplified version.
\[cor rho\] For all $i,j \ge 2$, we have $\rho \ge |(A_i+A_j)\cap A_{i+j-1}|$.
Follows from and Proposition \[deficit\].
Computing $W_0(S)$
------------------
The following formulas allow to evaluate $W_0(S)$ using the Apéry profile of $S$ as defined in Definition \[Ai\] and the decomposition ${\alpha}_q={\alpha}_q'+{\alpha}_q''$ given by Notation \[prim and dec\]. Recall that both $|L|$ and $|D_q|$ are involved in the expression of $W_0(S)$.
\[prop formulas\] Let $S \subseteq {\mathbb N}$ be a numerical semigroup with Apéry profile ${\alpha}(S)=({\alpha}_1,\dots,{\alpha}_{q-1})$. Then $$\begin{aligned}
|L| & = & q +(q-1){\alpha}_1+ \dots + {\alpha}_{q-1}, \\
|D_q| & = & {\alpha}_0+{\alpha}_1+\dots+\alpha_{q-1}+\alpha_q''.\end{aligned}$$
Let $s \in L$ be minimal in its class mod $m$. Then $s \in L \cap A$. Let $i \ge 0$ be the unique integer such that $s \in A_i$. Then $0 \le i \le q-1$ since $s \in L$. Let $z \in L$ be such that $z \equiv s \bmod m$. Then $z = s +jm$ with $0 \le j \le q-1-i$. It follows that $$|L \cap (s+m{\mathbb N})| = q-i.$$ Letting now $s$ run through all elements of $L$ which are minimal in their respective classes mod $m$, the above discussion implies $$|L|=q|A_0|+(q-1)|A_1|+ \dots +|A_{q-1}|,$$ yielding the first formula. Since $m=|P_q|+|D_q|={\alpha}_q'+|D_q|$, it follows that $$|D_q|=m-{\alpha}_q'=m-{\alpha}_q+{\alpha}_q''.$$ The second formula now follows from , i.e. $m=\sum_{i=0}^q {\alpha}_i$.
Notation {#subsec notation}
--------
We shall use the following notation throughout the remainder of the paper. Given a numerical semigroup $S$, we denote by $m$ its multiplicity, by $c$ its conductor, by $q$ its depth, by $\rho$ its offset, by $L$ its left part, by $P$ its set of primitive elements, by $D$ its set of decomposable elements, by $A = S \setminus (m+S)$ its Apéry set with respect to $m$, and by $${\alpha}(S) = ({\alpha}_1,\dots,{\alpha}_{q-1})$$ its Apéry profile, where ${\alpha}_i=|A_i|$ for all $0 \le i \le q$. For $i \ge 1$, we have ${\alpha}_i={\alpha}_i'+{\alpha}_i''$ where ${\alpha}_i'=|A_i \cap P|$ and ${\alpha}_i''=|A_i \cap D|$. We shall constantly use the formulas below to compute the ingredients involved in $$W_0(S) = |P \cap L||L|-q|D_q|+\rho,$$ namely $$|P \cap L| = 1 + \sum_{i=1}^{q-1} {\alpha}_i', \ \
|L| = \sum_{i=0}^{q-1} (q-i){\alpha}_i , \ \
|D_q| = \sum_{i=0}^{q-1} {\alpha}_i + {\alpha}_q''.$$
Finally, as in the preceding section, we denote by $C$ the set of compressed Apéry elements, i.e $$C = \{z \in A\cap D \mid \exists x,y \in A^*, \, z=x+y,\, \lambda(z) < \lambda(x)+\lambda(y)\}.$$ In a few cases, the estimate $\rho \ge |C|$ provided by Proposition \[deficit\] will be crucially needed in order to be able to conclude $W_0(S) \ge 0$.
An occurrence of $W_0(S) \ge 0$ {#section profile}
===============================
In this section, we establish $W_0(S) \ge 0$ for numerical semigroups $S$ under suitable assumptions on ${\alpha}(S)$ but not on $|L|$. We use the notation of Section \[subsec notation\] throughout.
\[thm:h\] Let $S$ be a numerical semigroup of depth $q \ge 4$. Let $h = \lceil q/2 \rceil$. Assume that $\alpha_i=0$ for all $1 \le i \le h-1$. Then $W_0(S) \ge 0$.
Since $P = \cup_{1 \le i \le q}$ and $P_1=\{m\} \sqcup A_1$, we have $|P|=1+{\alpha}_1+{\alpha}'_2+\dots+{\alpha}'_{q}$ and $$\label{P cap L}
|P \cap L| = 1+{\alpha}_1+{\alpha}'_2+\dots+{\alpha}'_{q-1}.$$
**[[$\bullet$ ]{}]{}Assume $q$ odd.** Then $h = (q+1)/2$. Since $A_1=\dots=A_{h-1}=\emptyset$, we have $$A \cap D \subseteq \bigcup_{i,j=h}^{q-1} (A_i+A_j).$$ By Proposition \[addition\], we have $
A_i+A_j \subseteq S_{i+j-1} \sqcup S_{i+j} \sqcup S_{i+j+1}.
$ Since $A_t = \emptyset$ for $t \ge q+1$, and since $2h=q-1$, it follows from the above that $$A \cap D \subseteq 2A_h.$$ Hence ${\alpha}_q'' \le |2A_h|$. It also follows that $A_i=P_i$ for all $h \le i \le q-1$. Hence $$\begin{aligned}
|P \cap L| & = & 1+{\alpha}_h+\dots+{\alpha}_{q-1} \\
& \ge & 1+{\alpha}_h, \\
|L| & = & q + (h-1){\alpha}_h+(h-2){\alpha}_{h+1}+\dots+{\alpha}_{q-1} \\
& \ge & q + (h-1){\alpha}_h, \\
|D_q| & = & 1+{\alpha}_h+\dots+{\alpha}_{q-1}+{\alpha}_q'' \\
& = & |P \cap L|+{\alpha}_q''.\end{aligned}$$ Thus, $$\begin{aligned}
W_0(S) & = & |P \cap L||L|-q|D_q|+\rho \\
& = & |P \cap L||L|-q(|P \cap L|+{\alpha}_q'')+\rho \\
& = & |P \cap L|(|L|-q)-q{\alpha}_q''+\rho \\
& = & (1+{\alpha}_h)(h-1){\alpha}_h-q{\alpha}_q''+\rho.\end{aligned}$$
Since $A \cap D_q \subseteq 2A_h$, and since $q=2h-1$, it follows from Lemma \[deficit\] that $A_q \cap D \subseteq C$, where $C$ is the subset defined in that Lemma. Consequently, we have $|A_q \cap D| \le |C| \le \rho$, whence $${\alpha}_q'' \le \min(|2A_h|, \rho).$$ Therefore $$W_0(S) \ge (1+{\alpha}_h)(h-1){\alpha}_h-q\min(|2A_h|, \rho)+\rho.$$ The following bound will take care of the last two summands.
**Claim.** For all $t \ge 0$, we have $$\label{easy}
-q\min(t,\rho)+\rho \ \ge \ -(q-1)t.$$ Indeed, if $\rho \le t$, then $-q\min(t,\rho)+\rho = -(q-1)\rho \ge -(q-1)t$. And if $\rho > t$, then $-q\min(t,\rho)+\rho = -qt+\rho > -(q-1)t$. This proves the claim.
Moreover, as a very crude estimate, we have $$|2A_h| \le {\alpha}_h({\alpha}_h+1)/2.$$ Hence, using and the above, we get $$W_0(S) \ge (1+{\alpha}_h)(h-1){\alpha}_h-(q-1){\alpha}_h({\alpha}_h+1)/2.$$ Since $(q-1)/2=h-1$, it follows that $W_0(S) \ge 0$, as desired.
**[[$\bullet$ ]{}]{}Assume $q$ even.** Then $h = q/2$. Since $A_1=\dots=A_{h-1}=\emptyset$, we have $$A \cap D \subseteq \bigcup_{i,j=h}^{q-1} (A_i+A_j).$$ By Proposition \[addition\], we have $
A_i+A_j \subseteq S_{i+j-1} \sqcup S_{i+j} \sqcup S_{i+j+1}.
$ Since $A_t = \emptyset$ for $t \ge q+1$, and since $2h=q$, it follows from the above that $$A \cap D \subseteq (2A_h) \cup (A_h+A_{h+1}).$$ Moreover, we have $$2A_h \cap A \subseteq A_{q-1} \cup A_q, \quad (A_h+A_{h+1}) \cap A \subseteq A_{q}.$$ Hence $$A \cap D \subseteq (2A_h \cap A_{q-1}) \cup (2A_h \cap A_{q}) \cup (A_h+A_{h+1}) \cap A_{q}.$$ We have $$\begin{aligned}
{\alpha}_{q-1}'' & = & |2A_h \cap A_{q-1}|, \\
{\alpha}_q'' & = & |(2A_h \cap A_{q}) \cup (A_h+A_{h+1}) \cap A_{q}|.\end{aligned}$$ Hence $$\label{q-1 et q}
{\alpha}_{q-1}''+{\alpha}_q'' \le |2A_h|+|(A_h+A_{h+1}) \cap A_{q}|.$$ Moreover, since $q=2h$, Corollary \[cor rho\] yields $$|(A_h+A_{h+1}) \cap A_{q}| \le \rho.$$ Moreover, we have $|(A_h+A_{h+1}) \cap A_{q}| \le {\alpha}_h{\alpha}_{h+1}$. Hence $$\label{h et h+1}
|(A_h+A_{h+1}) \cap A_{q}| \le \min(\rho, {\alpha}_h{\alpha}_{h+1}).$$ Combining and , we get $$\label{combin}
{\alpha}_{q-1}''+{\alpha}_q'' \le |2A_h|+\min(\rho, {\alpha}_h{\alpha}_{h+1}).$$ It also follows that $A_i=P_i$ for all $h \le i \le q-2$. Hence $$\begin{aligned}
|P \cap L| & = & 1+{\alpha}_h+\dots+{\alpha}_{q-2}+{\alpha}_{q-1}' \\
& \ge & 1+{\alpha}_h+{\alpha}_{h+1}, \\
|L| & = & q + h{\alpha}_h+(h-1){\alpha}_{h+1}+\dots+{\alpha}_{q-1} \\
& \ge & q + h{\alpha}_h+(h-1){\alpha}_{h+1}, \\
|D_q| & = & 1+{\alpha}_h+\dots+{\alpha}_{q-1}+{\alpha}_q'' \\
& = & |P \cap L|+{\alpha}_{q-1}''+{\alpha}_q''.\end{aligned}$$ Thus, $$\begin{aligned}
W_0(S) & = & |P \cap L||L|-q|D_q|+\rho \\
& = & |P \cap L||L|-q(|P \cap L|+{\alpha}_{q-1}''+{\alpha}_q'')+\rho \\
& = & |P \cap L|(|L|-q)-q({\alpha}_{q-1}''+{\alpha}_q'')+\rho \\
& \ge & (1+{\alpha}_h+{\alpha}_{h+1})(h{\alpha}_h+(h-1){\alpha}_{h+1}-q({\alpha}_{q-1}''+{\alpha}_q'')+\rho.\end{aligned}$$ Therefore, using , we get $$W_0(S) \ge (1+{\alpha}_h+{\alpha}_{h+1})(h{\alpha}_h+(h-1){\alpha}_{h+1})-q(|2A_h|+\min(\rho, {\alpha}_h{\alpha}_{h+1}))+\rho.$$ Using again for the last two summands, we have $$-q\min(\rho, {\alpha}_h{\alpha}_{h+1})+\rho \ge -(q-1){\alpha}_h{\alpha}_{h+1.}$$ We also have the very crude estimate $$|2A_h| \le {\alpha}_h({\alpha}_h+1)/2.$$ Hence $$W_0(S) \ge (1+{\alpha}_h+{\alpha}_{h+1})(h{\alpha}_h+(h-1){\alpha}_{h+1})-q{\alpha}_h({\alpha}_h+1)/2-(q-1){\alpha}_h{\alpha}_{h+1}.$$ Using $q=2h$, it follows that $$W_0(S) = (h-1){\alpha}_{h+1}({\alpha}_{h+1}+1).$$ Hence $W_0(S) \ge 0$, as desired.
We conclude this section with an easy particular case.
Let $S$ be a numerical semigroup of depth $q \ge 4$ such that $|P \cap L| \ge \max({\alpha}''_q,q)$. Then $W_0(S) \ge \rho$.
$W_0(S) = |P \cap L||L|-q|D_q|+\rho$. Recall that $|L|=\sum_{i=0}^{q-1} (q-i){\alpha}_i$ and that $|D_q|=\sum_{i=0}^{q-1}{\alpha}_i + {\alpha}''_q$. Hence $$W_0(S) = q(|P \cap L|-{\alpha}''_q)+\sum_{i=1}^{q-1}((q-i)|P \cap L|-q){\alpha}_i + \rho.$$ Since $|P \cap L| \ge {\alpha}_q''$ and $|P \cap L| \ge q$ by hypothesis, the claimed inequality follows.
Main result {#main}
===========
Let $S$ be a numerical semigroup. We use the notation of Section \[subsec notation\] throughout. Wilf’s conjecture has been successively settled for $|L| \le 4$ and $|L| \le 6$ in [@cota4] and [@WilfMacaulay], respectively. Here we extend these results to the case $|L| \le 12$. Even more so, we show that if $|L| \le 12$ then $W_0(S) \ge 0$. As mentioned earlier, this is best possible, since there are numerical semigroups $S$ satisfying $|L|=13$ and $W_0(S) <0$. At the time of writing, it remains an open problem to determine whether all numerical semigroups $S$ with $|L|=13$ satisfy Wilf’s conjecture. In this section we prove the following result.
\[main thm\] Let $S$ be a numerical semigroup such that $|L|\leq 12$. Then $W_0(S)\geq 0$. In particular, $S$ satisfies Wilf’s conjecture.
By Theorem \[q le 3\], the bound $W_0(S) \ge 0$ holds for all numerical semigroups of depth $q \le 3$. Consequently, in the sequel, we shall freely assume $q \ge 4$, since it suffices to prove Theorem \[main thm\] in that case. In fact, it also suffices to consider the case $q \le 7$, as follows from the following proposition.
\[prop q ge 8\] Let $S$ be a numerical semigroup of depth $q \ge 8$ such that $|L| \le 12$. Then $W_0(S) \ge 0$.
Let $h = \lceil q/2 \rceil$. Then $
12 \ge |L| \ge q+(q-1){\alpha}_1+\dots+(q-h+1){\alpha}_{h-1}.
$ This implies ${\alpha}_i=0$ for all $1 \le i \le h-1$. For if not, then $|L| \ge q+(q-h+1) $, and since $h \le (q+1)/2$, we would get $12 \ge |L| \ge 2q-(q+1)/2+1=(3q+1)/2$ and hence $3q+1 \le 24$, contrary to the hypothesis $q \ge 8$. It now follows from Theorem \[thm:h\] that $W_0(S) \ge 0$.
Finally, the following result strongly restricts the values of ${\alpha}_1$ to consider.
Let $S$ be a numerical semigroup of depth $q \ge 4$ such that $|L| \le 12$. Then ${\alpha}_1 \le 2$.
By Proposition \[prop formulas\], we have $|L| = q+(q-1){\alpha}_1+(q-2){\alpha}_2+\cdots+{\alpha}_{q-1}$. Hence $|L| \ge q + (q-1){\alpha}_1$. We have $q \ge 4$. If ${\alpha}_1 \ge 3$ then $|L| \ge 4+9=13$, contrary to the hypothesis on $|L|$.
The cases ${\alpha}_1=2$, $1$ and $0$ will now be treated successively. We shall occasionally use the following notation.
For all $i \ge 1$, we denote by ${\mathbf{1}}_i={\mathbf{1}}_{A_i}$ the indicator function of $A_i$.
When ${\alpha}_1=2$ {#a_1=2}
-------------------
Let $S$ be a numerical semigroup of depth $q \ge 4$ such that $|L| \le 12$. If ${\alpha}_1=2$, then $q=4$ and $W_0(S) \ge 0$.
Since ${\alpha}_1 = 2$, we have $12 \ge |L| \ge q + 2(q-1)=3q-2$. It follows that $q < 5$, whence $q=4$ since $q \ge 4$ by hypothesis. We have $$|P \cap L| \ge 3, \ \ |L| = 10+ 2{\alpha}_2+ {\alpha}_3, \ \ |D_4| = 3+{\alpha}_2+{\alpha}_3+{\alpha}''_4.$$ Hence $2{\alpha}_2+ {\alpha}_3 \le 2$ and so ${\alpha}_2 \le 1$.
[[$\bullet$ ]{}]{}If ${\alpha}_2=0$, then $|L|=10+{\alpha}_3$ and so ${\alpha}_3 \le 2$. Since $2A_1 \subset S_2 \cup S_3$, and since $A_2=\emptyset$ and $2S_3 \cap S_4 = \emptyset$, it follows that $$A_4 \cap D \subseteq (A_1+A_3).$$ Hence ${\alpha}_4'' \le 2 {\alpha}_3$. Therefore $$\begin{aligned}
W_0(S) & \ge & 3(10+{\alpha}_3)-4(3+{\alpha}_3+2{\alpha}_3)+\rho \\
& = & 18-9{\alpha}_3+\rho \\
& \ge & \rho. \end{aligned}$$
[[$\bullet$ ]{}]{}If ${\alpha}_2=1$, then $|L|=12+{\alpha}_3$ and so $|L|=12$ and ${\alpha}_3 =0$. Since $2A_1 \subset S_2 \cup S_3$ and $A_3=\emptyset$, it follows that $$A_4 \cap D \subseteq (A_1+A_2) \cup (2A_2).$$ Hence ${\alpha}''_4 \le 2+1=3$. Therefore $$\begin{aligned}
W_0(S) & \ge & 3\cdot 12-4(3+1+3)+\rho \\
& = & 8 + \rho. \qedhere\end{aligned}$$
A better lower bound on $W_0(S)$ may be obtained by splitting ${\alpha}_i$ as ${\alpha}_i'+{\alpha}_i''$ for $i=2, 3$ in the above proof. For instance, we have only used $|P \cap L| \ge 3$. But we could have used $|P \cap L| \ge 4$ if either ${\alpha}'_2$ or ${\alpha}'_3$ were assumed positive, while if ${\alpha}'_2={\alpha}'_3=0$, a sharper estimate on ${\alpha}''_4$ can been derived.
When ${\alpha}_1=1$ {#case a1=1}
-------------------
Since $12 \ge |L| \ge q +(q-1)=2q-1$, it follows that $q \le 6$. We shall successively treat the cases $q=4$, $5$ and $6$. Throughout Section \[case a1=1\], we set $$A_1=\{x\}.$$
### Case $q=4$
Then ${\alpha}(S)=(1,{\alpha}_2,{\alpha}_3)$. We have $|L|=4+3+2{\alpha}_2+{\alpha}_3$, whence $2{\alpha}_2+{\alpha}_3 \le 5$, implying ${\alpha}_2 \le 2$. We successively examine the cases ${\alpha}_2=2,1,0$. To start with, we have $$A_4 \cap D \subseteq (A_1+A_2) \cup (A_1+A_3) \cup (A_2+A_2) \cup (A_2+A_3).$$
**** Then ${\alpha}(S)=(1,2,{\alpha}_3)$. We have $|L|=11+{\alpha}_3$, whence ${\alpha}_3 \le 1$. Denote $$A_2=\{y_1, y_2\}.$$ Since $A_2 \cap D \subseteq 2A_1$, and since $|2A_1|=1$, we have ${\alpha}_2'' \le 1$ whence ${\alpha}_2' \in \{1,2\}$.
[[$\bullet$ ]{}]{}Assume first ${\alpha}_2'=1$. Say $y_1 \in P, y_2 \in D$. Then $y_2=2x$, and $|P \cap L|=3+{\alpha}_3'$. We have $|D_4|=4+{\alpha}_3+{\alpha}_4''$, and $$(A_1+A_2) \cup (2A_2) = \{4x, 3x, 2x+y_1,x+y_1, 2y_1\}.$$ Since $4x \notin S_3$ and since ${\alpha}_3 \le 1$, there are at most four possibilities for $A_3 \cap D$, listed below together with their consequences on $A_4 \cap D$. Note that Lemma \[downset\] plays a key role to deduce these consequences. For example, if $4x \in A_4$ or $2x+y_1 \in A_4$, then necessarily $3x \in A_3$ or $x+y_1 \in A_3$, respectively. Note also that the level function $\lambda=\lambda_S$ is nondecreasing. Consequently, in the last case $A_3=\{2y_1\}$ below, it follows that $x+y_1 \notin A_4$, for $\lambda(x+y_1) \le \lambda(2y_1)=3$ since $x < y_1$. Here then are the possibilities for $A_3 \cap D$:
1. If $A_3 \cap D = \emptyset$ then $A_4 \cap D \subseteq \{3x,x+y_1, 2y_1\}$.
2. If $A_3 \cap D = \{3x\}$ then $A_4 \cap D \subseteq \{4x,x+y_1, 2y_1\}$.
3. If $A_3 \cap D = \{x+y_1\}$ then $A_4 \cap D \subseteq \{3x,2x+y_1, 2y_1\}$.
4. If $A_3 \cap D = \{2y_1\}$ then $A_4 \cap D \subseteq \{3x, 2y_1\}$.
In either case, we have ${\alpha}_4'' \le 3$. Recall also that $\alpha_3 \le 1$ here. Hence $$\begin{aligned}
W_0(S) & = & (3+\alpha_3')(11+\alpha_3)-4(4+{\alpha}_3+{\alpha}_4'')+\rho\\
& \ge & (3+\alpha_3')(11+\alpha_3)-4(7+{\alpha}_3)+\rho\\
& \ge & 4 +11\alpha_3'+\alpha_3'\alpha_3+\rho.\end{aligned}$$
[[$\bullet$ ]{}]{}Assume now ${\alpha}_2'=2$, so that $y_1,y_2 \in P$. Then $|P \cap L|=4+{\alpha}_3'$. We have $|D_4|=4+{\alpha}_3+{\alpha}_4''$, and $$(2A_1) \cup (A_1+A_2) \cup (2A_2) = \{2x, x+y_1, x+y_2, 2y_1, y_1+y_2, 2y_2\}.$$ Up to permutation of $y_1, y_2$, and using ${\alpha}_3 \le 1$ and $2x \notin S_4$, here are the possibilities for $A_3 \cap D$, together with their consequences for $A_4 \cap D$:
1. If $A_3 \cap D = \emptyset$ then $A_4 \cap D \subseteq \{x+y_1, x+y_2, 2y_1, y_1+y_2, 2y_2\}$.
2. If $A_3 \cap D = \{2x\}$ then $A_4 \cap D \subseteq \{3x, x+y_1, x+y_2, 2y_1, y_1+y_2, 2y_2\}$.
3. If $A_3 \cap D = \{x+y_1\}$ then $A_4 \cap D \subseteq \{x+y_2, 2y_1, y_1+y_2, 2y_2\}$.
4. If $A_3 \cap D = \{2y_1\}$ then $A_4 \cap D \subseteq \{x+y_1, x+y_2, y_1+y_2, 2y_2\}$.
5. If $A_3 \cap D = \{y_1+y_2\}$ then $A_4 \cap D \subseteq \{x+y_1, x+y_2, 2y_1, 2y_2\}$.
In either case, we have ${\alpha}_4'' \le 6$. Hence $$\begin{aligned}
W_0(S) & = & (4+\alpha_3')(11+\alpha_3)-4(4+\alpha_3+\alpha_4'')+\rho \\
& \ge & (4+\alpha_3')(11+\alpha_3)-4(10+\alpha_3)+\rho \\
& = & 4+11\alpha_3'+\alpha_3'\alpha_3+\rho.\end{aligned}$$
**** Then ${\alpha}(S)=(1,1,{\alpha}_3)$. We have $|L|=9+{\alpha}_3$, whence ${\alpha}_3 \le 3$. We also have $|P \cap L| =2+{\alpha}_2'+{\alpha}_3'$ and $|D_4|=3+\alpha_3+\alpha_4''$. Denote $$A_2=\{y\}.$$
[[$\bullet$ ]{}]{}Assume first $y \in A_2 \cap D$. Then $y = 2x$, $|P \cap L| =2+{\alpha}_3'$ and $A_3 \cap D \subseteq \{3x\}$. Thus ${\alpha}_3'' \le 1$, and either $$A_4 \cap D \subseteq \{3x\} \cup (x+P_3) \,\,\, \textrm{ or }\,\,\, A_4 \cap D \subseteq \{4x\} \cup (2x+P_3) \cup (x+P_3).$$ We claim that ${\alpha}_4'' \le 1 + {\alpha}_3'$ in both cases. This is clear in the first one. In the second one, for all $z \in P_3$ we have $$|A_4 \cap \{2x+z, x+z\}| \le 1.$$ Therefore $|A_4 \cap ((2x+P_3) \cup (x+P_3))| \le |P_3|,$ implying ${\alpha}_4'' \le 1 + {\alpha}_3'$ here as well. Using ${\alpha}_3 ={\alpha}_3'+{\alpha}_3'' \le {\alpha}_3'+1$, we have $$\begin{aligned}
W_0(S) & = & (2+{\alpha}_3')(9+{\alpha}_3)-4(3+\alpha_3+\alpha_4'')+\rho\\
& \ge & (2+{\alpha}_3')(9+{\alpha}_3)-4(4+{\alpha}_3'+{\alpha}_3)+\rho\\
& = & 2+5{\alpha}_3'-2{\alpha}_3+{\alpha}_3{\alpha}_3'+\rho\\
& \ge & 2+5{\alpha}_3'-2({\alpha}_3'+1)+{\alpha}_3{\alpha}_3'+\rho\\
& \ge & 3{\alpha}_3'+{\alpha}_3{\alpha}_3'+\rho.\end{aligned}$$
[[$\bullet$ ]{}]{}Assume now $y \in A_2 \cap P$. Hence $|P \cap L| =3+{\alpha}_3'$. Here we have $$A_3 \cap D \subseteq \{2x, x+y, 2y\}$$ and so ${\alpha}_3'' \le 3$. Let us examine in turn the possibilities for $A_3 \cap D$ and their consequences for $A_4 \cap D$:
1. If $A_3 \cap D = \emptyset$ then $A_4 \cap D \subseteq \{x+y,2y\}\cup (\{x,y\}+P_3)$.
2. If $A_3 \cap D = \{2x\}$ then $A_4 \cap D \subseteq \{3x,x+y,2y\}\cup (\{x,y\}+P_3)$.
3. If $A_3 \cap D = \{x+y\}$ then $A_4 \cap D \subseteq \{2y\}\cup (\{x,y\}+P_3)$.
4. If $A_3 \cap D = \{2y\}$ then $A_4 \cap D \subseteq \{x+y,3y\}\cup (\{x,y\}+P_3)$.
5. If $A_3 \cap D = \{2x, x+y\}$ then $A_4 \cap D \subseteq \{3x,2x+y,2y\}\cup (\{x,y\}+P_3)$.
6. If $A_3 \cap D = \{2x, 2y\}$ then $A_4 \cap D \subseteq \{3x,x+y,2y\}\cup (\{x,y\}+P_3)$.
7. If $A_3 \cap D = \{x+y, 2y\}$ then $A_4 \cap D \subseteq \{x+2y,3y\}\cup (\{x,y\}+P_3)$.
8. If $A_3 \cap D = \{2x, x+y, 2y\}$ then $A_4 \cap D \subseteq \{3x,2x+y,x+2y,3y\}\cup (\{x,y\}+P_3)$.
Note that $|\{x,y\}+P_3| \le 2|P_3|=2{\alpha}_3'$. Consequently, distinguishing between the first seven cases and the last one, we have $$\begin{aligned}
{\alpha}_3'' \le 2 & \Rightarrow & {\alpha}_4'' \le 3+2{\alpha}_3' \\
{\alpha}_3'' = 3 & \Rightarrow & {\alpha}_4'' \le 4+2{\alpha}_3'.\end{aligned}$$
– Assume first ${\alpha}_3'' \le 2$, so that ${\alpha}_3 \le {\alpha}_3'+2$ and ${\alpha}_4'' \le 3+2{\alpha}_3'$. We have $$\begin{aligned}
W_0(S) & = & (3+{\alpha}_3')(9+{\alpha}_3)-4(3+{\alpha}_3+{\alpha}_4'')+\rho\\
& \ge & (3+{\alpha}_3')(9+{\alpha}_3)-4(6+{\alpha}_3+2{\alpha}_3')+\rho\\
& = & 3+{\alpha}_3'-{\alpha}_3+{\alpha}_3'{\alpha}_3+\rho\\
& \ge & 1+{\alpha}_3'{\alpha}_3+\rho.\end{aligned}$$
– Assume now ${\alpha}_3'' = 3$, so that ${\alpha}_4'' \le 4+2{\alpha}_3'$. Since ${\alpha}_3 \le 3$, it follows that ${\alpha}_3={\alpha}_3''=3$, whence ${\alpha}_3'=0$ and ${\alpha}_4'' \le 4$. We have $$\begin{aligned}
W_0(S) & = & (3+{\alpha}_3')(9+{\alpha}_3)-4(3+{\alpha}_3+{\alpha}_4'')+\rho\\
& \ge & 3\cdot 12-4\cdot 10+\rho\\
& = & -4+\rho.\end{aligned}$$ We now show that $\rho \ge 6$ here. Indeed, since $y \in A_2$ and $3y \in A_4$, we have $$2m-\rho \le y, \ \ 3y \le 5m-\rho-1.$$ Therefore $3(2m-\rho) \le 3y \le 5m-\rho-1$, implying $m \le 2\rho-1$. Now, since $m = 1+{\alpha}_1+{\alpha}_2+{\alpha}_3+{\alpha}_4 \ge 10$ by , it follows that $\rho \ge 6$, whence $W_0(S) \ge 2$ and we are done.
**** Then ${\alpha}(S)=(1,0,{\alpha}_3)$ here, and incidentally ${\alpha}_3 \le 5$ since $|L| \le 12$. We have $$|P \cap L| = 2+{\alpha}_3', \ \ |L| = 7+{\alpha}_3, \ \ |D_4| = 2+{\alpha}_3+{\alpha}_4''.$$
Recalling that $A_1=\{x\}$, we distinguish the cases where $2x \in A$ or not.
[[$\bullet$ ]{}]{}Assume first $2x \in A$. Since $\lambda(2x) \in \{2,3\}$ and since $A_2=\emptyset$, it follows that $2x \in A_3$. Hence ${\alpha}_3''=1$ and so ${\alpha}_3=1+{\alpha}_3'$. Thus $$|L| = 8+{\alpha}_3', \ \ |D_4| = 3+{\alpha}_3'+{\alpha}_4''.$$ Since $A_4 \cap D \subseteq \{3x\} \cup P_3$, it follows that ${\alpha}_4'' \le 1+{\alpha}_3'$ and hence $|D_4| \le 4+2{\alpha}_3'$. A straightforward computation then yields $$W_0(S) \ge \rho.$$
[[$\bullet$ ]{}]{}Assume now $2x \notin A$. It follows that $A_3 \subseteq P$, i.e. ${\alpha}_3={\alpha}_3'$, and $A_4 \cap D \subseteq x+A_3$. Therefore ${\alpha}_4'' \le {\alpha}_3$n and so $|D_4| = 2 + {\alpha}_3+{\alpha}_4'' \le 2+2{\alpha}_3$. This implies here $$W_0(S) \ge 6+\rho.$$
This concludes the case ${\alpha}_1=1$ and $q=4$.
### Case $q=5$
We now tackle the case ${\alpha}_1=1$ and $q=5$, i.e. ${\alpha}(S)=(1,{\alpha}_2,{\alpha}_3,{\alpha}_4)$. We have $|L|=9+3{\alpha}_2+2{\alpha}_3+{\alpha}_4$, whence $3{\alpha}_2+2{\alpha}_3+{\alpha}_4 \le 3$, implying ${\alpha}_2 \le 1$.
**** Then ${\alpha}(S)=(1,1,{\alpha}_3,{\alpha}_4)$. Then $|L|=12$ and ${\alpha}_3 = {\alpha}_4 =0$. Moreover, $|D_5|={\alpha}_0+{\alpha}_1+{\alpha}_2+{\alpha}_5''=3+{\alpha}_5''$. Set $$A_2=\{y\}.$$
[[$\bullet$ ]{}]{}Assume first $y \in D$. Then $y=2x$. Since $A_3=A_4=\emptyset$ by hypothesis, it follows that $|P \cap L|=2$ and $A_5 \cap D \subseteq \{3x\}$. Therefore ${\alpha}_5'' \le 1$. We conclude that $$\begin{aligned}
W_0(S) & = & 2\cdot 12 - 5(3+{\alpha}_5'')+\rho \\
& \ge & 4+\rho.\end{aligned}$$
[[$\bullet$ ]{}]{}Assume now $y \in P$. Then $|P \cap L|=3$ and $A_5 \cap D \subseteq \{2y\}$ since $\lambda(2x) \le 3$ and $\lambda(x+y) \le 4$. Hence ${\alpha}_5'' \le 1$ and $$\begin{aligned}
W_0(S) & = & 3\cdot 12 - 5(3+{\alpha}_5'')+\rho \\
& \ge & 16+\rho.\end{aligned}$$
**** Then ${\alpha}(S)=(1,0,{\alpha}_3,{\alpha}_4)$ and $|L|=9+2{\alpha}_3+{\alpha}_4$. Hence ${\alpha}_3 \le 1$. Moreover, $|D_5|={\alpha}_0+{\alpha}_1+{\alpha}_3+{\alpha}_4+{\alpha}_5''=2+{\alpha}_3+{\alpha}_4+{\alpha}_5''$. We have $A_2 = \emptyset$.
**** Then ${\alpha}(S)=(1,0,1,{\alpha}_4)$. We have $|L|=11+{\alpha}_4$, hence ${\alpha}_4 \le 1$. Since $A_2 = \emptyset$, we have $2x \in A$ if and only if $2x \in A_3$.
[[$\bullet$ ]{}]{}Assume first $2x \in A$. Then $A_3=\{2x\}$ since $A_2=\emptyset$ and ${\alpha}_3=1$. Hence $A_4 \cap D \subseteq \{3x\}$ and $A_5 \cap D \subseteq \{3x,4x\} \cup (x+A_4\cap P)$. Therefore $$\label{a5''}
{\alpha}_4''={\mathbf{1}}_4(3x), \ \ {\alpha}_5'' \le {\mathbf{1}}_5(3x)+{\mathbf{1}}_5(4x)+{\alpha}_4'.$$ Note also that if $4x \in A_5$, then $A_4=\{3x\}$ by Lemma \[downset\] and the bound ${\alpha}_4 \le 1$. We have $$\begin{aligned}
|P \cap L|=2+{\alpha}_4', \ \ |L|=11+{\alpha}_4, \ \ |D_5|=3+{\alpha}_4+{\alpha}_5''=3+{\alpha}_4'+{\mathbf{1}}_4(3x)+{\alpha}_5''.\end{aligned}$$ A straightforward computation, using , then yields $$\begin{aligned}
W_0(S) & = & (2+{\alpha}_4')(11+{\alpha}_4) - 5(3+{\alpha}_4'+{\mathbf{1}}_4(3x)+{\alpha}_5'')+\rho \\
& \ge & (2+{\alpha}_4')(11+{\alpha}_4) - 5(3+2{\alpha}_4'+{\mathbf{1}}_4(3x)+{\mathbf{1}}_5(3x)+{\mathbf{1}}_5(4x))+\rho \\
& \ge & 7-3\cdot {\mathbf{1}}_4(3x)-5\cdot {\mathbf{1}}_5(3x)-5\cdot {\mathbf{1}}_5(4x)+\rho.\end{aligned}$$ If $4x \notin A_5$ then $W_0(S) \ge 2+\rho$ since ${\mathbf{1}}_4(3x)+{\mathbf{1}}_5(3x) \le 1$ and we are done. If $4x \in A_5$, then $3x \in A_4$ as noted above, whence $W_0(S) \ge -1+\rho$. But then $4x \in C$, since $\lambda(4x)=5$ whereas $\lambda(2x)=3$. Thus $\rho \ge |C| \ge 1$, implying $W_0(S) \ge 0$, as desired.
[[$\bullet$ ]{}]{}Assume now $2x \notin A$. Then ${\alpha}_3'={\alpha}_3=1$ and $|P \cap L|=3+{\alpha}_4'$. We have $A_5 \cap D \subseteq (A_1+A_3) \cup (A_1+A_4)$, whence ${\alpha}_5'' \le 2$ since ${\alpha}_1={\alpha}_3=1$ and ${\alpha}_4 \le 1$. Therefore $|D_5|=3+{\alpha}_4+{\alpha}_5'' \le 6$, so that $$\begin{aligned}
W_0(S) & = & (3+{\alpha}_4')(11+{\alpha}_4) - 5|D_5|+\rho \\
& \ge & 3\cdot 11 -30+\rho\\
& \ge & 3+\rho.\end{aligned}$$
**** Then ${\alpha}(S)=(1,0,0,{\alpha}_4)$. We have $|L|=9+{\alpha}_4$, hence ${\alpha}_4 \le 3$ and $|D_5|=2+{\alpha}_4+{\alpha}_5''$. Since $A_2=A_3=\emptyset$, and since $2 \le \lambda(2x) \le 3$, it follows that $2x \notin A$ and $A_4=P_4$. Thus $A_5 \cap D \subseteq x+P_4$, so that ${\alpha}_5'' \le {\alpha}_4$ and $|P \cap L|=2+{\alpha}_4$. It follows that $$\begin{aligned}
W_0(S) & \ge & (2+{\alpha}_4')(9+{\alpha}_4) - 5(2+{\alpha}_4+{\alpha}_5'')+\rho \\
& \ge & (2+{\alpha}_4')(9+{\alpha}_4) - 5(2+2{\alpha}_4)+\rho \\
& \ge & 8+{\alpha}_4+{\alpha}_4^2+\rho.\end{aligned}$$
This concludes the case ${\alpha}_1=1$ and $q=5$.
### Case $q=6$
Still for ${\alpha}_1=1$, we now tackle the last case $q=6$. Then ${\alpha}(S)=(1,{\alpha}_2,{\alpha}_3,{\alpha}_4, {\alpha}_5)$. We have $|L|=11+4{\alpha}_2+3{\alpha}_3+2{\alpha}_4+{\alpha}_5$, whence ${\alpha}_2={\alpha}_3={\alpha}_4=0$ and ${\alpha}_5 \le 1$.
**** Then ${\alpha}(S)=(1,0,0,0,1)$. We have $|L|=12$ and $|D_6|= {\alpha}_0+{\alpha}_1+{\alpha}_5+{\alpha}_6''=3+{\alpha}_6''$. Recall that $A_1=\{x\}$. Since $\lambda(2x) \le 3$ and $A_2=A_3=A_4=\emptyset$, it follows that $2x \notin A$. Whence $A_5=P_5$, i.e. ${\alpha}_5'={\alpha}_5=1$, and $A_6 \cap D \subseteq x+A_5$. Hence $lP \cap Ll=4$ and ${\alpha}_6'' \le 1$. Consequently $|D_6| \le 4$, and $$\begin{aligned}
W_0(S) & \ge & 3\cdot 12- 6 \cdot 4+\rho \\
& = & 12+\rho.\end{aligned}$$
**** Then ${\alpha}(S)=(1,0,0,0,0)$. We have $|L|=11$ and $|D_6|=2+{\alpha}_6''$. Since $A_1=\{x\}$ and $A_2,A_3,A_4,A_5$ are all empty, it follows that $|P \cap L|=2$ and $A_6 \cap D=\emptyset$, i.e. ${\alpha}_6''=0$. Thus $$W_0(S) = 2 \cdot 11 -6 \cdot 2+\rho = 10+\rho.$$
Summarizing, we have shown that if $|L| \le 12$, ${\alpha}_1=1$ and $q=6$, then $W_0(S) \ge 10+\rho$. This concludes the case $|L| \le 12$ and ${\alpha}_1=1$.
When ${\alpha}_1=0$
-------------------
As noted at the beginning of Section \[main\], it suffices to consider the cases $4 \le q \le 7$.
### Case $q=4$
Then ${\alpha}(S)=(0,{\alpha}_2,{\alpha}_3)$. Since $\lceil q/2\rceil = 2$ here, Theorem \[thm:h\] yields $W_0(S) \ge 0$.
### Case $q=5$
Then ${\alpha}(S)=(0,{\alpha}_2,{\alpha}_3,{\alpha}_4)$. We have $12 \ge |L|=5+3{\alpha}_2+2{\alpha}_3+{\alpha}_4$. Hence ${\alpha}_2 \le 2$. We now examine successively the cases ${\alpha}_2=2,1,0$. Since $A_1=\emptyset$, it follows that $A_2 \subset P$.
**** Then ${\alpha}(S)=(0,2,{\alpha}_3,{\alpha}_4)$. We have $|L|=11+2{\alpha}_3+{\alpha}_4$. Hence ${\alpha}_3=0$ and ${\alpha}_4 \le 1$. Since $A_2 \subset P$, we have ${\alpha}_2'={\alpha}_2=2$ here. Thus $|P \cap L| \ge 3$. Set $$A_2=\{x_1,x_2\}.$$
Then ${\alpha}(S)=(0,2,0,1)$, so that $|L|=12$ and $|D_5|=4+{\alpha}_5''$. Denote $A_4=\{z\}$.
[[$\bullet$ ]{}]{}If $z \in P$, then $|P \cap L|=4$ and $2A_2 \cap (A_3 \cup A_4)=\emptyset$. Now $$A_5 \cap D \subseteq 2 A_2 \cup (A_2+A_4).$$ Since $|2A_2| \le 3$ and $|A_2+A_4| \le |A_2|$, it follows that ${\alpha}_5'' \le 5$. Hence $$\begin{aligned}
W_0(S) & = & |P \cap L||L|-5|D_5|+\rho \\
& \ge & 4 \cdot 12 -5 \cdot 9+\rho \\
& = & 3+\rho\end{aligned}$$ and we are done if $z \in P$.
[[$\bullet$ ]{}]{}If $z \in D$, then $|P \cap L|=3$ and $z \in A_2$ since $A_3=\emptyset$. Hence, up to renumbering, either $z=2x_1$ or $z=x_1+x_2$. If $z=2x_1$ then $A_5 \cap D \subseteq \{3x_1,x_1+x_2,2x_2\}$, whereas if $z=x_1+x_2$, then $A_5 \cap D \subseteq \{2x_1,2x_2\}$. In either case, we have ${\alpha}_5'' \le 3$. Therefore $|D_5|\le 7$. It follows that $$\begin{aligned}
W_0(S) & = & |P \cap L||L|-5|D_5|+\rho \\
& \ge & 3 \cdot 12 -5 \cdot 7+\rho \\
& = & 1+\rho\end{aligned}$$ and we are done as well if $z \in D$.
Then ${\alpha}(S)=(0,2,0,0)$, so that $|L|=11$ and $|D_5|=3+{\alpha}_5''$. In that case, we have $A_5 \subseteq 2A_2$, whence ${\alpha}_5'' \le 3$ and $|D_5| \le 6$. It follows that $$W_0(S) \ge 3 \cdot 11 - 5 \cdot 6 +\rho \ge 3+\rho.$$ That concludes the subcase ${\alpha}_2=2$, i.e. ${\alpha}(S)=(0,2,{\alpha}_3,{\alpha}_4)$ here.
**** Then ${\alpha}(S)=(0,1,{\alpha}_3,{\alpha}_4)$. We have $|L|=8+2{\alpha}_3+{\alpha}_4$. Hence ${\alpha}_3 \le 2$. Since $A_2 \subset P$, we have ${\alpha}_2'={\alpha}_2=1$ here. Set $$A_2=\{x\}.$$
**** Then ${\alpha}(S)=(0,1,2,{\alpha}_4)$. We have $|L|=12+{\alpha}_4$, whence $|L|=12$ and ${\alpha}_4=0$. Thus ${\alpha}(S)=(0,1,2,0)$. Hence $|D_5|=4+{\alpha}_5''$. Since $A_3 \cap D \subseteq 2A_2$, it follows that ${\alpha}_3'' \le 1$ and hence ${\alpha}_3' \in \{1,2\}$. Therefore $|P \cap L| \ge 3$. Set $$A_3=\{y_1,y_2\}.$$
[[$\bullet$ ]{}]{}Assume first ${\alpha}_3'=1$. Hence ${\alpha}_3''=1$, and up to renumbering, we may assume $y_1=2x$, $y_2 \in P$. Hence $|P \cap L|=3$ and $$A_5 \cap D \subseteq \{3x, x+y_2, 2y_2\}.$$ Therefore ${\alpha}_5'' \le 3$ and so $|D_5| \le 7$. We then have $$\begin{aligned}
W_0(S) & = & |P \cap L||L|-5|D_5|+\rho \\
& \ge & 3 \cdot 12 -5 \cdot 7+\rho \\
& = & 1+\rho\end{aligned}$$ and we are done here.
[[$\bullet$ ]{}]{}Assume now ${\alpha}_3'={\alpha}_3=2$. Thus $y_1,y_2 \in P$ and so $|P \cap L|=4$. We have $$A_5 \cap D \subseteq \{2x\} \cup (x+A_3) \cup (2A_3 \cap A_5).$$ Thus ${\alpha}_5'' \le 3 + |2A_3 \cap A_5|$, and of course $|2A_3 \cap A_5| \le |2A_3| \le 3$ since $|A_3|=2$. At this point we have $$|D_5| \le 7 + |2A_3 \cap A_5|.$$ Therefore $$\begin{aligned}
W_0(S) & = & |P \cap L||L|-5|D_5|+\rho \\
& \ge & 4 \cdot 12 -5 \cdot (7+|2A_3 \cap A_5|)+\rho \\
& = & 13-5|2A_3 \cap A_5|+\rho.\end{aligned}$$ Now $\rho \ge |2A_3 \cap A_5|$, since $2A_3 \cap A_5 \subseteq C$ and $|C| \le \rho$. Hence $$W_0(S) \ge 13-4|2A_3 \cap A_5|.$$ But $|2A_3 \cap A_5| \le 3$ as seen above. It follows that $W_0(S) \ge 1$ and we are done here.
**** Then ${\alpha}(S)=(0,1,1,{\alpha}_4)$. We have $$\label{general}
|P \cap L| = 2+{\alpha}_3'+{\alpha}_4', \ \ |L| = 10+{\alpha}_4, \ \ |D_5| = 3+{\alpha}_4+{\alpha}_5''.$$ We examine four subcases, depending on which multiples of $x$ belong to $A$.
**Case 1. $2x \in A_3, 3x \in A_4$.** Then $A_3=\{2x\}$ since ${\alpha}_3=1$, and $A_4 \cap D = \{3x\}$. Hence ${\alpha}_4={\alpha}_4'+1$. Moreover, $$A_5 \cap D \subseteq \{4x\} \cup (x+A_4 \cap P).$$ Whether $4x$ belongs to $A_5$ or not will be measured by ${\mathbf{1}}_5(4x) \in \{0,1\}$, where ${\mathbf{1}}_i={\mathbf{1}}_{A_i}$ denotes the indicator function of $A_i$ for all $i$. Thus, $${\alpha}_5'' \le {\mathbf{1}}_5(4x)+{\alpha}_4',$$ and so $$\label{data}
|P \cap L| = 2+{\alpha}_4', \ \ |L| = 11+{\alpha}_4', \ \ |D_5| = 4+{\alpha}_4' +{\alpha}_5'' \le 4+2{\alpha}_4'+{\mathbf{1}}_5(4x).$$ Plugging this data into , we get $$W_0(S) \ge 2+3{\alpha}_4'+{\alpha}_4'^2-5\cdot {\mathbf{1}}_5(4x)+\rho.$$ If either ${\alpha}_4' \not=0$ or ${\mathbf{1}}_5(4x)=0$, then $W_0(S) \ge \rho$ and we are done. If ${\alpha}_4'=0$ and ${\mathbf{1}}_5(4x)=1$, then $W_0(S) \ge -3+\rho$. But in this case, we claim that $\rho \ge 3$. Indeed, both $2x,3x$ are *compressed* Apéry elements, since $\lambda(x) = 2, \lambda(2x) = 3, \lambda(3x) =4$ and $2x=x+x, 3x=x+2x$. Moreover, since ${\mathbf{1}}_5(4x)=1$ here, then $4x$ is also a compressed Apéry element, using $\lambda(4x) =5$ and $4x=2x+2x$. Hence $\{2x,3x,4x\} \subseteq C$. Since $\rho \ge |C|$, this proves $\rho \ge 3$ as claimed and yields $W_0(S) \ge 0$, as desired.
**Case 2. $2x \in A_3, 3x \notin A_4$.** Then $A_3=\{2x\}$ and $A_4 \subset P$, so that ${\alpha}_4={\alpha}_4'$. We have $$A_5 \cap D \subseteq \{3x\} \cup (x+A_4).$$ Hence ${\alpha}_5'' \le 1+{\alpha}_4$, and so $$|P \cap L| = 2+{\alpha}_4, \ \ |L| = 10+{\alpha}_4, \ \ |D_5| = 3+{\alpha}_4 +{\alpha}_5'' \le 4+2{\alpha}_4.$$ Hence $W_0(S) \ge 2{\alpha}_4+{\alpha}_4^2+\rho$ and we are done in the present case.
**Case 3. $2x \in A_4$.** Then $A_3=A_3 \cap P$, so that ${\alpha}_3'={\alpha}_3=1$. We set $A_3=\{y\}$. We have $\{2x\} \subseteq A_4 \cap D \subseteq \{2x,x+y\}$ and $A_5 \cap D \subseteq \{3x,2y,x+y\}$. Since $x+y$ cannot belong to both $A_4, A_5$, it follows that ${\alpha}_4''+{\alpha}_5'' \le 4$.
Assume first ${\alpha}_4=2$. We then have $$|P \cap L| = 3+{\alpha}_4', \ \ |L| = 12, \ \ |D_5| = 3+{\alpha}_4'+{\alpha}_4'' +{\alpha}_5'' \le 7+{\alpha}_4'.$$ It follows that $W_0(S) \ge 12(3+{\alpha}_4') - 5(7+{\alpha}_4')+\rho \ge 1+\rho$ and we are done.
Assume now ${\alpha}_4=1$. Then $A_4=\{2x\}$ and $|L|=11$. If ${\alpha}_5'' \le 2$, then $$|P \cap L| = 3, \ \ |L| = 11, \ \ |D_5| = 4 +{\alpha}_5'' \le 6.$$ It follows that $W_0(S) \ge 3 \cdot 11 - 5 \cdot 6 +\rho \ge 3+\rho$ and we are done. But if ${\alpha}_5''=3$, i.e. if $A_5 \cap D = \{3x,2y,x+y\}$, then we only get $W_0(S) \ge -2+\rho$. But in this case, since $3x,2y \in C$, it follows that $\rho \ge 2$ whence $W_0(S) \ge 0$ as desired.
**Case 4. $2x \notin A$.** Then as above, $A_3=A_3 \cap P$, so that ${\alpha}_3'={\alpha}_3=1$, and we set $A_3=\{y\}$. If $x+y \in A$, then it belongs to either $A_4$ or $A_5$. But in any case, $x+2y \notin A$ since $\lambda(x+2y) \ge 6$, as easily seen. It follows that $A_4 \cap D \subseteq \{x+y\}$ and $A_5 \cap D \subseteq \{x+y,2y\}$. Hence $${\alpha}_4={\alpha}_4'+{\mathbf{1}}_4(x+y), \ \ {\alpha}_5'' \le 1+{\mathbf{1}}_5(x+y).$$ It follows that $$|P \cap L| = 3+{\alpha}_4', \ \ |L| = 10+{\alpha}_4'+{\mathbf{1}}_4(x+y), \ \ |D_5| \le 4+{\alpha}_4'+{\mathbf{1}}_4(x+y)+{\mathbf{1}}_5(x+y).$$ A straightforward computation, using ${\mathbf{1}}_4(x+y)+{\mathbf{1}}_5(x+y) \le 1$, then yields $$\begin{aligned}
W_0(S) & \ge & 10 +8{\alpha}_4'+{\alpha}_4'^2+({\alpha}_4'-2)\cdot {\mathbf{1}}_4(x+y)-5\cdot {\mathbf{1}}_5(x+y)+\rho \\
& \ge & 5+\rho\end{aligned}$$ and we are done.
**** Then ${\alpha}(S)=(0,1,0,{\alpha}_4)$. We have $$|P \cap L| = 2+{\alpha}_4', \ \ |L| = 8+{\alpha}_4, \ \ |D_5| = 2+{\alpha}_4+{\alpha}_5''.$$ Here $A_3=\emptyset$, and ${\alpha}_4 \le 4$ since $|L| \le 12$ by assumption.
**Case 1. $2x \in A$.** Then $2x \in A_4 \sqcup A_5$. Thus $$A_4\cap D \subseteq \{2x\}, \ \ A_5\cap D \subseteq \{2x,3x\} \cup (x+A_4\cap P).$$ Hence ${\alpha}_4={\alpha}_4'+{\alpha}_4''={\alpha}_4'+{\mathbf{1}}_4(2x)$ and ${\alpha}_5'' \le {\mathbf{1}}_5(2x)+{\mathbf{1}}_5(3x)+{\alpha}_4'$. We have $$|P \cap L| = 2+{\alpha}_4', \ \ |L| = 8+{\alpha}_4'+{\mathbf{1}}_4(2x), \ \ |D_5| \le 2+2{\alpha}_4'+{\mathbf{1}}_4(2x)+{\mathbf{1}}_5(2x)+{\mathbf{1}}_5(3x).$$ A straightforward computation then yields $$W_0(S) \ge 6-3\cdot {\mathbf{1}}_4(2x)-5\cdot {\mathbf{1}}_5(2x)-5\cdot {\mathbf{1}}_5(3x)+\rho.$$ If $3x \notin A$, then ${\mathbf{1}}_5(3x)=0$ and $W_0(S) \ge 1+\rho$. The case $3x \in A$ is more delicate. It implies $3x \in A_5$, i.e. ${\mathbf{1}}_5(3x)=1$, and then of course $2x \in A_4$, i.e. ${\mathbf{1}}_4(2x)=1$ and ${\mathbf{1}}_5(2x)=0$. Thus $W_0(S) \ge -2+\rho$. It remains to show $\rho \ge 2$. Since $(\lambda(x), \lambda(2x), \lambda(3x)) = (2,4,5)$ here, it follows that $3x \in C$, whence $\rho \ge 1$. This is not strong enough yet. However, since $(\lambda(2x), \lambda(3x)) = (4,5)$, we have $$4m-\rho \le 2x, \ \ 3x \le 6m-\rho-1.$$ Therefore $3(4m-\rho)/2 \le 3x \le 6m-\rho-1$, implying $\rho \ge 2$ as desired, and hence $W_0(S) \ge 0$.
**Case 2. $2x \notin A$.** Then $A_3=A_4 \cap D=\emptyset$. Hence $A_4=A_4 \cap P$ and $A_5 \cap D \subseteq x+A_4$. That is, we have ${\alpha}_4={\alpha}_4'$ and ${\alpha}_5'' \le {\alpha}_4$. Thus $|D_5| \le 2+2{\alpha}_4$, and a straightforward computation yields $$W_0(S) \ge 6 + \rho.$$
### Case $q=6$
Then ${\alpha}(S)=(0,{\alpha}_2,{\alpha}_3,{\alpha}_4,{\alpha}_5)$. We have $
12 \ge |L| = 6+4{\alpha}_2+3{\alpha}_3+2{\alpha}_4+{\alpha}_5$, whence ${\alpha}_2 \le 1$.
**** Then $|L|=10+3{\alpha}_3+2{\alpha}_4+{\alpha}_5$, implying ${\alpha}_3=0$ and ${\alpha}_4 \le 1$. Thus ${\alpha}(S)=(0,1,0,{\alpha}_4,{\alpha}_5)$. We set $$A_2 = \{x\},$$ with $x \in P$ since $A_2 \cap D = \emptyset$. We have $A_3=\emptyset$.
**** Then $|L|=12$, whence ${\alpha}_5=0$ and so ${\alpha}(S)=(0,1,0,1,0)$. Set $A_4=\{z\}$. Since $A_3=A_5=\emptyset$, it follows that $2x \in A$ if and only if $2x=z$. That is, either $z \in A\cap P$ and $2x \notin A$, or else $z=2x$. In either case, we have $$A_6 \cap D \subseteq x+A_4,$$ whence ${\alpha}_6'' \le 1$. Summarizing, we have $$|P \cap L| = 2+{\alpha}_4', \ \ |L| = 12, \ \ |D_6| = 3+{\alpha}_6'' \le 4.$$ It follows that $$W_0(S) \ge (2+{\alpha}_4')\cdot 12 - 6\cdot 4 + \rho \ge \rho$$ and we are done here.
**** Then ${\alpha}(S)=(0,1,0,0,{\alpha}_5)$. Since $\lambda(2x) \in [3,5]$ and $A_3=A_4=\emptyset$, we have $2x \in A$ if and only if $2x \in A_5$. Therefore $A_5 \cap D \subseteq \{2x\}$, i.e. ${\alpha}_5'' \le 1$. However $3x \notin A_6$. For if $3x \in A_6$, then $2x \in A_5$ by Lemma \[downset\], whence $$5m-\rho \le 2x, \ \ 3x \le 7m-\rho-1.$$ Thus $3(5m-\rho)/2 \le 3x \le 7m-\rho-1$, whence $\rho \ge m+2$, in contradiction with $\rho \in [0,m[$. Therefore $A_6 \cap D \subseteq x+(A_5 \cap P)$, i.e. ${\alpha}_6'' \le {\alpha}_5'$. Summarizing, we have $$|P \cap L| = 2+{\alpha}_5', \ \ |L| = 10+{\alpha}_5, \ \ |D_6| = 2+{\alpha}_5+{\alpha}_6'' \le 2+{\alpha}_5+{\alpha}_5'.$$ A straightforward computation then yields $$W_0(S) \ge 8 - 4{\alpha}_5''+\rho,$$ whence $W_0(S) \ge 4+\rho$ since ${\alpha}_5'' \le 1$.
**** Then ${\alpha}(S)=(0,0,{\alpha}_3, {\alpha}_4, {\alpha}_5)$ and Theorem \[thm:h\] yields $W_0(S) \ge 0$.
### Case $q=7$
Here ${\alpha}(S)=(0,{\alpha}_2,{\alpha}_3,{\alpha}_4,{\alpha}_5,{\alpha}_6)$. Then $12 \ge |L| \ge 7+5{\alpha}_2$, whence ${\alpha}_2 \le 1$.
[[$\bullet$ ]{}]{}If ${\alpha}_2=1$, then $|L|=12$ and ${\alpha}_3=\dots={\alpha}_6=0$. It follows that $A \cap D=\emptyset$, whence ${\alpha}_7''=0$. Thus $|P \cap L|=2$ and $|D_7| = 2$. Therefore $$W_0(S) = 2 \cdot 12 - 7\cdot 2+\rho=10+\rho$$ and we are done in this case.
[[$\bullet$ ]{}]{}If ${\alpha}_2=0$, then $|L|=7+4{\alpha}_3+3{\alpha}_4+2{\alpha}_5+{\alpha}_6$, whence ${\alpha}_3 \le 1$.
[[$\bullet\bullet$ ]{}]{}Assume ${\alpha}_3=1$. Then ${\alpha}(S)=(0,0,1,{\alpha}_4,{\alpha}_5,{\alpha}_6)$ and $|L|=11+2{\alpha}_5+{\alpha}_6$, whence ${\alpha}_4={\alpha}_5=0$ and ${\alpha}_6 \le 1$. Thus in fact, ${\alpha}(S)=(0,0,1,0,0,{\alpha}_6)$ and $|L|=11+{\alpha}_6$. We set $$A_3=\{x\}.$$ Of course $x \in P$ since $A_1=A_2=\emptyset$. Since $A_4=A_5=\emptyset$, it follows that $A_6 \cap D \subseteq \{2x\}$. Moreover, since $(S_3+S_6)\cap S_7=\emptyset$, we also have $A_7 \cap D \subseteq \{2x\}$. Thus $${\alpha}_6''={\mathbf{1}}_6(2x), \ \ {\alpha}_7''={\mathbf{1}}_7(2x).$$ Summarizing, we have $$|P \cap L| = 2+{\alpha}_6', \ \ |L| = 11+{\alpha}_6, \ \ |D_7| = 2+{\alpha}_6+{\alpha}_7'' = 2+{\alpha}_6'+{\mathbf{1}}_6(2x)+{\mathbf{1}}_7(2x).$$ A straightforward computation, using ${\mathbf{1}}_6(2x)+{\mathbf{1}}_7(2x) \le 1$, then yields $$W_0(S) \ge 1+\rho$$ and we are done.
[[$\bullet\bullet$ ]{}]{}Finally, assume ${\alpha}_3=0$. Then ${\alpha}(S)=(0,0,0,{\alpha}_4,{\alpha}_5,{\alpha}_6)$. In that case, since ${\alpha}_1=\dots={\alpha}_{h-1}=0$ where $h=\lceil q/2 \rceil = 4$, Theorem \[thm:h\] implies $W_0(S) \ge 0$.
### Case $q \ge 8$
For depth $q \ge 8$, Proposition \[prop q ge 8\] yields $W_0(S) \ge 0$.
Concluding remarks
------------------
Having examined above all possible Apéry profiles ${\alpha}(S)$ compatible with $|L| \le 12$, the proof of Theorem \[main\] is now complete.
As mentioned earlier, among the more than $10^{13}$ numerical semigroups of genus $g \le 60$, exactly five of them satisfy $W_0(S) < 0$. These five exceptions have depth $q=4$ and Apéry profile ${\alpha}(S)=(2,0,3)$, yielding $|L|=4+2\cdot 3+3=13$. Moreover, they satisfy $c=4m$, $|P \cap L|=|P_1|=3$ and ${\alpha}_4''=4$, whence $|D_4|=10$ and $W_0(S) = 3\cdot 13 - 4\cdot 10=-1$.
Infinite families of numerical semigroups satisfying $W_0(S)<0$ have been constructed in [@D; @Jean]. However, they all turn out to satisfy $W(S) \ge 0$. It would be very interesting to characterize those $S$ satisfying $W_0(S)<0$, but that will likely be hard to achieve. Less ambitiously, can one determine how many such cases occur in, say, genus $g \le 100$?
[99]{}
<span style="font-variant:small-caps;">W. Bruns, P. Garcia-Sanchez, C. O’Neill, D. Wilburne</span>, *Wilf’s conjecture in fixed multiplicity*, International Journal of Algebra and Computation (2020). <https://doi.org/10.1142/S021819672050023X>. <span style="font-variant:small-caps;">M. Delgado</span>, *On a question of Eliahou and a conjecture of Wilf*. Math. Z. 288 (2018) 595–627. <span style="font-variant:small-caps;">M. Delgado</span>, *Conjecture of Wilf: a survey*. arXiv:1902.03461 \[math.CO\] (2019). <span style="font-variant:small-caps;">D. Dobbs and G. Matthews</span>, *On a question of Wilf concerning numerical semigroups*. Focus on Commutative Rings Research, NovaSci.Publ., NewYork, 2006, pp.193–202. <span style="font-variant:small-caps;">S. Eliahou</span>, *Wilf’s conjecture and Macaulay’s theorem*. J. Eur. Math. Soc. 20 (2018) 2105–2129. DOI: 10.4171/JEMS/807. , *A graph-theoretic approach to Wilf’s conjecture*. Elec. J. Combin. 27(2) (2020), \#P2.15, 31 pp. <https://doi.org/10.37236/9106>. <span style="font-variant:small-caps;">S. Eliahou and J. Fromentin</span>, *Near-misses in Wilf’s conjecture*. Semigroup Forum 98 (2019) 285–298. <https://doi.org/10.1007/s00233-018-9926-5>. <span style="font-variant:small-caps;">R. Fröberg, C. Gottlieb and R. Häggkvist</span>, *On numerical semigroups*. Semigroup Forum 35 (1987) 63–83. <span style="font-variant:small-caps;">J. Fromentin and F. Hivert</span>, *Exploring the tree of numerical semigroups*. Math. Comp. 85 (2016) 2553–2568. <span style="font-variant:small-caps;">N. Kaplan</span>, *Counting numerical semigroups by genus and some cases of a question of Wilf*. J. Pure Appl. Algebra 216 (2012) 1016–1032. <span style="font-variant:small-caps;">J.I. García-García, D. Marín-Aragón and A. Vigneron-Tenorio</span>, *An extension of Wilf’s conjecture to affine semigroups*. Semigroup Forum 96 (2018) 396–408. <https://doi.org/10.1007/s00233-017-9906-1>. <span style="font-variant:small-caps;">A. Moscariello and A. Sammartano</span>, *On a conjecture by Wilf about the Frobenius number*. Math. Z. 280 (2015), no. 1–2, 47–53. <span style="font-variant:small-caps;">J.C. Rosales and P.A. García-Sánchez</span>, *Numerical semigroups*. Developments in Mathematics, 20. Springer, New York, 2009. <span style="font-variant:small-caps;">A. Sammartano</span>, *Numerical semigroups with large embedding dimension satisfy Wilf’s conjecture*. Semigroup Forum 85 (2012) 439–447. <span style="font-variant:small-caps;">J.J. Sylvester</span>, *Mathematical questions with their solutions*. Educational Times 41 (1884) 21. <span style="font-variant:small-caps;">H.S. Wilf</span>, *A circle-of-lights algorithm for the “money-changing problem”*. Amer. Math. Monthly 85 (1978), no. 7, 562–565.
Authors’ addresses:
- Shalom Eliahou, Univ. Littoral Côte d’Opale, UR 2597 - LMPA - Laboratoire de Mathématiques Pures et Appliquées Joseph Liouville, F-62228 Calais, France and CNRS, FR2037, France.\
`[email protected]`
- D. Marín-Aragón, Univ. Cádiz, Facultad de Ciencias Campus Universitario Río San Pedro s/n. 11510 Puerto Real, Cádiz, España\
`[email protected]`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We derive the photon propagator in light-shell gauge (LSG), introduced in [@Georgi:2010nq] in the context of light-shell effective theory.'
author:
- |
Howard Georgi,[^1] Greg Kestin,[^2] Aqil Sajjad,[^3]\
Center for the Fundamental Laws of Nature\
Jefferson Physical Laboratory\
Harvard University\
Cambridge, MA 02138
bibliography:
- 'shell.bib'
---
\
The Photon Propagator in Light-Shell Gauge
Introduction
============
In this paper we calculate the photon propagator in what we have called light-shell gauge (LSG). The motivation for working in LSG is described in depth in [@Georgi:2010nq] (and [@LSETintro]) where we discuss light-shell effective theory (LSET). We hope that LSET, for which LSG is an essential ingredient, may eventually provide another way of looking at high-energy scattering in gauge theories.
While perturbative computations in gauge theories are most commonly carried out in covariant gauges, where the procedure has been well established [@QCDprac] there *are* venues in which non-covariant gauges may be preferable. In this paper we derive the photon propagator in light-shell gauge, which is defined by the condition $$v_\mu A^\mu = 0
\label{LSG-condition1}$$ where $$v^\mu = \left(1, \hat r\right)^\mu
\label{vmu}$$ So, in terms of the scalar potential $A^0$ and the components of $\vec A$, (\[LSG-condition1\]) can be written as $$A^0 = A_r
\label{LSG-condition2}$$ where $$A_r \equiv \hat r\dot \vec A
$$ is the radial component of $\vec A$ (and is not to be confused with $A_r$ in the covariant tensor form). Note also that because (\[vmu\]) is not well-defined at the position space origin, many of our subsequent manipulations are ill-defined there, and we expect our propagator to make sense only in the punctured space from which the origin is excluded.
A gauge that shares some characteristics with LSG is radial (Fock-Schwinger) gauge [@Radrules] which is defined by the condition $$x_\mu A^\mu = 0,$$ and has found widespread use in QCD sum-rules [@QCDsum]. Shared characteristics between LSG and radial gauge include breaking translational invariance by choosing an origin and **coordinate dependent gauge condition**. As a result, it is often convenient to use a position space formulation rather than momentum space formulation. While these gauges share some characteristics, only LSG guarantees zero field strength off of the light-shell [@Georgi:2010nq] and allows for simplification of calculations in LSET [@LSETintro]. Another important difference is that the radial gauge condition is invariant under homogeneous Lorentz transformations, while LSG is only invariant under rotations about the origin.
Since we are at such an early stage (the first, as far as we know) in exploring this gauge, we restrict our analysis to QED where we can avoid complications that come with non-abelian theories.[^4] Even in QED, we cannot use standard techniques for calculating propagators in non-covariant gauges, such as LSG. We therefore, along the road to the LSG propagator, present a different derivation which we hope may prove useful in other gauges as well.
The basic outline of our derivation is as follows. We begin by writing the photon lagrangian in LSG in a matrix form, treating $\vec A$ and $\hat r$ as column vectors. In particular, we show that the photon’s kinetic energy can be written $${\mathcal{L}}= -\frac{1}{2}\,
\begin{pmatrix}
A_r&\vec A_\perp^T \cr
\end{pmatrix}\,M\,\begin{pmatrix}
A_r\cr\vec A_\perp \cr
\end{pmatrix}
\label{formM}$$ where we treat $\vec A$ as a column vector and write $$\vec A_\perp=\vec A-\hat r\,\hat r^T\vec A
=\vec A-\left(\hat r\cdot\vec A\right)\,\hat r
\label{aperp}$$ Then in the following sections we will show how from $M$ we are able to construct the LSG propagator. This is not simply a matter of inverting $M$ because $\vec A_\perp$ does not have a radial component. What we therefore need to compute is the inverse of $M$ restricted to the subspace from which we have projected out this (non-existent) radial component. We will see that doing so turns out to be non-trivial since $M$ does not commute with the projection operator in the radial direction. As a result, we cannot express $M$ in a diagonal basis and simply take the inverse on the relevant subspace to obtain the propagator. We therefore need to follow a slightly more involved procedure. Our technique, we hope, may also be applicable to other non-covariant gauges.
The Lagrangian in LSG
=====================
We will now find the matrix $M$ in equation (\[formM\]) starting with the standard form of the photon kinetic energy: $${\mathcal{L}}= -\frac{1}{4}F_{\mu\nu} F^{\mu\nu}
= \frac{1}{2}\left({\vec\nabla}A^0 + \partial_t \vec A\right)^2
-\frac{1}{2}\left({\vec\nabla}\times\vec A\right)^2$$ We then insert the LSG condition $A^0 = A_r$, giving $${\mathcal{L}}= \frac{1}{2}
\left(\vec\nabla A_r +\partial_t \vec A\right)^2
\,-\, \frac{1}{2}\left(\vec\nabla\times\vec A\right)^2$$ To arrive at the form given in (\[formM\]), we manipulate the above terms one at a time. The first term can be written as $$\left(\partial_t \vec A +{\vec\nabla}A_r\right)^2
= \left(\hat r \left(\partial_t +\hat r\dot{\vec\nabla}\right) A_r
+{\vec\nabla}_\perp A_r
+\partial_t \vec A_\perp\right)^2
\label{Lterm1}$$ where (not yet in a matrix notation) $${\vec\nabla}_\perp
= {\vec\nabla}- \hat r \left(\hat r\cdot {\vec\nabla}\right)
\label{vecnablaperp}$$
We expand (\[Lterm1\]) to get $$= \left(\left(\partial_t +\hat r\dot{\vec\nabla}\right) A_r\right)^2
+ \left(\partial_t \vec A_\perp\right)^2
+\left({\vec\nabla}_\perp A_r\right)^2
+ 2\left({\vec\nabla}_\perp A_r\right) \dot \partial_t \vec A_\perp$$ Integrating this by parts gives
$$\begin{array}
{c}
= - A_r (\partial_t +{\vec\nabla}\cdot\hat r) (\partial_t +\hat r\cdot{\vec\nabla}) A_r
- A_r {\vec\nabla}^2 A_r
+ A_r \left({\vec\nabla}\dot\hat r\right)\left(\hat r \dot{\vec\nabla}\right) A_r
\\-\vec A_\perp \cdot \partial_t^2 \vec A_\perp
- A_r \partial_t {\vec\nabla}_\perp \cdot A_\perp
- \vec A_\perp \cdot {\vec\nabla}_\perp \partial_t A_r
\end{array}
\label{Lagrangian-piece1}$$
For the $\left({\vec\nabla}\times \vec A\right)^2$ term we can write $$\left({\vec\nabla}\times \vec A\right)^2
= \left({\vec\nabla}\times A_r \hat r
+{\vec\nabla}\times \vec A_\perp\right)^2
\label{Lagrangian-piece2-full}$$ We can work out the $rr$, $r\perp$, $\perp r$ and $\perp\perp$ terms in this separately by writing all the cross products explicitly in terms of Cartesian indices and simplifying. The $rr$ term is $$\left({\vec\nabla}\times A_r \hat r\right)^2
= \left(\hat r\times{\vec\nabla}A_r\right)^2
= \left(\hat r_j \nabla_k A_r\right)\left(\hat r_j \nabla_k A_r\right)
- \left(\hat r_j \nabla_k A_r\right) \left(\hat r_k \nabla_j A_r\right)$$ $$= \left({\vec\nabla}A_r\right)^2
- \left(\hat r_k \nabla_k A_r\right) \left(\hat r_j \nabla_j A_r\right)$$ $$= \left({\vec\nabla}A_r\right)^2
- \left(\hat r \dot{\vec\nabla}A_r\right)^2$$ Integrating this by parts gives $$= -A_r \nabla^2 A_r
+ A_r \left({\vec\nabla}\dot\hat r\right)\left(\hat r \dot{\vec\nabla}\right) A_r
\label{Lagrangian-piece2a}$$ The $\perp\perp$ term is $$\left({\vec\nabla}\times\vec A_\perp\right)\dot\left({\vec\nabla}\times \vec A_\perp\right)
= \left(\nabla_j A_\perp^k\right) \left(\nabla_j A_\perp^k\right)
- \left(\nabla_j A_\perp^k\right) \left(\nabla_k A_\perp^j\right)$$ $$= -\vec A_\perp \dot \nabla^2 \vec A_\perp
+ \left(\vec A_\perp \dot {\vec\nabla}\right) \left({\vec\nabla}\dot \vec A_\perp\right)
\label{Lagrangian-piece2b}$$ Similarly, it can be shown that the $r\perp$ and $\perp r$ terms are $$\left({\vec\nabla}\times A_r \hat r\right) \dot \left({\vec\nabla}\times \vec A_\perp\right)
= A_r \left({\vec\nabla}\dot \hat r\right) \left({\vec\nabla}\dot \vec A_\perp\right)
\label{Lagrangian-piece2c}$$ and $$\left({\vec\nabla}\times \vec A_\perp\right) \dot \left({\vec\nabla}\times A_r \hat r\right)
= \left(\vec A_\perp \dot {\vec\nabla}\right) \left(\hat r\dot{\vec\nabla}A_r\right)
\label{Lagrangian-piece2d}$$
Combining all the terms from (\[Lagrangian-piece1\]), (\[Lagrangian-piece2a\]), (\[Lagrangian-piece2b\]), (\[Lagrangian-piece2c\]), and (\[Lagrangian-piece2d\]), we can write the Lagrangian in the matrix form in (\[formM\]) repeated below $${\mathcal{L}}= -\frac{1}{2}\,
\begin{pmatrix}
A_r&\vec A_\perp^T \cr
\end{pmatrix}\,M\,\begin{pmatrix}
A_r\cr\vec A_\perp \cr
\end{pmatrix}
\label{Matrix-form}$$ where we now know the matrix $M$ is given by $$M = \begin{pmatrix}
(\partial_t +{\vec\nabla}\dot\hat r)\cr
\vec\nabla \cr
\end{pmatrix}
\begin{pmatrix}
(\partial_t+\hat r\dot {\vec\nabla})&\vec\nabla^T \cr
\end{pmatrix}
+\begin{pmatrix}
0&0 \cr
0& I\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,\cr\end{pmatrix}
\label{matrix-lightshell-gauge}$$
Now things get a little complicated. The $4\times4$ matrix differential operator $M$ is invertible, but its inverse is not the propagator we want. The LSG propagator is the inverse of $M$ restricted to the subspace from which we have projected out the (non-existent) radial component of $\vec
A_\perp$. Let $P$ be the projection operator onto the radial direction of $\vec A$. Then the inverse we are looking for is the operator $D$ satisfying $$\begin{array}{c}
\displaystyle P\,D=D\,P=0
\\ \displaystyle
\left(I-P\right)\,M\,\left(I-P\right)\,D
=D\,\left(I-P\right)\,M\,\left(I-P\right)
=\left(I-P\right)
\end{array}
\label{D-subspace-inverse-property}$$ Because $P$ does not commute with $M$, we cannot simply invert $M$ and then project onto the relevant subspace. Instead, we will use a 2-step procedure. We will first show how the linear algebra of this 2-step procedure works in general, and then apply it to the LSG propagator in particular.
\[inversion\]Inversion on a subspace
====================================
Our aim is to take an invertible matrix $M$, and find its inverse restricted to the subspace projected onto by $\left(I-P\right)$, where $P$ is a projection operator onto a subspace and $I$ is the identity matrix. That is, we wish to find the matrix $D$ satisfying (\[D-subspace-inverse-property\]). There are two steps. Step one (which, for LSG, we will put off until later and relegate to an appendix) is to find the inverse of $M^{-1}$ on the space projected onto by $P$. That is, we find an operator $\nu$ satisfying $$\nu\,P=P\,\nu=\nu
\quad\quad\quad
\nu\,P\,M^{-1}\,P=P\,M^{-1}\,P\,\nu=P
\label{nu-condition}$$ Then in step two we consider the following operator: $$D = M^{-1}
- M^{-1} \,\nu\, M^{-1}
= M^{-1}
- M^{-1} \,P\,\nu\,P\, M^{-1},
\label{B}$$ It is straightforward to apply (\[nu-condition\]) to see that $D$ satisfies (\[D-subspace-inverse-property\]), and thus it is the desired inversion of $M$ on the subspace projected by $\left(I-P\right)$.
Returning to the LS gauge propagator
====================================
We now show how we can apply (\[B\]) to find the LSG propagator. In this and the following sections we will use an operator notation (discussed in more detail in appendix \[operator-notation\]) in which differential operators, their inverses, and ordinary functions of coordinates are all treated as linear operators acting on the tensor product space of our $4$-component index space and the space of functions of the coordinates.
In this language, the projection operator $P$ is $$P=\begin{pmatrix}0&0\cr0&\hat R\hat R^T\end{pmatrix}$$ Since the formula (\[B\]) for the inverse on a subspace involves the inverse of $M$ on the full space, we must begin by finding $M^{-1}$. For this purpose, it is convenient to note that $M$ can be written in terms of a diagonal matrix $M_d$ and a triangular matrix $T$ as (where $I_n$ is the $n\times n$ identity operator) $$M = TM_d T^\d,$$ where $$M_d = \begin{pmatrix}
\left(\partial_t+{\vec\nabla}^T\hat R\right) \left(\partial_t +\hat R^T\vec\nabla\right) & 0 \cr
0&{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\cr
\end{pmatrix},
\label{D-dagger}$$ $$T= \begin{pmatrix}
1&0\cr
{\vec\nabla}\left(\partial_t+\vec\nabla^T\hat R\right)^{-1}&I_{3} \cr
\end{pmatrix}
\label{T}$$ and $$T^\d = \begin{pmatrix}
1&\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}\vec\nabla^T\cr 0&I_{3} \cr
\end{pmatrix}
\label{T-dagger}$$ This makes inverting $M$ straightforward, and we get for $M^{-1}$ $$\begin{pmatrix}
\left(\partial_t +\hat R^T{\vec\nabla}\right)^{-1} \left(1 + {\vec\nabla}^T {\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1} {\vec\nabla}\right)
\left(\partial_t +{\vec\nabla}^T\hat R\right)^{-1}
& -\left(\partial_t+\hat R^T{\vec\nabla}\right)^{-1}{\vec\nabla}^T {\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}
\cr
-{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1} {\vec\nabla}\left(\partial_t+\vec\nabla^T\hat R\right)^{-1}
& {\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\cr\end{pmatrix}
\label{M-inverse}$$ The next ingredient we need is the inverse of $M^{-1}$ restricted to the subspace. Here it is useful to avoid the matrix structure and define a linear operator $\mu$, as $$\mu=\begin{pmatrix}0&\hat R^T\end{pmatrix}\, M^{-1}\,\begin{pmatrix}0\cr\hat R\end{pmatrix}
\label{def-mu}$$ whence $\nu$ in (\[nu-condition\]) is given by $$\nu=\begin{pmatrix}0&0\cr 0&\hat R\,\mu^{-1}\,\hat R^T\end{pmatrix}
$$
Now we can just use (\[B\]) and put the pieces together to formally compute the LSG propagator. Doing so and simplifying gives the following results: $$D_{rr}=\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}
\,\left(1+\vec\nabla\,^T\,C\,\vec\nabla\,\right)\,
\left(\partial_t+\vec\nabla^T\hat R\right)^{-1}
\label{radial-propagator}$$ $$D_{r\perp}=-\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}
\vec\nabla\,^T\,C\
\label{r-perp-propagator}$$ $$D_{\perp r}=
-C\,\vec\nabla\, \left(\partial_t+\vec\nabla^T\hat R\right)^{-1}
$$ $$D_{\perp\perp}=C
\label{prop-perp}$$ where $C$ is given by $$C = {\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1} -{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1} \,\hat R\, \mu^{-1}\, \hat R^T \,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}
\label{C-def}$$ Note that from this form we can see that $C$ is transverse; that is, if we act with the projection operator for the transverse subspace on either side of $C$, we get $C$. What remains to be done is to derive an explicit form for $C$, which is done in detail in appendix \[Cappendix\], with the result $$C=-R\,\vec\nabla_\perp\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,R\,\vec\nabla_\perp^T
+\vec L\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,\vec L^T
\label{C}$$ where $$R\equiv |\vec R|.
$$ Since this involves $L^{-2}$, we must show that this is well defined. We show in appendix \[L20\] that because of the operators that appear on either side of $L^{-2}$ in (\[C\]), the $L^{-2}$ operator never acts on an $L=0$ state, and the expression (\[C\]) makes sense.
Putting (\[C\]) into (\[radial-propagator\]-\[prop-perp\]) gives $$D_{rr}
=\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}
\,\left(1-R^{-1}\,L^2\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,R^{-1}\right)\,
\left(\partial_t+\vec\nabla^T\hat R\right)^{-1}
\label{2radial-propagator}$$ $$D_{r\perp}
=\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}
\,R^{-1}\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,\vec\nabla_\perp^T\,R
\label{2r-perp-propagator}$$ $$D_{\perp r}=
R\,\vec\nabla_\perp\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,R^{-1}
\, \left(\partial_t+\vec\nabla^T\hat R\right)^{-1}
$$ $$D_{\perp\perp}=
-R\,\vec\nabla_\perp\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,R\,\vec\nabla_\perp^T
+\vec L\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,\vec L^T
\label{2prop-perp}$$
We can also combine these into a $3\times3$ matrix form, call it $D_3$, appropriate for unconstrained $\vec A$ fields:
$$\begin{array}{c}
\displaystyle
\hat R\,\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}
\left(\partial_t+\vec\nabla^T\hat R\right)^{-1}\,\hat R^T
+\vec L\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,\vec L^T
\\ \displaystyle
-\Bigl(R\,\vec\nabla_\perp\,L^{-2}
-\hat R\,\left(\partial_t+\hat R^T\vec\nabla\right)^{-1}
\,R^{-1}\Bigr)
\,L^2\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,
\Bigl(L^{-2}\,R\,\vec\nabla_\perp^T-R^{-1}
\, \left(\partial_t+\vec\nabla^T\hat R\right)^{-1}\,\hat R^T\Bigr)
\end{array}
\label{unconstrained}$$
Conclusion
==========
Here we have derived the photon propagator in light-shell gauge. In the process of this derivation, we have presented a technique that may also be useful for calculations in other non-covariant gauges (and, we hope, other applications). LSG is a crucial part of the construction of the light-shell effective theory [@Georgi:2010nq], which we hope may provide a new viewpoint for high-energy scattering in gauge theories. We also hope that further insight can be gained once this method is extended to non-abelian gauge theories.
Acknowledgements {#acknowledgements .unnumbered}
================
We have benefited greatly from suggestions by Matthew Schwartz, Benjamin Grinstein, Randall Kelley, Aneesh Manohar and David Simmons-Duffin. Some of the initial work on this project by HG was done at the Aspen Center for Physics. He is grateful for the support of the Center and National Science Foundation grant \#1066293. This research has been supported at Harvard in part by the National Science Foundation under grants PHY-0804450 and and PHY-1067976.
Operator Notation
=================
Throughout we have used a notation that involves local and non-local operators. For example, when a local operator, such as $R^{-1}$ appears, it is $$R^{-1}\left(x_{1},x_{2}\right)=\frac{1}{r_{1}}\delta\left(x_{1}-x_{2}\right)$$ and when not written, the delta function and integrations over the arguments are implicit. We also come across the operators ${\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}$, $\left(\partial_t +\hat R\dot \vec\nabla\right)^{-1}$ and $\left(\partial_t +\vec\nabla\dot\hat R\right)^{-1}$. We know that ${\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}$ is the position space propagator for a massless scalar and is given by $${\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}(x-y)
= -\frac{i}{4\pi^2} \frac{1}{(x-y)^2 -i\e}$$ To find the expression for $(\partial_t + \hat R \dot \nabla)^{-1}$, we can consider the theory with the Lagrangian $${\mathcal{L}}=i \phi^* (\partial_t + \hat r \dot \nabla) \phi$$ Using canonical quantization to find the 2-point function for this theory gives $$(\partial_t + \hat R \dot \vec\nabla)^{-1}
= \frac{1}{{r'}^2} \theta(t-t')
\delta(t-r -t'+r') \delta(z-z') \delta(\phi-\phi')
\label{good-guy-inverse}$$ Similarly $$(\partial_t + \vec\nabla \dot \hat R)^{-1}
= \frac{1}{r^2} \theta(t-t')
\delta(t-r -t'+r') \delta(z-z') \delta(\phi-\phi')
\label{bad-guy-inverse}$$
Derivation of C {#Cappendix}
===============
We can find $C$ by brute force, but here we will instead use a slicker approach, which will take advantage of (\[D-subspace-inverse-property\]). Using the formula (\[matrix-lightshell-gauge\]) for $M$ and our result for the propagator (\[radial-propagator\])-(\[prop-perp\]), it is straightforward to see that $\left(I_{4}-P\right) M\left(I_{4}-P\right) D$ is $$\begin{pmatrix}
1 &
0\cr\vec\nabla_\perp\, \left(\partial_t+{\vec\nabla}^T\hat R\right)^{-1} -\left(I_3-P_{3}\right)\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,
\,C\,\vec\nabla\, \left(\partial_t+{\vec\nabla}^T\hat R\right)^{-1}
& \left(I_3-P_{3}\right)\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,\,C
\cr\end{pmatrix}$$ where $P_3=\hat R\,\hat R^T$, and we have used $\left(I_3-P_{3}\right)C=C$. For $D$ to be the LSG propagator, we want the 2nd row entries of $\left(I-P\right) M\left(I-P\right)D$ to be $0$ and $I_3-P_{3}$. Both these requirements are satisfied if $$\left(I_{3}-P_{3}\right)\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,C = I_{3}-P_{3}
\label{ccond}$$ We will now use this condition to find an explicit expression for $C$. Our approach will involve first finding a basis for the space perpendicular to $\hat R$, and then acting on (\[ccond\]) with various operators to find the components of $C$ in this basis. We begin by identifying the proper basis. Notice that $$R\,\vec\nabla_\perp^T=i\,\left(\vec L\times\hat R\right)^T
\label{rdperp}$$ So $\vec L$ and $R\,\vec\nabla_\perp$ are both orthogonal to $\hat R$ and orthogonal to one another, therefore forming our basis. We can express $\left(I_3-P_{3}\right)$ in terms of them. First note that from (\[rdperp\]) it follows that $$R\,\vec\nabla_\perp^T\,\vec\nabla_\perp\,R=-L^2
$$ so with proper normalization we have $$\left(I_3-P_{3}\right)=\vec L\,L^{-2}\,\vec L^T
-\vec\nabla_\perp\,R\,L^{-2}\,R\,\vec\nabla_\perp^T
\label{imp}$$ Now we want to find the components of $C$. The first, and easiest component to find is computed by acting on (\[ccond\]) with $\vec L$ on both sides to give $$\vec L^T\left(I_3-P_{3}\right)\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,C\vec L=\vec L^T\left(I_3-P_{3}\right)\vec L
$$ This is easy because $\vec L$ commutes with ${\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}$, so we get $$\vec L^T\,C\,\vec L={\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^2
\label{ll}$$ Acting on the left of (\[ccond\]) with $\vec L^T$ and on the right with $\vec\nabla_\perp\,R$ as follows $$\vec L^T\left(I_3-P_{3}\right)\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,C\,\vec\nabla_\perp\,R=\vec L^T\left(I_3-P_{3}\right)\vec\nabla_\perp\,R
$$ works similarly once we observe $\vec L^T\,\vec\nabla_\perp\,R=0$, giving $$\vec L^T\,C\,\vec\nabla_\perp\,R=0
\label{ld}$$ The final two matrix elements require the commutator $$\Bigl[R\,\vec\nabla_\perp^T\,,\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,\Bigr]
=2R^{-2}\,L^2\,\hat R^T
\label{crucom}$$ We now take a detour to demonstrate this commutator relation. We can write $${\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}=\partial_t^2-\left(\vec\nabla^T\hat R\right)\,\left(\hat R^T\vec\nabla\right)
+L^2\,R^{-2}
\label{bx}$$ The middle term in (\[bx\]) can be written $$\begin{array}{c}
\left(\vec\nabla^T\hat R\right)\,\left(\hat R^T\vec\nabla\right)
=\left(\vec\nabla^T\vec R\right)\,R^{-2}\,\left(\vec R^T\vec\nabla\right)
\\
=\left(\vec R^T\vec\nabla\right)\,R^{-2}\,\left(\vec R^T\vec\nabla\right)
+3R^{-2}\,\left(\vec R^T\vec\nabla\right)
=R^{-2}\,\left(\left(\vec R^T\vec\nabla\right)^2
+\left(\vec R^T\vec\nabla\right)\right)
\end{array}
$$ We chose this particular form because $\left(\vec R^T\vec\nabla\right)$ is a scaling operator that counts the total powers $R$ or $1/\vec\nabla$.[^5] So this term commutes with $R\,\vec\nabla_\perp^T$ and the only term in ${\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}$ that fails to commute is $L^2\,R^{-2}$.
The factors of $R$ commute with both $\vec\nabla_\perp^T$ and $L^2$, so we just need to consider $$\Bigl[R\,\vec\nabla_\perp^T\,,\,L^2\,\Bigr]
$$ Using (\[rdperp\]), we can write this in components, as $$\Bigl[\,i\,\epsilon_{abc}L_b\hat R_c\,,\,L_dL_d\,\Bigr]
$$ $$=i\,\epsilon_{abc}L_b\,\left(
L_d\,\Bigl[\,\hat R_c\,,\,L_d\,\Bigr]
+\Bigl[\,\hat R_c\,,\,L_d\,\Bigr]\,L_d
\right)
$$ $$=-\epsilon_{abc}\epsilon_{cde}L_b\,\left(
L_d\hat R_e+\hat R_eL_d
\right)
$$ $$=-L_b\,\left(\bigl[L_a,\hat R_b\bigr]+2\hat R_bL_a-2L_b\hat R_a
-\bigl[\hat R_a,L_b\bigr]\right)
\label{last}$$ The first and fourth terms in (\[last\]) cancel each another. The second term vanishes because $\vec L\cdot\hat R=0$. The third term gives $$\Bigl[R\,\vec\nabla_\perp^T\,,\,L^2\,\Bigr]
=2L^2\,\hat R^T
$$ or $$\Bigl[R\,\vec\nabla_\perp^T\,,\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,\Bigr]
=2R^{-2}\,L^2\,\hat R^T
\label{provedcrucom}$$ which is (\[crucom\]).
We now return to the derivation of $C$, but note that (\[provedcrucom\]) vanishes when acting on $C$. So, acting with $R\vec\nabla_\perp^T$ on the left and $\nabla_\perp\,R$ on the right gives $$R\,\vec\nabla_\perp^T\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,C\,\vec\nabla_\perp\,R=
{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,R\,\vec\nabla_\perp^T\,C\,\vec\nabla_\perp\,R=-L^2
\label{delcal2}$$ implying $$R\,\vec\nabla_\perp^T\,C\,\vec\nabla_\perp R=-{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^2
\label{dd}$$ In the same way we can see that the last component is zero $$R\,\vec\nabla_\perp^T\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,C\,\vec L=
{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}\,R\,\vec\nabla_\perp^T\,C\,\vec L=0
\label{dl}$$ Combining (\[ll\]), (\[ld\]), (\[dd\]) and (\[dl\]) with (\[imp\]) gives $$C=-R\,\vec\nabla_\perp\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,R\,\vec\nabla_\perp^T
+\vec L\,{\mathop{\raise-.8pt\hbox{\large$\Box$}}\nolimits}^{-1}\,L^{-2}\,\vec L^T
$$
Does $L^{-2}$ make sense? {#L20}
=========================
The derivation of $C$ (in appendix \[Cappendix\]) formally involves the inverse of $L^2$, and of course this makes no sense on $L=0$ states. But all we actually need is for (\[imp\]) to make sense acting on arbitrary functions, so that $$\left(I_3-P_{3}\right)\vec f(\vec r)=\vec L\,L^{-2}\,\vec L^T
\vec f(\vec r)-\vec\nabla\,R\,L^{-2}\,R\,\vec\nabla_\perp^T\vec f(\vec r)
\label{imp2}$$ This is perfectly well-defined, because if either the $\vec L^T\vec f(\vec r)$ or $R\,\vec\nabla_\perp^T\vec
f(\vec r)$ component has zero angular momentum, then that component itself is zero. This can been seen by first noting that if $L^2$ acting on either of these components is zero, then the component must be a function of the radius only, call it $g(r)$. If we integrate $g(r)$ over $d\Omega$, we get $4\pi g(r)$, but at the same time we see that integrating either component over $d\Omega$ must be zero because in both cases we are integrating a total derivative over a closed surface. Therefore $g(r)$, which denotes either component, is necessarily zero.
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
[^4]: We hope to extend this work to QCD and in the process describe attributes avoided herein (e.g. ghosts).
[^5]: Note also that the last form is trivial to remember because it vanishes for $r^a$ with $a=0$ or $-1$ as it should.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'I review the theoretical basis of the prediction that the decay $K_L \rightarrow \pi^+ \pi^- e^+ e^-$ should show a large $CP$- and $T$-violation, a prediction now confirmed by the KTeV experiment. The genesis of the effect lies in a large violation of $CP$- and $T$-invariance in the decay $K_L \rightarrow \pi^+ \pi^- \gamma$, which is encrypted in the polarization state of the photon. The decay $K_L \rightarrow \pi^+ \pi^- e^+ e^-$ serves as an analyser of the photon polarization. The asymmetry in the distribution of the angle $\phi$ between the $\pi^+ \pi^-$ and $e^+ e^-$ planes is a direct measure of the $CP$-odd, $T$-odd component of the photon’s Stokes vector. A complete study of the angular distribution can reveal further $CP$-violating features, which probe the non-radiative (charge-radius and short-distance) components of the $K_L \rightarrow \pi^+ \pi^- e^+ e^-$ amplitude.'
author:
- |
L. M. Sehgal\
*[Institute of Theoretical Physics, RWTH Aachen,]{}\
*[D-52056 Aachen, Germany]{}**
title: |
$CP$- and $T$-Violation in the Decay\
$K_L \rightarrow \pi^+ \pi^- e^+ e^-$ and Related Processes[^1]
---
Eight years ago, there appeared a report [@E731coll] by the E-731 experiment concerning the branching ratio and photon energy spectrum of the decays $K_{L,S} \rightarrow \pi^+ \pi^- \gamma$. It was found that while the $K_S$ decay could be well-reproduced by inner bremsstrahlung from an underlying process $K_S \rightarrow \pi^+ \pi^-$, the $K_L$ decay contained a mixture of a bremsstrahlung component ($I\!B$) and a direct emission component ($D\!E$), the relative strength being $D\!E/(D\!E+I\!B) = 0.68$ for photons above $20\, MeV$. The simplest matrix element consistent with these features is $$\begin{aligned}
\mathcal{M} (K_S \rightarrow \pi^+ \pi^- \gamma) & = & e f_S
\left[\frac{\epsilon \cdot p_+}{k \cdot p_+} -
\frac{\epsilon \cdot p_-}{k \cdot p_-} \right]
\label{matelem}\\
\mathcal{M} (K_L \rightarrow \pi^+ \pi^- \gamma) & = & e f_L
\left[\frac{\epsilon \cdot p_+}{k \cdot p_+} -
\frac{\epsilon \cdot p_-}{k \cdot p_-} \right] + e \frac{f_{DE}}{{M_K}^4}
\epsilon_{\mu\nu\rho\sigma} \epsilon^{\mu}k^{\nu}{p_+}^{\rho}{p_-}^{\sigma}
\nonumber\end{aligned}$$ where $$\begin{aligned}
f_L \equiv |f_S| g_{Br},\: g_{Br} = \eta_{+-} e^{i \delta_0(s={M_K}^2)},
\nonumber\\
f_{DE} \equiv |f_S| g_{M1},\: g_{M1} = i(0.76)e^{i \delta_1(s)}.
\label{whatmeansf}\end{aligned}$$ Here the direct emission has been represented by a $CP$-conserving magnetic dipole coupling $g_{M1}$, whose magnitude $|g_{M1}| = 0.76$ is fixed by the empirical ratio $DE/IB$. The phase factors appearing in $g_{Br}$ and $g_{M1}$ are dictated by the Low theorem for bremsstrahlung, and the Watson theorem for final state interactions. The factor $i$ in $g_{M1}$ is a consequence of $CPT$ invariance [@Sehgal:Costa:Lee]. The matrix element for $K_L \rightarrow \pi^+ \pi^- \gamma$ contains simultaneously electric multipoles associated with bremsstrahlung ($E1,\,E3,\,E5$ ...), which have $CP=+1$, and a magnetic $M1$ multipole with $CP = -1$. It follows that interference of the electric and magnetic emissions should give rise to $CP$-violation.
To understand the nature of this interference, we write the $K_L \rightarrow \pi^+ \pi^- \gamma$ amplitude more generally as $$\begin{aligned}
\mathcal{M}(K_L \rightarrow \pi^+ \pi^- \gamma) & = &
\frac{1}{{M_K}^3} \left\{ E(\omega,\cos \theta) \left[\epsilon \cdot p_+ \, k
\cdot p_- - \epsilon \cdot p_- \, k \cdot p_+ \right] \right. \nonumber\\
& & \left. \mbox{} + M(\omega, \cos \theta)
\epsilon_{\mu\nu\rho\sigma}\epsilon^{\mu}k^{\nu}{p_+}^{\rho}{p_-}^{\sigma}
\right\}\end{aligned}$$ where $\omega$ is the photon energy in the $K_L$ rest frame, and $\theta$ is the angle between $\pi^+$ and $\gamma$ in the $\pi^+ \pi^-$ rest frame. In the model represented by Eqs. (\[matelem\]) and (\[whatmeansf\]), the electric and magnetic amplitudes are $$\begin{aligned}
E & = & \left( \frac{2M_K}{\omega} \right)^2 \frac{g_{Br}}{1-\beta^2 \cos^2 \theta}
\nonumber\\
M & = & g_{M1}
\label{defEandM}\end{aligned}$$ where $\beta = (1- 4 {m_{\pi}}^2/s)^{1/2}$, $\sqrt{s}$ being the $\pi^+\pi^-$ invariant mass. The Dalitz plot density, summed over photon polarizations is $$\frac{d \Gamma}{d \omega \, d\! \cos \theta} = \frac{1}{512 \pi^3}
\left( \frac{\omega}{M_K} \right)^3 \beta^3 \left( 1- \frac{2 \omega}{M_K}
\right) \sin^2 \theta \left[ |E|^2 + |M|^2 \right]
\label{dGammaint}$$ Clearly, there is no interference between the electric and magnetic multipoles if the photon polarization is unobserved. Therefore, any $CP$-violation involving the interference of $g_{Br}$ and $g_{M1}$ is hidden in the polarization state of the photon.
The photon polarization can be defined in terms of the density matrix $$\rho = \left( \begin{array}{cc} |E|^2 & E^*M \\ EM^* & |M|^2 \end{array}
\right) = \frac{1}{2} \left( |E|^2+|M|^2 \right) \left[ 1 \!\!\!\!\:\: \mbox{l}
+ \vec{S} \cdot \vec{\tau} \right]$$ where $\vec{\tau} = (\tau_1, \, \tau_2, \, \tau_3)$ denotes the Pauli matrices, and $\vec{S}$ is the Stokes vector of the photon with components $$\begin{aligned}
S_1 & = & 2 Re \left( E^*M \right) / \left( |E|^2 + |M|^2 \right) \nonumber\\
S_2 & = & 2 Im \left( E^*M \right) / \left( |E|^2 + |M|^2 \right) \\
S_3 & = & \left(|E|^2 - |M|^2 \right) / \left( |E|^2 + |M|^2 \right). \nonumber\end{aligned}$$ The component $S_3$ measures the relative strength of the electric and magnetic radiation at a given point in the Dalitz plot. The effects of $CP$-violation reside in the components $S_1$ and $S_2$, which are proportional to $Re \, ({g_{Br}}^* g_{M1})$ and $Im \, ({g_{Br}}^* g_{M1})$, respectively. Of these $S_1$ is $CP$-odd, $T$-odd, while $S_2$ is $CP$-odd, $T$-even. Physically, $S_2$ is the net circular polarization of the photon: such a polarization is known to be possible in decays like $K_L \rightarrow \pi^+ \pi^- \gamma$ or $K_{L,S} \rightarrow \gamma \gamma$ whenever there is $CP$-violation accompanied by unitarity phases [@Sehgal]. To understand the significance of $S_1$, we examine the dependence of the $K_L \rightarrow \pi^+ \pi^- \gamma$ decay on the angle $\phi$ between the polarization vector $\vec{\epsilon}$ and the unit vector $\vec{n}_{\pi}$ normal to the $\pi^+ \pi^-$ plane (we choose coordinates such that $\vec{k} = (0,\, 0,\, k)$, $\vec{n}_{\pi} =
(1,\, 0, \, 0)$, $\vec{p}_+ = (0,\, p\sin \theta, \, p\cos \theta)$ and $\vec{\epsilon} = (\cos \phi, \, \sin \phi, \, 0)$): $$\frac{d \Gamma}{d \omega \, d\! \cos \theta \, d \phi} \sim
\left| E \sin \phi - M \cos \phi \right|^2 \sim
1 - \left[ S_3 \cos 2\phi + S_1 \sin 2\phi \right]$$ Thus the $CP$-odd, $T$-odd Stokes parameter $S_1$ appears as a coefficient of the term $\sin 2\phi$. The essential idea of Refs. [@Sehgal:Wanninger; @Heiliger:Sehgal] is to use in place of $\vec{\epsilon}$, the vector $\vec{n}_l$ normal to the plane of the Dalitz pair in the reaction $K_L \rightarrow \pi^+ \pi^- \gamma^* \rightarrow \pi^+ \pi^- e^+ e^-$. This motivates the study of the distribution $d\Gamma/d\phi$ in the decay $K_L \rightarrow \pi^+ \pi^- e^+ e^-$, where $\phi$ is the angle between the $\pi^+ \pi^-$ and $e^+ e^-$ planes.
To obtain a quantitative idea of the magnitude of $CP$- violation in $K_L \rightarrow \pi^+ \pi^- \gamma$, we show in Fig. \[SivsE\]a the three components of the Stokes vector as a function of the photon energy [@Sehgal:vanLeusen]. These are calculated from the amplitudes (\[defEandM\]) using weighted averages of $|E|^2$, $|M|^2$, $E^*M$ and $EM^*$ over $\cos \theta$. The values of $S_1$ and $S_2$ are remarkably large, considering that the only source of $CP$-violation is the $\epsilon$-impurity in the $K_L$ wave-function ($\epsilon = \eta_{+-}$). Clearly the $1/\omega^2$ factor in $E$ enhances it to a level that makes it comparable to the $CP$-conserving amplitude $M$. This is evident from the behaviour of the parameter $S_3$, which swings from a dominant electric behaviour at low $E_{\gamma}$ ($S_3 \approx 1$) to a dominant magnetic behaviour at large $E_{\gamma}$ ($S_3 \approx -1$), with a zero in the region $E_{\gamma} \approx 60 \, MeV$. To highlight the difference between the $T$-odd parameter $S_1$ and the $T$-even parameter $S_2$, we show in Fig. \[SivsE\]b the behaviour of the Stokes parameters in the “hermitian limit”: this is the limit in which the $T$-matrix or effective Hamiltonian governing the decay $K_L \rightarrow \pi^+ \pi^- \gamma$ is taken to be hermitian, all unitarity phases related to real intermediate states being dropped. This limit is realized by taking $\delta_0 , \, \delta_1 \rightarrow 0$, and $ar\!g \, \epsilon \rightarrow \pi /2$. The last of these follows from the fact that $\epsilon$ may be written as $$\epsilon = \frac{\Gamma_{12}-\Gamma_{21} + i \left( M_{12}-M_{21} \right)}
{\gamma_S - \gamma_L - 2 i \left( m_L - m_S \right)}$$ where $H_{eff} = M - i\Gamma$ is the mass matrix of the $K^0$-$\overline{K}^0$ system. The hermitian limit obtains when $\Gamma_{12} = \Gamma_{21} = \gamma_S
= \gamma_L = 0$. As seen from Fig. \[SivsE\]b, $S_2$ vanishes in this limit, but $S_1$ survives, as befits a $CP$-odd, $T$-odd parameter. Fig. \[SivsE\]c shows what happens in the $CP$-invariant limit $\epsilon \rightarrow 0$. It is clear that we are dealing here with a dramatic situation in which a $CP$-impurity of a few parts in a thousand in the $K_L$ wave-function gives rise to a huge $CP$-odd, $T$-odd effect in the photon polarization.
We can now examine how these large $CP$-violating effects are transported to the decay $K_L \rightarrow \pi^+ \pi^- e^+ e^-$. The matrix element for $K_L \rightarrow \pi^+ \pi^- e^+ e^-$ can be written as [@Sehgal:Wanninger; @Heiliger:Sehgal] $$\mathcal{M} (K_L \rightarrow \pi^+ \pi^- e^+ e^-) = \mathcal{M}_{br} +
\mathcal{M}_{mag} + \mathcal{M}_{CR} + \mathcal{M}_{SD}.$$ Here $\mathcal{M}_{br}$ and $\mathcal{M}_{mag}$ are the conversion amplitudes associated with the bremsstrahlung and $M1$ parts of the $K_L \rightarrow \pi^+ \pi^- \gamma$ amplitude. In addition, we have introduced an amplitude $\mathcal{M}_{CR}$ denoting $\pi^+ \pi^-$ production in a $J = 0$ state (not possible in a real radiative decay), as well as an amplitude $\mathcal{M}_{SD}$ associated with the short-distance interaction $s \rightarrow d\, e^+ e^-$. The last of these turns out to be numerically negligible because of the smallness of the $C\!K\!M$ factor $V_{ts} {V_{td}}^*$. The $s$-wave amplitude $\mathcal{M}_{CR}$, if approximated by the $K^0$ charge radius diagram, makes a small ($\sim 1 \%$) contribution to the decay rate. Thus the dominant features of the decay are due to the conversion amplitude $\mathcal{M}_{br} + \mathcal{M}_{mag}$.
Within such a model, one can calculate the differential decay rate in the form [@Heiliger:Sehgal] $$d \Gamma = I(s_{\pi}, \, s_l, \, \cos \theta_l, \, \cos \theta_{\pi}, \, \phi)
\, ds_{\pi} \, ds_l \, d\! \cos \theta_l \, d\! \cos \theta_{\pi} \, d\phi.$$ Here $s_{\pi}$ ($s_l$) is the invariant mass of the pion (lepton) pair, and $\theta_{\pi}$ ($\theta_l$) is the angle of the $\pi^+$ ($l^+$) in the $\pi^+ \pi^-$ ($l^+ l^-$) rest frame, relative to the dilepton (dipion) momentum vector in that frame. The all-important variable $\phi$ is defined in terms of unit vectors constructed from the pion momenta $\vec{p_{\pm}}$ and lepton momenta $\vec{k_{\pm}}$ in the $K_L$ rest frame: $$\begin{aligned}
\vec{n}_{\pi} = \left( \vec{p}_+ \times \vec{p}_- \right) /
\left| \vec{p}_+ \times \vec{p}_- \right|,
& \vec{n}_l = \left( \vec{k}_+ \times \vec{k}_- \right) /
\left| \vec{k}_+ \times \vec{k}_- \right|, \nonumber\\
\vec{z} = \left( \vec{p}_+ + \vec{p}_- \right) /
\left| \vec{p}_+ + \vec{p}_- \right|, \nonumber\end{aligned}$$ $$\begin{aligned}
\sin \phi \: = & \vec{n}_{\pi} \times \vec{n}_l \cdot \vec{z} & (CP = -, T = -), \\
\cos \phi \: = & \vec{n}_l \cdot \vec{n}_{\pi} & (CP = +, T = +). \nonumber\end{aligned}$$ In Ref. [@Sehgal:Wanninger], an analytic expression was derived for the 3-dimensional distribution $d\Gamma/ ds_l \, ds_{\pi} \, d\phi$, which has been used in the Monte Carlo simulation of this decay. In Ref. [@Heiliger:Sehgal], a formalism was presented for obtaining the fully differential decay function $I(s_{\pi},\, s_l, \, \cos
\theta_l, \, \cos \theta_{\pi}, \phi)$.
The principal results of the theoretical model discussed in [@Sehgal:Wanninger; @Heiliger:Sehgal] are as follows:
1\. Branching ratio: This was calculated to be [@Sehgal:Wanninger] $$\begin{aligned}
B\!R(K_L \rightarrow \pi^+ \pi^- e^+ e^-) & = & (1.3 \times 10^{-7})_{Br} +
(1.8 \times 10^{-7})_{M1} \nonumber \\
& & \mbox{} + (0.04 \times 10^{-7})_{CR} \approx 3.1 \times 10^{-7},\end{aligned}$$ which agrees well with the result $(3.32 \pm 0.14 \pm 0.28) \times 10^{-7}$ measured in the KTeV experiment [@KTeVcoll]. (A preliminary branching ratio $2.9 \times 10^{-7}$ has been reported by NA48 [@NA48coll]).
2\. Asymmetry in $\phi$ distribution: The model predicts a distribution of the form $$\frac{d\Gamma}{d\phi} \sim 1 - \left( \Sigma_3 \cos 2\phi + \Sigma_1 \sin 2\phi
\right)$$ where the last term is $CP$- and $T$-violating, and produces an asymmetry $$\mathcal{A} = \frac{\left( \int_{0}^{\pi/2} - \int_{\pi/2}^{\pi} +
\int_{\pi}^{3\pi/2} - \int_{3\pi/2}^{2\pi} \right) \frac{d\Gamma}{d\phi} d\phi}
{\left( \int_{0}^{\pi/2} + \int_{\pi/2}^{\pi} +
\int_{\pi}^{3\pi/2} + \int_{3\pi/2}^{2\pi}\right) \frac{d\Gamma}{d\phi} d\phi}
= - \frac{2}{\pi} \Sigma_1.$$ The predicted value [@Sehgal:Wanninger; @Heiliger:Sehgal] is $$|\mathcal{A}| = 15 \% \, \sin (\phi_{+-} + \delta_0({M_K}^2) - \overline{\delta}_1 )
\approx 14 \%$$ to be compared with the KTeV result [@KTeVcoll] $$|\mathcal{A}|_{KTeV} = (13.6 \pm 2.5 \pm 1.2) \%$$ The “Stokes parameters” $\Sigma_3$ and $\Sigma_1$ are calculated to be $\Sigma_3 = -0.133$, $\Sigma_1 = 0.23$. The $\phi$-distribution measured by KTeV agrees with this expectation (after acceptance corrections made in accordance with the model). It should be noted that the sign of $\Sigma_1$ (and of the asymmetry $\mathcal{A}$) depends on whether the numerical coefficient in $g_{M1}$ is taken to be $+0.76$ or $-0.76$. The data happen to support the positive sign chosen in Eq. (\[whatmeansf\]).
3\. Variation of Stokes parameters with $s_{\pi}$: As shown in Fig. \[SivsE\]d, the parameters $\Sigma_1$ and $\Sigma_3$ have a variation with $s_{\pi}$ that is in close correspondence with the variation of $S_1$ and $S_3$ shown in Fig. \[SivsE\]. (Recall that the photon energy $E_{\gamma}$ in $K_L \rightarrow \pi^+ \pi^- \gamma$ can be expressed in terms of $s_{\pi}$: $s_{\pi} = {M_K}^2 - 2M_K E_{\gamma}$.) In particular the zero of $\Sigma_3$ and the zero of $S_3$ occur at almost the same value of $s_{\pi}$. This variation with $s_{\pi}$ combined with the low detector acceptance at large $s_{\pi}$, has the consequence of enhancing the measured asymmetry ($23.3 \pm 2.3 \%$ in KTeV [@KTeVcoll], $20 \pm 5 \%$ in NA48 [@NA48coll]).
4\. Generalized Angular Distribution: As shown in Ref. [@Heiliger:Sehgal], a more complete study of the angular distribution of the decay $K_L \rightarrow \pi^+ \pi^- e^+ e^-$ can yield further $CP$-violating observables, some of which are sensitive to the non-radiative (charge-radius and short-distance) parts of the matrix element. In particular the two-dimensional distribution $d\Gamma / d \! \cos \theta_l d \phi$ has the form $$\begin{aligned}
\frac{d\Gamma}{d\! \cos \theta_l d\phi} & = & K_1 + K_2 \cos 2 \theta_l +
K_3 \sin^2 \theta_l \cos 2 \phi + K_4 \sin 2 \theta_l \cos \phi \nonumber \\
& & \mbox{} + K_5 \sin \theta_l \cos \phi + K_6 \cos \theta_l + K_7
\sin \theta_l \sin \phi \nonumber \\
& & \mbox{} + K_8 \sin 2 \theta_l \sin \phi + K_9 \sin^2 \theta_l \sin 2 \phi.
\label{defKi}\end{aligned}$$ Considering the behaviour of $\cos \theta_l$, $\sin \theta_l$, $\cos \phi$ and $\sin \phi$ under $CP$ and $T$, the various terms appearing in Eq. (\[defKi\]) have the following transformation:
$CP$ $T$
---------------------------- ------ -----
$K_1$, $K_2$, $K_3$, $K_5$ + +
$K_4$, $K_6$ $-$ +
$K_8$ + $-$
$K_7$, $K_9$ $-$ $-$
Note that $K_{4,6,8}$ have $(CP)(T)=-$, a signal that they vanish in the hermitian limit. If only the bremsstrahlung and magnetic dipole terms are retained in the $K_L \rightarrow \pi^+ \pi^- e^+ e^-$ amplitude, one finds $K_4 = K_5 = K_6 = K_7 = K_8 = 0$, the only non-zero coefficients being $K_1 = 1$ (norm), $K_2 = 0.297$, $K_3 = 0.180$, $K_9 = -0.309$. In this notation, the asymmetry in $d\Gamma / d\phi$ is $\mathcal{A} =
\frac{2}{\pi} \frac{\frac{2}{3} K_9}{1-\frac{1}{3}K_2} = - 14\%$. The introduction of a charge-radius term induces a new $CP$-odd, $T$-even term $K_4 \approx -1.3 \%$, while a short-distance interaction containing an axial vector electron current can induce the $CP$-odd, $T$-odd term $K_7$. The standard model prediction for $K_7$, however, is extremely small.\
We conclude with a list of questions that could be addressed by future research. In connection with $K_{L,S} \rightarrow \pi^+ \pi^- \gamma$: (i) Is there a departure from bremsstrahlung in $K_S \rightarrow \pi^+ \pi^- \gamma$ (evidence for direct $E1$)? (ii) Is there a $\pi^+ / \pi^-$ asymmetry in $K_L \rightarrow \pi^+ \pi^- \gamma$ (evidence for $E2$)? (iii) Is there a measurable difference between $\eta_{+-\gamma}$ and $\eta_{+-}$ (existence of direct $CP$-violating $E1$ in $K_L \rightarrow \pi^+ \pi^- \gamma$)? With respect to the decay $K_L \rightarrow \pi^+ \pi^- e^+ e^-$: (i) Is there evidence of an $s$-wave amplitude? (ii) Is there evidence for $K_4$ or $K_7$ types of $CP$-violation? On the theoretical front: (i) Can one calculate the $s$-wave amplitude, and the form factors in $K_L \rightarrow \pi^+ \pi^- \gamma^*$ [@Savage]? (ii) Can one understand the sign of $g_{M1}$? (iii) Can one explain why direct $E1$ in $K_S \rightarrow \pi^+ \pi^- \gamma$ is so small compared to direct $M1$ in $K_L \rightarrow \pi^+ \pi^- \gamma$ ($\left| g_{E1}/g_{M1} \right| < 5 \%$)?
[99]{} E. J. Ramberg *et al.*, Fermilab Report No. Fermilab-Conf-91/258,1991; Phys. Rev. Lett. **70**, 2525 (1993). T. D. Lee and C. S. Wu, Annu. Rev. Nucl. Sci. **16**, 511 (1966); G. Costa and P. K. Kabir, Nuovo Cimento A **61**, 564 (1967); L. M. Sehgal and L. Wolfenstein, Phys. Rev. **162**, 1362 (1967). L. M. Sehgal, Phys. Rev. D **4**, 267 (1971). L. M. Sehgal and M. Wanninger, Phys. Rev. D **46**, 1035 (1992); *ibid.*. D **46**, 5209(E) (1992). P. Heiliger and L. M. Sehgal, Phys. Rev. D **48**, 4146 (1993). L. M. Sehgal and J. van Leusen, (in preparation). KTeV collaboration, A. Alavi-Harati *et al.*, hep-ex/9908020, submitted to Phys. Rev. Lett.; A. Ledovskoy, these Proceedings. NA48 collaboration, S. Wronka, these Proceedings. M. Savage, these Proceedings.
[^1]: Talk given at the KAON 99 Conference, Chicago, June 21-26, 1999
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the *Complex Stone-Weierstrass Property* (CSWP), which is the complex version of the Stone-Weierstrass Theorem. If $X$ is a compact subspace of a product of three linearly ordered spaces, then $X$ has the CSWP if and only if $X$ has no subspace homeomorphic to the Cantor set. In addition, every finite power of the double arrow space has the CSWP. These results are proved using some results about those compact Hausdorff spaces which have scattered-to-one maps onto compact metric spaces.'
author:
- 'Kenneth Kunen[^1] [^2]'
title: 'Ordered Spaces, Metric Preimages, and Function Algebras [^3] '
---
Introduction {#sec-intro}
============
All topologies discussed in this paper are assumed to be Hausdorff. As usual, a subset of a space is *perfect* iff it is closed and non-empty and has no isolated points, so $X$ is *scattered* iff $X$ has no perfect subsets.
The usual version of the *Stone-Weierstrass Theorem* involves subalgebras of $C(X,{{\mathbb R}})$, and is true for all compact $X$. If one replaces the real numbers ${{\mathbb R}}$ by the complex numbers ${{\mathbb C}}$, the “theorem” is true for some $X$ and false for others, so it becomes a *property* of $X$:
\[def-cswp\] If $X$ is compact, then $C(X) = C(X,{{\mathbb C}})$ is the algebra of continuous complex-valued functions on $X$, with the usual supremum norm. ${{\mathcal A}}{\sqsubseteq}C(X)$ means that ${{\mathcal A}}$ is a subalgebra of $C(X)$ which separates points and contains the constant functions. ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$ means that ${{\mathcal A}}{\sqsubseteq}C(X)$ and ${{\mathcal A}}$ is closed in $C(X)$. $X$ has the *Complex Stone-Weierstrass Property (CSWP)* iff every ${{\mathcal A}}{\sqsubseteq}C(X)$ is dense in $C(X)$; equivalently, iff every ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$ equals $C(X)$.
The CSWP is easily seen to be true for finite spaces. The complex analysis developed in the 1800s shows that the CSWP is false for many compact subspaces of the plane; for example, it is false for the unit circle ${{\mathbb T}}$; the classic counter-example being the algebra of complex polynomials ${{\mathcal P}}{\sqsubseteq}C({{\mathbb T}})$. These remarks are subsumed by results of W. Rudin [@RUD1; @RUD2] from the 1950s:
\[thm-rudin\] Let $X$ be any compact space.
If $X$ contains a copy of the Cantor set, then $X$ fails the CSWP.
If $X$ is scattered, then $X$ satisfies the CSWP.
If a compact space is metrizable (equivalently, second countable), then it contains a Cantor subset iff it is not scattered, so as Rudin pointed out:
\[cor-metric\] If $X$ is compact metric, then $X$ satisfies the CSWP iff $X$ does not contain a copy of the Cantor set.
One might conjecture that this corollary holds for all compact $X$, but that was refuted in 1960 by Hoffman and Singer [@HS] (see also [@GA; @HO]); their results imply that any compactum containing $\beta{{\mathbb N}}$ fails the CSWP.
However, the corollary does hold for some more “reasonable” classes of spaces. Kunen [@KU] showed in 2004:
\[thm-lots\] If $X$ is a compact LOTS, then $X$ satisfies the CSWP iff $X$ does not contain a copy of the Cantor set.
As usual, a LOTS is a linearly ordered topological space. Of course, the $\rightarrow$ of this result is clear from Theorem \[thm-rudin\]; only the $\leftarrow$ was new. This theorem shows that there are some non-scattered spaces with the CSWP, such as the double arrow space of Alexandroff and Urysohn (see Definition \[def-das\], or [@AU], p. 76).
One can now ask whether there are further classes of “*reasonable*” spaces for which results such as Corollary \[cor-metric\] and Theorem \[thm-lots\] hold. We do not know the best possible result along this line, but we shall prove in Section \[sec-cswp\]:
\[thm-three-lots\] If $X$ is compact and $X \subseteq L_0 \times L_1 \times L_2$, where $L_0,L_1,L_2$ are LOTSes, then $X$ has the CSWP iff $X$ does not contain a copy of the Cantor set.
Here, we may assume that $L_0,L_1,L_2$ are compact (otherwise, replace them by the projections of $X$). It is unknown whether the product of two spaces with the CSWP must also have the CSWP. Even if this turns out to be true, Theorem \[thm-three-lots\] is not immediate from Theorem \[thm-lots\], since $X$ is an arbitrary compact subset of the product, and $L_0,L_1,L_2$ may fail the CSWP (i.e., have Cantor subsets).
By a slightly different argument, we shall show in Section \[sec-powers\]:
\[thm-das-power\] If $L$ is the double arrow space, then $L^n$ has the CSWP for every finite $n$.
Theorems \[thm-three-lots\] and \[thm-das-power\] are proved using some results from Section \[sec-tight\] about spaces which have scattered-to-one maps onto metric spaces. In Theorem \[thm-das-power\], there is a natural $f : L^n {\twoheadrightarrow}[0,1]^n$ for which the inverse of each point is scattered (and of size $2^n$). In Theorem \[thm-three-lots\], the $L_j$ need not have any scattered-to-one maps onto metric spaces, but a standard argument using measures reduces the proof of Theorem \[thm-three-lots\] to the case where the $L_j$ are separable (see Section \[sec-cswp-reduce\]), in which case $X$ must have an eight-to-one map onto a compact metric space.
If $L_0,L_1,L_2$ are separable in Theorem \[thm-three-lots\], then $X$ must also be first countable, and hence “small” in the cardinal functions sense (see Juhász [@JU]). However, we do not believe that there is a notion of “reasonable” involving only cardinal functions. In [@HK2] it is shown that in some models of set theory, there is a compact $X$ which does not contain Cantor subsets and which fails the CSWP, such that $X$ is both hereditarily separable and hereditarily Lindelöf (and hence also first countable). In these models, $2^{\aleph_0} = \aleph_1$ and the standard cardinal functions of our $X$ (all either $\aleph_0$ or $\aleph_1$) are the least possible among non-metric compacta.
Section \[sec-LOTS\] reviews some elementary fact about LOTSes. Section \[sec-remov\] discusses the notion of a *removable space* defined in [@HK1]; this is a strengthening of the CSWP used in Section \[sec-powers\].
\[def-local\] Let ${{\mathfrak K}}$ be a class of compact spaces. ${{\mathfrak K}}$ is *closed-hereditary* iff every closed subspace of a space in ${{\mathfrak K}}$ is also in ${{\mathfrak K}}$. ${{\mathfrak K}}$ is *local* iff ${{\mathfrak K}}$ is closed-hereditary *and* for every compact $X$: if $X$ is covered by open sets whose closures lie in ${{\mathfrak K}}$, then $X \in {{\mathfrak K}}$.
Classes of compacta which restrict cardinal functions (first countable, second countable, countable tightness, etc.) are clearly local, whereas the class of compacta which are homeomorphic to a LOTS is closed-hereditary, but not local.
It is easily seen that the CSWP is closed-hereditary; this is Lemma 1.3 of [@KU], but the proof is implicit in Rudin [@RUD1]. Thus, to prove part (1) of Theorem \[thm-rudin\] in [@RUD1], it was sufficient to show that the Cantor set itself fails the CSWP.
The removable spaces form a local class (see Section \[sec-remov\]). It is unknown whether the CSWP is a local property. A proof that it is local cannot be completely trivial. For example, locality would imply that the failure of the CSWP for ${{\mathbb T}}$ yields the failure of the CSWP for an arc $A \subseteq {{\mathbb T}}$. Now, $A$ does in fact fail the CSWP, since it contains a Cantor set, but we do not know how to construct a counter-example on $A$ directly from the polynomial algebra ${{\mathcal P}}{\sqsubseteq}C({{\mathbb T}})$; note that the restriction ${{\mathcal P}}{\mathord {\upharpoonright}}A {\sqsubseteq}C(A)$ is dense in $C(A)$ by Mergelyan’s Theorem.
Ordered Spaces {#sec-LOTS}
==============
We begin by defining the double arrow space and some variants thereof:
\[def-das\] $I = [0,1]$. If $\Lambda : I \to \omega$, then $I_\Lambda = \bigcup_{x \in I} \{x\} \times \{0,1,\ldots, \Lambda(x)\}$, which is given the lexicographic order and the usual order topology. If $S \subseteq (0,1)$, then $I_S = I_{{{\raise 2 pt \hbox{$\chi$}}}_S}$, where ${{\raise 2 pt \hbox{$\chi$}}}_S$ is the characteristic function; then for $x \in S$, let $x^- = (x,0)$ and $x^+ = (x,1)$; while if $x \notin S$, let $x^- = x^+ = (x,0)$. The *double arrow space* is $I_{(0,1)}$. For any $\Lambda$, the map $(x, \ell) \mapsto x$ is the *standard map* from $I_\Lambda$ onto $I$.
So, we form $I_\Lambda$ by splitting each $x \in S$ into $\Lambda(x) + 1$ neighboring points. For $I_S$, we split each $x \in S$ into two neighboring points, $x^-,x^+$, and we don’t split the points in $I \backslash S$; it is convenient to have $x^\pm$ defined for *all* $x \in I$, so, for example, we can say that for all $a < b$ in $I$, $(a^+, b^-)$ is an open interval in $I_S$. $I_S$ has no isolated points because $0,1\notin S$. The double arrow space is obtained by splitting all points other than $0,1$. $I_\emptyset \cong I$, and $I_{{{\mathbb Q}}\cap (0,1)}$ is homeomorphic to the Cantor set.
For each $S \subseteq (0,1)$, $I_S$ is a compact separable LOTS with no isolated points. $I_S$ is second countable iff $S$ is countable. Every $I_\Lambda$ is a compact first countable LOTS.
$I_\Lambda$ will not be separable unless $\{x : \Lambda(x) > 1\}$ is countable. The study of compact separable LOTSes can be reduced to spaces of the form $I_S$. First note, by Lutzer and Bennett [@LB]:
\[lemma-hs\] If $X$ is a separable LOTS, then $X$ is hereditarily separable and hereditarily Lindelöf.
Also, it is easy to check:
\[lemma-rel\] If $X$ is a LOTS and $H$ is a compact subset of $X$, then the relative topology and the order topology agree on $H$.
Relating this to our $I_S$:
\[lemma-LOTS-standard\] Let $X$ be a compact separable LOTS. Then
- If $X$ is perfect, then $X$ is homeomorphic to $I_S$ for some $S \subseteq (0,1)$.
- If $X$ is not second countable, then $X$ has a closed subspace which is homeomorphic to $I_S$ for some uncountable $S \subseteq (0,1)$.
- $X$ is homeomorphic to a subset of $I_S$ for some $S \subseteq (0,1)$.
For (1): Let $E \subseteq X$ be countable and dense in $X$ and contain the first and last elements of $X$. Let $B$ be the set of all $b \in E$ such that for some $a \in E$: $a < b$ and $(a,b) = \emptyset$. Let $D = E \backslash B$. Since $X$ has no isolated points, $D$ is also dense in $X$ and contains the first and last elements of $X$, and is also densely ordered. Let $f$ be an order isomorphism from $D$ onto ${{\mathbb Q}}\cap [0,1]$. Then $f$ extends in a natural way to a continuous $F: X {\twoheadrightarrow}[0,1]$, and $1 \le |F{^{-1}}\{r\}| \le 2$ for each $r \in [0,1]$. Let $S = \{r : |F{^{-1}}\{r\}| = 2\}$.
For (2): Since $X$ is hereditarily Lindelöf, the Cantor-Bendixson sequence of $X$ has countable length and removes countably many points. Thus, $X$ is not scattered, and, letting $H$ be the perfect kernel of $X$, $X \backslash H$ is countable. Then $H$ is separable and not second countable, so $H \cong I_S$ for some uncountable $S$.
For (3): Apply (1) to the space obtained from $X$ by replacing each isolated point by a copy of the double arrow space.
Note that $(I_S)^2$ is separable, but it is not hereditarily separable when $S$ is uncountable; in fact, more general $I_\Lambda$ occur naturally in such products. Fixing an uncountable $S \subseteq (0,1)$, let $L_n = I_{\Lambda_n}$, where $\Lambda_n(x) = n$ for $n \in S$, and $\Lambda_n(x) = 0$ for $n \notin S$. Then $L_n$ is not separable whenever $n \ge 2$, and the diagonal of $(I_S)^k$ is homeomorphic to $L_{2^k -1}$.
Tight Maps and Dissipated Spaces {#sec-tight}
================================
We recall some definitions and results from [@KU2]. As usual, $f : X \to Y$ means that $f$ is a *continuous* map from $X$ to $Y$, and $f : X {\twoheadrightarrow}Y$ means that $f$ is a continuous map from $X$ *onto* $Y$.
\[def-tight\] Assume that $X,Y$ are compact and $f: X \to Y$.
[43]{}
A *loose family* for $f$ is a disjoint family ${{\mathcal P}}$ of closed subsets of $X$ such that for some non-scattered $Q \subseteq Y$, $Q = f(P)$ for all $P \in {{\mathcal P}}$.
$f$ is $\kappa$–*tight* iff there are no loose families for $f$ of size $\kappa$.
$f$ is *tight* iff $f$ is $2$–tight.
This notion gets weaker as $\kappa$ gets bigger. $f$ is $1$–tight iff $f(X)$ is scattered, so that “$2$–tight” is the first non-trivial case. $f$ is trivially $|X|^+$–tight. The usual projection from $[0,1]^2$ onto $[0,1]$ is not $2^{\aleph_0}$–tight.
Some easy equivalents to “$\kappa$–tight” are described in Lemma 2.2 of [@KU2]:
\[lemma-tight-equiv\] Assume that $X,Y$ are compact and $f : X \to Y$. Then $(1) \leftrightarrow (2)$. If $\kappa$ is finite and $Y$ is metric, then all four of the following are equivalent:
There is a loose family of size $\kappa$.
There is a disjoint family ${{\mathcal P}}$ of perfect subsets of $X$ with $|{{\mathcal P}}| = \kappa$ and a perfect $Q \subseteq Y$ such that $Q = f(P)$ for all $P \in {{\mathcal P}}$.
For some metric $M$ and $\varphi\in C(X,M)$, $\{y \in Y : |\varphi(f{^{-1}}\{y\})| \ge \kappa\}$ is uncountable.
Statement $(3)$, with $M = [0,1]$.
If $X,Y$ are both compact metric, then $f: X \to Y$ is $\kappa$–tight iff $\{y \in Y : |f{^{-1}}\{y\}| \ge \kappa\}$ is countable (see Theorem 2.7 of [@KU2]). Of course, the $\leftarrow$ direction is trivial. The $\rightarrow$ direction for non-metric $X$ and $\kappa = 2$ is refuted by the standard map from the double arrow space onto $[0,1]$, which is tight by Lemma 2.3 of [@KU2]:
\[lemma-tight-LOTS\] If $X,Y$ are compact LOTSes and $f:X \to Y$ is order-preserving $(x_1 < x_2 \to f(x_1) \le f(x_2))$, then $f$ is tight.
One can estimate the tightness of product maps using Lemma 2.14 of [@KU2]:
\[lemma-tight-prod\] Assume that for $i = 0,1$: $X_i, Y_i$ are compact, $f_i:X_i \to Y_i$ is $(m_i + 1)$–tight, $m_i \le n_i < \omega$, and $|f_i{^{-1}}\{y\}| \le n_i$ for *all* $y \in Y_i$. Then $f_0 \times f_1 : X_0 \times X_1 \to Y_0 \times Y_1$ is $(\max(m_0 n_1 , m_1 n_0) + 1)$– tight.
The notion of a *dissipated* compactum (Definition \[def-dis\] below) involves tight maps onto metric compacta, ordered by *fineness*, so we define:
\[def-fine\] Assume that $X,Y,Z$ are compact, $f : X \to Y$, and $g : X \to Z$. Then $f \le g$, or $f$ is *finer than* $g$, iff there is a $\Gamma \in C(f(X),g(X))$ such that $g = \Gamma \circ f$.
\[lemma-fine-equiv\] Assume that $X,Y,Z$ are compact, $f : X \to Y$, and $g : X \to Z$. Then $f \le g$ iff $\; \forall x_1, x_2 \in X\, [ f(x_1) = f(x_2) \to g(x_1) = g(x_2)]$.
Assume $X$ is compact. Let ${{\mathfrak M}}(X)$, the *metric projections* of $X$, be the class of all maps $\pi$ such that $\pi : X \to Y$ for some compact metric $Y$. Then $\pi \in {{\mathfrak M}_{\mathfrak S}}(X) \subseteq {{\mathfrak M}}(X)$ iff in addition, each $\pi{^{-1}}\{y\}$ is scattered.
\[lemma-MS-down\] If $\pi,\sigma\in{{\mathfrak M}}(X)$ and $\pi \le \sigma \in {{\mathfrak M}_{\mathfrak S}}(X)$, then $\pi \in {{\mathfrak M}_{\mathfrak S}}(X)$.
Observe that in the definition of $f \le g$, it is irrelevant whether $f,g$ map $X$ *onto* $Y,Z$. Here, and in the definition of ${{\mathfrak M}}(X)$, we should really regard $f$ in the set-theoretic sense as a set of ordered pairs, not as a triple $(f,X,Y)$, so that $f: X \to Y$ and $f: X {\twoheadrightarrow}f(X)$ are exactly the same object. One could also define ${{\mathfrak M}}(X)$ and ${{\mathfrak M}_{\mathfrak S}}(X)$ as sets of closed equivalence relations on $X$.
\[lemma-count\] ${{\mathfrak M}}(X)$ is countably directed. That is, if $\sigma_n \in {{\mathfrak M}}(X)$ for $n \in \omega$, then there is a $\pi \in {{\mathfrak M}}(X)$ with $\pi \le \sigma_n$ for each $n$.
\[lemma-sing\] If $\sigma \in {{\mathfrak M}_{\mathfrak S}}(X)$, then there is a $\pi \in {{\mathfrak M}_{\mathfrak S}}(X)$ with $\pi \le \sigma$ and $\pi : X {\twoheadrightarrow}Y$, such that $\pi{^{-1}}\{b\}$ is a singleton for some $b\in Y$.
Say $\sigma : X {\twoheadrightarrow}Z$. Fix any $c \in Z$, and then fix $a \in \sigma{^{-1}}\{c\}$ such that $a$ is isolated in $\sigma{^{-1}}\{c\}$. Since $Z$ is metric, $\{a\}$ is a $G_\delta$ in $X$, so fix any $f\in C(X, [0,1])$ with $\{a\} = f{^{-1}}\{1\}$. Choose $\pi\in{{\mathfrak M}_{\mathfrak S}}(X)$ with $\pi \le \sigma$ and $\pi\le f$.
Only a scattered compactum $X$ has the property that *all* maps in ${{\mathfrak M}}(X)$ are tight: If $X$ is not scattered, then $X$ maps onto $[0,1]^2$; if we follow that map by the usual projection onto $[0,1]$, we get a map from $X$ onto $[0,1]$ which is not even ${{\mathfrak c}}$–tight. The *dissipated* compacta have the property that *cofinally* many of these maps are tight:
\[def-dis\] $X$ is *$\kappa$–dissipated* iff $X$ is compact and whenever $g \in {{\mathfrak M}}(X)$, there is a finer $\kappa$–tight $f \in {{\mathfrak M}}(X)$. $X$ is *dissipated* iff $X$ is $2$–dissipated.
So, the $1$–dissipated compacta are the scattered compacta. Metric compacta are dissipated because we can let $f$ be identity map. By Lemma 3.12 of [@KU2]:
\[lemma-local\] For any $\kappa$, the class of $\kappa$–dissipated compacta is a local class.
An easy example of a dissipated space is given by:
\[lemma-lots-dis\] If $X$ is a compact LOTS, then $X$ is dissipated
The proof (see Lemma 3.4 of [@KU2]) shows that given $g \in {{\mathfrak M}}(X)$, there is a finer $f \in {{\mathfrak M}}(X)$ such that $f(X)$ is a compact metric LOTS and $f$ is order-preserving.
Note that just having *one* tight map $g$ from $X$ onto some metric compactum $Z$ is not sufficient to prove that $X$ is dissipated, since the tightness of $g$ says nothing at all about the complexity of a particular $g{^{-1}}\{z\}$. However, if all $g{^{-1}}\{z\}$ are scattered, then just one tight $g$ is enough by Lemma 3.5 of [@KU2]:
\[lemma-one-enough\] Assume that some $g \in {{\mathfrak M}_{\mathfrak S}}(X)$ is $\kappa$–tight. Then all $f \le g$ are also $\kappa$–tight, so that $X$ is $\kappa$–dissipated.
This suggests the following definition:
\[def-super\] $\pi\in{{\mathfrak M}}(X)$ is $\kappa$–*supertight* iff $\pi$ is $\kappa$–tight and $\pi\in{{\mathfrak M}_{\mathfrak S}}(X)$. Then $X$ is $\kappa$–*superdissipated* iff some $\pi\in{{\mathfrak M}_{\mathfrak S}}(X)$ is $\kappa$–supertight.
Using Lemmas \[lemma-one-enough\], \[lemma-local\], and \[lemma-MS-down\] above:
\[lemma-lim-finer\] If $\pi, \sigma \in{{\mathfrak M}_{\mathfrak S}}(X)$, $\pi \le \sigma$, and $\sigma$ is $\kappa$–supertight, then $\pi$ is $\kappa$–supertight.
A compactum $X$ is $\kappa$–superdissipated iff $X$ is $\kappa$–dissipated and ${{\mathfrak M}_{\mathfrak S}}(X) \ne \emptyset$.
\[lemma-clh\] The class of $\kappa$–superdissipated compacta is a local class.
By Lemma \[lemma-tight-LOTS\]:
The standard map $\sigma : I_\Lambda {\twoheadrightarrow}I$ is $2$–supertight.
The situation for products is more complicated. By Lemma \[lemma-tight-prod\] and induction:
\[lemma-big-lim\] For any $n \ge 1$ and $S_i \subseteq I $ (for $i < n$): The standard map $\sigma : \prod_{i <n} I_{S_i} {\twoheadrightarrow}I^n$ is $(2^{n-1} + 1)$–supertight.
This result is best possible by Theorem 3.9 of [@KU2]; a product $\prod_{i< n}X_i$ is not $(2^{n-1} )$–dissipated if each $X_i$ is a compact separable LOTS, none of the $X_i$ is scattered, and at most one of the $X_i$ is second countable.
\[def-kernel\] The *perfect kernel*, $\ker(X)$, is $\emptyset$ if $X$ is scattered, and the largest perfect subset of $X$ otherwise.
By Lemma \[lemma-tight-equiv\], the tightness of $\pi : X \to Y$ can be expressed using perfect subsets of $X$, so that
\[lemma-ker-tight\] $\pi : X \to Y$ is $\kappa$–(super)tight iff $\pi{\mathord {\upharpoonright}}\ker(X)$ is $\kappa$–(super)tight, and the space $X$ is $\kappa$–(super)dissipated iff $\ker(X)$ is $\kappa$–(super)dissipated.
\[lemma-get-one\] Assume that $\pi: X \to Y$ is $(n+2)$–supertight, where $n \in \omega$, $X$ is compact and $Y$ is compact metric, and $\{P_0, \ldots, P_n\}$ is a loose family for $\pi$ of size $n+1$, with each $\pi(P_j) = Q$. Then each $\pi{\mathord {\upharpoonright}}P_j : P_j \to Q$ is $2$–supertight, $\ker (\pi{^{-1}}(Q)) \subseteq \bigcup_j P_j$, and $\pi{^{-1}}(Q)$ is $2$–superdissipated.
If tightness fails for $\pi{\mathord {\upharpoonright}}P_j$, then we could find uncountable closed $Q' \subseteq Q$ and disjoint closed $P_j^0, P_j^1 \subseteq P_j$ with $\pi(P_j^0) = \pi(P_j^1) = Q'$. If $P'_k = P_k \cap \pi{^{-1}}(Q')$, then the sets $P'_0, \ldots, P'_{j-1}, P_j^0, P_j^1, P'_{j+1}, \ldots, P'_n$ would be a loose family for $\pi$ of size $n+2$. If $\ker (\pi{^{-1}}(Q)) \not\subseteq \bigcup_j P_j$, we could find a perfect $R \subseteq \pi{^{-1}}(Q) \setminus \bigcup_j P_j$; then $\pi(R)$ is non-scattered (since all $\pi{^{-1}}\{y\}$ are scattered), and $R$ plus the $P_j$ would contradict the $(n+2)$–tightness of $\pi$. Finally, by Lemma \[lemma-ker-tight\], it is sufficient to prove that $\bigcup_j P_j$ is superdissipated, and this is done using the map into $Y \times \{0,1,\ldots,n\}$ which sends $x \in P_j$ to $(\pi(x),j)$.
Finally, we mention two lemmas for the case that $X$ does not contain a Cantor subset. $\pi: X \to Y$ is trivially $n$–supertight when all $|\pi{^{-1}}\{y\}| < n$, but also
\[lemma-lim-prop\] Assume that $\pi: X \to Y$, $X$ is compact, $Y$ is compact metric, each $|\pi{^{-1}}\{y\}| \le n$, and $X$ has no Cantor subsets. Then $\pi$ is $n$–supertight.
If not, let $P_0, \ldots, P_{n-1} \subseteq X $ be a loose family, with each $\pi(P_j) = Q$. Then $Q$ has a Cantor subset, and each $P_j$ is homeomorphic (via $\pi$) to $Q$.
By the next lemma, the spaces $X$ we consider are always totally disconnected:
\[lemma-discon\] Assume that $\pi : X \to Y$, where $X$ is compact and $Y$ is metric. Assume that each $\pi{^{-1}}\{y\}$ is totally disconnected and $X$ does not contain a copy of the Cantor set. Then $X$ is totally disconnected.
Assume that $X$ is not totally disconnected. Fix a metric on $Y$ for which ${\mathrm{diam}}(Y) \le 1$. Obtain $K_s$ for $s \in 2^{<\omega}$ to satisfy:
$K_s$ is an infinite closed connected subset of $X$.
${\mathrm{diam}}(\pi(K_s)) \le 2^{- {\mathrm{lh}}(s)}$.
$K_{s{^{\mathord{\frown}}}0}, K_{s{^{\mathord{\frown}}}1} \subset K_s$ and $K_{s{^{\mathord{\frown}}}0} \cap K_{s{^{\mathord{\frown}}}1} = \emptyset$.
Assuming that this can be done, define $K_f = \bigcap_{n \in \omega} K_{f{\mathord {\upharpoonright}}n}$ for $f \in 2^\omega$. By (2), $|\pi(K_f)| = 1$; say $\pi(K_f) = \{y_f\}$. But $K_f$ is connected and $\pi{^{-1}}\{y_f\}$ is totally disconnected, so $|K_f| = 1$; say $K_f = \{x_f\}$. Then $f \mapsto x_f$ is a homeomorphism from $2^\omega$ into $X$, contradicting our assumptions about $X$.
To build the $K_s$: For $K_{(\,)}$, just use the assumption that $X$ is not totally disconnected. Now, say we are given $K_s$. Choose $x_0,x_1\in K_s$ with $x_0 \ne x_1$. Then find disjoint relatively open $U_0,U_1 \subseteq K_s$ with each $x_\ell \in U_\ell$ and ${\mathrm{diam}}(\pi(U_\ell)) \le 2^{- {\mathrm{lh}}(s) - 1}$. Then find relatively open $V_\ell \subseteq K_s$ with $x_\ell \in V_\ell \subseteq \overline{V_\ell}\subseteq U_\ell$. Then, let $K_{s{^{\mathord{\frown}}}\ell}$ be the connected component of the point $x_\ell$ in the space $\overline{V_\ell}$, and note that $K_{s{^{\mathord{\frown}}}\ell}$ cannot be a singleton.
The CSWP: Two Reductions {#sec-cswp-reduce}
========================
These reductions were described in [@KU]: Using the standard theory of function algebras (see [@GAM; @GA]), we can reduce the CSWP to the study of idempotents, and we can reduce the study of the CSWP in LOTSes to the separable case.
If $f \in C(X)$, then $f$ is an *idempotent* iff $f^2 = f$; equivalently iff $f$ is the characteristic function of some clopen set. An idempotent is called *nontrivial* iff it is not the identically $0$ or the identically $1$ function. As with other proofs of the CSWP [@HK1; @KU], we shall proceed by considering idempotents. Following [@KU],
The compact space $X$ has the *NTIP* iff every ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$ contains a non-trivial idempotent.
So, the NTIP is trivially false of connected spaces. If $X$ is not connected, then the CSWP implies the NTIP. The following is Lemma 3.5 of [@KU]; it is also easy to prove from the Bishop Antisymmetric Decomposition (see [@BIS], or Theorem 13.1 in Chapter II of [@GAM]).
\[lemma-restrict-to-perfect\] Assume that $X$ is compact and every perfect subset of $X$ has the NTIP. Then $X$ has the CSWP.
Among the totally disconnected spaces, the NTIP is strictly weaker than the CSWP (see [@KU]). However, the lemma implies the following corollary, which is used to reduce proofs of the CSWP to proofs of the weaker NTIP:
\[cor-idem\] If ${{\mathfrak K}}$ is a closed-hereditary class of compact spaces and every perfect space in ${{\mathfrak K}}$ has the NTIP, then every space in ${{\mathfrak K}}$ has the CSWP.
In particular, if ${{\mathfrak K}}$ is the class of compact scattered spaces, then this corollary applies vacuously, so all spaces in ${{\mathfrak K}}$ have the CSWP. If ${{\mathfrak K}}$ contains some non-scattered spaces, then, as in [@KU; @HK1], we produce idempotents using:
\[lemma-get-clopen\] Suppose that ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$ and there is some $h \in {{\mathcal A}}$ such that either $\Re(h(X))$ or $\Im(h(X))$ is not connected. Then ${{\mathcal A}}$ contains a non-trivial idempotent.
This is easy to prove using Runge’s Theorem; see Lemma 2.5 of [@KU], but the method was also used in [@RUD1] and [@HS].
It remains to describe how to obtain such an $h$. If $X$ is scattered, then $\Re(h(X))$ is scattered also, so any $h$ for which $\Re(h(X))$ is not a singleton will do; this is essentially the argument of [@RUD2]. In some other cases, we can obtain $h$ using a tight map of $X$ onto a metric space; this is described in Section \[sec-cswp\].
We now turn to the second reduction. As in §5 of [@KU],
If $\mu$ is a regular complex Borel measure on the compact space $X$, then $|\mu|$ denotes its total variation, and ${\mathrm{supt}}(\mu) = {\mathrm{supt}}(|\mu|)$ denotes its (closed) support; that is, ${\mathrm{supt}}(\mu) = X \setminus \bigcup \{U\subseteq X :\; $U$ \mbox{ is open }
\ \ \& \ \ |\mu|(U) = 0\}$.
Considering measure orthogonal to ${{\mathcal A}}$, we get:
\[lemma-supt-cswp\] Assume that $X$ is compact and that ${\mathrm{supt}}(\mu)$ has the CSWP for all regular Borel measure $\mu$. Then $X$ has the CSWP.
By Corollary 5.4 of [@KU2], every such ${\mathrm{supt}}(\mu)$ is separable in the case that $X$ is $\aleph_0$–dissipated; for a LOTS $X$, this was a much earlier folklore result.
\[cor-CSWP-sep\] If $X$ fails the CSWP and is $\aleph_0$–dissipated, then some compact separable subspace of $X$ fails the CSWP.
Is there a compact space $X$ which fails the CSWP such that all compact separable subspaces of $X$ satisfy the CSWP?
This $X$ cannot be one of the three examples already known to fail the CSWP — namely, any space containing either the Cantor set [@RUD1] or $\beta{{\mathbb N}}$ [@HS] or the examples of [@HK2; @HK3] (obtained assuming $\diamondsuit$ or CH), since all these spaces are separable.
Now, considering products of LOTSes:
\[lemma-prod-cswp\] Assume that $X$ is a compact subset of $\prod_{\alpha < \kappa} L_\alpha$, where each $L_\alpha$ is a LOTS, and assume that $X$ does not have the CSWP. Then for some separable closed compact $H_\alpha \subseteq L_\alpha$, the space $X \cap \prod_{\alpha < \kappa} H_\alpha$ also fails the CSWP.
Let $\pi_\alpha : X \to L_\alpha$ be the usual coordinate projection. We may assume that each $L_\alpha = \pi_\alpha(X)$, so that $L_\alpha$ is compact. Fix $\mu$ on $X$ such that ${\mathrm{supt}}(\mu)$ fails the CSWP, let $\mu \pi_\alpha{^{-1}}$ be the induced measure on $L_\alpha$, let $H_\alpha = {\mathrm{supt}}(\mu \pi_\alpha{^{-1}})$, which is separable, and note that ${\mathrm{supt}}(\mu) \subseteq X \cap \prod_{\alpha < \kappa} H_\alpha$.
\[lemma-one-S\] For any $\kappa \le \omega$: Suppose that there is a compact $X \subseteq \prod_{\alpha < \kappa} L_\alpha$, where each $L_\alpha$ is a LOTS, $X$ has no Cantor subset, and $X$ does not have the CSWP. Then there is such an $X$ which is a subset of $(I_S)^\kappa$ for some $S \subseteq (0,1)$.
By Lemma \[lemma-prod-cswp\], we may assume that each $L_\alpha$ is separable and compact. Now, let $L$ be the compact separable LOTS obtained by placing the $L_\alpha$ end-to-end, adding a point $\infty$ in the case that $\kappa = \omega$. Then we may assume that $X \subseteq L^\kappa$. Finally, replace $L$ by an $I_S$ using Lemma \[lemma-LOTS-standard\](3).
The CSWP and Tightness {#sec-cswp}
======================
We show here how one can use the concepts from Section \[sec-tight\] to produce idempotents, and thus to prove the CSWP.
\[def-phat\] Assume that $\pi: X \to Y$, where $X,Y$ are compact. Then, for $f \in C(X)$, define $\widehat f = (\pi\times f)(X)$; that is, $$\widehat f \ =\ \{(\pi(x), f(x)) : x \in X \}
\ \subseteq\ Y \times {{\mathbb C}}\ \ .$$
Each $\widehat f$ is compact.
We plan to apply the next definition and lemma to sets of the form $\widehat f$:
Fix $E \subseteq Y \times {{\mathbb C}}$ and $\Phi: {{\mathbb C}}^m \to {{\mathbb C}}$. Then $E_y = \{z : (y,z) \in E\}$ and $$\Phi * E \ =\ \bigcup_{y \in Y} \Phi( (E_y)^m ) \ \subseteq\ {{\mathbb C}}\ \ .$$
\[lemma-open-cover\] Suppose that $F \subseteq Y \times {{\mathbb C}}$ is compact and $\Phi: {{\mathbb C}}^m \to {{\mathbb C}}$ is continuous. Let ${{\mathcal B}}$ be an open base for $Y \times {{\mathbb C}}$ which is closed under finite unions. Then $\Phi * F$ is compact and $\Phi * F = \bigcap\{ \Phi * \overline U : U \in {{\mathcal B}}\ \&\ F \subseteq U \}$.
\[lemma-scat-phi\] Assume that $\pi: X \to Y$ is $n$–supertight and $f \in C(X)$. Fix a continuous $\Phi : {{\mathbb C}}^{n} \to {{\mathbb C}}$ such that $\Phi(z_1, \ldots, z_n) = 0$ unless all $n$ of the $z_1, \ldots, z_n$ are different. Then $\Phi * \widehat f$ is compact and countable, and hence scattered.
Compactness follows from the compactness of $X,Y$. By $n$–tightness, $| \widehat f _y | < n$, and hence $\Phi(( \widehat f _y )^{n}) = \{0\}$, for all but countably many $y$ (see Lemma \[lemma-tight-equiv\]). But for all $y$, $\pi{^{-1}}\{y\}$ is scattered, so that $\Phi( ( \widehat f _y )^{n})$ is also scattered, and hence countable. Thus, the union of all these sets is also countable.
Dissipation is a notion of smallness, which is balanced by a notion of bigness, which is really a partition property:
\[def-big\] Fix a real $r > 0$. The compact space $X$ is $n$–*big* iff for all ${{\mathcal A}}{\sqsubseteq}C(X)$ and all partitions $\Upsilon : {{\mathcal A}}\to \omega$, there are $f_1, \ldots, f_n \in {{\mathcal A}}$ and a point $c \in X$ such that the $\Upsilon(f_j)$, for $j = 1, \ldots, n$, are all equal, and such that $|f_i(c) - f_j(c)| \ge r$ whenever $1 \le i < j \le n$.
Since ${{\mathcal A}}$ is a linear subspace, it does not matter which $r > 0$ we use. The notion of $1$–big is trivial, and $2$–big is easily characterized:
\[lemma-two-big\] The compact space $X$ is $2$–big iff $X$ is not second countable.
Note that $\exists c \,\big(|f_1(c) - f_2(c)| \ge r\big)$ holds iff $\|f_1 - f_2\| \ge r$. Also, if $X$ is not second countable then $C(X)$ is not separable, and hence any ${{\mathcal A}}{\sqsubseteq}C(X)$ is not separable, since the algebra generated by the functions in ${{\mathcal A}}$ and their complex conjugates is dense in $C(X)$ by the Stone-Weierstrass Theorem.
We relate this to the NTIP with the aid of:
\[def-S\] For each $n \ge 2$, define $\Xi_n : {{\mathbb C}}^{n} \to {{\mathbb C}}$ by: $$\Xi_n(z_1, \ldots, z_n) = 2 \cdot \prod_{1 \le i < j \le n} (z_j - z_i) \ \ .$$
$\Xi_n$ is a polynomial in $n$ variables. $\Xi(z_1, \ldots, z_n) = 0$ unless all $n$ of the $z_1, \ldots, z_n$ are different. If $|z_i - z_j| \ge 1$ for all $i < j \le n$, then either $|\Re( \Xi_n(z_1, \ldots, z_n)) | \ge 1$ or $|\Im( \Xi_n(z_1, \ldots, z_n)) | \ge 1$.
\[lemma-big-dis\] Assume that $X$ is compact, ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$, $H \subseteq V \subseteq X$, where $H,V$ are both clopen, and for some $n \ge 2$, $V$ is $n$–superdissipated and $H$ is $n$–big. Assume also that there is a $\psi \in {{\mathcal A}}$ such that $|\psi(x)| \le 1/2$ for all $x \in X \backslash V$ and $|\psi(x)| \ge 1$ for all $x \in H$. Then ${{\mathcal A}}$ has a non-trivial idempotent.
Fix $\pi : V {\twoheadrightarrow}Y$ which is $n$–supertight. Applying Lemmas \[lemma-sing\] and \[lemma-lim-finer\], we assume also that we have $b\in Y$ and $a \in V$ such that $\pi{^{-1}}\{b\} = \{a\}$. Let ${\mbox{\sc{r}}}^+( z_1, \ldots, z_n ) = -{\mbox{\sc{r}}}^-( z_1, \ldots, z_n ) =
\Re(\Xi_n( z_1, \ldots, z_n ))$ and ${\mbox{\sc{i}}}^+( z_1, \ldots, z_n ) = -{\mbox{\sc{i}}}^-( z_1, \ldots, z_n ) =
\Im(\Xi_n( z_1, \ldots, z_n ))$, so that ${\mbox{\sc{r}}}^+, {\mbox{\sc{r}}}^-, {\mbox{\sc{i}}}^+, {\mbox{\sc{i}}}^- : {{\mathbb C}}^{n} \to {{\mathbb R}}$. Call $(E, \rho, \tau)$ *good* iff:
$\rho, \tau \in {{\mathbb Q}}$ and $1/2 < \rho < \tau < 1$.
$E \subseteq Y \times {{\mathbb C}}$.
$[\rho,\tau]$ is disjoint from each of ${\mbox{\sc{r}}}^+ * E$, ${\mbox{\sc{r}}}^- * E$, ${\mbox{\sc{i}}}^+ * E$, ${\mbox{\sc{i}}}^- * E$.
$|\Xi_n( z_1, \ldots, z_n )| < \rho$ whenever $z_1, \ldots, z_n\in E_b$.
For $f\in C(X)$, use $\widehat f$ for $\widehat{\,f {\mathord {\upharpoonright}}V\,}$. Observe that for each $f\in C(X)$, we may choose $\rho,\tau$ so that $(\widehat f, \rho, \tau)$ is good: (4) is no problem since $\widehat f _b$ is a singleton. For the rest, note that each of ${\mbox{\sc{r}}}^+ * \widehat f$, ${\mbox{\sc{r}}}^- * \widehat f$, ${\mbox{\sc{i}}}^+ * \widehat f$, ${\mbox{\sc{i}}}^- * \widehat f$ is scattered by Lemma \[lemma-scat-phi\], so we may choose $\rho,\tau$ to make (1)(3) true.
Let ${{\mathcal B}}$ be a countable open base for $Y \times {{\mathbb C}}$ which is closed under finite unions. For each $f \in {{\mathcal A}}$, choose $s = s_f \in \omega$ so that $|(\psi(x))^s \, f(x)| \le 1/8$ for all $x\in X \backslash V$. Then, choose $\rho_f, \tau_f$ so that $(\widehat{\psi^{s_f} f\,}, \rho_f, \tau_f)$ is good. Then, applying Lemma \[lemma-open-cover\], choose a $U_f \in {{\mathcal B}}$ such that that $(U_f,\rho_f,\tau_f)$ is good and $\widehat{\psi^{s_f} f\,} \subseteq U_f$. Next, apply the definition of “$n$–big” using ${{\mathcal A}}{\mathord {\upharpoonright}}H$: Fix $c \in H $ and $f_1,f_2, \ldots , f_n \in {{\mathcal A}}$ and $(U,\rho,\tau,s)$ such that $(U_{f_j},\rho_{f_j},\tau_{f_j}) = (U,\rho,\tau)$ and $s_{f_j} = s$ for all $j$, and also $|f_j(c) - f_k(c)| \ge 1$, and hence $|\psi^s(c) f_j(c) - \psi^s(c) f_k(c)| \ge 1$, whenever $j \ne k$.
Let $h(x) = \Xi_n( (\psi(x))^s \, f_1(x), \ldots, (\psi(x))^s \, f_n(x))$; then $h \in {{\mathcal A}}$. Then, choose $\Phi \in \{{\mbox{\sc{r}}}^+, {\mbox{\sc{r}}}^-, {\mbox{\sc{i}}}^+, {\mbox{\sc{i}}}^-\}$ so that $\Phi( (\psi(c))^s f_1(c), \ldots, (\psi(c))^s f_n(c)) \ge 1$, and let $k(x) = \Phi( (\psi(x))^s \,f_1(x), \ldots,(\psi(x))^s \,f_n(x))$; so $k(x)$ is either $\pm\Re(h(x))$ or $\pm\Im(h(x))$.
Note that when $x \in X \backslash V$, each $|(\psi(x))^s \, f_j(x) - (\psi(x))^s \, f_k(x))| \le 1/4$ so (referring to the definition of $\Xi$), $|h(x)| \le 1/2$. Then $k(X) = k(V) \cup k(X \backslash V) \subseteq \Phi * U \cup [-1/2,1/2]$ is disjoint from $[\rho,\tau]$, but contains $k(c) > \tau$ and $k(a) < \rho$. Thus, either $\Re(h(X))$ or $\Im(h(X))$ is not connected, so ${{\mathcal A}}$ contains a non-trivial idempotent by Lemma \[lemma-get-clopen\].
In this section, we use only the special case of this lemma where $H = V = X$, in which case the hypotheses on $\psi$ are trivial, and the above proof can be simplified somewhat. The more general result will be needed in Section \[sec-remov\].
Setting $H = V = X$, we have:
\[lemma-big-ntip\] Suppose that $n \ge 2$ and $X$ is both $n$–big and $n$–superdissipated. Then $X$ has the NTIP.
Applying this and Lemma \[lemma-two-big\], we have:
If $X$ is $2$–superdissipated and is not second countable, then $X$ has the NTIP.
This theorem yields the NTIP for some spaces not covered by [@HK1; @KU], but the result on CSWP, obtained from Corollary \[cor-idem\], is contained in the results of [@HK1]:
\[cor-CSWP-onelim\] If $X$ is $2$–superdissipated and does not contain a Cantor subset, then $X$ has the CSWP.
The examples of [@HK2; @HK3] show (under $\diamondsuit$ or CH) that this need not hold if $X$ is merely $2$–dissipated. To extend this corollary to $3$–*super*dissipated spaces, we need a mechanism (Lemma \[lemma-three-big\]) for proving that a space is $3$–big. This notion, unlike $2$–big (see Lemma \[lemma-two-big\]), does not seem to have a simple equivalent in terms of standard cardinal functions; see Section \[sec-rem\].
\[lemma-get-good\] Assume that $n \ge 1$ and that $X$ is $(n+2)$–superdissipated but not $(n+1)$–superdissipated, and then fix $\sigma: X {\twoheadrightarrow}Z$ which is $(n+2)$–supertight, where $Z$ is compact metric. Assume that $X$ does not have a Cantor subset. Fix ${{\mathcal A}}{\sqsubseteq}C(X)$ and $\Upsilon : {{\mathcal A}}\to \omega$. Fix any disjoint open sets $V_0,V_1,V_2 \subseteq {{\mathbb C}}$ and any $\pi \in {{\mathfrak M}}(X)$. Then there are $f,g,a,d,c$ such that:
$f,g\in {{\mathcal A}}$ and $\Upsilon(f) = \Upsilon(g)$.
$a, d \in X$, $c = \sigma(a) = \sigma(d) \in Z$, and $\pi(a) = \pi(d)$.
$f(a) \in V_0$ and $g(a) \in V_1$.
$f(d) \in V_2$ and $g(d) \in V_2$.
For all $x \in \sigma{^{-1}}\{c\}$, $\ (f(x),g(x)) \in V_0 \times V_1 \;\cup\; V_2 \times V_2$.
First, replacing $\pi$ by a finer map, we may assume that $\pi \le \sigma$, so that $\pi \in {{\mathfrak M}_{\mathfrak S}}(X)$ and $\pi$ also is $(n+2)$–supertight (see Lemmas \[lemma-MS-down\], \[lemma-count\], and \[lemma-lim-finer\]). Say $\pi: X {\twoheadrightarrow}Y$; then fix $\Gamma \in C(Y,Z)$ with $\sigma = \Gamma \circ \pi$.
Since $\pi$ is not $(n+1)$–supertight, fix a loose family for $\pi$, $\{P_0, \ldots, P_n\}$, with each $\pi(P_j) = Q$ and each $P_j$ perfect (see Lemma \[lemma-tight-equiv\]). Then $\{P_0, \ldots, P_n\}$ is also a loose family for $\sigma$, with each $\sigma(P_j) = \Gamma(Q)$; note that $\Gamma(Q)$ cannot be scattered since $Q$ is not scattered and each $\Gamma{^{-1}}\{z\}$ is scattered. Then $\sigma{^{-1}}(\Gamma(Q)) = \pi{^{-1}}(\Gamma{^{-1}}(\Gamma(Q)))$ is superdissipated by Lemma \[lemma-get-one\], so it has the CSWP by Corollary \[cor-CSWP-onelim\]. Also, $X$ is totally disconnected by Lemma \[lemma-discon\]. Fix closed disjoint $\widetilde{P_j} \subseteq \sigma{^{-1}}(\Gamma(Q))$ such that each $\widetilde{P_j} \supseteq P_j$ and $\bigcup_j P_j = \sigma{^{-1}}(\Gamma(Q))$. Note that each $\sigma {\mathord {\upharpoonright}}\widetilde{P_j}$ is supertight by Lemma \[lemma-get-one\].
Choose $y_\xi \in Q$ for $\xi < \omega_1$ such that the $\Gamma(y_\xi)$ are all different and each $|\pi{^{-1}}\{y_\xi\} \cap P_0| \ge 2$; this is possible because $P_0$ does not have a Cantor subset. Then, applying the CSWP for $\sigma{^{-1}}(\Gamma(Q))$, choose $h_\xi \in {{\mathcal A}}$ for $\xi < \omega_1$ such that $h_\xi(\widetilde{P_0}) \subseteq V_0 \cup V_1$, $h_\xi(\widetilde{P_j}) \subseteq V_2$ when $j \ge 1$, and $h_\xi( \pi{^{-1}}\{y_\xi\} \cap P_0 )$ meets both $V_0$ and $V_1$. Since there are only countably many values for $\Upsilon$, we may assume that the $\Upsilon(h_\xi)$ are all the same. For each $\xi$, we have $P_0$ partitioned into two relatively clopen sets, $h_\xi{^{-1}}(V_0) \cap P_0$ and $h_\xi{^{-1}}(V_1) \cap P_0$, and both these sets meet $\pi{^{-1}}\{y_\xi\}$. If these clopen partition were the same for all $\xi$, we would contradict the tightness of $\pi{\mathord {\upharpoonright}}P_0$ (see Lemma \[lemma-get-one\]), so that we may fix $\xi \ne \eta$ with $H := h_\xi{^{-1}}(V_0) \cap h_\eta{^{-1}}(V_1) \cap P_0$ non-empty, and thus perfect. Let $f = h_\xi$ and $g = h_\eta$.
If we choose any $a \in H$, we may set $c = \sigma(a)$, and choose any $d \in P_1 \cap \pi{^{-1}}\{\pi(a)\}$. This will satisfy (1)(2)(3)(4), but (5) might fail, since there may be an $x_c \in \sigma{^{-1}}\{c\}$ such that $x_c \in \widetilde{P_0}$ and either $f(x_c) \in V_1$ or $g(x_c) \in V_0$. But note that we also have $a \in \sigma{^{-1}}\{c\}$ and $a \in \widetilde{P_0}$ and $f(a) \in V_0$ and $g(a) \in V_1$. Consider the map $(f,g) : X \to {{\mathbb C}}\times{{\mathbb C}}$. If (5) fails for every choice of $a \in H$, then there would be uncountably many $c \in \pi(H)$ such that $(f,g)$ takes more than one value on $\widetilde{P_0} \cap \sigma{^{-1}}\{c\}$, contradicting the tightness of $\sigma{\mathord {\upharpoonright}}\widetilde{P_0}$. Thus, we may choose $a,b,d$ so that (1)(2)(3)(4)(5) hold.
\[lemma-three-big\] Assume that $X$ is not dissipated, but that $X$ is $m$–superdissipated for some $m \in \omega$, and that $X$ does not have a Cantor subset. Then $X$ is $3$–big.
Fix ${{\mathcal A}}{\sqsubseteq}C(X)$ and $\Upsilon : {{\mathcal A}}\to \omega$. Fix any disjoint open sets $V_0,V_1,V_2 \subseteq {{\mathbb C}}$. To verify that $X$ is $3$–big, it is sufficient to find $h_0,h_1,h_2 \in {{\mathcal A}}$ and $x \in X$ such that each $h_j(x) \in V_j$.
Fix $n \ge 1$ such that $X$ is $(n+2)$–superdissipated but not $(n+1)$–superdissipated, and then fix $\sigma: X {\twoheadrightarrow}Z$ which is $(n+2)$–supertight Let ${{\mathcal B}}$ be a countable open base for $Z$. For $\pi \in {{\mathfrak M}}(X)$, call $F = (f, g, a, d, c ,s, U) = (f^F, g^F, a^F, d^F, c^F ,s^F, U^F)$ *good for* $\pi$ iff (1–5) from Lemma \[lemma-get-good\] hold together with:
$s \in \omega$ and $\Upsilon(f) = \Upsilon(g) = s$.
$c \in U$, $U \in {{\mathcal B}}$, and for all $x \in \sigma{^{-1}}(\overline U)$, $\ (f(x),g(x)) \in V_0 \times V_1 \;\cup\; V_2 \times V_2$.
Such an $F$ always exists. To see this, first get $(f, g, a, d, c)$ by Lemma \[lemma-get-good\] to satisfy (1–5). Then (6) is trivial, and we choose $U$ to satisfy (7) using the fact that $\{z \in Z : \forall x \in \sigma{^{-1}}\{z\}\,
[(f(x),g(x)) \in V_0 \times V_1 \;\cup\; V_2 \times V_2]\}$ is open.
Note that if $F$ is good for $\pi$ and $\pi \le \varphi$ then $F$ is good for $\varphi$.
Next, note that there are fixed $s$ and $U$ such that for all $\pi \in {{\mathfrak M}}(X)$, there is an $F$ good for $\pi$ with $s^F = s$ and $U^F = U$: If not, then for each $s,U$, choose $\varphi_{s,U}$ such that no $F$ good for $\varphi$ satisfies $s^F = s$ and $U^F = U$. Then fix $\pi$ such that $\pi \le \varphi_{s,U}$ for each $s,U$. An $F$ which is good for $\pi$ yields a contradiction.
For each $\pi$, choose $F^\pi$ good for $\pi$ with $s^{F^\pi} = s$ and $U^{F^\pi} = U$, and write $(f^\pi, g^\pi, a^\pi, d^\pi, c^\pi)$ for $(f^{F^\pi}, g^{F^\pi}, a^{F^\pi}, d^{F^\pi}, c^{F^\pi})$.
Now, for each $\pi$, we have $\sigma{^{-1}}(\overline U)$ partitioned into two relatively clopen sets, $A^\pi = \{x \in \sigma{^{-1}}(\overline U) :
(f^\pi(x), g^\pi(x)) \in V_0 \times V_1\}$ and $D^\pi =
\{x \in \sigma{^{-1}}(\overline U) :
(f^\pi(x), g^\pi(x)) \in V_2 \times V_2\}$. If these are all the same, say $A^\pi = A$ and $D^\pi = D$ for all $\pi$; then we may fix $\pi\in C(X,[0,1])$ which is $0$ on $A$ and $1$ on $D$, so $\pi(a^\pi) = 0$ and $\pi(d^\pi) = 1$, contradicting (2). Thus, we can choose $\pi, \varphi$ and an $x \in A^\pi \cap D^\varphi$; then $f^\pi(x) \in V_0$, $g^\pi(x) \in V_1$, $f^\varphi(x) \in V_2$, as required.
The “obvious” generalization of this would say that if $X$ does not have a Cantor subset and is $(n+2)$–superdissipated but not $(n+1)$–superdissipated, then $X$ is $(n+2)$–big. For $n=1$ this is Lemma \[lemma-three-big\], and for $n=0$ this is Lemma \[lemma-two-big\]. Unfortunately, this is not true in general; see Example \[ex-big-limprop\]. We do get:
\[thm-big-cswp\] Assume that $X$ is compact and is $3$–superdissipated and does not have a Cantor subset. Then $X$ has the CSWP.
Since “$3$–superdissipated” is closed-hereditary, it is sufficient, by Corollary \[cor-idem\], to assume that $X$ is also perfect and prove that $X$ has the NTIP. $X$ cannot be second countable, so $X$ is $2$–big by Lemma \[lemma-two-big\]. If $X$ is not $2$–superdissipated, then $X$ is $3$–big by Lemma \[lemma-three-big\]. Thus, whether or not $X$ is $2$–superdissipated, it has the NTIP by Lemma \[lemma-big-ntip\].
\[cor-two-lots\] If $X$ is compact and $X \subseteq L_0 \times L_1$, where $L_0,L_1$ are a LOTSes, then $X$ has the CSWP iff $X$ does not contain a copy of the Cantor set.
By Lemma \[lemma-one-S\], we may assume that $X \subseteq (I_S)^2$. Then $X$ is $3$–superdissipated by Lemma \[lemma-big-lim\], so $X$ has the CSWP by Theorem \[thm-big-cswp\].
We now can extend this to products of three LOTSes, using an argument which is much more specific to ordered spaces. First, we introduce a notation for lines, boxes, etc. in such products.
\[def-boxes\] Let $\prod_{\alpha < \kappa} L_\alpha$ be a product of LOTSes, and use $<$ for the order on each $L_\alpha$. Then:
["2B]{}
If $\beta < \kappa$ and $c$ is a point in $\prod_{\alpha \ne \beta} L_\alpha$, then ${\mathrm{line}}(\beta, c) = \{x \in \prod_{\alpha < \kappa} L_\alpha :
\forall \alpha \ne \beta \, [ x_\alpha = c_\alpha ]\}$. A *line* in $\prod_{\alpha < \kappa} L_\alpha$ is any set of the form ${\mathrm{line}}(\beta, c)$.
$<^+$ is $<$; $<^-$ is $>$; $\le^+$ is $\le$; $\le^-$ is $\ge$ .
${{\mathcal D}}= \{+, -\}^\kappa$ is the set of all *directions*. For $\Delta \in {{\mathcal D}}$ and $x,y \in \prod_{\alpha < \kappa} L_\alpha$, $x <^\Delta y$ iff $\forall \alpha \, [ x_\alpha <^{\Delta_\alpha}y_\alpha]$ and $x \le^\Delta y$ iff $\forall \alpha \, [ x_\alpha \le^{\Delta_\alpha}y_\alpha]$.
If $a,b \in \prod_{\alpha < \kappa} L_\alpha$, then ${\mathrm{box}}[a,b] = \prod_{\alpha < \kappa}
[\min(a_\alpha,b_\alpha), \max(a_\alpha,b_\alpha)] $, and a (closed) *box* is any set of this form.
If $a \in \prod_{\alpha < \kappa} L_\alpha$ and $\Delta \in {{\mathcal D}}$, then ${\mathrm{corn}}(a,\Delta) = \{x \in \prod_{\alpha < \kappa} L_\alpha
: a \le^\Delta x\}$.
If $a \in B \subseteq \prod_{\alpha < \kappa} L_\alpha$ and $\Delta \in {{\mathcal D}}$, then ${\mathrm{corn}}(a, B, \Delta) = B \cap {\mathrm{corn}}(a,\Delta)$.
For example, in ${{\mathbb R}}^3$: $(2,4,6) <^{+-+} (3,3,7) \le^{+-+} (4,2,7)$. Now, let $B = [0,9]^3 = {\mathrm{box}}[(0,0,0), (9,9,9)] = {\mathrm{box}}[(9,0,9), (0,9,0)]$. Then ${\mathrm{corn}}( (2,4,6), B, {\mathord{+}}{\mathord{-}}{\mathord{+}})$ is the box $[2,9]\times [0,4]\times [6,9]$. The directions $\Delta\in{{\mathcal D}}$ are also useful inside products of the form $(I_S)^\kappa$. Continuing the notation of Definition \[def-das\],
If $\sigma: (I_S)^\kappa {\twoheadrightarrow}I^\kappa$ is the standard map, $y \in I^\kappa$, and $\Delta\in {{\mathcal D}}$, then $y^\Delta = \langle y^{\Delta_\alpha} : \alpha < \kappa\rangle$.
For example, if $b = (b_0,b_1,b_2) \in I^3$, then $\sigma{^{-1}}\{b\}$ consists of the points, $b^{\pm\pm\pm} = (b_0^\pm,b_1^\pm,b_2^\pm)$; e.g., $b^{+-+}$ denotes the point $(b_0^+,b_1^-,b_2^+) \in (I_S)^3$. The size of $\sigma{^{-1}}\{b\}$ will be $8,4,2$ or $1$ depending on whether $3,2,1$ or $0$ of the $b_0,b_1,b_2$ lie in $S$.
The following lets us establish bigness for subsets of $(I_S)^n$ by checking a simpler geometric property:
\[lemma-big-helper\] Fix $S \subseteq (0,1)$, a closed $X \subseteq (I_S)^n$, and $m$ with $2^{n-1} < m \le 2^n$. Assume that whenever $\Upsilon : S^n \to \omega$, there are distinct $ \Delta_1, \Delta_2, \ldots, \Delta_m \in {{\mathcal D}}$, a point $x \in X$, and $ d_1, d_2, \ldots, d_m \in S^n$ such that $x \in {\mathrm{corn}}(d_j^{\Delta_j}, \Delta_j)$ for each $j$, and such that $ \Upsilon(d_1) = \Upsilon(d_2) = \cdots = \Upsilon(d_m)$. Then $X$ is $m$–big.
Note that for $d \in S^n$, the points $d^\Delta \in (I_S)^n$, for $\Delta \in {{\mathcal D}}$, are all distinct, and the ${\mathrm{corn}}(d^\Delta, \Delta)$, for $\Delta \in {{\mathcal D}}$, partition $(I_S)^n$ into $2^n$ clopen subsets.
Fix ${{\mathcal A}}{\sqsubseteq}C(X)$ and $\Upsilon : {{\mathcal A}}\to \omega$. Since finite spaces have the CSWP, we may choose, for each $d \in S^n$, an $f_d \in C((I_S)^n)$ with $f_d {\mathord {\upharpoonright}}X \in {{\mathcal A}}$ such that the $f_d(d^\Delta)$, for $\Delta \in {{\mathcal D}}$, are $2^n$ distinct integers. We shall verify the definition of “$m$–big” just by considering the functions $f_d {\mathord {\upharpoonright}}X$; the $r$ in Definition \[def-big\] will be $1/2$.
Each $f_d$ is continuous, so choose $p(d), q(d) \in {{\mathbb Q}}^n$ with $\forall \mu\,[ p(d)_\mu < d_\mu < q(d)_\mu]$ such that $\sup\{ | f_d(x) - f_d(d^\Delta) | :
x \in {\mathrm{corn}}(d^\Delta,\; {\mathrm{box}}[p(d)^{+},q(d)^{-}] ,\; \Delta)\} \le 1/4$ for each $\Delta \in {{\mathcal D}}$. Here, for $y \in I^n$, $y^+$ abbreviates $(y_0^+, \ldots, y_{n-1}^+)$ and $y^-$ abbreviates $(y_0^-, \ldots, y_{n-1}^-)$. Now, let $\Upsilon'(d) = (\Upsilon(f_d {\mathord {\upharpoonright}}X),\; p(d),\; q(d))$. Since ${\mathrm{ran}}(\Upsilon')$ is countable, we may apply the hypotheses of the lemma and fix distinct $ \Delta_1, \Delta_2, \ldots, \Delta_m \in {{\mathcal D}}$, along with $x \in X$ and $ d_1, d_2, \ldots, d_m \in S^n$, such that $x \in {\mathrm{corn}}(d_j^{\Delta_j}, \Delta_j)$ for each $j$, $\; \Upsilon(f_{d_1}{\mathord {\upharpoonright}}X) =
\Upsilon(f_{d_2}{\mathord {\upharpoonright}}X) = \cdots = \Upsilon(f_{d_m}{\mathord {\upharpoonright}}X)$, and also each $p(d_j) = p$ and $q(d_j) = q$ for some $p,q \in {{\mathbb Q}}^n$.
Note that all $d_j^\Delta \in {\mathrm{box}}[p^{+},q^{-}]$. Also, since $m > 2^{n-1}$, $\{ \Delta_1, \Delta_2, \ldots, \Delta_m\}$ contains both $\Delta$ and $-\Delta$ for some $\Delta$, which implies (using $x \in {\mathrm{corn}}(d_j^{\Delta_j}, \Delta_j)$) that $x \in {\mathrm{box}}[p^{+},q^{-}]$. Thus, $x \in {\mathrm{corn}}(d_j^{\Delta_j},\; {\mathrm{box}}[p^{+},q^{-}],\; \Delta_j)$, so $| f_{d_j}(x) - f_{d_j}(d_j^{\Delta_j}) | \le 1/4$ for each $j$, so that $|f_{d_j}(x) - f_{d_k}(x)| \ge 1/2$ when $j \ne k$.
Note that the points $ d_j^{\Delta_j}$ were not assumed to lie in $X$.
\[lemma-six-big\] Assume that $S \subseteq (0,1)$, $X$ is a closed subspace of $(I_S)^3$, $X$ is not $3$–dissipated, and $X$ does not contain a Cantor subset. Then $X$ is $6$–big.
We verify the hypotheses of Lemma \[lemma-big-helper\], so fix $\Upsilon : S^3 \to \omega$; we must find appropriate $ \Delta_1, \Delta_2, \ldots, \Delta_6 \in {{\mathcal D}}= \{+, -\}^3$, $x \in X$, and $ d_1, d_2, \ldots, d_6 \in S^3$.
Note that it is sufficient to find $x$ along with points $c_E, c_F, c_G \in S \times S$, numbers $u_E, w_E, u_G, w_G, u_F, v_F, w_F \in S$, and $\Delta \in \{+, -\}^2$ such that:
$c_E <^\Delta c_F <^\Delta c_G$.
$u_E, u_F, u_G < v_F$ and $v_F < w_E, w_F, w_G$.
$\Upsilon$ has the same value on the $6$ points: $
d_1 = (c_E, u_E),
d_2 = (c_E, w_E),\\
d_3 = (c_F, u_F),
d_4 = (c_F, w_F),
d_5 = (c_G, u_G),
d_6 = (c_G, w_G)
$.
$x \in X$ and $x$ is one of the four points $(c_F^{ \Gamma}, v_F^\pm)$, where $\Gamma \in \{+, -\}^2$ and $\Gamma$ is different from $\Delta$ and $-\Delta$.
Note that no ordering is assumed among $u_E, u_F, u_G$ or among $w_E, w_F, w_G$. To verify that (1–4) are sufficient, and to clarify our notation, assume WLOG that $\Delta = ++$, so $c_E <^{++} c_F <^{++} c_G$. Then $\Gamma$ is either $+-$ or $-+$; WLOG $\Gamma = +-$, so we are assuming $X$ contains at least one of the two points $(c_F^{+-}, v_F^\pm)$, denoted by $x$. But now we obtain the hypotheses of Lemma \[lemma-big-helper\]. Namely, $x \in {\mathrm{corn}}(d_j^{\Delta_j}, \Delta_j)$ for $j = 1,2,\ldots,6$, setting $
\Delta_1 = {\mathord{+}}{\mathord{+}}{\mathord{+}}\,,
\, \Delta_2 = {\mathord{+}}{\mathord{+}}{\mathord{-}}\,,
\, \Delta_3 = {\mathord{+}}{\mathord{-}}{\mathord{+}}\,,
\, \Delta_4 = {\mathord{+}}{\mathord{-}}{\mathord{-}}\,, \allowbreak
\, \Delta_5 = {\mathord{-}}{\mathord{-}}{\mathord{+}}\,,
\, \Delta_6 = {\mathord{-}}{\mathord{-}}{\mathord{-}}$
Now, to obtain (1–4): If $E \subseteq S$, let $\sigma_E: (I_S)^3 {\twoheadrightarrow}(I_E)^3$ be the natural map; so $\sigma_\emptyset = \sigma$. If also $E \in [S]^\omega$ (i.e., $|E| = \aleph_0$), then $(I_E)^3$ is a compact metric space, and we shall use the fact that *none* of these $\sigma_E$ are $3$–tight.
If $E_1 \subseteq E_2 \in [S]^\omega$ then $\sigma_{E_2} \le \sigma_{E_1}$ (see Lemma \[lemma-fine-equiv\]). Observe that $ [S]^\omega $ is countably directed upward. Call ${{\mathcal U}}\subseteq [S]^\omega$ *cofinal* iff $\forall E_1 \in [S]^\omega \, \exists E_2 \in {{\mathcal U}}\, ( E_1 \subseteq\nobreak E_2 )$; then ${{\mathcal U}}$ is also countably directed upward. We shall use this observation several times to show that a number of quantities dependent on $E$ can in fact be chosen uniformly, independently of $E$, on a cofinal set.
Temporarily fix an $E \in [S]^\omega$. Then we have $P_j = P^E_j \subseteq X \subseteq (I_S)^3 $ for $j = 0,1,2$ such that $\{P_0, P_1, P_2\}$ is a loose family. Then each $\sigma_E(P_j) = Q$, where $Q = Q^E\subseteq \sigma_E(X) \subseteq (I_E)^3$ is uncountable. We can now get such a $Q$ to be of a very simple form:
First, note that $Q$ must be a subset of finite union of lines. If not, then we may choose $y^\ell = (y^\ell_0, y^\ell_1, y^\ell_2) \in Q$ for $\ell \in \omega$ such that no two of the $y^\ell$ lie on the same line; that is, whenever $\ell < m < \omega$, the triples $y^\ell$ and $y^m$ differ on at least two coordinates. Now, we may thin the sequence and permute the coordinates and assume that each of the two sequences $\langle y^\ell_0 : \ell \in \omega \rangle$ and $\langle y^\ell_1 : \ell \in \omega \rangle$ is either strictly increasing or strictly decreasing, while $\langle y^\ell_2 : \ell \in \omega \rangle$ is either constant or strictly increasing or strictly decreasing. If $H$ is the set of limit points of the sequence of sets $\langle \sigma{^{-1}}\{y^\ell_i\} : \ell \in \omega \rangle$, then $|H| \le 2$, but $H$ must meet each of $P_0,P_1,P_2$, which is a contradiction.
Next, shrinking $Q$, along with $P_0,P_1,P_2$, we may assume that $Q = Q^E$ is a subset of one line; say $Q^E \subseteq {\mathrm{line}}(\beta_E, c_E)$, where $\beta_E < 3$.
$\beta_E$ depends on $E$, but since $ [S]^\omega $ is countably directed upward, there is a fixed $\beta$ such that $\beta_E = \beta$ on a cofinal set ${{\mathcal U}}\subseteq [S]^\omega$. By permuting coordinates, we may assume $\beta = 2$, so that $Q^E \subseteq {\mathrm{line}}(2, c_E) \subseteq (I_E)^3$, where $c_E = (a_E, b_E) \in (I_E)^2$. From now on, we shall delete the “2”; so ${\mathrm{line}}(c_E) = \{(a_E, b_E,u) : u \in I_E \}$. Then $Q^E = \{c_E\} \times \widetilde Q_E $, where $\widetilde Q_E \subseteq I_E$.
Again, fix $E$, and temporarily delete some of the sub/super-script $E$. Now $\sigma_E{^{-1}}({\mathrm{line}}( c)) \subseteq (I_S)^3$ is a union of 1, 2, or 4 lines in $(I_S)^3$. However, the existence of $Q,P_0,P_1,P_2$ implies that $\sigma_E : \sigma_E{^{-1}}({\mathrm{line}}(c)) {\twoheadrightarrow}{\mathrm{line}}(c)$ is not $3$–tight, so in fact $\sigma_E{^{-1}}({\mathrm{line}}( c))$ is a union of 4 lines, which means that $a,b \in S \backslash E$; that is, we may regard $a,b$ as real numbers which are not split in $I_E$, but which are split into $a^\pm, b^\pm$ in $I_S$, and $\sigma_E{^{-1}}(Q) \subseteq
{\mathrm{line}}( c^{++} )\cup
{\mathrm{line}}( c^{+-} )\cup
{\mathrm{line}}( c^{-+} )\cup
{\mathrm{line}}( c^{--} )
\subseteq (I_S)^3$. Now $\sigma_E{^{-1}}(Q) \cap {\mathrm{line}}( c^{++} ) \cap X$ is some closed subset of $\sigma_E{^{-1}}(Q) \cap {\mathrm{line}}( c^{++} )$, but replacing $Q$ by a smaller perfect set, we may assume that this closed subset is either empty or all of $\sigma_E{^{-1}}(Q) \cap {\mathrm{line}}( c^{++} )$. Repeating this argument three more times, we may assume that each of the four sets $\sigma_E{^{-1}}(Q) \cap {\mathrm{line}}( c^{\pm \pm } )$ is either contained in $X$ or disjoint from $X$. Again, the existence of $P_0,P_1,P_2$ implies that $\sigma_E : \sigma_E{^{-1}}(Q) \cap X {\twoheadrightarrow}Q$ is not $3$–tight, so at least three of the four sets $\sigma_E{^{-1}}(Q) \cap {\mathrm{line}}( c^{\pm \pm } )$ are contained in $X$. Which three or four depends on $E$; there is a cofinal set on which it is the same, although this is irrelevant now. More importantly, since $\widetilde Q_E \subseteq I_E$ and $E$ is countable, we may shrink $Q^E$ and assume that $\widetilde Q_E \cap E = \emptyset$; that is, we may regard $\widetilde Q_E$ as a perfect subset of $I \backslash E$. Note that $S$ must meet every perfect subset of $\widetilde Q$, since otherwise $X$ would contain a Cantor subset. In particular, $S \cap \widetilde Q$ is uncountable. Now $c_E = c = (a,b)$ is fixed, and for each $u \in S \cap \widetilde Q$, we have the triple $d = d_u = (a,b,u)$. We may now choose $t\in \omega$ and $u,v,w \in S \cap \widetilde Q$ such that $u < v < w$ and $\Upsilon(d_u) = \Upsilon(d_v) = \Upsilon(d_w) = t$. Also choose rational $\rho,\tau$ with $u < \rho < v < \tau < w$.
Of course, $t,\rho,\tau,u,v,w$ depend on $E$, but there are only $\aleph_0$ possibilities for $t,\rho,\tau$, so we may assume that for $E$ in our cofinal set ${{\mathcal U}}$, these are always the same, whereas $u,v,w$ are really $u_E,v_E,w_E$.
Choose an increasing $\omega_1$ sequence $\langle E_\xi : \xi < \omega_1 \rangle$ of elements of ${{\mathcal U}}$ such that $\xi < \eta \to c_{E_\xi} \in (E_\eta)^2$. Now $c_{E_\xi} = ( a_{E_\xi} b_{E_\xi})$ and $ a_{E_\xi}, b_{E_\xi} \notin E_\xi $, so $a_{E_\eta} \ne a_{E_\xi}$ and $b_{E_\eta} \ne b_{E_\xi}$ whenever $\xi \ne \eta$. It follows that we may find distinct $\xi_n < \omega_1$ for $n \in \omega$ and a fixed $\Delta \in \{+, -\}^2$ such that $m < n \to c_{E_{\xi_m}} <^\Delta c_{E_{\xi_n}}$. But, we only need three of these, so let $E,F,G$ denote $E_{\xi_0}, E_{\xi_1}, E_{\xi_2}$. Then we have $c_E <^\Delta c_F <^\Delta c_G$ as in (1) above. $u_E, u_F, u_G < \rho < v_F < \tau < w_E, w_F, w_G$, so (2) holds. (3) holds because $\Upsilon$ has the same value $t$ on all $( a_{E_\xi}, b_{E_\xi} , u_{E_\xi})$, $( a_{E_\xi}, b_{E_\xi} , v_{E_\xi})$, $( a_{E_\xi}, b_{E_\xi} , w_{E_\xi})$. Finally, we may choose $x$ to make (4) hold because at least three of the four sets $\sigma_F {^{-1}}(\widetilde Q _ F) \cap {\mathrm{line}}( c_F^{\Gamma} )$ (for $\Gamma \in \{+, -\}^2$) are contained in $X$ and $v_F \in S \cap \widetilde Q _ F $, and for these $\Gamma$, both points $(c_F^{ \Gamma}, v_F^\pm)$ lie in $X$.
[Theorem \[thm-three-lots\]]{} By Lemma \[lemma-one-S\], we may assume that $X \subseteq (I_S)^3$. Since the properties assumed of $X$ are closed-hereditary, it is sufficient, by Corollary \[cor-idem\], to assume that $X$ is also perfect and prove that $X$ has the NTIP. Note that “dissipated” is the same as “superdissipated” for these spaces. If $X$ is $3$–dissipated, then $X$ has the CSWP, and hence the NTIP, by Theorem \[thm-big-cswp\]. If $X$ is not $3$–dissipated, then $X$ is $5$–big by Lemma \[lemma-six-big\], but it also is $5$–dissipated by Lemma \[lemma-big-lim\], so $X$ has the NTIP by Lemma \[lemma-big-ntip\].
We do not know if the same theorem holds when $X$ is contained in a product of four LOTSes, but the analogue of Lemma \[lemma-six-big\] is false. That is, there is (see Example \[ex-seven\]) a closed $X \subseteq (I_S)^4$ such that $X$ is not $8$–dissipated and is not $7$–big. Of course, $X$ must be $9$–dissipated, but to prove the NTIP by our methods, $X$ would need to be $9$–big.
Removable Spaces {#sec-remov}
================
The property of a compact space being *removable*, defined in [@HK1], is a strengthening of the CSWP. Many of the spaces proved in Section \[sec-cswp\] to have the CSWP are in fact removable. We recall the definition, which is in terms of the Šilov boundary:
\[def-silov\] If ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$, then ${\mbox{\cyr Sh}}({{\mathcal A}})$ denotes the *Šilov boundary*; this is the smallest non-empty closed $H \subseteq X$ such that $\|f\| = \sup\{|f(x)| : x \in H\}$ for all $f\in {{\mathcal A}}$.
This is discussed in texts on function algebras; see [@GAM; @GA]. Note that ${\mbox{\cyr Sh}}({{\mathcal A}})$ cannot be finite unless $X$ is finite, in which case ${\mbox{\cyr Sh}}({{\mathcal A}}) = X$.
\[def-remove\] A compact space $K$ is *removable* iff for all $X,U,{{\mathcal A}}$, if:
["2B]{}
$X$ is compact, $U \subsetneqq X$, and $U$ is open,
$\overline U$ is homeomorphic to a subspace of $K$, and
${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$ and all idempotents of ${{\mathcal A}}$ are trivial,
then ${\mbox{\cyr Sh}}({{\mathcal A}}) \subseteq X\backslash U$.
The next four lemmas are clear from [@HK1]:
If $X$ is removable, then $X$ is totally disconnected and has the CSWP.
It is unknown whether the converse to this lemma is true. The removable spaces are of interest because one can prove some theorems about them which are currently unknown for the CSWP spaces. In particular, the removable spaces form a local class (see Definition \[def-local\]); this follows from:
If the compact $X$ is a finite union of closed sets, each of which is removable, then $X$ is removable.
More generally, one can do a type of Cantor-Bendixson analysis for a compact $X$, iteratively deleting open sets with removable closures; if one gets to $\emptyset$, then $X$ itself is removable and hence has the CSWP (see [@HK1], Lemma 2.15). This results in the next definition and lemma.
A compact space $P$ is *nowhere removable* iff $\overline W$ is not removable for all non-empty open $W \subseteq P$.
If $X$ is compact and not removable, then there is a non-empty closed $P \subseteq X$ such that $P$ is nowhere removable.
In particular, since the one-point space is removable,
Every compact scattered space is removable.
\[def-R\] ${{\mathfrak R}}$ is the class of all compact spaces $X$ such that for all perfect $H \subseteq X$: There is non-empty relatively clopen $U \subseteq H$ such that either $U$ is removable or for some finite $n \ge 2$, $U$ is both $n$–big and $n$–superdissipated.
If $X$ is removable, then $X \in {{\mathfrak R}}$, and we shall soon prove the converse statement. No space in ${{\mathfrak R}}$ can contain a Cantor subset (since the Cantor set is neither $2$–big nor removable). All spaces in ${{\mathfrak R}}$ are totally disconnected by Lemma \[lemma-discon\].
Our proof will use the following restatement of Definition \[def-remove\]:
\[lemma-restate\] Assume that ${{\mathfrak K}}$ is a closed-hereditary class of totally disconnected compact spaces, and assume that whenever $Z,V,{{\mathcal A}}$ satisfy:
["2B]{}
$Z$ is compact and infinite, ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(Z)$, and ${\mbox{\cyr Sh}}({{\mathcal A}}) = Z$.
$V \subseteq Z$, $V$ is clopen and non-empty, and $V \in {{\mathfrak K}}$.
then ${{\mathcal A}}$ contains a non-trivial idempotent. Then, all spaces in ${{\mathfrak K}}$ are removable.
Fix $K \in {{\mathfrak K}}$. Then fix $X,U,{{\mathcal A}}$ satisfying the hypotheses of Definition \[def-remove\]. Let $Z = {\mbox{\cyr Sh}}({{\mathcal A}})$. Assume that $Z \not\subseteq X \backslash U$. We shall derive a contradiction. Shrinking $U$, we may assume that $U$ is clopen. Clearly $U \ne \emptyset$, so $|X| \ge 2$ (by $U \subsetneqq X$), so $X$ is infinite (by all idempotents trivial), so $Z$ is infinite.
${{\mathcal A}}{\mathord {\upharpoonright}}Z {\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(Z)$ and ${\mbox{\cyr Sh}}({{\mathcal A}}{\mathord {\upharpoonright}}Z) = Z$. Let $V = Z \cap U$; then $V \ne \emptyset$. $V \in {{\mathfrak K}}$ because ${{\mathfrak K}}$ is closed-hereditary. So, ${{\mathcal A}}{\mathord {\upharpoonright}}Z $ contains a non-trivial idempotent, $f {\mathord {\upharpoonright}}Z$, where $f \in {{\mathcal A}}$. But then $f^2 - f$ is $0$ on $Z$ and hence on $X$, so $f$ is an idempotent, contradicting the hypotheses of Definition \[def-remove\].
\[thm-R-remov\] ${{\mathfrak R}}$ is the class of all removable spaces.
Since ${{\mathfrak R}}$ is clearly closed-hereditary, we may apply Lemma \[lemma-restate\] to prove that all spaces in ${{\mathfrak R}}$ are removable. Thus, assume that $X$ is compact and infinite, ${{\mathcal A}}{\, \mbox{\rlap{$\sqsubseteq$} \raise 2.3pt \hbox{\scriptsize $\mathrm{c}$}}\; }C(X)$, and ${\mbox{\cyr Sh}}({{\mathcal A}}) = X$, and $V \subseteq X$ is clopen and non-empty, and $V \in {{\mathfrak R}}$. We must show that ${{\mathcal A}}$ contains a non-trivial idempotent. We may assume that $V$ is nowhere removable, and in particular perfect, since otherwise the result is clear from the definition of “removable”. Applying the definition of ${{\mathfrak R}}$, whenever $U$ is a non-empty clopen subset of $V$, there is an $n_U \ge 2$ and a non-empty clopen $H$ with $H \subseteq U$ and $H$ both $n_U$–big and $n_U$–superdissipated. Taking a minimal $n_U$ and shrinking $V$, we may assume that $V$ itself is $n$–superdissipated, where $n \ge 2$, and that whenever $U$ is a non-empty clopen subset of $V$, there is a non-empty clopen $H$ with $H \subseteq U$ and $H$ $n$–big.
Since $X \backslash V$ is not a boundary, we may fix $\psi \in {{\mathcal A}}$ such that $\|\psi\| > 1$ but $|\psi(x)| \le 1/2$ for all $x \notin V$. Then fix a non-empty clopen $H \subseteq V$ such that $|\psi(x)| \ge 1$ for all $x \in H$. Shrinking $H$, we may assume that $H$ is $n$–big. We now get a non-trivial idempotent by Lemma \[lemma-big-dis\].
\[cor-3remov\] If $X \subseteq (I_S)^3$ is closed and does not contain a copy of the Cantor set, then $X$ is removable.
$X \in {{\mathfrak R}}$ by Lemmas \[lemma-big-lim\], \[lemma-six-big\], \[lemma-three-big\], and \[lemma-two-big\].
Powers of the Double Arrow Space {#sec-powers}
================================
Here we show that arbitrary finite powers of the double arrow space $I_{(0,1)}$ are removable, and hence have the CSWP. This argument works because there is a certain uniformity in the standard map from $(I_{(0,1)})^k$ onto $I^k$, which is captured by the next definition:
\[def-superdup\] For $n \ge 1$, $\pi : X {\twoheadrightarrow}Y$ is $n$–tight iff for $y \in Y$ and $0 \le j < n$, there are $K^j_y \subseteq X$ and $U^j_y \subseteq Y$ satisfying:
$X,Y$ are compact, $Y$ is metric, and the Cantor set does not embed into $X$.
For each $y$: The $K^j_y$, for $j < n$, form a clopen partition of $X$, and each $|K^j_y \cap \pi{^{-1}}\{y\}| \le 1$.
For each $j$: $\{(y,z) : z \in U^j_y\}$ is open in $Y^2$.
For each $y,j$: $\pi{^{-1}}(U^j_y) \subseteq K^j_y$.
For each $y,j$: $ K^j_y \setminus \pi{^{-1}}(U^j_y)$ is removable.
$X$ is $n$–dissipated iff $\pi: X {\twoheadrightarrow}Y$ is $n$–tight for some $\pi$ and $Y$.
Some of the $K^j_y$ and $U^j_y$ may be empty, so “$n$–dissipated” get weaker as $n$ gets bigger. Note that (2) implies that $|\pi{^{-1}}\{y\}| \le n$ for each $y$, so that $\pi$ is $n$–supertight by Lemma \[lemma-lim-prop\], and $X$ is totally disconnected by Lemma \[lemma-discon\]. $X$ is $1$–dissipated iff $X$ is compact and countable. The class of $n$–dissipated spaces is closed-hereditary, since if we have (1 – 5) and $\widetilde X$ is a closed subset of $X$, then we also have (1 – 5) for $\widetilde X$, using $\pi{\mathord {\upharpoonright}}\widetilde X : \widetilde X {\twoheadrightarrow}\widetilde Y = \pi(\widetilde X)$, $\widetilde K^j_y = K^j_y \cap \widetilde X$, and $\widetilde U^j_y = U^j_y \cap \widetilde Y$.
\[lemma-das-superdup\] If $(I_{(0,1)})^{k-1}$ is removable, then the standard map $\pi : (I_{(0,1)})^k {\twoheadrightarrow}I^k$ is $2^k$–tight.
As in Definition \[def-boxes\], let ${{\mathcal D}}= \{+, -\}^k$. For $y \in I^k$ and $\Delta \in {{\mathcal D}}$, let $U^\Delta_y = \{z \in I^k : y <^\Delta z\}$, and let $K^\Delta_y = \{t \in (I_{(0,1)})^{k} : y^\Delta \le^\Delta t\}$. Then properties (1 – 4) are easily verified, and (5) holds because $ K^j_y \setminus \pi{^{-1}}(U^j_y)$ is covered by finitely many homeomorphic copies of $(I_{(0,1)})^{k-1}$.
We shall eventually prove:
\[thm-superdup-remov\] If $n < \omega$ and $X$ is $n$–dissipated, then $X$ is removable.
It follows that $X$ is $n$–dissipated iff $X$ is removable and there is a $\pi : X {\twoheadrightarrow}Y$ such that $Y$ is compact metric and each $|\pi{^{-1}}\{y\}| \le n$. To prove the $\leftarrow$ direction: In Definition \[def-superdup\], take all $U^j_y = \emptyset$; the $K^j_y$ may simply be chosen arbitrarily to satisfy condition (2). Thus, the notion of “$n$–dissipated” becomes of little interest, but it was chosen to make the following proof work:
[Theorem \[thm-das-power\]]{} Each $(I_{(0,1)})^k$ is in fact removable. This follows by induction on $k$, using Lemma \[lemma-das-superdup\] and Theorem \[thm-superdup-remov\].
We shall now prove Theorem \[thm-superdup-remov\] by showing that $X \in {{\mathfrak R}}$ (see Definition \[def-R\]).
A compact space $P$ is *nowhere $n$–dissipated* iff $\overline W$ is not $n$–dissipated for all non-empty open $W \subseteq P$.
\[lemma-get-nowhere\] If $X$ is perfect and $n$–dissipated, then there is a non-empty clopen $V \subseteq X$ and an $m$ with $2 \le m \le n$ such that $V$ is $m$–dissipated and nowhere $(m-1)$–dissipated.
Theorem \[thm-superdup-remov\] will follow easily from the next two lemmas, about spaces which are $n$–dissipated and nowhere removable. Of course, the theorem implies that there are no such spaces.
\[lemma-nowhere\] Assume that $X$, $Y$, $n\ge 2$, $\pi$ and the $K^j_y$ and $U^j_y$ are as in Definition \[def-superdup\], with $X$ nowhere $(n-1)$–dissipated and nowhere removable.
For a fixed $j$ and non-empty open $V \subseteq Y$: $U^j_y \cap V \ne \emptyset$ for some $y \in V$.
For any $\varepsilon > 0$, the sets: $$\begin{aligned}
A^j_\varepsilon := \{z \in Y : \exists y [ z \in U^j_y
\ \&\ d(y,z) < \varepsilon ] \} \\
B^j_\varepsilon := \{y \in Y : \exists z [ z \in U^j_y
\ \&\ d(y,z) < \varepsilon ] \} \end{aligned}$$ are dense and open in $Y$.
For (1): Assume that $U^j_y \cap V = \emptyset$ for all $y \in V$. Let $W$ be a non-empty clopen subset of $\pi{^{-1}}\{V\}$, and consider the restriction $\pi{\mathord {\upharpoonright}}W : W {\twoheadrightarrow}\widetilde Y = \pi( W )$. $\widetilde U^j_y = U^j_y \cap \widetilde Y = \emptyset$ for each $y \in \widetilde Y$ and $\widetilde K^j_y = K^j_y \cap W = ( K^j_y \setminus \pi{^{-1}}(U^j_y) ) \cap W$ is empty for each $y \in \widetilde Y$ because it is clopen in $X$ and removable. But then, by deleting index $j$, we see that $W$ is $(n-1)$–dissipated; in the special case $n = 2$, $W$ would be countable because $X$ does not contain a Cantor subset.
For (2): They are open by (3) of Definition \[def-superdup\]. If one of them fails to be dense, then there is a non-empty open $V \subseteq Y$ such that $V$ is disjoint from either $A^j_\varepsilon$ or $B^j_\varepsilon$. In either case, we may assume that ${\mathrm{diam}}(V) < \varepsilon$ which implies that $z \notin U^j_y$ whenever $z,y \in V$, contradicting (1).
\[lemma-super-big\] If $n \ge 2$ and $X$ is $n$–dissipated and nowhere $(n-1)$–dissipated and nowhere removable, then $X$ is $n$–big.
Fix ${{\mathcal A}}{\sqsubseteq}C(X)$ and $\Upsilon : {{\mathcal A}}\to \omega$. We shall verify the conclusion of Definition \[def-big\] with $r = 1$, so we shall find $f_0, \ldots, f_{n-1} \in {{\mathcal A}}$ and $c \in X$ such that the $\Upsilon(f_j)$, for $j = 0, \ldots, n-1$, are all equal, and such that $|f_i(c) - f_j(c)| \ge 1$ whenever $0 \le i < j < n$. Let $Y$, $\pi$ and the $K^j_y$ and $U^j_y$ be as in Definition \[def-superdup\]. Let $G = \bigcap\{A^j_\varepsilon \cap B^j_\varepsilon:
\varepsilon > 0 \ \&\ j < n\}$; by Lemma \[lemma-nowhere\], $Y \backslash G$ is of first category in $Y$ because the intersection may be taken just over rational $\varepsilon$.
If $y \in G$, then $y$ is in the closure of each $U^j_y$, so that $\pi{^{-1}}\{y\}$ meets each $K^j_y$; let $x^j_y$ be the element of $\pi{^{-1}}\{y\} \cap K^j_y$. Since finite spaces have the CSWP, we may choose, for each $y \in G$, a $g_y \in {{\mathcal A}}$ such that $g_y(x^j_y) = 2j$ for each $j < n$. Then, chose a rational $\varepsilon_y > 0$ such that $|g_y(x) - 2j| < 1/2$ whenever $j < n$, $x \in K^j_y$, and $d(\pi(x),y) < \varepsilon_y$.
Now, fix $N \subseteq G$, $\varepsilon > 0$, and $\ell \in \omega$ such that $N$ is not of first category in $Y$ and $\varepsilon_y = \varepsilon$ and $\Upsilon(g_y) = \ell$ for all $y \in N$. Then, fix a point $d \in N$ and a $\delta$ with $0 < \delta < \varepsilon$ such that $N \cap B(d, \delta)$ is dense in $B(d, \delta)$. Let $c$ be any point in $\pi{^{-1}}\{d\}$. For each $j < n$, $\{y : d \in U^j_y\}$ is open, and this set meets $B(d, \delta)$ (since $d \in N \subseteq A^j_\delta$), so choose $y^j \in N \cap B(d, \delta)$ such that $d \in U^j_{y^j}$, and we can let $f_j = g_{y_j}$; note that $d \in U^j_{y^j} \to c \in K^j_{y^j} \to |f_j(c) - 2j| < 1/2$.
[Theorem \[thm-superdup-remov\]]{} Apply Theorem \[thm-R-remov\]; every $n$–dissipated space $X$ is in ${{\mathfrak R}}$ by Lemmas \[lemma-get-nowhere\] and \[lemma-super-big\].
Remarks and Questions {#sec-rem}
=====================
Regarding our notion of bigness: From the point of view of general topology, the use of the “${{\mathcal A}}{\sqsubseteq}C(X)$” in Definition \[def-big\] seems a bit artificial, although it was needed for the CSWP proofs. It would be more natural to restrict ${{\mathcal A}}$ to be only $C(X)$, which would result in a weaker property; but we do not know if it would really be strictly weaker. Of course, we can always replace ${{\mathcal A}}$ by ${\mathrm{cl}}({{\mathcal A}})$, so the two properties are equivalent when $X$ has the CSWP.
The degree of bigness of some LOTSes is easily calculated. Doing so lets us show (Example \[ex-big-limprop\]) that the “obvious” generalization of Lemma \[lemma-three-big\] is false. It is easy to see that $\omega_1 + 1$ is $n$–big for all $n$. But there is a class of LOTSes for which the bigness is bounded. We do not state the most general possible result, but just say enough to verify Example \[ex-big-limprop\], which uses the $I_\Lambda$ from Definition \[def-das\].
\[lemma-small-big\] Let $L = I_\Lambda$, where $\Lambda : I \to \omega$, and let $K$ be any compact space which is not $(n+1)$–big. Let $X = L \times K$. Then $X$ is not $(3n+1)$–big.
Let $\sigma: L {\twoheadrightarrow}I$ be the standard map. Also, applying the definition of “not $(n+1)$–big”, fix ${{\mathcal A}}{\sqsubseteq}C(K)$ and $\Upsilon : {{\mathcal A}}\to \omega$ such that for each $c \in K$ and each $f_0, f_1, \ldots, f_n \in {{\mathcal A}}$ with $\Upsilon(f_0) = \Upsilon(f_1) = \cdots = \Upsilon(f_n)$, there are $j < k \le n$ such that $|f_j(c) - f_k(c)| < 1/4$.
Let $M = C(K)$, with the usual $\sup$ norm. For $f \in C(X)$, define $\widetilde f \in C(L, M)$ by $(\widetilde f(u))(z) = f(u,z)$. Let ${{\mathcal B}}$ be the set of all $f \in C(X)$ such that $\widetilde f (u) \in {{\mathcal A}}$ for all $u \in L$. Then ${{\mathcal B}}{\sqsubseteq}C(X)$, and we shall define a partition $\Phi$ of ${{\mathcal B}}$ into $\aleph_0$ pieces demonstrating that $X$ is not $(3n+1)$–big. As a first approximation, for each $f \in {{\mathcal B}}$, choose $\Psi(f) = (m^f, \vec y^f, \vec r^f, \vec s^f, \vec t^f)$ so that:
$1 \le m^f \in \omega$.
$\vec y ^f = \langle y^f_i : 0 \le i \le 2m^f \rangle$, each $y^f_i \in I$, and $y^f_i \in {{\mathbb Q}}$ when $i$ is even.
$0 = y^f_0 < y^f_1 < \cdots < y^f_{2m^f} = 1$.
$\vec r ^f = \langle r^f_i : 0 \le i \le 2m^f \rangle$, where each $r^f_i = |\sigma{^{-1}}\{y^f_i\}| - 1 = \Lambda(y^f_i)$.
$\| \widetilde f(u) - \widetilde f(v)\| \le 1/4$ whenever $ \max(\sigma{^{-1}}\{y^f_i\}) \le u \le v \le \min(\sigma{^{-1}}\{y^f_{i+1}\})$.
$\vec s ^f = \langle s^f_{i,\mu} :
0 \le i \le 2m^f \ \&\ 0 \le \mu \le r^f_i\rangle$, where $\{s^f_{i,\mu} : 0 \le \mu \le r^f_i\} \subset L$ lists $\sigma{^{-1}}\{y^f_i\}$ in increasing order; so $s^f_{i,\mu} = (y^f_i, \mu)$.
$\vec t ^f = \langle t^f_{i,\mu} :
0 \le i \le 2m^f \ \&\ 0 \le \mu \le r^f_i\rangle$, where $t^f_{i,\mu} = \Upsilon(\widetilde f ( s^f_{i,\mu} ))$.
Such a $\Psi(f)$ may be chosen using compactness, plus continuity of $\widetilde f$. Of course, there are $2^{\aleph_0}$ possible values of $\Psi(f)$ because of the $y^f_i$ and $s^f_{i,\mu}$ for odd $i$, so we delete these and define $\Phi(f) = (m^f, \vec {\mathbbold{y}}^f, \vec r^f, \vec {\mathbbold{s}}^f, \vec t^f)$, where
$\vec {\mathbbold{y}}^f = \langle y^f_i : 0 \le i \le 2m^f
\ \&\ \mbox{$i$ is even} \rangle$.
$\vec {\mathbbold{s}}^f = \langle s^f_{i,\mu} :
0 \le i \le 2m^f \ \&\ \mbox{$i$ is even}\ \& \ 1 \le \mu \le r^f_i \rangle$.
There are only countably many possible values for $\Phi(f)$, so if $X$ were $(3n+1)$–big, we could fix a $(b,c) \in X = L \times K$ and $f_0, \ldots, f_{3n} \in {{\mathcal A}}$ such that the $\Phi(f_j)$ are all the same, and such that $|f_j(b,c) - f_k(b,c)| \ge 1$ whenever $j < k \le n$. We shall now derive a contradiction. Write $\Phi(f_j) = (m, \vec {\mathbbold{y}}, \vec r, \vec {\mathbbold{s}}, \vec t)$.
If $b = s^f_{i,\mu}$ for some even $i$, then the $\Upsilon(\widetilde f_j (b)) = t^f_{i,\mu}$ are all the same, and we contradict our assumptions on $\Upsilon$ just using $\widetilde f_0(b) , \ldots , \widetilde f_n(b)$. So, we may fix an even $i < 2m$ so that $ \max(\sigma{^{-1}}\{y_i\}) < b < \min(\sigma{^{-1}}\{y_{i+2}\})$. Now, for each $j \in \{0, 1, \ldots, 3n\}$, there are three cases:
$ \max(\sigma{^{-1}}\{y_i\}) < b < \min(\sigma{^{-1}}\{y^{f_j}_{i+1}\})$.
$b \in \sigma{^{-1}}\{y^{f_j}_{i+1}\}$.
$ \max(\sigma{^{-1}}\{y^{f_j}_{i+1}\}) < b <
\min(\sigma{^{-1}}\{y_{i+2}\})$.
So, one of these cases must happen for $n+1$ values of $j$. We shall assume that this is Case I, since the argument is essentially the same in the other two cases. Permuting the $f_j$, we may assume that Case I holds for $0 \le j \le n$. Fix $\mu = r_i$, so that $\max(\sigma{^{-1}}\{y_i\}) = s_{i,\mu}$, so $\Upsilon(\widetilde f_j (s_{i,\mu})) = t^f_{i,\mu}$ for each $j \le n$. By our assumptions on $\Upsilon$, we may fix $j < k \le n$ such that $| f_j (s_{i,\mu}, c) - f_k (s_{i,\mu}, c) | < 1/4$. Applying Condition (5) above, we have $| f_j (b, c) - f_k (b, c) | < 3/4$, contradicting $|f_j(b,c) - f_k(b,c)| \ge 1$.
In particular, letting $K$ be the $1$–point space, we see that an $I_\Lambda$ is not $4$–big. Then, proceeding by induction,
\[lemma-small-lotsprod\] $\prod_{j<m} L_{\Lambda_j}$ is not $(3^m + 1)$–big.
We remark that in the proof of Lemma \[lemma-small-big\], we could have replaced the “$<$” by “$\le$” in Cases I and III, although then they would not be disjoint from Case II. However, in the special case of $L = I_S$, Case II can now be eliminated, so that we can replace the “$(3n+1)$–big” by “$(2n+1)$–big”, obtaining:
\[lemma-smaller-big\] Let $L = I_S$, where $S \subseteq I$, and let $K$ be any compact space which is not $(n+1)$–big. Let $X = L \times K$. Then $X$ is not $(2n+1)$–big.
\[lemma-das-not-big\] $\prod_{j<m} I_{S_j}$ is not $(2^m + 1)$–big.
\[ex-big-limprop\] For any $n > 3$, there is an $X$ which does not have a Cantor subset and which is not $7$–big, such that $X$ is $(n + 2)$–superdissipated and not $(n + 1)$–superdissipated.
For $n \ge 1$, let $L_n = I_{\Lambda_n}$, where $\Lambda_n(x) = n$ for $n \in (0,1)$, and $\Lambda_n(0) = \Lambda_n(1) = 0$. Then $L_1$ is the double arrow space. Let $X = L_n \times L_1$. Then $X$ is not $7$–big by Lemmas \[lemma-small-lotsprod\] and \[lemma-smaller-big\]. $X$ is $(n + 2)$–dissipated by Lemma \[lemma-tight-prod\]. To prove that $X$ is not $(n + 1)$–dissipated, it is sufficient (by Lemma 3.6 of [@KU2]) to observe that for each $\varphi \in C(L_n, [0,1]^\omega)$ there is a $z \in [0,1]^\omega$ with $|\varphi{^{-1}}\{z\}| \ge n+1$.
It is easily seen using Lemma \[lemma-big-helper\] that $(I_S)^n$ is $2^n$–big when $S$ is uncountable, so it has the NTIP, since it is also $(2^{n-1} + 1)$–superdissipated. However, it is not clear whether it has the CSWP in the case that $S$ meets all Cantor sets, since the natural proof requires looking at arbitrary perfect subspaces of $(I_S)^n$.
\[ex-seven\] If $S \subseteq (0,1)$ meets all Cantor sets, then there is a perfect $X \subseteq (I_S)^4$ such that $X$ is not $8$–dissipated, is not $7$–big, and has no Cantor subsets.
Let $D \subset (I_S)^3$ be the diagonal, and let $X = D \times I_S$. Then $D$ is the same as the LOTS obtained from $I$ by replacing each point in $S$ by eight points. Since $7 = 2\cdot 3 + 1$, Lemmas \[lemma-small-lotsprod\] and \[lemma-smaller-big\] show that $X$ is not $7$–big. The proof of Example \[ex-big-limprop\] shows that $X$ is not $8$–dissipated.
This particular $X$ has the CSWP, and in fact is removable (see Section \[sec-remov\]), since $I_S$ is removable, and after removing the clopen copies of $I_S$ from $X$, we are left with a copy of $(I_S)^2$, which is also removable. We do not know whether $(I_S)^4$ itself must have the CSWP.
A simple example of $n$–big spaces is given by:
If $X$ is compact and $|X| > 2^{\aleph_0}$, then $X$ is $n$–big for all $n\in\omega$.
Fix ${{\mathcal A}}{\sqsubseteq}C(X)$, fix $n \in \omega$, and fix $\Upsilon : {{\mathcal A}}\to \omega$. We shall verify the conclusion of Definition \[def-big\] with $r = 1$.
Let ${{\mathbb P}}$ be the set of all finite partial functions from $X$ to $\omega$; so each $p \in {{\mathbb P}}$ is a function with ${\mathrm{dom}}(p)$ a finite subset of $X$ and ${\mathrm{ran}}(p) \subseteq \omega$. For $p \in {{\mathbb P}}$, choose an $f_p \in {{\mathcal A}}$ with $p \subset f_p$.
For each $c \in X$ and $s \in \omega$, let $E_{c,s} = \{p(c) :
p \in {{\mathbb P}}\ \&\ c \in {\mathrm{dom}}(p) \ \&\ \Upsilon(f_p)
= s\}\allowbreak\subseteq \omega$. If some $|E_{c,s}| \ge n$ then we are done, so assume that $|E_{c,s}| \le n-1$ for all $c,s$. There are only $ 2^{\aleph_0}$ possibilities for $\langle E_{c,s} : s \in \omega \rangle$, so we can fix an infinite $A \subseteq X$ and sets $E_s \in [\omega]^{<n}$ for $s\in \omega$ such that $E_{c,s} = E_s$ for all $c \in A$ and all $s \in \omega$. But then ${\mathrm{ran}}(p) \subseteq E_s$ whenever $p \in {{\mathbb P}}$ and $\Upsilon(f_p) = s$ and ${\mathrm{dom}}(p) \subseteq A$. Now choosing $p$ with ${\mathrm{dom}}(p) \subseteq A$ and $|{\mathrm{ran}}(p)| = n$ yields a contradiction.
Finally, the following Ramsey-type lemma might be of interest for studying products of LOTSes, although we never needed it in this paper. The proof uses the terminology from Definition \[def-boxes\].
Fix an uncountable $J \subseteq {{\mathbb R}}^n$, and assume that $$\forall x,y \in J \, [ x \ne y \to \forall i < n [x_i \ne y_i]]
\tag{$*$} \ \ .$$ Then there is a 1-1 function $\varphi: {{\mathbb Q}}\to J$ such that for all $i < n$: $$\forall p,q \, [ p < q \to \varphi(p)_i < \varphi(q)_i ]
\mbox{\ \ {\rm or} \ \ }
\forall p,q \, [ p < q \to \varphi(p)_i > \varphi(q)_i ] \ \ .
\tag{\dag}$$
Call a box $B = {\mathrm{box}}[a,b]$ *big* iff $B \cap J$ is uncountable; By $(*)$, this implies that $B^\circ \cap J$ is uncountable, where $B^\circ$ denotes the interior of $B$. For $\Delta \in {{\mathcal D}}$, let $-\Delta$ result from interchanging the signs $+$ and $-$ in $\Delta$. Call the box $B$ $\Delta$–*bad* iff $B$ is big and there is no $d \in B^\circ \cap J$ such that ${\mathrm{corn}}(d, B, \Delta)$ and ${\mathrm{corn}}(d, B, -\Delta)$ are both big. Observe, for any big box $B$:
$B$ is $\Delta$–bad iff $B$ is $(-\Delta)$–bad.
If $B$ is $\Delta$–bad and $A \subseteq B$ is a big box, then $A$ is $\Delta$–bad.
There is some $\Delta\in{{\mathcal D}}$ such that $B$ is not $\Delta$–bad.
\(1) and (2) are obvious. To prove (3), we note first that if we replaced ${{\mathbb Q}}$ by $\omega$ or a finite set in the statement of the lemma, then the result would be obvious by Ramsey’s Theorem. Now, let $Z$ be the set of points of $B^\circ \cap J$ which are condensation points of $J$. Obtain $\varphi: \{0,1,2\} \to Z$ so that () holds replacing ${{\mathbb Q}}$ by $\{0,1,2\}$. Let $a = \varphi(0)$, $b = \varphi(2)$, and $d = \varphi(1)$. By (), there is some $\Delta \in {{\mathcal D}}$ such that $a \in {\mathrm{corn}}(d, B, \Delta)$ and $b \in {\mathrm{corn}}(d, B, -\Delta)$, and then ${\mathrm{corn}}(d, B, \Delta)$ and ${\mathrm{corn}}(d, B, -\Delta)$ are both big.
Using (2) (sub-boxes go from bad to worse) and (3), we can fix a $\Delta \in {{\mathcal D}}$ and a big box $B$ such that for all big boxes $A \subseteq B$, $A$ is not $\Delta$–bad. We may now list ${{\mathbb Q}}$ in type $\omega$ and obtain $\varphi$ in $\omega$ steps. When $p < q$, we shall have $\varphi(p)_i < \varphi(q)_i $ when $\Delta_i = +1$ and $\varphi(p)_i > \varphi(q)_i $ when $\Delta_i = -1$.
[99]{}
P. Alexandroff and P. Urysohn, Mémoire sur les espaces topologiques compactes, *Verh. Akad. Wetensch. Amsterdam* 14 (1929) 1-96.
E. Bishop, A generalization of the Stone-Weierstrass theorem, *Pacific J. Math.* 11 (1961) 777-783.
T. W. Gamelin, *Uniform Algebras*, Second Edition, Chelsea Publishing Company, 1984.
J. B. Garnett, *Bounded Analytic Functions*, Academic Press, 1981.
J. Hart and K. Kunen, Complex function algebras and removable spaces, *Topology Appl.* 153 (2006) 2241-2259.
J. Hart and K. Kunen, Inverse limits and function algebras, *Topology Proceedings*, to appear.
J. Hart and K. Kunen, First countable continua and proper forcing, *Canadian Journal of Mathematics*, to appear.
K. Hoffman, *Banach Spaces of Analytic Functions*, Prentice-Hall, 1962.
K. Hoffman and I. M. Singer, Maximal algebras of continuous functions, *Acta Math.* 103 (1960) 217-241.
I. Juhász, [*Cardinal Functions in Topology – Ten Years Later*]{}, Mathematical Center Tracts \#123, Mathematisch Centrum, 1980.
K. Kunen, The complex Stone–Weierstrass property, *Fundamenta Mathematicae* 182 (2004) 151-167.
K. Kunen, Dissipated compacta, to appear; see *arXiv math.GN/0703429*.
D. J. Lutzer and H. R. Bennett, Separability, the countable chain condition and the Lindelöf property in linearly orderable spaces, *Proc. Amer. Math. Soc.* 23 (1969) 664-667.
W. Rudin, Subalgebras of spaces of continuous functions, *Proc. Amer. Math. Soc.* 7 (1956) 825-830.
W. Rudin, Continuous functions on compact spaces without perfect subsets, *Proc. Amer. Math. Soc.* 8 (1957) 39-42.
[^1]: University of Wisconsin, Madison, WI 53706, U.S.A., [email protected]
[^2]: Author partially supported by NSF Grant DMS-0456653.
[^3]: 2000 Mathematics Subject Classification: Primary 54C40, 46J10. Key Words and Phrases: Ordered space, function algebra.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this work, we introduce multiplicative drift analysis as a suitable way to analyze the runtime of randomized search heuristics such as evolutionary algorithms.
We give a multiplicative version of the classical drift theorem. This allows easier analyses in those settings where the optimization progress is roughly proportional to the current distance to the optimum.
To display the strength of this tool, we regard the classical problem how the (1+1) Evolutionary Algorithm optimizes an arbitrary linear pseudo-Boolean function. Here, we first give a relatively simple proof for the fact that any linear function is optimized in expected time $O(n \log n)$, where $n$ is the length of the bit string. Afterwards, we show that in fact any such function is optimized in expected time at most ${(1+o(1)) 1.39 {\mathrm{e}\xspace}n\ln (n)}$, again using multiplicative drift analysis. We also prove a corresponding lower bound of ${(1-o(1))e n\ln(n)}$ which actually holds for all functions with a unique global optimum.
We further demonstrate how our drift theorem immediately gives natural proofs (with better constants) for the best known runtime bounds for the (1+1) Evolutionary Algorithm on combinatorial problems like finding minimum spanning trees, shortest paths, or Euler tours.
author:
- Benjamin Doerr
- Daniel Johannsen
- |
Carola Winzen[^1]\
Max-Planck-Institut für Informatik\
Campus E1 4\
66123 Saarbrücken, Germany\
date: Submitted January 2011
title: Multiplicative Drift Analysis
---
Introduction {#sec:introduction}
============
An innocent looking problem is the question how long the (1+1) Evolutionary Algorithm ((1+1) EA) needs to find the optimum of a given linear function. However, this is in fact one of the problems that was most influential for the theory of evolutionary algorithms.
While particular linear functions like the functions [<span style="font-variant:small-caps;">OneMax</span>]{}and [<span style="font-variant:small-caps;">BinVal</span>]{}were easily analyzed, it took a major effort by Droste, Jansen and Wegener [@DrosteJW02] to solve the problem in full generality. Their proof, however, is highly technical.
A major breakthrough spurred by this problem is the work by He and Yao [@HeY01; @HeY04], who introduced *drift analysis* to the field of evolutionary computation. This allowed a significantly simpler proof for the linear functions problem. Even more important, it quickly became one of the most powerful tools for both proving upper and lower bounds on the expected optimization times of evolutionary algorithms. For example, see [@HeY04; @GielWstacs03; @GielL06; @HappJKN08; @NeumannOW09; @OlivetoW10].
Another great progress was made by Jägersküpper [@Jagerskupper08], who combined drift analysis with a clever averaging argument to determine reasonable values for the usually not explicitly given constants. More precisely, Jägersküpper showed that the expected optimization time of the (1+1) EA for any linear function defined on bit strings of length $n$ is bounded from above by $(1+o(1)) 2.02 {\mathrm{e}\xspace}n \ln(n)$.
Classical Drift Analysis
------------------------
The following method was introduced to the analysis of randomized search heuristics by He and Yao [@HeY04] and builds on a result of Hajek [@Hajek82]. When analyzing the optimization behavior of a randomized search heuristic over a search space, instead of tracking how the objective function improves, one uses an auxiliary potential function and tracks its behavior.
For example, consider the search space $\{0,1\}^n$ of bitstrings of length $n\in\mathbb{N}$.[^2] Suppose we want to analyze the (1+1) EA (which is introduced as Algorithm \[alg:oneoneea\] in Section \[sec:linear\]) minimizing a linear function $f\colon\{0,1\}^n\to\mathbb{R}$ with $$f(x)=\sum_{i=1}^n w_i x_i$$ and arbitrary positive weights $0<w_1\le\dots\le w_n$. (Note that we differ from previous works by always considering minimization problems. See Section \[sec:linear\] for a discussion why this does not influence the runtime analysis.) Then this potential function $h\colon \{0,1\}^n \to \mathbb{R}$ can be chosen as $$\label{eq:logpotential}
h(x)=\ln\Big(1+\sum_{j=1}^{\lfloor \frac{n}{2}\rfloor}x_j+\sum_{j=\lfloor \frac{n}{2}\rfloor +1}^n 2x_j\Big).$$
Though still needing some calculations, one can show the following (see, e.g., [@HeY04] where a variant of Algorithm \[alg:oneoneea\] is analyzed). Let $x \in \{0,1\}^n$. Let $y \in \{0,1\}^n$ be the result of one iteration (mutation and selection) of the (1+1) EA started in $x$. Then there exists a $\delta > 0$ such that $$\label{eq:drift}
{\ensuremath{\operatorname{E}}}[h(y)] \le h(x) - \delta/n.$$ Now, classical drift analysis tells us that in expectation after a number of $h(x)/(\delta/n) = O(n \log n)$ iterations, the potential value is reduced to zero. But $h(x)=0$ implies $f(x)=0$, that is, the (1+1) EA has found the desired optimum.
Using drift analysis to analyze a randomized search heuristic usually bears two difficulties. The first is guessing a suitable potential function $h$. The second, related to the first, is proving that during the search, $f$ and $h$ behave sufficiently similar, that is, we can prove some statement like inequality (\[eq:drift\]). Note that this inequality contains information about $f$ as well, namely implicitly in the fact that $y$ has an at least as good $f$–value as $x$.
A main difficulty in showing that $h$ in (\[eq:logpotential\]) is a suitable potential function is the logarithm around the simple linear function giving weights one and two to the bits. However, since the optimization progress for linear functions is faster if we are further away from the optimum, that is, have more one-bits, this seems difficult to avoid.
Multiplicative Drift Analysis
-----------------------------
We present a way to ease to use of drift analysis in such settings. Informally, our method applies if we have a potential function $g$ satisfying $$\label{eq:drift2}
{\ensuremath{\operatorname{E}}}[g(y)] \le (1 - \delta)g(x)$$ in the notation above. That is, we require a progress which *multiplicatively* depends on the current potential value. For this reason we call the method *multiplicative drift analysis*. We will see that for a number of problems such potential functions are a natural choice.
This new method allows us to largely separate the structural analysis of an optimization process from the actual calculation of a bound on the expected optimization time. Moreover, the runtime bounds obtained by multiplicative drift analysis are often sharper than those resulting from previously used techniques.
Our Results
-----------
We apply this new tool, multiplicative drift analysis, to the already mentioned problem of optimizing linear functions over $\{0,1\}^n$. This yields a simplified proof of the $O(n\log n)$ bound on the optimization time of the (1+1) EA. Similar to the proof using the classical drift theorem, we make use of the simple linear function $g\colon \{0,1\}^n \to \mathbb{R}$, chosen as $$g(x)=\sum_{i=1}^n \Big(1+\frac{i}{n}\Big)x_i.$$ This function $g$ serves us as a potential function for all linear functions $f\colon \{0,1\}^n \to \mathbb{R}$ with $$f(x)=\sum_{i=1}^n w_i x_i$$ and monotone weights $0<w_1\le\dots \le w_n$. Using parts of Jägersküpper’s analysis [@Jagerskupper08], we then improve his upper bound on the expected optimization time of the (1+1) EA on linear functions to $(1+o(1)) 1.39 {\mathrm{e}\xspace}n \ln(n)$.
We also give lower bounds for this problem. We show that, in the class of all functions with a unique global optimum, the function [<span style="font-variant:small-caps;">OneMax</span>]{}(see (\[eq:onemax\])) has the smallest expected optimization time. This extends the lower bound of $(1 - o(1)) {\mathrm{e}\xspace}n \ln(n)$ for the expected optimization time of the (1+1) EA on [<span style="font-variant:small-caps;">OneMax</span>]{} [@DoerrFW10] to all functions in that class (including all linear functions with non-zero coefficients).
Together with our upper bound, we thus obtain the remarkable result that all linear functions have roughly (within a 39% range) the same optimization time.
To further demonstrate the strength of multiplicative drift analysis, we give straight-forward analyses for three prominent combinatorial problems. We consider the problems of computing minimum spanning trees (MST), single-source shortest paths (SSSP), and Euler tours. Here, we reproduce the results obtained in [@NeumannW07] (cf. Theorem \[thm:mst\]), in [@BaswanaBDFKN09] (cf. Theorem \[thm:sssp\]), and in [@DoerrJ07] (cf. Theorem \[thm:euler\]), respectively. In doing so, we improve the leading constants of the asymptotic bounds.
Multiplicative Drift Analysis {#sec:drift}
=============================
Drift analysis can be used to track the optimization behavior of a randomized search heuristic over a search space by measuring the progress of the algorithm with respect to a *potential function*. Such a function maps each search point to a non-negative real number, where a potential of zero indicates that the search point is optimal.
\[Additive Drift [@HeY04]\] \[thm:additivedrift\] Let $ S\subseteq\mathbb{R}$ be a finite set of positive numbers and let $\{X^{(t)}\}_{t\in\mathbb{N}}$ be a sequence of random variables over $ S\cup\{0\}$. Let $T$ be the random variable that denotes the first point in time $t\in\mathbb{N}$ for which $X^{(t)}=0$.
Suppose that there exists a constant $\delta>0$ such that $$\label{eq:additivedrift}
{\ensuremath{\operatorname{E}}}\big[X^{(t)}-X^{(t+1)}{\,\big|\,}T>t\big]\ge \delta$$ holds. Then $${\ensuremath{\operatorname{E}}}\big[T{\,\big|\,}X^{(0)}\big] \le \frac{X^{(0)}}{\delta}.$$
This theorem tells us how to link the expected time at which the potential reaches zero to the first time the expected value of the potential reaches zero. If in expectation the potential decreases in each step by $\delta$ then after $X^{(0)}/\delta$ steps the expected potential is zero. Thus, one might expect the expected number of steps until the (random) potential reaches zero to be $X^{(0)}/\delta$, too. This is indeed the case in the setting of the previous theorem.
In order to apply the previous theorem to the analysis of randomized search heuristics over a (finite) search space $\mathcal{S}$, we define a potential function $h\colon\mathcal{S}\to\mathbb{R}$ which maps all optimal search points to zero and all non-optimal search points to values strictly larger than zero. We choose the random variable $X^{(t)}$ as the potential $h(x^{(t)})$ of the search point (or population) in the $t$-th iteration of the algorithm. Then the random variable $T$ becomes the optimization time of the algorithm, that is, the number of iterations until the algorithm finds an optimum.
When applying Theorem \[thm:additivedrift\], we call the expected difference between $h(x^{(t)})$ and $h(x^{(t+1)})$ the *drift* of the random process $\{x^{(t)}\}_{t\in\mathbb{N}}$ with respect to $h$. We say this drift is *additive* if condition (\[eq:additivedrift\]) holds.
Ideal Potential Functions for Additive Drift Analysis
-----------------------------------------------------
The application of additive drift analysis (Theorem \[thm:additivedrift\]) to the runtime analysis of randomized search heuristics requires a suitable potential function. The following lemma (Lemma 3 in [@HeY04]) tells us that if the random search points $x^{(0)},x^{(1)},x^{(2)},\dots$ generated by a search heuristic form a homogeneous absorbing Markov chain, then there always exists a potential function such that condition (\[eq:additivedrift\]) in Theorem \[thm:additivedrift\] holds with equality; namely the function that attributes to each search point the expected optimization time of the algorithm starting in that point.
\[lem:idealpotential\] Let $\mathcal{S}$ be a finite search space and $\{x^{(t)}\}_{t\in\mathbb{N}}$ the search points generated by a homogeneous absorbing Markov chain on $\mathcal{S}$. Let $T$ be the random variable that denotes the fist point in time $t\in\mathbb{N}$ such that $x^{(t)}$ is optimal.
Then the drift on the potential function $g\colon\mathcal{S}\to\mathbb{R}$ with $$g(x):={\ensuremath{\operatorname{E}}}[T\mid x^{(0)}=x]$$ satisfies $${\ensuremath{\operatorname{E}}}\big[g(x^{(t)})-g(x^{(t+1)}){\,\big|\,}T>t\big]=1.$$
In a way, ${\ensuremath{\operatorname{E}}}[T\mid x^{(0)}=x]$ is an “ideal” potential function for Theorem \[thm:additivedrift\]. It satisfies the additive drift condition (\[eq:additivedrift\]) with equality and results in precise upper bound on ${\ensuremath{\operatorname{E}}}[T\mid x^{(0)}=x]$. However, the previous theorem is not directly helpful in the runtime analysis of randomized search heuristics. In order to apply the previous theorem, we need to know the exact expected optimization time of a algorithm starting from every point in the search point. But with all this known, Theorem \[thm:additivedrift\] does not provide new information.
Still, the previous theorem indicates that potential functions which approximate the expected optimization time in the respective point are good candidates likely to satisfy the additive drift condition. In the next section, we will see such a potential function suitable for the analysis of the optimization behavior of the (1+1) Evolutionary Algorithm on linear functions.
A Multiplicative Drift Theorem
------------------------------
The drift theorem presented in this subsection can be considered as the multiplicative version of the classical additive result. Since we derive it from the original result, it is clear that the multiplicative version cannot be stronger than the original theorem.
\[Multiplicative Drift\] \[thm:multidrift\] Let $ S\subseteq\mathbb{R}$ be a finite set of positive numbers with minimum $s_{\min}$. Let $\{X^{(t)}\}_{t\in\mathbb{N}}$ be a sequence of random variables over $ S\cup\{0\}$. Let $T$ be the random variable that denotes the first point in time $t\in\mathbb{N}$ for which $X^{(t)}=0$.
Suppose that there exists a constant $\delta>0$ such that $$\label{eq:multidrift}
{\ensuremath{\operatorname{E}}}\big[X^{(t)}-X^{(t+1)}{\,\big|\,}X^{(t)}= s\big]\ge \delta s$$ holds for all $s\in S$ with ${\ensuremath{\mathrm{Pr}}}[X^{(t)}= s]>0$. Then for all $s_0\in S$ with ${\ensuremath{\mathrm{Pr}}}[X^{(0)}=s_0]>0$, $${\ensuremath{\operatorname{E}}}\big[T{\,\big|\,}X^{(0)}=s_0\big] \le \frac{1+\ln(s_0/s_{\min})}{\delta}.$$
Like for the notion of additive drift, we say that the drift of a random process $\{x^{(t)}\}_{t\in\mathbb{N}}$ with respect to a potential function $g$ is *multiplicative* if condition (\[eq:multidrift\]) holds for the associated random variables $x^{(t)}:=g(x^{(t)})$.
The advantage of the multiplicative approach is that it allows to use potential functions which are more natural. The most natural potential function, obviously, is the distance of the objective value of the current solution to the optimum. This often is a good choice in the analysis of combinatorial optimization problems. For example, in Section \[sec:combinatorial\] we see that the runtimes of the (1+1) EA on finding a minimum spanning tree, a shortest path tree, or an Euler tour can be bounded by analyzing this potential function.
Another potential function for which drift analysis has been successfully applied is the *distance in the search space* between the current search points and a (global) optimum.
The typical example for this is the drift analysis for linear functions in Section \[sec:linear\], where we use the (weighted) Hamming distance to the optimum as potential for all functions of this class. While being more difficult to analyze, this approach often gives tighter bounds which are independent of range of potential fitness values.
Note that multiplicative drift analysis applies to all situations where previously the so-called *method of expected weight decrease* was used. This method also builds on the observation that if the drift is multiplicative (that is, condition (\[eq:multidrift\]) holds), then at time $t=(1+\ln(s_0/s_{\min}))/\delta$ the expected potential $X^{(t)}$ is at most $s_0/{\mathrm{e}\xspace}$. Afterwards, various methods (variants of Wald’s identity in [@DrosteJW02; @Jagerskupper08] and Markov’s inequality in [@NeumannW07; @BaswanaBDFKN09]) are used to show that the expected stopping time ${\ensuremath{\operatorname{E}}}[T]$ is indeed in this regime. However, the bounds obtained in this way are not best possible. This is demonstrated in Section \[sec:jagerskupper\] where we replace for the proofs in [@Jagerskupper08] the method of expected weight decrease by the above multiplicative drift theorem. This results in an immediate improvement of the leading constant in the main runtime bound of [@Jagerskupper08].
Let $g\colon S\to\mathbb{R}$ be the function defined by $$g(s):=1+\ln\frac{s}{s_{\min}}.$$
Let $ R:=g(S)$ be the image of $g$ and let $\{Z^{(t)}\}_{t\in\mathbb{N}}$ be the sequence of random variables over $ R\cup\{0\}$ given by $$Z^{(t)}:=\begin{cases}
0 & \text{if }X^{(t)}=0,\\
g(X^{(t)}) & \text{otherwise.}
\end{cases}$$
Then $T$ is also the first point in time $t\in\mathbb{N}$ such that $Z^{(t)}=0$. Suppose $T>t$. Then we have $Z^{(t)}=g(X^{(t)})>0$. If $X^{(t+1)}=0$, then also $Z^{(t+1)}=0$ and $$\label{eq:multzero}
Z^{(t)}-Z^{(t+1)}=1+\ln\Big(\frac{X^{(t)}}{s_{\min}}\Big)\ge 1=\frac{X^{(t)}-X^{(t+1)}}{X^{(t)}}.$$ Otherwise, $Z^{(t+1)}=g(X^{(t+1)})$ and again $$\label{eq:multgeneral}
Z^{(t)}-Z^{(t+1)}=\ln\Big(\frac{X^{(t)}}{X^{(t+1)}}\Big)\ge\frac{X^{(t)}-X^{(t+1)}}{X^{(t)}},$$ where the last inequality follows from $$\frac{u}{w}=1+\frac{u-w}{w}\le {\mathrm{e}\xspace}{}^{\frac{u-w}{w}}$$ which implies $$\ln\Big(\frac{u}{w}\Big)\le\frac{u-w}{w}$$ and thus $$\ln\Big(\frac{w}{u}\Big)\ge\frac{w-u}{w}.$$ for all $u,w\in\mathbb{R}$.
Hence, by (\[eq:multzero\]) and (\[eq:multgeneral\]), independent of whether $Z^{(t+1)}=0$ or $Z^{(t+1)}\neq 0$, we have $$Z^{(t)}-Z^{(t+1)}\ge\frac{X^{(t)}-X^{(t+1)}}{X^{(t)}}.$$
Let $r\in R$. Since $g$ is bijective, there exist a unique $s\in S$ such that $r=g(s)$. Moreover, the events $Z^{(t)}=r$ and $X^{(t)}=s$ coincide. Hence, we have by condition (\[eq:multidrift\]) that $${\ensuremath{\operatorname{E}}}[Z^{(t)}-Z^{(t+1)}\mid Z^{(t)}=r]\ge\frac{{\ensuremath{\operatorname{E}}}[X^{(t)}-X^{(t+1)}\mid X^{(t)}=s]}{s}\ge\delta.$$
Finally, we apply Theorem \[thm:additivedrift\] for additive drift and obtain for $s\in S$ with ${\ensuremath{\mathrm{Pr}}}[X^{(0)}=s]>0$ that $${\ensuremath{\operatorname{E}}}[T\mid X^{(0)}=s]={\ensuremath{\operatorname{E}}}[T\mid Z^{(0)}=g(s)]\le\frac{g(s)}{\delta}\le\frac{1+\ln(s/s_{\min})}{\delta}$$ which concludes the proof of the theorem.
In Section \[sec:linear\] and Section \[sec:combinatorial\], we demonstrate the strength of this new tool by applying it to four well-known problems: the problem of minimizing linear pseudo-Boolean functions, the minimum spanning tree problem, the single-source shortest path problem, and the problem of finding Euler tours.
The Runtime of the (1+1) Evolutionary Algorithm on Pseudo-Boolean Functions {#sec:linear}
===========================================================================
Many optimization problems can be phrased as the problem of maximizing or minimizing a pseudo-Boolean function $f\colon \{0,1\}^n \to \mathbb{R}$ where $n$ is a positive integer. In the setting of randomized search heuristics, such a function $f$ is considered to be a black-box, that is, the optimization process can access $f$ only by evaluating it at limited number of points in $\{0,1\}^n$.
\[alg:oneoneea\] choose $x^{(0)}\in\{0,1\}^n$ uniformly at random
In this section, we analyze the (1+1) Evolutionary Algorithm ((1+1) EA) for pseudo-Boolean functions (Algorithm \[alg:oneoneea\]). This algorithm follows the neighborhood structure imposed by the hypercube on $\{0,1\}^n$ where two points are adjacent if they differ by exactly one bit, that is, if their Hamming distance is one. The (1+1) EA successively attempts to improve the so-far best search point by randomly sampling candidates over $\{0,1\}^n$ according to probabilities decreasing with the distance to the current optimum.
The *optimization time* of the (1+1) EA on a function $f$ is the random variable $T$ that denotes the first point in time $t\in\mathbb{N}$ such that $f(x^{(t)})$ is minimal.
One elementary linear pseudo-Boolean function for which the optimization time (1+1) EA has been analyzed (e.g., in [@Muehlenbein1992] and [@DrosteJW02]) is the function ${\textsc{OneMax}\xspace}\colon\{0,1\}^n\to\mathbb{N}$. This function simply counts the number of one-bits in $x$, that is, $$\label{eq:onemax}
{\textsc{OneMax}\xspace}(x):=|x|_1=\sum_{i=1}^n x_i.$$ Unlike indicated by the name of this function, we are interested in the time the (1+1) EA needs to find its minimum. Thus, in the selection step (Step 4) of each iteration, the (1+1) EA accepts the candidate solution $y^{(t)}$ if and only if the number of bits equal to $1$ does not increase.
Consider the progress $\Delta^{(t)}:={\textsc{OneMax}\xspace}(x^{(t)})-{\textsc{OneMax}\xspace}(x^{(t+1)})$ of the (1+1) EA in the $t$-th iteration. By construction of the (1+1) EA, $\Delta^{(t)}$ cannot be negative. By definition, the number of one-bits $x^{(t)}$ is ${\textsc{OneMax}\xspace}(x^{(t)})$. For each of these one-bits, there is a $(1/n)(1-1/n)^{n-1}\ge 1/({\mathrm{e}\xspace}n)$ chance that only this one-bit is flipped when sampling $y^{(t)}$, thus increasing the value of ${\textsc{OneMax}\xspace}(x^{(t)})$ by one. Hence, $${\ensuremath{\operatorname{E}}}\big[\Delta^{(t)}{\,\big|\,}x^{(t)}\big]\ge\frac{{\textsc{OneMax}\xspace}(x^{(t)})}{{\mathrm{e}\xspace}n}.$$
Thus, multiplicative drift analysis (Theorem \[thm:multidrift\]) immediately gives us the well-known result $${\ensuremath{\operatorname{E}}}[T_{{\textsc{OneMax}\xspace}}]\le{\mathrm{e}\xspace}n\Big(1+\ln{\ensuremath{\operatorname{E}}}\big[{\textsc{OneMax}\xspace}(x^{(0)})\big]\Big)={\mathrm{e}\xspace}n\Big(1+\ln\Big(\frac{n}{2}\Big)\Big).$$
Another elementary linear pseudo-Boolean function is [<span style="font-variant:small-caps;">BinVal</span>]{}. This function maps a bitstring to the binary value it represents (where $x_1$ represents the lowest and $x_n$ the highest bit). $$\label{eq:binval}
{\textsc{BinVal}\xspace}(x)=\sum_{i=1}^n{2^{i-1} x_i}.$$
Again, for $\Delta^{(t)}:={\textsc{BinVal}\xspace}(x^{(t)})-{\textsc{BinVal}\xspace}(x^{(t+1)})$, we have $${\ensuremath{\operatorname{E}}}\big[\Delta^{(t)}{\,\big|\,}x^{(t)}\big]\ge\frac{{\textsc{BinVal}\xspace}(x^{(t)})}{{\mathrm{e}\xspace}n}$$ and thus $${\ensuremath{\operatorname{E}}}[T_{{\textsc{BinVal}\xspace}}]\le{\mathrm{e}\xspace}n\Big(1+\ln{\ensuremath{\operatorname{E}}}\big[{\textsc{BinVal}\xspace}(x^{(0)})\big]\Big)={\mathrm{e}\xspace}n\Big(1+\ln\Big(\frac{2^n-1}{2}\Big)\Big).$$
Note, that the previous inequality gives us only a quadratic upper bound of ${\ensuremath{\mathord{O}}}(n^2)$ for the expected optimization time of the (1+1) EA on [<span style="font-variant:small-caps;">BinVal</span>]{}. However, it is known that for all linear functions — including [<span style="font-variant:small-caps;">BinVal</span>]{}— the expected optimization time of the (1+1) EA is ${\ensuremath{\mathord{O}}}(n\ln n)$. We discuss this in the following subsections and give a simplified proof using multiplicative drift analysis.
Linear Functions
----------------
A classical test problem for the runtime analysis of randomized search heuristics is the minimization of *linear functions*.
Let $n\in\mathbb{N}$ be a positive integer. A function $f\colon\{0,1\}^n\to\mathbb{R}$ on $n$ bits is *linear*, if there exists weights $w_1,\dots w_n \in \mathbb{R}$ such that $$f(x)=\sum_{i=1}^n w_i x_i$$ for all $x\in\{0,1\}^n$. In [@DrosteJW02] it has been argued and it is easily seen that in the analysis of upper bounds of the expected optimization time of the (1+1) EA on linear functions we may assume without loss of generality that the weights $w_i$ are all positive and sorted, that is, $$\label{eq:monotone}
0<w_1\le w_2\le\dots\le w_n.$$ We simply call such weights *monotone*. Moreover, for the runtime bounds we consider in this work it does not matter whether the (1+1) EA minimizes or maximizes the linear function. This is true since maximizing a function $f$ is equivalent to minimizing $-f$ and vice versa (for $-f$ we again have to invoke above argument which allows us to assume monotonicity of the weights).
Thus, from now on, we suppose that every linear function satisfies condition (\[eq:monotone\]). Furthermore, we formulate all results for the minimization problem, even if the referenced results originally considered the problem of maximizing linear functions.
We have already seen two prominent examples of linear functions, namely the functions [<span style="font-variant:small-caps;">OneMax</span>]{}and [<span style="font-variant:small-caps;">BinVal</span>]{}. When minimizing [<span style="font-variant:small-caps;">OneMax</span>]{}, the (1+1) EA accepts a new bit string in the selection step (Step 4) if the number of one-bits did not increase. In contrast, when minimizing [<span style="font-variant:small-caps;">BinVal</span>]{}, the inequality $2^k > \sum_{i=1}^{k-1}{2^{i-1}}$ implies that the (1+1) EA accepts a new bit string if and only if the highest-index bit that is touched in the mutation step (Step 3) is flipped from one to zero.
In spite of this difference in behavior, Droste, Jansen and Wegener showed in their seminal paper [@DrosteJW02] that for all linear functions the expected optimization time of the (1+1) EA is $\Theta(n \log n)$.
\[thm:ealinear\] For all positive integers $n\in\mathbb{N}$, the expected running time of the (1+1) EA on the class of linear functions with non-zero weights is $\Theta(n \log n)$.
The proof of Droste, Jansen and Wegener applies a level based argument to the potential function (called *artificial fitness function*) $g\colon \{0,1\}^n \to \mathbb{R}$ such that for all $x\in\{0,1\}^n$ $$\label{eq:droste}
g(x)=\sum_{i=1}^{\lfloor \frac{n}{2}\rfloor}x_i+\sum_{i=\lfloor \frac{n}{2}\rfloor +1}^n 2x_i.$$
A much easier proof avoiding partitioning arguments and instead working completely in the framework of drift analysis, was given by He and Yao in [@HeY04]. There, additive drift analysis is applied to the potential function $\widetilde{g}\colon \{0,1\}^n \to \mathbb{R}$ such that $$\label{eq:heyao}
\widetilde{g}(x)=\ln\Big(1+\sum_{i=1}^{\lfloor \frac{n}{2}\rfloor}x_i+\sum_{i=\lfloor \frac{n}{2}\rfloor +1}^n c x_i\Big)$$ for all $x\in\{0,1\}^n$. For this function, with $1<c\le 2$ chosen arbitrarily, they show that for all $x\in\{0,1\}^n\setminus\{(0,\dots,0)\}$ $${\ensuremath{\operatorname{E}}}\big[\widetilde{g}(x^{(t)})-\widetilde{g}(x^{(t+1)}){\,\big|\,}x^{(t)}=x\big]=\Omega(1/n).$$ Afterwards, they apply Theorem \[thm:additivedrift\] to show Theorem \[thm:ealinear\]. However, while this approach strongly reduced the complexity of the proof in [@DrosteJW02], introducing the natural logarithm into the potential function still resulted in unnecessary case distinctions and even inconsistencies in an early version of the proof [@HeY01; @HeY02AI].
The Drift for Linear Functions is Multiplicative
------------------------------------------------
In this subsection, we give a simple proof of the fact that the (1+1) EA optimizes any linear function in expected time ${\ensuremath{\mathord{O}}}(n \log n)$. Our proof is based on the theorem of multiplicative drift (Theorem \[thm:multidrift\]). Although proofs for Theorem \[thm:ealinear\] are known [@DrosteJW02; @HeY04; @Jagerskupper08], we present this alternative approach to demonstrate the strength of the multiplicative version of the classical drift theorem.
In order to apply Theorem \[thm:multidrift\] we need a suitable potential function. For this, we choose the function $g\colon \{0,1\}^n \to \mathbb{R}$ such that[^3]
$$g(x)=\sum_{i=1}^n\Big(1+\frac{i}{n}\Big)x_i$$
for all $x\in\{0,1\}^n$. This function defines the potential as the *weighted* distance of the current search point to the optimum (the all-zero string) in the search space. More precisely, it counts the number of one-bits, where each bit is assigned a weight between one and two, such that bits which have higher weight in the objective function $f$ also have higher weight in $g$.
We show that the drift of the (1+1) EA with respect to $g$ is multiplicative, that is, that condition (\[eq:multidrift\]) holds.
\[lem:pointwise\] Let $n\in\mathbb{N}$ be a positive integer. Let $f\colon\{0,1\}^n\to\mathbb{R}$ be a linear function with monotone weights and let $g\colon\{0,1\}^n\to\mathbb{R}$ be the potential function with $g(x)=\sum_{i=1}^n(1+i/n)x_i$ for all $x\in\{0,1\}^n$.
Let $x\in\{0,1\}^n$ and let $y\in\{0,1\}^n$ be randomly chosen by flipping each bit in $x$ with probability $1/n$. Let $\Delta(x):=g(x)-g(y)$ if $f(y)\le f(x)$ and $\Delta(x)=0$ otherwise. Then $${\ensuremath{\operatorname{E}}}[\Delta(x)]\ge\frac{g(x)}{4{\mathrm{e}\xspace}n}.$$
This lemma implies that *at every point in the search space* the drift is at least linear in the current potential value. Thus, the multiplicative drift condition (\[eq:multidrift\]) holds and Theorem \[thm:ealinear\] follows directly by applying Theorem \[thm:multidrift\].
Since ${\ensuremath{\operatorname{E}}}[\Delta(x)\mid f(y)>f(x)]=0$, we have by the law of total expectation that $$\label{eq:accepted}
{\ensuremath{\operatorname{E}}}[\Delta(x)]={\ensuremath{\operatorname{E}}}[g(x)-g(y)\mid f(y)\le f(x)]\,{\ensuremath{\mathrm{Pr}}}[f(y)\le f(x)].$$
Let $I=\{i\in\{1,\dots,n\}\colon x_i=1\}$. We may distinguish three events (cases).
- There is no index $i\in I$ such that $y_i=0$ and $f(y)\le f(x)$ holds, that is, $x=y$.
- There is exactly one index $i\in I$ such that $y_i=0$ and $f(y)\le f(x)$ holds.
- There are at least two different indices $j,\ell\in I$ such that $y_j=0$ and $y_\ell=0$ and $f(y)\le f(x)$ holds.
The only possibility for the event ($C_1$) to hold is if $x=y$. Therefore, $$\label{eq:casea}
{\ensuremath{\operatorname{E}}}[g(x)-g(y)\mid (C_1)]=0.$$
Next, suppose the event ($C_3$) holds. By linearity of expectation, we have $${\ensuremath{\operatorname{E}}}[g(x)-g(y)\mid (C_3)]=\sum_{i=1}^n
{\ensuremath{\operatorname{E}}}[g(x_i)-g(y_i)\mid (C_3)].$$
On the one hand, the event ($C_3$) implies that there are (at least) two indices $j$ and $\ell$ in $\{1,\dots,n\}$ for which $x_j=x_\ell=1$ and $y_j=y_\ell=1$. Since $g_j\ge 1$ and $g_\ell\ge 1$, we have $$\sum_{i\in I}{\ensuremath{\operatorname{E}}}[g(x_i)-g(y_i)\mid (C_3)]\ge 2.$$
On the other hand, if $i\in\{1,\dots,n\}\setminus I$ then $${\ensuremath{\operatorname{E}}}[g(x_i)-g(y_i)\mid (C_3)]=-g_i{\ensuremath{\mathrm{Pr}}}[y_1=0\mid(C_3)]\ge -\frac{g_i}{n},$$ since the condition ($C_3$) does not increase the probability of $1/n$ that the $y_i=0$. Therefore, since the $g_i$’s are at most two, we have $$\label{eq:casec}
{\ensuremath{\operatorname{E}}}[g(x)-g(y)\mid (C_3)]\ge 2-\frac{1}{n}\sum_{i\notin I}g_i\ge 0.$$
Therefore, by the law of total expectation and by (\[eq:accepted\]), (\[eq:casea\]) and (\[eq:casec\]), we have $$\label{eq:exactone}
{\ensuremath{\operatorname{E}}}[\Delta(x)]\ge{\ensuremath{\operatorname{E}}}[g(x)-g(y)\mid (C_2)]\,{\ensuremath{\mathrm{Pr}}}[(C_2)]$$ and can focus on the event ($C_2$).
Suppose that ($C_2$) holds. For every $i\in I$, we distinguish two events:
- The $i$-th bit is the only one-bit in $x$ that flips, none of the zero-bits at the positions larger than $i$ flips, and $f(y)\le f(x)$ holds.
- The $i$-th bit is the only one-bit in $x$ that flips, at least one of the zero-bits at the positions larger than $i$ that flips, and $f(y)\le f(x)$ holds.
We substitute the right side in (\[eq:exactone\]) and obtain $$\label{eq:ab}
{\ensuremath{\operatorname{E}}}[\Delta(x)]\ge\sum_{i\in I}{\ensuremath{\operatorname{E}}}[\Delta(x)\mid (A_i)]\,{\ensuremath{\mathrm{Pr}}}[(A_i)]+{\ensuremath{\operatorname{E}}}[\Delta(x)\mid (B_i)]\,{\ensuremath{\mathrm{Pr}}}[(B_i)]$$
Let $i\in I$ and suppose that the condition $(A_i)$ holds. Then we have $y_i=0$ and $y_j=x_j$ for all $j>i$. For a lower bound on ${\ensuremath{\operatorname{E}}}[\Delta(x)\mid (A_i)]$, we may suppose that $x_j=0$ for all $j<i$ and that every flip of a bit with index $j<i$ is accepted. Therefore, since $i\le n$ $${\ensuremath{\operatorname{E}}}[\Delta(x)\mid (A_i)]\ge 1+\frac{i}{n}-\sum_{j=1}^{i-1}\frac{1}{n}\Big(1+\frac{j}{n}\Big)=1+\frac{1}{n}-\frac{i(i-1)}{2n^2}\ge 1-\frac{i-3}{2n}.$$ and thus ${\ensuremath{\operatorname{E}}}[\Delta(x)\mid (A_i)]$ is positive. Furthermore, ${\ensuremath{\mathrm{Pr}}}[(A_i)]\ge\frac{1}{n}\big(1-\frac{1}{n}\big)^{n-1}$ which is the probability that only the $i$-th bit flips. Hence, $$\label{eq:ai}
{\ensuremath{\operatorname{E}}}[\Delta(x)\mid (A_i)]\,{\ensuremath{\mathrm{Pr}}}[(A_i)]\ge \frac{1}{n}\Big(1-\frac{1}{n}\Big)^{n-1}\Big(1-\frac{i-3}{2n}\Big).$$
Next, suppose that condition $(B_i)$ holds. Then we have $y_i=0$ and $y_\ell=1$ for all $\ell\in I\setminus\{i\}$, and there exists a $j>i$ with $j\notin I$ such that $y_j=1$. In order to satisfy $f(y)\le f(x)$, $w_j=w_i$ has to hold. This implies $x_\ell=y_\ell$ for all $\ell\in\{1,\dots,n\}\setminus\{i,j\}$. To see this, recall that the $w_i$’s are monotone and we condition on the event that the $i$-th bit is the only bit that flips from one to zero.
Let $J(i)=\{j\in\{i+1,\dots,n\}\colon x_j=0\text{ and }w_j=w_i\}$. For $j\in J(i)$ let $B_{i,j}$ be the event that $y_i=0$, $y_j=1$, and $y_\ell=x_\ell$ for $\ell$ not $i$ or $j$. Then $${\ensuremath{\operatorname{E}}}[\Delta(x)\mid (B_i)]\,{\ensuremath{\mathrm{Pr}}}[(B_i)]=\sum_{j\in J(i)}{\ensuremath{\operatorname{E}}}[\Delta(x)\mid B_{i,j}]\,{\ensuremath{\mathrm{Pr}}}[B_{i,j}].$$ We substitute ${\ensuremath{\operatorname{E}}}[\Delta(x)\mid B_{i,j}]=-\frac{j-i}{n}$ and ${\ensuremath{\mathrm{Pr}}}[B_{i,j}]=\frac{1}{n^2}\big(1-\frac{1}{n}\big)^{n-2}$ in the previous equation. Since these conditional expectations are always negative, we may pessimistically assume that $J(i)=\{i+1,\dots,n\}$ and get $$\sum_{j\in J(i)}{\ensuremath{\operatorname{E}}}[\Delta(x)\mid B_{i,j}]=-\frac{(n-i)}{n}\,\frac{(n+1-i)}{2}\ge -\Big(1-\frac{1}{n}\Big)\frac{(n+1-i)}{2}$$ and therefore $$\label{eq:bi}
{\ensuremath{\operatorname{E}}}[\Delta(x)\mid (B_i)]\,{\ensuremath{\mathrm{Pr}}}[(B_i)]\ge -\frac{1}{n}\Big(1-\frac{1}{n}\Big)^{n-1}\frac{n+1-i}{2n}.$$
Finally, we substitute (\[eq:ai\]) and (\[eq:bi\]) in (\[eq:ab\]) and derive $${\ensuremath{\operatorname{E}}}[\Delta(x)]\ge\frac{1}{n}\Big(1-\frac{1}{n}\Big)^{n-1}\sum_{i\in I}
1-\frac{i-3}{2n}-\frac{n+1-i}{2n}=\frac{1}{n}\Big(1-\frac{1}{n}\Big)^{n-1}\frac{n+2}{4n}\sum_{i\in I}2.$$ Since $g_i=1+i/n\le 2$ for all $i\in I$, we have $\sum_{i\in I}2\ge g(x)$ and therefore $${\ensuremath{\operatorname{E}}}[\Delta(x)]\ge \frac{g(x)}{4{\mathrm{e}\xspace}{}n}$$ which concludes the proof of the lemma.
Distribution-based Versus Point-wise Drift {#sec:jagerskupper}
------------------------------------------
In this subsection we show an almost tight upper bound on the expected optimization time of the (1+1) EA on linear functions.
If we take a closer look at Lemma \[lem:pointwise\], we see that it holds *point-wise*, that is, it guarantees $$\label{eq:pointwise}
{\ensuremath{\operatorname{E}}}\big[g(x^{(t)})-g(x^{t+1}){\,\big|\,}x^{(t)}=x\big]\ge\frac{g(x)}{4{\mathrm{e}\xspace}n}$$ for all $x\in\{0,1\}^n\setminus\{(0,\dots,0)\}$. This is far stronger than the positive *average* drift condition (\[eq:multidrift\]) which only requires
$$\label{eq:average}
{\ensuremath{\operatorname{E}}}\big[g(x^{(t)})-g(x^{t+1}){\,\big|\,}g(x^{(t)})=s, T>t\big]\ge\delta s$$
for all $s\in\mathbb{R}$ such that ${\ensuremath{\mathrm{Pr}}}[g(x^{(t)})=s, T>t]>0$.
The advantage of the stronger point-wise drift assumption is that it immediately guarantees that the result of Theorem \[thm:ealinear\] holds for all initial individuals.
The main reason, however, for not using the weaker condition (\[eq:average\]) is that this requires a deeper understanding of the probability distribution of $x^{(t)}$.
Let us stress that finding a potential function satisfying the stronger point-wise drift condition is usually very tricky. For example, one may ask why not take ${\textsc{OneMax}\xspace}(x)$ as potential function to bound the expected optimization time of the (1+1) EA for minimizing linear functions.
However, an easy observation reveals that there is an objective function $f$ and a search point $x$ such that $g$ yields to small a drift with respect to $f$. To see this, let $x=(x_1,\dots,x_n):=(0,\dots,0,1)$ and let $f:={\textsc{BinVal}\xspace}$ be the function to be minimized. Then the point-wise drift (\[eq:pointwise\]) with respect to ${\textsc{OneMax}\xspace}$ is only $1/n^2$. This example shows that finding a potential function yielding point-wise drift for all $x$ and all $f$ may be difficult. This observation is not to be confused with that in the discussion following (\[eq:binval\]). There, we determined the drift using the function [<span style="font-variant:small-caps;">BinVal</span>]{}itself as potential. Here, we use [<span style="font-variant:small-caps;">OneMax</span>]{}, that is, the 1-norm as potential function.
J[ä]{}gersk[ü]{}pper [@Jagerskupper08] was the first to overcome the difficulties of point-wise drift. While he still avoids completely analyzing the actual distribution of $x^{(t)}$, he does show the following property of this distribution which in turn allows him to use an average drift approach. In this way, he omits the need for point-wise drift. Jägersküpper’s simple observation is that at any time step $t$, the more valuable bits are more likely to be in the right setting (cf. Theorem 1 in [@Jagerskupper08]).
\[thm:jagerskupper\] Let $n\in\mathbb{N}$ be a positive integer and let $x^{(t)}$ denote the random individual (distributed over $\{0,1\}^n$) after $t\in\mathbb{N}$ iterations of the (1+1) EA minimizing a linear function $f\colon\{0,1\}^n\to\mathbb{R}$. Then $${\ensuremath{\mathrm{Pr}}}[x^{(t)}_1 = 0]\le\dots\le{\ensuremath{\mathrm{Pr}}}[x^{(t)}_n = 0].$$ Moreover, for all $k\in\{0,\dots,n\}$, this statement remains true if we condition on ${\textsc{OneMax}\xspace}(x)=k$.
Using this theorem, he was able to show a lower bound of ${\Omega(1/n)}$ for the drift of [<span style="font-variant:small-caps;">OneMax</span>]{}as potential function for any linear function.
\[lem:drift\] Let $n\in\mathbb{N}$ be a positive integer and let $f\colon\{0,1\}^n\to\mathbb{N}$ be a linear function. Let $x^{(t)}$ be the individual in the $t$-th iteration of the (1+1) EA minimizing $f$. Then $${\ensuremath{\operatorname{E}}}[{\textsc{OneMax}\xspace}(x^{(t)})-{\textsc{OneMax}\xspace}(x^{(t+1)})\mid {\textsc{OneMax}\xspace}(x^{(t)})=k]\ge\frac{({\mathrm{e}\xspace}-2)k}{{\mathrm{e}\xspace}n}.$$ holds for all $k\in\{0,\dots,n\}$ and $t\in\mathbb{N}$.
In addition to a more natural proof of the $O(n \ln(n))$ bound for expected optimization time of the (1+1) EA minimizing a linear function, J[ä]{}gersk[ü]{}pper was able to give a meaningful upper bound on the leading constant (cf. Theorem 2 in [@Jagerskupper08]).
\[thm:upperjagerskupper\] For all positive integers $n\in\mathbb{N}$, the expected optimization time of the (1+1) EA minimizing a linear function on $n$ bits is at most of order $(1+o(1)) 2.02 {\mathrm{e}\xspace}n\ln(n)$.
Using multiplicative drift analysis (Theorem \[thm:multidrift\]) on the result of Lemma \[lem:drift\] and thus replacing the halving argument employed by J[ä]{}gersk[ü]{}pper for the proof of Theorem \[thm:upperjagerskupper\], the constant of $2.02 {\mathrm{e}\xspace}$ in the upper bound of the previous theorem instantly improves to $1.39 {\mathrm{e}\xspace}$. In the light of our lower bound of $1.00{\mathrm{e}\xspace}$, to be proven in the next subsection, this is a considerable progress.
\[thm:upperbound\] For all positive integers $n\in\mathbb{N}$, the expected optimization time of the (1+1) EA minimizing a linear function on $n$ bits is at most of order $(1+o(1))\frac{{\mathrm{e}\xspace}}{{\mathrm{e}\xspace}-2}n\ln(n)\approx (1+o(1)) 1.39 {\mathrm{e}\xspace}n\ln(n)$.
The (1+1) EA Optimizes OneMax Faster than any Function with a Unique Global Optimum
-----------------------------------------------------------------------------------
In this section, we show that the expected optimization time of (1+1) EA on any pseudo-Boolean function with a unique global optimum is at least as large as its expected optimization time on the basic function [<span style="font-variant:small-caps;">OneMax</span>]{}. In particular, this is true for every linear function with non-zero coefficients.
In other words, if a function is easier to optimize than [<span style="font-variant:small-caps;">OneMax</span>]{}, then this can only be due to the fact that it has more than one global optimum. The general lower bound then follows from the following theorem by Doerr, Fouz and Witt [@DoerrFW10], which provides a lower bound for [<span style="font-variant:small-caps;">OneMax</span>]{}.
\[thm:LBOnemax\] For all positive integers $n\in\mathbb{N}$, the expected optimization time of the (1+1) EA minimizing [<span style="font-variant:small-caps;">OneMax</span>]{}on $n$ bits is at least $(1-{\ensuremath{\mathord{o}}}(1)) {\mathrm{e}\xspace}n\ln(n)$.
Thus, it remains to show that [<span style="font-variant:small-caps;">OneMax</span>]{}is optimized fastest. The result itself was announced by Scheder and Welzl [@SchederW08]. Their idea to prove this statement, however, differs from the one given below.
\[thm:lowerbound\] Let $n\in\mathbb{N}$ be a positive integer. The expected optimization time of the (1+1) EA on any function $f\colon\{0,1\}^n\to\mathbb{R}$ that has a unique global optimum is as least as large as its expected optimization time on [<span style="font-variant:small-caps;">OneMax</span>]{}.
The theorem can be formalized as follows: Let $f$ be a function with a unique global optimum. Let $\{x^{(t)}\}_{t \in \mathbb{N}}$ be the search points generated by the (1+1) EA minimizing $f$. Let $T_f:=\min\{t \in \mathbb{N}\mid f(x^{(t)}) =0\}$ be the optimization time of the (1+1) EA on $f$. Then ${\ensuremath{\operatorname{E}}}[T_f] \geq {\ensuremath{\operatorname{E}}}[T_{{\textsc{OneMax}\xspace}}].$
Theorems \[thm:LBOnemax\] and \[thm:lowerbound\] immediately yield the following.
\[cor:LBLinear\] For all positive integers $n\in\mathbb{N}$, the expected optimization time of the (1+1) EA minimizing a function with a unique global optimum on $n$ bits is at least $(1-{\ensuremath{\mathord{o}}}(1)) {\mathrm{e}\xspace}n\ln(n)$.
For the proof of Theorem \[thm:lowerbound\] we first show a preliminary lemma. It formalizes the following intuition. Let $x$ and $\tilde{x}$ be two search points such that $|x|_1\le|\tilde{x}|_1$. Then the probability that the (1+1) EA samples a new search point with exactly $j<|x|_1$ one-bits from $x$ is at least as big as from $\tilde{x}$.
\[lem:probabilities\] Let $n\in\mathbb{N}$ with $n \ge 1$. Let $x,\tilde{x} \in \{0,1\}^n$ with $|x|_1 \le |\tilde{x}|_1$. Let $y$ and $\tilde{y}$ two random points in $\{0,1\}^n$ obtained from $x$ and $\tilde{x}$ by independently flipping with probability $1/n$ each bit of $x$ and $\tilde{x}$, respectively.
Then for every $j\in\{0,\dots,|x|_1-1\}$, $${\ensuremath{\mathrm{Pr}}}\big[ |y|_1 = j\big] \ge {\ensuremath{\mathrm{Pr}}}\big[ |\tilde{y}|_1 = j\big] .$$
Let $k:=|x|_1$. The lemma holds trivially if $|\tilde{x}|_1=k$.
Suppose that $|\tilde{x}|_1=k+1$. Then $${\ensuremath{\mathrm{Pr}}}\big[ |y|_1 = j\big]=\hspace{-1em}
\sum_{i=0}^{\min\{j,n-k\}}\hspace{-0.4em}\binom{k}{j-i} \binom{n-k}{i}\Big(\frac{1/n}{1-1/n}\Big)^{k-j+2i}\big(1-1/n\big)^n$$ and $${\ensuremath{\mathrm{Pr}}}\big[ |\tilde{y}|_1 = j\big]=\hspace{-1em}
\!\sum_{i=0}^{\min\{j,n-k-1\}}\hspace{-0.4em}\binom{k+1}{j-i} \binom{n-k-1}{i}\Big(\frac{1/n}{1-1/n}\Big)^{k+1-j+2i}\big(1-1/n\big)^n.$$ As all summands in the previous two equations are positive, it suffices to see that the quotient $$\label{ineq:binom}
\frac{\binom{k}{j-i} \binom{n-k}{i}\big(\frac{1/n}{1-1/n}\big)^{k-j+2i}}{\binom{k+1}{j-i} \binom{n-k-1}{i} \big(\frac{1/n}{1-1/n}\big)^{k+1-j+2i}}
=\frac{(k+1-j+i)(n-k)(n-1)}{(k+1)(n-k-i)}$$ is minimal for $i=0$ and $j=k-1$ and therefore at least 1 for all values $0\leq i \leq \min\{j,n-k-1\}$.
Thus, for $|\tilde{x}|_1=k+1$ the lemma also holds. Finally, for $|\tilde{x}|_1>k+1$, the lemma follows by induction based on the case $|\tilde{x}|_1=k+1$.
To prove the main result of this section, Theorem \[thm:lowerbound\], we need some additional notation. Let $f$ be a function with a unique global optimum $x^*$. Without loss of generality, we may assume that $x^*:=(0,\dots,0)$ is the unique minimum of $f$. This is justified by the observation that the (1+1) EA treats the bit-values $0$ and $1$ symmetrically, that is, we might reinterpret one-bits in $x^*$ as zero-bits without changing the behavior of the algorithm.
Let $\mu(x) := {\ensuremath{\operatorname{E}}}[T_{{\textsc{OneMax}\xspace}} \mid x^{(0)}=x]$ and $\widetilde{\mu}(x) := {\ensuremath{\operatorname{E}}}[T_f \mid x^{(0)}=x]$ be the expected optimization times of the (1+1) EA starting in the point $x$ and minimizing [<span style="font-variant:small-caps;">OneMax</span>]{}and $f$, respectively.
For every $k \in\{0\dots,n\}$ let $$\mu_k := \min \{\mu(x) \mid x\in\{0,1\}^n, |x|_1 = k\}$$ be the optimization time of the (1+1) EA optimizing [<span style="font-variant:small-caps;">OneMax</span>]{}starting in a point with exactly $k$ one-bits.
Furthermore, let $$\widetilde{\mu}_k := \min \{\widetilde{\mu}(x) \mid x\in\{0,1\}^n, |x|_1 \geq k\}$$ be the minimum optimization time of the (1+1) EA minimizing $f$ and starting in a point $x$ with at *least* $k$ one-bits (note the difference to $\mu_k$).
Note that, due to the symmetry of the function [<span style="font-variant:small-caps;">OneMax</span>]{}, $\mu_k = \mu(x)$ for every $x\in\{0,1\}^n$ with exactly $k$ one-bits.
We inductively show for all $k$ that $\mu_k\le\widetilde{\mu}_k$. Clearly $\mu_0=0=\widetilde{\mu}_0$. Therefore, let $k \in \{0,\dots,n-1\}$ and suppose that $\mu_i\le\widetilde{\mu}_i$ for all $i\leq k$.
Let $x\in\{0,1\}^n$ with $|x|_1=k+1$ be arbitrary and let $y\in\{0,1\}^n$ be a random point generated by flipping each bit in $x$ independently with probability $1/n$.
The (1+1) EA minimizing [<span style="font-variant:small-caps;">OneMax</span>]{}and starting in $x$ accepts $y$ in the selection step (Step 4) if and only if $|y|_1 \leq |x|_1$. Furthermore, we have $\mu(x)=\mu(y)$ if $|y|_1=|x|_1=k+1$. Thus, $$\mu(x) = 1
+ \mu(x){\ensuremath{\mathrm{Pr}}}\big[|y|_1\ge k+1\big]
+ \sum_{j=0}^k{\ensuremath{\operatorname{E}}}\big[\mu(y){\,\big|\,}|y|_1=j\big]\,{\ensuremath{\mathrm{Pr}}}\big[|y|_1=j\big]$$ and therefore $$\label{eq:muonemax}
\mu_{k+1} = 1
+ \mu_{k+1} {\ensuremath{\mathrm{Pr}}}\big[|y|_1\ge k+1\big]
+ \sum_{j=0}^k \mu_j{\ensuremath{\mathrm{Pr}}}\big[|y|_1=j\big].$$
Next, let $\tilde{x}\in \{0,1\}^n$ be chosen arbitrarily such that $|\tilde{x}|_1\ge k+1$ and $\widetilde{\mu}_{k+1} = \widetilde{\mu}(\tilde{x})$. Furthermore, let $\tilde{y}\in\{0,1\}^n$ be a random point generated by flipping each bit in $\tilde{x}$ independently with probability $1/n$. Let $\tilde{z}=\tilde{y}$ if $f(\tilde{y})\le f(\tilde{x})$ and $\tilde{z}=\tilde{x}$ otherwise.
Then $$\begin{aligned}
\widetilde{\mu}(\tilde{x}) = 1
&+ {\ensuremath{\operatorname{E}}}\big[\widetilde{\mu}(\tilde{z}){\,\big|\,}|\tilde{z}|_1\ge k+1\big]\,{\ensuremath{\mathrm{Pr}}}\big[|\tilde{z}|_1\ge k+1\big]\\
&+ \sum_{j=0}^k{\ensuremath{\operatorname{E}}}\big[\widetilde{\mu}(\tilde{z}){\,\big|\,}|\tilde{z}|_1=j\big]\,{\ensuremath{\mathrm{Pr}}}\big[|\tilde{z}|_1=j\big] \end{aligned}$$ and therefore, by definition of $\widetilde{\mu}_j$, $$\label{eq:muf}
\widetilde{\mu}_{k+1}\ge 1
+ \widetilde{\mu}_{k+1}{\ensuremath{\mathrm{Pr}}}\big[|\tilde{z}|_1\ge k+1\big]
+ \sum_{j=0}^k \widetilde{\mu}_j{\ensuremath{\mathrm{Pr}}}\big[|\tilde{z}|_1=j\big].$$
Now, for all $0\le j\le k$, we have $${\ensuremath{\mathrm{Pr}}}\big[|\tilde{z}|_1=j\big]\le{\ensuremath{\mathrm{Pr}}}\big[|\tilde{y}|_1=j\big]\le{\ensuremath{\mathrm{Pr}}}\big[|y|_1=j\big].$$ The first inequality holds, since the event $|\tilde{z}|_1=j$ implies the event $|\tilde{y}|_1=j$. The second inequality follows from Lemma \[lem:probabilities\], since $|x|_1=k+1 \le |\tilde{x}|_1$.
Considering this relation and the fact that the $\widetilde{\mu}_i$ are monotonically increasing in $i$, we obtain from (\[eq:muf\]) that $$\widetilde{\mu}_{k+1}\ge 1
+ \widetilde{\mu}_{k+1}{\ensuremath{\mathrm{Pr}}}\big[|y|_1\ge k+1\big]
+ \sum_{j=0}^k \widetilde{\mu}_j{\ensuremath{\mathrm{Pr}}}\big[|y|_1=j\big].$$
Therefore, the induction hypothesis yields that $$\widetilde{\mu}_{k+1}\ge 1
+ \widetilde{\mu}_{k+1}{\ensuremath{\mathrm{Pr}}}\big[|y|_1\ge k+1\big]
+ \sum_{j=0}^k\mu_j{\ensuremath{\mathrm{Pr}}}\big[|y|_1=j\big] .$$
We subtract both sides of equation (\[eq:muonemax\]) from the previous inequality and immediately get $\widetilde{\mu}_{k+1}\ge \mu_{k+1}$ which concludes the induction.
Thus, for all $x \in \{0,1\}^n$, we have $$\mu(x)=\mu_{|x|_1}\le\widetilde{\mu}_{|x|_1}\le\widetilde{\mu}(x).$$ Consequently, ${\ensuremath{\operatorname{E}}}[ T_{{\textsc{OneMax}\xspace}}]\le{\ensuremath{\operatorname{E}}}[T_f]$ holds.
Multiplicative Drift on Combinatorial Problems {#sec:combinatorial}
==============================================
So far, we have seen that multiplicative drift analysis can be used to simplify the runtime analysis of the (1+1) EA on linear pseudo-Boolean functions while producing sharper bounds. In this section, we see that optimization processes with multiplicative drift occur quite naturally in combinatorial optimization, too. We demonstrate this claim on two prominent examples, the minimum spanning tree problem and the single source shortest path problem.
The Minimum Spanning Tree Problem {#subsec:MST}
---------------------------------
In this subsection, we consider the minimum spanning tree (MST) problem analyzed in [@NeumannW07]. Let $G=(V,E)$ be a connected graph with $n$ vertices, $m$ edges $e_1,\dots,e_m$, and positive integer edge weights $w_1,\dots,w_m$. In [@NeumannW07], a spanning tree is represented by a bit string $x\in\{0,1\}^m$ with $x_i=1$ marking the presence of the edge $e_i$ in the tree.
The fitness value of such a tree is defined by $w(x)=\sum_{i=1}^m w_i x_i+p(x)$, with $p(x)$ being a penalty term ensuring that once the (1+1) EA has found a spanning tree it does no longer accept bit strings that do not represent spanning trees (a new bit-strings is accepted if the fitness value decreases). The minimum weight of a spanning tree is denoted by $w_{\text{opt}}$ and the maximal edge weight by $w_{\text{max}}$.
In Lemma 1 of [@NeumannW07], Neumann and Wegener derive from [@Kano87] the following statement.
\[lem:mst\] Let $x\in\{0,1\}^m$ be a search point describing a non-minimum spanning tree. Then there exist a $k\in\{1,\dots,n-1\}$ and $k$ different accepted 2-bit flips such that the average weight decrease of these flips is at least $(w(x)-w_{\text{opt}})/k$.
Multiplicative drift analysis now gives us a reasonably small constant in the upper bound of the expected optimization time of the (1+1) EA on the MST problem.
\[thm:mst\] The expected optimization time of the (1+1) EA on the MST problem starting with an arbitrary spanning tree of a non-empty graph is at most $2{\mathrm{e}\xspace}m^2(1+\ln m+\ln w_{\text{max}})$.
For all $t\in\mathbb{N}$, let $x^{(t)}$ be the search point of the (1+1) EA for the MST problem at time $t$ and let $X^{(t)}=w(x^{(t)})-w_{\text{opt}}$. Then $$X^{(t)}-X^{(t+1)}=w(x^{(t)})-w(x^{(t+1)}).$$
Now, let $t\in\mathbb{N}$ and $x\in\{0,1\}^n\setminus\{(0,\dots,0)\}$ be fixed. Let the points $y_{(1)},\dots,y_{(k)}$ with $k\in\{0,\dots,n-1\}$ be the $k$ distinct search points in $\{0,1\}^m$ generated from $x$ by the $k$ different 2-bit flips according to Lemma \[lem:mst\]. That is, we have $w(y_{(i)})\le w(x)$ for all $i\in\{1,\dots,k\}$ and $$\label{eq:mstaverage}
\sum_{i=1}^k\big(f(x)-f(y_i)\big)\ge w(x)-w_{\text{opt}}.$$
Since the $y_{(i)}$’s are each generated from $x$ by a 2-bit flip, we have $$\label{eq:mstprob}
{\ensuremath{\mathrm{Pr}}}\big[x^{(t+1)}=y_{(i)}{\,\big|\,}x^{(t)}=x\big]=\Big(1-\frac{1}{m}\Big)^{m-2}\Big(\frac{1}{m}\Big)^2$$ for all $i\in\{1,\dots,k\}$. Furthermore $$\label{eq:mstexp}
{\ensuremath{\operatorname{E}}}\big[X^{(t)}-X^{(t+1)}{\,\big|\,}x^{(t)}=x,x^{(t+1)}=y_{(i)}\big]=w(x)-w(y_{(i)})$$ holds for all $i\in\{1,\dots,k\}$.
The (1+1) EA never increases the current $w$-value of a search point, that is, $X^{(t)}-X^{(t+1)}$ is non-negative. Thus, we have by (\[eq:mstprob\]) and (\[eq:mstexp\]) that $${\ensuremath{\operatorname{E}}}\big[X^{(t)}-X^{(t+1)}{\,\big|\,}x^{(t)}=x\big]\ge\sum_{i=1}^k\Big(w(x)-w(y_{(i)})\Big)\Big(1-\frac{1}{m}\Big)^{m-2}\Big(\frac{1}{m}\Big)^2$$ and therefore, by inequality (\[eq:mstaverage\]), we have for all $x\in\{0,1\}^m$ that $${\ensuremath{\operatorname{E}}}\big[X^{(t)}-X^{(t+1)}{\,\big|\,}x^{(t)}=x\big]
\ge
\frac{w(x)-w_{\text{opt}}}{{\mathrm{e}\xspace}m^2}.$$ In other words, $${\ensuremath{\operatorname{E}}}[X^{(t)}-X^{(t+1)}\mid X^{(t)}]\ge \frac{X^{(t)}}{{\mathrm{e}\xspace}m^2}$$ and the theorem follows from the Theorem \[thm:multidrift\] with $1\le X^{(t)}\le m w_{\text{max}}$.
The Single-source Shortest Path Problem {#subsec:SSSP}
---------------------------------------
In [@BaswanaBDFKN09], Baswana, Biswas, Doerr, Friedrich, Kurur, and Neumann study an evolutionary algorithm that solves the single-source shortest path (SSSP) problem on a directed graph with $n$ vertices via evolving a shortest-path tree. In the analysis of the upper bound for the expected optimization time, the authors introduce the *gap* $g_i$ in iteration $i$ as the difference in fitness between the current shortest-path tree candidate and an optimal shortest-path tree.
For every vertex in the tree, its *weight* in the tree is defined as the sum over the weights of edges in the paths leading to the root vertex, or as the penalty term $n w_{\text{max}}$ if the vertex is not connected to the root. The fitness of a shortest-path tree candidate is then the sum over the weights of all vertices in the tree. Thus the maximal gap is $n^2 w_{\text{max}}$. In Lemma 1 of [@BaswanaBDFKN09], the authors then provide the following statement.
Let $g_i$ denote the gap after $i$ mutations. Then it holds for the conditional expectation ${\ensuremath{\operatorname{E}}}[g_{i+1}\mid g_i=g]$ that $${\ensuremath{\operatorname{E}}}[g_{i+1}\mid g_i=g]\le g \Big(1-\frac{1}{3\cdot n^3}\Big).$$
To this, we can directly apply Theorem \[thm:multidrift\], taking the gap as a potential. We obtain the following result with a precise constant for the upper bound.
\[thm:sssp\] The expected optimization time of the (1+1) EA in [@BaswanaBDFKN09] on the SSSP problem starting with an arbitrary shortest-path tree candidate is at most $6 n^3 (1+2 \ln n+\ln w_{\text{max}})$.
The Euler Tour Problem {#subsec:Euler}
----------------------
The Euler tour problem is to find a Euler tour (a closed walk that visits every edge exactly once) in an input graph which permits such a tour.
In [@DoerrJ07], possible variants of the (1+1) EA for the Euler tour problem are analyzed. For the variant using the so-called edge-based distribution on cycle covers, the search space is given by *adjacency list matchings*, where each matching represents a cover of the input graph with edge-disjoint cycles. The fitness of a matching is given by the total number of cycles in the cover. Thus, a fitness of one implies that the graph is covered by a single cycle — an Euler tour.
Finding such a tour is then a minimization problem over this search space. For this setup, the following statement is implicitly shown in the proof of Theorem 3.
In a single iteration of the (1+1) EA in [@DoerrJ07] for the Euler tour problem using the edge-based distribution and starting with an arbitrary cycle cover, the probability to decrease the fitness $f(x)$ of the current search point $x$ by one (provided it was not minimal before) is at least $f(x)/{\mathrm{e}\xspace}{}m$ where $m$ is the number of edges of the input graph.
If we set the fitness minus one as potential, this lemma immediately implies that the expected drift is at least $f(x)/{\mathrm{e}\xspace}{}m$. Moreover, the starting potential is at most $m/3$ (each tour has hat least three edges). Again, we can apply Theorem \[thm:multidrift\] and reproduce the upper bound the expected optimization time, specifying the leading constant in the process.
\[thm:euler\] The expected optimization time of the (1+1) EA in [@DoerrJ07] for the Euler tour problem using the edge-based distribution and starting with an arbitrary cycle cover is at most ${\mathrm{e}\xspace}{}m\ln m$, where $m$ is the number of edges in the input graph.
Discussion and Outlook {#sec:Discussion}
======================
In this work, we showed that the multiplicative drift condition (\[eq:multidrift\]) occurs naturally in the runtime analysis of the (1+1) EA for number of prominent optimization problems (linear functions, minimum spanning trees, shortest paths, and Euler tours). In such situations our multiplicative drift theorem (Theorem \[eq:multidrift\]) yields good runtime bounds.
We applied this new tool to various settings. First, we used it to gain new insight in the classical problem of how the (1+1) EA optimizes linear functions.
We presented a simplified proof of the, by now, well-known fact that the (1+1) EA with mutation probability $1/n$ optimizes any linear function in time $O(n \log n)$. Moreover, we applied our result to the distribution-based drift analysis of Jägersküpper and obtained a new upper bound of $(1+o(1)) 1.39 {\mathrm{e}\xspace}n \ln(n)$ for the expected optimization time of the (1+1) EA for arbitrary linear functions.
We complement this upper bound by a lower bound of $(1-o(1)) {\mathrm{e}\xspace}n \ln(n)$. To do so, we showed that [<span style="font-variant:small-caps;">OneMax</span>]{}is the function easiest optimized by the (1+1) EA. By this we extended a recent lower bound of $(1-o(1)) {\mathrm{e}\xspace}n \ln(n)$ for the expected optimization time on [<span style="font-variant:small-caps;">OneMax</span>]{}to all functions having a unique global optimum.
Our upper and lower bounds for the expected optimization times of the (1+1) EA on arbitrary linear functions are relatively close. This raises the question if possibly all linear functions have the same expected optimization time of $(1+o(1)) {\mathrm{e}\xspace}n \ln(n)$.
Finally, we reviewed previous runtime analyses of the (1+1) EA on the combinatorial problems of finding a minimum spanning tree, shortest path tree, or Euler tour in a graph. For all three cases, we exhibited the appearance of multiplicative drift and determined the leading constants in the bounds of the expected optimization times.
In the light of these natural occurrences of multiplicative drift, we are optimistic to see applications of multiplicative drift analysis in the near future.
Acknowledgments {#acknowledgments .unnumbered}
---------------
We like to thank Dirk Sudholt for pointing out that Theorem \[thm:lowerbound\] holds for all pseudo-Boolean function with a unique global optimum rather than only for all linear functions.
Note Added in Proof {#note-added-in-proof .unnumbered}
-------------------
Recently, Doerr and Goldberg [@DoerrG10] have shown that in Theorem \[thm:multidrift\], the stopping time $T$ is with high probability at most of the same order as the upper bound on its expectation given in inequality (\[eq:multidrift\]), if $X^{(0)}$ is at least $\Omega(n)$. Thus, the implicit upper bound given in Theorem \[thm:upperbound\] and the bounds in Theorem \[thm:mst\] and Theorem \[thm:sssp\] also hold with high probability, if we allow a slightly larger leading constant.
[BBD[[$^{+}$]{}]{}09]{}
Surender Baswana, Somenath Biswas, Benjamin Doerr, Tobias Friedrich, Piyush P. Kurur, and Frank Neumann. Computing single source shortest paths using single-objective fitness. In [*[FOGA]{}’09: [P]{}roceedings of the 10th [ACM]{} [W]{}orkshop on [F]{}oundations of [G]{}enetic [A]{}lgorithms*]{}, pages 59–66. ACM, 2009.
Benjamin Doerr, Mahmound Fouz, and Carsten Witt. Quasirandom evolutionary algorithms. In [*[GECCO]{}’10: [P]{}roceedings of the 12th [A]{}nnual [G]{}enetic and [E]{}volutionary [C]{}omputation [C]{}onference*]{}, pages 1457–1464. ACM, 2010.
Benjamin Doerr and Leslie A. Goldberg. Drift analysis with tail bounds. In [*[PPSN]{}’10: [P]{}roceedings of the 11th [I]{}nternational [C]{}onference on [P]{}arallel [P]{}roblem [S]{}olving from [N]{}ature*]{}, pages 174–183. Springer, 2010.
Benjamin Doerr and Daniel Johannsen. Adjacency list matchings — an ideal genotype for cycle covers. In [*[GECCO]{}’07: [P]{}roceedings of the 9th [A]{}nnual [G]{}enetic and [E]{}volutionary [C]{}omputation [C]{}onference*]{}, pages 1203–1210. ACM, 2007.
Stefan Droste, Thomas Jansen, and Ingo Wegener. On the analysis of the (1+1) evolutionary algorithm. , 276(1–2):51–81, 2002.
Oliver Giel and Per Kristian Lehre. On the effect of populations in evolutionary multi-objective optimization. In [*[GECCO]{}’06: [P]{}roceedings of the 8th [A]{}nnual [G]{}enetic and [E]{}volutionary [C]{}omputation [C]{}onference*]{}, pages 651–658. ACM, 2006.
Oliver Giel and Ingo Wegener. Evolutionary algorithms and the maximum matching problem. In [*[STACS]{}’03: [P]{}roceedings of the 20th [A]{}nnual [S]{}ymposium on [T]{}heoretical [A]{}spects of [C]{}omputer [S]{}cience*]{}, volume 2607 of [ *Lecture Notes in Computer Science*]{}, pages 415–426. Springer, 2003.
Bruce Hajek. Hitting-time and occupation-time bounds implied by drift analysis with applications. , 14(3):387–403, 1982.
Edda Happ, Daniel Johannsen, Christian Klein, and Frank Neumann. Rigorous analyses of fitness-proportional selection for optimizing linear functions. In [*[GECCO]{}’08: [P]{}roceedings of the 10th [A]{}nnual [G]{}enetic and [E]{}volutionary [C]{}omputation [C]{}onference*]{}, pages 953–960. ACM, 2008.
Jun He and Xin Yao. Drift analysis and average time complexity of evolutionary algorithms. , 127(1):51–81, 2001.
Jun He and Xin Yao. Erratum to: drift analysis and average time complexity of evolutionary algorithms \[[A]{}rtificial [I]{}ntelligence 127 (2001) 57–85\]. , 140(1–2):245–248, 2002.
Jun He and Xin Yao. A study of drift analysis for estimating computation time of evolutionary algorithms. , 3(1):21–35, 2004.
Jens J[ä]{}gersk[ü]{}pper. A blend of [M]{}arkov-chain and drift analysis. In [*[PPSN]{}’08: [P]{}roceedings of the 10th [I]{}nternational [C]{}onference on [P]{}arallel [P]{}roblem [S]{}olving from [N]{}ature*]{}, pages 41–51. Springer, 2008.
Mikio Kano. Maximum and $k$-th maximal spanning trees of a weighted graph. , 7(2):205–214, 1987.
Heinz M[ü]{}hlenbein. How genetic algorithms really work. mutation and hill-climbing. In [*[PPSN]{}’92: [P]{}roceedings of the 2nd [I]{}nternational [C]{}onference on [P]{}arallel [P]{}roblem [S]{}olving from [N]{}ature*]{}, pages 15–25, 1992.
Frank Neumann, Pietro S. Oliveto, and Carsten Witt. Theoretical analysis of fitness-proportional selection: landscapes and efficiency. In [*[GECCO]{}’09: [P]{}roceedings of the 11th [A]{}nnual [G]{}enetic and [E]{}volutionary [C]{}omputation [C]{}onference*]{}, pages 835–842. ACM, 2009.
Frank Neumann and Ingo Wegener. Randomized local search, evolutionary algorithms, and the minimum spanning tree problem. , 378(1):32–40, 2007.
Pietro S. Oliveto and Carsten Witt. Simplified drift analysis for proving lower bounds in evolutionary computation. . In press.
Dominik Scheder and Emo Welzl. Private communications, 2008.
[^1]: Carola Winzen is a recipient of the Google Europe Fellowship in Randomized Algorithms, and this work is supported in part by this Google Fellowship.
[^2]: By $\mathbb{N}:=\{0,1,2,\dots\}$ we denote the set of integers including zero and by $\mathbb{R}$ we denote the set of real numbers.
[^3]: We might as well perform our analysis of $g$ as defined in (\[eq:droste\]). However, our choice of $g$ does not make the somewhat artificial binary distinction between bits with high and low indices and, thus, seems to be more natural.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A topological formulation of the eigenspace anholonomy, where eigenspaces are interchanged by adiabatic cycles, is introduced. The anholonomy in two-level systems is identified with a disclination of the director (headless vector) of a Bloch vector, which characterizes eigenprojectors. The covering map structure behind the exotic quantum holonomy and the role of the homotopy classification of adiabatic cycles are elucidated. The extensions of this formulation to nonadiabatic cycles and $N$-level systems are outlined.'
address:
- 'Department of Physics, Tokyo Metropolitan University, Hachioji, Tokyo 192-0397, Japan'
- ' Laboratory of Physics, Kochi University of Technology, Tosa Yamada, Kochi 782-8502, Japan'
author:
- Atushi Tanaka
- Taksu Cheon
title: 'Bloch vector, disclination and exotic quantum holonomy'
---
Introduction {#sec:Introduction}
============
An adiabatic cycle, where an external parameter of a system is varied infinitely slowly along a closed path, may induce a nontrivial change, which is sometimes referred to as an anholonomy. A famous example is the geometric phase in stationary states [@Berry-PRSLA-392-45]. This is also called as a phase holonomy, which is derived from the interpretation in terms of the differential geometry [@Simon-PRL-51-2167]. Besides the phase of state vector, it turned out that (quasi-)eigenenergies and eigenspaces of stationary states of a closed quantum system may exhibit anholonomies [@Cheon-PLA-248-285]. We call such a change an exotic quantum holonomy.
A more precise description of the exotic quantum holonomy is the following. Suppose that a system is initially in a stationary state, and undergo a unitary time evolution induced by an adiabatic cycle. The final state of the system is orthogonal to the initial state. In accordance to the correspondence between eigenstates and eigenenergies, the trajectory of the eigenenergy connects two different eigenenergies of the initial system. In other words, the adiabatic cycle interchanges the eigenspaces and eigenenergies [@Cheon-PLA-248-285].
Note that the exotic quantum holonomy is different from Wilczek-Zee’s holonomy [@Wilczek-PRL-52-2111], since the former occurs even when there is no spectral degeneracy. Also note that the exotic quantum holonomy is different from the interchange, or the flip, of eigenvectors induced by a cycle around a non-Hermitian degeneracy point [@Uzdin-JPA-44-435302], which is known as Kato’s exceptional point (EP) [@KatoExceptionalPoint; @biorthogonal], in spite of their resemblance explained below. This is because the decay due to the non-Hermiticity makes the stringent limitation to observe the flip in the adiabatic limit [@Uzdin-JPA-44-435302; @Berry-JPA-44-435303]. Namely, the flip due to EPs is observable only in the non-unitary time evolution whose timescale is shorter than the relevant lifetimes in the unstable system, and, in the parametric evolution [@Dembowski-PRL-86-787; @Lee-PRL-103-134101].
The earliest example of the exotic quantum holonomy is the one-dimensional particle under a generalized point potential [@Cheon-PLA-248-285; @Cheon-AP-294-1; @Tsutsui-JMP-42-5687]. Since then, examples are found in quantum maps [@Tanaka-PRL-98-160407; @Miyamoto-PRA-76-042115], quantum circuits [@Tanaka-EPL-96-10005], and quantum graphs [@Ohya-AP-331-299; @Ohya-AP-351-900; @Cheon-ActaPolytechnica-53-410]. It is also shown that the Lieb-Liniger model [@Lieb-PR-130-15; @Ichikawa-PRA-86-015602], which describes one-dimensional Bose systems, exhibits the exotic quantum holonomy along a cycle made of the confinement induced resonance [@Olshanii-PRL-81-938; @Haller-Science-325-1224], whose experimental realization should be feasible within the current state of the art [@Yonezawa-PRA-87-062113]. As an application, an acceleration of the adiabatic quantum computation was examined [@Tanaka-PRA-81-022320].
There are several theoretical works on the exotic quantum holonomy. Firstly, Fujikawa’s gauge theoretical formulation for the phase holonomy [@Fujikawa-PRD-72-025009; @Fujikawa-AP-322-1500] is extended so as to incorporate the eigenspace anholonomy, which is understood as a holonomy of basis vectors [@Cheon-EPL-85-20001; @TANAKA-AP-85-1340]. Secondly, Viennot proposed another gauge theoretical formulation based on the adiabatic Floquet theory, where the nontrivial Floquet block change is discussed in terms of gerbes [@Viennot-JPA-42-395302]. Thirdly, the exotic quantum holonomy is associated to the state flip induced by EPs through the generalized Fujikawa formalism [@Kim-PLA-374-1958; @Tanaka-JPA-46-315302]. Here the analytic continuation of a Hermitian (or unitary) adiabatic cycle provides non-Hermitian cycles, which relate the Riemann surface structure of eigenenergies [@MehriDehnavi-JMP-49-082105] to the exotic quantum holonomy.
In spite of these efforts, there still remain puzzling points on the nature of the exotic quantum holonomy. Firstly, the eigenspace anholonomy and the phase holonomy are mixed in the generalized Fujikawa formalism, and there is only an ad hoc procedure to disentangle them [@Tanaka-JPA-45-335305]. Also, it is not straightforward to extract the geometrical picture from the generalized Fujikawa formalism. Although a solution to this problem is given by Simon’s formulation as for the phase holonomy [@Simon-PRL-51-2167], there has been no known counterpart of the Simon’s formulation for the exotic quantum holonomy. Secondly, Viennot’s theory is for periodically driven systems and, is not applicable to the examples of the exotic quantum holonomy in autonomous (Hamiltonian) systems [@Cheon-PLA-374-144; @Yonezawa-PRA-87-062113; @Cheon-ActaPolytechnica-53-410]. Thirdly, the association of the exotic quantum holonomy to the non-Hermitian degeneracy requires the analyticity of the adiabatic parameter. However, it is easy to construct the examples that lack the analyticity [@Cheon-ActaPolytechnica-53-410]. Hence the EP interpretation of the exotic quantum holonomy is applicable only to a limited class of examples.
In this manuscript, we introduce a topological formulation of the exotic quantum holonomy, in particular the eigenspace anholonomy, to clarify these problems. This enables us to identify the quantity that predicts whether a given cycle exhibits the eigenspace anholonomy. The first key concept in our approach is [*the ordered set of eigenprojectors*]{}. We regard the eigenspace anholonomy as a permutation, which is induced by an adiabatic cycle, among the elements of the ordered set of eigenprojectors. This may be considered as a counterpart to Simon’s vector bundle formulation of the phase holonomy [@Simon-PRL-51-2167]. In a two-level system, we show that a permutation occurs only when the adiabatic cycle encloses a singular point odd times. The second key concept in our topological formulation is [*the cycles in a quantum dynamical variable*]{} that takes the place of conventional cycles in adiabatic parameters. It is shown that the topological nature of the eigenspace anholonomy has a direct link with the homotopy classification of cycles [@Mermin-RMP-51-591]. For example, the singular point mentioned above may be called a disclination, or line defect [@Director]. Furthermore, in common with Aharonov-Anandan formulation [@Aharonov-PRL-58-1593], the new definition of the cycles allows us to extend the eigenspace anholonomy to nonadiabatic cycles.
![ A disclination of eigenobjects of quantum kicked spin-$\frac{1}{2}$ (Eq. ) in the $(B_x,B_y)$-plane. (a) The Bloch vector ${\boldsymbol{a}}$ (Eq. ) at a circle $C$ ($|{\boldsymbol{B}}|=\pi$). The Bloch vector is not well-defined at the origin, which reflects the multiple-valuedness of ${\boldsymbol{a}}$. A “branch cut” is depicted by a wavy line. Here ${\boldsymbol{a}}_0={\boldsymbol{e}}_x$ is the normalized Bloch vector at the initial point $(\pi, 0)$ of $C$. The adiabatic time evolution along $C$ induces a flip of ${\boldsymbol{a}}$. (b) The director (headless vector) ${\boldsymbol{n}}$ of ${\boldsymbol{a}}$ at the circle $C$. Since ${\boldsymbol{n}}$ is single-valued in the $(B_x,B_y)$-plane, no branch cut needs to be drawn. Still, the line defect remains at the origin. []{data-label="fig:an_schematic"}](figure1a "fig:"){width="3.5cm"} ![ A disclination of eigenobjects of quantum kicked spin-$\frac{1}{2}$ (Eq. ) in the $(B_x,B_y)$-plane. (a) The Bloch vector ${\boldsymbol{a}}$ (Eq. ) at a circle $C$ ($|{\boldsymbol{B}}|=\pi$). The Bloch vector is not well-defined at the origin, which reflects the multiple-valuedness of ${\boldsymbol{a}}$. A “branch cut” is depicted by a wavy line. Here ${\boldsymbol{a}}_0={\boldsymbol{e}}_x$ is the normalized Bloch vector at the initial point $(\pi, 0)$ of $C$. The adiabatic time evolution along $C$ induces a flip of ${\boldsymbol{a}}$. (b) The director (headless vector) ${\boldsymbol{n}}$ of ${\boldsymbol{a}}$ at the circle $C$. Since ${\boldsymbol{n}}$ is single-valued in the $(B_x,B_y)$-plane, no branch cut needs to be drawn. Still, the line defect remains at the origin. []{data-label="fig:an_schematic"}](figure1b "fig:"){width="3.5cm"}
Quantum kicked spin-$\frac{1}{2}$ with two adiabatic parameters
===============================================================
Throughout the presentation of our formulation, we assume that the systems are described either by Hermitian Hamiltonians or by unitary Floquet operators. We also assume that there is no spectral degeneracy in the adiabatic cycles. For concreteness, we employ a periodically driven spin-$\frac{1}{2}$ to illustrate our formulation, which is immediately applicable to an arbitrary two-level system. An extension to an arbitrary $N$-level system is also to be shown. Let us suppose that the system is described by a time-periodic Hamiltonian[^1]: $
\hat{H}(t)
\equiv
\frac{1}{2}
{\boldsymbol{B}}\cdot\hat{{\boldsymbol{\sigma}}}
+ \frac{1}{2}
\lambda
({1 -\hat{\sigma}_z})
\sum_{m=-\infty}^{\infty}\delta(t-m)
,
$ where ${\boldsymbol{B}}$ is a static magnetic field and $\hat{{\boldsymbol{\sigma}}}
\equiv
{\boldsymbol{e}}_x\hat{\sigma}_x+{\boldsymbol{e}}_y\hat{\sigma}_y+{\boldsymbol{e}}_z\hat{\sigma}_z$ is a unimodular linear combination of Pauli matrices $\hat{\sigma}_j$ ($j=x,y,z$). The Hamiltonian $\hat{H}(t)$ contains a periodically pulsed rank-$1$ perturbation [@Combesqure-JSP-59-679] with strength $\lambda$. In the following, ${\boldsymbol{B}}$ is assumed to be restricted within $xy$-plane, i.e., ${\boldsymbol{B}}=B_x{\boldsymbol{e}}_x+B_y{\boldsymbol{e}}_y$. To parameterize ${\boldsymbol{B}}$ with the cylindrical coordinates $B$ and $\phi$, we introduce ${\boldsymbol{e}}_{\rho} \equiv {\boldsymbol{e}}_x\cos\phi+{\boldsymbol{e}}_y\sin\phi$ and ${\boldsymbol{e}}_{\phi} \equiv {\boldsymbol{e}}_y\cos\phi-{\boldsymbol{e}}_x\sin\phi$. Hence we have ${\boldsymbol{B}} = B{\boldsymbol{e}}_{\rho}$. We also impose $\lambda=\phi$ in the following. We introduce a Floquet operator, which describes a unit time evolution generated by $\hat{H}(t)$: $$\label{eq:U_def}
\hat{U}
\equiv
\exp\left(-i\phi\frac{1 -\hat{\sigma}_z}{2}\right)
\exp\left(-\frac{i}{2}
B{\boldsymbol{e}}_{\rho} \cdot{\hat{{\boldsymbol{\sigma}}}}\right)
.$$ It is straightforward to show that $\hat{U}$ is periodic in $\phi$ with the period $2\pi$ [@Tanaka-PRL-98-160407; @TANAKA-AP-85-1340]. Accordingly we identify the adiabatic parameter space of the model with a two-dimensional plane $(B_x, B_y)\equiv(B\cos\phi, B\sin\phi)$.
We now diagonalize $\hat{U}$. First, $\hat{U}$ is expanded as $
\hat{U}
=
e^{-i\phi/2}
\left[
\cos({\Delta}/{2})
-i\hat{{\boldsymbol{\sigma}}}\cdot\tilde{{\boldsymbol{a}}}
\right]
,
$ where $$\begin{aligned}
\Delta
&
\equiv& 2\arccos\left(\cos\frac{\phi}{2}\cos\frac{B}{2}\right)
,\\
\tilde{{\boldsymbol{a}}}
&
\equiv&
\left({\boldsymbol{e}}_{\rho}\cos\frac{\phi}{2}
- {\boldsymbol{e}}_{\phi}\sin\frac{\phi}{2}\right)
\sin\frac{B}{2}
\nonumber\\&&\quad{}
-{\boldsymbol{e}}_z\sin\frac{\phi}{2}\cos\frac{B}{2}
.
$$ Because $\tilde{{\boldsymbol{a}}}\cdot\tilde{{\boldsymbol{a}}} = \sin^2({\Delta}/{2})$ holds, the eigenvalues of $\hat{U}$ become degenerate when $\sin({\Delta}/{2}) = 0$ holds. Excluding the degeneracy points ${\boldsymbol{B}} = {\boldsymbol{0}}, 2\pi{\boldsymbol{e}}_x, 4\pi{\boldsymbol{e}}_x, \dots$, we can normalize $\tilde{{\boldsymbol{a}}}$: $$\label{eq:def_a}
{\boldsymbol{a}}
\equiv
\tilde{{\boldsymbol{a}}} / \sin({\Delta}/{2})
.$$ We obtain the spectral decomposition of $\hat{U}$ in the form $$\label{eq:spectral_decomposition}
\hat{U}
= z_+
\hat{P}({\boldsymbol{a}})
+ z_-
\hat{P}(-{\boldsymbol{a}})
,$$ where $z_{\pm}\equiv e^{-i(\phi \pm\frac{\Delta}{2})}$ are eigenvalues, and $$\hat{P}({\boldsymbol{a}})
\equiv \frac{1 +{\boldsymbol{a}}\cdot\hat{{\boldsymbol{\sigma}}}}{2}
,$$ is a projection operator parameterized by a unit vector ${\boldsymbol{a}}$, which is called a (normalized) Bloch vector. Eq. implies that $\hat{P}(\pm{\boldsymbol{a}})$ are the eigenprojectors of $\hat{U}$. In other words, for a given pair of $B_x$ and $B_y$, except at the degeneracy points, there are two normalized Bloch vectors $\pm{\boldsymbol{a}}$, which correspond to two eigenprojectors. We note that the spectral decomposition (Eq. ) is applicable to an arbitrary unitary Floquet operator or Hermitian Hamiltonian as long as the corresponding two level system has no spectral degeneracy. Hence the following argument is applicable to two-level systems in general.
Prior to the consideration of the exotic quantum holonomy in the kicked spin , a remark is due to the adiabatic time evolution in periodically driven systems, for which parametric evolution of an eigenvector of the Floquet operator $\hat{U}$ describes the adiabatic time evolution, in place of a Hamiltonian, up to a phase factor. The dynamical phase is determined by a quasienergy [@Zeldovich-JETP-24-1006] in place of eigenenergy. Proofs of the adiabatic theorem for Floquet systems are found in Refs. [@Young-JMP-11-3298; @Dranov-JMP-39-1340; @Tanaka-JPSJ-80-125002]. The corresponding adiabatic condition is governed by the gaps of quasienergies in place of eigenenergies [@Breuer-ZPD-11-1]. We examine the adiabatic time evolution of the eigenprojector $\hat{P}({\boldsymbol{a}})$ along a cycle $C$ in the $(B_x, B_y)$-plane. It is sufficient to examine the evolution of the normalized Bloch vector ${\boldsymbol{a}}$ instead of $\hat{P}({\boldsymbol{a}})$ due to their equivalence. We depict the parametric evolution of ${\boldsymbol{a}}$ in Fig. \[fig:an\_schematic\] (a). Let ${\boldsymbol{a}}={\boldsymbol{a}}_0$ at the initial point ${\boldsymbol{B}}_0$ on $C$. After a completion of the counterclockwise adiabatic rotation of ${\boldsymbol{B}}$ along $C$, ${\boldsymbol{a}}$ arrives at $-{\boldsymbol{a}}_0$ (see, Fig. \[fig:an\_schematic\](a)), which implies that the final eigenprojector $\hat{P}(-{\boldsymbol{a}}_0)$ is orthogonal to the initial one $\hat{P}({\boldsymbol{a}}_0)$. Hence $C$ induces the interchange of eigenprojectors $\hat{P}(\pm{\boldsymbol{a}}_0)$ resulting in the realization the eigenspace anholonomy. This fact is stable against the deformation of the adiabatic cycle, as long as $C$ encloses the origin only once. Let us next examine the case that $C$ does not enclose the origin. The simplest case is the one where $C$ start from ${\boldsymbol{B}}_0$ and keeps to stay ${\boldsymbol{B}}_0$, i.e., $C$ is a trivial cycle. The direction of the Bloch vector at the final point of the cycle agrees with the one at the initial point. Namely, the eigenprojector returns to the original one after the completion of the adiabatic cycle. This remains correct as long as $C$ does not enclose the origin $O$ in the $(B_x, B_y)$-plane. Also, the initial and final Bloch vectors are the same when $C$ encloses the origin even times.
Eigenspace anholonomy as an anholonomy of an ordered set of mutually orthogonal projection operators
======================================================================================================
Here, we propose a novel interpretation of the normalized Bloch vector ${\boldsymbol{a}}$ that allows the extension of our analysis to systems with an arbitrary number of levels. The central object is an ordered set of mutually orthogonal projection operators $$\label{eq:p_def}
p\equiv
\bigl(|\psi_0\rangle\langle\psi_0|, |\psi_1\rangle\langle\psi_1|\bigr)
,$$ which can be specified by a normalized Bloch vector $$p({\boldsymbol{a}})~= (\hat{P}({\boldsymbol{a}}),\hat{P}(-{\boldsymbol{a}}))
.$$ A given pair of $B_x$ and $B_y$, except at the degeneracy points, specifies two normalized Bloch vectors $\pm{\boldsymbol{a}}$. One of them, say ${\boldsymbol{a}}$, precisely determines $p$. Another normalized Bloch vector $-{\boldsymbol{a}}$ correspond to another ordered set of projection operators $p(-{\boldsymbol{a}})=(\hat{P}(-{\boldsymbol{a}}),\hat{P}({\boldsymbol{a}}))$, which is obtained by a permutation of the elements of $p({\boldsymbol{a}})$. As for two-level systems, we can identify $p$ with a normalized Bloch vector ${\boldsymbol{a}}$, which helps our geometric intuition. The $p$-space of two-level systems is equivalent to the sphere $S^2$.
In terms of the ordered set of projectors $p$, the eigenspace anholonomy is the permutation of the elements of $p$ induced by an adiabatic cycle. For example, let us start an adiabatic cycle $C$ that enclose the origin of $(B_x, B_y)$-plane in Fig. \[fig:an\_schematic\](a). After the completion of the cycle $C$, the elements of $p$ are interchanged. In other words, $C$ corresponds to a permutation of the elements of $p$. In this study, we make use of topological concepts of covering maps and homotopy groups [@Lee-ITT-2011; @Nakahara-GTP-1990]. These concepts have been utilized, for example, in the studies of topological defects [@Mermin-RMP-51-591] and exceptional points [@MehriDehnavi-JMP-49-082105]. A concise summary of the covering map is available in Ref. [@MehriDehnavi-JMP-49-082105]. It will be shown that the resultant interpretation of the eigenspace anholonomy resembles Simon’s interpretation of the phase holonomy in terms of vector bundles [@Simon-PRL-51-2167]. Although the following description is for the two-level kicked spin, the generalization to systems with an arbitrary number of levels is straightforward, as to be explained later. Let us be more precise on the relationship between the adiabatic parameter space and $p$-space. Let ${\mathcal{M}}$ denote the adiabatic parameter space where the spectral degeneracy points are excluded. For a given point $x$ in ${\mathcal{M}}$, let ${\mathcal{P}}_{x}$ denote the set of two possible values of $p$. Since ${\mathcal{P}}_{x}$ may be regarded as a fiber, we denote the corresponding fiber bundle as ${\mathcal{P}}$. The projection from ${\mathcal{P}}$ to ${\mathcal{M}}$ is known as a covering map in topological analysis of manifolds [@Lee-ITT-2011; @MehriDehnavi-JMP-49-082105].
Despite the apparent similarity, there are several crucial differences between the fiber bundle appearing here and the fiber bundle interpretation of the phase holonomy. Among them, we point out the discreteness of the structure group of the fibers for the eigenspace anholonomy. This comes from the fact that the fiber bundle originates from the covering map [@Lee-ITT-2011], so that $p$ is “quantized” at a given point in ${\mathcal{M}}$. Contrarily, the structure group for the phase holonomy is mostly continuous [@Bohm-GPQS-2003].
For the kicked spin , ${\mathcal{M}}$ is $(B_x, B_y)$-plane excluding the degeneracy points. For a given point, say ${\boldsymbol{B}}$, in ${\mathcal{M}}$, there are two Bloch vectors, say $\pm{\boldsymbol{a}}$, so that a fiber ${\mathcal{P}}_{{\boldsymbol{B}}}$ consists of the two points $({\boldsymbol{B}},
p(\pm{\boldsymbol{a}}))$. See, figure \[fig:cover2\] (a). The corresponding structure group is the symmetric group of two elements (i.e., the group that contains all permutations for two items), and coincides with ${{\mathbb{Z}}}_2$. In order to simplify the following argument, we restrict ${\mathcal{M}}$ to be an annulus whose center is the origin in $(B_x, B_y)$-plane (e.g., $\frac{\pi}{2} < B < \frac{3\pi}{2}$). (See, figures \[fig:an\_schematic\](a) and \[fig:cover2\] (a)). We now associate the time evolution of $p$ along an adiabatic cycle $C$ to the concept of lifting. For a given initial condition of $p$, the trajectory of $p$ induced by $C$ essentially determines the lifting $\tilde{C}$ of $C$ to ${\mathcal{P}}$ [@Bohm-GPQS-2003]. The situation is visualised in Fig. \[fig:cover2\] (a). As for the kicked spin, for a given initial point of $C$, say, ${\boldsymbol{B}}_0$, there are two possible normalized Bloch vectors $\pm{\boldsymbol{a}}_0$, which provides two different lifts of $C$. Suppose ${\boldsymbol{B}}_0$ is slightly varied. Correspondingly, the Bloch vectors $\pm {\boldsymbol{a}}_0$ are smoothly and slightly deformed. The repetition of this procedure determines the two lifts $\tilde{C}_{\pm}$ of $C$.
The lift $\tilde{C}$ tells us whether the eigenprojectors in $p$ are interchanged by the adiabatic cycle $C$. When $\tilde{C}$ is a closed path, each eigenprojector in $p$ also draws a closed path. This implies the absence of the eigenspace anholonomy. On the other hand, when $\tilde{C}$ is open, the initial and the final points of an eigenprojector are different, which is the case in which the eigenspace anholonomy occurs. Furthermore, the destination of $\tilde{C}$ precisely describes the permutation among the eigenprojectors.
Let $\phi_C(p)$ denote the final point of the lifted path $\tilde{C}$ of $C$, where $p$ is the initial point of $\tilde{C}$ path. Our question of the eigenspace anholonomy is now casted into the problem of determining $\phi_C$. The mapping $\phi_C$ is called a monodromy action in the analysis of covering map [@Lee-ITT-2011]. For a given initial point in ${\mathcal{P}}$, let $\Phi$ denote the set of $\phi_C$ generated by all possible adiabatic cycles. $\Phi$ is called the automorphism group of the covering [@Lee-ITT-2011], and is the counterpart of the holonomy group for the phase holonomy.
Not every detail of $C$ is required to determine $\phi_C$. This is to be expected from the fact that $\phi_C$ corresponds to a permutation, and thus is “quantized”. In the analysis of covering map, the homotopy classification of $C$ plays the central role [@Lee-ITT-2011; @MehriDehnavi-JMP-49-082105]. We say that a path $C$ is homotopic to another path $C'$, when $C$ can be smoothly deformed to $C'$ with the initial and final points kept unchanged. Let $[C]$ denote the class of paths that are homotopic to $C$ [@Lee-ITT-2011]. Since $\phi_C$ and $\phi_{C'}$ are the same as long as $C$ is homotopic to $C'$ because of the homotopy lifting property, the quantity $\phi_{C}$ may be denoted as $\phi_{[C]}$ [@Lee-ITT-2011]. Hence, it suffices to examine $\phi_{[C]}$ with $[C]$ belonging to the first fundamental group of ${\mathcal{M}}$, [*i.e.*]{}, $[C]\in\pi_1({\mathcal{M}})$.
It is important, at the same time, to observe that $\phi_{[C]}$ describes the identical permutation if and only if the lifted cycle $\tilde{C}$ of $C$ is homotopic to a closed path in ${\mathcal{P}}$, i.e., $[\tilde{C}]\in\pi_1({\mathcal{P}})$.
In order to express this condition in terms of ${\mathcal{M}}$, we consider the projection of all the elements $\pi_1({\mathcal{P}})$ to ${\mathcal{M}}$. We denote the resultant set as $H$, which is a subset of $\pi_1({\mathcal{P}})$ and is called an isotropy group. The quotient space $\pi_1({\mathcal{M}})/H$ precisely classifies $\phi_{[C]}$, i.e., $\Phi\simeq\pi_1({\mathcal{M}})/H$, according to the covering automorphism group structure theorem (Ref [@Lee-ITT-2011], Theorem 12.7).
For example, the kicked spin has $\pi_1({\mathcal{M}}) = \{[e], [\alpha], [\alpha^2], \ldots\}$, where $e$ is homotopic to a point and $\alpha$ encloses the origin $O$ once in ${\mathcal{M}}$ (see, figure \[fig:cover2\]). On the other hand, $\pi_1({\mathcal{P}}) = \{[e'], [\beta], [\beta^2], \ldots\}$ where $e'$ is homotopic to a point, and $\beta$ encloses the “hole” in ${\mathcal{P}}$. Because the projection of $\beta$ to ${\mathcal{M}}$ is homotopic to $\alpha^2$, we find $H=\{[e], [\alpha^2], [\alpha^4], \ldots\}$. Hence, $\Phi=\pi_1(M)/H$ consists of two classes. One corresponds to the cycles that encloses enclose $O$ even times, and makes $\phi_{[C]}$ the identity. The other is composed by the cycles that enclose $O$ odd times to make $\phi_{[C]}$ the cyclic permutation of the two items in $p$. Hence we conclude that $\Phi$, the automorphism group of the covering, is equivalent with ${{\mathbb{Z}}}_2$, the cyclic group whose order is $2$.
Definition of cycles by quantum dynamical variables instead of c-number parameters
====================================================================================
So far, cycles are parameterized by the adiabatic parameters $(B_x, B_y)$. This has been a common definition in the previous studies of the exotic quantum holonomy [@Cheon-PLA-248-285; @Cheon-EPL-85-20001]. Instead, we propose a way to define the cycles only in terms of quantum dynamical variables. The aim here is twofold. One is to complete a geometrical view of the eigenspace anholonomy. Another is to extend the exotic quantum holonomy into nonadiabatic cycles, which will be examined in the next section. To achieve this, we introduce [*a set of mutually orthogonal eigenprojectors*]{} $$\label{eq:b_def}
b
\equiv
\{|\psi_0\rangle\langle\psi_0|, |\psi_1\rangle\langle\psi_1|\}
,$$ where the order of the projector are disregarded.
As for the two level systems, we obtain a geometric interpretation of $b$ with the help of a normalized Bloch vector ${\boldsymbol{a}}$ $$b({\boldsymbol{a}})
=
\{\hat{P}({\boldsymbol{a}}),\hat{P}(-{\boldsymbol{a}})\}
,$$ which agrees with $b(-{\boldsymbol{a}})$, since the order of the elements in $b$ is ignored. In other words, we identify ${\boldsymbol{a}}$ and $-{\boldsymbol{a}}$ in the specification of $b$. In geometry, the identification of antipodal points on the sphere $S^2$ leads to the real projective plane ${{\mathbb{R}}}{}P^2$ [@Mermin-RMP-51-591; @Nakahara-GTP-1990]. Hence we identify $b$ with a point, which we denote as ${\boldsymbol{n}}$, in the projective plane. In Fig. \[fig:an\_schematic\] (b), ${\boldsymbol{n}}$ is depicted in the $(B_x, B_y)$-plane. We note that ${\boldsymbol{n}}$ is single-valued here. Still, we have a singularity at the origin $O$, where the value of ${\boldsymbol{n}}$ cannot be determined. This resembles a disclination of nematic liquid crystals [@Director]. In the studies of nematic liquid crystals, ${\boldsymbol{n}}$ is called as a [*director*]{}, or a headless vector [@Director]. The disclination in nematics is a discontinuity of ${\boldsymbol{n}}$-field in three-dimensional space and the discontinuity may extends along a line [@Mermin-RMP-51-591; @Nakahara-GTP-1990]. In this sense, the disclination is distinct from the effective monopole induced by the phase holonomy [@Berry-PRSLA-392-45]. We use ${\boldsymbol{n}}$, or equivalently $b$, to define an adiabatic cycle $C$. For the quantum kicked spin (Eq. ), we regard that the path $C$ resides in the ${\boldsymbol{n}}$-space rather than in the $(B_x, B_y)$-plane,
It is possible to examine all two-level systems that have no spectral degeneracy with ${{\mathbb{R}}}{}P^2$. This is because the director space ${{\mathbb{R}}}{}P^2$ can parameterize an arbitrary adiabatic cycle of two level systems including the kicked spin-$\frac{1}{2}$, as long as the adiabatic cycle do not encounter any spectral degeneracy. We have a complete classification of the adiabatic cycles of non-degenerate two-level systems in following manner.
Let ${\mathcal{B}}$ denote the adiabatic parameter space ${{\mathbb{R}}}{}P^2$. Now ${\mathcal{P}}$ is equivalent to the whole sphere $S^2$. There is a covering map from ${\mathcal{P}}$ to ${\mathcal{B}}$, in the sense that the inverse of the covering map of ${\boldsymbol{n}}$ is the set of two Bloch vectors $\pm{\boldsymbol{n}}$. The fundamental group of ${\mathcal{P}}$ is $\pi_1({\mathcal{P}}) = {\left\{ [e'] \right\}}$, where $e'$ is a cycle homotopic to a point in ${\mathcal{P}}$. The simplicity of $\pi_1({\mathcal{P}})$ compared to the previous case is the reflection of the fact that ${\mathcal{P}}$ is “larger” than the previous one, as we have “extended” the base space from ${\mathcal{M}}$, an annulus, to ${{\mathbb{R}}}{}P^2$. On the other hand, the fundamental group of the adiabatic parameter space is $\pi_1({{\mathbb{R}}}{}P^2) = {\left\{ [e], [\gamma] \right\}}$, where $e$ is homotopic to a point in ${{\mathbb{R}}}{}P^2$ and $\gamma$ a quantity not homotopic to $e$. We note that $\gamma^2$, i.e., the repetition of $\gamma$ twice is homotopic to $e$ (see, Fig. \[fig:cover2\] (b)). Because ${\mathcal{P}}$ is a universal cover of ${\mathcal{B}}$, there are only two classes of monodromy actions $\phi_{[C]}$, which is either $\phi_{[e]}$ or $\phi_{[\gamma]}$. The former and latter cases correspond to the absence and the presence of the eigenspace anholonomy respectively (see Fig. \[fig:cover2\] (b) again). We conclude that $\Phi=\pi_1({{\mathbb{R}}}{}P^2)$ is equivalent with ${{\mathbb{Z}}}_2$.
![ The covering map structures of two-level systems (schematic). (a) The adiabatic space ${\mathcal{M}}$ (bottom annulus) of the kicked spin-$\frac{1}{2}$ is chosen so as to avoid degeneracy points (e.g., $\frac{\pi}{2} < B < \frac{3\pi}{2}$ in Figure \[fig:an\_schematic\]). Since two edges depicted by dashed lines of ${\mathcal{P}}$ (the winding strip in the above) should be identified, ${\mathcal{P}}$ is a “double-winding” strip above ${\mathcal{M}}$. The projection from ${\mathcal{P}}$ to ${\mathcal{M}}$ is a covering map. The adiabatic cycle $C$ starts from $x_0$ and winds ${\mathcal{M}}$ in the counterclockwise direction. The corresponding lifting $\tilde{C}$ delivers $p_0$ to $p'_0$. Accordingly the monodromy action $\phi_C$ describe the motion in the fiber direction. (b) Adiabatic cycles of the two-level systems without spectral degeneracies can be parameterized by the projective plane ${\mathcal{M}}$, which is equivalent with ${{\mathbb{R}}}{}P^2$, shown in the bottom part. Two cycles (closed paths) $e$ and $\gamma$ are shown there. Two filled circles corresponding to ${\boldsymbol{n}}_0$, which is the initial point of these cycles, are identical in ${{\mathbb{R}}}{}P^2$. The cycle $e$ is homotopic to a zero-length cycle, and is not homotopic to $\gamma$. In the top, the lifts of $e$ and $\gamma$ to the $p$-space, which is equivalent to $S^2$, are shown. Because $S^2$ doubly covers ${{\mathbb{R}}}{}P^2$, there are two normalized Bloch vectors $\pm{\boldsymbol{a}}_0$ for a given director ${\boldsymbol{n}}_0$. Also, each cycle has two lifts (thick and dashed curves). The lifts $\tilde{e}_{\pm}$ of $e$ are closed, signifying the absence of eigenspace anholonomy. On the other hand, the lifts $\tilde{\gamma}_{\pm}$ are open. Along the adiabatic cycle $\gamma$, the initial point ${\boldsymbol{a}}_0$ of $\tilde{\gamma}_{+}$ is transposed to $-{\boldsymbol{a}}_0$, which is the initial point of $\tilde{\gamma}_{-}$, and vice versa. []{data-label="fig:cover2"}](figure2a "fig:"){width="4.3cm"} ![ The covering map structures of two-level systems (schematic). (a) The adiabatic space ${\mathcal{M}}$ (bottom annulus) of the kicked spin-$\frac{1}{2}$ is chosen so as to avoid degeneracy points (e.g., $\frac{\pi}{2} < B < \frac{3\pi}{2}$ in Figure \[fig:an\_schematic\]). Since two edges depicted by dashed lines of ${\mathcal{P}}$ (the winding strip in the above) should be identified, ${\mathcal{P}}$ is a “double-winding” strip above ${\mathcal{M}}$. The projection from ${\mathcal{P}}$ to ${\mathcal{M}}$ is a covering map. The adiabatic cycle $C$ starts from $x_0$ and winds ${\mathcal{M}}$ in the counterclockwise direction. The corresponding lifting $\tilde{C}$ delivers $p_0$ to $p'_0$. Accordingly the monodromy action $\phi_C$ describe the motion in the fiber direction. (b) Adiabatic cycles of the two-level systems without spectral degeneracies can be parameterized by the projective plane ${\mathcal{M}}$, which is equivalent with ${{\mathbb{R}}}{}P^2$, shown in the bottom part. Two cycles (closed paths) $e$ and $\gamma$ are shown there. Two filled circles corresponding to ${\boldsymbol{n}}_0$, which is the initial point of these cycles, are identical in ${{\mathbb{R}}}{}P^2$. The cycle $e$ is homotopic to a zero-length cycle, and is not homotopic to $\gamma$. In the top, the lifts of $e$ and $\gamma$ to the $p$-space, which is equivalent to $S^2$, are shown. Because $S^2$ doubly covers ${{\mathbb{R}}}{}P^2$, there are two normalized Bloch vectors $\pm{\boldsymbol{a}}_0$ for a given director ${\boldsymbol{n}}_0$. Also, each cycle has two lifts (thick and dashed curves). The lifts $\tilde{e}_{\pm}$ of $e$ are closed, signifying the absence of eigenspace anholonomy. On the other hand, the lifts $\tilde{\gamma}_{\pm}$ are open. Along the adiabatic cycle $\gamma$, the initial point ${\boldsymbol{a}}_0$ of $\tilde{\gamma}_{+}$ is transposed to $-{\boldsymbol{a}}_0$, which is the initial point of $\tilde{\gamma}_{-}$, and vice versa. []{data-label="fig:cover2"}](figure2b "fig:"){width="3.5cm"}
Nonadiabatic extension
======================
In our geometrical treatment of the eigenspace anholonomy so far, there has been no mention on the scale of temporal variations for the director ${\boldsymbol{n}}$, nor for the set of projectors $b$. In the adiabatic regime, we regard ${\boldsymbol{n}}$ and $b$ as adiabatic parameters. We can also consider the case in which ${\boldsymbol{n}}$ and $b$ are quantum dynamical variables. In this case, the time evolution defined by quantum theory naturally induces the time evolution of ${\boldsymbol{n}}$ and $b$ which may or may not be adiabatic. We can then define the cycles made out of nonadiabatic time evolution of ${\boldsymbol{n}}$ and $b$.
We give an example of a two level system described by a time-periodic Hamiltonian, whose time dependence may or may not be adiabatic. Let $\hat{F}$ denote a Floquet operator, which describes the time evolution during its period. We suppose that $\hat{F}$ has two eigenvectors ${|{}0{}\rangle}$ and ${|{}1{}\rangle}$. The trajectory of the Bloch vector of $({|{}0{}\rangle\langle{}0{}|},{|{}1{}\rangle\langle{}1{}|})$ forms a closed curve, and the trajectory of director of ${\left\{ {|{}0{}\rangle\langle{}0{}|}, {|{}1{}\rangle\langle{}1{}|} \right\}}$ is homotopic to a point in ${{\mathbb{R}}}{}P^2$, implying the absence of anholonomy.
We show that the nonadiabatic anholonomy occurs if $\hat{F}$ has two eigenvalues $\exp\{-i(\epsilon \pm \pi/2)\}$, where $\epsilon$ is an arbitrary real number. Let us write two eigenstates of ${\hat F}$ as ${|{}0{}\rangle}$ and ${|{}1{}\rangle}$, namely $$\begin{aligned}
\label{eq:eigF}
\hat{F}{|{}0{}\rangle}
= e^{-i(\epsilon + \pi/2)}{|{}0{}\rangle},
\quad
\hat{F}{|{}1{}\rangle}
= e^{-i(\epsilon - \pi/2)}{|{}1{}\rangle} .\end{aligned}$$ A pair of normalized vectors $$\begin{aligned}
\label{eq:defpm}
{|{}\pm{}\rangle}\equiv \frac{1}{\sqrt{2}}({|{}0{}\rangle}\pm{|{}1{}\rangle})\end{aligned}$$ exhibit nonadiabatic anholonomy. This is because, from Eq. , $\hat{F}{|{}+{}\rangle} = e^{-i\left(\epsilon+{\pi}/{2}\right)}{|{}-{}\rangle}$ and $\hat{F}{|{}-{}\rangle} = e^{-i\left(\epsilon+{\pi}/{2}\right)}{|{}+{}\rangle}$ hold. In other words, the trajectory of the Bloch vectors of $({|{}+{}\rangle\langle{}+{}|},{|{}-{}\rangle\langle{}-{}|})$ draws an open curve during the period of the Hamiltonian. The corresponding trajectory of the director must be closed and homotopic to $\gamma$ in ${{\mathbb{R}}}{}P^2$.
An example of $\hat{F}$ is provided by the kicked spin-$\frac{1}{2}$, which is described by the following Hamiltonian $$\begin{aligned}
\label{eq:anotherKickedSpin}
\hat{H}(t)\equiv
\frac{\pi}{2}\hat{\sigma}_z
+\frac{B}{2}\hat{\sigma}_y\sum_{n=-\infty}^{\infty}\delta(t-n)
,\end{aligned}$$ where we assume $0 < B < \pi$. The period of $H(t)$ is unity. The corresponding Floquet operator is $\hat{F}=
\exp(-i \pi\hat{\sigma}_z/2)\exp(-i B\hat{\sigma}_y/2)$, whose normalized eigenvectors are $$\begin{aligned}
\label{eq:AKSeigenvectors}
{|{}0{}\rangle} \equiv
\begin{pmatrix}
\cos\frac{B}{4}\\{}-\sin\frac{B}{4}
\end{pmatrix}
,\quad
{|{}1{}\rangle} \equiv
\begin{pmatrix}
\sin\frac{B}{4}\\{}\cos\frac{B}{4}
\end{pmatrix}
.\end{aligned}$$ Indeed, they satisfy Eq. with $\epsilon=0$. We depict the trajectories of the directors of two sets of projectors ${\left\{ {|{}0{}\rangle\langle{}0{}|}, {|{}1{}\rangle\langle{}1{}|} \right\}}$ and $({|{}+{}\rangle\langle{}+{}|},{|{}-{}\rangle\langle{}-{}|})$ during the unit time interval of the Hamiltonian in ${{\mathbb{R}}}{}P^2$ in Figure \[fig:nonad\].
). On the other hand, let $C_+$ be the cycle whose initial point ${\boldsymbol{n}}_{+}$ correspond to the set of projectors ${\left\{ {|{}+{}\rangle\langle{}+{}|}, {|{}-{}\rangle\langle{}-{}|} \right\}}$ (see, Eq. ). The end point of $C_+$ is ${\boldsymbol{n}}_{-}$, which is equivalent with ${\boldsymbol{n}}_{+}$. Note that $C_+$ is not homotopic to $C_0$. Suppose the system is initially at ${|{}+{}\rangle}$. Due to the time evolution along $C_+$, the system arrives at ${|{}-{}\rangle}$ to exhibit the nonadiabatic anholonomy. []{data-label="fig:nonad"}](figure3){width="3.2cm"}
$N$-dimensional extension
=========================
We close our argument by outlining the extension to cases that involves an arbitrary number, say $N$, of levels. First, let us note that the definitions of $p$ and $b$ (Eqs. and ), which consist of mutually orthogonal projection operators, and also the definitions of ${\mathcal{P}}$ and ${\mathcal{B}}$ are applicable to an arbitrary $N$. We do not attempt, however, to extend the concepts of Bloch vectors and directors for now. We emphasize that ${\mathcal{P}}$ is a covering space of ${\mathcal{B}}$ [@Lee-ITT-2011; @MehriDehnavi-JMP-49-082105], irrespective of the value of $N$. The quantity ${\mathcal{P}}$ is referred to as a flag manifold [@Adelman-FP-23-211; @FlagText].
It is still an open problem to provide a complete classification of the cycles for an arbitrary $N$-level system. This is due to the difficulty in identifying the fundamental groups of ${\mathcal{P}}$ and ${\mathcal{M}}$ for the $N$-level case.
Non the less, it is possible to provide a working example of the present formulation with an arbitrary $N$ if we choose appropriate sub-family of the general $N$-level systems. Let us examine a family of quantum maps whose unit time evolution operator is given by $
\hat{U}(\lambda) \equiv \hat{U}_0\exp(-i\lambda{|{}v{}\rangle\langle{}v{}|}),
$ in which $\lambda$ is an adiabatic parameter, $\hat{U}_0$ a nondegenerate unitary operator, and ${|{}v{}\rangle}$ a normalized vector. We also assume that ${|{}v{}\rangle}$ is not an eigenvector of $\hat{U}_0$. The following is a summary of the result of Ref. [@Miyamoto-PRA-76-042115] on the exotic quantum holonomy in this model. Let ${|{}n{}\rangle}$ denote the $n$-th eigenstate of $\hat{U}_0$, where the quantum number $n$ is assigned in the increasing order of quasienergy. Since $\hat{U}(\lambda)$ is $2\pi$-periodic in $\lambda$, the adiabatic parameter space ${\mathcal{M}}$ is identified with a circle $S^1$. Let $C$ denote an adiabatic cycle where $\lambda$ is increased from $0$ to $2\pi$. It is shown in Ref. [@Miyamoto-PRA-76-042115] that the eigenspace anholonomy is induced by $C$ because $C$ delivers ${|{}n{}\rangle}$ to ${|{}n+1{}\rangle}$.
Applying the present formulation to this model, we first note that we have $\pi_1({\mathcal{M}}) = \{[e], [\alpha], [\alpha^2], \ldots\}$, where $e$ and $\alpha$ are a point and a cycle in ${\mathcal{M}}$, as ${\mathcal{M}}$ is regarded as $S^1$ [@Nakahara-GTP-1990]. Also we find $\pi_1({\mathcal{P}}) = \{[e'], [\beta], [\beta^2], \ldots\}$, where $e'$ is homotopic to a point, and $\beta$ encloses the “hole” in ${\mathcal{P}}$ (cf. Figure \[fig:cover2\](a)). The $N$ dependence appears in the fact that the projection of $\beta$ is homotopic to $\alpha^N$. Hence the isotropy group is $H=\{[e], [\alpha^N], [\alpha^{2N}], \ldots\}$. We conclude $\Phi(=\pi_1(M)/H)$ is equivalent with ${{\mathbb{Z}}}_N$, $N$-th cyclic group.
Discussion
==========
We discuss the relationship of this work to the previous works on the eigenspace anholonomy. In the present formulation, we have identified the layers in the hierarchy of the quantum holonomy, which consists of the $b$-space, $p$-space and the space consists of frames $({|{}\psi_0{}\rangle},{|{}\psi_1{}\rangle},\dots)$. A lift of a closed cycle of $b$ to the $p$-space involves the anholonomy in $p$. In contrast, a lift of a path in $p$-space to the frame space involves the phase holonomy. In particular, Aharonov-Anandan phase is induced by the cycle in $p$-space. On the other hand, the phase holonomy associated with a open path in $p$-space correspond to the off-diagonal geometric phase [@Manini-PRL-85-3067; @Mukunda-PRA-65-012102]. In a gauge theoretical approach introduced in Ref. [@Cheon-EPL-85-20001], the eigenspace anholonomy and the off-diagonal geometric phase are treated together. These two concepts are disentangled and assigned to the different layers through the present formulation.
A remark is due to the relationship between the exotic quantum holonomy and Kato’s exceptional point (EP), which is a branch point of the Riemann surface of eigenenergies, in non-Hermitian quantum theory [@biorthogonal; @KatoExceptionalPoint]. The adiabatic time evolution under the presence of the eigenspace anholonomy resembles a parametric evolution that encloses an EP, in the sense that these evolutions permutate eigenspaces. An analytic continuation of adiabatic cycle in Hermitian Hamiltonian and unitary Floquet systems has enabled to interpret the exotic quantum holonomy as the result of parametric encirclement of EP in the complex plane [@Kim-PLA-374-1958; @Tanaka-JPA-46-315302]. Although such a correspondence is valid only when an analytic continuation of the adiabatic cycle is available, the topological formulation is applicable regardless of the analytic continuation. Also, we do not know how the non-adiabatic extension of exotic quantum holonomy can be associated with EPs. On the other hand, we remind the readers that the relationship between the phase holonomy and EPs is established through the analysis of the Riemann surface of (quasi-)eigenenergy [@MehriDehnavi-JMP-49-082105]. Because the covering space structure naturally resides in the Riemann surfaces, an extension of the present approach to non-Hermitian systems should be of interest.
Acknowledgments {#acknowledgments .unnumbered}
===============
AT wish to thank Professor Ali Mostafazadeh for a useful conversation and Professor Akira Shudo for comments. This research was supported by the Japan Ministry of Education, Culture, Sports, Science and Technology under the Grant number 24540412.
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
M. V. Berry, Quantal phase factors accompanying adiabatic changes, Proc. R. Soc. London A 392 (1984) 45–57.
B. Simon, [H]{}olonomy, the [Q]{}uantum [A]{}diabatic [T]{}heorem, and [B]{}erry’s [P]{}hase, Phys. Rev. Lett. 51 (24) (1983) 2167–2170.
T. Cheon, Double spiral energy surface in one-dimensional quantum mechanics of generalized pointlike potentials, Phys. Lett. A 248 (1998) 285–289.
F. Wilczek, A. Zee, Appearance of gauge structure in simple dynamical systems, Phys. Rev. Lett. 52 (1984) 2111–2114.
R. Uzdin, A. Mailybaev, N. Moiseyev, [On the observability and asymmetry of adiabatic state flips generated by exceptional points]{}, Journal of Physics A: Mathematical and Theoretical 44 (43) (2011) 435302.
T. Kato, Perturbation Theory for Linear Operators, corrected printing of the second Edition, Springer-Verlag, Berlin, 1980, Ch. II.
See, e.g., N. Moiseyev, [*Non-Hermitian Quantum Mechanics*]{} (Cambridge Univ. Press, New York, 2011).
M. V. Berry, R. Uzdin, [Slow non-Hermitian cycling: exact solutions and the Stokes phenomenon]{}, Journal of Physics A: Mathematical and Theoretical 44 (43) (2011) 435303.
C. Dembowski, H.-D. Gr[ä]{}f, H. Harney, A. Heine, W. Heiss, H. Rehfeld, A. Richter, Experimental observation of the topological structure of exceptional points, Phys. Rev. Lett. 86 (2001) 787–790.
S.-B. Lee, J. Yang, S. Moon, S.-Y. Lee, J.-B. Shim, S. Kim, J.-H. Lee, K. An, Observation of an exceptional point in a chaotic optical microcavity, Phys. Rev. Lett. 103 (2009) 134101.
T. Cheon, T. F[ü]{}l[ö]{}p, I. Tsutsui, Symmetry, duality and anholonomy of point interaction in one dimension, Ann. Phys. (N.Y.) 294 (2001) 1–23.
I. Tsutsui, T. F[ü]{}l[ö]{}p, T. Cheon, Moebius structure of the spectral space of schroedinger operators with point interaction, J. Math. Phys. 42 (2001) 5687–5697.
A. Tanaka, M. Miyamoto, Quasienergy anholonomy and its application to adiabatic quantum state manipulation, Phys. Rev. Lett. 98 (2007) 160407–1–160407–4.
M. Miyamoto, A. Tanaka, Cheon’s anholonomies in [Floquet]{} operators, Phys. Rev. A 76 (2007) 042115–1–042115–9.
A. Tanaka, S. W. Kim, T. Cheon, Eigenvalue and eigenspace anholonomies in hierarchical systems, Europhys. Lett. 96 (2011) 10005–p1–10005–p6.
S. Ohya, Parasupersymmetry in quantum graphs, Ann. Phys. (NY) 331 (2013) 299.
S. Ohya, Non-abelian monopole in the parameter space of point-like interactions, Ann. Phys. (NY) 351 (2014) 900.
T. Cheon, A. Tanaka, O. Turek, Examples of quantum holonomy with topology changes, Acta Polytechnica 53 (2013) 410.
E. H. Lieb, W. Liniger, Exact [A]{}nalysis of an [I]{}nteracting [B]{}ose [G]{}as. [I]{}. [T]{}he [G]{}eneral [S]{}olution and the [G]{}round [S]{}tate, Phys. Rev. 130 (4) (1963) 1605–1616.
T. Ichikawa, I. Tsutsui, N. Yonezawa, Equation of state for the one-dimensional attractive $\delta$-potential [B]{}ose gas in the weak-coupling regime, Phys. Rev. A 86 (2012) 015602.
M. Olshanii, Atomic [S]{}cattering in the [P]{}resence of an [E]{}xternal [C]{}onfinement and a [G]{}as of [I]{}mpenetrable [B]{}osons, Phys. Rev. Lett. 81 (1998) 938.
E. Haller, M. Gustavsson, M. J. Mark, J. G. Danzl, , R. Hart, G. Pupillo, H.-C. N[ä]{}gerl, Realization of an [E]{}xcited, [S]{}trongly [C]{}orrelated [Q]{}uantum [G]{}as [P]{}hase, Science 325 (2009) 1224.
N. Yonezawa, A. Tanaka, T. Cheon, Quantum holonomy in the [L]{}ieb-[L]{}iniger model, Phys. Rev. A 87 (2013) 062113.
A. Tanaka, K. Nemoto, Adiabatic quantum computation along quasienergies, Phys. Rev. A 81 (2010) 022320–1–022320–8.
K. Fujikawa, Geometric phases and hidden local gauge symmetry, Phys. Rev. D 72 (2005) 025009.
K. Fujikawa, Geometric phases for mixed states and decoherence, Ann. Phys. (N.Y.) 322 (2007) 1500–1517.
T. Cheon, A. Tanaka, New anatomy of quantum holonomy, Europhys. Lett. 85 (2009) 20001–p1–20001–p5.
A. Tanaka, T. Cheon, A unified theory of quantum holonomies, Ann. Phys. (N.Y.) 324 (2009) 1340–1359.
D. Viennot, [G]{}eometric phases in adiabatic [F]{}loquet theory, [A]{}belian gerbes and [C]{}heon’s anholonomy, J. Phys. A. 42 (2009) 395302, (22pp).
S. W. Kim, T. Cheon, A. Tanaka, Exotic quantum holonomy induced by degeneracy hidden in complex parameter space, Phys. Lett. A 374 (2010) 1958–1961.
A. Tanaka, N. Yonezawa, T. Cheon, Exotic quantum holonomy and non-[H]{}ermitian degeneracies in the two-body [L]{}ieb-[L]{}iniger model, Journal of Physics A: Mathematical and Theoretical 46 (31) (2013) 315302.
H. Mehri-Dehnavi, A. Mostafazadeh, Geometric phase for non-hermitian hamiltonians and its holonomy interpretation, J. Math. Phys. 49 (2008) 082105.
A. Tanaka, T. Cheon, S. W. Kim, Gauge invariants of eigenspace and eigenvalue anholonomies: [E]{}xamples in hierarchical quantum circuits, Journal of Physics A: Mathematical and Theoretical 45 (33) (2012) 335305–1–335305–20.
T. Cheon, A. Tanaka, S. W. Kim, Exotic quantum holonomy in [Hamiltonian]{} systems, Phys. Lett. A 374 (2009) 144–149.
N. D. Mermin, The topological theory of defects in ordered media, Rev. Mod. Phys. 51 (1979) 591.
P. G. de Gennes, J. Prost, [The Physics of Liquid Crystals]{}, 2nd Edition, Oxford University Press, Oxford, 1993, Ch. 1.
Y. Aharonov, J. Anandan, Phase change during a cyclic quantum evolution, Phys. Rev. Lett. 58 (1987) 1593–1596.
M. Combescure, Spectral properties of a periodically kicked quantum hamiltonian, J. Stat. Phys. 59 (1990) 679–690.
Y. B. Zel[’]{}dovich, The quasienergy of a quantum-mechanical system subjected to a periodic action, J. Exp. Theor. Phys. 24 (6) (1967) 1006–1008.
R. H. Young, W. J. [Deal Jr.]{}, Adiabatic [R]{}esponse to an [O]{}scillatory [F]{}ield, J. Math. Phys. 11 (11) (1970) 3298–3306.
A. Dranov, J. Kellendonk, R. Seiler, [Discrete time adiabatic theorems for quantum mechanical systems]{}, Journal of Mathematical Physics 39 (3) (1998) 1340.
A. Tanaka, Adiabatic theorem for discrete time evolution, Journal of the Physical Society of Japan 80 (2011) 125002–1–125002–2.
H. Breuer, M. Holthaus, Adiabatic processes in the ionization of highly excited hydrogen atoms, Z. Phys. D 11 (1989) 1–14.
J. M. Lee, Introduction to Topological Manifolds, 2nd Edition, Vol. 202 of Graduate Texts in Mathematics, Springer, New York, 2011.
M. Nakahara, Geometry, Topology and Physics, IOP Publishing Ltd, Bristol, 1990.
A. Bohm, A. Mostafazadeh, H. Koizumi, Q. Niu, Z. Zwanziger, The Geometric Phase in Quantum Systems, Springer, Berlin, 2003.
M. Adelman, J. V. Corbett, C. A. Hurst, The geometry of state space, Found. Phys. 23 (2) (1993) 211.
I. Bengtsson, K. [Ż]{}yczkowski, Geometry of Quantum States, Cambridge University Press, Cambridge, 2006, Ch. 4.
N. Manini, F. Pistolesi, Off-[D]{}iagonal [G]{}eometric [P]{}hases, Phys. Rev. Lett. 85 (2000) 3067–3071.
N. Mukunda, Arvind, S. Chaturvedi, R. Simon, Bargmann invariants and off-diagonal geometric phases for multilevel quantum systems: [A]{} unitary-group approach, Phys. Rev. A 65 (2001) 012102–1–012102–10.
[^1]: The exotic quantum holonomy in an autonomous Hamiltonian system requires either a level crossing or the divergence of eigenenergy [@Cheon-PLA-374-144]. To avoid complications from such singularities, we here examine periodically driven systems.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The commensurate charge density wave (CDW) in the layered compound $1T$-TaS$_2$ has hitherto mostly been treated as a quasi two-dimensional phenomenon. Recent band structure calculations have, however, predicted that the CDW coexists with a nearly one-dimensional metallic dispersion perpendicular to the crystal planes. Using synchrotron radiation based angle-resolved photoemission spectroscopy, we show that this metallic band does in fact exist. Its occupied band width is in excellent agreement with predictions for a simple $\tau_c$ stacking order of the CDW between adjacent layers and its periodicity in the $c$ direction is $2 \pi / c$.'
author:
- 'Arlette S. Ngankeu'
- Kevin Guilloy
- 'Sanjoy K. Mahatha'
- Marco Bianchi
- 'Charlotte E. Sanders'
- Kai Rossnagel
- 'Jill A. Miwa'
- Philip Hofmann
title: 'Quasi one-dimensional metallic band dispersion in the commensurate charge density wave of $1T$-TaS$_2$'
---
The $1T$ polytype of TaS$_2$ is one of the most studied layered transition metal dichalcogenides (TMDCs). Its rich electronic phase diagram involves several charge density wave (CDW) transitions driven by strong electronic correlations and electron-phonon coupling [@Wilson:1975aa; @Rossnagel:2011aa]. Particular focus has been on the ground state below 180 K which is a commensurate CDW phase with a $\sqrt{13} \times \sqrt{13}$ so-called Star of David reconstruction that is rotated by $13.9^{\circ}$ against the lattice. In this phase, the large-scale periodic lattice distortion is thought to coexist with a Mott insulating ground state arising from the single electron localized on the centre atom of the Star of David [@Wilson:1975aa; @Fazekas:1979aa]. While the research on CDWs in layered TMDCs is more than 40 years old, renewed interest has been driven by the possibility to elucidate transitions between different CDW states using ultrafast techniques [@Perfetti:2006aa; @Perfetti:2008aa; @Hellmann:2010aa; @Eichberger:2010aa; @Petersen:2011ab; @Stojchevska:2014aa]; by the observation of metastable “hidden states” [@Stojchevska:2014aa]; and by the experimental accessibility of metallic TMDCs as single layers [@Ugeda:2016aa; @Sanders:2016aa].
The realization that CDWs could be different in single layer TMDCs than in analogous bulk materials has drawn attention to the fact that viewing the bulk materials’ electronic properties as essentially two-dimensional might be an oversimplification. While reduced dimensionality has a significant impact on electronic instabilities, due to increased electronic correlations and electron-phonon coupling, interlayer coupling also appears to be essential for a full understanding of the electronic properties of these materials [@Bovet:2003aa; @Rossnagel:2005ab; @Freericks:2009ab; @Darancet:2014aa; @Ritschel:2015aa; @Lazar:2015aa]. Specifically, several calculations predict a one-dimensional metallic band formation along the $\Gamma-A$ direction of the Brillouin zone in the ground state CDW of $1T$-TaS$_2$ (i.e., perpendicular to the planes). This is found in density functional theory calculations [@Bovet:2003aa; @Ritschel:2015aa; @Lazar:2015aa], even when electronic correlations are taken into account [@Darancet:2014aa; @Freericks:2009ab]. While the metallic band along $\Gamma-A$ is universally found in all calculations, the details of the dispersion depend on the stacking order of the CDW unit cell between adjacent planes [@Darancet:2014aa; @Ritschel:2015aa].
Angle-resolved photoemission spectroscopy (ARPES) is an experimental technique capable of determining the three-dimensional band structure of crystalline solids, and numerous ARPES studies have been performed on $1T$-TaS$_2$ (for a review see Ref. [@Rossnagel:2011aa]). However, very little attention has been paid to the possibly three-dimensional character of the band structure. In this Letter, we report a systematic determination of the band structure perpendicular to the planes of $1T$-TaS$_2$ from high-quality crystals in the commensurate CDW phase, with special focus on the possible metallicity of the compound in this direction. We do observe the theoretically predicted one-dimensional metallic band. The occupied band width and the observed periodicity agree with a simple $\tau_c$ stacking order of the CDW—i.e., a stacking in which the center atoms of the Stars of David are placed directly on top of each other in adjacent layers.
The $1T$-TaS$_2$ crystals were grown from high purity elements by chemical vapor transport using iodine as a transport agent; for details see Ref. [@Haupt:2016aa]. The crystals were cleaved at $\approx$30 K in ultra-high vacuum, and ARPES data were collected on the SGM-3 end station of ASTRID2 [@Hoffmann:2004aa]. The energy resolution varied from $\approx$50 to $\approx$130 meV for the lowest and highest photon energies, respectively. The angular resolution was better than 0.2$^{\circ}$. The temperature during the ARPES experiments was $\approx$30 K.
A first suggestion of the three-dimensional character of the $1T$-TaS$_2$ band structure in the commensurate CDW is given in Fig. \[fig:1\], which shows the photoemission intensity along high-symmetry directions in the surface Brillouin zone for two different photon energies (96 and 75 eV for Fig. \[fig:1\](a) and (b), respectively). Assuming free-electron final states and using a procedure outlined below [@SMAT], the photon energies have been chosen such that data are collected approximately in the $\Gamma-M-K$ plane and $A-L-H$ plane of the bulk Brillouin zone for (a) and (b), respectively. For a definition of the high-symmetry points see Fig. \[fig:1\](c). Note, however, that only the parallel component of the crystal momentum $k_{\parallel}$ is conserved in the photoemission process. The high-symmetry points of the *surface* Brillouin zone are reached exactly. The $\Gamma$ and $A$ points in the *bulk* Brillouin are also reached rather precisely, but the $M,K,L,H$ points at finite $k_{\parallel}$ only approximately [@SMAT].
![(Color online) (a),(b) Photoemission intensity along high symmetry lines of the surface Brillouin zone for photon energies of 96 and 75 eV, respectively. Dark corresponds to high photoemission intensity. The high symmetry points in the surface Brillouin zone (noted with a bar over the letter) are reached exactly but the bulk high symmetry points only approximately (see text and Ref. [@SMAT]). (c) Sketch of the first Brillouin zone of $1T$-TaS$_2$ and its projection on the (0001) surface. (d) Energy dispersion curves at the $\bar{\Gamma}$ point for photon energies of 96 (red) and 75 eV (green), corresponding to the bulk $\Gamma$ and $A$ points, respectively.[]{data-label="fig:1"}](Fig1 "fig:"){width="45.00000%"}\
At first glance, the dispersions are very similar to each other and in good agreement with previous ARPES results [@Pillo:1999ac; @Arita:2004aa; @Bovet:2004ab; @Perfetti:2005aa; @Rossnagel:2005ab; @Clerc:2006aa; @Ritschel:2015aa]. The states close to the Fermi energy $E_F$ are broad due to the strongly correlated character of the material, with a lack of any clear Fermi level crossings. However, upon closer inspection, some differences between Fig. \[fig:1\](a) and (b) can be noted. The deeper lying, sulphur 2p-derived [@Mattheiss:1973aa; @Manzke:1988aa] states are expected to be less two-dimensional and do indeed show a different dispersion, for example around $\bar{M}$. The states near $E_F$, on the other hand, mostly differ in their intensity. Note, for example that the flat band immediately below $E_F$ near $\Gamma$ in Fig. \[fig:1\](a) is very well defined, while it is almost absent at $A$. This difference is best seen in a direct comparison of energy distribution curves through $\Gamma$ and $A$, as given in Fig. \[fig:1\](d).
Subtle differences in the states near the Fermi level are especially well seen in the $k_{\parallel}$-dependence of the photoemission intensity. Fig. \[fig:1b\](a) and (b) show such intensity plots at a binding energy of 90 meV, taken for the same photon energies as the data in Fig. \[fig:1\](a) and (b). The plots show distinct differences. In particular, the photoemission intensity reaches a maximum at normal emission (marked as $\bar{\Gamma}_0$) in (a) while it shows a local minimum in (b), indicating a dependence of the electronic structure on the perpendicular crystal momentum $k_{\perp}$.
![(Color online) (a),(b) Photoemission intensity 90 meV below the Fermi energy for photon energies of 96 and 75 eV, respectively. $\bar{\Gamma}_0$ corresponds to normal emission and to the bulk $\Gamma$ and $A$ points in (a) and (b) respectively. []{data-label="fig:1b"}](Fig2 "fig:"){width="45.00000%"}\
The possible existence of the predicted quasi one-dimensional band along $\Gamma-A$ can be established by collecting the photoemission intensity in normal emission as a function of photon energy $h\nu$ and binding energy $E_{bin}$. The result of such a scan is shown in Fig. \[fig:2\](a). The photoemission intensity $I$ in the figure has been converted from the raw data ($I$ measured as a function of $h\nu$ and $E_{bin}$) to a function of $k_{\perp}$ and $E_{bin}$, using the assumption of free-electron final states [@SMAT]. The most important feature in the data is the small electron pocket appearing near the Fermi energy around the $\Gamma$ points, i.e. at integer multiples of $k_{\perp}= 2 \pi / c$ ($c=5.86$ [Å]{} [@Wilson:1969]). For clarity, the intensity close to $E_F$ is magnified in Fig. \[fig:2\](b) and a detailed view of the situation at the highest $k_{\perp}$ is given in Fig. \[fig:4\]. The experimental observation of this metallic band is the central result of this paper.
![(Color online) (a) Photoemission intensity measured in normal emission as a function of photon energy $h\nu$, here converted to $k_{\perp}$ using free electron final states. $k_{\perp}$ values are given in units of the reciprocal lattice vector $2 \pi / c$. The $h\nu$ values given on the upper axis refer to the photon energy for the states at the Fermi energy. The greyscale is logarithmic. The dashed red lines mark the maximum binding energy of the small electron pocket near $\Gamma$ and the $k_{\perp}$ values mid-way between two $\Gamma$ points. (b) Magnification of the intensity in the vicinity of the Fermi energy. The intensity is normalized by an exponential function of the photon energy. The dashed-dotted green lines are the result of the calculation from Ref. [@Darancet:2014aa].[]{data-label="fig:2"}](Fig3 "fig:"){width="50.00000%"}\
It should be noted that recovering an approximate initial state $k_{\perp}$ using free-electron final states requires a choice of the solid’s inner potential $V_{0}$ and work function $\Phi$ [@Plummer:1982aa]. Here $V_{0}=20$ eV and $\Phi=4.5$ eV were chosen in order to place the periodically appearing electron pocket close to the $\Gamma$ point of the Brillouin zone. While somewhat different values for $V_0$ have been used in very early investigations [@Manzke:1988aa], we emphasize that the precise choice of this parameter is not critical. Indeed, due to the symmetry of the electron pocket’s dispersion, it can only be placed at either $\Gamma$ or $A$ and no choice of $V_0$ below 40 eV would lead to the electron pocket being found at $A$. Moreover, the choice of the inner potential is also confirmed by the symmetry of the deeper lying bands, especially the sulphur p-bands that show a binding energy minimum at $\Gamma$ [@Mattheiss:1973aa]. This minimum is most clearly identified at a binding energy of $\approx$1.7 eV for $k_{\perp}= 12 \pi / c$.
![(Color online) (a) Magnification of the data in Fig. \[fig:2\] around $k_{\perp}=12\pi /c$ with a momentum distribution curve ($k$-dependent photoemission intensity) at the Fermi energy on the top. (b) Energy distribution curves taken at the arrows in (a), i.e. near the Fermi level crossings and at maximum binding energy of the band. The curves are normalized and vertically displaced. []{data-label="fig:4"}](Fig4 "fig:"){width="50.00000%"}\
Fig. \[fig:2\](b) also shows the calculated band structure from Ref. [@Darancet:2014aa] superimposed on the data as dashed lines. The agreement is excellent. Other published calculations show very similar dispersions [@Bovet:2003aa; @Freericks:2009ab; @Ritschel:2015aa; @Lazar:2015aa]. Moreover, the observed $k_{\perp}$ periodicity of $2 \pi / c$ suggests a real space periodicity of only one unit cell and hence the $\tau_c$ stacking. This is consistent with calculations, most of which were made under the assumption of $\tau_c$ stacking.
On the other hand, a regular $\tau_c$ stacking does not agree with a substantial amount of structural information available on the commensurate CDW phase—see, e.g., Refs. [@Tanda:1984aa; @Nakanishi:1984aa; @Ishiguro:1991aa]. Data from different techniques reveal considerable disorder in the $c$ direction, accompanied by different stackings. This does not, however, imply a contradiction between the ARPES observations and the structural data. First of all, only periodic contributions to the stacking give rise to any regular band structure, while disorder merely increases the background intensity. Moreover, we note that the observed bands are rather broad, as seen in the momentum and energy distribution curves of Fig. \[fig:4\], suggesting that $k_{\perp}$ is not well defined. This can be partly due to disorder, in addition to the intrinsic uncertainty in $k_{\perp}$ that stems from the short inelastic mean free path of electrons in solids and the localisation accompanying this [@Plummer:1982aa].
An alternative way to view the $2 \pi / c$ periodicity is that, in the absence of a CDW, this would be expected for every band in the $1T$ polytype. Since the CDW is, after all, only a minor distortion of the lattice [@Rossnagel:2011aa], such a periodicity could still be present. This interpretation is supported by an accurate band structure calculation for the undistorted $1T$ structure (including the significant spin-orbit coupling [@Rossnagel:2006aa]) that shows a very similar metallic dispersion in the $\Gamma-A$ direction, even though the in-plane dispersion is completely different from the CDW case [@Darancet:2014aa].
The observed electron pocket around $\Gamma$ also appears to explain the distinct differences in the constant energy surfaces of Fig. \[fig:1b\], since the band is occupied at $\Gamma$ but empty at $A$. Indeed, Ritschel *et al.* have challenged the common view that this band is the lower Hubbard band of the Mott insulating state because it can be reproduced by a calculation not including correlations [@Ritschel:2015aa]. However, the situation is more complex because the metallic band dispersion appears to coexist with a part of the spectral weight at the original peak position of the lower Hubbard band, as seen in the energy distribution curves of Fig. \[fig:1\](d). Indeed, the peak that is usually assigned to the lower Hubbard band significantly changes its intensity but never entirely disappears, not even at the $A$ point of the bulk Brillouin zone (corresponding to h$\nu$=75 eV) where the strongly dispersing metallic band is predicted to be well above the Fermi level. Such a complex behaviour is not unexpected given the partially disordered character of the CDW along $c$ [@Tanda:1984aa; @Nakanishi:1984aa; @Ishiguro:1991aa], which should limit the formation of a well-defined band structure in this direction.
The observation of a metallic band is partly consistent with the reported transport phenomena in the material. In the temperature region immediately below the transition to the commensurate CDW ($\approx$50 – 80 K), a metallic temperature dependence of the resistivity is observed, consistent with the remaining metallic band, but at very low temperature the resistivity increases strongly [@Salvo:1977aa]. This has been ascribed to disorder-induced Anderson localization [@Salvo:1977aa], an interpretation that appears consistent with the observed disorder in the $c$ direction [@Tanda:1984aa; @Nakanishi:1984aa; @Ishiguro:1991aa].
Given the one-dimensional metallic dispersion, one would expect that the resistivity $\rho_c$ in the $c$ direction would be lower than the resistivity $\rho_a$ in the plane, but the opposite is found. In fact, a direct measurement of $\rho_c / \rho_a$ gives a value of approximately 500 [@Hambourger:1980aa], even in the temperature range of metallic conductance. The apparent contradiction of a metallic band with a lack of metallic conduction could be due to a gap opening in the one-dimensional dispersion near $E_F$. This is not supported by the detailed view on the dispersion in Fig. \[fig:4\] which appears to show clear Fermi level crossings. However, the rather broad features do not allow us to draw a definite conclusion about this type of gap formation. Moreover, the simultaneous presence of the dispersing band and the lower Hubbard band throughout the bulk Brillouin zone increase the difficulty of identifying a clear gap opening.
In conclusion, we have observed a one-dimensional metallic band in the $c$ direction of $1T$-TaS$_2$. This has recently been predicted by several calculations but hitherto never been observed. The result gives strong experimental support to the notion that viewing the TMDC CDW materials as quasi two-dimensional is an oversimplification. It also suggests that new rich physics can be expected from truly two-dimensional single layers of these compounds, not only because of the absence of interaction with neighboring crystal planes but also because the electronic properties can be influenced by substituting these planes with other materials of choice.
This work was supported by the Danish Council for Independent Research, Natural Sciences under the Sapere Aude program (Grants No. DFF-4002-00029 and DFF-6108-00409) and by VILLUM FONDEN via the Centre of Excellence for Dirac Materials (Grant No. 11744).
[35]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , ****, ().
, ****, ().
, ****, ().
, , , , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , , ****, ().
, , , , , , , ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , , , , , , ****, ().
, , , ****, ().
, , , , , , , ****, ().
, , , , , ****, ().
, , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , , , ****, ().
, , , , , ****, ().
, , , , , , , , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we address the challenging problem of optimal experimental design (OED) of constrained inverse problems. We consider two OED formulations that allow to reduce the experimental costs by minimizing the number of measurements. The first formulation assumes a fine discretization of the design parameter space and uses sparsity promoting regularization to obtain an efficient design. The second formulation parameterizes the design and seeks optimal placement for these measurements by solving a small-dimensional optimization problem. We consider both problems in a Bayes risk as well as an empirical Bayes risk minimization framework. For the unconstrained inverse state problem, we exploit the closed form solution for the inner problem to efficiently compute derivatives for the outer OED problem. The empirical formulation does not require an explicit solution of the inverse problem and therefore allows to integrate constraints efficiently. A key contribution is an efficient optimization method for solving the resulting, typically high-dimensional, bilevel optimization problem using derivative-based methods. To overcome the lack of non-differentiability in active set methods for inequality constraints problems, we use a relaxed interior point method. To address the growing computational complexity of empirical Bayes OED, we parallelize the computation over the training models. Numerical examples and illustrations from tomographic reconstruction, for various data sets and under different constraints, demonstrate the impact of constraints on the optimal design and highlight the importance of OED for constrained problems.'
author:
- 'Lars Ruthotto[^1]'
- 'Julianne Chung[^2]'
- 'Matthias Chung[^3]'
title: '[[Optimal Experimental Design for Constrained Inverse Problems]{}]{}[^4]'
---
experimental design, constrained optimization, tomographic reconstruction, Bayes risk, and empirical Bayes risk
62K05, 65F22, 80M50
Introduction {#sec:introduction}
============
The ability to optimally configure or design an experimental setup, or optimal experimental design (OED), can have significant benefits and improvements in a wide range of scientific and engineering applications [@Atkinson2007; @Pukelsheim2006]. Only recently has the focus shifted from OED for well-posed problems to OED for ill-posed inverse problems, where an unresolved challenge is how to efficiently include constraints on the state parameters (i.e., the solutions of the inverse problem) [@HaberEtAl2008OED; @HaberEtAl2010]. In this work, we investigate the impact of state constraints on the optimal design, and we propose a unified OED framework for linear inverse problems with linear equality and inequality constraints in the context of tomographic reconstruction.
Before introducing the OED problem, we first describe the discrete inverse problem. Given a design parameter $\bfp \in \bbR^\ell$ that describes the experiment setup, the observations are given as $$\bfd(\bfp) = \bfM(\bfp) \bff_{{\rm true}}+ \bfvarepsilon(\bfp),
\label{eq:forwardmodel}$$ where $\bff_{{\rm true}}\in \bbR^n$ is the exact model that we wish to reconstruct, $\bfM: \bbR^{\ell} \to \bbR^{m \times n}$ is the design-dependent forward operator with matrix $\bfM(\bfp)$ describing the parameter-to-observation map, and $\bfvarepsilon(\bfp)\in\bbR^{m}$ represents additive noise. Some examples of $\bfM(\bfp)$ include the map from image to sinogram in tomography (see, e.g., [@hsieh2003computed; @kak2001principles; @natterer2001mathematics]) or the map onto an observation space of the solution of a partial or ordinary differential equation (see, e.g., [@Arridge1999; @CheneyEtAl1999; @Haber2015]). In this work we assume that the measurements $\bfd(\bfp) \in\bbR^m$ depend on the underlying design of the experiment (e.g., $\bfp$ may determine the positions of the sources and/or detectors or represent the times at which measurements are taken). The noise can come from various sources, e.g., measurement errors, modeling errors, and numerical rounding errors. For simplicity we assume that $\bfvarepsilon(\bfp)$ is normally distributed with zero mean and known symmetric positive definite covariance matrix $\bfGamma_\bfvarepsilon(\bfp)\in \bbR^{m \times m}$ for any design $\bfp$.
We focus on ill-posed inverse problems where regularization in the form of prior knowledge is required to compute stable, reasonable approximations of $\bff_{{\rm true}}$ [@hansen2010discrete]. Following a Bayesian framework [@Calvetti:2007dq; @Kaipio:2006gx], we treat $\bff_{{{\rm true}}}$ as a random variable with a truncated multivariate normal distribution with probability density given as $$\pi(\bff_{{\rm true}}) =
\begin{cases}
c\,{\textnormal{e}}^{-\tfrac{\gamma^2}{2} (\bff_{{{\rm true}}}-\bfmu)\t\bfGamma_\bff^{-1}(\bff_{{{\rm true}}}-\bfmu) } , & \text{if } \bfC_{\rm e} \bff_{{{\rm true}}} - \bfc_{\rm e} = \bfzero$ and $\bfC_{\rm i} \bff_{{{\rm true}}} - \bfc_{\rm i} \geq \bfzero, \\
0, & \text{otherwise,}
\end{cases}$$ with appropriate constants $c,\gamma > 0$ and given positive definite covariance matrix $\bfGamma_\bff \in \bbR^{n\times n}$ and mean $\bfmu \in \bbR^n$. Here, the pairs $\bfC_{\textnormal{e}}\in \R^{m_e \times n}, \bfc_{\textnormal{e}}\in \R^{m_e}$ and $\bfC_{\rm i}\in \R^{m_i \times n}$, $\bfc_{\rm i}\in \R^{m_i}$ define linear equality and inequality constraints on $\bff$, respectively. Denoting by $\bfI_n \in \R^{n\times n}$ the identity matrix, for example, bound constraints correspond to choosing $\bfC_{\rm i} = [\bfI_n; -\bfI_n] \in \R^{2n\times n}$ and $\bfc_{\rm i} = [\bff_{\rm L}; -\bff_{\rm H}]\in\R^{2n}$ where $\bff_{\rm L}$ and $\bff_{\rm H}$ contain lower and upper bounds respectively (e.g., non-negativity constraints are reasonable when reconstructing density images). Furthermore, setting $\bfC_{\textnormal{e}}= \bfe^\top$, where $\bfe$ is a vector of all ones, allows one to fix the integral (or mass) of $\bff$, which might be helpful in applications such as emission tomography. A wide range of prior knowledge can be included to estimate $\bff_{{{\rm true}}}$ in this formulation. The unconstrained case reduces to a simple Gaussian distribution on the entire domain [@calvetti2007introduction] and can provide insight in the Bayes formulation.
For fixed $\bfp,$ the maximum a-posteriori (MAP) estimate $\widehat \bff(\bfp)$ provides an estimator to and can be obtained by minimizing the negative log likelihood of the posterior probability distribution function, i.e., $$\begin{gathered}
\begin{split}\label{eq:ip}
\widehat \bff(\bfp) = \argmin_{\bff} \tfrac{1}{2}{\left\lVert[\right\rVert}\bfGamma^{-1}_\bfvarepsilon(\bfp)]{\bfM(\bfp)\bff -\bfd(\bfp)}^2 + \tfrac{\gamma^2}{2}{\left\lVert[\right\rVert}2]{\bfL(\bff -\bfmu)}^2\\
\text{subject to } \ \bfC_{\rm e} \bff - \bfc_{\rm e} = \bfzero \text{ and } \bfC_{\rm i} \bff - \bfc_{\rm i} \geq \bfzero\,,
\end{split}\end{gathered}$$ where $ \bfL\t\bfL = \bfGamma_\bff^{-1}$. In general, obtaining $\widehat \bff(\bfp)$ requires solving a constrained optimization problem, which in this case is a convex quadratic programming problem. In the absence of inequality constraints, the optimal solution to depends linearly on the data, which can be used for efficient optimization of the design parameter; see [@HaberEtAl2008OED]. In the presence of inequality constraints however, the optimality condition becomes nonlinear in the data and, thus, solving becomes computationally challenging. We assume a unique minimizer of the constrained optimization problem exists.
In this work we consider optimal design problems that aim at selecting an experimental design that not only leads to most accurate estimates of $\bff_{{\rm true}}$ but also minimizes experimental costs. Finding a balance between these often conflicting goals can be challenging. For example, in tomographic reconstruction, obtaining more projection data may lead to higher resolution image reconstructions, but more scans may lead to more harmful radiation to the patient, longer scan times, and higher operational costs.
For our OED problem, the goal is to obtain design parameters $\widehat\bfp$ that minimize the Bayes risk, i.e., the expected value of the mean squared reconstruction errors, while simultaneously minimizing measurement costs. This requires solving an optimization problem of the form, $$\label{eq:designproblem}
\begin{split}
\min_{\bfp \in \Omega}& \quad \calJ(\bfp) = \bbE \ \tfrac{1}{2}{\left\lVert[\right\rVert}2]{\widehat\bff(\bfp) - \bff_{{\rm true}}}^2 + \calR(\bfp),
\end{split}$$ where $\Omega\subset\R^p$ is the set of feasible design parameters, $\bbE$ is the expected value where $\bff_{{\rm true}}$ and $\bfvarepsilon(\bfp)$ are random variables, and the functional $\calR$ encodes the measurement costs. Two formulations in the context of tomographic reconstruction will be described in Section \[sec:problem\_setup\].
The literature on OED methodologies is vast, with a range of techniques tailored to various optimality criteria in both Bayesian [@chaloner1995bayesian] and non-Bayesian settings. Classic OED works include [@Atkinson2007; @pazman1986foundations; @Pukelsheim2006], but a recent large effort has been made to develop efficient algorithms for obtaining designs that minimize a loss function of the Fisher information matrix and to new applications, e.g., in biology and exploratory drilling [@HaberEtAl2008OED; @HaberEtAl2010; @haber2012numerical; @Nowak2010]. Many of these approaches follow an empirical Bayes risk minimization framework and exploit the case where the model parameters, here $\widehat \bff(\bfp)$, depend linearly on the observables $\bfd(\bfp).$ This is not necessarily the case for the constrained inverse problems of interest here. Extensions to nonlinear problems have been considered in [@alexanderian2016fast; @Chung2012; @huan2013simulation], and an approach based on consistent Bayesian inference was developed in [@butler2017consistent] and used for OED in [@walsh2017optimal].
The primary goal in many design problems is to enforce sparse sampling in the design parameters. Sparsity enforcing regularization in the context of ODE was considered in [@alexanderian2014optimal; @Chung2012; @haber2012numerical; @haber2015optimal], but no additional state constraints were considered. The main contribution of this work is the inclusion of state constraints in OED frameworks, which is a critical yet missing piece in the shift from OED for well-posed to OED for ill-posed problems. Two open questions include (1) how does the inclusion of state constraints *impact* the optimal design? And (2) how does one *efficiently* incorporate constraints in an OED framework? In this paper, we address both of these questions by developing and investigating a unified OED framework for constrained inverse problems.
An overview of this paper is as follows. In Section \[sec:problem\_setup\] we describe two problem formulations for OED in the context of tomographic reconstruction, one of which assumes a fine discretization of the parameter space and enforces sparsity of the design and the other assumes a fixed number of design parameters and seeks optimal locations for these measurements. In Section \[sec:problemformulation\] we investigate various problem formulations for the OED problems that demonstrate the range of problems and constraints that can be addressed. We describe efficient computational approaches in Section \[sec:computational\]. In Section \[sec:numerical\_experiments\], we demonstrate the impact of state constraints on the design, which is something that has not been investigated before, and we provide numerical results that demonstrate the effectiveness of our approach. Conclusions are provided in Section \[sec:conclusions\].
Motivating Application: Tomographic Reconstruction {#sec:problem_setup}
==================================================
Although the methods described in this paper extend to more general problems and applications, we focus our discussion on two problem setups in the context of tomographic reconstruction.
Tomography is a widely used imaging technique where penetrating waves (e.g., x-ray or acoustic waves) transverse an object and are collected or detected [@hsieh2003computed; @kak2001principles; @natterer2001mathematics]. Oftentimes, multiple projections are made as the wave source rotates around the object. Then given these observed measurements, the goal of the inverse problem is to reconstruct the interior properties (e.g., densities) of the object at different locations; see Figure \[fig:tomo\].
Mathematically, we can model the transmission process for one projection via a (sparse) matrix $\bfT \in \bbR^{n_r \times n}$ where $n_r$ is the number of rays. Then the noise-free projection data obtained by rotating the source $\theta$ degrees clockwise can be modeled as $\bfd(\theta) = \bfT \bfR(\theta) \bff_{{\rm true}}\,,$ where $\bfR(\theta)\in \bbR^{n \times n}$ rotates the object by $\theta$ degrees counterclockwise. Typically we assume $\theta \in [0, 180].$ Next we describe two OED problems.
#### OED Problem A
In the first scenario, we assume that a fine discretization of the set of projection angles $\bftheta = [\theta_1, \ldots, \theta_\ell]\t$ is given and aim at identifying the angles that provide the most important measurements. To this end, we introduce the design parameters $\bfp\in \bbR^\ell$ with $\bfp \geq \bfzero$ whose components encode the importance of each measurement. Define the ordered index set $\calI(\bfp) = \{i \, : \, p_i > 0 \}$ and denote $k(\bfp)$ to be the cardinality of $\calI(\bfp).$ We define matrix $\bfE(\bfp) \in \bbR^{\ell \times k(\bfp)}$ to contain the $i$-th standard basis vectors for $i \in \calI(\bfp).$ Then, $\bfM(\bfp)$ and $\bfd(\bfp)$ are given by $$\bfM(\bfp) =
\bfP\left( \bfI_\ell \kron \bfT \right) \begin{bmatrix} \bfR(\theta_1)\\ \vdots \\ \bfR(\theta_\ell) \end{bmatrix}\,, \qquad \text{and} \qquad \bfd(\bfp) = \bfP\bfd,$$ where $\bfP = \left(\bfE(\bfp)\t \diag(\bfp)\right) \otimes \bfI_{n_r} \in \bbR^{k(\bfp) n_r \times \ell n_r}.$ Notice that $\bfE(\bfp)\t \diag(\bfp)$ essentially extracts all of the non-zero rows of $\diag(\bfp).$ We assume that $\bfGamma_\bfvarepsilon(\bfp)^{-1} = \sigma^2 \bfI_{k(\bfp) n_r}$.
Since we would like to keep the number of projection angles low, we incorporate a sparsity-inducing prior on the design parameters. That is, we use the $\ell_1$-norm $\calR(\bfp) = \beta {\left\lVert[\right\rVert}1]{\bfp}$ with $\beta>0$ large enough so that we can get a relaxation of $\|\cdot\|_0$ and enforce sparsity in the design parameters [@candes2006stable; @candes2005decoding]. In the Bayesian framework, this is equivalent to imposing a Laplace prior distribution on $\bfp$ (see e.g., [@huang2013optimal; @johnson1972distributions]). Furthermore, we assume $\Omega = \R^{\ell}_+$ (non-negative orthant) such that continuous optimization methods can be used.
In summary, *OED problem A* can be written as the bilevel optimization problem $$\begin{gathered}
\min_{\bfp \geq \bfzero} \quad \bbE \ \tfrac{1}{2}{\left\lVert[\right\rVert}2]{\widehat\bff(\bfp) - \bff_{{\rm true}}}^2 + \beta {\left\lVert[\right\rVert}1]{\bfp} \label{eq:designproblemA}\\
\text{subject to} \nonumber\\
\hat \bff(\bfp) = \argmin_{\bff} \tfrac{1}{2}{\left\lVert[\right\rVert}2]{\bfM(\bfp)\bff -\bfd(\bfp)}^2 + \tfrac{\alpha^2}{2}{\left\lVert[\right\rVert}2]{\bfL(\bff -\bfmu)}^2 \label{eq:ipA}\\
\text{subject to } \ \bfC_{\rm e} \bff - \bfc_{\rm e} = \bfzero \text{ and } \bfC_{\rm i} \bff - \bfc_{\rm i} \geq \bfzero\,, \nonumber\end{gathered}$$ where $\alpha = \gamma/\sigma$ and zero values in $\bfp$ correspond to zeroing out measurements for the corresponding angle. After computing such an $\ell_1$-regularized design, the solution can be used to identify important components, and a second optimization can be done to optimize the weights of the non-zero components (see [@HaberEtAl2008OED]). While the sparsity of the design generally depends on the choice of $\beta$ (e.g., the larger $\beta$ the fewer non-zero elements in the design vector) a relevant issue in some applications is identifying $\beta$ so that a design with a given number of measurements is obtained.
#### OED Problem B
In the second scenario, we present a method that optimizes the design parameters for a fixed number of measurements. Suppose $\bfp\in \Omega \subset \bbR^\ell$, $\ell < m,$ e.g., for tomography it contains angles $ \bfp_{\rm L}\leq \bfp \leq \bfp_{\rm H}$ corresponding to locations of the sources. Then in the inverse problem, $$\bfM(\bfp) = \left( \bfI_\ell \kron\bfT \right) \bfR(\bfp), \quad \mbox{ where } \quad \bfR(\bfp) = \begin{bmatrix} \bfR(p_1) \\ \vdots \\ \bfR(p_\ell) \end{bmatrix}$$ and $\bfGamma_\bfvarepsilon = \sigma^2\bfI_m$. Assuming that no additional regularization is required for $\bfp$ (other than the above described reparameterization), *OED problem B* reads $$\begin{gathered}
\min_{\bfp \in \Omega} \quad \bbE \ \tfrac{1}{2}{\left\lVert[\right\rVert}2]{\widehat\bff(\bfp) - \bff_{{\rm true}}}^2 \label{eq:designproblemB}\\
\text{subject to} \nonumber\\
\hat \bff(\bfp) = \argmin_{\bff} \tfrac{1}{2}{\left\lVert[\right\rVert}2]{\bfM(\bfp)\bff -\bfd(\bfp)}^2 + \tfrac{\alpha^2}{2}{\left\lVert[\right\rVert}2]{\bfL(\bff -\bfmu)}^2 \label{eq:ipB}\\
\text{subject to } \ \bfC_{\rm e} \bff - \bfc_{\rm e} = \bfzero \text{ and } \bfC_{\rm i} \bff - \bfc_{\rm i} \geq \bfzero \,. \nonumber
\end{gathered}$$
Note that mathematically OED problem A and B differ in the particular choice of $\bfM(\bfp)$, $\bfd(\bfp)$ and the sparsity regularization term. We refer to problems and as the design problems and problems and as the underlying inverse or state problems.
Design Problem Formulations {#sec:problemformulation}
===========================
In this section, we consider OED problems and and discuss two problem formulations, each of which may have advantageous properties depending on problem assumptions and constraints.
#### Bayes risk minimization
Notice that the expected values in and are defined in terms of the distributions of $\bff_{{\rm true}}$ and $\bfvarepsilon(\bfp)$. For problems where such knowledge is available or can be well approximated (e.g., by the sample mean or sample covariance), we investigate Bayes risk minimization for both design problems. This approach assumes that no inequality constraints are included on the inverse state problems and . While the theory can be extended to equality constrained problems, we consider unconstrained inverse problems for simplicity. Under these assumptions the MAP estimate is given by $$\label{eq:lsproblem}
\widehat\bff(\bfp) = \bfQ(\bfp)^{-1}\left( \bfM(\bfp)\t \left(\bfM(\bfp)\bff_{\rm true} + \bfvarepsilon(\bfp)\right)+\alpha^2 \bfL\t \bfL \bfmu \right),$$ where $\bfQ(\bfp) = \bfM(\bfp)\t \bfM(\bfp) + \alpha^2 \bfL\t\bfL\,.$ Then the design objective (for convenience omitting the design costs expressed by $\calR$) can be written as $$\begin{aligned}
\calJ(\bfp) & = \tfrac{1}{2} \bbE {\left\lVert[\right\rVert}2]{\bfQ(\bfp)^{-1} \left( \bfM(\bfp)\t \left(\bfM(\bfp)\bff_{\rm true} + \bfvarepsilon(\bfp)\right)+\alpha^2 \bfL\t \bfL \bfmu \right) - \bff_{\rm true}}^2\\
& = \tfrac{1}{2} \bbE {\left\lVert[\right\rVert}2]{\left(\bfQ(\bfp)^{-1} \bfM(\bfp)\t \bfM(\bfp) -\bfI_n \right) \bff_{\rm true} + \bfQ(\bfp)^{-1} (\bfM(\bfp)\t \bfvarepsilon(\bfp)+\alpha^2 \bfL\t \bfL \bfmu)}^2\,.\end{aligned}$$ Let $\bfK(\bfp) = \bfQ(\bfp)^{-1} \bfM(\bfp)\t \bfM(\bfp) -\bfI_n$ and denote $\trace(\cdot)$ to be the trace of a matrix, then by utilizing the quadratic form property $\bbE(\bfdelta\t\bfLambda\bfdelta) = \bfmu_\bfdelta\t\bfLambda \bfmu_\bfdelta + \trace(\bfLambda\bfGamma_\bfdelta)$ along with the above assumptions, we get $$\begin{aligned}
2 \calJ(\bfp) =\ & \bbE\Big( \bff_{\rm true}\t \bfK(\bfp)\t\bfK(\bfp) \bff_{{\rm true}}+2 \bff_{\rm true}\t \bfK(\bfp)\t \bfQ(\bfp)^{-1} \left(\bfM(\bfp)\t \bfvarepsilon(\bfp)+\alpha^2 \bfL\t \bfL \bfmu \right) \\
+& \ \left(\bfM(\bfp)\t \bfvarepsilon(\bfp)+\alpha^2 \bfL\t \bfL \bfmu\right)\t \bfQ(\bfp)^{-\top} \bfQ(\bfp)^{-1} \left(\bfM(\bfp)\t \bfvarepsilon(\bfp)+\alpha^2 \bfL\t \bfL \bfmu\right) \Big)\\
=\ & \bfmu\t \bfK(\bfp)\t\bfK(\bfp) \bfmu + \gamma^{-2} \trace\left(\bfK(\bfp)\t\bfK(\bfp) \bfGamma_\bff\right) + 2 \alpha^2 \bfmu\t \bfK(\bfp)\t \bfQ(\bfp)^{-1} \bfL\t \bfL \bfmu \\
+& \ \sigma^{-2} \trace\left(\bfM(\bfp) \bfQ(\bfp)^{-\top} \bfQ(\bfp)^{-1} \bfM(\bfp)\t\right) + \alpha^4 \bfmu\t \bfL\t \bfL \bfQ(\bfp)^{-\top} \bfQ(\bfp)^{-1} \bfL\t \bfL \bfmu.\end{aligned}$$ Notice that the first, third, and fifth term sum up to ${\left\lVert[\right\rVert}2]{\left(\bfK(\bfp)+\alpha^2 \bfQ(\bfp)^{-1} \bfL\t \bfL\right)\bfmu}^2 = 0$. Thus, we get $$\begin{aligned}
\calJ(\bfp) & = \tfrac{1}{2\gamma^2 } {\left\lVert[\right\rVert}\rm F]{\bfK(\bfp)\bfL^{-1}}^2 + \tfrac{1}{2\sigma^2} {\left\lVert[\right\rVert}\rm F]{\bfQ(\bfp)^{-1} \bfM(\bfp)\t}^2\\
& = \tfrac{1}{2\sigma^2} {\left\lVert[\right\rVert}\rm F]{\bfQ(\bfp)^{-1} \begin{bmatrix} \bfM(\bfp) \\ \alpha \bfL \end{bmatrix}\t }^2 \\
& = \tfrac{1}{2\sigma^2} {\left\lVert[\right\rVert}\rm F]{\bfM_\alpha^\dagger (\bfp) }^2\,,\end{aligned}$$ where $\bfM_\alpha(\bfp) = \begin{bmatrix} \bfM(\bfp) \\ \alpha \bfL \end{bmatrix}$, the $\dagger$ in the last equation denotes the Moore-Penrose pseudoinverse, and ${\left\lVert[\right\rVert}\rm F]{\,\cdot\,}$ denotes the Frobenius norm. Recall that the goal of OED is to find design parameters that minimize $\calJ(\bfp)$. Thus, the significance of this result is that in the Bayes formulation with an unconstrained state problem, the optimal design parameters will correspond to a regularized coefficient matrix that has smallest pseudoinverse in the Frobenius norm sense (with an additional regularization term for OED problem A). Due to the nonlinear dependence on $\bfp,$ the design problem does not admit a closed-form solution; however, numerical methods and computational simplifications can be used to obtain optimal parameters (see Section \[sec:computational\]).
#### Empirical Bayes risk minimization
For problems where distributions of $\bff_{{\rm true}}$ and $\bfvarepsilon(\bfp)$ may be unknown or not obtainable, we consider empirical Bayes risk design problems, where training or calibration data $\bff_{{\rm true}}^{(k)}$, $k = 1, \ldots, K$ are used to approximate the expected value. Such data are often readily available, and in the case of tomography provide a clear understanding of how images may look. Stochastic programming methods such as stochastic average approximation (SAA) or stochastic approximation (SA) can be used to incorporate these training data in OED problems A and B. Solving stochastic optimization problems often requires computationally intensive techniques in order to obtain a good approximation of the expected value [@carlin1997bayes; @chung2017stochastic; @shapiro2009lectures]. Here we consider an SAA approach, but remark that for very large numbers of training data, this approach may not be feasible. However, a benefit of this formulation, compared to the Bayes risk minimization procedure described above, is the ability to incorporate constraints on the state problem and take advantage of existing constrained optimization algorithms.
We follow the *empirical* Bayes approach in [@Chung2012; @HaberEtAl2008OED; @HaberEtAl2010] that treats given training data as samples from the distribution and uses the sample mean to approximate the expected value. Note that, due to the presence of the regularization, the estimator given by will be biased unless the true solution is in the nullspace of $\bfL$. Thus, design choices cannot solely be based on properties of the forward operator $\bfM(\bfp)$ (see also the the discussion in [@HaberEtAl2008OED]).
Assume that we are given a set of training data that consists of $N$ true models $\bff^{(1)}_{{\rm true}}, \ldots, \bff^{(N)}_{{\rm true}}\in\R^n$. For a fixed design parameter $\bfp$, we can simulate datasets $\bfd^{(1)}(\bfp),\ldots,\bfd^{(N)}(\bfp) \in\R^m$ using and obtain reconstructions $\widehat{\bff}^{(1)}(\bfp),\ldots,\widehat{\bff}^{(N)}(\bfp) $ by solving the constrained inversion problem using that data. Then we can approximate the Bayes risk OED problem with the following empirical Bayes risk OED problem, $$\label{eq:OED}
\begin{split}
\min_{\bfp \in \Omega} &\quad \calJ_N(\bfp) = \tfrac{1}{2N} \sum_{i=1}^N \left \| \widehat{\bff}^{(i)}(\bfp) - \bff_{{\rm true}}^{(i)} \right\|^2 + \calR(\bfp)\\
\text{ subject to} & \quad \widehat{\bff}^{(i)}(\bfp) \text{ solves}~\eqref{eq:ip} \text{ for data } \bfd^{(i)}(\bfp).
\end{split}$$
The bilevel optimization problem could be solved as a large constrained optimization problem. Instead we follow a technique commonly used in the PDE constrained optimization literature where we eliminate the constraint by solving for $\widehat{\bff}^{(i)}(\bfp), $ for $i=1,2,\ldots,N$ yielding the *reduced* problem. Although both approaches have their merits, the reduced problem might be more attractive, especially if $N\gg1$. Furthermore, the reduced problem allows for parallel computing, since the constraints can be eliminated independently.
Assuming that $\Omega \subset \R^{\ell}$ is closed and convex, can be solved, for example, using a Projected Steepest Descent or Projected Gauss-Newton method (see also [@HaberEtAl2010]). In both cases, we need to compute the gradient of the design objective function, $$\label{eq:OEDgradient}
\nabla_{\bfp} \calJ_N(\bfp) = \tfrac{1}{N} \sum_{i=1}^N \nabla_{\bfp}\widehat{\bff}^{(i)} \left(\widehat{\bff}^{(i)}(\bfp) - \bff_{{\rm true}}^{(i)} \right) + \nabla_{\bfp} \calR\,,$$ where $\nabla_{\bfp}\widehat{\bff}^{(i)}$ contains the derivatives of the reconstructed image $\widehat{\bff}^{(i)}(\bfp)$ with respect to the design parameters. For OED problem A, where $\calR(\bfp) =\beta \|\bfp\|_1$ and $\bfp \geq \bf0$, computing the gradient for the regularization term is straightforward. However, the more challenging computation involves the sensitivity matrix $\nabla_{\bfp}\widehat{\bff}^{(i)}$ that is solver-dependent and outlined in the following section.
Computational approaches for OED {#sec:computational}
================================
In this section, we describe efficient computational approaches for computing solutions to OED problems A and B.
#### Bayes risk minimization
We begin with computational simplifications that can be used for the Bayes risk minimization problem. First we consider the case where $\bfL =\bfI_n$ (standard Tikhonov regularization). Assume $\bfM(\bfp)\in \bbR^{m \times n}$ with rank $r \leq \min(m,n)$ and let $\bfM(\bfp) = \bfU(\bfp) \bfSigma(\bfp) \bfV(\bfp)\t$ be the SVD where the non-zero diagonal elements of $\bfSigma(\bfp)$ are $\sigma_1(\bfp) \geq \ldots \geq \sigma_r(\bfp)$. Then, with some algebraic manipulations, it can be shown that design objective (again omitting the regularization term for convenience) can be written as $$\begin{aligned}
\calJ(\bfp) & = \tfrac{1}{2\sigma^2} {\left\lVert[\right\rVert}\rm F]{(\bfSigma(\bfp)\t \bfSigma(\bfp) + \alpha^2 \bfI_n)^{-1} \begin{bmatrix}\bfSigma(\bfp)\t & \alpha \bfI_n \end{bmatrix} }^2 \\
& = \tfrac{1}{2\sigma^2} \left(\sum_{i=1}^r \frac{1}{\sigma_i(\bfp)^2 + \alpha^2} + \frac{n-r}{\alpha^2} \right)\,.\end{aligned}$$ Note that the second term should not be ignored since the rank of $\bfM(\bfp)$ may depend on $\bfp.$
For the general case where $\bfL\neq \bfI_n,$ the generalized SVD could be used to simplify the objective function as $$\calJ(\bfp) = \tfrac{1}{2\sigma^2} {\left\lVert[\right\rVert}\rm F]{ \bfX(\bfp)^{-1}(\bfSigma(\bfp)\t \bfSigma(\bfp) + \alpha^2 \bfPsi(\bfp)\t \bfPsi(\bfp))^{-1/2} }^2\,,$$ where $$\bfM(\bfp) = \bfU(\bfp) \bfSigma(\bfp) \bfX(\bfp) \t \quad \mbox{and} \quad \bfL = \bfV(\bfp) \bfPsi(\bfp) \bfX(\bfp)\t$$ and $\bfU(\bfp)\in \bbR^{m\times m}$ and $\bfV(\bfp)\in \bbR^{n\times n}$ are orthogonal matrices, $\bfSigma(\bfp)\in \bbR^{m\times n}$ and $\bfPsi(\bfp)\in \bbR^{n\times n}$ are diagonal matrices and $\bfX(\bfp) \in \bbR^{n \times n}$ is a nonsingular matrix [@stewart2001matrix]. A simpler approach for $\bfL\neq \bfI_n$ is to compute the singular values of $\bfM_\alpha(\bfp)$ denoted $\sigma_{\alpha,1}(\bfp), \ldots, \sigma_{\alpha,n}(\bfp) > 0$, in which case $$\calJ(\bfp) = \tfrac{1}{2\sigma^2} \sum_{i=1}^n \left(\sigma_{\alpha,i}(\bfp)\right)^{-2}\,.$$
Next we exploit the specific structure of OED problem A to analyze the dependence on $\bfp$. Assume that all $\bfp >\bfzero$ so that $\bfE(\bfp) = \bfI_\ell$, then $$\bfM(\bfp) = (\diag(\bfp) \otimes \bfI_{n_r}) \begin{bmatrix} \bfT \bfR(\theta_1) \\ \vdots \\ \bfT \bfR(\theta_\ell)\end{bmatrix} = \begin{bmatrix} p_1 \bfT \bfR(\theta_1) \\ \vdots \\ p_\ell \bfT \bfR(\theta_\ell)\end{bmatrix}\,.$$ The singular values of $\bfM(\bfp)$ are given in the following Lemma.
Let $\bfA_j \in \bbR^{m \times n}$ for $j = 1, \ldots, \ell$, then the unsorted singular values of $(\diag(\bfp) \kron \bfI) \begin{bmatrix} \bfA_1 \\ \vdots \\ \bfA_\ell\end{bmatrix}$ are given by $\sqrt{\sum_{j=1}^\ell \left(\sigma_i^{(j)}\right)^2 p_j^2}, i=1, \ldots, n$ where $\sigma_i^{(j)}$ is the $i$-th singular value of $\bfA_j$.
The result follows from the fact that the eigenvalues of $\sum_{j=1}^\ell p_j^2 \bfA_j\t \bfA_j$ are given as $\sum_{j=1}^\ell \left(\sigma_i^{(j)}\right)^2 p_j^2$.
Thus, if we let $\bfA_j = \bfT \bfR(\theta_j)$ and define $\bfPi \in \bbR^{n \times \ell}$ such that the $j$-th column of $\bfPi$ contains the squares of the singular values of $\bfA_j$, then the squares of the singular values of $\bfM(\bfp)$ are given in the vector $\bfh = \bfPi \diag(\bfp)\bfp.$ This formulation enables us to compute the gradient as $$ \nabla_\bfp \calJ(\bfp) = - \sigma^{-2} \diag(\bfp) \bfPi\t \bfh_\alpha + \beta {\rm sign}(\bfp),$$ where $\bfh_\alpha = \begin{bmatrix} (h_1 + \alpha^2)^{-2}, \ldots, (h_n + \alpha^2)^{-2}\end{bmatrix}\t$ with $h_i$ being the $i$-th element of $\bfh$. Notice that a critical point occurs if $\bfp=\bfzero$ (ignoring the non-differentiability of ${\left\lVert[\right\rVert}1]{\bfp}$ at $\bfp = \bfzero$), but this implies that $\bfM(\bfp)=\bfzero$ which is not reasonable. Thus, to obtain optimal design parameters for OED problem A, a nonlinear solver must be used.
#### Empirical Bayes risk minimization
In the absence of a closed form expression for the MAP estimate (e.g., due to the presence of inequality constraints) we follow an SAA approach and use gradient-based optimization methods to solve the empirical Bayes risk problem . To ensure differentiability, we propose using interior point methods for solving the constrained inverse state problems and . Although active set methods could also be used for solving problems such as , we prefer interior point methods because the ultimate goal is fast optimization of the design, and interior point methods enable fast sensitivity computations, i.e., methods for computing the derivatives of the reconstructed solution with respect to the (possibly relaxed) design parameters. Before deriving the sensitivities we describe the interior point method used in our experiments, which is a standard primal dual interior point method for quadratic programming based on Mehrotra’s predictor-corrector approach; a more detailed description can be found, e.g., in [@NocedalWright2006 Ch.16].
We first rewrite the constrained optimization problem as a quadratic program $$\min_{\bff} {\tfrac{1}{2}}\bff^\top \bfQ(\bfp) \bff + \bfb(\bfp)^\top \bff
\quad\text{ subject to }\quad \bfC_{\textnormal{e}}\bff - \bfc_{\textnormal{e}}= \bfzero, \quad \bfC_{\textnormal{i}}\bff - \bfc_{\textnormal{i}}\geq \bfzero \,,$$ where $\bfQ(\bfp) = \bfM(\bfp)^\top \bfM(\bfp) + \alpha^2 \bfL^\top \bfL$ and $\bfb(\bfp) = -\bfM(\bfp)^\top \bfd(\bfp)-\alpha^2 \bfL\t \bfL \bfmu$. To deal with the linear inequality constraints, we introduce slack variables $\bfs \in \R^{m_i}$, yielding the equivalent problem $$\label{eq:QPslack}
\min_{\bff,\bfs} {\tfrac{1}{2}}\bff^\top \bfQ(\bfp) \bff + \bfb(\bfp)^\top \bff
\quad\text{ subject to }\quad \bfC_{\textnormal{e}}\bff - \bfc_{\textnormal{e}}= \bfzero, \quad \bfC_{\textnormal{i}}\bff - \bfc_{\textnormal{i}}- \bfs = \bfzero,\quad \bfs \geq \bfzero.$$ This is a convex quadratic optimization problem, whose objective function is strictly convex if ${\rm null}(\bfM(\bfp)) \cap {\rm null}(\bfL) = \emptyset$. The Lagrangian is given by $$\label{eq:Lag}
\mathcal{L}(\bff, \bflambda_{\textnormal{e}},\bfs, \bflambda_{\textnormal{i}}) = {\tfrac{1}{2}}\bff^\top \bfQ(\bfp) \bff + \bfb(\bfp)^\top \bff - \bflambda_{\textnormal{e}}^\top (\bfC_{\textnormal{e}}\bff - \bfc_{\textnormal{e}}) - \bflambda_{\textnormal{i}}^\top (\bfC_{\textnormal{i}}\bff - \bfc_{\textnormal{i}}- \bfs).$$ Necessary and sufficient conditions for a global minimizer are the KKT conditions. Here, we consider the perturbed conditions for some centrality parameter $\delta \geq 0$, $$\label{eq:KKTCond}
F(\bff,\bflambda_{\textnormal{e}},\bfs,\bflambda_{\textnormal{i}},\delta,\mu) = \left[
\begin{array}{c}
\bfQ(\bfp) \bff + \bfb(\bfp) - \bfC_{\textnormal{e}}^\top \bflambda_{\textnormal{e}}- \bfC_{\textnormal{i}}^\top \bflambda_{\textnormal{i}}\\
\bfC_{\textnormal{e}}\bff - \bfc_{\textnormal{e}}\\
\bfC_{\textnormal{i}}\bff - \bfc_{\textnormal{i}}- \bfs\\
\bfS \bfLambda_{\textnormal{i}}\bfe - \delta \mu \bfe
\end{array}
\right] = \bfzero,\quad \quad \bflambda_{\textnormal{i}}, \bfs \geq \bfzero.$$ Here, $\bfS = \diag(\bfs)$, $\bfLambda_{\textnormal{i}}= \diag(\bflambda_{\textnormal{i}})$ and $\bfe \in \R^{m_i}$ is a vector of all ones, and $\mu = \frac{\bfs^\top \bflambda_{\textnormal{i}}}{m_i}$ is a complementarity measure (it is zero at a KKT point). In interior point methods, the goal is to iteratively approximate a root of $F$ using Newton’s method and some line search that ensures strict positivity of the slack and the Lagrange multiplier associated with the inequality constraints (i.e., $\bfs > \bfzero$ and $\bflambda_{\textnormal{i}}> \bfzero$).
At the $k$-th iteration of the interior point method, we denote the current iterates as $(\bff^k, \bflambda_{\textnormal{e}}^k, \bfs^k, \bflambda_{\textnormal{i}}^k)$, linearize the optimality conditions , and solve the linear system $$\label{eq:KKTnewton}
\left[
\begin{array}{c|c|c|c}
\vdots & \vdots & \vdots & \vdots \\
\nabla_{\bff} F & \nabla_{\bflambda_{\textnormal{e}}} F & \nabla_{\bfs} F & \nabla_{\bflambda_{\textnormal{i}}} F\\
\vdots & \vdots & \vdots & \vdots
\end{array}
\right] \left[
\begin{array}{c}
\Delta \bff \\ \Delta \bflambda_{\textnormal{e}}\\ \Delta \bfs \\ \Delta \bflambda_{\textnormal{i}}\end{array}
\right]
=
\left[
\begin{array}{c}
- \bfr_d(\bfp)\\
- \bfr_{\textnormal{e}}\\
- \bfr_{\textnormal{i}}\\
- \bfS \bfLambda_{\textnormal{i}}\bfe + \delta \mu \bfe
\end{array}
\right],$$ where $$\label{eq:gradF}
\left[
\begin{array}{c|c|c|c}
\vdots & \vdots & \vdots & \vdots \\
\nabla_{\bff} F & \nabla_{\bflambda_{\textnormal{e}}} F & \nabla_{\bfs} F & \nabla_{\bflambda_{\textnormal{i}}} F\\
\vdots & \vdots & \vdots & \vdots
\end{array}
\right]
=
\left[
\begin{array}{cccc}
\bfQ(\bfp) & -\bfC_{\textnormal{e}}\t & \bfzero & -\bfC_{\textnormal{i}}\t \\
\bfC_{\textnormal{e}}& \bfzero & \bfzero & \bfzero \\
\bfC_{\textnormal{i}}& \bfzero & -\bfI_{m_i} & \bfzero \\
\bfzero & \bfzero & \bfLambda_{\textnormal{i}}& \bfS
\end{array}
\right].$$ Here, the dual residual is $$\bfr_d(\bfp) = \bfQ(\bfp) \bff^k + \bfb(\bfp) - \bfC_{\textnormal{e}}^\top \bflambda_{\textnormal{e}}^k - \bfC_{\textnormal{i}}^\top \bflambda_{\textnormal{i}}^k$$ and the primal residuals for the equality and inequality constraints are, respectively, $$\bfr_{\textnormal{e}}= \bfC_{\textnormal{e}}\bff^k - \bfc_{\textnormal{e}}\quad \text{ and }\quad \bfr_{\textnormal{i}}= \bfC_{\textnormal{i}}\bff^k - \bfc_{\textnormal{i}}- \bfs^k.$$
The crucial components of the primal dual interior point method include an efficient linear solver for , the step length selection (here we use the largest step size in $[0,1)$ that ensures that both $\bflambda_{\textnormal{i}}$ and $\bfs$ remain sufficiently far from the boundary), and the choice of the centrality parameter $\delta$. For the latter, we use Mehrotra’s predictor-corrector approach as described in [@NocedalWright2006]. First, in the predictor step, we compute an *affine scaling step* (i.e., we solve for $\delta = 0$) and perform a line search. Then, in the corrector step, we compute the final direction by solving for $$\delta = \left( \frac{(\bff + \alpha^{\rm aff} \Delta \bff^{\rm aff})^\top (\bfs + \alpha^{\rm aff} \Delta \bfs^{\rm aff})}{m_e \mu }\right)^3.$$ Thus, each iteration requires two linear solves.
#### Computing sensitivities
In order to enable fast optimization of the design parameters (i.e., optimization for the outer problem), we need to differentiate the solutions of the quadratic program with respect to the design parameters. To do this, we use implicit differentiation of the optimality condition . Let $\widehat\bff(\bfp)$ be the computed solution to the quadratic programming problem. Then, we are interested in computing the sensitivity matrix $\bfJ_{\widehat\bff} (\bfp) \in \R^{n \times \ell}$ such that $$\widehat \bff(\bfp+ \Delta \bfp) = \widehat \bff(\bfp) + \bfJ_{\widehat \bff}(\bfp) \Delta \bfp + \mathcal{O}(\| \Delta \bfp\|^2)$$ for all $\Delta \bfp \in \R^\ell$.
To this end, we differentiate both sides of around the current KKT point $(\widehat\bff(\bfp),\bflambda_{\textnormal{e}}(\bfp),\bfs(\bfp), \bflambda_{\textnormal{i}}(\bfp))$ with respect to $\bfp$ and obtain $$\begin{aligned}
\bf0 & = \nabla_\bfp F(\widehat\bff(\bfp), \bflambda_{\textnormal{e}}(\bfp),\bfs(\bfp),\bflambda_{\textnormal{i}}(\bfp),\delta,\mu)\\
& = \nabla_{\bfp} F(\widehat\bff(\bfp),\bflambda_{\textnormal{e}}, \bfs, \bflambda_{\textnormal{i}},\delta, \mu) + \left[
\begin{array}{c|c|c|c}
\vdots & \vdots & \vdots & \vdots \\
\nabla_{\widehat \bff} F & \nabla_{\bflambda_{\textnormal{e}}} F & \nabla_{\bfs} F & \nabla_{\bflambda_{\textnormal{i}}} F\\
\vdots & \vdots & \vdots & \vdots
\end{array}
\right] \left[
\begin{array}{c}
\bfJ_{\widehat\bff}(\bfp) \\ \bfJ_{\bflambda_{\textnormal{e}}}(\bfp) \\ \bfJ_{\bfs}(\bfp) \\ \bfJ_{\bflambda_{\textnormal{i}}}(\bfp)
\end{array}
\right].\end{aligned}$$ Assuming that the linear system in is invertible (that is, there is a unique KKT point) we obtain $$\begin{bmatrix}
\bfJ_{\widehat\bff}(\bfp) \\ \bfJ_{\bflambda_{\textnormal{e}}}(\bfp) \\ \bfJ_{\bfs}(\bfp) \\ \bfJ_{\bflambda_{\textnormal{i}}}(\bfp)
\end{bmatrix}
= -
\begin{bmatrix}
\bfQ(\bfp) & -\bfC_{\textnormal{e}}\t & \bf0 & -\bfC_{\textnormal{i}}\t \\
\bfC_{\textnormal{e}}& \bf0 & \bf0 & \bf0 \\
\bfC_{\textnormal{i}}& \bf0 & -\bfI_{m_i} & \bf0 \\
\bf0 & \bf0 & \bfLambda_{\textnormal{i}}& \bfS
\end{bmatrix}^{-1}
\begin{bmatrix} \nabla_{\bfp}(\bfQ(\bfp) \widehat \bff(\bfp) + \bfb(\bfp)) \\ \bf0 \\ \bf0 \\ \bf0
\end{bmatrix}.$$ In the case of OED Problem A with $\bfE(\bfp) = \bfI_\ell$ we have $$\nabla_{\bfp}(\bfQ(\bfp) \bff + \bfb(\bfp)) = 2 \begin{bmatrix} \bfR(\theta_1)\t \bfT\t & & \\
& \ddots & \\ & & \bfR(\theta_\ell)\t \bfT\t
\end{bmatrix} (\bfM(\bfp)\bff - \bfd(\bfp)).$$
Computing the complete sensitivity matrix of interest $\bfJ_{\widehat\bff}$ requires $\ell$ linear solves and thus might be infeasible when $\ell \gg 1$. Thus, as is also common in PDE parameter estimation [@Haber2015], we provide matrix-free implementations that compute matrix vector products $\bfv \mapsto \bfJ_{\widehat\bff}\bfv $ and $\bfw \mapsto \bfJ_{\widehat \bff}^\top \bfw $ at the cost of one linear solve per training sample.
For some applications and in particular for OED Problem B, it may be beneficial to re-parameterize the angles contained in $\bfp$ by taking a reference angle $\delta_ 1 = p_1$ and non-negative increments $\delta_j\geq 0 $, $j = 2,\dots, n$ such that $p_k = \sum_{j = 1}^k \delta_j$. This results in the following additional constraints $$\delta_1 -a \geq 0, \quad
\delta_j \geq 0 \qquad \text{for } j = 2,\dots, n, \quad \mbox{ and } \quad
b - \sum_{j = 1}^n \delta_j \geq 0,$$ where $a$ is the lower and $b$ the upper bound for the angles and derivatives are given by $$\frac{\partial \bfp}{\partial \bfdelta} =
\begin{bmatrix}
1 & 0 & 0 & \ldots & 0 \\
1 & 1 & 0 & \ldots & 0 \\
\vdots & & \ddots & & \vdots \\
1 & 1 & 1 & \ldots & 1 \\
\end{bmatrix}\,.$$
Numerical Experiments {#sec:numerical_experiments}
=====================
In this section, we provide various examples for both OED problems A and B, using the tomography reconstruction problems described in Section \[sec:problem\_setup\]. We consider various assumptions (including different training data sets) and investigate how different constraints on the imaging state problem may affect the optimal design parameters.
Implementation {#sub:implementation}
--------------
Our numerical framework for minimizing the empirical Bayes risk is implemented as an add-on to the parameter estimation package <span style="font-variant:small-caps;">jInv</span> [@RuthottoTreisterHaber2017]. To benefit from <span style="font-variant:small-caps;">jInv</span>’s existing capabilities, the routines for solving the lower-level problem and computing sensitivities are implemented as a module, extending the abstract forward problem type. In particular, this allows parallel and distributed-memory evaluation of the constrained inverse problems for different training data. The design problem is formulated and solved using the misfit, regularization, and optimization methods provided in <span style="font-variant:small-caps;">jInv</span>’s. Our module will be made freely available.
Data sets and Constraints {#sub:trainingdata}
-------------------------
We consider four sets of training data shown in Figure \[fig:testData\] and described below. To obtain each projection, we use model with parallel rays where the number of rays is equal to the pixel size.
Rectangles:
: The data set consists of $20$ binary images of size $40\times 40$ that show randomly spaced and sized rectangles. The rectangles’ edges are aligned with the coordinate axes. Note that given the knowledge on the binary constraints, exact reconstruction is possible using two measurements with angles of $0$ and $90$ degrees, respectively.
Pentagons:
: A second data set consists of binary images of pentagons. As before, the training data consists of $20$ discrete images with $40\times 40$ pixels. The location and the size of the pentagons change randomly from image to image; however, the angles between the edges and the coordinate axes are the same. Thus, the object can be reconstructed exactly with appropriate constraints and from five projection angles.
Shapes:
: As a non-binary example we consider a synthetically generated dataset containing $20$ discrete images of size $40\times 40$ that are obtained by evaluating a random smooth function on a randomly chosen supporting set.
Phantom:
: As a more realistic example, we generate a training data set consisting of $20$ gray valued images resembling random variations of the Modified Shepp-Logan phantom [@jain1989fundamentals; @shepp1974fourier]. The Shepp-Logan phantom resembles basic head characteristics: exterior, skull, left and right ventricles, as well as tumors. We randomly varied head features, such as skull size, head & ventricle size and orientation, intensity, and number of tumors. Our Matlab implementation will be publicly available. Here, we discretize the data on a $64\times64$ regular grid.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Simulated training data used in the numerical experiments. Each data set consists of 20 images with intensity values between $0$ (black) and $1$ (white). The blue lines divide the plot into the individual examples.[]{data-label="fig:testData"}](OED_A_rect-xtrue "fig:"){width="\iwidth"} ![Simulated training data used in the numerical experiments. Each data set consists of 20 images with intensity values between $0$ (black) and $1$ (white). The blue lines divide the plot into the individual examples.[]{data-label="fig:testData"}](OED_B_pent "fig:"){width="\iwidth"}
1\) rectangles 2\) pentagons
\[2ex\] ![Simulated training data used in the numerical experiments. Each data set consists of 20 images with intensity values between $0$ (black) and $1$ (white). The blue lines divide the plot into the individual examples.[]{data-label="fig:testData"}](OED_A-xtrue "fig:"){width="\iwidth"} ![Simulated training data used in the numerical experiments. Each data set consists of 20 images with intensity values between $0$ (black) and $1$ (white). The blue lines divide the plot into the individual examples.[]{data-label="fig:testData"}](OED_shepp "fig:"){width="\iwidth"}
3\) shapes 4\) phantom
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
These data sets can be used directly in empirical Bayes risk minimization OED problems for approximating the expected value, as discussed in Section \[sec:problemformulation\]. Furthermore, we can treat the images as samples from some underlying distributions and either use a prior assumption on the covariance $\bfGamma_\bff$ or a sample approximation to solve the Bayes OED problems.
The main interest of this work is to investigate how different constraints on the imaging problem may affect the optimal design parameters. Therefore, we consider four common choices of constraints in .
unconstrained:
: No constraints are imposed on the discrete image, $\bff$. In this case the optimality condition in is a linear system with a positive definite matrix and (for small-scale problems) is solved using a Cholesky factorization. For this case, we compare the optimal Bayes design with the optimal empirical Bayes design.
equality constraints:
: We assume that the sum of the intensity values equals a known constant. To this end, we set $\bfC_{\textnormal{e}}= \begin{bmatrix} 1, & \ldots ,& 1\end{bmatrix} \in \R^{1 \times n}$ and $\bfc_{\textnormal{e}}= \bfC_{\textnormal{e}}\bff_{\rm true}$. Similar to the previous case the optimality condition is linear and requires solving a saddle point problem. To this end, we use an LU factorization.
non-negativity:
: A physically meaningful constraint in many imaging problems is to enforce reconstructed intensity values to be non-negative. In our formulation we use $\bfC_{\textnormal{i}}= \bfI_n \in \R^{n\times n}$ and $\bfc_{\textnormal{i}}= \bfzero$.
bound constraints:
: In addition to non-negativity, we enforce an upper bound and restrict the intensity values of each pixel to be between $0$ and $1$ by using $\bfC_{\textnormal{i}}= [\bfI_n; -\bfI_n] \in \bbR^{2n\times n}$ and $\bfc_{\textnormal{i}}= [\bfzero; -\bfe] \in \bbR^{2n}$.
Since the most plausible choice of a constraint will depend on the particular application our framework supports a variety of constraints. Due to the relative simplicity of OED problem B (e.g., no additional design regularization and potentially much fewer parameters to optimize), we begin with some investigations on the impact of lower level constraints on the overall optimal experimental design.
OED Results for Problem B {#sub:comparison_of_objective_functions}
-------------------------
For simplicity and for visualization purposes, we consider OED problem B with $\ell=2$; that is, we aim to find the $2$ projection angles, where the resulting reconstructions minimize the mean-squared error. Note that computing reconstructions in this case is highly under-determined.
We first investigate the OED objective function $\calJ(\bfp)$ for projection angles $\bfp \in [0,180]^2$ in intervals of $1$ degree, assuming $p_1 \geq p_2$. Note that no regularization is included in the outer optimization problem of OED problem B, and thus the objective function measures the mean squared error of the reconstruction. We begin with no constraints on the inner problem and provide the Bayes risk values in Figure \[fig:Bayes\] for various covariance matrices where the underlying images are $64 \times 64$ pixels. First, we provide the Bayes risk for $\bfL = \bfI_{4096}$, where the minimum occurs around angles $45$ and $135$ degrees. Then for both the rectangles and phantoms data sets, we generated 1,000,000 sample images and computed the sample covariance matrices (with a small regularization to ensure positive definiteness). Using the computed factor $\bfL$, we computed Bayes risks and provide them along with their corresponding minima in Figure \[fig:Bayes\]. In general, we see that $p_1=p_2$ as well as $p_1=0, p_2=180$ correspond to higher Bayes risk, which seems intuitive for tomography. However, it is evident that the choice of $\bfL$ does affect the optimal design.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\bfL = \bfI$ rectangles phantom
![Bayes risk for OED Problem B with two projection angles $p_1$ and $p_2$ for $\bfL$ being the identity matrix and for $\bfL$ coming from covariance matrices obtained from realizations of the rectangles and the phantoms datasets. Red dots correspond to minimum values. These results correspond to the unconstrained state problem with $\alpha =1$ and $\sigma = 1$. []{data-label="fig:Bayes"}](eye.png "fig:"){width=".33\textwidth"} ![Bayes risk for OED Problem B with two projection angles $p_1$ and $p_2$ for $\bfL$ being the identity matrix and for $\bfL$ coming from covariance matrices obtained from realizations of the rectangles and the phantoms datasets. Red dots correspond to minimum values. These results correspond to the unconstrained state problem with $\alpha =1$ and $\sigma = 1$. []{data-label="fig:Bayes"}](rect.png "fig:"){width=".33\textwidth"} ![Bayes risk for OED Problem B with two projection angles $p_1$ and $p_2$ for $\bfL$ being the identity matrix and for $\bfL$ coming from covariance matrices obtained from realizations of the rectangles and the phantoms datasets. Red dots correspond to minimum values. These results correspond to the unconstrained state problem with $\alpha =1$ and $\sigma = 1$. []{data-label="fig:Bayes"}](phantom.png "fig:"){width=".33\textwidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Next we provide the *empirical* Bayes risk (i.e., values of the sampled objective function) for all four training data sets and four options for constraints in Figure \[fig:OEDB\_plotObjective\]. The two best designs are marked with red dots. For all training data, the design improvement is most pronounced for the box-constrained lower-level problem. Further, for the pentagon example, the global optima are obtained at different points for unconstrained/equality constrained and the inequality constrained problems.
-- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
unconstrained equality non-negativity box
[![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_rect_unconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_rect_eqconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_rect_nonneg "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_rect_box "fig:"){width="31mm"}]{}
[![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_pent_unconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_pent_eqconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_pent_nonneg "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_pent_box "fig:"){width="31mm"}]{}
[![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_blobs_unconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_blobs_eqconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_blobs_nonneg "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_blobs_box "fig:"){width="31mm"}]{}
[![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_shepp_unconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_shepp_eqconstrained "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_shepp_nonneg "fig:"){width="31mm"}]{} [![Mean squared reconstruction error for OED Problem B with two projection angles plotted along the $x$ and $y$ axis. For each data set we show the reduction of MSE for the unconstrained, equality constrained, non-negativity constrained and box constrained OED problem. The two best designs are indicated by red dots. To allow for comparison, the same color axis is in the sub plots of each row. It can be seen that a considerably larger reduction of the reconstruction error is achieved for the constrained problems. []{data-label="fig:OEDB_plotObjective"}](OED_shepp_box "fig:"){width="31mm"}]{}
-- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
We investigate the impact of the value of the regularization parameter, $\alpha$, on the reconstruction error for the optimal designs identified in the previous steps. To this end, we compute the MSE for 20 logarithmically equal spaced values of $\alpha$ between $10^{-4}$ and $10^3$ using the optimal projection angles determined in the previous step. We see that for the unconstrained and equality constrained problem, the MSE is less sensitive to the choice of $\alpha$ and thus the global minima are difficult to identify. For the inequality constrained problems substantial improvement of the design can be obtained. For all datasets, we found that thee box constraints resulted in smaller MSE values overall. Since a good choice of $\alpha$ may not be available a priori, one could consider an approach that incorporates $\alpha$ as a design parameter, so that optimal design also includes optimizing for $\alpha$.
![Investigation on the impact of regularization parameter on MSE for different choices of constraints and different data sets. []{data-label="fig:plotObjectiveBLambda"}](plotObjectiveB){width="\textwidth"}
OED Results for Problem A {#sub:solution_approach_for_oed_problem_a}
-------------------------
Given the training data and a fine discretization of the tomography operator, we generate data, by evaluating the forward problem and adding 0.1% Gaussian white noise. Then, the OED problem is solved twice. First, we aim at eliminating the number of rows in $\bfA$ by solving the OED problem A with an $\ell_1$-regularizer on the measurement parameters. Second, we adjust the weights of the non-zero weights by re-solving the OED problem without regularization. This procedure resembles the method introduced in [@HaberEtAl2008OED].
We first investigate the impact of the sparsity parameter $\beta$ on the design. In the top row of Figure \[fig:OEDA\_beta\], we provide the number of projections $\ell$ for various values of $\beta$, and as expected, we see that with a larger sparsity parameter, we obtain fewer projections. The more interesting result is in the second row where we provide the MSE as a function of $\beta.$ Here we see that even with fewer projections, reconstructions obtained by imposing constraints correspond to smaller MSE values.
![Results for OED problem A using the four test data sets for the tomography problem (column wise). First row depicts the number of projection for the optimal design in dependence on the sparsity parameter $\beta$. The second row shows the optimal mean squared error (MSE).[]{data-label="fig:OEDA_beta"}](plotHIsOEDA){width="\textwidth"}
In Figures \[fig:OEDA\_binary\] and \[fig:OEDA\_blobs\], we provide four sample reconstructions and error images for each dataset and constraint. We observe that, overall, fewer projections are required and smaller reconstruction errors are possible if box or non-negativity constraints are included on the lower problem. This distinction is most prominent with the datasets of binary images, where only a few projections are needed.
Intuitively, the optimal angles for the rectangle images should be 0 and 90 degrees and the optimal angles for the pentagon images should be 27, 63, 99, 135, and 171 degrees. This is because the training images all share the same orientation, and angles orthogonal to the edges may be considered optimal (see Figure \[fig:testData\]). In Figure \[fig:OEDA\_binary\], we see that for the rectangles, the non-negative and box-constrained OED parameters are $\hat\bfp = [0,90]\t$ and $\hat\bfp = [0, 90]\t$ degrees respectively, and for the pentagons, $\hat\bfp = [25, 64, 99, 136, 171]\t$ and $\hat\bfp = [27, 62, 99, 134, 171]\t$ degrees respectively. These results illustrate the benefits of incorporating proper constraints to improve the optimal experimental design.
![OED results for binary datasets (rectangles and pentagons respectively) comparing unconstrained, equality constrained, non-negativity constrained, and box constrained formulations of OED problem A. Reconstructions and differences to the true training image for 4 randomly chosen images are provided, where the color axis is identical in each row to simplify comparison. The red dots on the half circle correspond to the projection angles of the optimal designs. []{data-label="fig:OEDA_binary"}](OEDA_RectImages "fig:"){width=".9\textwidth"}\
![OED results for binary datasets (rectangles and pentagons respectively) comparing unconstrained, equality constrained, non-negativity constrained, and box constrained formulations of OED problem A. Reconstructions and differences to the true training image for 4 randomly chosen images are provided, where the color axis is identical in each row to simplify comparison. The red dots on the half circle correspond to the projection angles of the optimal designs. []{data-label="fig:OEDA_binary"}](OEDA_PentImages "fig:"){width=".9\textwidth"}
![OED results for blobs and brain phantoms datasets comparing unconstrained, equality constrained, non-negativity constrained, and box constrained formulations of OED problem A. Reconstructions and differences to the true training image for 4 randomly chosen images are provided, where the color axis is identical in each row to simplify comparison. The red dots on the half circle correspond to the projection angles of the optimal designs. []{data-label="fig:OEDA_blobs"}](OEDA_BlobsImages "fig:"){width=".9\textwidth"}\
![OED results for blobs and brain phantoms datasets comparing unconstrained, equality constrained, non-negativity constrained, and box constrained formulations of OED problem A. Reconstructions and differences to the true training image for 4 randomly chosen images are provided, where the color axis is identical in each row to simplify comparison. The red dots on the half circle correspond to the projection angles of the optimal designs. []{data-label="fig:OEDA_blobs"}](OEDA_SheppImages "fig:"){width=".9\textwidth"}
Conclusions {#sec:conclusions}
===========
In this work, we consider optimal experimental design problems in the context of tomographic reconstruction and investigate the impact of state constraints on the optimal design. We examine two problem formulations for enforcing sparse sampling in the design parameters, where OED problem A employs a sparsity enforcing regularization term and OED problem B achieves a desired level of sparsity by construction. We investigate Bayes risk and empirical Bayes risk minimization techniques for OED. For problems with known or well-approximated (e.g., from very large data sets) mean and covariance matrix, a reformulation of the Bayes risk can lead to efficient methods to obtain optimal designs for the unconstrained case. However, the empirical risk minimization framework allows for incorporation of state constraints. The primary challenge toward efficient optimization of the empirical problem is computing derivatives of the reconstructed image with respect to the design parameters. We obtain these by using implicit differentiation of the KKT conditions within interior point methods and by exploiting parallel computing in Julia. Our numerical results on various datasets demonstrate that including state constraints does indeed impact the optimal design, in that fewer projections are required, and smaller MSE values can be obtained. Some items for future work include considering integer or binary constraints on the design parameters, extensions to nonlinear inverse problems, and extensions to other applications.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was initiated as a part of the SAMSI Program on Optimization 2016-2017. The material was based upon work partially supported by the National Science Foundation under Grant DMS-1127914 to the Statistical and Applied Mathematical Sciences Institute. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.
[^1]: Department of Mathematics and Computer Science, Emory University, Atlanta, GA (, <http://www.mathcs.emory.edu/~lruthot/>).
[^2]: Department of Mathematics, Computational Modeling and Data Analytics Division, Academy of Integrated Science, Virginia Tech, Blacksburg, VA (, <http://www.math.vt.edu/people/jmchung/>).
[^3]: Department of Mathematics, Computational Modeling and Data Analytics Division, Academy of Integrated Science, Virginia Tech, Blacksburg, VA (, <http://www.math.vt.edu/people/mcchung/>).
[^4]: Submitted to the editors .
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the minor process of (Hermitian) matrix diffusions with constant diagonal drifts. At any given time, this process is determinantal and we provide an explicit expression for its correlation kernel. This is a measure on the Gelfand-Tsetlin pattern that also appears in a generalization of Warren’s process [@War07], in which Brownian motions have level-dependent drifts. Finally, we show that this process arises in a diffusion scaling limit from an interacting particle system in the anisotropic KPZ class in $2+1$ dimensions introduced in [@BF08]. Our results generalize the known results for the zero drift situation.'
author:
- 'Patrik L. Ferrari[^1]'
- 'René Frings[^2]'
title: |
Perturbed GUE Minor Process\
and Warren’s Process with Drifts
---
Introduction
============
In this paper we determine a determinantal point process living on the Gelfand-Tsetlin cone ${\mathrm{GT}}_N$, $${\mathrm{GT}}_N = \{(x^1,x^2,\dotsc,x^N) \in {\mathbb{R}}^1 \times {\mathbb{R}}^2 \times \dotsb \times {\mathbb{R}}^N : x_k^{n+1} \leq x_k^n \leq x_{k+1}^{n+1} \}$$ arising both from random matrices diffusions and interacting particle systems. An element $x \in {\mathrm{GT}}_N$ is called a Gelfand-Tsetlin pattern. Here is a graphical representation of $x \in {\mathrm{GT}}_4$, which illustrates the interlacing condition on $x$, $$\arraycolsep-.1ex
\begin{array}{ccccccccccccc}
x_1^4 & & & & x_2^4 & & & & x_3^4 & & & & x_4^4 \\[-.5ex]
& \rotatebox[origin=c]{-45}{$\le$} & & \rotatebox[origin=c]{45}{$\le$} & & \rotatebox[origin=c]{-45}{$\le$} & & \rotatebox[origin=c]{45}{$\le$} & & \rotatebox[origin=c]{-45}{$\le$} & & \rotatebox[origin=c]{45}{$\le$}\\[-.5ex]
& & x_1^3 & & & & x_2^3 & & & & x_3^3 \\[-.5ex]
& & & \rotatebox[origin=c]{-45}{$\le$} & & \rotatebox[origin=c]{45}{$\le$} & & \rotatebox[origin=c]{-45}{$\le$} & & \rotatebox[origin=c]{45}{$\le$} \\[-.5ex]
& & & & x_1^2 & & & & x_2^2 \\[-.5ex]
& & & & & \rotatebox[origin=c]{-45}{$\le$} & & \rotatebox[origin=c]{45}{$\le$} \\[-.5ex]
& & & & & & x_1^1
\end{array}$$ Measures on Gelfand-Tsetlin patterns naturally appear in several fields of mathematics like (a) random matrix theory [@Bar01; @JN06; @FN08; @OR06] where the question of universality was recently approached in [@Met11], (b) random tiling problems [@BG08; @OR06; @NY11], (c) representation theory [@BK07; @BK09], and (d) interacting particle systems [@BF08; @Nor08; @NY11] and diffusions [@War07; @WW08]. Probably the most famous example which belongs to more than one of these classes is the Aztec diamond. Indeed, the measure on the Aztec diamond both comes from a random tiling problem [@EKLP92] and can be obtained through a Markov chain [@Nor08] on a (discrete) Gelfand-Tsetlin pattern, which itself can be seen as a special case of the more general Markov chain construction in [@BF08]. A continuous space analogue is for instance Warren’s process, a system of interacting standard Brownian diffusions [@War07; @WW08]. Recently, the link between the “drift-less” case of [@BF08] and Warren’s process has been studied in [@GS12].
In this paper we consider GUE matrix diffusions with drifts. Its eigenvalue process at a fixed time is a determinantal point process and we explicitly determine its correlation kernel. We then show that the point process arises from a diffusion scaling limit of an interacting particle system in the anisotropic KPZ class in $2+1$ dimensions [@BF07; @BF08], as well as in a generalization of Warren’s process [@War07] if we let the Brownian motions to have level-dependents drifts. The analogue results for the zero-drift case were all previously known, see [@JN06; @GS12; @FF10].
Remark that the GUE minor process and the Warren process are not the same if considered as stochastic processes. Indeed, this is true already for the zero-drift case. Without drift, the GUE minor process was described in [@JN06], while Warren’s process was introduced in [@War07]. It is known that the two processes coincide when projected on “space-like” paths [@FF10], in which case they both are Markovian and determinantal. However, in the whole “space-time” the two processes are different [@ANvM10b]. Here we focus on the fixed-time process although the connection will certainly hold along “space-like” paths as for the zero-drift case, see [@FF10].
### Matrix diffusions {#matrix-diffusions .unnumbered}
The first model we study on ${\mathrm{GT}}_N$ is a variant of the GUE minor process which has been introduced in [@JN06]. Consider an $N \times N$ Hermitian matrix $H$ with eigenvalues $\lambda_1^N\leq \dotsb \leq \lambda_N^N$. Denote by $H^n$ the submatrix obtained by keeping the first $n$ rows and columns of $H$, and its ordered eigenvalues by $\lambda^n_1\leq \dotsb \leq \lambda_n^n$. The collection of all these eigenvalues $(\lambda^1,\dotsc,\lambda^N)$ then forms a Gelfand-Tsetlin pattern, with $\lambda^n=(\lambda_1^n,\dotsc,\lambda_n^n)$. In this paper we take $H(t)$ to be a GUE matrix diffusion perturbed by a deterministic drift matrix $M={\operatorname{diag}}(\mu_1,\dotsc,\mu_N)$, i.e., we consider $G(t)=H(t)+t M$ with $H$ evolving as standard Brownian Motion starting from $0$. The eigenvalues’ point process $\xi$ has support on ${\mathbb{R}}\times \{1,\dotsc,N\}$, $$\xi({\mathrm{d}}x,m)=\sum_{1\leq k \leq n \leq N}\delta_{n,m} \delta_{\lambda_k^n}({\mathrm{d}}x)$$ and its correlation function is given as follows (see Section \[sec:2\] for the proof).
\[thm:1\] For a fixed time $t>0$ consider the eigenvalues’ point process on the $N$ submatrices of $H(t)$. Then, its $m$-point correlation function $\varrho^m_t$ is given by $$\label{eq:24}
\varrho^m_t((x_1,n_1),\dots,(x_m,n_m))=\det[K_t((x_i,n_i),(x_j,n_j))]_{1\leq i,j\leq m},$$ with $(x_i,n_j) \in {\mathbb{R}}\times \{1,\dots,N\}$ and correlation kernel $$\label{eq:26}
K_t((x,n),(x',n'))=-\phi^{(n,n')}(x,x')+\sum_{k=1}^{n'} \Psi_{n-k}^{n,t}(x) \Phi_{n'-k}^{n',t}(x'),$$ where[^3] $$\begin{aligned}
\phi^{(n,n')}(x,x') &= \frac{(-1)^{n'-n}}{2\pi{\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-}{\mathrm{d}}z\, \frac{{\mathrm{e}}^{z(x'-x)}}{(z-\mu_{n+1})\dotsb(z-\mu_{n'})}\,{\mathbbm{1}}_{[n<n']}, \label{eq:9}\\
\Psi_{n-k}^{n,t}(x) & = \frac{(-1)^{n-k}}{2\pi {\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} {\mathrm{d}}z \, {\mathrm{e}}^{tz^2/2-xz} \, \frac{(z-\mu_1)\dotsb (z-\mu_n)}{(z-\mu_1)\dotsb(z-\mu_k)}, \label{eq:9a} \\
\Phi_{n-\ell}^{n,t}(x) &= \frac{(-1)^{n-\ell}}{2\pi {\mathrm{i}}} \oint_{\Gamma_{\mu_1,\dotsc,\mu_N}} {\mathrm{d}}w\, {\mathrm{e}}^{-tw^2/2+x w} \, \frac{(w-\mu_1)\dotsb(w-\mu_{\ell-1})}{(w-\mu_1)\dotsb(w-\mu_n)} \label{eq:9b}\end{aligned}$$ with $\mu_-<\min\{\mu_1,\dots,\mu_N\}$.
The integral for $\phi^{(n,n')}$ in is only well-defined for $n'-n>1$. For $n'-n=1$ we set $\phi^{(n-1,n)}(x,x'):=\phi_n(x,x')={\mathrm{e}}^{\mu_n(x'-x)}{\mathbbm{1}}_{[x>x']}$ instead.
In an independent work [@AvMW13] on minors of random matrices by Adler, van Moerbeke, and Wang appeared on the arXiv after this work, the same kernel is computed and a double integral expression is also provided.
### Warren’s process with drifts {#warrens-process-with-drifts .unnumbered}
Our second model is Warren’s process with drifts that describes the dynamics of a system of Brownian motions $\{B_k^n,1\leq k \leq n \leq N\}$ on ${\mathrm{GT}}_N$, where $B_1^1$ is a standard Brownian motion with drift $\mu_1$ starting from the origin. The Brownian motions $B_1^2$ and $B_2^2$ are Brownian motions with drifts $\mu_2$ conditioned to start at the origin and, whenever they touch $B_1^1$, they are reflected off $B_1^1$. Similarly for $n\geq 2$, $B_k^n$ is a Brownian motion with drift $\mu_n$ conditioned to start at the origin and being reflected off $B_k^{n-1}$ (for $k\leq n-1$) and $B_{k-1}^{n-1}$ (for $k\geq 2$). The process with $\mu_1=\dotsb=\mu_N=0$ was introduced and studied by Warren in [@War07].
The correlation functions of this process at a fixed time agree with those of the perturbed GUE minor process (the proof is in Section \[sec:3\]).
\[thm:2\] For a fixed time $t>0$ consider the point process of the positions of the Brownian motions $\{B_k^n(t):1\leq k\leq n\leq N\}$ described above. Then, its $m$-point correlation function $\varrho^m_t$ is also given by .
### Interacting particle system {#interacting-particle-system .unnumbered}
Finally we introduce a discrete model giving rise to Warren’s process with drifts under a diffusion scaling limit. This model is a generalization of TASEP with particle-dependent jump rates [@BF07] to the $2+1$ dimensional particle system with Markov dynamics introduced in [@BF08]. We denote by $x_k^n \in {\mathbb{Z}}$ the position of a particle labeled by $(k,n)$, with $1\leq k \leq n \leq N$, and call $n$ the “level” of the particle. Particle $(k,n)$ performs a continuous time random walk with one-sided jumps (to the right) and with rate $v_n$. Particles with smaller level evolve independently from the ones with higher level like in the Brownian motion model described above. More precisely, the interaction between levels is the following: (a) if particle $(k,n)$ tries to jump to $x$ and $x_{k-1}^{n-1}=x$, then the jump is suppressed, and (b) when particle $(k,n)$ jumps from $x-1$ to $x$, then all particles labeled by $(k+\ell,n+\ell)$ (for some $\ell\geq 1$) which were at $x-1$ are forced to jump to $x$, too. This is a particle system with state space in a discrete Gelfand-Tsetlin pattern.
Consider the diffusion scaling with appropriate scaled jump rates $$\label{scaling}
t=\tau T,\quad x_k^n=\tau T-\sqrt{T} \lambda_k^n, \quad v_n=1-\frac{\mu_n}{\sqrt{T}}.$$ Then, in the $T\to\infty$ limit, the particle process $\{x_k^n(t)\}$ converges to the GUE minor process with drift $\{\lambda_k^n(\tau)\}$.
More precisely, let us denote by $\widetilde {\mathbb{P}}^v$ the probability measure on these particles with jump rates $v=(v_1,\dotsc,v_N)$ given in . We fix $\tau>0$ and set $$\nu_T(A) = \widetilde {\mathbb{P}}^{v}\biggl(-\frac{x_k^n(\tau T)-\tau T}{\sqrt T} \in A_k^n \text{ for all } 1\leq k\leq n\leq N\biggr)$$ where $A_k^n \subseteq {\mathbb{R}}$ are Borel sets, $A=\prod_{1\leq k\leq n \leq N}A_k^n$. Moreover, we define $$\nu(A) = {\mathbb{P}}^\mu\bigl(\lambda_k^n(\tau)\in A_k^n \text{ for all } 1\leq k\leq n\leq N \bigr)$$ where ${\mathbb{P}}^\mu$ is the GUE minor measure with drift ${\operatorname{diag}}(\mu_1,\dotsc,\mu_N)$. In Section \[sec:4\] we show the following result.
\[thm:3\] As $T\to \infty$, $\nu_T$ converges to $\nu$ in total variation, i.e., $$\label{eq:31}
\lim_{T\to \infty} \sup_{\substack{A \subseteq {\mathbb{R}}^{N(N+1)/2},\\ A \, \mathrm{ Borel }}} \lvert \nu_T(A) - \nu(A)\rvert = 0.$$ In particular, $\nu_T \to \nu$ weakly.
### Acknowledgments {#acknowledgments .unnumbered}
The authors are grateful to M. Shkolnikov to point out how to use [@GS12] to show Theorem \[thm:2\] and to J. Warren for useful discussions. This work was supported by the German Research Foundation via the SFB 611–A12 and SFB 1060–B04 projects.
GUE minor process with drift {#sec:2}
============================
Model and measure
-----------------
Let $(H(t):t\geq 0)$ be a process on the $N\times N$ Hermitian matrices defined by $$H_{k\ell}(t) =
\begin{cases}
b_{kk}(t)+\mu_k t, & \text{if } 1\leq k \leq N, \\
\frac1{\sqrt{2}} (b_{k\ell}(t)+{\mathrm{i}}\tilde b_{k\ell}(t)), & \text{if } 1\leq k < \ell \leq N, \\
\frac1{\sqrt{2}} (b_{k\ell}(t)-{\mathrm{i}}\tilde b_{k\ell}(t)), & \text{if } 1\leq \ell < k \leq N,
\end{cases}$$ where $\{b_{kk}, b_{k\ell}, \tilde b_{k\ell}\}$ are independent one-dimensional standard Brownian motions[^4]. Denote by $M={\operatorname{diag}}(\mu_1,\dots,\mu_N)$ the diagonal drift matrix added to the matrix $H$. Then, the probability measure on these matrices at time $t$ is given by $$\label{eq:2}
{\mathbb{P}}(H\in{\mathrm{d}}H)={\mathrm{const}}\times \exp\left(-\frac{{\operatorname{Tr}}(H-t M)^2}{2t}\right){\mathrm{d}}H$$ where ${\mathrm{d}}H=\prod_{i=1}^N {\mathrm{d}}H_{ii}\prod_{1\leq j < k \leq N} {\mathrm{d}}{\operatorname{Re}}(H_{j,k}) {\mathrm{d}}{\operatorname{Im}}(H_{j,k})$ and ${\mathrm{const}}$ is the normalization constant.
Since we are interested in the statistics of the eigenvalues’ minors at time $t$, we first determine the measure on the eigenvalues of the $N\times N$ matrix.
\[lem:3\] Assume that $\mu_1,\dotsc,\mu_N$ are all distinct. Then under , the joint probability measure of the eigenvalues $\lambda_1,\dots,\lambda_N$ of $H$ is given by $$\begin{gathered}
\label{eq:8}
{\mathbb{P}}(\lambda_1 \in {\mathrm{d}}\lambda_1, \dots \lambda_N \in {\mathrm{d}}\lambda_N)\\ = {\mathrm{const}}\times \det\bigl[ {\mathrm{e}}^{-(\lambda_i-t \mu_j)^2/(2t)}\bigr]_{1\le i,j\le N} \, \frac{\Delta(\lambda_1,\dots,\lambda_N)}{\Delta(\mu_1,\dots,\mu_N)} \,{\mathrm{d}}\lambda_1\dotsb{\mathrm{d}}\lambda_N\end{gathered}$$ with ${\mathrm{const}}$ a normalization constant and $\Delta(x_1,\dots,x_m) = \prod_{1\leq i < j \leq m}(x_j-x_i)$ the Vandermonde determinant.
If $\mu_1,\dotsc,\mu_N$ are not all distinct, we have to take limits in . For instance, if $\mu_1=\dotsb=\mu_N\equiv \mu$, then $$\begin{gathered}
{\mathbb{P}}(\lambda_1 \in {\mathrm{d}}\lambda_1, \dots \lambda_N \in {\mathrm{d}}\lambda_N)\\ = {\mathrm{const}}\times \biggl(\prod_{k=1}^N {\mathrm{e}}^{-(\lambda_k-t \mu)^2/(2t)}\biggr) \Delta^2(\lambda_1,\dots,\lambda_N) \,{\mathrm{d}}\lambda_1\dotsb{\mathrm{d}}\lambda_N.\end{gathered}$$
We diagonalize $H=U\Lambda U^\ast$ with a unitary matrix $U$ and the diagonal matrix $\Lambda={\operatorname{diag}}(\lambda_1,\dotsc,\lambda_N)$. Then, $$\label{eq:6}
{\mathrm{e}}^{-{\operatorname{Tr}}(H-t M)^2/(2t)} {\mathrm{d}}H = {\mathrm{const}}\times {\mathrm{e}}^{-{\operatorname{Tr}}(U \Lambda U^\ast -tM)^2/(2t)}\Delta^2(\lambda)\, {\mathrm{d}}U\,{\mathrm{d}}\lambda,$$ where ${\mathrm{d}}U$ is the Haar measure on the unitary group $\mathcal U$. Moreover, since $${\operatorname{Tr}}(U\Lambda U^\ast - tM)^2 = {\operatorname{Tr}}\Lambda^2+t^2 {\operatorname{Tr}}M^2-2t{\operatorname{Tr}}(U\Lambda U^\ast M)$$ by integrating over $\mathcal U$ in and using the Harish-Chandra-Itzykson-Zuber formula, we obtain the desired expression.
Now we focus on the minor process. For $1\leq n\leq N$ let us denote by $H^n(t)$ the $n\times n$ principal submatrix of $H(t)$ which is obtained from $H(t)$ by keeping only the $n$ first rows and columns. In particular, $H^1(t)=H_{11}(t)$ and $H^N(t)=H(t)$. We denote by $\lambda_1^n(t)\leq \dots \leq \lambda_n^n(t)$ the ordered eigenvalues of $H^n(t)$. It is then a classical fact of linear algebra that at any time $t$, the process $(\lambda^1,\dots,\lambda^N)(t)$ lies in the Gelfand-Tsetlin cone of order $N$, $${\mathrm{GT}}_N = \{ (x^1,\dots,x^N) \in {\mathbb{R}}^1 \times \dotsb \times {\mathbb{R}}^N : x^n\preceq x^{n+1} \text{ for all } 1\leq k \leq N-1\},$$ where $x^n \preceq x^{n+1}$ means that $x^n$ and $x^{n+1}$ interlace, i.e., $$x_k^{n+1} \leq x^n_k \leq x^{n+1}_{k+1} \quad \text{ for all } 1 \leq k \leq n.$$
The induced measure on $\{\lambda_k^n:1\leq k \leq n \leq N\}$ is the following.
Fix $t>0$. Then, under the measure , the joint density of the eigenvalues of $\{H^n:1\leq n \leq N\}$ on ${\mathrm{GT}}_N$ is given by $$\label{eq:1}
{\mathrm{const}}\times \prod_{k=1}^N {\mathrm{e}}^{-t\mu_k^2/2}\prod_{k=1}^N {\mathrm{e}}^{-(\lambda_k^N)^2/(2t)} \Delta(\lambda^N)\prod_{\substack{1\leq n \leq N \\ 1 \leq k \leq n}} {\mathrm{e}}^{\mu_n \lambda_k^n} \prod_{\substack{2\leq n\leq N\\1\leq k \leq n-1}} {\mathrm{e}}^{-\mu_n\lambda_k^{n-1}},$$ where the normalization constant does not depend on $\mu_1,\dots,\mu_N$.
We first derive under the assumption that the $\mu_1,\dots,\mu_N$ are all distinct; the case where some of the $\mu_i$ are equal is then recovered by taking the limit. We prove the statement inductively and follow the presentation in [@FN08b]. For $N=1$, the density is clearly proportial to $\exp(-(\lambda_1^1-\mu_1t)^2/(2t))$. For $N\geq 2$, we consider an $N \times N$ matrix $H^N$ distributed according to which we write as $$H^N - t M = \begin{pmatrix} H^{N-1} & w \\ w^\ast & x \end{pmatrix} - t \begin{pmatrix} M^{N-1} & 0 \\ 0 & \mu_N \end{pmatrix},$$ where $M^{N-1}$ denotes the $(N-1)\times (N-1)$ principal submatrix of $M$, $w \in {\mathbb{C}}^{N-1}$ is a Gaussian vector and $x \in {\mathbb{R}}$ is a Gaussian variable. Then we diagonalize $H^{N-1}$, i.e., we choose a unitary matrix $U$ such that $H^{N-1}=U \Lambda U^\ast$ with $\Lambda = {\operatorname{diag}}(\lambda^{N-1}_1,\dotsc,\lambda^{N-1}_{N-1})$ the diagonal matrix for the eigenvalues. Since the Gaussian distribution is invariant under unitary rotations and $w$ is independent of $H^{N-1}$, we have $$\begin{pmatrix} U^\ast & 0 \\ 0 & 1 \end{pmatrix} (H^N-tM) \begin{pmatrix} U & 0 \\ 0 & 1 \end{pmatrix} \overset{d}{=} \begin{pmatrix} \Lambda & w \\ w^\ast & x \end{pmatrix}-t\begin{pmatrix} U^\ast M^{N-1}U & 0 \\ 0 & \mu_N \end{pmatrix},$$ where $\overset{d}{=}$ denotes equality in distribution. Applying the map $H^{N-1} \mapsto (\Lambda,U)$, we get that measure on $H^N$ is proportional to $$\begin{gathered}
\label{eq:3}
\exp\left(-\frac1{2t}{\operatorname{Tr}}\left[\begin{pmatrix} \Lambda & w \\ w^\ast & x \end{pmatrix}-t\begin{pmatrix} U^\ast M^{N-1}U & 0 \\ 0 & \mu_N \end{pmatrix}\right]^2\right) \Delta^2(\lambda^{N-1}) \\
\times {\mathrm{d}}U \, {\mathrm{d}}w \, {\mathrm{d}}x \, {\mathrm{d}}\lambda^{N-1},\end{gathered}$$ where ${\mathrm{d}}U$ is the Haar measure on the unitary group $\mathcal U_{N-1}$. We consider only the part of that depends on $U$ and integrate over $\mathcal U_{N-1}$, using the Harish-Chandra-Itzykson-Zuber formula, $$\int_{\mathcal U_{N-1}} {\mathrm{d}}U \, {\mathrm{e}}^{{\operatorname{Tr}}(\Lambda U^\ast M^{N-1} U)} = {\mathrm{const}}\times \frac{\det[{\mathrm{e}}^{\lambda_i^{N-1}\mu_j}]_{1\leq i,j\leq N-1}}{\Delta(\lambda^{N-1})\Delta(\mu_1,\dotsc,\mu_{N-1})}.$$ After this integration the measure reads $$\label{eq:4}
{\mathrm{const}}\times {\mathbb{P}}(\lambda^{N-1}\in{\mathrm{d}}\lambda^{N-1}) \, {\mathrm{e}}^{x \mu_N-t\mu_N^2/2} \prod_{k=1}^{N-1} {\mathrm{e}}^{-\lvert w_k \rvert^2/t} {\mathrm{d}}x\,{\mathrm{d}}w.$$ We focus on the measure on $w_k$ and represent the variables in polar coordinates, $w_k = r_k{\mathrm{e}}^{{\mathrm{i}}\varphi_k}$ with $r_k \in {\mathbb{R}}_+$ and $\varphi_k \in [0,2\pi)$. Since the Jacobian of this transformation is given by $r_1\dotsb r_{N-1}$, we get $$\prod_{k=1}^{N-1} {\mathrm{e}}^{-\lvert w_k\rvert^2/t}{\mathrm{d}}w_k = \prod_{k=1}^{N-1} r_k {\mathrm{e}}^{-r_k^2/t}{\mathrm{d}}r_k \,{\mathrm{d}}\varphi_k,$$ where ${\mathrm{d}}r_k$ and ${\mathrm{d}}\varphi_k$ are Lebesgue measures on ${\mathbb{R}}_+$ and $[0,2\pi)$. Then we can express $r_k$ and $x$ in terms of the eigenvalues of $H^{N-1}$ and $H^N$, see e.g. [@FN08b] for details, $$\begin{aligned}
r_k^2 &= - \frac{\prod_{j=1}^N (\lambda_k^{N-1}-\lambda_j^N)}{\prod_{j=1,j\neq k}^{N-1}(\lambda_k^{N-1}-\lambda_j^{N-1})}\, {\mathbbm{1}}_{[\lambda^{N-1} \preceq \lambda^N]},\\
x & = {\operatorname{Tr}}(H^N-H^{N-1})=\sum_{i=1}^N \lambda_k^N-\sum_{k=1}^{N-1} \lambda_k^{N-1}.
\end{aligned}$$ The Jacobian of the transformation $T:(r_1,\dotsc,r_{N-1},x) \mapsto \lambda^N$ is then given by $$r_1\dotsb r_{N-1}\lvert \det T' \rvert = \frac{\Delta(\lambda^N)}{\Delta(\lambda^{N-1})} \, {\mathbbm{1}}_{[\lambda^{N-1} \preceq \lambda^N]},$$ and hence, given $\lambda^{N-1}$, we have $$\begin{gathered}
\label{eq:5}
{\mathrm{e}}^{x \mu_N} \prod_{k=1}^{N-1} {\mathrm{e}}^{-\lvert w_k \rvert^2/t} {\mathrm{d}}x\,{\mathrm{d}}w
= \prod_{k=1}^N {\mathrm{e}}^{-(\lambda_k^N)^2/(2t)+\mu_N \lambda_k^N} \prod_{k=1}^{N-1} {\mathrm{e}}^{(\lambda_k^{N-1})^2/(2t)-\mu_N \lambda_k^{N-1}}\\
\times \frac{\Delta(\lambda^N)}{\Delta(\lambda^{N-1})} \, {\mathbbm{1}}_{[\lambda^{N-1}\preceq \lambda^N]} \, {\mathrm{d}}\lambda^N \, {\mathrm{d}}\varphi.\end{gathered}$$ Here we used that $2 (r_1^2+\dotsb+r_{N-1}^2) = {\operatorname{Tr}}(H^N)^2-{\operatorname{Tr}}(H^{N-1})^2$. Moreover, by the induction assumption for $N-1$ we have $$\begin{gathered}
\label{eq:30}
{\mathbb{P}}(\lambda^{N-1}\in{\mathrm{d}}\lambda^{N-1}) = {\mathrm{const}}\times \prod_{k=1}^{N-1}{\mathrm{e}}^{-t \mu_k^2/2} \prod_{k=1}^{N-1} {\mathrm{e}}^{-(\lambda^{N-1}_k)^2/(2t)}\Delta(\lambda^{N-1}) \\ \times \prod_{\substack{1\leq n\leq N-1 \\ 1\leq k\leq n}} {\mathrm{e}}^{\mu_n\lambda_k^n}\prod_{\substack{2\leq n\leq N-1\\1\leq k\leq n}}{\mathrm{e}}^{-\mu_n\lambda_k^{n-1}} \prod_{n=1}^{N-1}{\mathrm{d}}\lambda^n.\end{gathered}$$ Finally, inserting and into and integrating out $\varphi$ (which multiplies the measure by a finite constant) results in the claimed formula .
Correlation functions
---------------------
Now we determine the correlation functions of the point process on the eigenvalues $\{\lambda_k^n:1\leq k\leq n\leq N\}$, and for that purpose, we rewrite the density in as a product of determinants. We set $\phi_n(x,y)={\mathrm{e}}^{\mu_n(y-x)}{\mathbbm{1}}_{\{x>y\}}$ and introduce “virtual” variables $\lambda_n^{n-1}=\text{virt}$ with the property that $\phi_n(\text{virt},y)={\mathrm{e}}^{\mu_n y}$. Then in we have, up to a set of measure zero, $$\det[\phi_n(\lambda_i^{n-1},\lambda_j^n)]_{1\leq i,j\leq n} = \prod_{j=1}^n {\mathrm{e}}^{\mu_n \lambda_j^n} \prod_{j=1}^{n-1} {\mathrm{e}}^{-\mu_n \lambda_j^{n-1}} {\mathbbm{1}}_{[\lambda^n \preceq \lambda^{n+1}]}.$$ Moreover, for $k=1,\dots,N$ we set $$\Psi_{N-k}^{N,t}(x)=\frac{{\mathrm{e}}^{-x^2/(2t)}}{\sqrt{2\pi t}}\,t^{-(N-k)/2}\,p_{N-k}\left(\frac{\mu_{k+1}t-x}{\sqrt t},\dots,\frac{\mu_Nt-x}{\sqrt t}\right),$$ where $p_n$ are symmetric polynomials of degree $n$ in $n$ variables defined by $p_0\equiv 1$ and $$p_n(x_1,\dots,x_n)=\frac{(-1)^n}{{\mathrm{i}}\sqrt{2\pi}}\int_{{\mathrm{i}}{\mathbb{R}}}{\mathrm{d}}w\, {\mathrm{e}}^{w^2/2}(w-x_1)\dotsb(w-x_n) \quad \text{ for } n\geq 1.$$ Hence we have that $$\prod_{k=1}^N {\mathrm{e}}^{-(\lambda_k^N)^2/(2t)} \Delta(\lambda^N) = {\mathrm{const}}\times \det\bigl[\Psi^{N,t}_{N-k}(\lambda_\ell^N)\bigr]_{1\leq k,\ell\leq N},$$ which means that we can rewrite as $$\label{eq:10}
{\mathrm{const}}\times \prod_{n=1}^N \det \bigl[\phi_n (\lambda_i^{n-1},\lambda_j^n)\bigr]_{1\leq i,j\leq n} \prod_{k=1}^N {\mathrm{e}}^{-t\mu_k^2/2} \det\bigl[\Psi^{N,t}_{N-k}(\lambda_\ell^N)\bigr]_{1\leq k,\ell\leq N} .$$ Note that by a change of variable $w = (t z-x)/\sqrt t$ we have $$\label{eq:11}
\Psi_{N-k}^{N,t}(x)=\frac{(-1)^{N-k}}{2\pi{\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}} {\mathrm{d}}z\, {\mathrm{e}}^{tz^2/2-xz} (z-\mu_{k+1})\dotsb(z-\mu_N).$$ A measure of the form has determinantal correlation functions and the kernel can be computed with Lemma 3.4 of [@BFPS06], see the appendix.
Proof of Theorem \[thm:1\]
--------------------------
We prove the theorem first for $\mu_1<\dots<\mu_N$ and then use analytic continuation. Note that for $n=N$, the function $\Psi_{n-k}^{n,t}$ in is the same as $\Psi_{N-k}^{N,t}$ in .
\[lem:1\] The following identities hold.
- For all $n\in\{1,\dots,N\}$, $k\in{\mathbb{Z}}$ and $t>0$, we have $\phi_n \ast \Psi^{n,t}_{n-k} = \Psi^{n-1,t}_{n-1-k}$.
- For $n<n'$, we have $\phi_{n+1}\ast\dots\ast\phi_{n'} = \phi^{(n,n')}$ with $\phi^{(n,n')}$ given in .
Because of ${\operatorname{Re}}z < \mu_n$ we can exchange the two integrals, $$\begin{aligned}
(\phi_n &\ast \Psi^{n,t}_{n-k})(x) \\
& = \int_{-\infty}^x{\mathrm{d}}y\, {\mathrm{e}}^{\mu_n(y-x)} \frac{(-1)^{n-k}}{2\pi {\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} \hspace{-1em} {\mathrm{d}}z \, {\mathrm{e}}^{tz^2/2-yz} \, \frac{(z-\mu_1)\dots (z-\mu_n)}{(z-\mu_1)\dots(z-\mu_k)} \\
& = \frac{(-1)^{n-k}}{2\pi {\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} \hspace{-1em} {\mathrm{d}}z\, {\mathrm{e}}^{tz^2/2-\mu_nx} \, \frac{(z-\mu_1)\dots (z-\mu_n)}{(z-\mu_1)\dots(z-\mu_k)} \int_{-\infty}^x{\mathrm{d}}y\, {\mathrm{e}}^{y(\mu_n-z)} \\
& = \frac{(-1)^{n-1-k}}{2\pi {\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-}{\mathrm{d}}z\, {\mathrm{e}}^{tz^2/2-x z} \, \frac{(z-\mu_1)\dots (z-\mu_{n-1})}{(z-\mu_1)\dots (z-\mu_k)} \\
& = \Psi^{n-1,t}_{n-1-k}(x).
\end{aligned}$$ This proves the first statement. To show (ii), we first consider the case $n'-n=2$. A simple calculation gives $$\phi^{(n-2,n)}(x,x')=(\phi_{n-1} \ast \phi_n)(x,x')= -\left( \frac{{\mathrm{e}}^{\mu_n(x'-x)}}{\mu_n-\mu_{n-1}}+\frac{{\mathrm{e}}^{\mu_{n-1}(x'-x)}}{\mu_{n-1}-\mu_n}\right){\mathbbm{1}}_{[x>x']},$$ which has the following contour integral representation, $$\phi^{(n-2,n)}(x,x') = \frac1{2\pi{\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} {\mathrm{d}}z\,\frac{{\mathrm{e}}^{z(x'-x)}}{(z-\mu_{n-1})(z-\mu_n)}.$$ For $n'-n>2$, we get inductively that $$\begin{aligned}
(\phi_n &\ast \phi^{(n,n')})(x,x') \\
& = \frac{(-1)^{n'-n}}{2\pi{\mathrm{i}}} \int_{-\infty}^x{\mathrm{d}}y \, {\mathrm{e}}^{\mu_n(y-x)} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} \hspace{-1em} {\mathrm{d}}z\, \frac{{\mathrm{e}}^{z(x'-y)}}{(z-\mu_{n+1})\dotsb(z-\mu_{n'})} \\
& = \frac{(-1)^{n'-n}}{2\pi{\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} \hspace{-1em} {\mathrm{d}}z\, \frac{{\mathrm{e}}^{zx'-\mu_nx}}{(z-\mu_{n+1})\dotsb(z-\mu_{n'})} \int_{-\infty}^x {\mathrm{d}}y\, {\mathrm{e}}^{y(\mu_n-z)} \\
& = \frac{(-1)^{n'-n+1}}{2\pi{\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-}{\mathrm{d}}z\, \frac{{\mathrm{e}}^{z(x'-x)}}{(z-\mu_n)(z-\mu_{n+1})\dotsb(z-\mu_{n'})} \\
& = \phi^{(n-1,n')}(x,x'),
\end{aligned}$$ where, as before, we could exchange the integrals because of ${\operatorname{Re}}z< \mu_n$.
Next we consider the $n$-dimensional space $V_n$ spanned by the set of functions $$\{\phi_1 \ast \phi^{(1,n)}(x_1^0,\cdot),\dots,\phi_{n-1}\ast \phi^{(n-1,n)}(x_{n-1}^{n-2},\cdot),\phi_n(x_n^{n-1},\cdot)\}.$$ According to Lemma 3.4 of [@BFPS06] we need to find a basis $\{\Phi_{n-k}^{n,t}:1\leq k\leq n\}$ of $V_n$ that is biorthogonal to the set $\{\Psi_{n-k}^{n,t}:1\leq k\leq n\}$, i.e., $$\int_{\mathbb{R}}{\mathrm{d}}x \, \Psi_{n-k}^{n,t}(x)\Phi_{n-\ell}^{n,t}(x) = \delta_{k\ell}, \quad 1 \leq k,\ell\leq n.$$ The form of the biorthogonal functions can be guessed, with some experience, from the form of the kernel [@BP07].
\[lem:1b\] We have:
- $V_n$ is spanned by $\{x \mapsto {\mathrm{e}}^{\mu_k x}: 1\leq k \leq n\}$.
- The functions $\{\Phi_{n-k}^{n,t}:1\leq k\leq n\}$ are given by .
For any ${\varepsilon}>0$ we have $$\begin{gathered}
(\phi_k \ast \phi^{(k,n)})(x_k^{k-1},x) \\
= \int_{\mathbb{R}}{\mathrm{d}}y\, {\mathrm{e}}^{\mu_k y} \frac{(-1)^{n-k}}{2\pi{\mathrm{i}}}\int_{{\mathrm{i}}{\mathbb{R}}+\mu_{k+1}-{\varepsilon}} {\mathrm{d}}z\,\frac{{\mathrm{e}}^{z(x-y)}}{(z-\mu_{k+1})\dotsb(z-\mu_n)}.\end{gathered}$$ We split the $y$-integral into one over ${\mathbb{R}}_+$ and one over ${\mathbb{R}}_-$. Then we can exchange the integrals over ${\mathbb{R}}_-$ and the imaginary axis provided that and use $\int_{{\mathbb{R}}_-} {\mathrm{d}}y \, {\mathrm{e}}^{y(\mu_k-z)}=\frac{1}{z-\mu_k}$. In the same way we integrate over ${\mathbb{R}}_+$ taking $z$ such that $\mu_k<{\operatorname{Re}}z < \mu_{k+1}$. This gives $\int_{{\mathbb{R}}_+}{\mathrm{d}}y\,{\mathrm{e}}^{y(\mu_k-z)}=-\frac1{z-\mu_k}$. Putting these two integrals together we get $$(\phi_k \ast \phi^{(k,n)})(x_k^{k-1+1},x) = \frac{(-1)^{n-k}}{2\pi{\mathrm{i}}} \oint_{\Gamma_{\mu_k}}{\mathrm{d}}z\, \frac{{\mathrm{e}}^{x z}}{(z-\mu_k)(z-\mu_{k+1})\dotsb(z-\mu_n)},$$ which is a constant multiple of ${\mathrm{e}}^{\mu_k x}$. This proves (i). For (ii) we proceed similarly. Using that $1\leq k,\ell\leq n$ we have $$\begin{gathered}
\label{eq:12}
\int_{\mathbb{R}}{\mathrm{d}}x \, \Psi_{n-k}^{n,t}(x)\Phi_{n-\ell}^{n,t}(x) \\
= \frac{(-1)^{k+\ell}}{(2\pi{\mathrm{i}})^2} \int_{\mathbb{R}}{\mathrm{d}}x\int_{{\mathrm{i}}{\mathbb{R}}} {\mathrm{d}}z \oint_{\Gamma_{\mu_\ell,\dots,\mu_n}}{\mathrm{d}}w\, \frac{{\mathrm{e}}^{tz^2-x z}}{{\mathrm{e}}^{tw^2-x w}}\,\frac{(z-\mu_{k+1})\dotsb(z-\mu_n)}{(w-\mu_\ell)\dotsb(w-\mu_n)}.\end{gathered}$$ When integrating $x$ over ${\mathbb{R}}_-$, we take the $z$-integral such that ${\operatorname{Re}}z<{\operatorname{Re}}w$, and when we integrate $x$ over ${\mathbb{R}}_+$, we choose ${\operatorname{Re}}z> {\operatorname{Re}}w$. Thus, reduces to $$\frac{(-1)^{k+\ell}}{2\pi{\mathrm{i}}} \oint_{\Gamma_{\mu_\ell,\dots,\mu_n}} {\mathrm{d}}w\, \frac{(w-\mu_{k+1})\dotsb(w-\mu_n)}{(w-\mu_\ell)\dotsb(w-\mu_n)} = \delta_{k\ell}.$$ Finally, note that $$\Phi_{n-\ell}^{n,t}(x)= \sum_{i=\ell}^n b_i {\mathrm{e}}^{\mu_i x}\quad \text{ with} \quad b_i= \prod_{\substack{j=\ell\\j\neq i}}^n \frac{{\mathrm{e}}^{-t \mu_i^2/2}}{\mu_i-\mu_j}.$$ which shows that the set $\{\Phi_{n-k}^{n,t}:1\leq k\leq n\}$ spans $V_n$.
Next we verify Assumption (A) from Lemma 3.4 in [@BFPS06]. Indeed, $$\Phi_0^{n,t}(x) = \frac1{2\pi{\mathrm{i}}}\oint_{\Gamma_{\mu_n}} {\mathrm{d}}w\, \frac{{\mathrm{e}}^{-tw^2/2+xw}}{w-\mu_n} = c_n \phi_n(x_n^{n-1},x)$$ with $c_n={\mathrm{e}}^{-t\mu_n^2/2}\neq 0$ for $n=1,\dots,N$.
Finally, we can also determine the value of the normalization constant in , since it is given by $1/\det [M_{k\ell}]_{1\leq k,\ell\leq N}$ with $$M_{k\ell}=(\phi_k \ast \dotsb \ast \phi_N \ast \Psi_{N-\ell}^{N,t})(x_k^{k-1}).$$
\[lem:2\] We have $\det M = \prod_{n=1}^N {\mathrm{e}}^{t\mu_n^2/2}$, in particular $\det M >0$.
By Lemma \[lem:1\] (i) we may write $M_{k\ell} = (\phi_k \ast \Psi_{k-\ell}^{k,t})(x_k^{k-1})$. Thus, for $k\geq \ell$, $$M_{k\ell} = \int_{\mathbb{R}}{\mathrm{d}}y\, {\mathrm{e}}^{\mu_k y}\, \frac{(-1)^{k-\ell}}{2\pi {\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}} {\mathrm{d}}z\, {\mathrm{e}}^{tz^2/2-y z}\, (z-\mu_{\ell+1})\dotsb(z-\mu_k).$$ Once again, we let run the $y$-integral over ${\mathbb{R}}_-$ and ${\mathbb{R}}_+$ separately. In the first case we take the $z$-integral such that ${\operatorname{Re}}z < \mu_k$, in the second case such that ${\operatorname{Re}}z > \mu_k$. This allows us to exchange the integrals, which gives $$M_{k\ell} = \frac{(-1)^{k-\ell}}{2 \pi {\mathrm{i}}} \oint_{\Gamma_{\mu_k}} {\mathrm{d}}z\, {\mathrm{e}}^{tz^2/2} \, \frac{(z-\mu_{\ell+1})\dotsb (z-\mu_k)}{z-\mu_k}.$$ Since the integrand has no poles for $k>\ell$, we have $M_{k\ell}=0$ in this case, while for $k=\ell$ we get $M_{kk} = {\mathrm{e}}^{-t \mu_k^2/2}$. Thus, $M$ is upper triangular and the claim follows.
With the results of Lemma \[lem:1\] and Lemma \[lem:1b\], Theorem \[thm:1\] follows directly from Lemma 3.4 of [@BFPS06].
We have shown that Theorem \[thm:1\] holds when we impose $\mu_1 < \dotsb < \mu_N$. In particular, the joint density (\[eq:1\]) is given by an $(N(N+1)/2)$-point correlation function: With $m=N(N+1)/2$ we have $$\label{eq42}
(\ref{eq:1})=m!\, \rho^{(m)}(\{(\lambda_k^n,n),1\leq k \leq n \leq N\}).$$ Let $M>0$ be any fixed real number. The density (\[eq:1\]) is analytic in each of the $\mu_j$ in $[-M,M]$, $j=1,\dotsc,N$. The same holds for the correlation kernel (take e.g., $\mu_-=-M-1$). From this it follows that also the r.h.s. of (\[eq42\]) is analytic in each of the variables $\mu_1,\dotsc,\mu_N$. Since this holds for any $M$, by analytic continuation it follows that Theorem \[thm:1\] holds for any given drift vector $(\mu_1,\dotsc,\mu_N)\in{\mathbb{R}}^N$.
$2+1$ dynamics with different jump rates {#sec:4}
========================================
In this section we show that the correlation functions that we obtained for the GUE matrix diffusion with drifts can be obtained as a limit from an ${\mathrm{GT}}_N$-extension of TASEP with particle-dependent jump rates. This latter process was introduced in [@BF08]. Before we come to the convergence result, let us describe the model.
\[cl\][$x_1^1$]{} \[cl\][$x_1^2$]{} \[cl\][$x_1^3$]{} \[cl\][$x_1^4$]{} \[cl\][$x_1^5$]{} \[cl\][$x_2^2$]{} \[cl\][$x_2^3$]{} \[cl\][$x_2^4$]{} \[cl\][$x_2^5$]{} \[cl\][$x_3^3$]{} \[cl\][$x_3^4$]{} \[cl\][$x_3^5$]{} \[cl\][$x_4^4$]{} \[cl\][$x_4^5$]{} \[cl\][$x_5^5$]{} \[cl\][$n=1$]{} \[cl\][$n=2$]{} \[cl\][$n=3$]{} \[cl\][$n=4$]{} \[cl\][$n=5$]{} ![(Left) Initial particles configuration. (Right) A possible particles configuration after some time; in this configuration, if particle $(1,3)$ tries to jump, the move is suppressed (blocked by particle $(1,2)$), while if particle $(2,2)$ jumps, then also particles $(3,3)$ and $(4,4)$ move by one unit to the right. Particles at level $n$ have a jump rate $v_n$.[]{data-label="FigureParticles"}](FigParticles.eps "fig:"){width="\textwidth"}
At a fixed time $t$, let us denote by $x(t)=(x_k^n(t))_{1\leq k\leq n \leq N} \in {\mathrm{GT}}_N$ the positions of the $N(N+1)/2$ particles at time $t$. We choose initial conditions $x_k^n(0)=k-n-1$ and let the particles evolve as follows: Each particle $x_k^n$ has an independent exponential clock of rate $v_n>0$, i.e., particles on the same level have the same jump rates. When the $x_k^n$-clock rings, the particle jumps to the right by one, provided that $x_k^n < x_k^{n-1}-1$, otherwise we say that $x_k^n$ is blocked by $x_k^{n-1}$. If the $x_k^n$-particle can jump, we take the largest $c\geq 1$ such that $x_k^n = x_{k+1}^{n+1}=\dotsb=x_{k+c-1}^{n+c-1}$, and all $c$ particles in this string jump to the right by one, see Figure \[FigureParticles\] for an example). This ensures that at any time $t$, all the particles are in ${\mathrm{GT}}_N$. More precisely, these dynamics imply that the particles stay in a discrete version of ${\mathrm{GT}}_N$, namely $$\widetilde {\mathrm{GT}}_N=\{(x^1,x^2,\dotsc,x^N) \in {\mathbb{Z}}^1 \times {\mathbb{Z}}^2 \times \dotsb \times {\mathbb{Z}}^N : x_k^{n+1} < x_k^n \leq x_{k+1}^{n+1} \}.$$
The joint distribution of the particles has been calculated in Theorem 4.1 of [@BF07], and the result is $$\label{eq:22}
{\mathrm{const}}\times \det\bigl[\widetilde \Psi^{N,t}_{N-k}(x_\ell^N)\bigr]_{1\leq k,\ell\leq N} \prod_{n=1}^N \det\bigl[\widetilde \phi_n(x_i^{n-1},x_j^n)\bigr]_{1\leq i,j\leq n},$$ where $$\begin{aligned}
\widetilde \Psi^{N,t}_{N-k}(x) & = \frac1{2\pi{\mathrm{i}}} \oint_{\Gamma_0}{\mathrm{d}}z\, {\mathrm{e}}^{t/z} z^{x+N-1}(1-v_{k+1}z)\dotsb(1-v_Nz), \\
\widetilde \phi_n(x,y) & = (v_n)^{y-x}{\mathbbm{1}}_{[y\geq x]} \quad \text{ and } \quad \widetilde \phi_n(x_n^{n-1},y)=(v_n)^y.
\end{aligned}$$ Actually, Theorem 4.1 of [@BF07] is a statement about the marginal of a (possibly signed) measure. However, this model is the continuous time limit of a generic Markov chain introduced in Section 2 of [@BF08], from which it follows that the measure with fully packed initial conditions $y_n = x_1^n(0)=-n$ for $1\leq n \leq N$ is actually a probability distribution. The formulation of follows then from the theorem by taking $a(t)=t$ and $b(t)=0$ for all $t\geq 0$. Also note that we put the transition from time $t=0$ to time $t$ (which is encoded by $\mathcal T_{t,0}$ in the theorem) into $\Psi_{N-k}^N$. As shown in [@BF07], the correlation functions of this point process are determinantal, so what remains to do is the biorthogonalization for the generic jump rates.
\[prop1\] Consider a system of particles on $\widetilde {\mathrm{GT}}_N$ with fully packed initial conditions and dynamics described above. Then, at fixed time $t$, the corresponding point process has $m$-point correlation function $\tilde \varrho^m_t$ given by $$\tilde \varrho^m_t((x_1,n_1),\dotsc,(x_m,n_m)) = \det[\widetilde K^v_t((x_i,n_i),(x_j,n_j))]_{1\leq i,j\leq m}$$ with $(x_i,n_i)\in {\mathbb{R}}\times \{1,\dotsc,N\}$ and correlation kernel $$\widetilde K^v_t((x,n),(x',n'))= - \widetilde \phi^{(n,n')}(x,x')+\sum_{k=1}^{n'} \widetilde \Psi_{n-k}^{n,t}(x) \widetilde \Phi_{n'-k}^{n',t}(x'),$$ where $$\begin{aligned}
\widetilde \phi^{(n,n')}(x,x') & = \frac{1}{2\pi{\mathrm{i}}} \oint_{\Gamma_{0,v}} {\mathrm{d}}z\, \frac{1}{z^{x-x'+1}}\, \frac{z^{n'-n}}{(z-v_{n+1})\dotsb(z-v_{n'})} \, {\mathbbm{1}}_{[n<n']} \label{eq:27}, \\
\widetilde \Psi_{n-k}^{n,t}(x) & = \frac{1}{2\pi{\mathrm{i}}} \oint_{\Gamma_{0,v}} {\mathrm{d}}z \, \frac{{\mathrm{e}}^{tz}}{z^{x+n+1}} \, \frac{(z-v_1)\dotsb(z-v_n)}{(z-v_1)\dotsb(z-v_k)}, \label{eq:23} \\
\widetilde \Phi_{n-\ell}^{n,t}(x) & = \frac{1}{2\pi{\mathrm{i}}} \oint_{\Gamma_v}{\mathrm{d}}w \, \frac{w^{x+n}}{{\mathrm{e}}^{t w}} \, \frac{(w-v_1)\dotsb(w-v_{\ell-1})}{(w-v_1)\dotsb(w-v_n)}.\end{aligned}$$
By Proposition 3.1 of [@BF07], we have $$\widetilde \Psi_{n-k}^{n,t}(x) = \frac1{2\pi{\mathrm{i}}} \oint_{\Gamma_0} {\mathrm{d}}w\, w^{x+k-1} {\mathrm{e}}^{t/w} \, \frac{(1-v_1w)\dotsb(1-v_nw)}{(1-v_1w)\dotsb(1-v_kw)}$$ for $k\geq 1$. A change of variable $z=1/w$ then yields . Next we need to verify that $\{\widetilde \Psi_{n-k}^{n,t}:1\leq k \leq n\}$ is biorthogonal to $\{\widetilde \Phi_{n-\ell}^{n,t}:1\leq \ell\leq n\}$ (see Eq. (3.5) of [@BF07]). We split the sum over ${\mathbb{Z}}$ into two parts, one over $x\geq 0$ and one over $x<0$. Then, $$\begin{gathered}
\sum_{x \geq 0} \widetilde \Psi_{n-k}^{n,t}(x)\widetilde\Phi_{n-k}^{n,t}(x) \\
= \sum_{x \geq 0} \frac{1}{(2\pi{\mathrm{i}})^2} \oint_{\Gamma_v}{\mathrm{d}}w\oint_{\Gamma_0}{\mathrm{d}}z\, \frac{{\mathrm{e}}^{t z}}{{\mathrm{e}}^{t w}}\, \frac{w^{x+n}}{z^{x+n+1}} \, \frac{(z-v_{k+1})\dotsb(z-v_n)}{(w-v_\ell)\dotsb(w-v_n)}.\end{gathered}$$ We choose $\Gamma_0$ and $\Gamma_v$ such that $\lvert w \rvert \leq \lvert z \rvert$ which allows us to put the sum inside the integrals. This gives $$\begin{gathered}
\sum_{x \geq 0} \widetilde\Psi_{n-k}^{n,t}(x)\widetilde\Phi_{n-k}^{n,t}(x)\\
= \frac{1}{(2\pi{\mathrm{i}})^2} \oint_{\Gamma_v}{\mathrm{d}}w \oint_{\Gamma_{0,w}} {\mathrm{d}}z\, \frac{{\mathrm{e}}^{tz}}{{\mathrm{e}}^{tw}}\,\frac{w^n}{z^n} \, \frac{(z-v_{k+1})\dotsb(z-v_n)}{(w-v_k)\dotsb(w-v_\ell)} \, \frac{1}{z-w}.\end{gathered}$$ For $x<0$ we choose $\Gamma_0$ and $\Gamma_v$ such that they satisfy $\lvert w \rvert > \lvert z \rvert$ which gives $$\begin{gathered}
\sum_{x < 0} \widetilde\Psi_{n-k}^{n,t}(x)\widetilde\Phi_{n-k}^{n,t}(x) \\
= - \frac{1}{(2\pi{\mathrm{i}})^2} \oint_{\Gamma_0}{\mathrm{d}}z \oint_{\Gamma_{v,z}} {\mathrm{d}}w\, \frac{{\mathrm{e}}^{tz}}{{\mathrm{e}}^{tw}}\,\frac{w^n}{z^n} \, \frac{(z-v_{k+1})\dotsb(z-v_n)}{(w-v_k)\dotsb(w-v_\ell)} \, \frac{1}{z-w}.\end{gathered}$$ Thus, $$\sum_{x\in{\mathbb{Z}}} \widetilde\Psi_{n-k}^{n,t}(x)\widetilde\Phi_{n-k}^{n,t}(x) = \frac{1}{2\pi{\mathrm{i}}} \oint_{\Gamma_v} {\mathrm{d}}w \, \frac{(w-v_{k+1})\dotsb(w-v_n)}{(w-v_\ell)\dotsb(w-v_n)} = \delta_{k\ell}.$$ Finally, we show that $\{\widetilde \Phi_{n-\ell}^{n,t}:1\leq \ell\leq n\}$ spans the space of functions $V_n$. We denote by $u_1<\dotsb<u_\nu$ the different values of $v_1,\dotsc,v_n$ and $\alpha_k$ the multiplicity of $u_k$, i.e., $\alpha_1+\dotsb+\alpha_\nu = n$. Then, we may write $$\begin{aligned}
\widetilde \Phi_{n-1}^{n,t}(x) & = \frac{1}{2\pi{\mathrm{i}}} \oint_{\Gamma_v}{\mathrm{d}}w \, \frac{w^{x+n}}{{\mathrm{e}}^{t w}} \, \frac{1}{(w-u_1)^{\alpha_1}\dotsb(w-u_\nu)^{\alpha_\nu}} \\
& = \sum_{i=1}^\nu\frac1{(\alpha_i-1)!}\,\frac{{\mathrm{d}}^{\alpha_i-1}}{{\mathrm{d}}w^{\alpha_i-1}}\bigg|_{w=u_i} \biggl( \frac{w^{x+n}}{{\mathrm{e}}^{t w}} \prod_{j \neq i} \frac1{(w-u_j)^{\alpha_j}}\biggr) \\
& = \sum_{i=1}^\nu (u_i)^x \sum_{j=1}^{\alpha_i} c_{i,j} x^{j-1}.
\end{aligned}$$ For $\ell=2,\dotsc,n$, we can represent $\widetilde \Phi_{n-\ell}^{n,t}$ in the same way, but with exponents $\alpha_{\ell,i} \le\alpha_i$, $1\leq i\leq \nu$. Since $(\alpha_{k,1},\dotsc,\alpha_{k,\nu}) \neq (\alpha_{\ell,1},\dotsc,\alpha_{\ell,\nu})$ for $k\neq \ell$, this shows that $${\operatorname{span}}\{\widetilde \Phi_{n-\ell}^{n,t} : 1\leq \ell \leq n\} = {\operatorname{span}}\{x\mapsto (u_i)^x x^{j-1} :1\leq u \leq \nu, 1 \leq j \leq \alpha_i\},$$ which is $V_n$.
We continue by establishing the convergence result under the scaling (\[scaling\]). Correspondingly, we rescale (and conjugate) the kernel $\widetilde K_t$ and define the rescaled kernel as $$K_{\tau,T,\mathrm{resc}}^\mu((\xi,n),(\xi',n'))
=\frac{T^{n'/2}}{T^{n/2}} \, \sqrt T \, \widetilde K_{\tau T}^{\mu_T}\bigl(([\tau T - \xi \sqrt{T}],n),([\tau T-\xi' \sqrt{T}],n')\bigr)$$ where $[\,\cdot\,]$ denotes the integer part, and the drift $v$ is now $\mu_T=1-\mu/\sqrt T$. Of course, $T$ is assumed to be so large that $\mu_T>0$ is satisfied.
\[prop:1\] For any fixed $L>0$, the rescaled kernel $K_{\tau,T,\mathrm{resc}}^\mu$ converges, uniformly for $\xi,\xi'\in[-L,L]$, as $$\lim_{T\to\infty} K_{\tau,T,\mathrm{resc}}^\mu((\xi,n),(\xi',n')) = K_\tau^\mu((\xi,n),(\xi',n'))$$ with $K_\tau^\mu\equiv K_\tau$ given in .
Let us define the rescaled functions $$\begin{aligned}
\phi^{(n,n')}_{T,\mathrm{resc}}(\xi,\xi') & = T^{-(n'-n+1)/2} \, \widetilde \phi^{(n,n')}(\tau T + \xi \sqrt T,\tau T + \xi' \sqrt T), \\
\Psi_{n-k, T,\mathrm{resc}}^{n,\tau}(\xi) & = T^{(n-k+1)/2} {\mathrm{e}}^{-\tau T} \, \widetilde \Psi_{n-k}^{n,\tau T}(\tau T + \xi \sqrt T), \\
\Phi_{n-k, T,\mathrm{resc}}^{n,\tau}(\xi') & = T^{-(n-k)/2} {\mathrm{e}}^{\tau T} \, \widetilde \Phi_{n-k}^{n,\tau T}(\tau T + \xi' \sqrt T),
\end{aligned}$$ where we also rescale the jump rates as in . We have to show that these functions converge to their analogues from –. We first verify that $\phi^{(n,n')}_{T,\mathrm{resc}}(\xi,\xi')\to \phi^{(n,n')}(\xi,\xi')$ with $n<n'$. For $y\geq y'$, the integrand of $\widetilde \phi^{(n,n')}(y,y')$ in has residue $0$ at infinity and thus the whole integral vanishes, while for $y<y'$, there is no pole at $z=0$ and therefore $$\widetilde \phi^{(n,n')}(y,y')=\sum_{i=n+1}^{n'} v_i^{(y'-y)+(n'-n)-1} \prod_{j\neq i} \frac{1}{v_i-v_j} \, {\mathbbm{1}}_{[y<y']}.$$ Hence, for its rescaled version, $$\phi_{T,\mathrm{resc}}^{(n,n')}(\xi,\xi')= \sum_{i=n+1}^{n'} \left( 1-\frac{\mu_i}{\sqrt T }\right)^{(\xi-\xi')\sqrt T +(n'-n)-1} \prod_{j\neq i}\frac{1}{\mu_j-\mu_i}\,{\mathbbm{1}}_{[\xi>\xi']},$$ which, as $T\to \infty$, converges to $$\sum_{i=n+1}^{n'} {\mathrm{e}}^{\mu_i(\xi'-\xi)} \prod_{j\neq i} \frac{1}{\mu_j-\mu_i}{\mathbbm{1}}_{\{\xi>\xi'\}} = \frac{(-1)^{n'-n}}{2\pi{\mathrm{i}}} \int_{{\mathrm{i}}{\mathbb{R}}+\mu_-} \hspace{-1em}{\mathrm{d}}z \, \frac{{\mathrm{e}}^{z(\xi'-\xi)}}{(z-\mu_{n+1})\dotsb(z-\mu_{n'})}.$$
Next we show that $\Psi_{n-k, T,\mathrm{resc}}^{n,\tau}(\xi)\to \Psi_{n-k}^{n,\tau}(\xi)$ uniformly for $\xi \in [-L,L]$. We have $$\begin{aligned}\label{eq:28}
\Psi_{n-k, \mathrm{resc}}^{n,\tau,T}(\xi) & = \frac{\sqrt T}{2\pi {\mathrm{i}}} \oint_{\Gamma_{0,v}} {\mathrm{d}}z \, \frac{{\mathrm{e}}^{\tau T(z-1)}}{z^{\tau T-\xi\sqrt T+n+1}} \, g(z) \\
& = \frac{\sqrt T}{2\pi {\mathrm{i}}} \oint_{\Gamma_{0,v}} {\mathrm{d}}z\, {\mathrm{e}}^{\tau T f_0(z)+\sqrt T f_1(z) +f_2(z)} g(z)
\end{aligned}$$ with $f_0(z)=z-1-\ln z$, $f_1(z)=\xi \ln z$, $f_2(z)=-(n+1)\ln z$, and $$g(z)= \frac{(\sqrt T(z-1)+\mu_1)\dotsb (\sqrt T(z-1)+\mu_n)}{(\sqrt T(z-1)+\mu_1)\dotsb(\sqrt T(z-1)+\mu_k)}.$$ A Taylor expansion around the double critical point of $f_0$, i.e., around $z_c=1$ gives $$\begin{aligned}
f_0(z) & = \frac12(z-1)^2 + {\mathcal{O}}((z-1)^3), \\
f_1(z) & = \xi(z-1) + {\mathcal{O}}((z-1)^2), \\
f_2(z) & = 0+{\mathcal{O}}(z-1).
\end{aligned}$$
Fix $r>1-\mu_-/\sqrt{T}$ and deform $\Gamma_{0,v}$ to the contour $\gamma= \gamma_1 \cup \gamma_2$ with $$\gamma_1 = 1-\mu_-/\sqrt{T} + {\mathrm{i}}[-r,r], \quad \gamma_2 = \{\lvert z \rvert = r\} \cap \{{\operatorname{Re}}z < 1-\mu_-/\sqrt{T}\}.$$ Let us verify that $\gamma$ is a steep descent path for $f_0$. On the segment $\gamma_1$, we have that ${\operatorname{Re}}f_0(x+{\mathrm{i}}y)=x-1-\frac12\ln(x^2+y^2)$, so $$\frac{{\mathrm{d}}{\operatorname{Re}}f_0(x+{\mathrm{i}}y)}{{\mathrm{d}}y} = - \frac{y}{x^2+y^2}, \qquad y \in [-r,r],$$ with $x=1-\mu_-/\sqrt{T}$. Thus $f_0$ is strictly increasing on $1-\frac{\mu_-}{\sqrt T}+{\mathrm{i}}[-r,0)$ and strictly decreasing on $1-\frac{\mu_-}{\sqrt T}+{\mathrm{i}}(0,r]$. On $\gamma_2$ we compute $${\operatorname{Re}}f_0(r {\mathrm{e}}^{{\mathrm{i}}\varphi}) = r \cos \varphi -1 -\ln r, \qquad \varphi \in (\arccos \tfrac1r,2\pi-\arccos \tfrac1r),$$ which means that $f_0$ is strictly decreasing on $\gamma_2 \cap \{{\operatorname{Im}}z>0\}$ and strictly increasing on $\gamma_2 \cap \{{\operatorname{Im}}z < 0\}$. Thus $\gamma$ is a steepest descent path for $f_0$ and the major contribution comes from a line segment $\gamma_\delta = 1-\frac{\mu_-}{\sqrt T}+{\mathrm{i}}[-\delta,\delta]$ for any $\delta \in (0,1)$. Indeed, the error we make when we integrate along $\gamma_\delta$ instead of $\gamma$ is of order ${\mathcal{O}}({\mathrm{e}}^{-c T})$ with $c\sim \delta^2$. We therefore consider the integral on $\gamma_\delta$ only, $$\label{eq:29}
\frac{\sqrt T}{2\pi{\mathrm{i}}} \int_{\gamma_\delta} {\mathrm{d}}z \, g(z) {\mathrm{e}}^{\xi \sqrt T (z-1)+\frac {\tau T} 2 (z-1)^2} {\mathrm{e}}^{{\mathcal{O}}\left(z-1,\sqrt T (z-1)^2,T (z-1)^3\right)}.$$ Using $\lvert {\mathrm{e}}^x-1\rvert \leq \lvert x \rvert {\mathrm{e}}^{\lvert x \rvert}$, the difference between and the same integral without the error term can be bounded by $$\begin{gathered}
\frac{\sqrt T}{2\pi}\int_{\gamma_\delta} {\mathrm{d}}z\, \Big| {\mathrm{e}}^{c_1 \xi \sqrt T (z-1)+c_2\frac {\tau T} 2 (z-1)^2}\\
\times {\mathcal{O}}\left(z-1,\sqrt T (z-1)^2,T(z-1)^3,T^{(n-k)/2}(z-1)^{n-k}\right)\Big|\end{gathered}$$ for some constants $c_1$ and $c_2$ that can be chosen arbitrarily close to $1$ as $\delta \to 0$. By a change of variable $Z = \sqrt T(1-z)$ one then sees that this error is of order ${\mathcal{O}}(T^{-1/2})$. Hence we can consider the integral in without the error term, which simplifies to $$\frac{\sqrt T}{2\pi{\mathrm{i}}} \int_{\gamma_\delta} {\mathrm{d}}z\,{\mathrm{e}}^{\tau T(z-1)^2/2+\xi \sqrt T(z-1)}g(z).$$ The error we make if we extend $\gamma_\delta$ to $1-\frac{\mu_-}{\sqrt T}+{\mathrm{i}}{\mathbb{R}}$ is of order ${\mathcal{O}}({\mathrm{e}}^{-c T})$. All together the integral from agrees, up to an error ${\mathcal{O}}({\mathrm{e}}^{-c T},T^{-1/2})$ uniform in $\xi \in [-L,L]$, with $$\frac{\sqrt T}{2\pi{\mathrm{i}}} \int_{1-\frac{\mu_-}{\sqrt T}+{\mathrm{i}}{\mathbb{R}}} {\mathrm{d}}z \, {\mathrm{e}}^{\tau T (z-1)^2/2+\xi \sqrt T(z-1)} g(z),$$ where the poles of $g$ lie on the left of the integration axis. After a change of variable $Z=-\sqrt T(z-1)$ this integral can be identified as $\Psi_{n-k}^{n,\tau}(\xi)$.
Finally we show that $\Phi_{n-k,T,\mathrm{resc}}^{n,\tau}(\xi') \to \Phi_{n-k}^{n,\tau}(\xi')$. We have $$\begin{gathered}
\Phi_{n-k, T,\mathrm{resc}}^{n,\tau}(\xi') = \frac{\sqrt T}{2\pi{\mathrm{i}}} \oint_{\Gamma_v}{\mathrm{d}}w \, {\mathrm{e}}^{\tau T(\ln w-w+1)-\sqrt T \xi' \ln w + n \ln w} \\
\times \frac{(\sqrt T(w-1)+\mu_1)\dotsb(\sqrt T(w-1)+\mu_{k-1})}{(\sqrt T(w-1)+\mu_1)\dotsb(\sqrt T(w-1)+\mu_n)},\end{gathered}$$ and by a change of variable $W=-\sqrt T(w-1)$ and a Taylor expansion in the exponent we get $$\begin{gathered}
\Phi_{n-k,T,\mathrm{resc}}^{n,\tau}(\xi') \\
= \frac{(-1)^{n-k+1}}{2\pi{\mathrm{i}}} \oint_{\Gamma_a} {\mathrm{d}}W\, {\mathrm{e}}^{-\tau W^2/2+\xi' W+{\mathcal{O}}(T^{-1/2})} \, \frac{(w-\mu_1)\dotsb (w-\mu_{k-1})}{(w-\mu_1)\dotsb(w-\mu_n)},\end{gathered}$$ which converges uniformly for $\xi'\in[-L,L]$ to $\Phi_{n-k}^{n,\tau}(\xi')$.
With the above results we can now prove Theorem \[thm:3\].
Set $m=N(N+1)/2$ and define $n_1,\dotsc,n_m$ by $$n_1=1, \enspace n_2=n_3=2, \enspace n_4=n_5=n_6 = 3, \enspace \dotsc, \enspace n_{m-N+1}=\dotsb = n_m =N.$$ For $A \subseteq {\mathbb{R}}^m$ we set $A_T = (\tau T-\sqrt T A)\cap {\mathbb{Z}}$. Then, we have $$\begin{aligned}
\nu_T&(A) = \sum_{(x_1,\dotsc,x_m) \in A_T} \det\bigl[ \widetilde K_{\tau T}^{\mu_T} ((x_i,n_i),(x_j,n_j)) \bigr]_{1\leq i,j\leq m} \\
& = T^{m/2} \int_A{\mathrm{d}}^m x \det\bigl[\widetilde K_{\tau T}^{\mu_T} ((\tau T - [x_i] \sqrt T,n_i),(\tau T-[x_j]\sqrt T,n_j))\bigr]_{1\leq i,j\leq m} \\
& = \int_A{\mathrm{d}}^m x \det\bigl[\widetilde K_{\tau,T,\mathrm{resc}}(([x_i],n_i),([x_j],n_j))\bigr]_{1\leq i,j\leq m}
\end{aligned}$$ and $$\nu(A) = \int_A {\mathrm{d}}^m x \det\bigl[K_\tau^\mu((x_i,n_i),(x_j,n_j))\bigr]_{1\leq i,j\leq m}.$$ Since the determinants are continuous functions of the kernels, we have by Proposition \[prop:1\] that $$\begin{gathered}
\lim_{T\to \infty}\det\bigl[\widetilde K^\mu_{\tau,T,\mathrm{resc}}(([x_i],n_i),([x_j],n_j))\bigr]_{1\leq i,j\leq m}\\
= \det\bigl[K_\tau^\mu((x_i,n_i),(x_j,n_j))\bigr]_{1\leq i,j\leq m}\end{gathered}$$ for all $x_1,\dotsc,x_m \in {\mathbb{R}}$. Thus we have shown that the densities of the probability measures in question converge pointwise to each other. Then, is a direct consequence of Scheffé’s theorem, see e.g. [@Bil68].
Warren’s process with drifts {#sec:3}
============================
We have seen in Section \[sec:2\] that the eigenvalues’ density can be written as a product of determinants, and, in Lemma \[lem:2\], we calculated the normalization constant, so that the probability measure on the eigenvalues reads $$\label{eq:19}
{\mathbb{P}}\biggl( \bigcap_{1\leq k\leq n \leq N} \{\lambda_k^n \in {\mathrm{d}}\lambda_k^n\}\biggr) = \tilde p_t(\lambda) \, {\mathrm{d}}\lambda$$ with ${\mathrm{d}}\lambda=\prod_{1\leq k\leq n\leq N} {\mathrm{d}}\lambda^n_k$, and $$\label{eqDensity}
\tilde p_t(\lambda)= \det\bigl[\Psi_{N-k}^{N,t}(\lambda_\ell^N)\bigr]_{1\leq k,\ell \leq N} \prod_{n=1}^N {\mathrm{e}}^{-t \mu_n^2/2} \prod_{n=1}^N \det\bigl[\phi_n(\lambda_i^{n-1},\lambda_j^n)\bigr]_{1\leq i,j\leq n}$$ In this section we explain the connection to a system of Brownian motions in ${\mathrm{GT}}_N$. More precisely, we consider Brownian motions $\{B_k^n, 1\leq k \leq n \leq N\}$ in ${\mathrm{GT}}_N$ starting from $0$, with drift $\mu_n$, and interacting as follows:
- The evolution of $B_k^n$ does not depend on the Brownian motions with higher upper index ($B_\ell^m$ for $m\geq n+1$, and any $\ell$);
- $B_k^{n}$ is reflected off $B_k^{n-1}$ and $B_{k-1}^{n-1}$.
These reflections are sometimes called oblique reflections [@WW84], since in the $(x_k^{n-1},x_k^n)$-plane (resp. $(x_{k-1}^{n-1},x_k^n)$-plane) the reflection directions are not normal, but oblique as indicated in Figure \[FigReflections\]. Note that the projection on $\{B_1^n,1\leq n \leq N\}$ differs from the process studied in [@PP08], where the reflections are in the normal direction.
![The two reflection types in our system. They correspond to the boundary condition (\[eq:21\]).[]{data-label="FigReflections"}](bc1)
![The two reflection types in our system. They correspond to the boundary condition (\[eq:21\]).[]{data-label="FigReflections"}](bc2)
Let us now describe the system of Brownian motions. Denote by $p_t$ be the probability density of the Brownian motions in ${\mathrm{GT}}_N$ (its existence will be a consequence of our result). Following [@HW87], where Brownian motions with oblique reflections were studied, for a Brownian motion with drift $\mu$ reflected at the boundary in the direction $v$, the boundary conditions on the density function may be expressed as follows. Denote by $n$ the normal vector of the boundary, let $v$ be normalized such that $n\cdot v=1$ and let $q=v-n$. Moreover, set $\nabla_T=\nabla - n (n\cdot \nabla)$, $D^*=n\cdot \nabla - q\cdot \nabla_T$. Then, the boundary condition can be written as $$D^* p_t = (\nabla_T \cdot q + 2 \mu \cdot n) p_t \quad \textrm{on the boundary}.$$ Specializing to our case, we get $$\label{eq:21}
\frac{\partial}{\partial x_k^n} \, p_t(x) + (\mu_{n+1}-\mu_n)p_t(x)=0,$$ whenever $x_k^n=x_k^{n+1}$ or $x_k^n=x_{k+1}^{n+1}$, for $1\leq k\leq N-1$.
This process, without drifts, was introduced by Warren in [@War07], where he determined the transition probability for any initial condition and also showed that the process is well-defined when starting from $0$. We here consider a system of Brownian motions with constant (bounded) drifts, which can be expressed as follows, $$\begin{aligned}
B_1^1(t)&=\mu_1 t + b_1^1(t),\\
B_1^n(t)&=\mu_n t + b_1^n(t)-L_{B_1^{n-1}-B_1^n}(t),\quad n=2,\dotsc,N,\\
B_k^n(t)&=\mu_n t + b_k^n(t)-L_{B_k^{n-1}-B_k^n}(t)+L_{B_k^n-B_{k-1}^{n-1}}(t),\quad 2\leq k < n \leq N,\\
B_n^n(t)&=\mu_n t + b_n^n(t)+L_{B_n^n-B_{n-1}^{n-1}}(t),\quad n=2,\dotsc,N,
\end{aligned}$$ where the $b_k^n$, $1\leq k \leq n \leq N$, are independent standard Brownian motions and $L_{X-Y}(t)$ is twice the semimartingale local time at zero of $X(t)-Y(t)$. The question of well-definedness was related to the, a priori possible, presence of triple collisions. Bounded drifts do not influence this property as can be seen by applying Girsanov’s theorem like in the works [@IK10; @KPS12].
Reflected Brownian motions can be also defined as follows. A standard one-dimensional reflected Brownian motion can also be defined to be the image under the Skorokhod map of standard Brownian motion. More precisely, one define a Brownian motion, $B(t)$, starting from $y\in{\mathbb{R}}$ and being reflected at some continuous function $f(t)$ with $f(0)<y$ is via the Skorokhod representation [@Sko61; @AO76] $$\begin{aligned}
B(t)&=y+b(t)-\min\big\{0,\inf_{0\leq s \leq t}(y+b(s)-f(s))\big\} \\
&= \max\big\{y+b(t),\sup_{0\leq s \leq t}(f(s)+b(t)-b(s))\big\},
\end{aligned}$$ where $b$ is a standard Brownian motion starting at $0$. In this paper we use as definition of the Warren process with drifts to be the image of independent Brownian motions under the extended Skorokhod map introduced by Burdzy, Kang and Ramanan (see Theorem 2.6 of [@BKR09] for an explicit formula).
Consider a particle system as in Section \[sec:4\] but where the particles evolves independently, i.e., $\tilde x_k^n(0)=-n+k-1$ for $1\leq k \leq n\leq N$ and the evolution of $\tilde x_k^n(t)$ is a continuous time random walk with jump rate $v_n$. Consider now the scaling (\[scaling\]) $$t=\tau T,\quad \tilde B_k^n=\frac{\tilde x_k^n-\tau T}{-\sqrt{T}}, \quad v_n=1-\frac{\mu_n}{\sqrt{T}}.$$ The $\tilde x_k^n$’s are independent, so in the $T\to\infty$ limit, converges weakly to a -dimensional Brownian motion , where $B_k^n$ has drift $\mu_n$ (see Donsker’s theorem). As shown in [@GS12] by Gorin and Shkolnikov, the particle with the blocking/pushing dynamics converges wearkly as $T\to\infty$ to the Warren process with level-dependent drifts. To be precise, they first showed the convergence for the drift-less case, where the limit process is the Warren process. However the same proof applies for more generic cases including the one of this paper, see Remark 10 of [@GS12].
In Proposition \[prop1\] we have proven that the correlation functions has a limit as $T\to\infty$. Further, the integral of the density is one, so that no mass is lost at infinity or localized in some Dirac mass. Thus, the $n$-point correlation function of the reflected Brownian motion is the $T\to\infty$ limit of the $n$-point correlation function for the interacting particle system.
For completeness, let us remark that the transition density $p_t$ in ${\mathrm{GT}}_N$ satisfies:\
(1) the Fokker-Planck equation (or Kolmogorov forward equation) $$\label{eq:16}
\frac{\partial}{\partial t}\, p_t(x)=\sum_{n=1}^N \sum_{k=1}^n \left( \frac{1}{2}\, \frac{\partial^2}{\partial (x_k^n)^2}-\mu_n \frac{\partial}{\partial x_k^n}\right) p_t(x),$$ (2) the initial condition $$\label{eq:20}
\lim_{t\searrow 0} p_t(x) {\mathrm{d}}x = \prod_{1\leq k\leq n\leq N} \delta_{x_k^n},$$ (3) the boundary condition (\[eq:21\]).
\[prop:2\] Denote by $p_t: {\mathrm{GT}}_N \to [0,1]$ be the probability density defined in (\[eqDensity\]). Inside ${\mathrm{GT}}_N$, this density satisfies the Fokker-Planck equation (\[eq:16\]), the initial condition (\[eq:20\]), and the boundary condition (\[eq:21\]).
First observe that by setting $\tilde \Psi_{N-k}^{N,t}(x) = {\mathrm{e}}^{\mu_N x} \Psi_{N-k}^{N,t}(x)$, we can rewrite as a probability measure on ${\mathrm{GT}}_N$ with density $$\tilde p_t(x) = \det\left[ \tilde \Psi_{N-k}^{N,t}(x_\ell^N)\right]_{1\leq k,\ell\leq N} \prod_{k=1}^N {\mathrm{e}}^{-t\mu_k^2/2} \prod_{n=1}^{N-1} \prod_{k=1}^n {\mathrm{e}}^{(\mu_n-\mu_{n+1})x_k^n}.$$ for $x=(x_k^n)_{1\leq k\leq n \leq N}\in {\mathrm{GT}}_N$. The double product only depends on $(x_k^n)_{1\leq k\leq n\leq N-1}$, while the determinant is a function of $(x_k^N)_{1\leq k\leq N}$. We have $$\frac{1}{2} \frac{\partial^2}{\partial x^2}\,\tilde \Psi_{N-k}^{N,t}(x)= \frac{\partial}{\partial t}\,\tilde \Psi_{N-k}^{N,t}(x)+\mu_N \frac{\partial}{\partial x} \tilde \Psi_{N-k}^{N,t}(x) -\frac{\mu_N^2}{2} \tilde \Psi_{N-k}^{N,t}(x),$$ from which follows that $$\label{eq:13}
\frac{1}{2} \sum_{\ell=1}^N \frac{\partial^2}{\partial (x_\ell^N)^2}\,\tilde p_t(x) = \frac{\partial}{\partial t}\,\tilde p_t(x)+\mu_N \sum_{\ell=1}^N \frac{\partial}{\partial x_\ell^N}\,\tilde p_t(x)+\frac12 \left(\sum_{n=1}^N \mu_n^2-N \mu_N^2\right)\tilde p_t(x).$$ For $k=1,\dots,N-1$, we have $$\label{eq:14}
\frac{\partial}{\partial x_k^n} \, \tilde p_t(x)=(\mu_n-\mu_{n+1})\tilde p_t(x), \quad \frac{\partial^2}{\partial (x_k^n)^2} \, \tilde p_t(x)=(\mu_n-\mu_{n+1})^2 \tilde p_t(x),$$ and thus, putting and together, $$\begin{gathered}
\label{eq:15}
\frac{1}{2} \sum_{n=1}^N\sum_{k=1}^n \frac{\partial^2}{\partial(x_k^n)^2}\,\tilde p_t(x) = \frac{\partial}{\partial t}\,\tilde p_t(x) +\mu_N \sum_{k=1}^N \frac{\partial}{\partial x_k^N}\,\tilde p_t(x) \\
+ \frac{1}{2} \left(\sum_{n=1}^N \mu_n^2-N \mu_N^2+ \sum_{n=1}^{N-1} n(\mu_n-\mu_{n+1})^2 \right)\tilde p_t(x).\end{gathered}$$ Using that $$\label{eq:17}
N \mu_N^2-\sum_{n=1}^N \mu_n^2 = \sum_{n=1}^{N-1}n(\mu_{n+1}^2-\mu_n^2) = \sum_{n=1}^{N-1}n(\mu_n-\mu_{n+1})^2 - 2 \sum_{k=1}^{N-1} n \mu_n(\mu_n-\mu_{n+1})$$ the expression between the brackets in simplifies to $2\sum n\mu_n(\mu_n-\mu_{n+1})$. On the other hand, $$\label{eq:18}
\sum_{n=1}^N \sum_{k=1}^n \mu_n \frac{\partial}{\partial x_k^n} \,\tilde p_t(x)= \sum_{n=1}^{N-1}n\mu_n(\mu_n-\mu_{n+1})\tilde p_t(x)+\mu_N \sum_{k=1}^N \frac{\partial}{\partial x_k^N} \, \tilde p_t(x).$$ Then, follows from , and . The initial condition is satified because as $t\searrow 0$, we obtain the Dirac measure at $x_k^N=0$ for $1\leq k \leq N$ and since we consider $\tilde p_t$ on ${\mathrm{GT}}_N$, this immediately implies that $x^k_n=0$ for all $1\leq k \leq n \leq N-1$. Finally, the boundary condition holds trivially by .
The three conditions in Proposition \[prop:2\] are, in general, not enough to prove that $\tilde p_t=p_t$. For that, one would need the backwards equation.
Determinantal correlations
==========================
Since we refer several times to Lemma 3.4 of [@BFPS06], we report it here.
Assume we have a signed measure on $\{x_i^n,n=1,\dotsc,N,i=1,\dotsc,n\}$ given in the form, $$\label{Sasweight}
\frac{1}{Z_N}\prod_{n=1}^{N-1} \det[\phi_n(x_i^n,x_j^{n+1})]_{1\leq i,j\leq n+1} \det[\Psi_{N-i}^{N}(x_{j}^N)]_{1\leq i,j \leq N},$$ where $x_{n+1}^n$ are some “virtual” variables and $Z_N$ is a normalization constant. If $Z_N\neq 0$, then the correlation functions are determinantal.
To write down the kernel we need to introduce some notations. Define $$\label{Sasdef phi12}
\phi^{(n_1,n_2)}(x,y)=
\begin{cases} (\phi_{n_1} \ast \dotsb \ast \phi_{n_2-1})(x,y),& n_1<n_2,\\
0,& n_1\geq n_2,
\end{cases}.$$ where $(a* b)(x,y)=\sum_{z\in{\mathbb{Z}}}a(x,z) b(z,y)$, and, for $1\leq n<N$, $$\label{Sasdef_psi}
\Psi_{n-j}^{n}(x) := (\phi^{(n,N)} * \Psi_{N-j}^{N})(y), \quad j=1,\dotsc,N.$$ Set $\phi_0(x_1^0,x)=1$. Then the functions $$\{ (\phi_0*\phi^{(1,n)})(x_1^0,x), \dots,(\phi_{n-2}*\phi^{(n-1,n)})(x_{n-1}^{n-2},x), \phi_{n-1}(x_{n}^{n-1},x)\}$$ are linearly independent and generate the $n$-dimensional space $V_n$. Define a set of functions $\{\Phi_j^{n}(x), j=0,\dotsc,n-1\}$ spanning $V_n$ defined by the orthogonality relations $$\label{Sasortho}
\sum_x \Phi_i^n(x) \Psi_j^n(x) = \delta_{i,j}$$ for $0\leq i,j\leq n-1$.
Further, if $\phi_n(x_{n+1}^n,x)=c_n \Phi_0^{(n+1)}(x)$, for some $c_n\neq 0$, , then the kernel takes the simple form $$\label{SasK}
K(n_1,x_1;n_2,x_2)= -\phi^{(n_1,n_2)}(x_1,x_2)+ \sum_{k=1}^{n_2} \Psi_{n_1-k}^{n_1}(x_1) \Phi_{n_2-k}^{n_2}(x_2).$$
[10]{}
M. Adler, E. Nordenstam, and P. van Moerbeke, *[Consecutive Minors for Dyson’s Brownian Motions]{}*, arXiv:1007.0220 (2010).
M. Adler, P. van Moerbeke, and D. Wang, *Random matrix minor processes related to percolation theory*, arXiv:1301.7017 (2013).
R.F. Anderson and S. Orey, *Small random perturbation of dynamical systems with reflecting boundary*, Nagoya Math. J. **60** (1976), 189–216.
Y. Baryshnikov, *[GUEs and queues]{}*, Probab. Theory Relat. Fields **119** (2001), 256–274.
P. Billingsley, *[Convergence of Probability Measures]{}*, Wiley ed., New York, 1968.
A. Borodin and P.L. Ferrari, *[Anisotropic Growth of Random Surfaces in $2+1$ Dimensions]{}*, To appear in Comm. Math. Phys.; arXiv:0804.3035 (2008).
A. Borodin and P.L. Ferrari, *[Large time asymptotics of growth models on space-like paths I: PushASEP]{}*, Electron. J. Probab. **13** (2008), 1380–1418.
A. Borodin, P.L. Ferrari, M. Pr[ä]{}hofer, and T. Sasamoto, *[Fluctuation properties of the TASEP with periodic initial configuration]{}*, J. Stat. Phys. **129** (2007), 1055–1080.
A. Borodin and V. Gorin, *Shuffling algorithm for boxed plane partitions*, Adv. Math. **220** (2009), 1739–1770.
A. Borodin and J. Kuan, *[Asymptotics of Plancherel measures for the infinite-dimensional unitary group]{}*, Adv. Math. **219** (2008), 894–931.
A. Borodin and J. Kuan, *[Random Surface Growth with a Wall and Plancherel Measures for $O(\infty)$]{}*, Comm. Pure Appl. Math. **63** (2010), 831–894.
A. Borodin and S. Péché, *[Airy Kernel with Two Sets of Parameters in Directed Percolation and Random Matrix Theory]{}*, J. Stat. Phys. **132** (2008), 275–290.
K. Burdzy, W. Kang, and K. Ramanan, *[The Skorokhod problem in a time-dependent interval]{}*, Stoch. Processes Appl. **119** (2009), 428–452.
N. Elkies, G. Kuperbert, M. Larsen, and J. Propp, *[Alternating-Sign Matrices and Domino Tilings I and II]{}*, J. Algebr. Comb. **1** (1992), 111–132.
P.L. Ferrari and R. Frings, *[On the Partial Connection Between Random Matrices and Interacting Particle Systems]{}*, J. Stat. Phys. **141** (2010), 613–637.
P. J. Forrester and T. Nagao, *[Determinantal Correlations for Classical Projection Processes]{}*, J. Stat. Mech. (2011), P08011.
P. J. Forrester and E. Nordenstam, *[The Anti-Symmetric GUE Minor Process]{}*, Mosc. Math. J. **9** (2009), 749––774.
V. Gorin and M. Shkolnikov, *[Limits of multilevel TASEP and similar processes]{}*, To appear in Ann. Inst. H. Poincaré; arXiv:1206.3817 (2012).
J.M. Harrison and R.J. Williams, *[Multidimensional reflected Brownian motions having exponential stationary distributions]{}*, Ann. Probab. (1987), 115–137.
T. Ichiba and I. Karatzas, *[On collisions of Brownian particles]{}*, Ann. Appl. Prob. (2010), 951–977.
K. Johansson and E. Nordenstam, *[Eigenvalues of GUE minors]{}*, Electron. J. Probab. **11** (2006), 1342–1371.
I. Karatzas, S. Pal, and M. Shkolnikov, *[Systems of Brownian particles with asymmetric collisions]{}*, arXiv:1210.0259 (2012).
A. Metcalfe, *[Universality properties of Gelfand-Tsetlin patterns]{}*, Probab. Theory Relat. Fields (2011), online first.
E. Nordenstam, *[On the Shuffling Algorithm for Domino Tilings]{}*, Electron. J. Probab. **15** (2010), 75–95.
E. Nordenstam and B. Young, *[Domino Shuffling on Novak Half-Hexagons and Aztec Half-Diamonds]{}*, Electr. J. Comb. **18** (2011).
A. Okounkov and N. Reshetikhin, *[The Birth of a Random Matrix]{}*, Mosc. Math. J. **6** (2006), 553––566.
S. Pal and J. Pitman, *[One-dimensional Brownian particle systems with rank-dependent drifts]{}*, Ann. Appl. Prob. (2008), 2179–2207.
A.V. Skorokhod, *Stochastic equations for diffusions in a bounded region*, Theory Probab. Appl. (1961), 264–274.
S.R.S. Varadhan and R.J. Williams, *[Brownian motion in a wedge with oblique reflection]{}*, Comm. Pure Appl. Math. (1984), 405–443.
J. Warren, *[Dyson’s Brownian motions, intertwining and interlacing]{}*, Electron. J. Probab. **12** (2007), 573–590.
J. Warren and P. Windridge, *[Some Examples of Dynamics for Gelfand Tsetlin Patterns]{}*, Electron. J. Probab. **14** (2009), 1745–1769.
[^1]: Institute for Applied Mathematics, Bonn University, Endenicher Allee 60, 53115 Bonn, Germany. E-mail: [[email protected]]{}
[^2]: Institute for Applied Mathematics, Bonn University, Endenicher Allee 60, 53115 Bonn, Germany. E-mail: [[email protected]]{}
[^3]: For a set $S$, the notation $\frac{1}{2\pi{\mathrm{i}}} \oint_{\Gamma_S}{\mathrm{d}}w\,f(w)$ means that the integral is taken over any positively oriented simple contour that encloses only the poles of $f$ belonging to $S$.
[^4]: Here, standard Brownian motions start from $0$ and are normalized to have variance $t$ at time $t$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Field-theoretic models have been used extensively to study the phase behavior of inhomogeneous polymer melts and solutions, both in self-consistent mean-field calculations and in numerical simulations of the full theory capturing composition fluctuations. The models commonly used can be grouped into two categories, namely [*species*]{} models and [*exchange*]{} models. Species models involve integrations of functionals that explicitly depend on fields originating both from species density operators and their conjugate chemical potential fields. In contrast, exchange models retain only linear combinations of the chemical potential fields. In the two-component case, development of exchange models has been instrumental in enabling stable complex Langevin (CL) simulations of the full complex-valued theory. No comparable stable CL approach has yet been established for field theories of the species type. Here we introduce an extension of the exchange model to an arbitrary number of components, namely the multi-species exchange (MSE) model, which greatly expands the classes of soft material systems that can accessed by the complex Langevin simulation technique. We demonstrate the stability and accuracy of the MSE-CL sampling approach using numerical simulations of triblock and tetrablock terpolymer melts, and tetrablock quaterpolymer melts. This method should enable studies of a wide range of fluctuation phenomena in multiblock/multi-species polymer blends and composites.'
author:
- Dominik Düchs
- 'Kris T. Delaney'
- 'Glenn H. Fredrickson'
bibliography:
- 'mse\_refs.bib'
title: 'A multi-species exchange model for fully fluctuating polymer field theory simulations'
---
[^1]
[^2]
Introduction
============
Self-consistent field theory (SCFT) is a versatile and well-established numerical method for studying the mesoscopic self-assembly and equilibrium thermodynamic behavior of polymeric melts and solutions[@fredrickson06; @schmid98]. In contrast to microscopic simulation methods like molecular dynamics[@frenkel] or coarse-grained Monte Carlo[@landau00], partition functions containing particle-particle interactions are transformed mathematically such that the fundamental objects in SCFT are auxiliary chemical potential fields and their conjugate monomer species densities. While the field theories underpinning SCFT are formulated in continuous space, a range of numerical methods are available including projection of the fields onto a finite set of basis functions[@matsen94; @matsen03], real-space solutions via finite differences on a computational lattice[@fleer79], or pseudo-spectral solutions on a collocation grid invoking fast Fourier Transforms (FFTs)[@rasmussen02; @tzeremes02]. In all of these techniques the continuum SCFT equations are transformed into a large but finite set of non-linear equations that are solved iteratively by relaxation methods. The stationary solutions are discrete representations of the saddle-points of the field theory, and hence constitute *mean-field approximations* to the full theory. For each iteration of the field relaxation, a modified diffusion equation must be solved to deduce the statistical properties of non-interacting chains in a set of static external chemical-potential fields; this constitutes the vast majority of the computational cost of the method.
For inhomogeneous polymer systems, a frequent topic of study is the self-assembly of constituents into a variety of complex mesoscale morphologies (e.g., lamellae, bicontinuous networks, or hexagonally packed cylinders). SCFT allows phase diagrams to be constructed in the space of architectural parameters (e.g., the block molecular weights in copolymers) and segmental interaction parameters between the constituent species. Being a mean-field approximation, however, SCFT neglects fluctuations in composition profiles; this approximation is valid only for very dense systems, such as melts, in regions of the phase diagram that are sufficiently removed from critical phase transitions. To capture fluctuation effects, simulations of the full field theory are needed.
A major impediment to a straightforward implementation of simulations of the full theory is what is known in the quantum field-theory literature as the [*sign problem*]{}: the occurrence of rapidly oscillating complex exponential factors making the direct evaluation of statistical averages intractable. Until recently, strategies for handling the sign problem in polymer field theories had been developed and applied only for two-species systems. In one attempt the [*partial saddle-point method*]{} approximates the full theory by restricting fluctuations to fields without rapid oscillations[@duechs03; @stasiak/matsen:2013]. This approach is suitable mainly for two-species incompressible systems as it neglects fluctuations in the pressure mode governing total local densities, but it nevertheless constitutes an uncontrolled approximation. A more comprehensive solution, which captures fluctuations of all fields in the full theory, involves applying the complex Langevin (CL) method[@parisi83; @klauder83] for generating importance-sampled sequences of field configurations. This method tolerates the sign problem by analytically continuing the theory into the space of complex fields, and sampling stochastically along high-dimensional paths that are locally of near-constant phase. Estimates for physical properties (thermal averages) remain real valued. This method has previously been used to study a variety of problems including the fluctuation-induced shift in the order-disorder transition in symmetric[@ganesan01] and asymmetric[@lennon08_2] diblock copolymers, and the complexation of oppositely charged polyelectrolytes[@lee08].
Not surprisingly, a simulation method capturing the full statistical field theory faces several numerical challenges. As in SCFT, diffusion equations must be solved at each CL field iteration step. In contrast to SCFT however, where the purpose of iterating the fields is to approach a saddle-point solution, field updates in CL serve to sample representative field configurations with probability consistent with the partition function. An important efficiency consideration then is to ensure that field updates are performed in a manner that is both stable and accurate, and to propagate the system as quickly as possible between statistically uncorrelated field configurations. Lennon *et al*. have discussed the performance of several explicit and implicit field-update algorithms in selected two-component test cases[@lennon08]. Another numerical concern is the choice of a suitable algorithm to solve the diffusion equations in the presence of rough, stochastically generated fields. Recently Audus *et al*. demonstrated that higher-order diffusion solvers can be less stable than low-order solvers when subjected to fields containing white noise[@audus13]. While these considerations are important from a practical standpoint, a parallel concern is the stability of the CL equations in multi-component systems, which are of growing interest in the design and fabrication of novel nanostructured materials[@BatesScience2012]. In particular: *How can one construct a field-theoretic model and a corresponding CL stochastic dynamics for sampling the fields that is stable for systems containing more than two monomer species?* We address this challenge in the present article by generalizing the commonly used exchange model to polymer systems with more than two species, demonstrating for the first time that CL can be applied to polymeric fluids with many chemically distinct components; the method in principle extends to an arbitrary number of species.
A detailed discussion of field-theoretic models for polymeric melts and solutions can be found, e.g., in Ref. . We here briefly review the two broad classes of models, which are in principle thermodynamically equivalent. In density-explicit species-type models, a pair of thermodynamically conjugate complex fields arises *for each monomer/segment species*: an auxiliary chemical-potential field and a density field. In addition, incompressible models feature a “pressure” field conjugate to the total density at each point in space. Hence, for $S$ species, a typical density-explicit species model formally involves $2S+1$ independent fields. In contrast, *exchange-type models*, which prior to this work have been formulated only for $S=2$ component systems, decouple the interactions using the normal modes of the interaction matrix rather than the species fields themselves; the decoupling allows redundant degrees of freedom to be integrated out of the problem (provided the pair interaction is invertible). As a result, exchange-type models utilize only $S$ fields for $S$ components, and are therefore more efficient in their representation of a multi-component system. This is a first indication that exchange models are better suited to CL than species-type models, as imposing independent fluctuations on an over-complete set of fields can be imagined to produce numerical instabilities. Indeed, even for a simple two-species system, such as an incompressible AB diblock copolymer melt, one can show that the linearized CL equations for the species model possess eigenvalues of mixed sign, symptomatic of an unstable scheme. In contrast, the exchange formulation of the model has been shown to enable stable CL simulations of the same diblock copolymer melt model[@ganesan01; @lennon08; @lennon08_2]. We note that a third type of simulation method has been recently explored[@Koski/Huikuan/Riggleman:2013], utilizing an exchange mapping for *each pair* of species. The method was demonstrated to be stable for conducting simulations of a nanocomposite melt of linear AB polymer chains interacting with nanoparticles of a third distinct species, though the method requires $2S+1$ fields for $S$ species. We expect that the MSE approach presented here will be more efficient and stable as the number of distinct species is increased.
This paper is organized as follows. In Section II, we generalize the two-species exchange model to an arbitrary number of species for incompressible and weakly compressible melt models. In Section III we devise a complex Langevin (CL) sampling scheme for conducting field-theoretic simulations based on the new [*multi-species exchange (MSE)*]{} model, and discuss numerical aspects that affect efficiency. In section IV, we apply our model and methods to symmetric ABC triblock and ABCA$^\prime$ tetrablock terpolymer melts, and to ABDC tetrablock quaterpolymer melts; we show that CL sampling is stable both for disordered melts at low segregation strength and for ordered mesophases at intermediate segregation strength. We further demonstrate that the sampling is accurate by comparing structure factors in the disordered melt to those obtained by the random phase approximation, and through the recovery of the correct mean-field limit. Finally, we demonstrate that the compressible model fully recovers the incompressible one in the appropriate limit, while in some cases permitting more stable and efficient simulations.
Theory {#sec:theory}
======
The multi-species exchange model of strictly incompressible melts
-----------------------------------------------------------------
#### Coarse-grained particle model: {#coarse-grained-particle-model .unnumbered}
We begin with the following canonical partition function for a coarse-grained particle model of a multi-species polymer melt:
$$\mathcal{Z}_c = \frac{1}{\prod_{p=1}^{P}\lambda_T^{3 n_pN_p} n_p! }\prod_{i=1}^n\int\mathcal{D}{{\bf {r}}}_i e^{-\beta U_0 -\beta U_1}\delta\left[\sum_{j=1}^S\hat{\rho}_j{{\left({{\bf {r}}}\right)}}-\rho_0\right].
\label{eqn:ce_incmelt_particlepartfn}$$
In a volume $V$, there are a total of $n$ molecules with $P$ distinct molecule types. Each molecule type $p\in\left[1,P\right]$ has $n_p$ indistinguishable copies, such that $n=\sum_{p=1}^Pn_p$, and $N_p$ degrees of freedom. Assuming the molecules are all linear polymer chains (without loss of generality on the field-theory transformation), the degree of polymerization of each chain type is $N_p$, and we choose arbitrarily a reference $N$ such that $N_p = N\alpha_p$. The overall density of monomers in the melt is $\rho_0 = \sum_{p=1}^P n_p N_p/V$, and we assume that all statistical segments have a common reference volume $v_0 = 1/\rho_0$. $\lambda_T$ is the thermal wave length. A total of $S$ distinct chemical (“monomer”) species are distributed in arbitrary sequences in the chains. For a continuous-chain model, as employed throughout this work, the species density operators are $$\hat{\rho}_j{{\left({{\bf {r}}}\right)}}= \sum_{p=1}^{P} \sum_{i=1}^{n_p} \int_{s\in j}ds\, \delta\left({{\bf {r}}}-{{\bf {r}}}^p_i\left(s\right)\right),$$ where the integration bound $s\in j$ includes all statistical segments of chemical species $j$ in the chain-contour integral (i.e., repeated blocks, e.g., in linear ABA polymers, are handled implicitly by repeated summation). ${{\bf {r}}}_i^p\left(s\right)$ is the space-curve of polymer chain $i$ of type $p$, parameterized as a continuous linear filament with contour variable $s$. The functional integrals $\mathcal{D}{{\bf {r}}}_i$ run over the possible space curves ${{\bf {r}}}\left(s\right)$ for the $i$th polymer chain. The delta functional in Eqn. \[eqn:ce\_incmelt\_particlepartfn\] strictly enforces incompressibility in the melt by projecting onto a configuration subspace for which the microscopic densities sum to a uniform constant segment density ($\rho_0$) throughout space. The energy weights in Eqn. \[eqn:ce\_incmelt\_particlepartfn\] are $\beta U_0$, the conformational statistics of a single chain in free space (e.g., chain stretching term such as the Wiener weight for a continuous Gaussian chain), and $\beta U_1$, the potential energy for interactions between distinct segments. For linear continuous chains with Gaussian stretching statistics, the conformation term is $$\beta U_0 = \sum_{p=1}^{P}\sum_{i=1}^{n_p} \int_0^{N\alpha_p}ds\, \frac{3}{2b\left(s\right)^2}\left|\frac{d{{\bf {r}}}^p_i\left(s\right)}{ds}\right|^2,$$ where $b\left(s\right)$ is the statistical segment length of a monomer at contour position $s$. For Flory-like contact interactions, the potential energy term is $$\begin{aligned}
\beta U_1 & = & \frac{1}{2\rho_0 N}\sum_{i,j=1}^S{{\int d{{\bf {r}}}\,}}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}\chi_{ij}N\hat{\rho}_j{{\left({{\bf {r}}}\right)}}\,\\
& = & \frac{1}{2\rho_0 N}{{\int d{{\bf {r}}}\,}}{{\underline{\rho}}}^T{{\underline{\underline{\chi}}}}N{{\underline{\rho}}},\end{aligned}$$ where ${{\underline{\rho}}}^T = \left(\hat{\rho}_1,\ldots,\hat{\rho}_S\right)$ and ${{\underline{\underline{\chi}}}} = \left(\chi_{ij}\right)$ with $i,j\in\left\{1,\ldots,S\right\}$. In this model it is assumed that $\chi_{ii}=0 \enskip \forall i \in \left\{1,\ldots,S\right\}$ and $\chi_{ij} = \chi_{ji} \enskip \forall i,j \in \left\{1,\ldots,S\right\}$. $\beta U_1$ is responsible for the many-body character of the problem, and will be replaced by particle-field interactions using Hubbard-Stratonovich transformations. However, since the interaction matrix $\chi_{ij}N$ is usually not positive definite, direct Hubbard-Stratonovich transformation is usually not possible, and must instead be accompanied by decomposition into normal modes (“exchange mapping”).
#### Field Theory: {#field-theory .unnumbered}
Appendix \[sec:appendix\_incomp\] details the derivation of an exchange-mapped field theory corresponding to the canonical partition function Eqn. \[eqn:ce\_incmelt\_particlepartfn\]. The resulting partition function is $$\label{eqn:Z_incomp_mse_body}
\mathcal{Z}_c = \mathcal{Z}_0 {{\int\mathcal{D}}}\mu_1\ldots{{\int\mathcal{D}}}\mu_{S-1}{{\int\mathcal{D}}}\mu_+\, e^{-H\left[\left\{\mu_i\right\},\mu_+\right]}\\$$ with
$$\label{eqn:H_incomp_mse_body}
H\left[\left\{\mu_i\right\},\mu_+\right] = C\left[-\sum_{i=1}^{S-1}\frac{1}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}+ \sum_{i,j=1}^{S-1}\frac{O_{ji}\chi_{jS}N}{d_i} {{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}- {{\int d{{\bf {r}}}\,}}\mu_+{{\left({{\bf {r}}}\right)}}- \sum_{p=1}^P \frac{V\phi_p}{\alpha_p} \ln Q_p\left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right]\right],$$
and $\mathcal{Z}_0$ is a constant, including ideal gas terms, fully defined in Appendix \[sec:appendix\_incomp\]. Eqns. \[eqn:Z\_incomp\_mse\_body\] and \[eqn:H\_incomp\_mse\_body\] constitute a full description of the field theoretic canonical partition function. The fields $\mu_1,\ldots,\mu_{S-1},\mu_+$ are auxiliary *exchange*-mapped chemical potential fields that are related to the *species* chemical potential fields $\psi_A{{\left({{\bf {r}}}\right)}}N,\ldots,\psi_S{{\left({{\bf {r}}}\right)}}N$ by a linear transformation. $\mu_+$ is responsible for enforcing incompressibility of the melt. $d_i$ and $O_{ij}$ are respectively the eigenvalues and eigenvectors of the $\left(S-1\right)\times\left(S-1\right)$ matrix $X_{ij} = \chi_{ij}N-2\chi_{iS}N$. All lengths are non-dimensionalized using the radius of gyration of a reference ideal polymer chain, $R_g=b\left(N/6\right)^{1/2}$, with reference statistical segment length $b$ and reference polymerization degree $N$. $C=\rho_0 R_g^3/N$ is a dimensionless polymer chain number density parameter. $Q_p$ is the partition function of a single molecule of type $p$ experiencing the $S$ $\mu$ fields, while $\phi_p$ is the overall volume fraction of the system occupied by molecules of type $p$ (i.e., $\sum_p \phi_p = 1$). The functional integrals $\int\mathcal{D}\mu_i$ are over all configurations of real fields if $d_i<0$, and over purely imaginary fields otherwise (i.e., fields corresponding to repulsive interactions, which have saddle point configurations of physical relevance that are purely imaginary, have been Wick rotated to bring the saddle-point onto the real axis, as discussed in Appendix \[sec:appendix\_incomp\]). The $S\times S$ matrix ${{\underline{\underline{A}}}}$ specifies the linear transformation from exchange-mapped fields to the species chemical potential fields, i.e., $\psi_i{{\left({{\bf {r}}}\right)}}N = \sum_j A_{ij}\mu_j{{\left({{\bf {r}}}\right)}}$, which enter the single-chain partition function: $$A_{ij} =
\begin{pmatrix} O_{ij} & 1 \\ 0 & 1 \end{pmatrix}.
\label{A}$$ The inverse transformation ${{\underline{\underline{A}}}}^{-1}$ takes the form $$A^{-1}_{ij} = \left(\begin{array}{cc}
O_{ji} & X_i\\
0 & 1
\end{array}\right),$$ where $X_i = -\sum_{k=1}^{S-1}O_{ki}$. The need for the inverse transformation arises only during initialization of $\left\{\mu_i\right\}, \mu_+$ based on stored species-field configurations $\left\{\psi_i N\right\}$. It is usually more convenient to initialize calculations from species fields because the precise meaning of exchange fields can change significantly with system parameters. Once initialized, however, a CL or SCFT simulation will proceed without requiring the inverse transformation.
The single-chain partition functions $Q_p\left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right]$ are evaluated for continuous linear chains according to $$Q_p = \frac{1}{V} {{\int d{{\bf {r}}}\,}}q_p\left({{\bf {r}}}, \alpha_p; \left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right]\right),
\label{Qp}$$ where the propagators $q_p({{\bf {r}}}, s)$ are governed by the diffusion equation $$\begin{aligned}
\frac{\partial}{\partial s} q_p\left({{\bf {r}}},s\right) &=& \left[\frac{b\left(s\right)^2}{b^2}\nabla^2-N\psi_p\left({{\bf {r}}},s\right)\right] q_p\left({{\bf {r}}},s\right),\\
q_p\left({{\bf {r}}},0\right) &=& 1,
\label{diff_forward}\end{aligned}$$ which must be solved along the contour variable $s\in\left[0,\alpha_p\right]$. The external field $N\psi_p\left({{\bf {r}}}, s\right)$ is equal to $\psi_{A}{{\left({{\bf {r}}}\right)}}N$, $\psi_{B}{{\left({{\bf {r}}}\right)}}N$, etc., depending on the chemical identity of the segment at contour position $s$, where $\psi_i{{\left({{\bf {r}}}\right)}}N = \sum_j A_{ij} \mu_j{{\left({{\bf {r}}}\right)}}$ are the “species” chemical potential fields. The forward propagator, $q_p({{\bf {r}}},s)$, can be combined with a backward propagator, $q_p^{\dagger}({{\bf {r}}},s)$, governed by $$\begin{aligned}
\frac{\partial}{\partial s} q_p^{\dagger}\left({{\bf {r}}},s\right) &=& \left[\frac{b\left(s\right)^2}{b^2}\nabla^2-N\psi_p^\dagger\left({{\bf {r}}},s\right)\right] q_p^{\dagger}\left({{\bf {r}}}, s\right),\\
q_p^{\dagger}\left({{\bf {r}}},0\right) &=& 1,
\label{diff_backward}\end{aligned}$$ with $\psi_p^{\dagger}$ defined by traversing polymer $p$ in the contour parameter $s$ from the opposite chain end, to deduce dimensionless monomeric density operators, $\varphi_i\left({{\bf {r}}}\right)$, which are required for CL and mean-field relaxation dynamics. The density operators are $$\label{rho_i}
\varphi_j({{\bf {r}}}; [\{\mu_{i}\},\mu_{+}] ) = \sum_p \frac{\phi_p}{Q_p \alpha_p} \int_{s\in j} ds \, q_p({{\bf {r}}},s) q_p^{\dagger}({{\bf {r}}}, s),$$ where $s\in j$ denotes contour segments along chain $p$ occupied by monomers of species type $j$. These segment densities are normalized so that their expectation values over all field configurations weighted by $\exp (-H)$ yield the *local* volume fraction of type-$j$ segments, i.e., $\rho_0\left<\varphi_j{{\left({{\bf {r}}}\right)}}\right>_{\left\{\mu\right\}} = \left<\hat{\rho}_j\right>_{\left\{{{\bf {r}}}_i^p\right\}}$, where the former is a field-based operator and the latter is a microscopic density operator averaged over particle coordinates. It is straightforward to extend this definition of propagators and density operators to non-linear/branched polymer architectures, and to discrete-chain models.
Although not used in the present study, we note for completeness the form of the action in the grand canonical ensemble:
$$H_G\left[\left\{\mu_i\right\},\mu_+\right] = C\left[-\sum_{i=1}^{S-1}\frac{1}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}+ \sum_{i,j=1}^{S-1}\frac{O_{ji}\chi_{jS}N}{d_i} {{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}- {{\int d{{\bf {r}}}\,}}\mu_+{{\left({{\bf {r}}}\right)}}\right] - \sum_{p=1}^P z_p V \ln Q_p\left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right],$$
where $z_p$ is the activity of chain $p$, related to the chemical potential $\mu_p$ by $z_p = \exp\left(\beta \mu_p\right)$. For a strictly incompressible melt, the compositional redundancy manifests as an arbitrary scaling of chain activities: only ratios of $z_p$ affect the thermodynamics. All other symbols have the same definitions as in the canonical ensemble.
#### Complex Langevin Sampling: {#complex-langevin-sampling .unnumbered}
It is now appropriate to discuss the use of the [*MSE*]{} model for complex Langevin (CL) simulations of polymer systems containing an arbitrary number of chains $P$ and species $S$. Key to devising a stable CL scheme is writing Langevin equations not for the species potentials, but for the generalized eigenmode potentials $\left\{\mu_j\right\}, \mu_+$. In the canonical “diagonal” CL dynamics adopted here, the relaxational part of the field dynamics is in the direction of the generalized force acting on that same field. Including random force terms, the full stochastic CL equation for fields $\left\{\mu_j\right\}, \mu_+$ are: $$\label{eqn:CL}
\frac{\partial\mu_{i}\left({{\bf {r}}}, t\right)}{\partial t} = -\lambda_{i}\gamma_{i}^2\frac{\delta H\left[\left\{\mu_{i}\right\}\right]}{\delta \mu_{i}\left({{\bf {r}}},t\right)}
+ \gamma_{i} \eta_{i}\left({{\bf {r}}},t\right),$$ where $t$ is a fictitious simulation time variable, $\lambda_{i}>0$ is a real-valued relaxation rate parameter, and $i$ is an $S$-dimensional index spanning $\left[1,S-1\right]$ and $+$. The parameters $\gamma_{i}$ arise from the Wick rotation of the pressure and pressure-like fields, defined by $$\begin{aligned}
\gamma_{+} &=& i, \\
\gamma_{j} &=& \left\lbrace{\begin{array}{c}
1, \, \; d_{j} < 0 \\
i, \, \; d_{j} > 0
\end{array}} \right.,
\label{gamma}\end{aligned}$$ where $i=\sqrt{-1}$. Finally, the fields $\eta_{i}({\bf r}, t)$ are purely real and independent Gaussian white (in both space and time) noise terms obeying the fluctuation dissipation theorem: $$\begin{aligned}
\left<\eta_{i}\left({{\bf {r}}},t\right)\right> &=& 0 \nonumber \\
\left<\eta_{i}\left({{\bf {r}}},t\right)\eta_{i}\left({{\bf {r}}}^\prime,t^\prime\right) \right> &=& 2 \lambda_{i} \delta\left({{\bf {r}}}-{{\bf {r}}}^\prime\right)\delta\left(t-t^\prime\right).
\label{fluc_diss}\end{aligned}$$
The CL equations (Eqns. \[eqn:CL\]) involve thermodynamic forces that are expressed as functional derivatives of the action with respect to the generalized field variables. The relevant derivatives can be expressed in terms of the generalized potentials and the species density operators by the following expressions: $$\begin{aligned}
\frac{\delta H\left[\left\{\mu_i\right\},\mu_+\right]}{\delta \mu_{i} ({\bf r})} &=&
\frac{C}{d_i}\left[-\mu_{j}\left({{\bf {r}}}\right) + \sum_{j=1}^{S-1}O_{ji}\chi_{jS}N\right]\nonumber\\
& +& C\sum_{j=1}^{S-1} O_{ji}\varphi_{j}\left({{\bf {r}}}\right)\\
\frac{\delta H\left[\left\{\mu_j\right\},\mu_+\right]}{\delta \mu_{+} \left({{\bf {r}}}\right)} &=&
C\left[\sum_{j=1}^{S}\varphi_{j}({\bf r}) - 1\right]
\label{eqn:fplus}\end{aligned}$$
An interesting observation from Eqn. \[eqn:fplus\] is that the force on the $\mu_+$ field has no explicitly linear component. Furthermore, the spatial average of the sum of normalized species density operators is guaranteed to be unity for *any* configuration of the fields $\left\{\mu_i\right\},\mu_+$. Hence, the spatial average of the force on $\mu_+$ is zero for *any* $\mu_+$ configuration. This reflects an invariance of the species densities to a shift in the average value of $\mu_+$, which originates from the invariance of the exponential form of the incompressibility delta functional to such a shift. A practical consequence is that upon application of noise, the average value of the $\mu_+$ field will drift arbitrarily. However, all physical properties should be invariant to this drift. This invariance can be seen in $H$, because the linear term in $-n\ln Q_p$ (Sec. \[sec:wie\]) exactly cancels the explicit $\mu_+$ integral in $H$. We have nevertheless found it convenient to constrain ${{\int d{{\bf {r}}}\,}}\mu_+{{\left({{\bf {r}}}\right)}}=0$ during our CL and SCFT simulations, a choice with no thermodynamic consequences.
An interesting feature of the complex Langevin method is that instantaneous samples of thermodynamic operators are complex. However, as required on physical grounds, all observables become real upon thermal (time) averaging.
We note that eliminating the noise terms in Eqns. \[eqn:CL\] leads to deterministic relaxation equations that have mean-field saddle-point solutions as the long-time stationary state; i.e., all fields relax until $$0 = \left.\frac{\delta H\left[\left\{\mu_i\right\},\mu_+\right]}{\delta \mu_{i,+}{{\left({{\bf {r}}}\right)}}}\right|_{\left\{\mu_i\right\}=\left\{\mu_i^\star\right\}, \mu_+=\mu_+^\star}$$ is satisfied. One convenient aspect of SCFT is the immediate availability of the Helmholtz free energy $\beta F = -\ln \mathcal{Z}_c \approx -\ln \mathcal{Z}_0 + H^\star$, since under the saddle-point approximation $\mathcal{Z}_c \approx \mathcal{Z}_0 \exp\left(-H\left[\left\{\mu_i^\star\right\},\mu_+^\star\right]\right)$. This feature makes constructing SCFT phase diagrams straightforward for a predetermined set of candidate phases.
Finally, it is important to recognize that all saddle-point field configurations are independent of the $C$ chain number density in this model. This invariance means that the molecular weight of polymers in the melt enters only in unison with the interaction parameters through the $\chi N$ terms. In contrast, inclusion of composition fluctuations generates a family of models for different $C$ chain densities, where the $C$ parameter depends implicitly on the number of statistical segments present in the polymer chains. Hence, beyond-mean-field simulations acquire a molecular-weight dependence beyond $\chi N$ to all thermodynamic observables.
Weakly compressible multi-species melt
--------------------------------------
#### Coarse-grained particle model: {#coarse-grained-particle-model-1 .unnumbered}
In this case, the following particle model is used:
$$\mathcal{Z}_c = \frac{1}{\prod_{p=1}^{P} \lambda_T^{3 n_pN_p}n_p! }\int\ldots\int\prod_{i=1}^{n}\mathcal{D}{{\bf {r}}}_i \exp\left(-\beta U_0 -\beta U_1 - \beta U_2\right).
\label{eqn:ce_hcmelt_particlepartfn}$$
$\beta U_0$ and $\beta U_1$ are identical to the previous model, while the delta functional that enforced incompressibility has been replaced by a Helfand[@Helfand75] weak compressibility penalty $\beta U_2 = \frac{\zeta N}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\left(\sum_{j=1}^S\hat{\rho}_j{{\left({{\bf {r}}}\right)}}-\rho_0\right)^2$. The Helfand compressibility ensures that the thermally and spatially averaged density is $\rho_0$ (mass conservation), but local fluctuations around the spatial average density are permitted in the thermally averaged density profiles. This type of model can be numerically advantageous over the strictly incompressible model by making the complex Langevin dynamics less stiff, at the expense of introducing one extra parameter ($\zeta$).
#### Field Theory: {#field-theory-1 .unnumbered}
Appendix \[sec:appendix\_comp\] details the derivation of the exchange-mapped field theory for the compressible melt canonical partition function (Eqn. \[eqn:ce\_hcmelt\_particlepartfn\]). The resulting partition function is $$\label{eqn:Z_comp_mse_body}
\mathcal{Z}_c = \mathcal{Z}_0 {{\int\mathcal{D}}}\mu_1\ldots{{\int\mathcal{D}}}\mu_S \,e^{-H\left[\left\{\mu_i\right\}\right]},$$ with action
$$H\left[\left\{\mu_i\right\}\right] = C\left[-\sum_{i=1}^S \frac{\left(\zeta N\right)^{-\frac{1}{2}}}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}- \sum_{i,j=1}^S\frac{O_{ji}}{d_i\left(\zeta N\right)^{-\frac{1}{2}}}{{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}- V\sum_{p=1}^P \frac{\phi_p}{\alpha_p}\ln Q_p\left[{{\underline{\underline{O}}}}{{\underline{\mu}}}\right]\right],$$
$\mathcal{Z}_0$ a constant, and $d_i$ and $O_{ij}$ are eigenvalues and eigenvectors of the $S\times S$ matrix $\left({{\underline{\underline{\chi}}}}N\left(\zeta N\right)^{-\frac{1}{2}} + \left(\zeta N\right)^\frac{1}{2}{{\underline{\underline{\openone}}}}\right)$, where ${{\underline{\underline{\openone}}}}$ is an $S\times S$ matrix with all entries equal to $1$. Other symbols are as defined in the previous section. Note that all sums include a full $S$ count of exchange-mapped fields, and there is no separate pressure field. The transformation matrices from exchange to species fields and vice versa are directly ${{\underline{\underline{O}}}}$ and ${{\underline{\underline{O}}}}^T$ respectively. For reference, the action of the grand partition function in this model is
$$H_G\left[\left\{\mu_i\right\}\right] = C\left[-\sum_{i=1}^S \frac{\left(\zeta N\right)^{-\frac{1}{2}}}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}- \sum_{i,j=1}^S\frac{O_{ji}}{d_i\left(\zeta N\right)^{-\frac{1}{2}}}{{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}\right] - V\sum_{p=1}^P z_p Q_p\left[{{\underline{\underline{O}}}}{{\underline{\mu}}}\right].$$
In contrast with the incompressible case, the absolute magnitude of chain activities $\left\{z_p\right\}$ is not arbitrary; the total mass of the system is not specified by parameter constraints, and the absolute values of $z_p$ combine with the compressibility parameter to determine the equilibrium monomer density of the system. In this case, the quantity $\rho_0$ should be considered to be the inverse of a reference volume of a statistical segment.
#### Complex Langevin Sampling: {#complex-langevin-sampling-1 .unnumbered}
We use the following equation of motion for all fields in the Helfand-type MSE model: $$\partial_t \mu_i{{\left({{\bf {r}}}\right)}}= -\lambda_i\gamma_i^2\frac{\delta H}{\delta \mu_i{{\left({{\bf {r}}}\right)}}} + \gamma_i\eta_i
\label{eqn:comp_eom}$$ where the forces are given by $$\begin{aligned}
\frac{\delta H}{\delta \mu_i{{\left({{\bf {r}}}\right)}}} & = & C\left[-\frac{\left(\zeta N\right)^{-1/2}}{d_i}\mu_i{{\left({{\bf {r}}}\right)}}-\frac{1}{d_i\left(\zeta N\right)^{-\frac{1}{2}}} \sum_{j=1}^SO_{ji}\right. \nonumber\\
&+&\left.\sum_{j=1}^{S}O_{ji}\varphi_j\left({{\bf {r}}}\right)\right].
\label{eqn:comp_forces}\end{aligned}$$
One point of caution is that the spectrum of eigenvalues becomes very broad if $\zeta N$ becomes large (approaching the incompressible limit), so that the relaxation rates of different fields are strongly disparate. Efficient sampling can often be restored by adjusting the $\lambda_i$ parameters appropriately. An analysis of efficiency and mobility optimization is provided in Section \[sec:results\].
Numerical Methods {#sec:numerics}
=================
Our numerical methods are based on pseudospectral collocation with periodic boundary conditions and second-order operator splitting for contour-stepping the modified diffusion equations[@rasmussen02; @tzeremes02; @audus13; @lennon08], which are solved as an inner loop for each instantaneous configuration of the auxiliary fields $\left\{\mu\right\}$. Pseudospectral collocation with plane waves is ideally suited for complex Langevin simulations because it makes accessible large simulation cells (e.g., beyond the solution/melt correlation length) and does not rely on symmetric field configurations. Consequently, spatially decorrelated noise can be applied to the fields discretized on the computational lattice to affect unbiased stochastic sampling.
We find counter to intuition that solving the diffusion equation with a low-order (second) solver in the presence of rough fields makes for more stable CL trajectories[@audus13] when compared with nominally higher-order solvers. In all of the present work, unless otherwise specified, we use second-order operator splitting with a contour step size of $\Delta s = 0.01N$ for computing the chain propagators (Eqns. \[diff\_forward\], \[diff\_backward\]). It now remains to discuss the relevant discretizations for the CL equations of motion, which stochastically evolve field configurations in an outer loop.
The complex Langevin (CL) method[@fredrickson06; @lennon08; @ganesan01; @ghfreview02] allows sampling the full partition function (e.g., Eqns. \[eqn:Z\_incomp\_mse\_body\] and \[eqn:Z\_comp\_mse\_body\]) by moving along paths of approximate local constant phase, thus avoiding adverse efficiency loss from strongly oscillating terms. To numerically implement CL, the equations of motion must be discretized in space and time. All of the CL equations of motion provided in Sections \[sec:theory\]A and B have the generic continuum form $$\partial_t \mu\left({{\bf {r}}},t\right) = -\lambda\gamma^2\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)} + \gamma\eta\left({{\bf {r}}},t\right),$$ with Gaussian white noise statistics $\left<\eta\left({{\bf {r}}},t\right)\right>=0$, $\left<\eta\left({{\bf {r}}},t\right)\eta\left({{\bf {r}}}^\prime,t^\prime\right)\right> = 2\lambda\delta\left({{\bf {r}}}-{{\bf {r}}}^\prime\right)\delta\left(t-t^\prime\right)$. For spatial discretization, we first generate all functional derivatives in the continuum representation, then discretize the resulting functions onto the collocation mesh[^3]. The resulting equation of motion on the computational lattice takes the form $$\partial_t \mu^{{\bf {r}}}\left(t\right) = -\lambda\gamma^2\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]^{{\bf {r}}} + \gamma\eta^{{\bf {r}}}\left(t\right),$$ now with noise statistics $\left<\eta^{{\bf {r}}}\left(t\right)\right>=0$, $\left<\eta^{{\bf {r}}}\left(t\right)\eta^{{{\bf {r}}}^\prime}\left(t^\prime\right)\right> = 2\lambda\left(\Delta V\right)^{-1}\delta_{{{\bf {r}}},{{\bf {r}}}^\prime}\delta\left(t-t^\prime\right)$, where $\Delta V = V/M$, and $M$ is the number of cells on the spatial collocation mesh. Notice that the equations of motion are local in space, which makes the computational effort of solving the CL field-update equations insignificant compared to the inner loop consisting of diffusion equations (Eqns. \[diff\_forward\], \[diff\_backward\]).
Complex Langevin time integrators
---------------------------------
We now discuss a variety of time integration algorithms with a range of computational costs and accuracy, stability, and efficiency profiles. In practice, we have found that when used in tandem with appropriate semi-implicit time integration schemes, the discretized CL dynamics schemes presented in the previous section can be made both stable *and* accurate for arbitrary numbers of species and choices of interaction parameters in the weak and intermediate segregation regime. A number of the algorithms discussed here were presented in the context of two-species simulations by Lennon *et al.*[@lennon08], and all are compared in the context of multi-species simulations in Section \[sec:results\].
#### Euler-Maruyama: {#euler-maruyama .unnumbered}
The simplest discrete time stepper for evolving the complex Langevin equations is the Euler-Maruyama method, which is the stochastic analog of forward Euler stepping. The discretized equation of motion is $$\mu^{{{\bf {r}}},t+\Delta t} = \mu^{{{\bf {r}}},t}-\lambda\Delta t\gamma^2\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]^{{{\bf {r}}},t} + \gamma\eta^{{{\bf {r}}},t},$$ with $\left<\eta^{{{\bf {r}}},t}\right>=0$, $\left<\eta^{{{\bf {r}}},t}\eta^{{{\bf {r}}}^\prime,t^\prime}\right> = 2\lambda\Delta t\left(\Delta V\right)^{-1}\delta_{{{\bf {r}}},{{\bf {r}}}^\prime}\delta_{t,t^\prime}$, where $\Delta t$ is the time-step size. This algorithm is accurate only to first order in the relaxation term and, being an explicit scheme, suffers from poor stability at very large time steps. However, the small time step required to control inaccuracies from the time-step bias mean that stability is less of a concern than accuracy when using this algorithm in CL simulations. (SCFT, in contrast, is primarily limited by stability, because accurate trajectories are unimportant provided the method eventually converges to a stationary saddle-point solution).
#### 1S first-order semi-implicit: {#s-first-order-semi-implicit .unnumbered}
A stochastic extension to the SIS method introduced by Ceniceros and Fredrickson[@ceniceros04], the 1S semi-implicit method takes the linear part of the relaxation term to the future time: $$\begin{aligned}
\mu^{{{\bf {r}}},t+\Delta t} = \mu^{{{\bf {r}}},t} & -&\lambda\Delta t\gamma^2\left(\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]^{{{\bf {r}}},t} + \left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]_\mathrm{lin}^{{{\bf {r}}},t+\Delta t} \right.\nonumber\\
&-&\left.\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]_\mathrm{lin}^{{{\bf {r}}},t}\right) + \gamma\eta^{{\bf {r}}}\left(t\right), \end{aligned}$$ with $\left<\eta^{{{\bf {r}}},t}\right>=0$, $\left<\eta^{{{\bf {r}}},t}\eta^{{{\bf {r}}}^\prime,t^\prime}\right> = 2\lambda\Delta t\left(\Delta V\right)^{-1}\delta_{{{\bf {r}}},{{\bf {r}}}^\prime}\delta_{t,t^\prime}$. By using the linear response kernel for the homogeneous state, which is isotropic and translationally invariant (see Sections \[sec:numerics\]B and C), the linearized part of the force can be written in the form $$\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]_\mathrm{lin} \approx \left(\kappa\star\mu\right){{\left({{\bf {r}}}\right)}}$$ with Fourier transform $\hat{\kappa}{{\left({{\bf {k}}}\right)}}\hat{\mu}{{\left({{\bf {k}}}\right)}}$, where $\left\{{{\bf {k}}}\right\}$ are vectors on the reciprocal lattice, and $\hat{\kappa}{{\left({{\bf {k}}}\right)}}$ is a linear-response kernel that can be computed analytically for any polymer model. The 1S time stepper can then be rearranged to solve for $\mu^{t+\Delta t}$ in reciprocal space, yielding $$\hat{\mu}^{{{\bf {k}}},t+\Delta t} = \hat{\mu}^{{{\bf {k}}},t} - \frac{\Delta t \lambda \gamma^2}{1+\Delta t \lambda \gamma^2 \hat{\kappa}\left({{\bf {k}}}\right)}\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} + \gamma\hat{\eta}^{{{\bf {k}}},t}$$ The term preceding the Fourier transformed force is an effective ${{\bf {k}}}$-dependent time step that can be precomputed (and recomputed whenever the cell shape or volume changes). Note that only force terms for which $\gamma^2 \hat{\kappa}{{\left({{\bf {k}}}\right)}}>0$ for all ${{\bf {k}}}$ should be made semi-implicit in this fashion; other modes are destabilized by making this transformation. In addition to linearized forces, explicitly linear terms in the force can be treated implicitly in the same way. All forces in the MSE scheme, except for $\delta H/\delta \mu_+$ in the strictly incompressible case, have explicitly linear terms that can be treated implicitly.
#### Exponential time differencing: {#exponential-time-differencing .unnumbered}
Exponential time differencing was introduced in the context of complex Langevin simulations by Villet and Fredrickson (Refs. ). The method takes the linearized and explicitly linear parts of the relaxational force as an integrating factor over the interval $t\rightarrow t+\Delta t$, resulting in $$\begin{aligned}
\mu^{{{\bf {k}}},t+\Delta t} = \mu^{{{\bf {k}}},t} &-&\left(\frac{1-e^{-\lambda\Delta t\gamma^2\hat{\kappa}\left({{\bf {k}}}\right)}}{\hat{\kappa}\left({{\bf {k}}}\right)}\right)\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} \nonumber\\
&+&\left(\frac{1-e^{-\lambda\Delta t\gamma^2\hat{\kappa}\left({{\bf {k}}}\right)}}{2\lambda\Delta t\gamma^2\hat{\kappa}{{\left({{\bf {k}}}\right)}}}\right)\eta^{{{\bf {k}}},t},\end{aligned}$$
#### Weak second-order predictor-corrector methods: {#weak-second-order-predictor-corrector-methods .unnumbered}
One of the disadvantages of the algorithms presented in the previous sections is the reliance on linearization of the relaxational terms of the CL dynamics to improve both stability and accuracy over each time step. In practice, the linearization is performed with an isotropic, translationally invariant kernel that is accurate only for weak perturbations around a homogeneous state. In the limit of strongly segregated (inhomogeneous) polymer melts, the stability and accuracy of methods employing such kernels is less certain. One option for moving beyond this limitation is to employ predictor-corrector algorithms, which divide the propagation of fields into two steps and average the force between present and estimated future times for the corrected time step. No assumption about the nature or amplitude of inhomogeneity is required. These schemes offer weak second-order convergence to the continuous-time Langevin dynamics.
In the scheme due to Öttinger [@oettinger96] and Petersen[@petersen98] (PO), the predictor step consists of an explicit forward step which, taken alone, reduces to the Euler-Maruyama scheme. The corrector step then not only introduces a stabilizing implicit dependence but, in addition, adds and subtracts the approximately linearized and explicitly linear derivatives at the future ($t + \Delta t$) and the predictor ($\tilde t$) time steps, respectively. The predictor and corrector update steps are: $$\begin{aligned}
\label{PO}
\hat{\mu}^{{{\bf {k}}},\tilde t} = \hat{\mu}^{{{\bf {k}}},t} &-& \lambda\Delta t\gamma^2 \widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} + \gamma\hat{\eta}^{{{\bf {k}}},t} \\
\hat{\mu}^{{{\bf {k}}},t+\Delta t} = \hat{\mu}^{{{\bf {k}}},t} & -&\frac{\lambda\Delta t\gamma^2}{2}\left(\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} +\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},\tilde{t}} \right.\nonumber\\
&-&\left.\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}_\mathrm{lin}^{{{\bf {k}}},\tilde{t}} +\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}_\mathrm{lin}^{{{\bf {k}}},t+\Delta t}\right) \nonumber\\
&+& \gamma\hat{\eta}^{{{\bf {k}}},t},\end{aligned}$$ Note that this method requires two full sets of solutions of the chain propagator modified diffusion equations to complete a time step, so that the computational cost is doubled, but importantly the *same stochastic realization* of $\eta_{i}^{t}$ applies in both the predictor and corrector steps.
The PO method is significantly more accurate than the first order methods. If the linear response kernel is not employed in the corrector step, the method reduces to an Euler-Maruyama Predictor Corrector (EMPEC) method: $$\begin{aligned}
\label{eqn:EMPEC}
\hat{\mu}^{{{\bf {k}}},\tilde t} = \hat{\mu}^{{{\bf {k}}},t} &-& \lambda\Delta t\gamma^2 \widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} + \gamma\hat{\eta}^{{{\bf {k}}},t} \\
\hat{\mu}^{{{\bf {k}}},t+\Delta t} = \hat{\mu}^{{{\bf {k}}},t} & -&\frac{\lambda\Delta t\gamma^2}{2}\left(\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} +\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},\tilde{t}}\right)\nonumber\\
&+& \gamma\hat{\eta}^{{{\bf {k}}},t}, \end{aligned}$$ which performs similarly well in many cases, but does not rely on force linearization.
Finally, a third type of predictor-corrector algorithm, ETDPEC, was introduced by Villet and Fredrickson[@Villet2010]. ETDPEC uses exponential time differencing with force linearization for *both* the predictor and corrector steps:
$$\begin{aligned}
\mu^{{{\bf {k}}},\tilde t} & =& \mu^{{{\bf {k}}},t} -\left(\frac{1-e^{-\lambda\Delta t\gamma^2\hat{\kappa}\left({{\bf {k}}}\right)}}{\hat{\kappa}\left({{\bf {k}}}\right)}\right)\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} +\left(\frac{1-e^{-\lambda\Delta t\gamma^2\hat{\kappa}\left({{\bf {k}}}\right)}}{2\lambda\Delta t\gamma^2\hat{\kappa}{{\left({{\bf {k}}}\right)}}}\right)\eta^{{{\bf {k}}},t}\\
\mu^{{{\bf {k}}},t+\Delta t} &=& \mu^{{{\bf {k}}},t} -\frac{1}{2}\left(\frac{1-e^{-\lambda\Delta t\gamma^2\hat{\kappa}\left({{\bf {k}}}\right)}}{\hat{\kappa}\left({{\bf {k}}}\right)}\right)
\left(\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},t} + \widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu\left({{\bf {r}}},t\right)}\right]}^{{{\bf {k}}},\tilde t} \right)+\left(\frac{1-e^{-\lambda\Delta t\gamma^2\hat{\kappa}\left({{\bf {k}}}\right)}}{2\lambda\Delta t\gamma^2\hat{\kappa}{{\left({{\bf {k}}}\right)}}}\right)\eta^{{{\bf {k}}},t}.\end{aligned}$$
We now discuss the derivation of weak-inhomogeneity linear response kernels, $\hat{\kappa}\left({{\bf {k}}}\right)$, for linear multiblock polymer chains.
Weak inhomogeneity expansion {#sec:wie}
----------------------------
A weak inhomogeneity expansion assumes that local field deviations from a homogeneous saddle-point are weak, such that a power-series expansion of the action in a small perturbation parameter can be truncated beyond the quadratic term. The explicit field dependences of the action of both incompressible and weakly compressible melt models appear only to second order, but the $\ln Q_p$ term is a non-linear, non-local functional containing all powers of the fields. An expansion of $\ln Q_p$ to second order in the field perturbations yields the following weak inhomogeneity expansion for the MSE-mapped action of an incompressible melt of *linear* multiblock polymer chains: $$\begin{aligned}
H &\approx& H_{0} - \frac{C V}{2}\sum_{{{\bf {k}}}\ne 0} \left(\sum_{j=1}^{S-1}\frac{1}{d_j}
\hat{\mu}_j\left(-{{\bf {k}}}\right) \hat{\mu}_j\left({{\bf {k}}}\right)\right.
\nonumber \\
&+& \left.\sum_p \frac{\phi_p}{\alpha_p}
\sum_{ij=1}^{N_B}\sum_{lm=1}^S \hat{g}^p_{ij}{{\left({{\bf {k}}}\right)}}A_{S_il}\hat{\mu}_{l}\left(-{{\bf {k}}}\right)A_{S_jm}\hat{\mu}_{m}{{\left({{\bf {k}}}\right)}}\right),\nonumber\\\end{aligned}$$ where $H_0$ is evaluated for the homogeneous saddle-point configuration, and we have used periodic boundary conditions with the Fourier transform conventions $\hat{f}^{{\bf {k}}} = V^{-1}{{\int d{{\bf {r}}}\,}}f{{\left({{\bf {r}}}\right)}}\exp\left(-i{{\bf {k}}}.{{\bf {r}}}\right)$, $f{{\left({{\bf {r}}}\right)}}= \sum_{{\bf {k}}} \hat{f}^{{\bf {k}}}\exp\left(i{{\bf {k}}}.{{\bf {r}}}\right)$. $N_B$ is the number of distinct blocks along the copolymer backbone, and $S_i$ is a length-$N_B$ map vector that specifies the chemical species index for block $i$. $\phi_{p}$ is the volume fraction of the $p$th polymer chain type, and $\alpha_p$ is its chain length relative to $N$. $g^p_{ij}({\bf k})$ is a Debye pair correlation factor in reciprocal space applied between blocks $i$ and $j$ of uniform species index on polymer chain type $p$: $$\hat{g}^p_{ij}{{\left({{\bf {k}}}\right)}}= \left\lbrace{\begin{array}{lr}
\hat{g}\left(\left|{{\bf {k}}}\right|,f_i^p\right) , & i = j\\
\hat{h}\left(\left|{{\bf {k}}}\right|,f_i^p\right) \hat{h}\left(\left|{{\bf {k}}}\right|,f_j^p\right) \hat{l}\left(\left|{{\bf {k}}}\right|,d_{ij}^p\right), & i \ne j
\end{array}} \right.$$ with the scalar-valued functions $$\begin{aligned}
\hat{g}\left(k,f\right) &=& 2k^{-4} \left( e^{-fk^2} + fk^2 - 1 \right)\\
\hat{h}\left(k,f\right) &=& k^{-2}\left(1 - e^{-fk^2}\right) \\
\hat{l}\left(k,d\right) &=& e^{-d k^2}\end{aligned}$$ where $f^p_{i}$ is the normalized contour length (such that $\sum_i f^p_i = 1\, \forall\, p$ regardless of $\alpha_p$) of a sequential block of segments of species $i$ on chain $p$, and $d_{ij}$ is the sum of all normalized block lengths between, but not including, blocks $i$ and $j$ along the chain contour. Notice that $\hat{g}_{ij}^p\left({\bf k}\right)$ is only evaluated for block pairs $(i, j)$ belonging to the same polymer component $p$. Higher-order inter-chain terms do not appear, because those are handled by the explicit interaction terms in the action, while $\ln Q_p$ is the normalized partition function of non-interacting chains in external fields. Similar expressions can be derived for branched polymer chain architectures using perturbation theory[@fredrickson06].
For a weakly compressible melt, the weak inhomogeneity expansion is: $$\begin{aligned}
H &\approx& H_{0} - \frac{C V}{2}\sum_{{{\bf {k}}}\ne 0} \left(\sum_{j=1}^{S}\frac{\left(\zeta N\right)^{-\frac{1}{2}}}{d_j}
\hat{\mu}_j\left(-{{\bf {k}}}\right) \hat{\mu}_j\left({{\bf {k}}}\right)\right.
\nonumber \\
&+& \left.\sum_p \frac{\phi_p}{\alpha_p}
\sum_{ij=1}^{N_B}\sum_{lm=1}^S \hat{g}^p_{ij}{{\left({{\bf {k}}}\right)}}O_{S_il}\hat{\mu}_{l}\left(-{{\bf {k}}}\right)O_{S_jm}\hat{\mu}_{m}{{\left({{\bf {k}}}\right)}}\right),\nonumber\\
\label{eqn:wie_comp}\end{aligned}$$ with the same Debye correlation functions as in the incompressible model.
Linearized derivatives
----------------------
The linearized forces employed in Sec. \[sec:numerics\]A can be obtained by functional derivatives of the second-order weak-inhomogeneity expansions provided in the previous section. For the incompressible model: $$\begin{aligned}
\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu_{l}{{\left({{\bf {r}}}\right)}}}\right]}^{{\bf {k}}}_{lin} &=&
-\frac{C\hat{\mu}_l{{\left({{\bf {k}}}\right)}}}{d_{l}} \nonumber \\
& -& \sum_p \frac{C\phi_p}{\alpha_p} \sum_{ij=1}^{N_B}\sum_{m=1}^{S} \hat{g}^p_{ij}{{\left({{\bf {k}}}\right)}}A_{S_il} A_{S_jm} \hat{\mu}_m{{\left({{\bf {k}}}\right)}}\nonumber \\
\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu_{+}{{\left({{\bf {r}}}\right)}}}\right]}^{{\bf {k}}}_{lin} &=& - \sum_p \frac{C\phi_p}{\alpha_p} \sum_{ij}^{N_B}\sum_{m=1}^{S} \hat{g}^p_{ij}{{\left({{\bf {k}}}\right)}}A_{S_i+} A_{S_jm} \hat{\mu}_m{{\left({{\bf {k}}}\right)}}. \nonumber \\
\label{linderiv}\end{aligned}$$
For the purposes of stabilizing the time stepping algorithms presented earlier, it is advantageous to only consider those parts of the linearized derivatives that are proportional to the field whose derivative is calculated; contributions proportional to other fields are in general not stabilizing. Hence, $$\widehat{\left[\frac{\delta H\left[\left\{\mu\right\}\right]}{\delta \mu_{i}{{\left({{\bf {r}}}\right)}}}\right]}^{{\bf {k}}}_{lin} \approx \hat{\kappa}_{i}{{\left({{\bf {k}}}\right)}}\mu_{i}{{\left({{\bf {k}}}\right)}},
\label{linderiv_num}$$ where $\hat{\kappa}$ is a linear-response kernel. Moreover, instead of the full functions $\hat{\kappa}_{i}{{\left({{\bf {k}}}\right)}}$ we retain only those terms whose sign suggest that they are stabilizing for implicit time stepping, resulting in $$\begin{aligned}
\label{eqn:kappaj}
\hat{\kappa}_{j}{{\left({{\bf {k}}}\right)}}&=& -\frac{C}{d_{j}}\theta\left(-d_i\right) -\sum_{p}\frac{C\phi_p}{\alpha_p}\sum_{il=1}^S \hat{g}^p_{il}{{\left({{\bf {k}}}\right)}}A_{ij}A_{lm}\theta\left(A_{ij}A_{lm}d_{j}\right)\nonumber\\\\
\hat{\kappa}_{+}{{\left({{\bf {k}}}\right)}}&=& -\sum_p \frac{C\phi_p}{\alpha_p}\hat{g}\left(\left|{{\bf {k}}}\right|,\alpha_p\right)
\label{eqn:kappa+}\end{aligned}$$ with the Heaviside function $\theta\left(x\right)=1$ for $x>0$ and $=0$ otherwise. The $\theta$ terms are introduced so that only terms with the correct sign for stable implicit time stepping are included in $\kappa$. For the results presented in Section \[sec:results\], we use only $\hat{\kappa}_+$, together with all explicitly linear terms, for implicit time stepping. Further improvements in time stepping stability and accuracy might be achieved by also using Eqn. \[eqn:kappaj\]. Analogous expressions can be developed for the compressible model, though the model in general lacks a pure pressure-like field analogous to $\mu_+$. Our present strategy mirrors the non-pressure modes of the compressible model: we take only the stabilizing explicitly linear terms for implicit time stepping.
Results {#sec:results}
=======
Mean-field Theory
-----------------
As discussed in Section \[sec:theory\]A, the complex Langevin dynamics used for sampling the field theoretic partition function will return mean-field saddle-point solutions if the driving noise is eliminated. This approximation becomes asymptotically exact for $C\rightarrow\infty$, a limit for which the noise strength is naturally vanishing[@ghf87; @ganesan01; @lennon08] relative to $\delta H/\delta \mu$. Although the primary focus of the present work is the development of stable schemes for propagating CL dynamics, as a first test of the correctness of the MSE mapping we here explore mean-field simulations for linear ABC triblock melts and compare to known results.
#### Consistency of $ABC$ mean-field phase diagram: {#consistency-of-abc-mean-field-phase-diagram .unnumbered}
Throughout this work we consider only conformationally symmetric systems (that is, those for which the statistical segment lengths are all equal to the reference $b$); this restriction is easily lifted. The parameter space for incompressible melts of linear, monodisperse triblock polymers is then five dimensional: two block fractions ($f_A$, $f_B$, and $f_C=1-f_A-f_B$), and three $\chi N$ parameters ($\chi_{AB}N$, $\chi_{AC}N$, $\chi_{BC}N$). For a three-species incompressible melt, the reduced interaction matrix $X_{ij} = \chi_{ij}N-2\chi_{iS}N$ becomes a $2\times 2$ matrix. If we further restrict our initial tests to the case of $\chi N = \chi_{AB}N = \chi_{AC}N = \chi_{BC}N$, we find $$\begin{aligned}
{{\underline{\underline{X}}}} &= &\chi N\left(
\begin{array}{c c}
-2 & -1 \\
-1 & -2 \\
\end{array}
\right),\\
{{\underline{\underline{O}}}} & = &\frac{1}{\sqrt{2}}\left(
\begin{array}{c c}
-1 & -1 \\
-1 & 1 \\
\end{array}
\right),\\
d_1 & = & -3\chi N, \quad d_2 = -\chi N.\end{aligned}$$ These quantities fully describe the MSE mapping and the meaning of the exchange-mapped fields $\mu_1{{\left({{\bf {r}}}\right)}}, \mu_2{{\left({{\bf {r}}}\right)}}, \mu_+{{\left({{\bf {r}}}\right)}}$ for this specific system.
A partial SCFT phase diagram was first presented by Matsen[@matsen98b] for symmetric triblock polymers ($f_A = f_C$). As a first test of the correctness of the MSE mapping, Matsen’s ABC phase diagram is partially reproduced in Fig. \[fig:SCFT\_ABCPhases\]. Excellent agreement between the MSE-SCFT results and Matsen’s species-model SCFT phase diagram is observed.
[![Phase diagram for an incompressible symmetric ABC triblock melt with $\chi N = \chi_{AB}N = \chi_{BC}N = \chi_{AC}N$. Panel (a) shows the SCFT free energy, computed as $H$ of Eqn. \[eqn:H\_incomp\_mse\_body\] at the saddle-point field configuration ($\left\{\mu^\star\right\}$) for each phase under consideration, plotted against the A block fraction $f_A (=f_C)$ at $\chi N = 45$. Lines are cubic-spline interpolants. Phase transitions are shown as vertical grey lines. Panel (b) shows a reproduction of Matsen’s phase diagram[@matsen98b] with selected phase boundary points from the MSE-SCFT method denoted as red circles. The dashed line shows the cut at $\chi N = 45$ used to generate panel (a). []{data-label="fig:SCFT_ABCPhases"}](ABC_PhDia_2panel "fig:"){width="0.9\columnwidth"}]{}
#### Invariance of thermodynamic properties to eliminated species: {#invariance-of-thermodynamic-properties-to-eliminated-species .unnumbered}
A feature of the MSE model of an incompressible melt is the elimination of $\hat{\rho}_S$ by employing the delta functional subspace projection. An important test case then is to demonstrate that the thermodynamic properties computed from SCFT and CL simulations are invariant to the choice of species that is eliminated. For example, in an ABC triblock melt, the C ($=S$) species, being the highest indexed, is eliminated by default, but we are free to eliminate any other by permuting the species labels. This test, shown in Fig. \[fig:SCFT\_HvsV\], is made with SCFT calculations of the alternating gyroid phase of an ABC triblock melt, using the free-energy variation with lattice parameter and the stress tensor as example thermodynamic properties.
[![Helmholtz free energy per chain and stress tensor as a function of cubic lattice parameter for the alternating double-gyroid phase with space group $I4_132$, demonstrating the invariance of thermodynamic observables to the species that is eliminated in the incompressible MSE mapping. Calculations were conducted at $\chi N = 35$ and $f_A = f_C = 0.19$. The stress tensor is related to the derivative of the intensive free energy with respect to lattice parameters at *constant concentration*, and is used as the relevant operator in our variable-cell method[@barrat05]. The stress-free configuration has lattice parameter $6.53$$R_g$. Inset shows isosurfaces of $A$ and $C$ normalized density equal to $0.85$.[]{data-label="fig:SCFT_HvsV"}](GYR_Composite "fig:"){width="0.9\columnwidth"}]{}
#### Reduction to two-species exchange-mapped field theory: {#reduction-to-two-species-exchange-mapped-field-theory .unnumbered}
As a further test of the MSE mapping, we consider the special case of a system consisting of two species. In this case the interaction matrix $X_{ij}$ has only one entry $X_{11}=-2\chi_{AB}N$. The parameters that enter the action are therefore $d_1 = -2\chi_{AB}N$ and $O_{11} = 1$, and the action itself is $$\begin{aligned}
H &=& C\left[\frac{1}{4\chi_{AB}N}{{\int d{{\bf {r}}}\,}}\mu_1{{\left({{\bf {r}}}\right)}}- {{\int d{{\bf {r}}}\,}}\left(\mu_+{{\left({{\bf {r}}}\right)}}+ \frac{\mu_1{{\left({{\bf {r}}}\right)}}}{2}\right)\right.\nonumber\\
&-&\left.\sum_{p=1}^PV\frac{\phi_p}{\alpha_p}\ln Q_p\right].\end{aligned}$$ The species chemical potential fields that enter the $Q_p$ functionals are defined in terms of the exchange fields through the $A_{ij}$ matrix as $\psi_A{{\left({{\bf {r}}}\right)}}N = \mu_1{{\left({{\bf {r}}}\right)}}+\mu_+{{\left({{\bf {r}}}\right)}}$, and $\psi_B{{\left({{\bf {r}}}\right)}}N = \mu_+{{\left({{\bf {r}}}\right)}}$. The traditional two-species exchange-mapped field theory[@fredrickson06] has the relationships $\psi_A{{\left({{\bf {r}}}\right)}}N = \mu_+^{(2)}{{\left({{\bf {r}}}\right)}}+\mu_-^{(2)}{{\left({{\bf {r}}}\right)}}$, and $\psi_B{{\left({{\bf {r}}}\right)}}N = \mu_+^{(2)}{{\left({{\bf {r}}}\right)}}-\mu_-^{(2)}{{\left({{\bf {r}}}\right)}}$. We therefore expect to find consistent saddle-point field configurations between the two methods if we identify $\mu_+{{\left({{\bf {r}}}\right)}}+ \mu_1{{\left({{\bf {r}}}\right)}}/2 = \mu_+^{(2)}$ and $-\mu_1{{\left({{\bf {r}}}\right)}}/2 = \mu_-^{(2)}$. Fig. \[fig:2specreduction\] shows such an example for the saddle-point solutions of an AB diblock copolymer melt in the lamellar phase. The normalized species densities (upper panel) are immediately equal between the reduced MSE and traditional two-species exchange mapping (as are scalar thermodynamic quantities such as the free energy and components of the stress tensor), but the auxiliary exchange chemical potential fields (lower panel) require the mapping above to be brought into agreement. This test verifies that MSE correctly reduces to the traditional two-species exchange mapping in the appropriate limit, up to an arbitrary shift in the $\mu_+{{\left({{\bf {r}}}\right)}}$ field.
[![Density profiles (upper panel) and exchange field configurations (lower panel) for mean-field configurations of an incompressible AB diblock copolymer melt in the lamellar phase ($\chi N=25$, $f_A=0.4$) using both the traditional[@fredrickson06] two-species exchange mapping (dashed lines) and MSE (solid lines). The density profiles match between the two methods. The fields are in agreement with the appropriate mapping (see text) apart from an arbitrary shift of the pressure field, as discussed in Section \[sec:theory\]A, CL.[]{data-label="fig:2specreduction"}](Comparison_Stacked "fig:"){width="0.8\columnwidth"}]{}
#### Finite compressibility: {#finite-compressibility .unnumbered}
Here we demonstrate the effect on SCFT morphologies of permitting weak violation of the incompressibility condition by the introduction of a Helfand compressibility penalty. Figure \[fig:scftcompressible\] demonstrates the monomer density profiles in the saddle-point approximation of an ABC triblock polymer melt in the lamellar phase. The compressible model (black lines) has a non-uniform total density $\hat{\rho}_A{{\left({{\bf {r}}}\right)}}+ \hat{\rho}_B{{\left({{\bf {r}}}\right)}}+ \hat{\rho}_C{{\left({{\bf {r}}}\right)}}\neq \rho_0$, with strongest violations in the vicinity of interfaces. Fig. \[fig:scftcompressible\_incomplimit\] shows the continuous approach to the incompressible limit of both the free energy and total density profiles of the same ABC triblock melt. Relaxing the incompressibility constraint can be numerically advantageous in a number of situations, especially in cases for which the incompressible constraint is difficult to satisfy, such as in systems *constrained* to have non-uniform total density, e.g., due to boundary conditions.
[![Cross-sectional profiles of SCFT monomer densities for an incompressible melt (grey) and weakly compressible melt (black), with Helfand penalty $\zeta N=100$, of an ABC triblock polymer melt in the lamellar phase. The system parameters are $\chi_{AB}N = \chi_{BC}N = \chi_{AC}N = 30$ and $f_A = f_C =0.3$.[]{data-label="fig:scftcompressible"}](CompressibleLAM "fig:"){width="0.9\columnwidth"}]{}
[![The same system as in Fig. \[fig:scftcompressible\]. Upper panel: sections of total density profiles ($\rho_A{{\left({{\bf {r}}}\right)}}+ \rho_B{{\left({{\bf {r}}}\right)}}+ \rho_C{{\left({{\bf {r}}}\right)}}$) approach the incompressible limit ($=\rho_0$ for all ${{\bf {r}}}$) as the $\zeta N$ penalty is increased. Lower panel: the SCFT free energy per chain continuously approaches that generated by the fully incompressible model as $\zeta N$ is increased.[]{data-label="fig:scftcompressible_incomplimit"}](IncompLimitRecovery_SCFT "fig:"){width="0.9\columnwidth"}]{}
#### Efficiency of the saddle-point search in a compressible model: {#sec:scft_compressible_results .unnumbered}
For simplicity we make our analysis of time stepping efficiency for the same system as in Fig. \[fig:scftcompressible\]: an ABC linear triblock polymer melt in the lamellar phase with all binary $\chi N$ interactions equal. This specific choice makes the interaction matrix, ${{\underline{\underline{X}}}} = \left({{\underline{\underline{\chi N}}}}\left(\zeta N\right)^{-\frac{1}{2}} + \left(\zeta N\right)^{\frac{1}{2}}{{\underline{\underline{\openone}}}}\right)$ introduced in Appendix \[sec:appendix\_comp\], have one eigenmode equal to $\left(1,\ldots,1\right)$, mirroring the character of the pure pressure field in the incompressible model. Furthermore, the ${{\underline{\underline{X}}}}$ matrix has only two distinct eigenvalues: $d_i = -\chi N/\sqrt{\zeta N}$ for $i \in \left[1,S-1\right]$, and $d_S = \left(S\zeta N + \left(S-1\right)\chi N\right)/\sqrt{\zeta N}$ for the pressure-like mode. Consequently, $d_S$ diverges as the incompressible limit is approached, while all other $d_i \rightarrow 0$. This expansion of the eigenvalue spectrum leads to a critical slowing of the relaxation rate of the compressible model as the incompressible limit is approached.
Studying the form of Equations \[eqn:comp\_eom\] and \[eqn:comp\_forces\] indicates that scaling the mobility $\lambda_i \approx d_i$ might largely compensate for the spectrum broadening. However, this proposed startegy is oversimplified, because the density fields $\varphi_i{{\left({{\bf {r}}}\right)}}$, or equivalently the single-chain conformational term $\ln Q$, entangles all modes to all orders, as revealed even in the lowest-order expansion of the forces in Eqn. \[eqn:wie\_comp\]. Lacking a general strategy for efficient selection of mobility coefficients, we instead derive understanding by studying rates of convergence numerically. Figure \[fig:comp\_convergence\_map\] shows an analysis of a map of the number of mean-field relaxation steps required to converge to the lamellar saddle-point solution as a function of $\lambda_i$ mobility coefficients. Mirroring the separation of the eigenvalue spectrum, we explore scalar-valued mobility coefficients $\lambda_i \ne \lambda_S$ with $\Delta t=1$, and determine the stability envelope as well as the number of SCFT iterations required to reach the saddle point configuration.
[{width="25.00000%"} {width="74.00000%"}]{}
For the incompressible model, the fastest convergence to the saddle-point configuration (to a tolerance of less than $10^{-6}$ in the L2 norm of all forces) was achieved with $\lambda_i \Delta t=0.75$, with approximately $15,000$ iterations of the SIS field updater. For moderately larger $\lambda_i \Delta t$, the number of iterations is increased due to a damped oscillatory approach to the solution. Further increases in the mobility coefficient lead to non-damped oscillations or divergent trajectories. From the compressible data, the following observations can be made:
- Almost without exception, the pressure-like mode, $\mu_3$, required a mobility at least as large as the other modes for stable field trajectories.
- For weak compressibility penalty (more compact eigenvalue spectrum), optimal performance can be achieved with $\lambda_3 \approx \lambda_{1,2}$. In contrast, strong compressibility penalties (wide eigenvalue spectrum) favor a strong asymmetry in mobility coefficients. In that case, increasing $\lambda_i$ for $i\in\left[1,S-1\right]$ does not significant increase the rate of convergence; increasing $\lambda_S$ does.
- With an appropriate choice of $\left\{\lambda_i\right\}$, the inefficiency related to the sparse eigenvalue spectrum on the approach to the incompressible limit is significantly ameliorated, while a poor choice leads to either increased computational cost or instability.
- If $\lambda_i\Delta t=1.00$ is chosen, the $\zeta N=100$, $1,000$ and $10,000$ models required respectively $\sim 62,000$, $\sim 70,000$, and $\sim 510,000$ iterations of the SIS field updater. In contrast, tuning to the optimal mobility coefficients in the space of $\lambda_1 = \lambda_2 \neq \lambda_3$ results in a minimum iteration count of $\sim 2,500$ for $\zeta N=100$, $\sim 11,500$ for $\zeta N=1,000$, and $\sim 85,000$ for $\zeta N=10,000$, at least a factor $5$ improvement.
- Fig. \[fig:scftcompressible\_incomplimit\] shows that $\zeta N=1000$ differs from the incompressible limit by $\sim 0.05\%$ in the SCFT free energy and by a peak of $\sim 1\%$ in the spatial variation of the total segment density. The $\zeta N=1,000$ model required approximately 30% fewer time steps than the incompressible model in this test case.
Complex Langevin Simulations
----------------------------
#### Time-integration algorithms: {#time-integration-algorithms .unnumbered}
Figure \[fig:CLDT\] shows the variation of the thermal average of a trial operator, $\left<H\right>$, as the time integration algorithm and time-step size are both varied for a selection of multi species incompressible melts. Only simulations that did not yield divergent trajectories have data included. A crucial finding is that all methods are stable (no divergent trajectories observed) and accurate (time-step error controlled compared to statistical uncertainties) for sufficiently small values of $\Delta t$. However, working with small $\Delta t$ is very inefficient, because the correlation time, and therefore the number of discrete time steps required to move between statistically distinct configurations, is long. The number of statistically decorrelated configurations used in computing thermal averages is a crucial quantity for finding the standard error of the mean, and therefore determining stochastic uncertainty on predictons of observables. The standard error of the mean decreases as the square root of the number of decorrelated field configurations sampled. For Fig. \[fig:CLDT\], simulations of the incompressible model were ran long enough to yield standard errors smaller than or equal to the symbol size (i.e., smaller $\Delta t$ simulations completed more time steps). Stability and accuracy trends are similar for the compressible model.
[{width="\textwidth"}]{}
Fig. \[fig:CLDT\] shows that Euler-Maruyama and 1S methods, while quite stable at the $\Delta t$ range of interest, are least effective from the viewpoint of sampling efficiency due to a large time-step bias. Exponential time differencing shows significant improvements in accuracy, permitting approximately one order of magnitude improvement in efficiency. Each of these methods has essentially equal computational cost per CL time step, which is dominated by solving the modified diffusion equations for chain propagators. EMPEC, the simplest predictor-corrector method, improves accuracy further while doubling the runtime per CL time step. Finally, Petersen-Öttinger and ETDPEC methods demonstrate univerally excellent accuracy in all tests, with time step sizes being limited by stability rather than accuracy. Table \[tab:FUData\] summarizes a quantitative comparison of sampling efficiency for each time-step method. It is evident that predictor-corrector methods are crucial for efficient access to advanced applications of CL sampling.
Time Stepper $\Delta t_\mathrm{max}$[^4] $N_t^\mathrm{min}$[^5] $\left<H\right>$ $\tau_\mathrm{corr}\left(\Delta t_\mathrm{max}\right)$ runtime (sec)
------------------- -------------- ----------------------------- ------------------------ ------------------ -------------------------------------------------------- ---------------
EM 0.002 16500 265.72(3) 420 760
Linear ABC 1S 0.002 20000 256.45(3) 630 880
DIS phase ETD 0.020 2450 256.45(3) 85 108
$\chi N=10,14,10$ EMPEC 0.050 1250 256.41(3) 70 103
$C=30$ PO 0.100 500 256.59(3) 60 44
ETDPEC 0.100 500 256.52(3) 55 44
EM 0.005 180000 700.98(3) 260 12200
Linear ABC 1S 0.002 321500 700.41(3) 400 16600
GYR phase ETD 0.020 49000 700.42(3) 65 3580
$\chi N=30,30,30$ EMPEC 0.050 18000 700.61(3) 60 1465
$C=50$ PO 0.050 13600 700.63(3) 50 1320
ETDPEC 0.020 33850 700.64(3) 65 3260
Now that we have demonstrated controlled stability and accuracy of the stochastic sampling, we address the physical correctness of the CL-MSE method.
#### Structure factors: {#structure-factors .unnumbered}
One physical test of the CL-MSE simulations involves comparing the melt structure factors in the disordered phase to the RPA predictions of the same quantity. Since RPA structure factors contain fluctuations to Gaussian order, the numerical and analytical predictions should be in good agreement, provided the fluctuations are not sufficiently strong to engage higher-order terms. This will be the case for high molecular-weight melts with weak interactions far from an order-disorder transition.
The structure factor components are defined for translationally invariant systems as $$\begin{aligned}
S_{ij}{{\left({{\bf {k}}}\right)}}= \frac{1}{V}\int d{{\bf {r}}}\int d{{\bf {r}}}^\prime e^{-i{{\bf {k}}}.\left({{\bf {r}}}-{{\bf {r}}}^\prime\right)}\left<\delta\hat{\rho}_i{{\left({{\bf {r}}}\right)}}\delta\hat{\rho}_j\left({{\bf {r}}}^\prime\right)\right>\quad\end{aligned}$$ where $\delta \hat{\rho}_i{{\left({{\bf {r}}}\right)}}= \hat{\rho}_i{{\left({{\bf {r}}}\right)}}- \left<\hat{\rho}_i{{\left({{\bf {r}}}\right)}}\right>$, and $i, j\in\left[1,S\right]$ are species labels. Using our Fourier transform conventions detailed in Section \[sec:theory\]B, the structure factors become $$\begin{aligned}
S_{ij}{{\left({{\bf {k}}}\right)}}= V\left<\delta\hat{\rho}_i\left({{\bf {k}}}\right)\delta\hat{\rho}_j\left(-{{\bf {k}}}\right)\right>.\end{aligned}$$ Since the microscopic density operators $\hat{\rho}_i$ are integrated out of our field theory, we must use the methods described in Reference to derive an appropriate field-based structure-factor operator. We arrive at $$\begin{aligned}
\frac{S_{ij}{{\left({{\bf {k}}}\right)}}}{\rho_0 N} & = & CV\left<\sum_{k=1}^{S-1} \frac{A^{-1}_{ki}\hat{\mu}_k{{\left({{\bf {k}}}\right)}}}{d_k}\hat{\varphi}_j\left(-{{\bf {k}}}\right)\right>\nonumber\\
& - & CV\left<\sum_{k=1}^{S-1} \frac{A^{-1}_{ki}\hat{\mu}_k{{\left({{\bf {k}}}\right)}}}{d_k}\right>\Bigg<\hat{\varphi}_j\left(-{{\bf {k}}}\right)\Bigg>\end{aligned}$$ for incompressible systems, and $$\begin{aligned}
\frac{S_{ij}{{\left({{\bf {k}}}\right)}}}{\rho_0 N} & = & CV\left<\sum_{k=1}^{S} \frac{O_{ik}\hat{\mu}_k{{\left({{\bf {k}}}\right)}}}{\left(\zeta N\right)^{1/2}d_k}\hat{\varphi}_j\left(-{{\bf {k}}}\right)\right>\nonumber\\
& - & CV\left<\sum_{k=1}^{S} \frac{O_{ik}\hat{\mu}_k{{\left({{\bf {k}}}\right)}}}{\left(\zeta N\right)^{1/2}d_k}\right>\Bigg<\hat{\varphi}_j\left(-{{\bf {k}}}\right)\Bigg>\end{aligned}$$ for the compressible case, where $\varphi_j{{\left({{\bf {r}}}\right)}}$ is defined in Eqn. \[rho\_i\]. These expressions were derived using a single functional integration by parts, which yields the least noisy estimator out of those tested. Fig. \[fig:CLSK\] shows the structure factor for an ABC triblock melt with $f_A = f_C = 0.2$ and $\chi_{AB}N = \chi_{BC}N = 10$, $\chi_{AC}N=14$ for both a normalized chain number density of $C=100$, corresponding to very high molecular weight and weak composition fluctuations, and $C=20$, smaller molecular weight with stronger fluctuations. The disorder-phase structure factor contains a single peak resulting from short-length-scale correlations in the fluid. CL data are almost statistically indistinguishable from RPA curves for this system, which is sufficiently removed from the order-disorder transition.
[![Six structure factor components of an ABC symmetric triblock melt in the disordered phase with $f_B=0.6$, $\chi_{AB}N=\chi_{BC}N=10$, $\chi_{AC}N=14$. The simulation cell is a cube of side $8$$R_g$ with a spatial collocation mesh of $32\times 32\times 32$. High contour resolution is required to properly converge the high-$k$ decay of the structure factor; in this case $400$ contour samples were used with a second-order operator splitting pseudospectral algorithm[@audus13]. CL calculations were conducted at $C=100$ (weak fluctuations) and $C=20$ (moderately strong fluctuations). Increasing fluctuation strength weakly dampens the amplitude of the dominant scattering peak (inset). For comparison, the RPA structure factor is plotted (solid lines), which contains composition fluctuations on the Gaussian level[@fredrickson06].[]{data-label="fig:CLSK"}](ABC_SKvsRPA "fig:"){width="0.9\columnwidth"}]{}
#### Efficiency of weakly compressible CL simulations: {#efficiency-of-weakly-compressible-cl-simulations .unnumbered}
[![Complex Langevin simulations using the ETDPEC field updater for weakly compressible melts of ABC tiblock copolymers in the alternating gyroid phase, with $f_A = f_C = 0.2$ and all $\chi N = 30$. The upper panels show the error introduced in the estimate of $\left<H\right>$ as the mobility coefficients are varied, with reference value taken from the small $\lambda_i$ limit with the same compressibility. The lower panels show the relative efficiency, measured by $\epsilon_{\zeta N}^2\left(\left\{\lambda_i\right\}\right)/\epsilon^2_{\infty,\mathrm{ref}}$, where $\epsilon$ is the standard error of the mean of $\left<H\right>$ (here computed from $50,000$ CL time steps for each simulation). $\epsilon_{\infty,\mathrm{ref}}$ is a reference standard error of the mean taken from an incompressible simulation with optimal time step and $\lambda_i=1 \quad \forall i$. The efficiency measures the relative length of simulation (number of samples) required to achieve a stochastic uncertainty of the estimate of $\left<H\right>$ that is equal to the reference. Optimal choice of mobility coefficients can improve sampling efficiency by $\sim 10\times$–$100\times$.[]{data-label="fig:CLCompressible"}](Compressible_CL_dH "fig:"){width="\columnwidth"}]{} [![Complex Langevin simulations using the ETDPEC field updater for weakly compressible melts of ABC tiblock copolymers in the alternating gyroid phase, with $f_A = f_C = 0.2$ and all $\chi N = 30$. The upper panels show the error introduced in the estimate of $\left<H\right>$ as the mobility coefficients are varied, with reference value taken from the small $\lambda_i$ limit with the same compressibility. The lower panels show the relative efficiency, measured by $\epsilon_{\zeta N}^2\left(\left\{\lambda_i\right\}\right)/\epsilon^2_{\infty,\mathrm{ref}}$, where $\epsilon$ is the standard error of the mean of $\left<H\right>$ (here computed from $50,000$ CL time steps for each simulation). $\epsilon_{\infty,\mathrm{ref}}$ is a reference standard error of the mean taken from an incompressible simulation with optimal time step and $\lambda_i=1 \quad \forall i$. The efficiency measures the relative length of simulation (number of samples) required to achieve a stochastic uncertainty of the estimate of $\left<H\right>$ that is equal to the reference. Optimal choice of mobility coefficients can improve sampling efficiency by $\sim 10\times$–$100\times$.[]{data-label="fig:CLCompressible"}](Compressible_CL_eff "fig:"){width="\columnwidth"}]{}
Fig. \[fig:CLCompressible\] presents data collected from complex Langevin simulations of a weakly compressible ABC triblock melt in the alternating gyroid phase with $f_A = f_C = 0.2$ and all binary $\chi N = 30$. Simulation trajectories are more accurate for $\lambda_i \Delta t\rightarrow 0$, but the result is a slower time evolution and longer correlation time so that a significant increase in the number of time steps is required to accurately evaluate thermally averaged operators. The cost of simulations as mobility coefficients are varied can be evaluated using a relative efficiency metric, $\epsilon_{\zeta N}^2\left(\left\{\lambda_i\right\}\right)/\epsilon^2_{\infty,\mathrm{ref}}$, where $\epsilon$ is the standard error of the mean of $\left<H\right>$. $\epsilon_{\infty,\mathrm{ref}}$ is a reference standard error taken from an incompressible simulation with optimal time step ran for the same number of iterations. An efficiency greater than $1.0$ implies that the simulation can be ran for proportionally fewer time steps to achieve the same standard error as the reference. In a compressible model with all binary $\chi N$ equal, the MSE eigenvalue spectrum is split as detailed in Section \[sec:scft\_compressible\_results\]. Similarly grouping the mobility coefficients $\lambda_1 = \lambda_2 \neq \lambda_3$ in this case can significantly improve both simulation stability *and* efficiency, as shown in the lower panels of Fig. \[fig:CLCompressible\].
Conclusions
===========
We have introduced a multi-species-exchange field-theoretic model for multi-species, multiblock polymers and blends with arbitrary chain architectures. We have demonstrated that the method can be used as a basis for stable and accurate numerical simulations of the full field theory, including composition fluctuations, using the complex Langevin sampling scheme. Such fully fluctuating simulations were demonstrated to reproduce RPA structure factors for high-molecular-weight triblock copolymer melts deep in the disordered phase. In addition, mean-field (SCFT) solutions of the MSE model are readily available within the same framework and were shown to reproduce known phase boundaries in an ABC triblock terpolymer melt
We have detailed the efficiency and accuracy of a variety of field update algorithms for complex Langevin sampling, and explored the use of asymmetric mobility coefficients to recover stability and efficiency in weakly compressible melt models in both mean-field and non-mean-field simulations.
It is anticipated that the MSE method will be an important vehicle for the efficient study of fluctuation phenomena, including beyond-mean-field corrections to phase diagrams of multi-species melts. In addition, although a comparison has not been made in the present work, it is conceivable that the MSE approach may perform better than traditional density-explicit species methods for mean-field saddle-point searches, due in part to the relaxation along normal modes of the interaction matrix.
We note that the ultra-violet sensitivity[@Villet2012; @wang2010; @fredrickson06] present in many polymer field theories remains present in the simulations presented here. However, a systematic strategy is available to regularize the field theory, which is entirely compatible with the MSE scheme.[@Villet2012; @wang2010] We will explore this aspect in future work.
Acknowledgements
----------------
We acknowledge support from the Center for Scientific Computing at the CNSI and MRL: an NSF MRSEC (DMR-1121053) and NSF CNS-0960316. This work was partially supported by the CMMT program of the National Science Foundation under Award No. DMR-1160895. KTD was partially supported by the NSF DMREF program under Award DMR-1332842. DD was partially supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under DU 1100/1-1.
Exchange Mapping and Field Theory Transformation for Incompressible Multi-species Melt {#sec:appendix_incomp}
======================================================================================
We begin with the canonical partition function from Eqn. \[eqn:ce\_incmelt\_particlepartfn\]. Incompressibility can be enforced explicitly in the interaction term by replacing $\hat{\rho}_S = \rho_0 - \sum_{j=1}^{S-1} \hat{\rho}_j$. Defining ${{\underline{P}}}^T = \left(\hat{\rho}_1,\ldots,\hat{\rho}_{S-1}\right)$, ${{\underline{\underline{X}}}} = \left(\chi_{ij}N-2\chi_{iS}N\right)$, and ${{\underline{X}}}_S=\left(\chi_{iS}N\right)$ for $i,j\in\left[1,S-1\right]$, and using $\chi_{SS}=0$, we have $$\begin{aligned}
\beta U_1 & = & \frac{1}{2\rho_0N}\sum_{i,j=1}^{S-1}{{\int d{{\bf {r}}}\,}}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}\chi_{ij}N\hat{\rho}_j{{\left({{\bf {r}}}\right)}}\nonumber\\
&+& \frac{1}{\rho_0N}\sum_{i=1}^{S-1}{{\int d{{\bf {r}}}\,}}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}\chi_{iS}N\left(\rho_0-\sum_{j=1}^{S-1}\hat{\rho}_j{{\left({{\bf {r}}}\right)}}\right)\\
& = & \frac{1}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\left( {{\underline{P}}}^T{{\underline{\underline{X}}}}{{\underline{P}}} + 2 {{\underline{P}}}^T{{\underline{X}}}_S\rho_0\right)\\
& = & \frac{1}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\left({{\underline{P}}}+{{\underline{\underline{X}}}}^{-1}{{\underline{X}}}_S\rho_0\right)^T{{\underline{\underline{X}}}}\left({{\underline{P}}}+{{\underline{\underline{X}}}}^{-1}{{\underline{X}}}_S\rho_0\right) \nonumber\\
&-& \frac{1}{2\rho_0N}{{\int d{{\bf {r}}}\,}}{{\underline{X}}}_S^T{{\underline{\underline{X}}}}^{-1}{{\underline{X}}}_S\rho_0^2.
\label{eqn:intterm_incompmelt_nondiag}\end{aligned}$$ where the latter expression is obtained by completing the square.
We now diagonalize the ${{\underline{\underline{X}}}}$ matrix using a similarity transform: ${{\underline{\underline{X}}}}={{\underline{\underline{O}}}}{{\underline{\underline{D}}}}{{\underline{\underline{O}}}}^T$, where ${{\underline{\underline{O}}}}$ is an orthogonal matrix with columns equal to the eigenvectors of ${{\underline{\underline{X}}}}$. ${{\underline{\underline{D}}}}$ is a diagonal matrix containing the eigenvalues of ${{\underline{\underline{X}}}}$, $\left\{d_i\right\}$. Without loss of generality, we here assume that ${{\underline{\underline{X}}}}$ has no zero eigenvalues, i.e., all $d_{i}$ are nonzero. The presence of any $d_{i}=0$ indicates a singular interaction matrix, pointing to redundancies in the way the species are labeled. For example, a model with three species A, B, and C for which $\chi_{AB}=\chi_{AC}$ and $\chi_{BC}=0$ should be rewritten as a two-species model because B and C are indistinguishable. All zero-eigenvalue modes can thus be deleted and the remaining modes renumbered. Inserting the similarity transform into Eqn. \[eqn:intterm\_incompmelt\_nondiag\], we find $$\begin{aligned}
\beta U_1 & = & \frac{1}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\sum_{i=1}^{S-1}d_i\left[{{\underline{\underline{O}}}}^T\left({{\underline{P}}}+\rho_0{{\underline{\underline{X}}}}^{-1}{{\underline{X}}}_S\right)\right]_i^2 \nonumber\\
&-& \frac{\rho_0}{2N}{{\int d{{\bf {r}}}\,}}\sum_{i=1}^{S-1}\left[{{\underline{\underline{O}}}}^T{{\underline{X}}}_S\right]_i^2 d_i^{-1},
\label{eqn:incomp_uint_diagonalized}\end{aligned}$$ where we have made use of the fact that ${{\underline{\underline{O}}}}^{-1} = {{\underline{\underline{O}}}}^T$, and $\left[\ldots\right]_i$ selects the $i$th component of a vector. We now split Eqn. \[eqn:incomp\_uint\_diagonalized\] into two parts: $\beta U_1 = \beta\bar{U}_1 + \beta\bar{U}_\mathrm{ref}$, where $\beta\bar{U}_\mathrm{ref}$ depends only on the number of species and their mutual $\chi$ interactions. With convenient non-dimensionalizations: $$\beta\bar{U}_\mathrm{ref} = -\frac{CV}{2}\sum_{i=1}^{S-1}\left(\sum_{j=1}^{S-1} O_{ji}\chi_{jS}N\right)^2d_i^{-1},$$ which depends only on the $\left\{ \chi N\right\}$ parameters together with the dimensionless chain density, $C$, and the cell volume, $V$. This constant term, which can be precomputed, enters SCFT and CL calculations only as a free-energy shift; it does not depend on $n_p$ for any of the molecular constituents, and consequently does not affect the chemical potential references. The remaining part of the interaction term can be written as $$\beta \bar{U}_1 = \frac{1}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\sum_{i=1}^{S-1} d_i \sigma_i^2{{\left({{\bf {r}}}\right)}},
\label{eqn:quadraticform}$$ where $\sigma_i{{\left({{\bf {r}}}\right)}}= \left[{{\underline{\underline{O}}}}^T\left({{\underline{P}}}{{\left({{\bf {r}}}\right)}}+\rho_0{{\underline{\underline{X}}}}^{-1}{{\underline{X}}}_S\right)\right]_i$ are linear combinations of the original microscopic density operators, $\hat{\rho}_i{{\left({{\bf {r}}}\right)}}$, with ${{\underline{\underline{\chi}}}} N$-dependent constant shifts originating from completing the square. The purely quadratic form of Eqn. \[eqn:quadraticform\] permits a direct decoupling of interactions through Hubbard-Stratonovich transformation. For this task we make use of the following Gaussian functional integral (GFI) identity (Eqns. C.27 & C.28 of Ref. ):
$$\exp\left(\frac{\gamma^2}{2}{{\int d{{\bf {r}}}\,}}\int d{{\bf {r}}}^\prime J{{\left({{\bf {r}}}\right)}}A^{-1}\left({{\bf {r}}},{{\bf {r}}}^\prime\right)J\left({{\bf {r}}}^\prime\right)\right) = \frac{\int \mathcal{D}f \exp\left(-\frac{1}{2}{{\int d{{\bf {r}}}\,}}\int d{{\bf {r}}}^\prime f{{\left({{\bf {r}}}\right)}}A\left({{\bf {r}}},{{\bf {r}}}^\prime\right)f\left({{\bf {r}}}^\prime\right) - \gamma{{\int d{{\bf {r}}}\,}}J{{\left({{\bf {r}}}\right)}}f{{\left({{\bf {r}}}\right)}}\right)}{\int \mathcal{D}f \exp\left(-\frac{1}{2}{{\int d{{\bf {r}}}\,}}\int d{{\bf {r}}}^\prime f{{\left({{\bf {r}}}\right)}}A\left({{\bf {r}}},{{\bf {r}}}^\prime\right)f\left({{\bf {r}}}^\prime\right)\right)},
\label{eqn:HSDefinition}$$
where $\gamma = \pm i$ or $\pm 1$. In this definition of the Hubbard-Stratonovich transformation, ${{\int\mathcal{D}}}f$ is a functional integral over real-valued configurations of the field $f{{\left({{\bf {r}}}\right)}}$. Let $A^{-1}\left({{\bf {r}}},{{\bf {r}}}^\prime\right) = \frac{\left|d_i\right|}{\rho_0N}\delta\left({{\bf {r}}}-{{\bf {r}}}^\prime\right)$, then $$\begin{aligned}
e^{-\beta\bar{U}_1} & = & \prod_{i=1}^{S-1} \exp\left(\frac{\left|d_i\right|\gamma_i^2}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\sigma_i^2{{\left({{\bf {r}}}\right)}}\right)\\
&=& \prod_{i=1}^{S-1}\left(D_i^{-1} \int \mathcal{D}w_i \exp\left(-\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}- \gamma_i{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\sigma_i{{\left({{\bf {r}}}\right)}}\right)\right),\end{aligned}$$ where we have used $\gamma_i = 1$ for $d_i<0$ and $\gamma_i=i$ otherwise. $w_i{{\left({{\bf {r}}}\right)}}$ is an auxiliary chemical potential field that is conjugate to $\sigma_i{{\left({{\bf {r}}}\right)}}$ – the exchange-mapped linear combination of species microscopic density operators. $D_i$ is a Gaussian functional integral denominator that contributes to the overall measure of the partition function: $$D_i = \int \mathcal{D}w_i \exp\left(-\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}\right).$$ Hence, $$\mathcal{Z}_c = \mathcal{Z}_\mathrm{ig} \left(\prod_{k=1}^{S-1}D^{-1}_k\right)\int\prod_{i=1}^n\mathcal{D}{{\bf {r}}}_i\int\prod_{j=1}^{S-1}\mathcal{D}w_j \exp\left[-\beta U_0 - \sum_i\left(\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}+ \gamma_i {{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\sigma_i{{\left({{\bf {r}}}\right)}}\right)\right]\delta\left[\hat{\rho}_+{{\left({{\bf {r}}}\right)}}-\rho_0\right],$$ where $\hat{\rho}_+{{\left({{\bf {r}}}\right)}}= \sum_i^S\hat{\rho}_i{{\left({{\bf {r}}}\right)}}$ and $\mathcal{Z}_\mathrm{ig} = \exp\left(-\beta \bar{U}_\mathrm{ref}\right)\left(\prod_{p=1}^P\lambda_T^{3n_pN_p}n_p!\right)^{-1}$.
We now transform the functional delta to exponential form[@fredrickson06]. This leads to $$\begin{aligned}
\mathcal{Z}_c = \mathcal{Z}_\mathrm{ig}^\prime \int\prod_{i=1}^n\mathcal{D}{{\bf {r}}}_i\,e^{-\beta U_0}\int\prod_{j=1}^{S-1}\mathcal{D}w_j\int \mathcal{D}w_+ &\exp&\left[-\sum_{i=1}^{S-1}\left(\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}+\gamma_i {{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\sigma_i{{\left({{\bf {r}}}\right)}}\right)\right.\nonumber\\
&-&\left.i{{\int d{{\bf {r}}}\,}}w_+{{\left({{\bf {r}}}\right)}}\left(\hat{\rho}_+{{\left({{\bf {r}}}\right)}}-\rho_0\right)\right],\end{aligned}$$ with $$\mathcal{Z}_\mathrm{ig}^\prime = \frac{1}{\prod_{p=1}^P\lambda_T^{3n_pN_p} n_p!}e^{-\beta\bar{U}_\mathrm{ref}}\left(\int\prod_{i=1}^{S-1}\mathcal{D}w_i\,e^{-\sum_{i=1}^{S-1}\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}}\right)^{-1}\left(\int \mathcal{D}w\, e^{-\frac{1}{2}{{\int d{{\bf {r}}}\,}}w^2{{\left({{\bf {r}}}\right)}}}\right)^{-2}.$$ Upon re-inserting the full expression for $\sigma_i$, we find $$\begin{aligned}
\mathcal{Z}_c &=& \mathcal{Z}_\mathrm{ig}^\prime \int\prod_{j=1}^{S-1}\mathcal{D}w_j{{\int\mathcal{D}}}w_+\, e^{-\sum_{i=1}^{S-1}\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}+ i{{\int d{{\bf {r}}}\,}}w_+{{\left({{\bf {r}}}\right)}}\rho_0}\nonumber\\
&\times& \int\prod_{i=1}^n\mathcal{D}{{\bf {r}}}_i\,e^{-\beta U_0 - \sum_{i,j=1}^{S-1}\gamma_i O_{ji}{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\left(\hat{\rho}_j{{\left({{\bf {r}}}\right)}}+\rho_0\sum_{k=1}^{S-1} X_{jk}^{-1}\chi_{kS}N\right) - i{{\int d{{\bf {r}}}\,}}w_+{{\left({{\bf {r}}}\right)}}\sum_{i=1}^{S}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}}\\
\Rightarrow \mathcal{Z}_c &=& \mathcal{Z}_\mathrm{ig}^\prime \int\prod_{j=1}^{S-1}\mathcal{D}w_j{{\int\mathcal{D}}}w_+\, e^{-\sum_{i=1}^{S-1}\frac{\rho_0N}{2\left|d_i\right|}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}- \rho_0\sum_{i,j=1}^{S-1} \gamma_i O_{ji}\sum_{k=1}^{S-1}X_{jk}^{-1}\chi_{kS}N{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}+ i{{\int d{{\bf {r}}}\,}}w_+{{\left({{\bf {r}}}\right)}}\rho_0}\nonumber\\
&\times& \int\prod_{i=1}^n\mathcal{D}{{\bf {r}}}_i\,e^{-\beta U_0 - \sum_{i,j=1}^{S-1}\gamma_i O_{ji}{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\hat{\rho}_j{{\left({{\bf {r}}}\right)}}- i{{\int d{{\bf {r}}}\,}}w_+{{\left({{\bf {r}}}\right)}}\sum_{i=1}^{S}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}}.
\label{eqn:ZcInt}\end{aligned}$$
The final part of Eqn. \[eqn:ZcInt\] is a partition function of non-interacting chains subject to external fields $\left\{w_i{{\left({{\bf {r}}}\right)}}\right\}$. Note that $\psi_j{{\left({{\bf {r}}}\right)}}:= \sum_i O_{ji}\gamma_i w_i{{\left({{\bf {r}}}\right)}}+iw_+{{\left({{\bf {r}}}\right)}}$ is conjugate to $\hat{\rho}_j{{\left({{\bf {r}}}\right)}}$, and as such $\psi_j$ can be defined as “species” chemical potential fields.
We now make the following transformation to the exchange-mapped auxiliary fields $\left\{w_{i,+}\right\}$: $\mu_i{{\left({{\bf {r}}}\right)}}= \gamma_i w_i{{\left({{\bf {r}}}\right)}}N$, $\mu_+{{\left({{\bf {r}}}\right)}}= iw_+{{\left({{\bf {r}}}\right)}}N$. This mapping includes a Wick rotation for pressure-like fields – those with $d_i>0$ and $w_+$. We do this in all terms in $\mathcal{Z}_c$, including the Gaussian functional integral denominators in $\mathcal{Z}_\mathrm{ig}^\prime$, so that the implied Jacobians cancel regardless of discretization strategy. Note that Wick rotations of the subset of the fields with pressure character changes their integration domain so that $\int\mathcal{D}\mu$ is over purely imaginary fields for that subset. With this mapping, the partition function becomes
$$\begin{aligned}
\mathcal{Z}_c &=& \mathcal{Z}_{ig}^\prime \int\prod_{j=1}^{S-1}\mathcal{D}\mu_j{{\int\mathcal{D}}}\mu_+\, e^{\sum_{i=1}^{S-1}\frac{C}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}- C\sum_{i,j=1}^{S-1} O_{ji}\sum_{k=1}^{S-1}X_{jk}^{-1}\chi_{kS}N{{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}+ C{{\int d{{\bf {r}}}\,}}\mu_+{{\left({{\bf {r}}}\right)}}}\nonumber\\
&\times& \int\prod_{i=1}^n\mathcal{D}{{\bf {r}}}_i\,e^{-\beta U_0 - \sum_{i,j=1}^{S} A_{ij}{{\int d{{\bf {r}}}\,}}\mu_j{{\left({{\bf {r}}}\right)}}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}/N},
\label{eqn:beforeparticleintegrate}\end{aligned}$$
where $C=\rho_0R_g^3/N$ is a dimensionless polymer chain number density, and we have non-dimensionalized all lengths with respect to the radius of gyration of a reference ideal polymer chain, $R_g = b\left(N/6\right)^{1/2}$ with a reference statistical segment length $b$ and degree of polymerization $N$. We have made use of the fact that $\left|d_i\right|\gamma_i^2 = -d_i \enskip \forall i$. The factor $1/N$ in the latter terms can be absorbed by rescaling the contour variable to $\left[0,\alpha_p\right]$ in $\hat{\rho}$. We have defined the matrix $A_{ij}$ that transforms from exchange fields, $\left(\left\{\mu_i{{\left({{\bf {r}}}\right)}}\right\}, \mu_+{{\left({{\bf {r}}}\right)}}\right)$, to the species fields $\left(\psi_A{{\left({{\bf {r}}}\right)}}N, \ldots, \psi_S{{\left({{\bf {r}}}\right)}}N\right)$ that are directly conjugate to $\hat{\rho}_i{{\left({{\bf {r}}}\right)}}$, $i\in\left[1,S\right]$: $$A_{ij} = \left(
\begin{array}{cc}
O_{ij} & e_i\\
0 & 1
\end{array}
\right),$$ where ${{\underline{e}}}$ is a column vector containing ones with $S-1$ entries. Note that $\mu_+{{\left({{\bf {r}}}\right)}}= \psi_S{{\left({{\bf {r}}}\right)}}N$. Due to the orthogonal nature of ${{\underline{\underline{O}}}}$, the inverse of $A$ is readily identified to be $$A^{-1}_{ij} = \left(\begin{array}{cc}
O_{ji} & X_i\\
0 & 1
\end{array}\right),$$ where $X_i$ is a (column) vector $X_i = -\sum_{k=1}^{S-1}O_{ki}$ with $S-1$ entries. Though $A^{-1}_{ij}$ is never needed during the course of a field theoretic simulation, in which the fundamental degrees of freedom are the fields $\left\{\mu_i\right\}$, $\mu_+$ entering the functional integrals, having this transformation can be useful for seeding a calculation using the species fields. Initialization with species fields is preferred, because then field definitions are independent of the exchange-mapping strategy (e.g., elimination of a species other than $S$ when applying the incompressibility constraint).
Isolating the single-chain part of Eqn. \[eqn:beforeparticleintegrate\], we have
$$\int \prod_{p=1}^P\prod_{l=1}^{n_p} \mathcal{D}{{\bf {r}}}_l^p \exp\left(-\sum_{p=1}^{P}\sum_{l=1}^{n_p}\int_0^{\alpha_p}ds\,\frac{3N}{2b\left(s\right)^2}\left|\frac{d{{\bf {r}}}_l^p\left(s\right)}{ds}\right|-\sum_p^P\sum_l^{n_p}\sum_i^S\int_{s\in i}ds\,N\psi_i\left({{\bf {r}}}_l^p\left(s\right)\right)\right),$$
where $i$ indexes chemical species, $p$ indexes chain types, and $l$ indexes members of each type. The contour variable $s$ has been rescaled by $N$. In this expression, the microscopic density operator has been substituted and the $\delta$ functions over particle coordinates evaluated, ${{\underline{\psi}}}={{\underline{\underline{A}}}}{{\underline{\mu}}}$ Evaluating $\sum_i$ and expanding the $\exp\left(\ldots\right)$ leads to
$$\prod_{p=1}^P\prod_{l=1}^{n_p}\left[\int \mathcal{D}{{\bf {r}}}_l^p \exp\left(-\int_0^{\alpha_p}ds\,\frac{3N}{2b\left(s\right)^2}\left|\frac{d{{\bf {r}}}_l^p\left(s\right)}{ds}\right|-\int_0^{\alpha_p}ds\,N\psi_p\left({{\bf {r}}}_l^p\left(s\right);s\right)\right)\right]:=\prod_{p=1}^P\left(V g_{N_p} Q_p\left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right]\right)^{n_p}$$
where $Vg_{N_p}$ is the reference single-chain partition function of an ideal polymer in free space of volume $V$ (Ref. ), and $\psi_p\left({{\bf {r}}}\left(s\right);s\right):=\psi_i\left({{\bf {r}}}\left(s\right)\right)$ for $s$ in the $i$ block (i.e., $\psi_p$ has a parametric dependence on $s$, switching between the different species fields depending on the statistical segment on which it is acting).
Now use the following relation: $\sum_j O_{ji}X^{-1}_{jk} = {{\underline{\underline{O}}}}^T{{\underline{\underline{X}}}}^{-1} = \left({{\underline{\underline{X}}}}{{\underline{\underline{O}}}}\right)^{-1} = {{\underline{\underline{D}}}}^{-1}{{\underline{\underline{O}}}}^T$ to give $C\sum_{i,j,k=1}^{S-1} O_{ji}X^{-1}_{jk}\chi_{kS} = C\sum_{i,j=1}^{S-1}\left(d_i\right)^{-1}O_{ji}\chi_{jS}N$. Then the partition function becomes $$\begin{aligned}
\mathcal{Z}_c &=& \mathcal{Z}_0 {{\int\mathcal{D}}}\mu_1\ldots{{\int\mathcal{D}}}\mu_{S-1}{{\int\mathcal{D}}}\mu_+\, e^{-H\left[\left\{\mu_i\right\},\mu_+\right]}\\
\label{eqn:H_incomp_mse}
H\left[\left\{\mu_i\right\},\mu_+\right] &=& -\sum_{i=1}^{S-1}\frac{C}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}+ \sum_{i,j=1}^{S-1}\frac{C}{d_i} O_{ji}\chi_{jS}N{{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}- C{{\int d{{\bf {r}}}\,}}\mu_+{{\left({{\bf {r}}}\right)}}- \sum_{p=1}^P n_p \ln Q_p\left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right],\end{aligned}$$
where $\mathcal{Z}_0 = \mathcal{Z}_{ig}^\prime V^n \prod_{p=1}^P g_{N_p}^{n_p}$. (i.e., $\mathcal{Z}_0$ now contains all terms required for an ideal gas of polymers chains, together with the Gaussian denominators introduced during Hubbard-Stratonovich transformations). It is convenient to always work with volume fractions instead of explicit particle/molecule numbers. The $n_p$ of the final term can be replaced as follows: $$\begin{aligned}
\rho_0 &=& \sum_p \frac{n_pN_p}{V} = CN \\
\Rightarrow CV &= &\sum_p n_p \alpha_p\\
\phi_p &:= & \frac{n_p N_p}{\sum_q n_q N_q} \equiv \frac{n_p \alpha_p}{\sum_q n_q \alpha_q}\\
\Rightarrow n_p & = & \frac{CV\phi_p}{\alpha_p},\end{aligned}$$ where $\phi_p$ is the chain volume fraction (i.e., the fraction of the total count of monomers originating on chains of type $p$), and $\alpha_p$ is the chain length relative to the reference $N$. Similar derivations for branched polymers result in the $\alpha_p$ denominator being replaced by a sum of sub-chain lengths relative to $N$.
The normalized single-chain partition function for a linear chain in a collection of external fields can be written $$Q_p\left[{{\underline{\underline{A}}}}{{\underline{\mu}}}\right] = \frac{1}{V}{{\int d{{\bf {r}}}\,}}q_p\left({{\bf {r}}},\alpha_p\right),$$ where $q_p$ is computed through $$\partial_s q\left({{\bf {r}}},s\right) = \frac{b\left(s\right)^2}{b^2} \nabla^2 q\left({{\bf {r}}},s\right) - N\psi_p\left({{\bf {r}}};s\right)q\left({{\bf {r}}},s\right),$$ with $s\in\left[0,\alpha_p\right]$, and $q\left({{\bf {r}}},0\right) = 1$ is the initial condition for untethered, unconstrained polymer chains. All lengths are scaled to units of $R_g = b\left(N/6\right)^{1/2}$.
Exchange Mapping and Field Theory Transformation for Compressible Multi-species Melt {#sec:appendix_comp}
====================================================================================
In the case of a weakly compressible melt, we do not eliminate one of the species in favor of the total polymer density — all spatially inhomogeneous microscopic density operators must be retained. The strategy here involves completing the square in the exponent, diagonalizing the resulting coupling matrix, and making a Hubbard-Stratonovich transformation of all normal-mode density operators. We will use the symbol ${{\underline{e}}}=\left(1,\ldots,1\right)^T$, a rank-one vector with $S$ elements, and an $S\times S$ matrix with all entries equal $1$. Consider $\beta U_\mathrm{int} = \beta U_1 + \beta U_2$ from Eqn. \[eqn:ce\_hcmelt\_particlepartfn\]:
$$\begin{aligned}
\beta U_\mathrm{int} & = & \frac{1}{2\rho_0}{{\int d{{\bf {r}}}\,}}\left[{{\underline{\rho}}}^T{{\left({{\bf {r}}}\right)}}{{\underline{\underline{\chi}}}}{{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+ \zeta\left({{\underline{\rho}}}^T{{\left({{\bf {r}}}\right)}}{{\underline{e}}}-\rho_0\right)^2\right]\\
& = & \frac{1}{2\rho_0}{{\int d{{\bf {r}}}\,}}\left[{{\underline{\rho}}}^T{{\left({{\bf {r}}}\right)}}{{\underline{\underline{\chi}}}}{{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+ \zeta{{\underline{\rho}}}^T{{\left({{\bf {r}}}\right)}}{{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}- 2\zeta\rho_0{{\underline{\rho}}}^T{{\left({{\bf {r}}}\right)}}{{\underline{e}}}+\zeta\rho_0^2\right]\\
& = & \frac{1}{2\rho_0 N}{{\int d{{\bf {r}}}\,}}\left[\left({{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+{{\underline{\nu}}}\right)^T\left({{\underline{\underline{\chi}}}}N+\zeta N\right)\left({{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+ {{\underline{\nu}}}\right) - {{\underline{\nu}}}^T\left({{\underline{\underline{\chi}}}}N + \zeta N\right){{\underline{\nu}}}+\zeta N\rho_0^2\right]\end{aligned}$$
where the last step comes from completing the square, which demands that ${{\underline{\nu}}} = -\rho_0\zeta N\left({{\underline{\underline{\chi}}}}N+\zeta N\right)^{-1}{{\underline{e}}}$. The symbols ${{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}$, ${{\underline{\underline{\chi}}}}$ and $\rho_0$ are as defined in Appendix \[sec:appendix\_incomp\]. One option to proceed from here is to directly diagonalize the matrix $\left({{\underline{\underline{\chi}}}}N+\zeta N\right)$ to decompose $\exp\left(-\beta U_\mathrm{int}\right)$ into normal modes. However, the large difference in the common magnitude of $\left\{\chi_{ij} N\right\}$ versus that of $\zeta N$ can cause numerical problems. The matrix becomes singular for $\zeta N$ large (approaching the incompressible limit) which makes the inversion involved in ${{\underline{\nu}}}$ problematic. Likewise, extracting a factor of $\zeta N$ from the matrix results in $\chi N / \zeta N$, which approaches zero (numerical under-run) when a strong compressibility penalty is imposed, also resulting numerically in a singular matrix. For typical magnitudes of the matrix entries, extracting a factor of $\left(\zeta N\right)^{-\frac{1}{2}}$ tends to make the eigenvalues close to order $1$. Then $$\begin{aligned}
\beta U_\mathrm{int} &=& \frac{\left(\zeta N\right)^\frac{1}{2}}{2\rho_0 N}{{\int d{{\bf {r}}}\,}}\left[\left({{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+{{\underline{\nu}}}\right)^T{{\underline{\underline{X}}}}\left({{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+ {{\underline{\nu}}}\right) - {{\underline{\nu}}}^T{{\underline{\underline{X}}}}{{\underline{\nu}}}\right]\nonumber\\
&+&\frac{\rho_0V\zeta N}{2N},\end{aligned}$$ with ${{\underline{\underline{X}}}} = \left({{\underline{\underline{\chi}}}}N\left(\zeta N\right)^{-\frac{1}{2}} + \left(\zeta N\right)^\frac{1}{2}{{\underline{\underline{\openone}}}}\right)$, and ${{\underline{\nu}}} = -\rho_0 \left(\zeta N\right)^\frac{1}{2}{{\underline{\underline{X}}}}^{-1}{{\underline{e}}}$. We now diagonalize the matrix using ${{\underline{\underline{X}}}} ={{\underline{\underline{O}}}}{{\underline{\underline{D}}}}{{\underline{\underline{O}}}}^T$. $$\begin{aligned}
\beta U_\mathrm{int} & = & \frac{\left(\zeta N\right)^\frac{1}{2}}{2\rho_0 N}{{\int d{{\bf {r}}}\,}}\left[\left({{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+{{\underline{\nu}}}\right)^T{{\underline{\underline{O}}}}{{\underline{\underline{D}}}}{{\underline{\underline{O}}}}^T\left({{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}+ {{\underline{\nu}}}\right) \right.\nonumber\\
&-& \left. {{\underline{\nu}}}^T{{\underline{\underline{X}}}}{{\underline{\nu}}}+\rho_0^2\left(\zeta N\right)^\frac{1}{2}\right]\\
& = & \frac{\left(\zeta N\right)^\frac{1}{2}}{2\rho_0 N}{{\int d{{\bf {r}}}\,}}\left[\sum_{i=1}^S d_i \sigma_i^2{{\left({{\bf {r}}}\right)}}+ K\right]\end{aligned}$$ where $\sigma_i{{\left({{\bf {r}}}\right)}}= \sum_{j=1}^S O_{ji}\left(\hat{\rho}_j{{\left({{\bf {r}}}\right)}}+ \nu_j\right)$ and $K = \rho_0^2\left(\zeta N\right)^\frac{1}{2} - \rho_0^2 \zeta N\sum_{i,j}^SX^{-1}_{ij}$, where the last term comes from ${{\underline{\nu}}}^T{{\underline{\underline{X}}}}{{\underline{\nu}}} = \rho_0^2 {{\underline{e}}}^T\left(\zeta N\right)^\frac{1}{2}\left({{\underline{\underline{X}}}}^{-1}\right)^T{{\underline{\underline{X}}}}{{\underline{\underline{X}}}}^{-1}\left(\zeta N\right)^\frac{1}{2}{{\underline{e}}} = \rho_0^2\zeta N\sum_{i,j}X^{-1}_{ij}$. Now we split the interaction energy into a constant reference shift and a remainder that depends on the system configuration. $$\begin{aligned}
\beta U_\mathrm{ref}&=&\frac{CV}{2\left(\zeta N\right)^{-1}}\left(1-\frac{1}{\left(\zeta N\right)^{-\frac{1}{2}}}\sum_{i,j=1}^SX_{ij}^{-1}\right)\\
&=&\frac{CV}{2\left(\zeta N\right)^{-1}}\left(1-\frac{1}{\left(\zeta N\right)^{-\frac{1}{2}}}\sum_{i=1}^S d_i^{-1}\left(\sum_{j=1}^S O_{ji}\right)^2\right).\nonumber\\\end{aligned}$$ The Hubbard-Stratonovich transformations introduced in Equation \[eqn:HSDefinition\] can be used with the positive-definite interaction potential $A_i^{-1} \left({{\bf {r}}},{{\bf {r}}}^\prime\right) = \frac{\left|d_i\right|\left(\zeta N\right)^\frac{1}{2}}{\rho_0N}\delta\left({{\bf {r}}}-{{\bf {r}}}^\prime\right)$. $-d_i = \left|d_i\right|\gamma_i^2$ with $\gamma_i = \pm i$ if $d_i>0$, $\gamma_i = \pm 1$ otherwise.
$$\begin{aligned}
e^{-\beta U_\mathrm{int}} & = & e^{-\beta U_\mathrm{ref}}\prod_{i=1}^S\left[\exp\left(\gamma_i^2\frac{\left|d_i\right|\left(\zeta N\right)^\frac{1}{2}}{2\rho_0N}{{\int d{{\bf {r}}}\,}}\sigma_i^2{{\left({{\bf {r}}}\right)}}\right)\right]\\
& = & e^{-\beta U_\mathrm{ref}}\prod_{i=1}^S\left[D_i^{-1}\int \mathcal{D}w_i \exp\left(-\frac{\rho_0N}{2\left|d_i\right|\left(\zeta N\right)^\frac{1}{2}}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}-\gamma_i{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\sigma_i{{\left({{\bf {r}}}\right)}}\right)\right]\end{aligned}$$
The Gaussian functional integral denominator is $$D_i = \int \mathcal{D} w \exp\left(-\frac{C}{2\left|d_i\right|\left(\zeta N\right)^\frac{1}{2}}{{\int d{{\bf {r}}}\,}}N^2 w^2{{\left({{\bf {r}}}\right)}}\right)$$ Now substituting ${{\underline{\sigma}}}{{\left({{\bf {r}}}\right)}}$ and collecting terms that depend on particle coordinates: $$\begin{aligned}
e^{-\beta U_\mathrm{int}} & = & e^{-\beta U_\mathrm{ref}}{{\int\mathcal{D}}}\left\{w_i\right\} \exp\left(-\sum_i^S\frac{\rho_0 N}{2\left|d_i\right|\left(\zeta N\right)^\frac{1}{2}}{{\int d{{\bf {r}}}\,}}w_i^2{{\left({{\bf {r}}}\right)}}-\sum_i^S\gamma_i{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\left[{{\underline{\underline{O}}}}^T{{\underline{\nu}}}\right]_i\right)\nonumber\\
&\times& \exp\left(-\sum_{i=1}^S\gamma_i{{\int d{{\bf {r}}}\,}}w_i{{\left({{\bf {r}}}\right)}}\left[{{\underline{\underline{O}}}}^T{{\underline{\rho}}}{{\left({{\bf {r}}}\right)}}\right]\right)\prod_{i=1}^SD_i^{-1},\\\end{aligned}$$
Since ${{\underline{\underline{O}}}}$ is orthogonal, ${{\underline{\underline{X}}}} = {{\underline{\underline{O}}}}{{\underline{\underline{D}}}}{{\underline{\underline{O}}}}^T$, and ${{\underline{\nu}}} = -\rho_0\left(\zeta N\right)^\frac{1}{2}{{\underline{\underline{X}}}}^{-1}{{\underline{e}}}$, then ${{\underline{\underline{O}}}}^T{{\underline{\nu}}} = -\rho_0\left(\zeta N\right)^\frac{1}{2}\left({{\underline{\underline{O}}}}{{\underline{\underline{D}}}}\right)^{-1}{{\underline{e}}}$ and $\left[\left({{\underline{\underline{O}}}}{{\underline{\underline{D}}}}\right){{\underline{e}}}\right]_i=\sum_j d_i^{-1} O_{ji}$. We now make the replacement $\gamma_iNw_i{{\left({{\bf {r}}}\right)}}= \mu_i{{\left({{\bf {r}}}\right)}}$, which includes a Wick-rotation for repulsive modes and a factor to scale out the reference polymer chain length.
$$\begin{aligned}
e^{-\beta U_\mathrm{int}} & = & e^{-\beta U_\mathrm{ref}}{{\int\mathcal{D}}}\left\{\mu_i\right\} \exp\left(\sum_i^S\frac{C\left(\zeta N\right)^{-\frac{1}{2}}}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}+\frac{C}{\left(\zeta N\right)^{-\frac{1}{2}}}\sum_{i,j=1}^S{{\int d{{\bf {r}}}\,}}\frac{O_{ji}\mu_i{{\left({{\bf {r}}}\right)}}}{d_i}\right)\nonumber\\
&\times& \exp\left(-\sum_{i=1}^S{{\int d{{\bf {r}}}\,}}\mu_i{{\left({{\bf {r}}}\right)}}\sum_{j=1}^S O_{ji}\hat{\rho}_j{{\left({{\bf {r}}}\right)}}/N\right)\prod_{i=1}^SD_i^{-1},\end{aligned}$$
The decoupled interaction potential is inserted into the partition function such that $$\begin{aligned}
\mathcal{Z}_c & = & \mathcal{Z}_\mathrm{ig} e^{-\beta U_\mathrm{ref}}{{\int\mathcal{D}}}\left\{\mu_i\right\}\,
\exp\left(\sum_{i=1}^S\frac{C\left(\zeta N\right)^{-\frac{1}{2}}}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}+\frac{C}{\left(\zeta N\right)^{-\frac{1}{2}}}\sum_{i,j=1}^S{{\int d{{\bf {r}}}\,}}\frac{O_{ji}\mu_i{{\left({{\bf {r}}}\right)}}}{d_i}\right)\nonumber\\
&\times & {{\int\mathcal{D}}}\left\{{{\bf {r}}}_i\right\} \exp\left(-\beta U_0 - \sum_{i=1}^S{{\int d{{\bf {r}}}\,}}\psi_i{{\left({{\bf {r}}}\right)}}\hat{\rho}_i{{\left({{\bf {r}}}\right)}}/N\right)\end{aligned}$$ where $\psi_i{{\left({{\bf {r}}}\right)}}N := \sum_j O_{ij} \mu_j{{\left({{\bf {r}}}\right)}}$, i.e., defining a species field that is conjugate to $\hat{\rho}_i{{\left({{\bf {r}}}\right)}}$. $\mathcal{Z}_\mathrm{ig} = \left(\prod_{p=1}^P\lambda_T^{3n_pN_p}n_p!\right)^{-1}\prod_{i=1}^S D_i^{-1}$.
Finally, $$\mathcal{Z}_c = \mathcal{Z}_0 {{\int\mathcal{D}}}\left\{\mu_i\right\} e^{-H\left[\left\{\mu_i\right\}\right]},$$ where $$\begin{aligned}
\mathcal{Z}_0 & = & \left[\prod_{p=1}^P\frac{V^{n_p}g_{N_p}^{n_p}}{\lambda_T^{3n_pN_p}n_p!}\right]e^{-\beta U_\mathrm{ref}}\prod_{i=1}^SD_i^{-1}\\
H\left[\left\{\mu_i\right\}\right] & = & -\sum_{i=1}^S \frac{C\left(\zeta N\right)^{-\frac{1}{2}}}{2d_i}{{\int d{{\bf {r}}}\,}}\mu_i^2{{\left({{\bf {r}}}\right)}}-\frac{C}{\left(\zeta N\right)^{-\frac{1}{2}}}\sum_{i,j=1}^S{{\int d{{\bf {r}}}\,}}\frac{O_{ji}\mu_i{{\left({{\bf {r}}}\right)}}}{d_i} - CV\sum_{p=1}^P \frac{\phi_p}{\alpha_p}\ln Q_p\left[{{\underline{\underline{O}}}}{{\underline{\mu}}}\right]\\
Q_p\left[\left\{\psi_i\right\}\right] & = & \left(Vg_{N_p}\right)^{-1}\left[{{\int\mathcal{D}}}{{\bf {r}}}^p\exp\left(-\frac{3N}{2b^2}\int_0^{\alpha_p}ds\,\left|\frac{d{{\bf {r}}}^p}{ds}\right|^2-\sum_{i=1}^S\int_{s\in i}ds\,\psi_i\left({{\bf {r}}}^p\left(s\right)\right)\right)\right]\\
& = & \frac{1}{V}{{\int d{{\bf {r}}}\,}}q_p\left({{\bf {r}}},s=\alpha_p\right)\end{aligned}$$
Note that the compressibility penalty appears as $\left(\zeta N\right)^{-1/2}$ everywhere, including in the ${{\underline{\underline{X}}}}$ matrix. We use $\left(\zeta N\right)^{-1}$ as a convenient control parameter that recovers the incompressible limit at $0$.
[^1]: These authors contributed equally to this work.
[^2]: These authors contributed equally to this work.
[^3]: A subtle issue should be noted here. The dimensions of a functional derivative $\delta H/\delta \mu{{\left({{\bf {r}}}\right)}}$ are $\left[H\right]\left[\mu\right]^{-1}\left[V\right]^{-1}$. Hence, if the $H$ functional is discretized first, the partial derivative $\left(\Delta V\right)^{-1} \partial H\left({{\bf {\mu}}}\right)/\partial \mu_i$ is the one that correctly approaches $\delta H/\delta \mu{{\left({{\bf {r}}}\right)}}$ in the continuum limit.
[^4]: $\Delta t$ required to converge $\left<H\right>$ to within 0.05% of low-$\Delta t$ value.
[^5]: $N_t\left(\Delta t_\mathrm{max}\right)$ is the number of time steps required to reduce the standard error of the mean ($\epsilon$) sufficiently so that $3\epsilon$ is no larger than $1$ in the 4th significant figure of $\left<H\right>$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this article, we introduce the notion of cluster automorphism of a given cluster algebra as a ${\mathbb{Z}}$-automorphism of the cluster algebra that sends a cluster to another and commutes with mutations. We study the group of cluster automorphisms in detail for acyclic cluster algebras and cluster algebras from surfaces, and we compute this group explicitly for the Dynkin types and the euclidean types.'
author:
- 'Ibrahim Assem, Ralf Schiffler and Vasilisa Shramchenko'
title: Cluster Automorphisms
---
[Introduction]{} Cluster algebras were introduced by Fomin and Zelevinsky in [@FZ1; @FZ2] in the context of canonical bases and total positivity. These are $
\mathbb{Z}$-algebras whose generators are grouped into sets called clusters, and one passes from one cluster to another using an operation called mutation.
We are interested in the question whether one can study cluster algebras as a category. This means understanding the morphisms which preserve their very particular structure, that is, which keep invariant the grouping of generators into clusters and are compatible with mutations. As a first step in this direction, we study here what we call cluster automorphisms. We define a cluster automorphism of a given cluster algebra as an automorphism of $\mathbb{Z}$-algebras sending a cluster to another and commuting with mutations. Thus, in this paper, we study the symmetries of a given cluster algebra and compute the cluster automorphism group for the best known classes of cluster algebras, those arising from an acyclic quiver and those arising from a surface. Observe that, in [@FZ2], Fomin and Zelevinsky have considered a related notion of strong isomorphisms, by which they mean an isomorphism of the cluster algebras which maps every seed to an isomorphic seed. As will follow from our results, a strong automorphism of a cluster algebra is what we call here a direct cluster automorphism.
Let $\mathcal{A}=\mathcal{A}(\mathbf{x},Q)$ be a cluster algebra. Among the most interesting properties of the automorphism group ${\textup{Aut}\,\mathcal{A}}$ of $\mathcal{A}$ is the fact that an element of this group sends the quiver $Q$ either to itself or to the opposite quiver $Q^{\textup{op}}$. This allows to define a subgroup ${\textup{Aut}^+\mathcal{A}}$ of ${\textup{Aut}\,\mathcal{A}}$ consisting of those automorphisms sending $Q$ to itself. We prove that the index of ${\textup{Aut}^+\mathcal{A}}$ in ${\textup{Aut}\,\mathcal{A}}$ is two if and only if $Q$ is mutation equivalent to $Q^{\textup{op}}$ and otherwise ${\textup{Aut}\,\mathcal{A}}= {\textup{Aut}^+\mathcal{A}}$. We first compute these groups in the context of acyclic cluster algebras. In this case, the combinatorics of the cluster algebra is nicely encoded in the cluster category introduced in [@BMRRT] and, for type $\mathbb{A}$, also in [@CCS]. In particular, we recall that the Auslander-Reiten quiver of the cluster category of an acyclic cluster algebra has a particular connected component, called the transjective component.
Let $\mathcal{A}$ be an acyclic cluster algebra and ${\Gamma}_{tr}$ the transjective component of the Auslander-Reiten quiver of the associated cluster category. Then ${\textup{Aut}^+\mathcal{A}}$ is the quotient of the group $\textup{Aut}({\Gamma}_{tr})$ of the quiver automorphisms of ${\Gamma}_{tr}$, modulo the stabiliser $\textup{Stab}({\Gamma}_{tr})_0$ of the points of this component. Moreover, if ${\Gamma}_{tr} \cong \mathbb{Z}\Delta$, where $\Delta$ is a tree or of type $\tilde{\mathbb A}$ then ${\textup{Aut}\,\mathcal{A}}= {\textup{Aut}^+\mathcal{A}}\rtimes {\mathbb{Z}}_2$ and this semidirect product is not direct.
As an easy consequence, we compute the automorphism groups of the cluster algebras of Dynkin and euclidean types.
We next consider the case of the cluster algebras arising from an oriented marked surface $(S,M)$ with $p$ punctures in the sense of Fomin, Shapiro and Thurston [@FST] (see also [@FG]). We define the marked mapping class group ${\mathcal{MG}_{\bowtie}(S,M)}$ of $(S,M)$ to be a semidirect product of a power of ${\mathbb{Z}}_2$ and the quotient ${\mathcal{MG}(S,M)}$ of the group ${\textup{Homeo}^+(S,M)}$ of orientation preserving homeomorphisms from the surface $S$ to itself which map the set of marked points $M$ to itself, modulo the subgroup ${\textup{Homeo}_0(S,M)}$ consisting of those $f\in {\textup{Homeo}^+(S,M)}$ which are isotopic to the identity via an isotopy that fixes $M$ pointwise. We then prove our second theorem.
Let $(S,M)$ be a marked surface. Then ${\mathcal{MG}_{\bowtie}(S,M)}$ is isomorphic to a subgroup of ${\textup{Aut}^+\mathcal{A}}$. Furthermore, if $(S,M)$ is a disc or an annulus without punctures, then ${\mathcal{MG}(S,M)}\cong{\textup{Aut}^+\mathcal{A}}$.
We also consider the case of the disc with one or two punctures and show that (except for two exceptional cases) we also have ${\mathcal{MG}_{\bowtie}(S,M)}\cong{\textup{Aut}^+\mathcal{A}}$.
As a consequence of our results, we show that if the cluster algebra $\mathcal{A}$ is acyclic or arising from a surface, then the group ${\textup{Aut}\,\mathcal{A}}$ is finite if and only if $\mathcal{A}$ is a cluster algebra of Dynkin type.
The paper is organised as follows. In section \[sect 1\], we define our notion of cluster automorphism and prove some of its elementary properties, section \[sect\_acyclic\] is devoted to the case of acyclic cluster algebras and section \[sect 3\] to that of cluster algebras arising from surfaces. Finally in section \[sect 4\], we consider the finiteness of the automorphism group.
[**Acknowledgements.**]{}[The authors thank the referee for pointing out an error in an earlier version of the paper, as well as for many useful suggestions.]{}
Cluster automorphisms {#sect 1}
=====================
Cluster algebras
-----------------
We recall that a [*quiver*]{} is a quadruple $Q=(Q_0,Q_1,s,t)$ consisting of two sets: $Q_0$ (whose elements are called *points*) and $Q_1$ (whose elements are called *arrows*), and of two maps $s,t:Q_1\to Q_0$ associating to each $\alpha \in Q_1$ its *source* $s(\alpha)$ and its *target* $t(\alpha)$, respectively. Given a point $i\in Q_0$, we denote by $i^-=\{ \alpha\in Q_1\mid t(\alpha)=i \}$ the set of arrows ending in $i$, and by $i^+=\{ \alpha\in Q_1\mid s(\alpha)=i \}$ the set of arrows starting in $i$.
Let now $Q$ be a connected finite quiver without oriented cycles of length one or two. Let $n=|Q_0|$ denote the number of points in $Q$, let $\amas=\{x_1,\dots,x_n\}$ be a set of $n$ variables, and denote the points by $Q_0=\{1,\ldots,n \}$, where we agree that the point $i$ corresponds to the variable $x_i$. We consider the field ${\cal F} = \mathbb{Q}(x_1,\dots,x_n)$ of rational functions in $x_1,\dots,x_n$, which we call the [*ambient field*]{}. The [*cluster algebra*]{} ${\cal A} = {\cal A}(\amas,Q)$ is a $\mathbb{Z}$-subalgebra of ${\cal F}$ defined by a set of generators obtained recursively from $\amas$ in the following manner. Let $i$ be such that $1\leq i \leq n.$ The [*mutation*]{} $\mu_{x_i,\amas}$ of $(\amas,Q)$ (or $\mu_{x_i}$ or $\mu_i$ for brevity if there is no ambiguity) is defined as follows. Firstly, $Q$ becomes a new quiver $Q^\prime$ obtained from $Q$ by:
- inserting a new arrow $k\to j$ for each path $k\to i \to j$ of length two with midpoint $i$;
- inverting all arrows of $Q$ passing through $i$;
- deleting each occurrence of a cycle of length two.
Secondly, $\amas$ becomes a new set of variables $\amas^\prime=(\amas\setminus\{x_i\}) \cup \{x_i^\prime\}$ where $x_i^\prime\in{\cal F}$ is defined by the so-called [*exchange relation*]{}: $$x_i x_i^\prime = \prod_{\alpha\in i^+} x_{t(\alpha)} + \prod_{\alpha\in i^-} x_{s(\alpha)}.$$ Let ${\cal X}$ be the union of all possible sets of variables obtained from $\amas$ by successive mutations. Then ${\cal A} = {\cal A}(\amas, Q)$ is the $\mathbb{Z}$-subalgebra of ${\cal F}$ generated by ${\cal X}$.
Each pair $(\tilde{\amas}, \tilde{Q})$ obtained from $(\amas, Q)$ by successive mutations is called a [*seed*]{}, and $\tilde{\amas}$ is called a [*cluster*]{}. The elements $\tilde{x}_1,\dots, \tilde{x}_n$ of a cluster $\tilde{\amas}$ are [*cluster variables*]{}. Each cluster is a transcendence basis for the ambient field $\mathcal{F}$. The pair $(\amas, Q)$ is the [*initial seed*]{} and $\amas$ is the [*initial cluster*]{}.
It has been shown in [@GSV Theorem 3] that for every seed $(\tilde\amas,\tilde Q)$ the quiver $\tilde Q$ is uniquely defined by the cluster $\tilde \amas$, and we use the notation $Q(\tilde\amas)$ for the quiver of the cluster $\tilde\amas$. More precisely, there is a canonical bijection $p$ from the cluster $\tilde \amas$ to the set of points of the quiver $Q(\tilde\amas)$. We write $p_x$ for the point in $Q(\tilde\amas)$ corresponding to the cluster variable $x\in \tilde\amas.$
We recall two results from the theory of cluster algebras. The so-called [Laurent phenomenon]{} is the fact that each cluster variable can be expressed as a Laurent polynomial in the $x_i$, with $i=1,\dots,n$, so that ${\cal A} \subseteq {\mathbb{Z}}[x_1^{\pm 1},\dots, x_n^{\pm 1}]$, see [@FZ1]. Also, ${\cal A}$ is [*of finite type*]{}, that is the set $\cal{X}$ of all cluster variables of $\cal A$ is finite, if and only if there exists a sequence of mutations transforming $Q$ into a Dynkin quiver [@FZ2]. In the latter case (as well as in many others), the so-called [*positivity conjecture*]{} holds true, that is the cluster variables are Laurent polynomials with non-negative coefficients or, equivalently, the cluster variables are contained in ${\mathbb{Z}_{\geq 0}}[x_1^{\pm 1},\dots, x_n^{\pm 1}]$, see [@MSW; @N; @AsRS].
Main definition
---------------
We define cluster automorphisms as follows.
\[def\_main\] Let $\cal A$ be a cluster algebra, and let $f:{\cal A} \to {\cal A}$ be an automorphism of ${\mathbb Z}$-algebras. Then $f$ is called a *cluster automorphism* if there exists a seed $(\amas,Q)$ of ${\cal A}$, such that the following conditions are satisfied:
- $f(\amas)$ is a cluster; \[def1\]
- $f$ is compatible with mutations, that is, for every $x\in \amas$, we have $$f(\mu_{x,\amas}(x)) = \mu_{f(x),f(\amas)}(f(x)).$$ \[def2\]
\[rmk\_initial\]
- As we shall see in Proposition \[prop1\] below, if $\mathcal{A}$ is a cluster algebra, and $f:\mathcal{A\rightarrow A}$ is an automorphism of $\mathbb{Z}$-algebras, then $f$ is a cluster automorphism if and only if it satisfies properties and for *every* seed $
\left( \mathbf{x}^{\prime},Q^{\prime }\right) $ of $\mathcal{A}$, thus it sends clusters to clusters and commutes with any sequence of mutations.
- Every cluster automorphism $f$ is uniquely determined by its value on the initial cluster variables $x_1,\dots, x_n$ and thus extends in a unique way to an automorphism of the ambient field ${\cal F} = \mathbb{Q}(x_1, \dots, x_n)$ by $$\frac{p(x_1,\dots,x_n)}{q(x_1,\dots,x_n)} \mapsto \frac{p(f(x_1), \dots, f(x_n))}{q(f(x_1), \dots, f(x_n))},$$ for all polynomials $p,q$.
The converse, however, is not true, as shown in the following Example.
\[ex 1\] \[ex\_mutation\] We give an example of a $\mathbb Z$-automorphism of the ambient field $\cal F$ which does not restrict to a cluster automorphism of $\cal A$. Let $Q$ be the following quiver $$\xymatrix{ 1&2\ar[l]&3\ar[l]}
$$ and $\amas=(x_1,x_2,x_3)$. Clearly, any change of transcendence basis of ${\cal F} = {\mathbb Q}(x_1,x_2,x_3)$ induces an automorphism of $\cal F$, and such a change is induced, for instance, by a mutation. Let us define $f:{\cal F}\to {\cal F}$ by $f=\mu_{x_1}$, that is $$\begin{aligned}
&&f(x_1) = \mu_{x_1}(x_1)= \frac{1+x_2}{x_1},\\
&&f(x_2) = \mu_{x_1}(x_2) = x_2,\\
&&f(x_3) = \mu_{x_1}(x_3) = x_3. \end{aligned}$$ Then $f(\amas) = \mu_{x_1}(\amas)$ is a cluster. On the other hand, a straightforward calculation gives that $$f\mu_{x_2, \amas}(x_2) = \frac{1+x_2+x_1x_3}{x_1x_2}$$ while $$\mu_{f(x_2),f(\amas)}f(x_2) = \frac{x_1+x_3+x_2x_3}{x_1x_2}.$$ Thus condition is not satisfied and $f$ is not a cluster automorphism of $\cal A$.
The above automorphism does not even map cluster variables to cluster variables. Indeed, one of the nine cluster variables in the cluster algebra $\mathcal{A}$ has the Laurent polynomial expansion $$\frac{x_1 + x_1x_2 + x_3}{x_2x_3}$$ and applying $f$ to this cluster variable gives
$$\frac{1+2x_2+x_2^2 + x_1x_3}{x_1x_2x_3}$$ which is not a cluster variable in $\mathcal{A}$.
Equivalent characterisations
----------------------------
Since mutations induce maps of a cluster algebra onto itself, we may wonder why in the previous example we do not obtain a cluster automorphism. As we see below, the reason is that the associated quiver $Q(f(\amas))$ $$\xymatrix{p_{f(x_1)}\ar[r]&p_{f(x_2)}&p_{f(x_3)}\ar[l]},
$$ is isomorphic neither to the original quiver $Q$ nor to its opposite $Q^{\textup{op}}$.
When we say that a sequence of mutations $\mu$ transforms a quiver $Q$ into itself, we mean that for any $\alpha \in Q_1$ we have $\mu(s(\alpha))=s(\mu(\alpha))$ and $\mu(t(\alpha))=t(\mu(\alpha))$. When we say that $\mu$ transforms $Q$ into $Q^{op}$, we mean that for any $\alpha\in Q_1$ we have $\mu(s(\alpha))=t(\mu(\alpha))$ and $\mu(t(\alpha))=s(\mu(\alpha))$.
\[lemma1\] Let $f$ be a ${\mathbb{Z}}$-algebra automorphism of $\mathcal{A}$. Then $f$ is a cluster automorphism if and only if there exists a seed $(\amas,Q)$ such that $f(\amas)$ is a cluster and one of the following two conditions is satisfied:
- there exists an isomorphism of quivers $\varphi : Q\longrightarrow Q(f(\amas))$ such that $\varphi(p_{x})=p_{f(x)}$ for all $p_{x}\in Q_0$, or
- there exists an isomorphism of quivers $\varphi : Q^{\textup{op}}\longrightarrow Q(f(\amas))$ such that $\varphi(p_{x})=p_{f(x)}$ for all $p_x\in Q_0$.
Consider a cluster $\amas$ in $\mathcal{A}$. For a cluster variable $x_i$ in $\amas$, corresponding, say, to the point $i\in Q_0$, the exchange relation reads: $$\mu_{x_i,\amas}(x_i) = \frac{1}{x_i} \left( \prod_{\alpha \in i^+ } x_{t(\alpha)} + \prod_{\alpha \in i^-} x_{s(\alpha)} \right).$$ Since $f$ is an algebra homomorphism, this implies $$f(\mu_{x_i,\amas}(x_i)) = \frac{1}{f(x_i)}\left(\prod_{\alpha \in i^+ } f(x_{t(\alpha)}) + \prod_{\alpha \in i^-} f(x_{s(\alpha)}) \right).
\label{temp1}$$ On the other hand, $f$ induces a map $\varphi :Q_0 \to Q(f(\amas))_0$ defined by $\varphi(p_{x})=p_{f(x)}$. Hence $$\mu_{f(x_i),f(\amas)} (f(x_i))
= \frac{1}{f(x_i)}\left( \prod_{\beta \in \varphi(i)^+ } f(\amas)_{t(\beta)} + \prod_{ \beta \in \varphi(i)^-} f(\amas)_{s(\beta)} \right) ,
\label{temp2}$$ where we denote by $f(\amas)_j$ the cluster variable in the cluster $f(\amas)$ which corresponds to the point $j\in Q(f(\amas))_0$. Then $f$ is a cluster automorphism if and only if the expressions (\[temp1\]) and (\[temp2\]) coincide for every $i$. This is the case if and only if we have one of the following two situations: either
- $$\prod_{\alpha \in i^+} f(x_{t(\alpha)}) = \prod_{ \beta\in \varphi(i)^+ } f(\amas)_{t(\beta)} \qquad \mbox{and} \qquad
\prod_{ \alpha\in i^-} f(x_{s(\alpha)}) = \prod_{ \beta \in \varphi(i)^- } f(\amas)_{s({\beta)}};$$
or
- $$\prod_{ \alpha \in i^+} f(x_{t(\alpha)}) =\prod_{ \beta \in \varphi(i)^- } f(\amas)_{s({\beta)}} \qquad \mbox{and} \qquad
\prod_{ \alpha\in i^-} f(x_{s(\alpha)}) = \prod_{ \beta\in \varphi(i)^+ } f(\amas)_{t(\beta)}.$$
Since the set $f\left( \mathbf{x}\right) $ is a transcendance basis of the ambient field $\mathcal{F}$, this implies that we have either
- $i^{+}=\varphi (i)^{+}$ and $i^{-}=\varphi (i)^{-}$\
or
- $i^{+}=\varphi (i)^{-}$and $i^{-}=\varphi (i)^{+}\\ $ for every $i=1,\dots,n.$
Let us now prove that if one of the two situations (i) or (ii) holds for the point $i$, then the same situation holds for every point of $Q$.
Suppose that we are in the situation (i) for the point $i$, and let $j$ be a neighbour of $i$. Without loss of generality, we can assume that there is an arrow $\alpha
:j\rightarrow i$ in the quiver $Q$, that is, $\alpha \in i^{-}$. This implies that there is no arrow from $i$ to $j$ because $Q$ has no cycles of length two. From the bijection between $i^{-}$ and $\varphi \left(
i^{-}\right) $, we get an arrow $\varphi \left( \alpha \right) :\varphi
\left( j\right) \rightarrow \varphi (i),$ and so $\varphi \left(\alpha \right)
\in \varphi (j)^{+}$. Since $\alpha \in j^{+}$, this implies that there is a bijection $j^{+}\cong \varphi \left( j\right) ^{+}$. The proof is entirely similar if we choose $\alpha$ in $i^{+}.$ Proceeding in this way from neighbour to neighbour, we see that the map $\varphi $ between points extends to an isomorphism of quivers $Q\cong Q(f(x)).$
Analogously, the situation (ii) yields an isomorphism of quivers $Q^{op}\rightarrow Q(f(\mathbf{x}))$.
In the sequel, we shall mostly need quiver isomorphisms satisfying one of the conditions of Lemma \[lemma1\]. Accordingly, if $f$ is a cluster automorphism of the cluster algebra $\mathcal{A}\left( \amas,Q\right) $, then $f$ induces a map (actually a bijection) between the points $\varphi
:Q_{0}\rightarrow Q\left( f\left( \amas \right) \right) _{0}$ by $
\varphi \left( p_{x}\right) =p_{f\left( x\right) }$ for every $x\in
\amas$. If this bijection $\varphi $ extends to an isomorphism of quivers $\varphi :Q\rightarrow Q\left( f\left( \amas \right) \right)
$ then we say that the latter is *induced* by $f$, and that $f$ is a *direct cluster automorphism*. Similarly, if $\varphi $ extends to an isomorphism of quivers $\varphi :Q^{op}\rightarrow Q\left( f\left(
\amas\right) \right) $ then we also say that $\varphi $ is *induced* by $f$ but then $f$ is called an *inverse cluster automorphism*.
\[prop1\] Let $f$ be a cluster automorphism. Then $f$ satisfies conditions and for *every* seed.
Any seed is obtained from the seed $(\amas,Q)$ of Definition \[def\_main\] by a finite sequence of mutations. It is therefore enough to show that if (CA1), (CA2) hold for a seed $(\amas,Q)$, then they hold for any seed $(\amas',Q')$ that is obtained from $(\amas,Q)$ by a single mutation. Let $(\amas',Q')$ be such a seed. Then $(\amas',Q')=\mu_{x,\amas}(\amas,Q)$, for some $x\in \amas$, thus $$\amas'=\left(\amas \setminus\{x\}\right)\cup\{x'\},$$ with the exchange relation $$x'=\frac{1}{x}\left(\prod_{p_{x}\to p_{x_i} \in Q_1} x_i +\prod_{p_{x}{\leftarrow}p_{x_j} \in Q_1} x_j \right).$$
It follows that $f(\amas') =\left( f(\amas) \setminus\{f(x)\}\right)\cup\{f(x')\}.$ By (CA1), $f(\amas)$ is a cluster and by (CA2) $$f(x')=f(\mu_{x,\amas}(x))= \mu_{f(x),f(\amas)}(f(x)).$$ Therefore $f(\amas')=\mu_{f(x),f(\amas)}(f(\amas))$; in particular, $f(\amas')$ is a cluster. This shows (CA1).
Let us show that condition (a) or (b) of Lemma \[lemma1\] is satisfied for the seed $(\amas',Q')$. We have $Q'=\mu_x Q$. On the other hand, $$Q(f(\amas'))=Q(f(\mu_{x,\amas}(\amas))) = Q(\mu_{f(x),f(\amas)}f(\amas))
= \mu_{f(x)}(Q(f(\amas))),$$ where the second equality follows from the condition (CA2) for the seed $(\amas,Q)$. Now, one of the conditions (a) or (b) holds for the seed $(\amas,Q)$, thus, if (a) holds, then there is an isomorphism $\varphi:Q\to Q(f(\amas))$ induced by $f$, that is, such that $\varphi(p_{x})=p_{f(x)}$, for every $x\in \amas$, and therefore $$Q'=\mu_x Q\cong \mu_{f(x)}Q(f(\amas)) = Q(f(\amas')),$$ and this isomorphism sends every point $p_{x_i'}$ in $Q'$ to the point $p_{f(x_i')}$ in $Q(f(\amas'))$. In other words, the condition (a) holds for the seed $(\amas',Q')$. On the other hand, if condition (b) holds for $(\amas,Q)$, then there is an isomorphism $\varphi:Q^{\textup{op}}\to Q(f(\amas))$ induced by $f$, that is, such that $\varphi(p_{x})=p_{f(x)}$, for every $x\in \amas$, and therefore $$Q'^{\textup{op}}=\mu_x Q^{\textup{op}}\cong \mu_{f(x)}Q(f(\amas)) = Q(f(\amas')),$$ and this isomorphism sends any point $p_{x_i'}$ in $Q'^{\textup{op}}$ to the point $p_{f(x_i')}$ in $Q(f(\amas'))$; thus condition (b) holds for the seed $(\amas',Q')$.
Therefore since the structure of $f(Q^\prime)$ coincides with that of $Q^\prime$ or $Q'^{\textup{op}}$, the expressions analogous to (\[temp1\]) and (\[temp2\]) for the cluster $\amas'$ are equal, thus (CA2) is satisfied for the cluster $f\left( \mathbf{x}^{\prime }\right).$
\[lemma\_property\] Let ${\cal A=A}(\amas,Q)$ and $f:\cal A\to A$ be a cluster automorphism. Then
- If $f$ is direct, then it induces a quiver isomorphism $Q'\cong Q(f(\amas'))$, for any seed $(\amas',Q')$.
- If $f$ is inverse, then it induces a quiver isomorphism $Q'^{op}\cong Q(f(\amas'))$, for any seed $(\amas',Q')$.
This follows from the proof of Proposition \[prop1\] and (CA2).
\[rmk\_seeds\] It follows from (CA1) that a cluster automorphism amounts to a replacement of a cluster by another cluster of the same algebra. Since seeds are uniquely determined by clusters, a cluster automorphism can equivalently be considered as a “change of seed". Condition (CA2) says that this change is compatible with mutations so that if $(\amas,Q)$ is a seed of $\cal A$ and $x\in \amas$, then we have the following commutative diagram: $$\xymatrix@R35pt@C75pt{ (\amas,Q) \ar[r]^{{f}}\ar[d]^{\mu_{(x,\amas)}} & (f(\amas),Q(f(\amas))\ar[d]^{ \mu_{(f(x),f(\amas))} } \\
(\mu_{x,\amas}(\amas), \mu_{x,\amas}(Q))\ar[r]^{f} & (f\mu_{x,\amas}(\amas), Q(f\mu_{x,\amas}(\amas)) .}$$
We end this subsection with one more characterisation of cluster automorphisms.
\[cor2.7bis\] Let $\mathcal{A}=\mathcal{A}\left(
Q\right)$ and let $f:\mathcal{A}\rightarrow \mathcal{A}$ be a ${\mathbb{Z}}$-algebra automorphism. Then $f$ is a cluster automorphism if and only if $f$ maps each cluster to a cluster.
Assume that for every cluster $ \mathbf{x} $, $f(\mathbf{x)}$ is a cluster. We must prove that for every $x\in \mathbf{x}$ we have a commutative diagram as in the remark above. Let $x^{\prime }$ be the variable obtained from $x$ by mutation, then $\mathbf{x}^{\prime }=(\mathbf{x}\setminus \left\{ x\right\} )\cup \left\{ x^{\prime
}\right\} $ is a cluster. Because of our hypothesis, $f(\mathbf{x}^{\prime })=(f(\mathbf{x)}
\setminus\left\{ f(x)\right\} )\cup \left\{ f(x^{\prime
})\right\} $ is a cluster as well. On the other hand, mutating in $f(x)$ the cluster $f(\mathbf{x)}$ yields the cluster $(f(\mathbf{x)}\setminus\left\{ f(x)\right\} )\cup \left\{ y^{\prime }\right\}$. These two clusters are obtained from $f(\mathbf{x)}$ by mutating in the same variable $f(x)$ therefore $y^{\prime }=f(x^{\prime }).$
As a consequence, we see that the notion of direct cluster automorphism coincides with that of strong automorphism of a cluster algebra [@FZ2], that is, an isomorphism of $\mathbb{Z}$-algebras that maps every seed to an isomorphic seed.
The group of cluster automorphisms
----------------------------------
Examples and construction techniques for cluster automorphisms are given below. Clearly, the identity on $\cal A$ is a cluster automorphism. In fact, the following lemma holds.
The set ${\textup{Aut}\,\mathcal{A}}$ of all cluster automorphisms of $\cal A$ is a group under composition.
Let $f,g \in {\textup{Aut}\,\mathcal{A}}$. By Remark \[rmk\_seeds\], a cluster automorphism amounts to replacing the initial seed $\left( \mathbf{x},Q\right)
$ by another seed whose quiver is isomorphic to either $Q$ or $Q^{op},$ therefore $f^{-1} \in {\textup{Aut}\,\mathcal{A}}$ and $gf \in {\textup{Aut}\,\mathcal{A}}$.
\[lemma\_index\] The set $\Autp$ of all direct cluster automorphisms of $\cal A$ is a normal subgroup of ${\textup{Aut}\,\mathcal{A}}$ of index at most two.
Clearly, the identity of $\cal A$ is a direct automorphism. Also, if $f,g \in \Autp$, then $fg^{-1} \in \Autp$, therefore $\Autp$ is a subgroup of ${\textup{Aut}\,\mathcal{A}}$. The normality follows from the fact that if $f\in\Autp$ and $g\in{\textup{Aut}\,\mathcal{A}}$, then $gfg^{-1}$ induces an automorphism of $Q$ even if $g$ induces an anti-isomorphism.
In order to prove the statement about the index, let us consider a map $\phi: {\textup{Aut}\,\mathcal{A}}\to {\mathbb{Z}}_2$ defined by $$\label{2.3}
\phi(f) = \left\{\begin{array}{l} \bar{0}, \;\mbox{if} \;\; f\in \Autp \\\bar{1}, \;\mbox{if} \;\; f\notin \Autp \end{array}\right. .$$
The map $\phi$ is a group homomorphism. Indeed, if $f,g\in{\textup{Aut}\,\mathcal{A}}$, then $\phi(fg) = \bar{0}$ if and only if $fg \in \Autp$, that is, if and only if $f$ and $g$ are both direct or both inverse. The latter condition may be written as $\phi(f)=\phi(g)$, which holds if and only if $\phi(f)+\phi(g)=\bar{0}.$ Thus $\phi(fg) = \phi(f)+\phi(g)$, and $\phi$ is a group homomorphism. Since $\textup{Ker}\, \phi=\Autp$ and ${\rm Im} \,\phi \subseteq {\mathbb{Z}}_2,$ the lemma is proved.
Here is an example of an inverse cluster automorphism. Let $Q$ be the following quiver of type $\A_3$: $$\xymatrix{p_{x_1}&p_{x_2}\ar[l]\ar[r]&p_{x_3}}
$$ and $\amas=\{x_1,x_2,x_3\}$. The cluster variables computed inside the cluster category ${\cal C}_Q$ (see [@BMRRT] or Section \[sect\_acyclic\] below) are as follows: $$\scalebox{0.95}{$\begin{array}{ccccccccccccccc}
& & x_3 & & & & \frac{1+x_2+x_1x_3}{x_2x_3} & & & & \frac{1+x_2}{x_1}& & & &x_1\\
& \nearrow & & \searrow & & \nearrow & & \searrow& & \nearrow & & \searrow& & \nearrow & \\
x_2& & & & \frac{1+x_1x_3}{x_2} &&&&\frac{x_2^2 + 2x_2+1+x_1x_3}{x_1x_2x_3}&&&&x_2 \\
& \searrow & & \nearrow & & \searrow & & \nearrow& & \searrow & & \nearrow& & \searrow & \\
& &x_1& &&&\frac{1+x_2+x_1x_3}{x_1x_2}&&&&\frac{1+x_2}{x_3}&&&&x_3
\end{array}$}$$ Define a map $f:\cal A\to A$ to be induced by the mutation $\mu_{x_2}$, so that on the initial cluster we have $$f(x_1) = x_1, \qquad f(x_2) = \frac{1+x_1x_3}{x_2}, \qquad f(x_3) = x_3.$$ Then $f$ extends to an algebra homomorphism. A straightforward computation gives the images under $f$ of the remaining cluster variables of the algebra: $$\begin{aligned}
&& f \left( \frac{1+x_1x_3}{x_2} \right) = x_2, \qquad \qquad \qquad f \left( \frac{1+x_1x_3+x_2}{x_2x_3} \right) = \frac{x_2+1}{x_3}, \\
&& f \left( \frac{1+x_1x_3+x_2}{x_1x_2} \right) = \frac{x_2+1}{x_1}, \qquad f \left( \frac{x_2^2 + 2x_2+1+x_1x_3}{x_1x_2x_3} \right) = \frac{x_2^2 + 2x_2+1+x_1x_3}{x_1x_2x_3}, \\
&& f \left( \frac{1+x_2}{x_1} \right) = \frac{x_2+1+x_1x_3}{x_1x_2}, \qquad f \left( \frac{1+x_2}{x_3} \right) = \frac{x_2+1+x_1x_3}{x_2x_3} .\end{aligned}$$ Thus $f$ is a cluster automorphism sending the seed $$(\{x_1,x_2,x_3\},
\xymatrix{p_{x_1}&p_{x_2}\ar[l]\ar[r]&p_{x_3}} )$$ to the seed $$(\{f(x_1),f(x_2),f(x_3)\},\xymatrix{p_{f(x_1)}\ar[r]&p_{f(x_2)}&p_{f(x_3)}\ar[l]})$$ hence $f$ induces an isomorphism of quivers $Q(\amas)^{op}\cong Q(f(\amas))$.
Two quivers $Q$ and $Q'$ are called *mutation equivalent* if there exists a sequence of mutations transforming $Q$ to $Q'$.
\[thm ref\] Let $\mathcal{ A}=\mathcal{A}(\amas,Q)$.
- [If $Q$ and $Q^{op}$ are mutation equivalent then the index of ${\textup{Aut}^+\mathcal{A}}$ in ${\textup{Aut}\,\mathcal{A}}$ is two.]{}
- If $Q$ and $Q^{op}$ are not mutation equivalent then ${\textup{Aut}\,\mathcal{A}}=\Autp$.
[ If $Q$ and $Q^{op}$ are mutation equivalent then there exists a sequence of mutations $\mu$ such that $\mu(\amas,Q(\amas))=(\amas',Q(\amas'))$ together with two isomorphisms of quivers $\varphi:Q(\amas)\stackrel{\cong }{\to}Q$ and $\varphi':Q^{op}\stackrel{\cong }{\to}Q(\amas')$. [Observe that there is no reason for the isomorphisms $\varphi$ and $\varphi'$ to be compatible with the canonical bijection between the points $Q_0\cong Q(\amas')_0$.]{} Define a map $f:\amas\to\amas'$ as the following composition of the bijections on points $$Q(\amas)_0\stackrel{\varphi}{{\longrightarrow}} Q_0 \stackrel{=}{{\longrightarrow}} (Q^{op})_0\stackrel{\varphi'}{{\longrightarrow}} Q(\amas')_0.$$ Since $\amas$ and $\amas'$ are transcendence bases of the ambient field, then $f$ extends to a ${\mathbb{Z}}$-algebra automorphism of $\mathcal{A}$, and, moreover, $\varphi' :Q^{op}\stackrel{\cong }{\to}Q(\amas')$ is an isomorphism of quivers that satisfies $\varphi'(p_x)=p_{f(x)}$, and it follows from Lemma 2.3 that $f\in {\textup{Aut}\,\mathcal{A}}\setminus\Autp$. This shows (a).]{}
In order to show (b), suppose that $Q$ and $Q^{op}$ are not mutation equivalent. Suppose that there exists $f\in{\textup{Aut}\,\mathcal{A}}\setminus\Autp$. Then there exists a seed $(\amas,Q(\amas))$ in $\mathcal{A}$ with $Q(\amas)\cong Q$ whose image $(f(\amas),Q(f(\amas)))$ under $f$ is a seed in $\mathcal{A}$ whose quiver is isomorphic to $Q^{op}$. Thus $Q$ and $Q^{op}$ are mutation equivalent, a contradiction.
\[cor two new\] [If there exists $\sigma\in{\textup{Aut}\,\mathcal{A}}\setminus{\textup{Aut}^+\mathcal{A}}$ of order two, then ${\textup{Aut}\,\mathcal{A}}={\textup{Aut}^+\mathcal{A}}\rtimes {\mathbb{Z}}_2$.]{}
[If ${\textup{Aut}\,\mathcal{A}}= \Autp $ then there is an exact sequence of groups. ]{} $$1 \to \Autp \to {\textup{Aut}\,\mathcal{A}}\stackrel{\phi}{\to} {\mathbb{Z}}_2\to1$$ [where $\phi$ is defined in equation (\[2.3\]). We have ${\textup{Aut}\,\mathcal{A}}\cong \Autp \rtimes{\mathbb{Z}}_2$ if and only if this sequence splits, and this is the case if and only if there exists an inverse automorphism $\sigma$ of order 2. ]{}
\[ex torus\] Let $Q$ be the quiver $$\xymatrix@R20pt@C25pt{1\ar@<2pt>[rd]\ar@<-2pt>[rd] &&2 \ar@<2pt>[ll]\ar@<-2pt>[ll]\\
&3\ar@<2pt>[ru]\ar@<-2pt>[ru]
}$$ and $\mathcal{A}=\mathcal{A}(\{x_1,x_2,x_3\},Q\})$ its cluster algebra. Then the mutation of $Q$ in $1$ is the quiver $$\xymatrix@R20pt@C25pt{1\ar@<2pt>[rr]\ar@<-2pt>[rr] &&2 \ar@<2pt>[ld]\ar@<-2pt>[ld]\\
&3\ar@<2pt>[lu]\ar@<-2pt>[lu]
},$$ and thus it induces an inverse cluster automorphism $f$ given by $$f(x_1)=\frac{x_2^2+x_3^2}{x_1}, \ f(x_2)=x_2 \textup{ and } f(x_3)=x_3.$$
Mutating once more, this time in $x_2$ yields back the quiver $Q$, and thus it induces a direct cluster automorphism $g$ given by $$g(x_1)=\frac{x_2^2+x_3^2}{x_1}, \ g(x_2)=\frac{1}{x_2} \left(\left(\frac{x_2^2+x_3^2}{x_1}\right)^2+x_3^2\right)\textup{ and } g(x_3)=x_3 .$$
Note that in this particular example, any sequence of mutations will either produce the quiver $Q$ (if the number of mutations is even) or its opposite (if the number of mutations is odd), and thus, in this example, any sequence of mutations induces a cluster automorphism.
\[ex ref\] Let $\mathcal{A}=\mathcal{A}(\amas,Q)$, where $Q$ is the quiver $$\xymatrix{&2\ar@<3pt>[rd]\ar@<0pt>[rd]\\1\ar@<3pt>[rr]\ar@<-3pt>[rr]\ar@<0pt>[rr] \ar[ur] &&3\ .}$$ Then $Q $ is not mutation equivalent to $Q^{op}$ (and therefore ${\textup{Aut}^+\mathcal{A}}={\textup{Aut}\,\mathcal{A}}$). This can be seen using the associated cluster category, see [@BMRRT] or Section 3 below. Indeed, assume that $Q$ is mutation equivalent to $Q^{op}$. There exists a local slice $\Sigma$ (in the sense of [@ABS2]) in the cluster category ${\cal C}_Q$ of $Q$ whose quiver is isomorphic to $Q$. Since $Q$ and $Q^{op}$ are acyclic, the slice $\Sigma$ lies in the transjective component $\Gamma_{tr}$ of the Auslander-Reiten quiver of ${\cal C}_Q$, and this transjective component is of the form ${\mathbb{Z}}Q^{op}$. In particular, the point (corresponding to) $3$ in $\Sigma$ is the target of five arrows in $\Gamma_{tr}$. The point $1$ in $\Sigma$ is the source of three of the arrows of target $3$, and the point $2$ is the source of the remaining two arrows. Since $\Gamma_{tr}\cong {\mathbb{Z}}Q^{op}$, this implies the existence of an arrow from $2$ to $1$, and then the quiver of $\Sigma$ is not isomorphic to $Q$, a !
contradiction.
Cluster automorphisms induced by quiver automorphisms
-----------------------------------------------------
We now show how any automorphism of the quiver $Q$ induces a direct cluster automorphism of ${\cal A} = {\cal A}(\amas,Q)$. Let $\sigma \in {\rm Aut}\, Q$. Define a map $f_\sigma: \amas \to \amas$ by $f_\sigma(x)= x'$, for $x \in \amas$, where $x'\in\amas $ is the unique cluster variable such that $\sigma(p_x)=p_{x'}$.
Then $f_\sigma$ permutes the cluster variables in $\amas$ and clearly extends to a unique automorphism $f_\sigma: \cal F \to F$ of the ambient field. We now show that $f_\sigma$ is a direct cluster automorphism of $\cal A.$
\[prop 2.13\] \[prop\_kernel\] The map $F:\sigma \mapsto f_\sigma$ is a group homomorphism from ${\rm Aut}\, Q$ to $\Autp$ whose kernel is given by the stabiliser ${\rm Stab} \, Q_0$ of the points of $Q.$
In order to show that $f_\sigma$ is a cluster automorphism we must prove that conditions (CA1) and (CA2) are satisfied for the initial cluster $\amas$. Since $f_\sigma(\amas) = \amas$, the first condition obviously holds.
Let now $x \in \amas$ and consider the mutation $\mu_x = \mu_{x,\amas}$. Since $\sigma \in {\rm Aut}\, Q$, we have $\sigma Q = Q$ and hence $\mu_{f_\sigma(x),\amas}$ is a mutation of the seed $(\amas, Q)$.
We want to show (CA2), that is $$f_\sigma(\mu_{x,\amas}(x))=\mu_{f_\sigma(x),\amas}(f_\sigma(x)).$$ Using the exchange relations, we have $$\label{eq25}
f_\sigma(\mu_{x,\amas}(x))
= \frac{1}{f_{\sigma}(x)} \left(\,\prod_{{\alpha}\in p_x^+}f_\sigma(x_{t(\alpha)}) + \prod_{{\alpha}\in p_x^-}f_\sigma(x_{s(\alpha)})\right) ,$$ while $$\label{eq26}
\mu_{f_\sigma(x),\amas}(f_\sigma(x))
= \frac{1}{f_\sigma(x)} \left(\prod_{{\beta}\in \sigma(p_x)^+} x_{t({\beta})} + \prod_{{\beta}\in \sigma(p_x)^-}x_{s({\beta})}\right) .$$ Since $\sigma $ is an automorphism of $Q$, we have $\sigma(t({\alpha}))=t(\sigma({\alpha}))$, $\sigma(s({\alpha}))=s(\sigma({\alpha}))$, and $\sigma(p_x^+)=\sigma(p_x)^+$, $\sigma(p_x^-)=\sigma(p_x)^-$. Therefore $$\begin{aligned}
\prod_{{\alpha}\in p_x^+}f_\sigma(x_{t(\alpha)}) = \prod_{{\alpha}\in p_x^+}x_{t(\sigma(\alpha))} =\prod_{{\beta}\in \sigma(p_x)^+} x_{t({\beta})} && \textup{and}
\\
\prod_{{\alpha}\in p_x^-}f_\sigma(x_{s(\alpha)}) = \prod_{{\alpha}\in p_x^-}x_{s(\sigma(\alpha)) }=\prod_{{\beta}\in \sigma(p_x)^-} x_{s({\beta})},\end{aligned}$$ which shows that the right hand sides of equations (\[eq25\]) and (\[eq26\]) are equal, and hence (CA2). This completes the proof that $f_\sigma$ is a cluster automorphism. It is direct because $\sigma Q \cong Q$. This shows that the map $F$ is well-defined. It is easy to see that $F$ is a group homomorphism with kernel ${\rm Stab} \, Q_0.$
The automorphism of the Kronecker quiver $$\xymatrix{1&2\ar@<2pt>[l]\ar@<-2pt>[l]}
$$ which fixes the points and interchanges the arrows lies in the kernel of $F$.
We finally observe that an anti-automorphism of the quiver induces in the same way an inverse cluster automorphism. Let $\mathcal{A}\left( \mathbf{x},Q\right) $ be a cluster algebra and $\sigma $ be an anti-automorphism of $
Q$. We define $f_{\sigma }:\mathbf{x}\rightarrow \mathbf{x}$ by setting $
f_{\sigma }\left( x\right) =x^{\prime }$ where $x^{\prime }$ is the unique cluster variable such that $\sigma \left( p_{x}\right) =p_{x^{\prime }}.$ Then $f_{\sigma }$ is clearly an automorphism of the ambient field $\mathcal{
F}$.
\[prop anti\] With the above notation, the map $
f_{\sigma }$ is an inverse cluster automorphism of $\mathcal{A}
\left( \mathbf{x},Q\right) $.
The proof is entirely similar to that of Proposition \[prop 2.13\] and will be omitted.
\[cor anti new\] [Let $\mathcal{A}$ be a cluster algebra with a seed $(\amas,Q)$ such that $Q$ admits an anti-automorphism. Then ${\textup{Aut}\,\mathcal{A}}=\Autp\rtimes{\mathbb{Z}}_2$.]{}
[In this situation, Proposition 2.15 yields that there exists $\sigma\in{\textup{Aut}\,\mathcal{A}}\setminus{\textup{Aut}^+\mathcal{A}}$ of order two, and the result follows from Corollary \[cor two new\].]{}
\[ex anti atilde\] [The cluster algebras of euclidean type $\tilde{\mathbb{A}}$ admit an inverse cluster automorphism. Indeed, for any cluster algebra of type $\tilde{\mathbb{A}}$, there exist integers $p,q$ such that the cluster algebra has a seed whose quiver is of the following form.]{} $$\xymatrix{&2\ar[r]&3\ar[r]&\cdots\ar[r]&p\ar[rd]\\
1\ar[ur]\ar[dr]&&&&&p+q\\
&p+1\ar[r]&p+2\ar[r]&\cdots\ar[r]&p+q-1\ar[ur]
}$$ [ This quiver has an anti-automorphism given on the points by the permutation that exchanges 1 with $p+q$, $\ell $ with $p+2-\ell$ for $\ell=2,3,\cdots,p$, and $p+m$ with $p+q-m$ for $m=1,2,\ldots,q-1$. Therefore Proposition \[prop anti\] implies that the cluster algebra has an inverse cluster automorphism. ]{}
The acyclic case {#sect_acyclic}
================
The cluster category
--------------------
In this section, we assume that $Q$ is an acyclic quiver. In this case, the combinatorics of cluster variables are encoded in the cluster category. Let $k$ be an algebraically closed field, $kQ$ the path algebra of $Q$ and ${\rm mod} \, kQ$ the category of finitely generated right $kQ$-modules. We denote by ${\rm ind}\, kQ$ a full subcategory of ${\rm mod} \, kQ$ consisting of exactly one object from each isomorphism class of indecomposable $kQ$-modules. For $x \in Q_0$, we denote by $P_x$ the corresponding indecomposable projective $kQ$-module. For properties of ${\rm mod}\, kQ$ and its Auslander-Reiten quiver $\Gamma ({\rm mod}\, kQ)$, we refer the reader to [@ARS; @ASS].
We denote by ${\cal D}^b({\rm mod}\,kQ)$ the bounded derived category over ${\rm mod} \, kQ$. This is a triangulated Krull-Schmidt category having Serre duality and hence almost split triangles. Since $kQ$ is hereditary, the Auslander-Reiten quiver $\Gamma({\cal D}^b({\rm mod}\,kQ))$ of ${\cal D}^b({\rm mod}\,kQ)$ is well-understood [@H]. The cluster category is defined to be the orbit category of ${\cal D}^b({\rm mod}\, kQ)$ under the action of the automorphism $\tau^{-1}[1]$, where $\tau$ is the Auslander-Reiten translation and $[1]$ is the shift of ${\cal D}^b({\rm mod}\, kQ)$, see [@BMRRT]. Then ${\cal C}_Q$ is also a triangulated Krull-Schmidt category having almost split triangles, and the projection functor ${\cal D}^b({\rm mod} \,kQ) \to {\cal C}_Q$ is a functor of triangulated categories commuting with the Auslander-Reiten translation [@K]. Moreover, ${\cal C}_Q$ is a $2$-Calabi-Yau category [@BMRRT]. Let ${\rm ind} \,{\cal C}_Q$ denote a full subcat!
egory of ${\cal C}_Q$ consisting of exactly one object from each isomorphism class of indecomposable objects in ${\cal C}_Q$, then ${\rm ind}\, {\cal C}_Q$ can be identified with the disjoint union of ${\rm ind} \,kQ$ and $kQ[1] = \{P_x[1] | x\in Q_0\}$, the shifts of the indecomposable projective $kQ$-modules. We always use this identification in the sequel. The Auslander-Reiten quiver $\Gamma({\cal C}_Q)$ of ${\cal C}_Q$ is the quotient of $\Gamma({\cal D}^b({\rm mod}\,kQ))$ under the action of the quiver automorphism $\tau^{-1}[1]$. This Auslander-Reiten quiver has always a unique component containing all the objects of $kQ[1]$. This is the *transjective* component of $\Gamma({\cal C}_Q)$ and is denoted by $\Gamma_{tr}$. If $Q$ is a Dynkin quiver, then $\Gamma({\cal C}_Q) \cong \Gamma_{tr}. $ Otherwise, $\Gamma_{tr}$ is isomorphic to the repetition quiver $\Gamma_{tr} \cong {\mathbb Z}Q$ of $Q$ (see [@ASS]), and there are infinitely many so-called *regular*!
components which are either stable tubes (if $Q$ is euclidean) or of type $\mathbb{ZA}_\infty$ (if $Q$ is wild).
Let $n = |Q_0|$. There exists a map $$X_? : ({\cal C}_Q)_0 \to \mathbb Z [x_1^{\pm 1}, \dots, x_n^{\pm 1}]$$ called the *canonical cluster character*, or the *Caldero-Chapoton map*. The map $X_?$ induces a bijection between the $M$ in ${\rm ind}\, {\cal C}_Q$ which have no self-extensions, and the cluster variables $X_M$, see [@CK06]. Under this bijection, the clusters correspond to the so-called [*tilting*]{} objects (also known as cluster-tilting objects) in ${\cal C}_Q$. In practice, the map $X_?$ is difficult to compute explicitly. An easier method for computing the cluster variables is via the frieze functions [@AsRS; @AD; @ADSS].
Cluster automorphisms in the acyclic case
-----------------------------------------
In this section, we prove that if $Q$ is acyclic, then the cluster automorphisms of ${\cal A=A}(\amas, Q)$ are entirely determined by the quiver automorphisms of the transjective component $\Gamma_{tr}$.
\[lemma 3.1\] Let $f$ be a cluster automorphism of ${\cal A}={\cal A}(\amas,Q)$, where $Q$ is acyclic.
- If $f$ is direct, then it induces a triangle equivalence $f_{{\cal D}}: {\cal D}^b({\rm mod} \; kQ) \to {\cal D}^b({\rm mod} \; kQ) $.
- If $f$ is inverse, then it induces a triangle equivalence $f_{{\cal D}}: {\cal D}^b({\rm mod} \; kQ) \to {\cal D}^b({\rm mod} \; kQ^{op}) $.
Let $X_?$ denote as before the canonical cluster character. For each $x$ in $\mathbf{x}
$, there exists a unique indecomposable object $M_{x}$ in the cluster category $\mathcal{C}_{Q}$ such that $f(x)=X_{M_{x}}$. Because $\mathbf{x}$ is a cluster, then $M=\oplus _{x\in \mathbf{x}}M_{x}$ is a tilting object in $\mathcal{C}_{Q}$. [ If $f$ is direct then $\textup{End}\, M\cong kQ$.]{} Therefore, the set $\left\{
M_{x}\mid x\in \mathbf{x}\right\} $ forms a local slice in $
\mathcal{C}_{Q}$. Due to [@ABS2], there exists a slice $\Sigma $ in a transjective component of ${\cal D}^{b}\left( \textup{mod}\,kQ\right) $, isomorphic to $Q$, such that, for each $x\in \mathbf{x}$, the object $M_{x}$ lifts to $\widetilde{M_{x}}$ in $\Sigma$. Because $\widetilde{M}=\oplus _{x\in
\mathbf{x}}\widetilde{M_{x}}$ is a slice complex in ${\cal D}^{b}\left( \textup{mod}\,
kQ\right)$, it is also a tilting complex. Therefore the triangle functor $
f_{\mathcal{D}}=-\otimes _{kQ}^{\mathbb{L}}\widetilde{M}[-1]:\mathcal{D}^{b}\left( \textup{mod}\,kQ\right) \rightarrow \mathcal{D}^{b}\left( \textup{mod}\,kQ\right) $ is a triangle equivalence. [This shows (a). If $f$ is inverse then $\textup{End}\, M\cong kQ^{op}$, and again the set $\left\{
M_{x}\mid x\in \mathbf{x}\right\} $ forms a local slice in $
\mathcal{C}_{Q}$. Again due to [@ABS2], there exists a slice $\Sigma $ in a transjective component of ${\cal D}^{b}\left( \textup{mod}\,kQ\right) $, isomorphic to $Q^{op}$, such that, for each $x\in \mathbf{x}$, the object $M_{x}$ lifts to $\widetilde{M_{x}}$ in $\Sigma$. Because $\widetilde{M}=\oplus _{x\in
\mathbf{x}}\widetilde{M_{x}}$ is a slice complex in ${\cal D}^{b}\left( \textup{mod}\,
kQ\right)$, it is also a tilting complex. Again the triangle functor $
f_{\mathcal{D}}=-\otimes _{kQ}^{\mathbb{L}}\widetilde{M}[-1]:\mathcal{D}^{b}\left( \textup{mod}\,kQ\right) \rightarrow \mathcal{D}^{b}\left( \textup{mod}\,kQ\right) $ is a triangle equivalence. ]{}
Recall that a morphism of translation quivers is a morphism of quivers which commutes with the translation.
Let $f$ be a cluster automorphism of ${\cal A=A}(\amas,Q)$, where $Q$ is acyclic.
- If $f$ is direct, then it induces a quiver automorphism of the transjective component $\Gamma_{tr}$ of $ \Gamma({\cal C}_Q)$.
- If $f$ is inverse, then it induces a quiver anti-automorphism of the transjective component $\Gamma_{tr}$ of $ \Gamma({\cal C}_Q)$.
We only prove (a), because the proof of (b) is similar. Let $f$ be direct. As seen in Lemma \[lemma 3.1\], it induces a triangle equivalence $f_{\mathcal{
D}}:\mathcal{D}^{b}\left( \textup{mod}\,kQ\right) \rightarrow\mathcal{D}^{b}\left( \textup{mod}\,
kQ\right) $ mapping the slice $\mathcal{P}=\left\{ P_{x}\mid
x\in \mathbf{x}\right\} $ consisting of the indecomposable projective $kQ$-modules to the isomorphic slice $\Sigma =\left\{ \widetilde{M_{x}}\mid x\in \mathbf{x}\right\}$, and both slices may be assumed to lie in the same transjective component $\Gamma $ of $\Gamma \left( \mathcal{D}^{b}\left(
\textup{mod}\,kQ\right) \right)$. Clearly, $f_{\mathcal{D}}\left( P_{x}\right) =\widetilde{M_{x}}$, and moreover $f_{\mathcal{D}}$ induces a quiver isomorphism $\mathcal{P}\cong \Sigma ,$ which extends uniquely to a quiver automorphism of $\Gamma$. [On the other hand, $f_{\mathcal{D}}$ is a triangle equivalence, hence commutes with the shift of $ \mathcal{D}^{b}\left(
\textup{mod}\,kQ\right)$ and the Auslander-Reiten translation of $\Gamma$. Therefore it induces an automorphism of the translation quiver $\Gamma_{tr}$. ]{}
\[rem 3.3\] As a direct consequence of the above proofs, if $f$ is nontrivial, then it acts nontrivially on the transjective component.
[To illustrate Remark \[rem 3.3\] we give an example of a cluster automorphism defined by permuting two regular components of the cluster category and show that this automorphism induces a nontrivial action on the transjective component. ]{}
[ Let $(\amas,Q)$ be the seed with cluster $\amas=\{x_1,x_2,x_3,x_4\}$ and quiver $$\xymatrix@R=10pt{&2\ar[ld]\\1\ar@<2pt>[rr]\ar@<-2pt>[rr] &&4\ar[ul]\ar[dl]\ . \\ &3\ar[lu]}$$ This is a seed of type $\tilde {\mathbb{A}}_{2,2}$: indeed, mutating the seed in $x_2$ and then in $x_3$ one gets a new seed with cluster $\amas'=\{x_1,\frac{x_1+x_4}{x_2},\frac{x_1+x_4}{x_3},x_4\}$ and quiver $$\xymatrix@R=10pt{&2\ar[rd]\\1\ar[ru]\ar[rd] &&4\ . \\ &3\ar[ru]}$$ ]{}
[In the corresponding cluster category, each of the objects associated with $x_2$ and $x_3$ sits in the mouth of a tube of rank 2. Therefore permuting the two tubes induces a cluster automorphism given on the cluster $\amas$ by the permutation of the two variables $x_2 $ and $x_3$. On the cluster $\amas'$ this automorphism acts by permutation of the variables $\frac{x_1+x_4}{x_2}$ and $\frac{x_1+x_4}{x_3}$. Therefore the induced action on the transjective component is nontrivial.]{}
We now relate cluster automorphisms to automorphisms of the original quiver. We call a mutation an *APR-mutation* if it is applied to a source or to a sink. The letters APR stand for Auslander, Platzeck and Reiten and evoke the similarity between such mutations and the so-called APR-tilts [@APR]. Equivalently, one may think of APR-mutations as a generalisation of the reflection functors of Bernstein, Gelfand and Ponomarev [@BGP]. We also note that, given an APR-mutation, the new cluster variable is obtained from the old ones by using the frieze function.
For our purposes, it is useful to understand how an APR-mutation translates into terms of the cluster category. Let $\mathcal{A}\left( \mathbf{x},Q\right)
$ be a cluster algebra, with $Q$ acyclic, and $\mu _{x}$ be an APR-mutation corresponding to a source (say) in $Q$. Then $\mu _{x}$ maps $\mathcal{A}\left( \mathbf{x},Q\right) $ to $\mathcal{A}\left( \mu _{x}\mathbf{x},\mu
_{x}Q\right)$. Identifying $Q$ with the full subquiver $kQ[1]=\left\{
P_{x}[1]\mid x\in \mathbf{x}\right\} $ of the transjective component $\Gamma _{tr}$ of $\Gamma \left( \mathcal{C}_{Q}\right) $, the application of $\mu _{x}$ to $kQ[1]$ clearly amounts to replacing $kQ[1]$ by the new tilting object $(kQ[1]\setminus \left\{ P_{x}[1]\right\} )\cup
\left\{ P_{x}\right\} $ whose quiver is also a slice in $\Gamma _{tr}$ , though not isomorphic to $Q$.
\[lemma\_APR\] Let $f$ be a cluster automorphism of ${\cal A} (\amas,Q)$, where $Q$ is acyclic. Then there exists a sequence $\mu$ of APR mutations such that $\mu(\amas)=f(\amas)$ [as sets]{}.
Let $\amas=\{x_1,\ldots,x_n\}$. If $f\in {\textup{Aut}\,\mathcal{A}}$, then $(f(\amas), Q(f(\amas)) $ is a seed such that $Q(f(\amas))$ is isomorphic to $Q$ or $Q^{\textup{op}}$. In particular, $Q(f(\amas))$ is acyclic. Therefore, the cluster $f(\amas)$ corresponds to a tilting object $T=\oplus_{i=1}^n T_i$ in ${\cal C}_Q$ such that the indecomposable summand $T_i$ corresponds to the variable $f(x_i)$ for all $i$ such that $1\le i\le n$. Then $Q(f(\amas))$ is the ordinary quiver of the cluster-tilted algebra ${\rm End_{{\cal C}_Q}}T$, and since $Q(f(\amas))$ is acyclic, it follows that ${\rm End_{{\cal C}_Q}}T$ is hereditary. Therefore, $T$ is a local slice in the transjective component $\Gamma_{tr}$ of $\Gamma({\cal C}_Q)$, see [@ABS2 Corollary 20], and hence there exists a sequence $\mu$ of APR-mutations such that $\mu(\amas)=f(\amas)$ [as sets. Observe that if one lifts the slice $(kQ[1]\setminus \{P_x [1]\}) \cup \{P_x \}$ to a slice $\Sigma$, say, in the derived category, then the endomorphism algebra of $\Sigma$ is obtained from $kQ$ by an APR-tilt (see [@APR]).]{}
Lemma \[lemma\_APR\] states that the action of a cluster automorphism on a cluster is given by a composition of APR-mutations. However, it is not true in general that the automorphism $f$ itself is given by a sequence of APR-mutations, as we show in the following example.
\[ex atilde\] Let $Q$ be the quiver $
\xymatrix{p_{x_1}\ar@/^10pt/[rr]\ar[r]&p_{x_2}\ar[r]&p_{x_3}}$. The transjective component ${\Gamma}_{tr}$ of the Auslander-Reiten is the infinite quiver illustrated in Figure \[fig atilde\]. The position of the initial cluster $\{x_1,x_2,x_3\}$ as well as one other variable $u$ are depicted in the figure. There is a direct cluster automorphism $f\in{\textup{Aut}^+\mathcal{A}}$ defined by $f(x_1)=x_2, f(x_2)=x_3$ and $f(x_3)=u$. Note that $f$ induces an isomorphism of quivers $Q(\amas)\cong Q(f(\amas))$. The corresponding APR-mutation $\mu$ of Lemma \[lemma\_APR\] is the mutation in $x_1$. We have $\mu_{x_1}(\amas)=\{u,x_2,x_3\}=f(\amas)$. But the mutation $\mu$ fixes the variables $x_2,x_3$ and sends $x_1$ to $u$, which shows that the automorphism $f$ is not given by $\mu$. Also note that $\mu$ sends the sink $p_{x_3}$ in the quiver $Q$ to a point $p_{x_3}$ of the quiver $Q(f(\amas))$ which is neither a source nor a sink, whence $\mu$ does not induce a quiver isomorphism between $Q$ and $Q(f(\amas))$ or between $Q^{\textup{op}}$ and $Q(f(\amas))$; consequently, $\mu$ does not induce a cluster automorphism.
$$\xymatrix@R=10pt@C=10pt{
&&&\cdot\ar[rrd]\ar@/_5pt/[rdd]&&&x_3\ar[rrd]\ar@/_5pt/[rdd]&&&\cdot\ar[rrd]\ar@/_5pt/[rdd]&&&\cdot\ar@/_5pt/[rdd]&\\
\cdots&&\cdot\ar[ru]\ar[rrd]&&&x_2\ar[ru]\ar[rrd]&&&\cdot\ar[ru]\ar[rrd]&&&\cdot\ar[ru]\ar[rrd]&&\cdots\\
&\cdot\ar[ru]\ar@/^12pt/[rruu]&&&x_1\ar[ru]\ar@/^12pt/[rruu]&&&u\ar[ru]\ar@/^12pt/[rruu]&&&\cdot\ar[ru]\ar@/^12pt/[rruu]&&&
}$$
We have seen in Examples \[ex\_mutation\] and \[ex atilde\] that not every sequence of APR-mutations is a cluster automorphism. As follows from Lemma \[lemma1\], if a sequence of mutations $\mu$ transforms $Q$ into itself or its opposite $Q^{op}$ in such a way that the bijection $\mu: Q_0 \to \mu(Q)_0$ extends to an isomorphism of quivers of the form either $\mu: Q \to \mu(Q)$ or $\mu:
Q^{op} \to \mu(Q)$, then $\mu$ induces a cluster automorphism $f_{\mu }\in
\textup{Aut}\, \mathcal{A}\left( \mathbf{x},Q\right) $ defined on the initial cluster by $f_{\mu }\left( x_{i}\right) =\mu \left( x_{i}\right) .$ Note that this applies also to quivers which are not acyclic.
Now we are in a position to prove the first main theorem of this section.
\[theorem\_main\] The map $\phi: {\rm Aut}(\Gamma_{tr}) \to \Autp$ defined by $\phi(g)(x_i) = X_{g(P_i[1])}$ for $g\in {\rm Aut}(\Gamma_{tr}) $ and $1\leq i\leq n,$ is a surjective group homomorphism, whose kernel equals the stabiliser ${\rm Stab}(\Gamma_{tr})_0$ of the points in $\Gamma_{tr}$.
\(a) First we prove that the map $\phi$ is well defined. Let $f=\phi(g)$. Now, since $g\in {\rm Aut} (\Gamma_{tr})$, then the local slice $g(kQ[1])$ of $\Gamma_{tr}$ has underlying quiver isomorphic to $Q$. Therefore $(f(\amas),Q(f(\amas))$ is a seed with quiver isomorphic to $Q$. This shows that $f$ satisfies the condition (CA1). Moreover, $f$ satisfies the condition (CA2) because $g$ is an automorphism of ${\Gamma}_{tr}$, and the canonical cluster character map commutes with mutations. This proves that $f$ is a cluster automorphism. Since $g$ is an automorphism of ${\Gamma}_{tr}$, and so preserves the orientation of the arrows between the $P_i[1]$, then $f$ is a direct automorphism. (b) We now prove that $\phi$ is a group homomorphism. Let $g_1,\, g_2 \in {\rm Aut}\, \Gamma_{tr}$, then the following equalities hold: $$\begin{aligned}
\phi(g_1 g_2)(x_i) = X_{g_1g_2(P_i[1])}, \\
\phi(g_1)\phi(g_2)(x_i) = \phi(g_1)(X_{g_2(P_i[1])})\end{aligned}$$ with $1\leq i\leq n.$ We need to establish the equality of these two expressions. Note that $X_{g_2(P_i[1])}$ is a Laurent polynomial which we denote by $L_{2,i}(x_1,\dots,x_n)$ and, similarly, $X_{g_1(P_i[1])}$ is a Laurent polynomial which we denote by $L_{1,i}(x_1,\dots,x_n)$.
It follows directly from the definition of a morphism of translation quivers that $g_1g_2kQ[1]$ is a local slice with quiver isomorphic to $Q$. Thus, denoting by $g_2(i)$ the image in $g_2kQ[1]$ of the point $P_i[1]$, we get $$X_{g_1g_2(P_i[1])} = L_{1,g_2(i)}( L_{2,1}(x_1,\dots,x_n) ,\dots, L_{2,n}(x_1,\dots,x_n)).$$ On the other hand, $g_2kQ[1]$ is also a local slice with quiver isomorphic to $Q$, therefore $$\phi(g_1)X_{g_2(P_i[1])} = L_{1,g_2(i)}( L_{2,1}(x_1,\dots,x_n) ,\dots, L_{2,n}(x_1,\dots,x_n)),$$ which completes the proof.
\(c) We now prove that the kernel of $\phi$ equals ${\rm Stab}(\Gamma_{tr})_0$. Assume that $g, \,g^\prime \in {\rm Aut}\,\Gamma_{tr}$ are such that $\phi(g)=\phi(g^\prime)$. By definition of $\phi$, this means that $g(P_i[1]) = g^\prime(P_i[1])$, for every $i$. Thus $g$ and $g^\prime$ coincide on the initial slice. Therefore $g(M)=g^\prime(M)$ for every indecomposable object $M$ of ${\cal C}_Q$, that is, $g^{-1}g^\prime \in {\rm Stab}(\Gamma_{tr})_0$.
\(d) To show that $\phi$ is surjective, let $f\in\Autp$, where ${\cal A=A}(\amas,Q)$. Then $Q(f(\amas))\cong Q$. Let $M_i$ be an object in ${\rm ind}\, {\cal C}_Q$ such that $X_{M_i} = f(x_i) \in f(\amas)$ with $1\leq i\leq n.$ Due to the isomorphism $Q\cong Q(f(\amas)),$ the $M_i$’s form a local slice in the category ${\cal C}_Q$ and, in particular, $M_i$ lies in the transjective component of ${\cal C}_Q$. The correspondence $P_i[1] \mapsto M_i$ extends to an automorphism $g$ of $\Gamma_{tr}$ (actually to an automorphism of ${\cal C}_Q$) and we get $\phi(g)(x_i) = X_{g(P_i[1])} = X_{M_i} = f(x_i)$ for each $i.$ Therefore $\phi(g)=f.$
\[rem 3.11\] There is always a distinguished automorphism of the transjective component, induced by the Auslander-Reiten translation $\tau $. If $Q$ is a Dynkin quiver, then $\tau $ is periodic so this automorphism is of finite order. If $Q$ is a representation-infinite quiver, then it is of infinite order and $\mathbb{Z}$ is a subgroup of Aut$^{+}\mathcal{A}$.
\[lemma\_reflection\] Let $\mathcal{A}$ be a cluster algebra with a seed $(\amas,Q)$ such that $Q$ is a tree. Then there exists an anti-automorphism $\sigma$ of $\Gamma_{tr}$ given by a reflection with respect to a vertical axis. Moreover, we have $\sigma^2=1$ and $\sigma \tau^m \sigma = \tau^{-m}$ for any $m \in {\mathbb Z}$, where $\tau$ stands for the Auslander-Reiten translation on $\Gamma_{tr}$.
We may assume without loss of generality that the quiver $Q$ of the initial seed is a tree. Let $M$ be an arbitrary point in a transjective component $\Gamma $ of $\Gamma \left( \mathcal{D}^{b}\left( \textup{mod}\,kQ\right) \right)$. We recall that $\Gamma \cong \mathbb{Z}Q$ and we assume fixed a polarisation of $\Gamma $ (see [@ASS p. 131]). We define a reflection $\sigma _{\mathcal{D}}$ “along the vertical axis passing through $M$” in the following way.
There exists a unique slice $\Sigma ^{+}$ in $\Gamma $ having $M$ as its unique source. This slice is the full subquiver of $\Gamma $ consisting of all the points $N$ in $\Gamma $ such that there exists a path from $M$ to $N$ and every such path is sectional. Dually, one constructs the unique slice $\Sigma ^{-}$ in $\Gamma $ having $M$ as its unique sink. Now there exists an obvious bijection between the sets of points of $\Sigma ^{+}$ and $\Sigma
^{-}$, mapping each point of $\Sigma ^{+}$ to the unique point in $\Sigma
^{-}$ lying in the same $\tau$-orbit. Since $Q$ is a tree, the very definition of $\mathbb{Z}Q$ implies that this bijection first extends to an anti-isomorphism between $\Sigma ^{+}$ and $\Sigma ^{-}$ and then to an anti-automorphism $\sigma _{\mathcal{D}}$ of $\Gamma$.
Because $\sigma _{\mathcal{D}}$ is clearly compatible with the functor $\tau
^{-1}[1],$ it induces an anti-automorphism $\sigma $ of the transjective component $\Gamma _{tr}$ of $\Gamma \left( \mathcal{C}_{Q}\right) .$ Finally, the asserted relations for $\sigma $ and $\tau $ follow easily.
\[theorem\_semidirect\] [If $\mathcal{A}$ is a cluster algebra with a seed $(\amas,Q)$ such that $Q$ is a tree then the group of cluster automorphisms is the semidirect product ${\rm Aut} \, {\cal A} = {\rm Aut}^+ {\cal A} \rtimes {\mathbb{Z}}_2$. This product is not direct.]{}
[The anti-automorphism $\sigma $ given by Lemma \[lemma\_reflection\] induces a cluster automorphism $f_\sigma\in{\textup{Aut}\,\mathcal{A}}\setminus{\textup{Aut}^+\mathcal{A}}$ of order two, and, therefore, Corollary \[cor two new\] implies that ${\rm Aut} \, {\cal A} = {\rm Aut}^+ {\cal A} \rtimes {\mathbb{Z}}_2$. This product is not direct because $f_\sigma$ does not commute with $\tau$.]{}
[ Let $\mathcal{A}$ be a cluster algebra of Dynkin or euclidean type. Then ${\textup{Aut}\,\mathcal{A}}=\Autp\rtimes{\mathbb{Z}}_2$.]{}
[ The only case that is not a tree is the euclidean type $\tilde{\mathbb{A}}$, and, for this case, the result follows from Corollary \[cor anti new\] and Example \[ex anti atilde\].]{}
Computing the automorphism groups for quivers of types ${\mathbb A}, {\mathbb D}, {\mathbb E}, \tilde{{\mathbb A}}, \tilde{{\mathbb D}}, \tilde{{\mathbb E}}$ {#sect 3.3}
-------------------------------------------------------------------------------------------------------------------------------------------------------------
Using Theorems \[theorem\_main\] and \[theorem\_semidirect\], we are able to compute the cluster automorphism groups for the cluster algebras of Dynkin and euclidean types explicitly by computing the automorphism groups of the corresponding transjective component $\Gamma_{tr}$ of the cluster category. This computation is straightforward and is done by using the fact that under such an automorphism, the local structure of the quiver is preserved. We refer the reader to [@Ri] for a similar calculation. The results are collected in Table \[table\]. As an example, we discuss the case of the quiver $\tilde{\mathbb D}_n$ in more detail.
[c|c|c]{} Dynkin Type & ${\rm Aut}^+ {\cal A}$ & ${\rm Aut}\, {\cal A}$\
${\mathbb A}_n, n>1$ & ${\mathbb Z}_{n+3}$ & $D_{n+3}$\
${\mathbb A}_1$ & ${\mathbb Z}_{2}$ & $\mathbb{Z}_{2}$\
${\mathbb D}_n,\;n>4$ & ${\mathbb Z}_n\times {\mathbb Z}_2$ & ${\mathbb Z}_n\times {\mathbb Z}_2 \rtimes {\mathbb Z}_2$\
${\mathbb D}_4$ & ${\mathbb Z}_4 \times S_3$ & $D_4 \times S_3$\
${\mathbb E}_6$ & ${\mathbb Z}_{14}$& $D_{14}$\
${\mathbb E}_7 $ &${\mathbb Z}_{10}$& $D_{10}$\
$ {\mathbb E}_8$ & ${\mathbb Z}_{16}$& $D_{16}$\
\
Euclidean Type & ${\rm Aut}^+ {\cal A}$ & ${\rm Aut}\, {\cal A}$\
$\tilde{\mathbb A}_{p,q},\;p\neq q$ & $H_{p,q}$ & $H_{p,q} \rtimes {\mathbb Z}_2$\
$\tilde{\mathbb A}_{p,p}$ & $H_{p,p}\rtimes {\mathbb Z}_2$ & $H_{p,p} \rtimes {\mathbb Z}_2 \rtimes {\mathbb Z}_2$\
$\tilde{\mathbb D}_{n-1}, \; n\neq5$ & $G$ & $G\rtimes {\mathbb Z}_2$\
$\tilde{\mathbb D}_4$ & ${\mathbb Z} \times S_4$ & ${\mathbb Z} \times S_4\rtimes {\mathbb Z}_2$\
$ \tilde{\mathbb E}_6$ & ${\mathbb Z} \times S_3$ & ${\mathbb Z} \times S_3 \rtimes {\mathbb Z}_2$\
$\tilde{\mathbb E}_7$ & ${\mathbb Z} \times {\mathbb Z}_2$ & $({\mathbb Z}\times {\mathbb Z}_2) \rtimes {\mathbb Z}_2$\
$\tilde{\mathbb E}_8$ & ${\mathbb Z}$ & ${\mathbb Z} \rtimes {\mathbb Z}_2$
\[table\]
The notation in Table \[table\] is as follows: $D$ stands for the dihedral group, $H_{p,q} = \langle r_1, r_2| r_1r_2 = r_2 r_1, r_1^p=r_2^q \rangle$ and $G$ is given by equation (\[G\]) below.
Note that for the type $\mathbb{A}_1$, the quiver $Q$ has no arrows, and therefore there are no anti-automorphisms in this case. Moreover, the cluster algebra has exactly two clusters, each consisting of a single cluster variable and the Auslander-Reiten translation in the cluster category is of order 2.
In the case of $\tilde{\mathbb{A}}_{p,q}$, the automorphisms $r_{1}$ and $
r_{2}$ are defined as follows. Let $\Sigma ^{+}$, as in the proof of Lemma \[lemma\_reflection\], be the unique slice in the transjective component having $M$ as its only source. Let $\omega $ denote the unique path of length $p$ from $M$ to the only sink in $\Sigma^+$. The automorphism $r_{1}$ is defined as follows. We send $M$ to the target $N$ of the only arrow on $\omega $ with source $M$, and send $\Sigma
^{+}$ isomorphically to the unique slice having $N$ as its only source. This map extends to an automorphism $r_{1}$ of the transjective component. We define $r_{2}$ similarly, using the path of length $q$ from $M$ to the sink of $\Sigma ^{+}$. Note that $\tau^{-1} =r_{1}r_{2}.$
Consider a cluster algebra ${\cal A=A}(\amas, Q)$, where $Q$ is the following euclidean quiver of type $\widetilde{\mathbb D}_{n-1}$ with $n\neq 5$ (the number of points is $n$).
$$\xymatrix@R10pt{
1\ar[rd]&&&&&&n-1\\
&3\ar[r]&4\ar[r]&\quad\cdots\quad\ar[r]&n-3\ar[r]&n-2\ar[ru]\ar[rd]\\
2\ar[ru]&&&&&&n
}
$$
The transjective component $\Gamma_{tr}$ is represented by the infinite translation quiver $\mathbb{Z}Q$, an example with $n=8$ is given in Figure \[fig 22\].
$$\xymatrix@R10pt@C10pt@!{&&\cdot\ar[rd]&&\cdot\ar[rd]&&1\ar[rd] &&\cdot\ar[rd] &&\cdot\ar[rd]\\
&\cdot \ar[ru]\ar[r]\ar[rd]&\cdot\ar[r]&\cdot \ar[ru]\ar[r]\ar[rd] & \cdot \ar[r] &{\sigma}6 \ar[ru]\ar[r]\ar[rd] & 2
\ar[r] &3 \ar[ru]\ar[r]\ar[rd] & \cdot \ar[r] &\cdot \ar[ru]\ar[r]\ar[rd] & \cdot \ar[r] &\cdot\\
\cdots&&\cdot \ar[ru]\ar[rd] && {\sigma}5\ar[ru]\ar[rd] &&\cdot \ar[ru]\ar[rd] &&4\ar[ru]\ar[rd] &&\cdot \ar[ru]\ar[rd]&& \cdots \\
&\cdot\ar[ru]\ar[rd] &&{\sigma}4\ar[ru]\ar[rd] &&\cdot\ar[ru]\ar[rd] &&\cdot\ar[ru]\ar[rd] &&5\ar[ru]\ar[rd] &&\cdot\\
&{\sigma}2\ar[r] &{\sigma}3 \ar[ru]\ar[rd]\ar[r] & \cdot\ar[r] & \cdot \ar[ru]\ar[rd]\ar[r] & \cdot\ar[r] & \cdot \ar[ru]\ar[rd]\ar[r] & \cdot\ar[r] & \cdot \ar[ru]\ar[rd]\ar[r] & \cdot\ar[r] & 6 \ar[ru]\ar[rd]\ar[r] & 7 &\\
&{\sigma}1\ar[ru] &&\cdot\ar[ru] &&\cdot\ar[ru] &&\cdot\ar[ru] &&\cdot\ar[ru] &&8
}$$
Recall that automorphisms of a translation quiver are by definition the quiver automorphisms commuting with the translation $\tau$. Besides $\tau$ we introduce three more generators of the automorphism group.
Let $\rho_1$ denote the automorphism which interchanges the corresponding points in the $\tau$-orbits of $1$ and $2$ in ${\Gamma}_{tr}$ and fixes all other points. Let $\rho_n$ denote the automorphism which interchanges the corresponding points in the $\tau$-orbits of $n-1$ and $n$ in ${\Gamma}_{tr}$ and fixes all other points. Finally, let ${\sigma}$ be the automorphism given by the translation of the plane that sends the point $n$ to the point $1$ followed by the reflection with respect to the horizontal line through the point $1$; we have indicated the action of ${\sigma}$ on the points $1,2,\ldots,6$ in Figure \[fig 22\], and ${\sigma}8=1$ and ${\sigma}7=2$.
Since for every point of the quiver $kQ$ the number of incoming and outgoing arrows is preserved under a quiver automorphism, one sees that every automorphism of $kQ$ can be expressed as a combination of $\tau,\rho_1,\rho_n$, and ${\sigma}.$
We note the following relations between these generators:
1. the translation $\tau$ commutes with all automorphisms and is of infinite order;
2. $\rho_1$ and $\rho_n$ are of order two and commute with each other;
3. ${\sigma}^2=\tau^{n-3}$;
4. $\rho_1{\sigma}={\sigma}\rho_n$ and ${\sigma}\rho_1=\rho_n{\sigma}$.
Thus we get a presentation of the group of automorphisms of ${\Gamma}_{tr}$ as $$\label{G}
G=\left\langle \tau,{\sigma},\rho_1,\rho_n \left| \begin{array}{c}
\rho_i^2=1 ,\tau \rho_i=\rho_i \tau \ (i=1,n)\\
\tau{\sigma}={\sigma}\tau,\ {\sigma}^2=\tau^{n-3} \\
\rho_1{\sigma}={\sigma}\rho_n, \ {\sigma}\rho_1=\rho_n{\sigma}\end{array}\right.\right\rangle$$
[Cluster algebras from surfaces]{}\[sect 3\] Following [@FST], we describe the construction of cluster algebras from surfaces.
Let $S$ be an oriented Riemann surface with or without boundary, and let $M\subset S$ be a finite set of marked points such that $M$ contains at least one point of every connected component of the boundary. If the boundary is empty then the surface is said to be *closed*. Points in $M$ that are in the interior of $S$ are called *punctures*. We call the pair $(S,M)$ simply a *surface*.
For technical reasons, we require that $(S,M)$ is not a sphere with one, two or three punctures; a disc with one, two or three marked points on the boundary; or a punctured disc with one marked point on the boundary. Some simple examples of surfaces are given in Table \[table 1\].
g b c p surface type
------- --------- --------- ------- -------------------------- ----------------------------
0 1 n+3 0 polygon $\mathbb{A}_n$
0 1 n 1 once punctured polygon $\mathbb{D}_n$
0 1 n-3 2 twice punctured polygon $\tilde{\mathbb{D}}_{n-1}$
0 2 n 0 annulus $\tilde{\mathbb{A}}_{n-1}$
0 2 n-3 1 punctured annulus not acyclic
0 0 0 4 sphere with 4 punctures not acyclic
1 0 0 1 punctured torus not acyclic
3 0 n-3 0 pair of pants not acyclic
: Examples of surfaces, $g$ is the genus, $b$ is the number of boundary components, $p$ the number of punctures, $c$ the number of marked points on the boundary and $n$ is the rank of the cluster algebra.[]{data-label="table 1"}
Arcs and triangulations
-----------------------
An *arc* ${\gamma}$ in $(S,M)$ is the isotopy class of a curve in $S$ such that
- the endpoints of the curve are in $M$;
- the curve does not cross itself, except that its endpoints may coincide;
- except for the endpoints, the curve is disjoint from $M$ and from the boundary of $S$,
- the curve does not cut out an unpunctured monogon or an unpunctured bigon.
For any two arcs ${\gamma},{\gamma}'$ in $S$, let $e({\gamma},{\gamma}')$ be the minimal number of crossings of curves ${\alpha}$ and ${\alpha}'$, where ${\alpha}$ and ${\alpha}'$ range over all curves in the isotopy classes ${\gamma}$ and ${\gamma}'$, respectively. We say that two arcs ${\gamma}$ and ${\gamma}'$ are *compatible* if $e({\gamma},{\gamma}')=0$.
An *ideal triangulation* is a maximal collection of pairwise compatible arcs. The arcs of an ideal triangulation cut the surface into *ideal triangles*.
![Two ideal triangulations of a punctured annulus related by a flip of the arc $6$. The triangulation on the right hand side has a self-folded triangle.[]{data-label="fig 2"}](fig2.eps)
Examples of ideal triangulations are given in Figure \[fig 2\]. There are two types of ideal triangles: triangles that have three distinct sides and triangles that have only two. The latter are called *self-folded* triangles. For example, the triangle formed by the arcs $6$ and $1$ on the right hand side of Figure \[fig 2\] is self-folded. In a self-folded triangle the arc incident to the puncture is called *radius* and the other arc is called *loop*.
Ideal triangulations are connected to each other by sequences of [*flips*]{}. Each flip replaces a single arc $\gamma$ in a triangulation $T$ by the (unique) arc $\gamma' \neq \gamma$ that, together with the remaining arcs in $T$, forms a new ideal triangulation.
Note that an arc $\gamma$ that is the radius inside a self-folded triangle in $T$ cannot be flipped.
In [@FST], the authors associated a cluster algebra to any bordered surface with marked points. Roughly speaking, the cluster variables correspond to arcs, the clusters to triangulations, and the mutations to flips. However, because arcs inside self-folded triangles cannot be flipped, the authors were led to introduce the slightly more general notion of [*tagged arcs*]{}. They showed that ordinary arcs can all be represented by tagged arcs and gave a notion of flip that applies to all tagged arcs.
A [*tagged arc*]{} is obtained by taking an arc that does not cut out a once-punctured monogon and marking (“tagging") each of its ends in one of two ways, [*plain*]{} or [*notched*]{}, so that
- each end connecting to a marked point on the boundary of $S$ must be tagged plain;
- if both ends of an arc connect to the same point then they must be tagged in the same way.
\[compatible\] Two tagged arcs $\alpha$ and $\beta$ are called [*compatible*]{} if the arcs $\alpha^0$ and $\beta^0$ obtained from $\alpha$ and $\beta$ by forgetting the taggings are compatible and
- if $\alpha^0=\beta^0$ then at least one end of $\alpha$ must be tagged in the same way as the corresponding end of $\beta$;
- if $\alpha^0\neq \beta^0$ but they share an endpoint $a$, then the ends of $\alpha$ and $\beta$ connecting to $a$ must be tagged in the same way.
One can represent an ordinary arc $\beta$ by a tagged arc $\iota(\beta)$ as follows. If $\beta$ does not cut out a once-punctured monogon, then $\iota(\beta)$ is simply $\beta$ with both ends tagged plain. Otherwise, $\beta$ is a loop based at some marked point $a$ and cutting out a punctured monogon with the sole puncture $b$ inside it. Let $\alpha$ be the unique arc connecting $a$ and $b$ and compatible with $\beta$. Then $\iota(\beta)$ is obtained by tagging $\alpha$ plain at $a$ and notched at $b$. Figure \[figtags\] shows the tagged triangulation corresponding to the triangulation on the right hand side of Figure \[fig 2\]. The notching is indicated by a bow tie.
A maximal collection of pairwise compatible tagged arcs is called a [*tagged triangulation*]{}.
![Tagged triangulation of the punctured annulus corresponding to the ideal triangulation of the right hand side of Figure \[fig 2\].[]{data-label="figtags"}](figtags.eps)
We are now ready to define the cluster algebra associated to the surface. For that purpose, we choose an ideal triangulation $T$ and then define a quiver $Q_T$ without loops or 2-cycles, or, equivalently, a skew-symmetric integer matrix $B_T$. Let $\tau_1,\tau_2,\ldots,\tau_n$ be the $n$ arcs of $T$. For any triangle $\Delta$ in $T$ which is not self-folded, we define a matrix $B^\Delta=(b^\Delta_{ij})_{1\le i\le n, 1\le j\le n}$ as follows.
- $b_{ij}^\Delta=1$ and $b_{ji}^{\Delta}=-1$ in each of the following cases:
- $\tau_i$ and $\tau_j$ are sides of $\Delta$ with $\tau_j$ following $\tau_i$ in the clockwise order;
- $\tau_j$ is a radius in a self-folded triangle enclosed by a loop $\tau_\ell$, and $\tau_i$ and $\tau_\ell$ are sides of $\Delta$ with $\tau_\ell$ following $\tau_i$ in the clockwise order;
- $\tau_i$ is a radius in a self-folded triangle enclosed by a loop $\tau_\ell$, and $\tau_\ell$ and $\tau_j$ are sides of $\Delta$ with $\tau_j$ following $\tau_\ell$ in the clockwise order;
- $b_{ij}^\Delta=0$ otherwise.
Then define the matrix $ B_{T}=(b_{ij})_{1\le i\le n, 1\le j\le n}$ by $b_{ij}=\sum_\Delta b_{ij}^\Delta$, where the sum is taken over all triangles in $T$ that are not self-folded.
Note that $B_{T}$ is skew-symmetric and each entry $b_{ij}$ is either $0,\pm 1$, or $\pm 2$, since every arc $\tau$ is in at most two triangles. We associate a quiver $Q_T$ to the matrix $B_T$ as follows. The points of $Q_T$ are labeled by $1,2,\ldots,n$ and the number of arrows from $i$ to $j$ equals $b_{ij}$, with the convention that if $b_{ij}$ is a negative number, then having $b_{ij}$ arrows from $i$ to $j$ means having $|b_{ij}|$ arrows from $j$ to $i$. For example, the quiver corresponding to the triangulation on the right of Figure \[fig 2\] is $$\xymatrix@C50pt@R10pt{1\ar[rd]&&3\ar[rd]\ar[dd]\\&2\ar[ru]&&4\\
6\ar[ru]&&5\ar[ul]\ar[ru]}$$ Since the matrix $B_T$ is skew-symmetric, it follows that $Q_T$ has no oriented cycles of length at most two.
The cluster algebra $\mathcal{A}=\mathcal{A}(\mathbf{x},Q_T)$ given by the quiver $Q_T$ is said to be the *cluster algebra* (with trivial coefficients) *associated to the surface $(S,M)$*.
\[remtag\] [It has been shown in [@FST] that]{} if the surface $(S,M)$ is not a closed surface with exactly one puncture then there is a bijection between tagged arcs in $(S,M)$ and cluster variables in the cluster algebra, such that compatible tagged arcs correspond to compatible cluster variables, and tagged triangulations correspond to clusters. The mutations in the cluster algebra are given by the flips in the tagged triangulations.
If $(S,M)$ is a closed surface with exactly one puncture, for example a once punctured torus, then there is a bijection between arcs (not tagged arcs) and cluster variables. The reason for this is that a flip can not change the tagging at the endpoint of a given arc because all arcs are incident to the unique puncture, thus changing the tagging on one of the arcs would not be compatible with the others.
This fact will be important when we consider the cluster automorphisms induced by change of taggings, see Lemma \[lemtag\] and Theorem \[theorem mg\].
Mapping class groups {#sect 4.2}
--------------------
In this section, we give the definitions and some basic properties of mapping class groups. For further details we refer the reader to [@FM].
Let $\textup{Homeo}^+(S)$ be the group of orientation preserving homeomorphisms from $S$ to $S$ and let $\textup{Homeo}^+(S,\partial S)$ be the subgroup of all $f\in \textup{Homeo}^+(S)$ such that the restriction $f|_{\partial S}$ of $f$ to the boundary is equal to the identity $1_{\partial S}$.
Two homeomorphisms $f,g$ of $S$ are *isotopic* if there is a continuous function $H:S\times [0,1]\to S$ such that $H(x,0)=f$ and $H(x,1)=g$ and such that for each $t\in [0,1]$ the map $H(x,t):S\to S$ is a homeomorphism.
Let $\textup{Homeo}_0(S,\partial S)$ be the subgroup of all $f\in \textup{Homeo}^+(S,\partial S)$ that are isotopic to the identity $1_S$ via an isotopy $H$ fixing $\partial S$ pointwise, thus $H(x,t)=x$ for all $x\in \partial S$ and $t\in [0,1]$.
The *mapping class group* ${\mathcal{M}od\,S}$ of the surface $S$ is defined as the quotient group $${\mathcal{M}od\,S}=\textup{Homeo}^+(S,\partial S)/\textup{Homeo}_0(S,\partial S).$$
We now define the mapping class group of the surface with marked points $(S,M)$ in a similar way. Let ${\textup{Homeo}^+(S,M)}$ be the group of orientation preserving homeomorphisms from $S$ to $S$ which map $M$ to $M$. Note that we do *not* require that the points in $M$ are fixed, neither that the points on the boundary of $S$ are fixed, nor that each boundary component is mapped to itself. However if a boundary component $C_1$ is mapped to another component $C_2$ then the two components must have the same number of marked points. We say that a homeomorphism $f$ is *isotopic to the identity relative to $M$*, if $f$ is isotopic to the identity via an isotopy $H$ that fixes $M$ pointwise, thus $H(x,t)=x$ for all $x\in M$ and $t\in [0,1]$. Let ${\textup{Homeo}_0(S,M)}$ be the subgroup of all $f\in {\textup{Homeo}^+(S,M)}$ that are isotopic to the identity relative to $M$.
We define the *mapping class group* ${\mathcal{MG}(S,M)}$ of the surface $(S,M)$ to be the quotient $${\mathcal{MG}(S,M)}={\textup{Homeo}^+(S,M)}/{\textup{Homeo}_0(S,M)}.$$
The two mapping class groups are related as follows.
\[mgg\] ${\mathcal{M}od\,S}$ is isomorphic to a subgroup of ${\mathcal{MG}(S,M)}$.
Clearly $\textup{Homeo}_0(S,\partial S)$ is a subgroup of ${\textup{Homeo}_0(S,M)}$, and thus there is a map from ${\mathcal{M}od\,S}$ to ${\mathcal{MG}(S,M)}$ sending the class of a homeomorphism $f$ in ${\mathcal{M}od\,S}$ to its class in ${\mathcal{MG}(S,M)}$. This map is an injective group homomorphism.
Next we review *Dehn twists*. If $S$ is an annulus, we can parametrise $S$ as $S^1\times [0,1]$, where $S^1$ denotes the circle of radius one, such that the two boundary components of $S$ are $S^1\times\{0\}$ and $S^1\times\{1\}$. Then the map $T:S^1\times [0,1]\to S^1\times[0,1], (\theta,t)\mapsto (
\theta +2\pi t, t)$ is an orientation preserving homeomorphism that fixes both boundary components of $S$ pointwise. $T$ is called *Dehn twist* on $S$. Thus $T\in{\textup{Homeo}^+(S,\partial S)}$, and since $T$ is not isotopic to the identity relative to $\partial S$, the class of $T$ in ${\mathcal{M}od\,S}$ is non-zero. It is clear that the class of $T$ has infinite order in ${\mathcal{M}od\,S}$, and hence it generates an infinite cyclic subgroup of ${\mathcal{M}od\,S}$. Since $S$ is an annulus, one can show that the class of $T$ actually generates the whole group ${\mathcal{M}od\,S}$. One can think of this Dehn twist as cutting the annulus along the equator $S^1\times\{1/2\}$, performing a full rotation of one end (keeping the boundary fixed) and gluing the two pieces back together, see Figure \[fig dehn\].
![Dehn twist on the annulus, the curve a is mapped to the curve a’; the equator is drawn as a dashed line[]{data-label="fig dehn"}](dehn1.eps)
Now suppose that $S$ is any Riemann surface, and that $c$ is a closed simple curve in $S$. Then one can define a Dehn twist about $c$ in $S$ by performing the Dehn twist $T$ on a regular neighbourhood $N$ of $c$ in $S$ which is homeomorphic to an annulus, see Figure \[fig regular\].
![Regular neighborhood of a closed curve $c$[]{data-label="fig regular"}](regular.eps)
\[rem mg1\] If the surface $S$ has genus at least one or if $S$ has genus zero and at least two boundary components, then there exists a Dehn twist that generates an infinite cyclic subgroup of ${\mathcal{M}od\,S}$.
\[rem mg2\] We list the mapping class groups of some Riemann surfaces:
S disc annulus punctured disc torus
--------------------- ------ ---------------- ---------------- ------------------------------- --
$\mathcal{M}od (S)$ $0$ ${\mathbb{Z}}$ $0$ $\textup{SL}(2,{\mathbb{Z}})$
S $\textup{sphere with $3$ punctures}$ $\textup{disc with $p$ punctures}$
--------------------- -------------------------------------- ------------------------------------ -- -- --
$\mathcal{M}od (S)$ $S_3$ $B_p$
where $S_3$ denotes the symmetric group on $3$ letters and $B_p$ the braid group on $p$ strands.
In general mapping class groups are very difficult to compute and known only for a few cases.
Marked mapping class group
--------------------------
In order to describe the cluster automorphism group of a cluster algebra corresponding to a marked surface $\left( S,M\right) $, we need a group that contains the mapping class group ${\mathcal{MG}(S,M)}$ from the previous subsection, but also contains automorphisms that change the taggings at the punctures. We call this group the *marked mapping class group.*
A *marked mapping class* $(\bar f, \cal P) $ is an element $\bar f\in {\mathcal{MG}(S,M)}$ together with a subset $\cal P$ of the set of punctures of $(S,M)$.
If the set $\cal P$ consists of a single element $z$, then we write $(\bar f, z)$ instead of $(\bar f, \{z\})$ for the marked mapping class.
A marked mapping class acts on the set of arcs of the surface by applying the homeomorphism $f$ and changing the tagging at each puncture in the set $\mathcal P$. We can define a product on the set of marked mapping classes by $$(\bar f_1,\mathcal{P}_1)(\bar f_2, \mathcal {P}_2) =(\bar f_1\bar f_2, \mathcal {P}_1\ominus f_1(\mathcal {P}_2)),$$ where $\ominus $ denotes the symmetric difference $A\ominus B= (A\cup B)\setminus (A\cap B)$.
The set of marked mapping classes forms a group under the product above.
Associativity follows from the associativity of the symmetric difference, the identity is given by $(1, \emptyset)$, where $1$ denotes the identity of ${\mathcal{MG}(S,M)}$, and inverses are given by $(\bar f,\mathcal{P})^{-1}=(\bar f^{-1},f^{-1}(\mathcal{P}))$.
The *marked mapping class group* ${\mathcal{MG}_{\bowtie}(S,M)}$ of the surface $(S,M)$ is the group of all marked mapping classes of $(S,M)$.
We can also define ${\mathcal{MG}_{\bowtie}(S,M)}$ as a semidirect product as follows. Let $\{z_1,\ldots,z_p\}$ be the set of punctures and let $\mathcal{Z}$ be the group of powersets of $\{z_1,\ldots,z_p\}$ with respect to the group operation $\ominus$. Note that $\mathcal{Z}\cong{\mathbb{Z}}_2^p$. For each ${\overline{f}}\in {\mathcal{MG}(S,M)}$, the homeomorphism $f$ induces an automorphism of $\mathcal{Z}$. This defines an action of ${\mathcal{MG}(S,M)}$ on $\mathcal{Z}$.
\[lem sd\] ${\mathcal{MG}_{\bowtie}(S,M)}$ is isomorphic to the semidirect product $\mathcal{Z}\rtimes {\mathcal{MG}(S,M)}$ with respect to the above action.
The product in $\mathcal{Z}\rtimes{\mathcal{MG}(S,M)}$ is defined as $$({\overline{f}}_1,\mathcal{P}_1)({\overline{f}}_2,\mathcal{P}_2)=({\overline{f}}_1{\overline{f}}_2,\mathcal{P}_1\ominus f_1(\mathcal{P}_2)),$$ which proves the statement.
\[rem 38\] If the surface has precisely one puncture, then for each ${\overline{f}}\in {\mathcal{MG}(S,M)}$ the homeomorphism $f$ must fix the puncture, hence the action of ${\mathcal{MG}(S,M)}$ on $\mathcal{Z}$ is trivial, whence ${\mathcal{MG}_{\bowtie}(S,M)}$ is the direct product $\mathcal{Z}\times {\mathcal{MG}(S,M)}$.
\[cor 37\]
- ${\mathcal{MG}(S,M)}$ is a subgroup of ${\mathcal{MG}_{\bowtie}(S,M)}$;
- $\mathcal{Z}$ is a normal subgroup of ${\mathcal{MG}_{\bowtie}(S,M)}$;
- ${\mathcal{MG}_{\bowtie}(S,M)}$ is generated by the elements $(\bar f, z)$ where $\bar f$ runs over all elements of ${\mathcal{MG}(S,M)}$ and $z$ runs over all punctures;
- $(1,z)(1,z)=(1,\emptyset)$, more generally $(1,\mathcal{ P})^m=(1, \emptyset)$ if $m$ is even, and $(1,\mathcal{ P})^m=(1,\mathcal{P})$ if $m$ is odd;
- If $z,z'$ are two punctures such that $f$ maps $z$ to $z'$, then $$\big(1, z'\big)\,\big(\bar f,\emptyset\big) = \big(\bar f,\emptyset\big)\,\big(1,z\big).$$
(1),(2) and (3) are direct consequences of Lemma \[lem sd\], and (4) and (5) are easy computations.
Mapping class group and cluster automorphism group
--------------------------------------------------
We now show that the group of cluster automorphisms has a subgroup isomorphic to ${\mathcal{MG}_{\bowtie}(S,M)}$. The change of tagging induces a cluster automorphism which is described in the following Lemma.
Let $z$ be a puncture in $(S,M)$ and ${\alpha}$ any arc. We denote by ${\alpha}^z$ the arc that is isotopic to ${\alpha}$ and has the opposite tagging at each endpoint that is equal to $z$. Essentially there are three different cases which are illustrated in Figure \[figtag\], namely ${\alpha}$ can have one endpoint, both endpoints or no endpoint equal to $z$.
\[lemtag\] Assume $(S,M)$ is not a closed surface with exactly one puncture, and assume $T$ is a triangulation of $(S,M)$. Then, for every puncture $z$ of $(S,M)$, the automorphism $\psi_z:\mathcal{A}\to\mathcal{A}$ defined by $\psi_z(x_\tau)=x_{\tau^z}$, for every arc $\tau\in T$, where $x_\tau$ is the cluster variable corresponding to $\tau$, and extended to the other cluster variables by the algebra homomorphism properties, is a cluster automorphism in ${\textup{Aut}^+\mathcal{A}}$.
The cluster algebra $\mathcal{A}$ has an initial seed $(\mathbf{x}_T,B_T)$ associated to the triangulation $T$. As noted in [@FST Definition 9.2], the compatibility of tagged arcs is invariant with respect to a simultaneous change of all tags at a given puncture, therefore the set $T'=\{\tau_1^z,\ldots,\tau_n^z\}$ also is a triangulation, and hence it defines another seed $(\mathbf{x}_{T'},B_{T'})$ of $\mathcal{A}$, where $$B_{T'}=\left(b_{\tau_i^z\tau_j^z}\right)_{ij}=\left(b_{\tau_i\tau_j}\right)_{ij} =B_T.$$ The automorphism $\psi_z$ sends the seed $(\mathbf{x}_T,B_T)$ to the seed $(\mathbf{x}_{T'},B_{T'})$, which shows (CA1). Moreover, since $B_{T^{^{\prime }}}=B_{T},$ the quivers corresponding to these two matrices are equal. By Lemma \[lemma1\], $\psi _{z}\in $ Aut$^{+}\mathcal{A}$.
In the case of a closed surface with exactly one puncture, $\psi_z$ is not defined. Indeed, in this case all arcs of a triangulation start and end at the puncture and thus must be tagged in the same way. Therefore flips do not change the tagging.
\[theorem mg\] Let $(S,M)$ be a surface with $p$ punctures. Then ${\mathcal{MG}(S,M)}$ is isomorphic to a subgroup of ${\textup{Aut}^+\mathcal{A}}$. If moreover $p\ge 2$, or if $\partial S\ne\emptyset$, then ${\mathcal{MG}_{\bowtie}(S,M)}$ is isomorphic to a subgroup of ${\textup{Aut}^+\mathcal{A}}$.
We start by showing that ${\mathcal{MG}(S,M)}$ is isomorphic to a subgroup of ${\textup{Aut}^+\mathcal{A}}$. Fix a triangulation $T$ and let $\mathbf{x}_T$ be the corresponding cluster. Denote the elements of $T$ by $\tau_1,\ldots,\tau_n$ and those of $\mathbf{x}_T$ by $x_{\tau_1},\ldots,x_{\tau_n}$. Then $\{x_{\tau_1},\ldots,x_{\tau_n}\}$ is a transcendence basis of the ambient field $\mathcal{F}$ of the cluster algebra.
For any element $f\in {\textup{Homeo}^+(S,M)}$, let ${\overline{f}}\in{\mathcal{MG}(S,M)}$ denote its class in the mapping class group. Define a map $\phi:{\mathcal{MG}(S,M)}\to {\textup{Aut}^+\mathcal{A}}$ by letting $\phi({\overline{f}})$ be the map from $\mathcal{A}$ to $\mathcal{A}$ defined on the basis by $\phi({\overline{f}})(x_{\tau_i})=x_{f(\tau_i)}$ and extended to $\mathcal{A}$ by the algebra homomorphism properties. We show that $\phi$ is an injective group homomorphism.
To show that the definition of $\phi $ does not depend on the choice of the representative $f$, assume that $f$ is isotopic to the identity relative to $M$. Then $f(\tau_i)$ is isotopic to $\tau_i$ relative to $M$, and thus, $f(\tau_i)$ and $\tau_i$ represent the same arc. It follows that $\phi$ does not depend on the choice of $f$.
Next we show that $\phi({\overline{f}})$ is a cluster automorphism. Since $f$ is a homeomorphism, any two arcs ${\alpha},{\beta}$ in $(S,M)$ which are compatible, have compatible images $f({\alpha}),f({\beta})$ in $(S,M)$. Therefore any triangulation of $(S,M)$ is mapped under $f$ to a triangulation of $(S,M)$. Thus $\phi \left( \overline{f}\right) $ maps clusters to clusters. By Corollary \[cor2.7bis\], $\phi \left( \overline{f}\right) $ is a cluster automorphism. Moreover, since $f$ is actually an orientation preserving homeomorphism, we have $\phi({\overline{f}})\in {\textup{Aut}^+\mathcal{A}}$, and this shows that $\phi$ is well-defined.
Now we show that the definition of $\phi$ does not depend on the choice of the triangulation $T$. That is, let us show that for any arc ${\alpha}$ in $(S,M)$ $$\label{stern}
x_{f({\alpha})}=\phi({\overline{f}})(x_{\alpha}).$$ Indeed, let ${\alpha}$ be an arc in $(S,M)$. Then there is a sequence of flips $\mu_T$ such that $\mu_T\,\tau={\alpha}$ for some $\tau\in T$. Let ${\overline{\mu}}$ denote the corresponding sequence of mutations in $\mathcal{A}$. Then $x_{f({\alpha})}= x_{f(\mu_T\,\tau)}=x_{\mu_{f(T)}f(\tau)}$, where the last identity holds because $f$ commutes with flips. Since flips correspond to mutations, we get $x_{f({\alpha})}= {\overline{\mu}}_{\phi({\overline{f}})(\mathbf{x}_T)}x_{f(\tau)}$, which by definition of $\phi({\overline{f}})$ is equal to ${\overline{\mu}}_{\phi({\overline{f}})(\mathbf{x}_T)}\phi({\overline{f}})(x_{\tau})$. Now using the fact that $\phi({\overline{f}})$ is a cluster automorphism, we get $x_{f({\alpha})}= \phi({\overline{f}})({\overline{\mu}}_{\mathbf{x}_T}(x_\tau))$, which is equal to $\phi({\overline{f}})(x_{\mu_{T}(\tau)})=\phi({\overline{f}})(x_{\alpha})$, because flips correspond to mutations.
To show that $\phi$ is a group homomorphism, let ${\overline{f}},{\overline{g}}\in {\mathcal{MG}(S,M)}$, then for any arc ${\alpha}$, $$\phi({\overline{f}}\circ{\overline{g}})(x_{\alpha})=x_{f\circ g ({\alpha})} =\phi({\overline{f}})(x_{g({\alpha})})=\phi({\overline{f}})\big(\phi({\overline{g}})(x_{\alpha})\big).$$
Finally, we show that $\phi $ is injective. Let ${\overline{f}}\in\textup{Ker}\,\phi$. Then $\phi({\overline{f}})=1_{\mathcal{A}}$, and for any arc ${\alpha}$ in $(S,M)$, we have $x_{\alpha}=\phi({\overline{f}})(x_{\alpha})=x_{f({\alpha})}$, where the last identity holds by equation (\[stern\]). Thus $f({\alpha})$ is isotopic to ${\alpha}$, for every arc $
{\alpha}$ in $(S,M)$, and in particular, $f $ fixes each point in $M$. Thus for every triangle ${\Delta}$, $f$ fixes the points of ${\Delta}$ and maps the arcs of ${\Delta}$ to isotopic arcs. Therefore $f$ is isotopic to the identity on each triangle, and hence on the whole surface. This shows that $f$ is zero in ${\mathcal{MG}(S,M)}$, and hence $\phi$ is injective. This shows that ${\mathcal{MG}(S,M)}$ is isomorphic to a subgroup of ${\textup{Aut}^+\mathcal{A}}$.
Now suppose that $p\ge2$ or $\partial S \ne \emptyset$, For any puncture $z$ let $\psi_z$ be the cluster automorphism of Lemma \[lemtag\]. Define a map $\chi:{\mathcal{MG}_{\bowtie}(S,M)}\to {\textup{Aut}^+\mathcal{A}}$ by $\chi({\overline{f}},\mathcal{P})=(\prod_{z\in\mathcal{P}}\psi_z)\phi({\overline{f}})$. In order to show that $\chi$ is a group homomorphism, we compute $$\label{eq44} \chi( ({\overline{f}}_1,\mathcal{P}_1)({\overline{f}}_2,\mathcal{P}_2) ) = (\prod_{z\in \mathcal{P}_1\ominus f_1(\mathcal{P}_2)} \psi_z) \phi({\overline{f}}_1{\overline{f}}_2)$$ and on the other hand $$\begin{aligned}
\chi ({\overline{f}}_1,\mathcal{P}_1)\chi({\overline{f}}_1,\mathcal{P}_1)&=& (\prod_{z\in \mathcal{P}_1} \psi_z) \phi({\overline{f}}_1)(\prod_{z\in \mathcal{P}_2} \psi_z) \phi({\overline{f}}_2) \nonumber \\
&=&(\prod_{z\in \mathcal{P}_1} \psi_z) (\prod_{z\in f_1(\mathcal{P}_2)} \psi_z) \phi({\overline{f}}_1)\phi({\overline{f}}_2),\label{eq45}\end{aligned}$$ where the last identity follows from the equation $\phi({\overline{f}})\psi_z=\psi_{f(z)}\phi({\overline{f}})$. The equality of the expressions in equations (\[eq44\]) and (\[eq45\]) now follows because $\phi $ is a homomorphism and because $\psi_z^2=1$. This shows that $\chi $ is a homomorphism. To show that $\chi $ is injective, suppose that $\chi({\overline{f}},\mathcal{P}) $ is the identity automorphism. Then we have $\mathcal{P}=\emptyset$ and $\chi({\overline{f}},\emptyset)=\phi({\overline{f}})$, and from the injectivity of $\phi$ we get that ${\overline{f}}=1$. Thus $\chi $ is injective.
[The theorem does not describe the whole group ${\textup{Aut}^+\mathcal{A}}$ but only a subgroup, and it is not true in general that ${\textup{Aut}^+\mathcal{A}}= {\mathcal{MG}_{\bowtie}(S,M)}$. However, the only cases we know where this equality does not hold are the surfaces corresponding to the acyclic types $\mathbb{D}_4$ and $\tilde{\mathbb{D}}_4$. These two types correspond to star shaped quivers with $3$ and $4$ branches which have $S_3$-symmetry and $S_4$-symmetry, respectively. However the corresponding surfaces do not have such symmetries. We conjecture that these are the only exceptions, since we know no other surface that gives rise to a quiver having an $S_\ell$-symmetry, with $\ell>2$. ]{}
\[conj\] Let $(S,M)$ be any surface different from the disc with exactly one puncture and four marked points on the boundary or the disc with exactly two punctures and two marked points on the boundary. Then
1. if $(S,M)$ is not a closed surface with exactly one puncture then $${\textup{Aut}^+\mathcal{A}}= {\mathcal{MG}_{\bowtie}(S,M)}.$$
2. if $(S,M)$ is a closed surface with exactly one puncture then $${\textup{Aut}^+\mathcal{A}}= {\mathcal{MG}(S,M)}.$$
We can prove this conjecture using the results from section \[sect 3.3\] in the cases where the cluster algebra from the surface is of acyclic type.
\[theorem a\] Let $(S,M)$ be a disc or an annulus without punctures. Then $${\textup{Aut}^+\mathcal{A}}={\mathcal{MG}(S,M)}={\mathcal{MG}_{\bowtie}(S,M)}.$$
Since $(S,M)$ has no punctures, we have ${\mathcal{MG}(S,M)}={\mathcal{MG}_{\bowtie}(S,M)}$. Suppose first that $(S,M)$ is a disc, and let the number of marked points be $n+3$. Then the cluster algebra is of Dynkin type $\mathbb{A}_n$, and we know from section \[sect 3.3\] that ${\textup{Aut}^+\mathcal{A}}\cong {\mathbb{Z}}_{n+3}$. Thus we only need to show that ${\mathcal{MG}(S,M)}\cong{\mathbb{Z}}_{n+3}$.
We may assume without loss of generality that the marked points are the points of a regular polygon, so that any rotation about the centre of the disc of angle $k \frac{2\pi}{n+3}$, with $k\in {\mathbb{Z}}$, maps $M$ to $M$. Each of these rotations is an orientation preserving homeomorphism of $(S,M)$, and it is isotopic to the identity relative to $M$ if and only if it fixes each point in $M$. This shows that the rotations form a subgroup of ${\mathcal{MG}(S,M)}$ isomorphic to ${\mathbb{Z}}_{n+3}$.
Each element of ${\mathcal{MG}(S,M)}$ is determined by its values on $M$, because if two orientation preserving homeomorphisms $f,g$ agree on $M$, then $fg^{-1}$ fixes $M$, and we may suppose without loss of generality that $fg^{-1}\in {\textup{Homeo}^+(S,\partial S)}$. Since $\mathcal{M}od(S,M)=0$, it follows that $fg^{-1}\in \textup{Homeo}_0(S,\partial S)$, and therefore $fg^{-1}\in {\textup{Homeo}_0(S,M)}$. Since each element of ${\mathcal{MG}(S,M)}$ maps the boundary to itself and preserves orientation, each element can be represented by a rotation. This shows that ${\mathcal{MG}(S,M)}\cong{\mathbb{Z}}_{n+3}$.
Now suppose that $(S,M)$ is an annulus, let $C_1,C_2$ be the two boundary components of $(S,M)$, and let $p$ be the number of marked points on $C_1$ and $ q$ be the number of marked points on $C_2$. Then the cluster algebra is of euclidean type $\tilde{\mathbb{A}}_{p,q}$, and we know from section \[sect 3.3\] that $${\textup{Aut}^+\mathcal{A}}\cong \left\{ \begin{array}
{ll} H_{p,q} &\textup{if $p\ne q$;}\\
H_{p,p}\times {\mathbb{Z}}_2 &\textup{if $p= q$;}
\end{array}
\right.$$ where $H_{p,q}=\langle r_1,r_2\mid r_1r_2=r_2r_1, r_1^p=r_2^q\rangle.$ As in the case of the disc above, the rotations of each boundary component form a subgroup of ${\mathcal{MG}(S,M)}$. Note however that these subgroups are infinite cyclic. We can choose two generators $r_1 $ for the group given by rotating $C_1$ and $r_2 $ for the group given by rotating $C_2$ such that $r_1^p$ and $ r_2^q$ fix every point in $M$ and $r_1^p=r_2^q$. (Thus $r_1,r_2$ are rotations in opposite directions.) Moreover $r_1r_2=r_2r_1$. This shows that $H_{p,q}$ is a subgroup of ${\mathcal{MG}(S,M)}$.
Note that $r_1^p$ and $r_2^q$ are Dehn twists of the annulus described in section \[sect 4.2\].
Suppose first that $p\ne q$. Then each element of ${\mathcal{MG}(S,M)}$ maps each boundary component to itself and, in particular, on each boundary component it is given by a rotation. Moreover, each element of ${\mathcal{MG}(S,M)}$ is determined by its values on $M$ up to composition with $r_1^p$, because if two elements $f,g$ agree on $M$, then $fg^{-1}$ fixes $M$, hence without loss of generality $fg^{-1}$ fixes each point on the boundary. It follows that $fg^{-1}\in{\mathcal{M}od\,S}$ and therefore $fg^{-1}$ is a power of a Dehn twist by Remark \[mgg\], hence a power of $r_1^p$. This shows that ${\mathcal{MG}(S,M)}\cong H_{p,q}$.
Now suppose that $p=q$. Then the elements of ${\mathcal{MG}(S,M)}$ may map one boundary component to the other. Exchanging the boundary components twice maps each boundary component to itself, and by the same argument as in the case $p\ne q$, we see that such an element is given by the rotations. Thus exchanging the boundary components corresponds to a subgroup of order two, whence ${\mathcal{MG}(S,M)}\cong H_{p,p}\rtimes {\mathbb{Z}}_2$, as required.
\[theorem d\] Let $(S,M)$ be a disc with $p$ punctures with $p$ equal to $1$ or $2$, and suppose that the number of marked points on the boundary is at least $5$ if $p=1$ and at least $3$ if $p=2$. Then$${\textup{Aut}^+\mathcal{A}}= {\mathcal{MG}_{\bowtie}(S,M)}.$$
Suppose first that $(S,M)$ is a disc with one puncture let $n$ be the number of marked points on the boundary. By our assumption, we have $n>4$. Then the cluster algebra is of type $\mathbb{D}_n$ and we know from section \[sect 3.3\] that ${\textup{Aut}^+\mathcal{A}}\cong {\mathbb{Z}}_n\times {\mathbb{Z}}_2$. On the other hand, the mapping class group of the once punctured disc is equal to the mapping class group of the unpunctured disc, see Remark \[rem mg2\], thus ${\mathcal{MG}(S,M)}\cong{\mathbb{Z}}_n$. Now it follows from Remark \[rem 38\] that ${\mathcal{MG}_{\bowtie}(S,M)}\cong{\mathbb{Z}}_n\times {\mathbb{Z}}_2$ as required.
Suppose now that $(S,M)$ is a disc with two punctures, and let $n-3$ be the number of marked points on the boundary. By our assumption, we have $n>5$. Then the cluster algebra is of type $\tilde{\mathbb{D}}_{n-1}$, and we know from section \[sect\_acyclic\] that ${\textup{Aut}^+\mathcal{A}}\cong G$, where $$\label{GG}
G=\left\langle \tau,{\sigma},\rho_1,\rho_n \left| \begin{array}{c}
\rho_i^2=1 ,\tau \rho_i=\rho_i \tau \ (i=1,n)\\
\tau{\sigma}={\sigma}\tau,\ {\sigma}^2=\tau^{n-3} \\
\rho_1{\sigma}={\sigma}\rho_n, \ {\sigma}\rho_1=\rho_n{\sigma}\end{array}\right.\right\rangle$$
The mapping class group $\mathcal{M}od$ of the disc with $p$ punctures (without any marked points on the boundary) is isomorphic to the braid group $B_p$ on $p$ strands, see Remark \[rem mg2\], thus in our situation it is isomorphic to $B_2\cong {\mathbb{Z}}$. Let $s$ be a generator of $B_2$. Then $s $ maps one puncture to the other and $s^2$ is isotopic to a rotation of the boundary by the angle $2\pi$ that fixes the punctures.
On the other hand, the elements of ${\mathcal{MG}(S,M)}$ which are induced by the rotations of the boundary (fixing the punctures) form an infinite cyclic group, and we can choose a generator $r$ such that $r^{n-3}=s^2$. Clearly, $rs=s r$.
Up to composition with $r^{n-3}$ and $s$, any element of ${\mathcal{MG}(S,M)}$ is determined by its values on $M$, because if $f,g\in {\mathcal{MG}(S,M)}$ agree on $M$, then $fg^{-1}$ fixes each point in $M$, hence we can suppose without loss of generality that $fg^{-1}$ fixes each point on the boundary. Thus since ${\mathcal{M}od\,S}$ is generated by $s$, it follows that $fg^{-1}$ is a power of $s$. This shows that ${\mathcal{MG}(S,M)}=\langle r,s\mid s^2=r^{n-3}, rs=sr\rangle$.
We must show that $G\cong{\mathcal{MG}_{\bowtie}(S,M)}$. Denote the two punctures by $z_1,z_2$, and let $\phi$ be the map from $G$ to ${\mathcal{MG}_{\bowtie}(S,M)}$ defined on the generators by $$\begin{array} {rcl}
\phi(\tau) &=&(r\,,\,\{z_1,z_2\})\\
\phi({\sigma})&=& \left\{\begin{array}{ll}
(s,\emptyset ) &\textup{if $n$ odd}\\
(s,z_1) &\textup{if $n$ even}
\end{array}\right.\\
\phi(\rho_1) &=&(1,z_1)\\
\phi(\rho_n)&=& (1,z_2)
\end{array}$$ and extended to $G$ by the homomorphism property. One can easily check that $\phi $ preserves the relations of the group, for example if $n$ is even then $$\phi({\sigma}^2)=(s,z_1)^2=(s^2,z_1\ominus s(z_1))=(s^2, \{z_1,z_2\})$$ which is equal to $$\phi(\tau^{n-3})=(r,\{z_1,z_2\})^{n-3} =(r^{n-3},\{z_1,z_2\}),$$ where the last identity follows from Corollary \[cor 37\] (4), since $n$ is even.
To show that $\phi $ is injective suppose that $x=\tau^a{\sigma}^b\rho_1^c\rho_n^d\in \textup{Ker}\, \phi$ for some integers $a,b,c,d$. Then $(1,\emptyset)=\phi(x)$, and by computing the first coordinate of this equation, we get $1=r^a s^b $. Consequently, since $\tau$ and $\sigma$ satisfy the same relations as $r$ and $s$, we have $1=\tau^a {\sigma}^b$, and therefore $x=\rho_1^c\rho_n^d$. Thus $$(1,\emptyset)=\phi(x)=(1,\{z_1\}^c\ominus\{z_2\}^d),$$ which implies that $c$ and $d$ are even, by Corollary \[cor 37\].(4), and thus $x=1$. This shows that $\phi$ is injective.
It remains to show that $\phi$ is surjective. Let $x=(r^as^b,\mathcal{P})\in {\mathcal{MG}_{\bowtie}(S,M)}$. Then $\phi(\tau^a{\sigma}^b)=(r^as^b,\overline{\mathcal{P}})$, for some subset $\overline{\mathcal{P}}\subset \{z_1,z_2\}$, and multiplying with $\phi(\rho_1)$ or $\phi(\rho_2)$ if necessary, we see that $x$ lies in the image of $\phi$. This shows that $\phi $ is surjective, and thus $\phi $ is an isomorphism.
We end this section with another look at Example \[ex torus\]. The quiver $$\xymatrix@R15pt@C10pt{x_1\ar@<2pt>[rd]\ar@<-2pt>[rd] &&x_2 \ar@<2pt>[ll]\ar@<-2pt>[ll]\\
&x_3\ar@<2pt>[ru]\ar@<-2pt>[ru]
}$$ corresponds to a triangulation of the torus with one puncture, which can be seen easily using the plane as a universal cover and the triangulation shown on the left hand side of Figure \[fig torus\]. The edges are labeled $1,2,3$ instead of $x_1,x_2,x_3$ for brevity. Edges that have the same label are to be identified, and each point in Figure \[fig torus\] is identified to the puncture. Thus in the triangulation shown on the left hand side of Figure \[fig torus\] there are exactly two triangles, both formed by edges $1,2,3$ and both having the same orientation.
The picture in the middle of Figure \[fig torus\] shows the triangulation corresponding to the seed obtained from the initial seed by mutating in $x_1 $, while the image on the right hand side of Figure \[fig torus\] shows the triangulation corresponding to the seed obtained by mutating once more, this time in $x_2$.
Geometrically, one can deform the picture on the left into the picture on the right by dragging the right end upwards and the left end downwards. In the torus, this “deformation" corresponds to two Dehn twists along the closed curve labeled $3$. On the other hand, there is no orientation preserving homeomorphism transforming the picture on the left into the one in the middle. Thus this mutation is not given by a mapping class. Of course we could have deduced this simply from the observation in Example \[ex torus\] that this mutation corresponds to an inverse cluster automorphism and not to a direct one.
![Three triangulations of the torus[]{data-label="fig torus"}](figtorus.eps)
[Finiteness of the automorphism group]{}\[sect 4\] In this section, we introduce the notion of *automorphism finite* cluster algebras and prove that for acyclic cluster algebras and for cluster algebras from surfaces it is equivalent to the notion of finite type cluster algebras. We say that a cluster algebra $\mathcal{A}$ is *automorphism finite* if its automorphism group ${\textup{Aut}\,\mathcal{A}}$ is finite.
\[finite\] Let $\mathcal{A}$ be a cluster algebra arising from an acyclic quiver or from a surface. Then $\mathcal{A}$ is automorphism finite if and only if $\mathcal{A}$ is of Dynkin type.
Sufficiency follows from Table \[table\]. To prove necessity, suppose first that $\mathcal{A}$ is arising from an acyclic quiver $Q$. By Theorem \[theorem\_main\], ${\textup{Aut}^+\mathcal{A}}$ is isomorphic to the quotient of the group of automorphisms of the transjective component ${\Gamma}_{tr}$ of the Auslander-Reiten quiver of the cluster category $\mathcal{C}_Q$ modulo the stabiliser of the points in ${\Gamma}_{tr}$.
If $Q$ is not of Dynkin type, then the Auslander-Reiten translation induces an element of $\textup{Aut}({\Gamma}_{tr})$ of infinite order, acting freely on the points of ${\Gamma}_{tr}$. Thus ${\textup{Aut}\,\mathcal{A}}$ is infinite if $Q$ is acyclic and not Dynkin.
Suppose now that $\mathcal{A}$ arises from a surface $(S,M)$. By Lemma \[mgg\], the mapping class group ${\mathcal{M}od\,S}$ of the surface $S$ is a subgroup of ${\mathcal{MG}(S,M)}$, which in turn is isomorphic to a subgroup of ${\textup{Aut}\,\mathcal{A}}$, by Theorem \[theorem mg\]. So in order to show that ${\textup{Aut}\,\mathcal{A}}$ is infinite, it suffices to find an infinite subgroup in ${\mathcal{M}od\,S}$.
By Remark \[rem mg1\], there exists a Dehn twist which generates an infinite cyclic subgroup of ${\mathcal{M}od\,S}$ if $S$ has genus at least one or $S$ has genus zero and two or more boundary components.
There remain the cases where $S$ is a disc or a sphere. If $S$ is a disc with $p\ge 2$ punctures, then the braid group $B_p$ is an infinite subgroup of the mapping class group of $S$. In the cases where $p=0$ or $1$, we have that $\mathcal{A}$ is of Dynkin type $\mathbb{A}$ or $\mathbb{D}$, respectively.
Finally, if $S$ is a sphere with $p$ punctures, then by our assumption $p\ge 4$, and then it is known that the mapping class group of $S$ has a free subgroup, see [@FM 4.2].
For a sphere with 3 punctures, the mapping class group is $S_3$, which is a finite group. However the sphere with 3 or less punctures is excluded in the construction of cluster algebras from surfaces in [@FST].
[99]{}
, Cluster-tilted algebras and slices, *J. of Algebra* [**319**]{} (2008), 3464-3479. I. Assem and G. Dupont, Friezes and the construction of the euclidean cluster variables, J. Pure and Applied Algebra [**215**]{} (2011), 2322-2340. I. Assem, G. Dupont, R. Schiffler and D. Smith, Friezes, strings and cluster variables, to appear in *Glasgow J. Math.*
I. Assem, C. Reutenauer, and D. Smith, Friezes, Adv. in Math. [**225**]{} (2010), 3134-3165.
I. Assem, D. Simson and A. Skowroński, [*Elements of the Representation Theory of Associative Algebras, 1: Techniques of Representation Theory*]{}, London Mathematical Society Student Texts 65, Cambridge University Press, 2006.
M. Auslander, M. I. Platzeck and I. Reiten, Coxeter functors without diagrams, *Trans. Amer. Marh. Soc.* [**250**]{} (1979), 1-46. , *Representation Theory of Artin Algebras*, Cambridge Studies in Advanced Math. 36, Cambridge University Press, Cambridge, 1995.
Bernstein, I. N., Gelfand, I. M. and Ponomarev, V. A., Coxeter functors and Gabriel’s theorem, *Uspekhi Mat. Nauk.* [**28**]{} (1973), 19–33, English translation in *Russian Math. Surveys* [**28**]{} (1973), 17-32.
, Tilting theory and cluster combinatorics, *Adv. Math.* [**204**]{} (2006), no. 2, 572-618.
, Quivers with relations arising from clusters ($A_n$ case), *Trans. Amer. Math. Soc.* [**358**]{} (2006), no. 3, 1347-1364.
, From triangulated categories to cluster algebras, *Invent. Math.* [**172**]{} (2008), no. 1, 169-211.
P. Caldero and B. Keller. From triangulated categories to cluster algebras II. *Ann. Sc. Ec. Norm. Sup.* [**39**]{} (2006), no. 4, 83-100.
D. Dummit and R. Foote, Abstract Algebra, third edition, Wiley and Sons, 2003.
V. Fock and A. Goncharov, Moduli spaces of local systems and higher Teichmüller theory. [*Publ. Math. Inst. Hautes Études Sci.*]{} (2006), no. 103, 1-211.
B. Farb and D. Margalit, A Primer on Mapping Class Groups, [http://www.math.uchicago.edu/ margalit/mcg/mcgv400.pdf]{}
S. Fomin, M. Shapiro, and D. Thurston, Cluster algebras and triangulated surfaces. Part I: Cluster complexes, *Acta Mathematica* [**201**]{} (2008), 83-146.
, Cluster algebras I. Foundations, *J. Amer. Math. Soc.* [**15**]{} (2002), no. 2, 497-529 (electronic)
S. Fomin and A. Zelevinsky, Cluster algebras II: Finite type classification, *Invent. Math.* [**154**]{}, (2003), 63-121. M. Gekhtman, M. Shapiro and A. Vainshtein, On the properties of the exchange graph of a cluster algebra, Math. Res. Lett. [**15**]{} (2008), no. 2, 321–330.
, *Triangulated Categories in the Representation Theory of Finite Dimensional Algebras*, London Mathematical Society. Lecture Notes Series 119, Cambridge University Press, Cambridge, 1988.
, On triangulated orbit categories, *Documenta Math.* [**10**]{} (2005), 551-581.
G. Musiker, R. Schiffler, and L. Williams, Positiviy for cluster algebras from surfaces, [*Adv. Math.*]{} [**227**]{}, (2011), 2241–2308. H. Nakajima, Quiver varieties and cluster algebras (preprint) [arXiv:0905.0002]{}.
C. Riedtmann, Algebren, Darstellungsköcher, Überlagerungen und zurück, [*C*omment. Math. Helv.]{} [**55**]{} (1980), 199-224.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Tangent categories provide an axiomatic framework for understanding various tangent bundles and differential operations that occur in differential geometry, algebraic geometry, abstract homotopy theory, and computer science. Previous work has shown that one can formulate and prove a wide variety of definitions and results from differential geometry in an arbitrary tangent category, including generalizations of vector fields and their Lie bracket, vector bundles, and connections.
In this paper we investigate differential and *sector* forms in tangent categories. We show that sector forms in any tangent category have a rich structure: they form a symmetric cosimplicial object. This appears to be a new result in differential geometry, even for smooth manifolds. In the category of smooth manifolds, the resulting complex of sector forms has a subcomplex isomorphic to the de Rham complex of differential forms, which may be identified with *alternating* sector forms. Further, the symmetric cosimplicial structure on sector forms arises naturally through a new equational presentation of symmetric cosimplicial objects, which we develop herein.
address: |
Department of Mathematics and Computer Science,\
Mount Allison University, Sackville, Canada.
author:
- 'G. S. H. Cruttwell and R. B. B. Lucyshyn-Wright'
bibliography:
- 'simplicialDeRham.bib'
title: |
A simplicial foundation for\
differential and sector forms\
in tangent categories
---
Introduction
============
Tangent categories [@rosicky; @sman3] provide an axiomatization of one of the key structures in differential geometry: the tangent bundle. Tangent categories are useful for a number of reasons. First, constructions of objects like the tangent bundle appear in a variety of categories, some related to the category of smooth manifolds, others to categories in algebraic geometry, and others to categories in homotopy theory and computer science. Thus, it is helpful to have a single axiomatization which can deal with all these examples simultaneously. Secondly, a variety of definitions and constructions in differential geometry are closely linked to the tangent bundle. For example, vector fields, the Lie bracket, connections, and differential forms can all be viewed as certain maps in the category of smooth manifolds which take as domain or codomain the tangent bundle (or bundles related to it). Thus, one can hope to give definitions and prove results about these objects in an arbitrary tangent category.
This paper is a contribution to the second aspect of this program; in particular, in this paper we are interested in determining how to define differential forms, their exterior derivative, and the resulting cochain complex of de Rham in an arbitrary tangent category. However, to do so requires a close inspection of the nature of differential forms. This inspection reveals an interesting structure, a simplicial object of *sector forms*, of which de Rham cohomology can be seen as a simple consequence.
There is a relatively straightforward analog of the notion of differential form in any tangent category. Classical differential $n$-forms on a smooth manifold $M$ can be viewed as multilinear, alternating maps $$T_nM \rightarrow \mathbb{R}$$ where $T_nM$ is the object of consisting of all “$n$-tuples of tangent vectors at a common point on $M$”. (That is, $T_nM$ is the fibre product of $n$ copies of the tangent bundle $TM \rightarrow M$ over $M$.) These objects exist in any tangent category, and thus one can define (classical) differential forms in any tangent category as above, with $\mathbb{R}$ replaced by a suitable coefficient object.
However, a difficulty arises when attempting to define a direct analog of the exterior derivative of such forms in an arbitrary tangent category. In the category of smooth manifolds, the exterior derivative of an $n$-form $\omega: T_nM \rightarrow \mathbb{R}$ is an $(n+1)$-form ${\ensuremath{\partial}}\omega$, which can be defined locally on open subsets $U \cong {\ensuremath{\mathbb R}\xspace}^n$. In the case where $M = {\ensuremath{\mathbb R}\xspace}^n$ one can define $\partial \omega$ as an alternating sum of certain maps $T_{n+1}M \rightarrow {\ensuremath{\mathbb R}\xspace}$ [@natural 7.8], each expressed in terms of the Jacobian derivative $T(\omega):T(T_nM) \rightarrow T{\ensuremath{\mathbb R}\xspace}\cong {\ensuremath{\mathbb R}\xspace}\times {\ensuremath{\mathbb R}\xspace}$ by pre-composing with a certain canonical map $\kappa:T_{n+1}M \rightarrow T(T_nM)$. A similar definition applies in any Cartesian differential category [@deRhamCartDiffCats]. However, in an arbitrary tangent category, the objects need not be manifolds, and the local definition cannot be mimicked globally for want of a suitable map $\kappa$ to mediate between the intended domain of $\partial \omega$ (namely $T_{n+1}M$) and the domain of $T(\omega)$ (namely $T(T_nM)$).
One solution to this problem can be found by considering how synthetic differential geometry (SDG) handles differential forms. In SDG one finds categories which have *representable* tangent structure; these are categories with an object $D$ for which there is a tangent functor $T$ defined by $TM := M^D$. Various definitions and results have been transplanted from classical differential geometry to models of SDG; see, e.g., [@kock; @lavendhomme; @reyes]. In a typical model of SDG, as in a tangent category, the objects need not be locally isomorphic to some $\mathbb{R}^n$. Thus, for a general object in such categories, the exterior derivative also cannot be defined by mimicking the classical definition directly.
In SDG, the solution to this problem is to look at a different type of map: instead of considering multilinear alternating maps from $T_nM \rightarrow R$, one instead considers multilinear alternating maps $$T^nM \to R,$$ where $T^nM$ is the $n$th iterate of the tangent bundle of $M$. Such maps were first considered in [@Kock:DiffFormsSDG; @KRV] and referred to as *singular forms* in [@lavendhomme Definition 4.1]; we shall use that name here to distinguish them from other notions of form we shall consider. In contrast to the classical case, the Jacobian derivative $T(\omega)$ of a singular $n$-form $\omega$ does have the expected domain, namely $T^{n+1}M$. The references above show how to define an exterior derivative for such forms; the definition involves an alternating sum of permutations of the Jacobian derivative. Moreover, it has been shown that for a particular model of SDG which contains the category of smooth manifolds, if $M$ is a smooth manifold then singular forms are in bijective correspondence with classical differential forms, and their exterior derivatives agree [@reyes IV, Proposition 3.7]. This shows that models of SDG have a notion of de Rham complex which generalizes the classical notion.
Thus, for tangent categories, a natural point of investigation is to look at multilinear alternating maps from $T^nM$ to some coefficient object $E$. We show in this paper that such maps indeed have an exterior derivative that generalizes the definition from SDG, and has the required properties (Proposition \[prop:complexOfSectorForms\]). Thus, this shows that tangent categories have a notion of de Rham cohomology (namely, the cohomology of the resulting complex) which generalizes the SDG notion, and hence also generalizes the classical notion.
However, there is much more to say about maps from $T^nM$ to $E$ in a tangent category. In particular, none of the results that need to be proved to show that such maps have an exterior derivative require that the maps from $T^nM$ be alternating; multilinearity suffices. Thus, one is led to consider maps $$\omega: T^nM \rightarrow E$$ (for a suitable coefficient object $E$) which are multilinear but not necessarily alternating. Such maps do not appear in published accounts of SDG, but have appeared in differential geometry [@white; @sectorFormsPhysicsArticle]. They are known as *sector forms* (for some basic examples of sector forms, see Section \[sec:results\]). The exterior derivative of singular forms works for sector forms, and so in addition to the complex of singular forms, tangent categories have complexes of sector forms (\[def:compl\_sector\_forms\]).
However, there is much more structure to these sector forms than a cochain complex. We show that for each $n$, there are $n+1$ ‘derivative’ or *co-face* operations which take sector $n$-forms to sector $(n+1)$-forms (Theorem \[thm:sym\_cosimpl\_cmon\_sector\_forms\]), there are $n-1$ *symmetry* operations which take sector $n$-forms to sector $n$-forms, and there are $n-1$ *co-degeneracy* operations which take sector $n$-forms to sector $(n-1)$-forms (Proposition \[thm:cod\_symm\_sector\_forms\])[^1]. Taken together, these operations constitute the structure of an *(augmented) symmetric cosimplicial object* [@Barr; @Gra:Symm] of sector forms (Theorem \[thm:sym\_cosimpl\_cmon\_sector\_forms\]); that is, there is a functor on the category of finite cardinals. This is a remarkably rich structure, and has not previously appeared in either ordinary differential geometry or synthetic differential geometry[^2].
Thus, we view the symmetric cosimplicial object of sector forms as the primary object of interest in relation to the various notions of differential forms considered above. In particular, from this cosimplicial object one can obtain as a simple corollary the complex of sector forms and the complex of singular forms[^3]. Moreover, by generalizing to maps which are not necessarily alternating, one also generalizes covariant tensors (multilinear maps with domain $T_n$) which have numerous uses throughout differential geometry [@white Section 3.1]. In other words, sector forms generalize three important ideas in differential geometry: differential forms, covariant tensors, and singular forms. Thus, it is important to understand the structure of sector forms, and this paper represents a substantial advance in the study of these objects in the general setting of tangent categories.
-------------- ------------------- ------------------
Alternating Not alternating
Domain $T_n$ differential form covariant tensor
Domain $T^n$ singular form sector form
-------------- ------------------- ------------------
It is also worth noting that this paper contains two other points of independent interest. First, to establish the symmetric cosimplicial structure of sector forms, it becomes natural to give an alternative presentation of symmetric cosimplicial objects, and in particular to give an alternative presentation of the category of finite cardinals. The standard presentation [@Gra:Symm] involves co-face maps, symmetry maps, and co-degeneracy maps. However, for each $n$, the $n+1$ co-face maps from $n$ to $n+1$ can all be obtained by applying symmetries to a single co-face map, and thus one can show (Theorem \[thm:fund\_cof\_pres\_fincard\]) that the category of finite cardinals can be presented by symmetries, co-degeneracies, and a single co-face map for each $n$.
The second point of interest relates to methodology in tangent categories in general. The definition of the symmetry and co-degeneracy maps of sector forms involves various combinations of the *lift* natural transformation $\ell: T \to TT$ and the *canonical flip* transformation $c: TT \to TT$ (which are part of the definition of a tangent category). To establish the various identities that are required of the symmetries and co-degeneracies, one then must perform various complicated calculations with these maps. One way to handle the complexity of such calculations is to use string diagrams, as was done in previous work on tangent categories [@jacobiProof]. Another way to handle the complexity is to use a recently discovered embedding theorem for tangent categories [@garnerEmbedding] (for more on this approach, see the discussion after \[tangent-category-examples\]). However, here we use a different approach. Diagrams involving the maps $\ell$ and $c$ in a tangent category ${\ensuremath{\mathscr X}\xspace}$ can be viewed as the application of a certain functor from the category of finite cardinals and surjections (written as ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$) to the category of endofunctors on ${\ensuremath{\mathscr X}\xspace}$ (Example \[exa:symm\_deg\_it\_tang\]). Thus, to establish the commutativity of such a diagram of natural transformations, it suffices to establish the commutativity of a certain diagram in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$, and this is typically straightforward. For examples of this proof technique, see Proposition \[thm:cod\_symm\_sector\_forms\] and Theorem \[thm:fund\_deriv\_sector\_form\].
The paper is laid out as follows. In section \[sec:tanCats\], we review the definitions of tangent categories and *differential objects*, which are the coefficient objects in which the forms will take their values. Before going into the various details required of many of the proofs, in section \[sec:results\] we give an overview of the key definitions and results of the paper, providing more detail than in the discussion above, and also providing some examples of sector forms. In sections \[sec:monSemigroups\], \[sec:symCosimp\], and \[sec:presFundamentalCoface\], we study symmetric cosimplicial objects and related notions, emphasizing their relations to categories of finite cardinals and establishing equational presentations of some of these key categories. Throughout these sections, we show how some of the structure of these categories is present in the category of endofunctors on a tangent category. In section \[sec:cosimpOfSectorForms\], we look at sector forms, their *fundamental derivative*, and how they have the structure of a symmetric cosimplicial object. In section \[sec:complexes\], we obtain the complexes of sector forms and singular forms as simple consequences of the symmetric cosimplicial structure on sector forms. In section \[sec:relFormsClSDG\], we study forms in the presence of representable tangent structure, and we show how our definitions of sector forms and singular forms in a tangent category relate to existing definitions in classical and synthetic differential geometry. Finally, in section \[sec:futureWork\], we look at various ways to extend or add to the results we have presented.
Tangent categories and differential objects {#sec:tanCats}
===========================================
Notation {#sec:notn}
--------
Throughout this paper, composition in diagrammatic order is indicated with a semicolon, so that $f$, followed by $g$, is written as $f {\ensuremath{\mathop{;}}}g$. When $F$ and $G$ are functors, we will sometimes denote the composite $F {\ensuremath{\mathop{;}}}G$ instead by $GF$, so that juxtaposition of functors denotes classical right-to-left composition. Given an object $C$ of a category ${\ensuremath{\mathscr C}\xspace}$, we denote by ${\ensuremath{\textnormal{Aut}}}_{\ensuremath{\mathscr C}\xspace}(C)$ the group of automorphisms of $C$ in ${\ensuremath{\mathscr C}\xspace}$. Rather than straying from convention by defining multiplication in ${\ensuremath{\textnormal{Aut}}}_{\ensuremath{\mathscr C}\xspace}(C)$ in terms of the diagrammatic composition order, we instead take the view that groups are certain one-object categories, and we define composition in ${\ensuremath{\textnormal{Aut}}}_{\ensuremath{\mathscr C}\xspace}(C)$ as in ${\ensuremath{\mathscr C}\xspace}$.
Additive bundles
----------------
If $M$ is an object in a category ${\ensuremath{\mathscr X}\xspace}$, an **additive bundle over $M$** consists of a map $q: E \to M$ such that (i) $q$ admits finite *fibre powers*, i.e., for each $n \in {\ensuremath{\mathbb N}\xspace}$ there is a fibre product $E_n \rightarrow M$ of $n$ copies of $(E,q)$ over $M$ with projections $\pi_1,...,\pi_n:E_n \rightarrow E$, and (ii) $(E,q)$ is a commutative monoid in the slice category ${\ensuremath{\mathscr X}\xspace}/M$. In particular, this means there is an addition operation, which will often be written as $\sigma: E_2 \to E$ and must satisfy the usual requirements of commutativity and associativity, and a unit for this addition, which will often be written as $\zeta:M \to E$. A map between such bundles will, in general, just be a commutative square $$\xymatrix{E \ar[rr]^e \ar[d]_q && E' \ar[d]^{q'} \\ B \ar[rr]_b && B'}$$ written $(e,b): q \to q'$. If, in addition, such a map of bundles preserves the addition – that is $e_2; \sigma' = \sigma; e$ and $b; \zeta' = \zeta; e$ – then we shall say that $(e,b)$ is an [**additive bundle morphism**]{}.
\[defnTangentCategory\] For a category ${\ensuremath{\mathscr X}\xspace}$, tangent structure ${\ensuremath{\mathbb T}\xspace}= (T,p,0,+,\ell,c)$ on ${\ensuremath{\mathscr X}\xspace}$ consists of the following data:
- (**tangent functor**) a functor $T: {\ensuremath{\mathscr X}\xspace}\to {\ensuremath{\mathscr X}\xspace}$ with a natural transformation $p: T \to I$ such that each $p_M: T(M) \to M$ admits finite fibre powers that are preserved by each $T^n:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$;
- (**additive bundle**) natural transformations $+:T_2 \to T$ (where $T_2$ is the second fibre power of $p$) and $0: I \to T$ making each $p_M: TM \to M$ an additive bundle;
- (**vertical lift**) a natural transformation $\ell:T \to T^2$ such that for each $M$ $$(\ell_M,0_M): (p: TM \to M,+,0) \to (Tp: T^2M \to TM,T(+),T(0))$$ is an additive bundle morphism;
- (**canonical flip**) a natural transformation $c:T^2 \to T^2$ such that for each $M$ $$(c_M,1): (Tp: T^2M \to TM, T(+),T(0)) \to (p_T: T^2M \to TM,+_T,0_T)$$ is an additive bundle morphism;
- (**coherence of $\ell$ and $c$**) $c{\ensuremath{\mathop{;}}}c = 1$ (so $c$ is an isomorphism), $\ell{\ensuremath{\mathop{;}}}c = \ell$, and the following diagrams commute: $$\xymatrix{T \ar[r]^{\ell} \ar[d]_{\ell} &T^2 \ar[d]^{T\ell} \\
T^2 \ar[r]_{\ell T} & T^3}
~~~
\xymatrix{T^3 \ar[r]^{T c} \ar[d]_{c T} & T^3 \ar[r]^{c T} & T^3 \ar[d]^{Tc} \\
T^3 \ar[r]_{Tc} & T^3 \ar[r]_{cT} & T^3}
~~~
\xymatrix{T^2 \ar[d]_{c} \ar[r]^{\ell T} & T^3 \ar[r]^{Tc} & T^3 \ar[d]^{cT} \\
T^2 \ar[rr]_{T\ell} & & T^3}$$
- (**universality of vertical lift**) defining $v: T_2M \to T^2M$ by $v := {\langle}{\pi}_1{\ensuremath{\mathop{;}}}\ell, {\pi}_2{\ensuremath{\mathop{;}}}0_T{\rangle}T(+)$, the following diagram is a pullback that is preserved by each $T^n$: $$\xymatrix{T_2(M) \ar[d]_{\pi_1 {\ensuremath{\mathop{;}}}p\:=\:\pi_2 {\ensuremath{\mathop{;}}}p} \ar[rr]^{v} & & T^2(M) \ar[d]^{T(p)} \\ M \ar[rr]_{0} & & T(M)}$$
A category with tangent structure, $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$, is a [**tangent category**]{}.
\[tangent-category-examples\] Here are several important examples of tangent categories, the first four of which are drawn from [@sman3] and [@diffBundles].
1. Finite dimensional smooth manifolds with the usual tangent bundle structure.
2. Convenient manifolds with the kinematic tangent bundle [@convenient Section 28].
3. The infinitesimally and vertically linear objects in any model of synthetic differential geometry [@kock] form a tangent category: if $D$ is the object of square-zero infinitesimals, then we take $TM := M^D$.
4. The opposite of the category of finitely presented commutative rings (or more generally commutative rigs[^4]) is another example of a category with representable tangent structure: here $D$ is the ‘rig of infinitesimals’, ${\ensuremath{\mathbb N}\xspace}[\varepsilon] := {\ensuremath{\mathbb N}\xspace}[x]/(x^2=0)$ and again $TA := A^D$.
5. A very different example of a tangent category arises from abstract homotopy theory, in particular in work on *Abelian functor calculus* [@abFunCalc]. In [@abFuncCalcDirDeriv], the authors show that a certain operation in abelian functor calculus gives rise to a *Cartesian differential category* [@cartDiff]. As every Cartesian differential category is a tangent category [@sman3 Section 4.2], this example is also a tangent category; this insight was useful in providing a straightforward proof of the existence of certain higher-order chain rules for abelian functor calculus (see the discussion at the top of page 5 in [@abFuncCalcDirDeriv]).
6. Other examples of tangent categories that arise as Cartesian differential categories include the models of the differential $\lambda$-calculus that appear in computer science (for example, see [@manzonetto] and [@cockettGallagher]).
More examples can be found in [@sman3] and [@diffBundles]. In addition to these examples, recent work of Leung [@leungClassifyingTanStructures] and Garner [@garnerEmbedding] establishes certain equivalent formulations of tangent categories that provide new perspectives on the axioms. Leung’s work shows that tangent categories are closely related to categories of Weil algebras, while Garner’s work builds on this result to show not only that tangent categories can be seen as certain types of enriched categories, but also that every tangent category can be embedded in a tangent category that is *representable*, in the sense that the functor $T$ is representable (for more on representable tangent categories, see [@sman3 Section 5] and also \[def:rep\_tan\] below).
This last point allows one to work in an arbitrary tangent category as if $T$ was representable, allowing for calculations in a tangent category which closely resemble those in SDG. One may ask, for example, whether this could simplify the proofs of some of the results in this paper. However, we have not found that this was the case. Certain of our initial attempts at proofs of the main results of this paper indeed used representable tangent categories, but the resulting calculations were no less lengthy than those recorded herein; indeed, by observing that the transformations $\ell$ and $c$ generate a model of a certain *PROP*, in Mac Lane’s sense [@MacL:CatAlg], we have reduced many of these calculations to showing that certain diagrams of finite sets commute. More importantly still, it was only by working with the relatively restrictive stuctures of tangent categories and associated PROPs that we discovered the main results of this paper, including the result that sector forms carry cosimplicial structure.
Commutative monoids in a Cartesian tangent category {#sec:cmons_in_tngnt_cat}
---------------------------------------------------
The coefficient objects of our forms will in particular be commutative monoids, so it is useful to first make some remarks about commutative monoids in a tangent category.
A tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ is said to be **Cartesian** if ${\ensuremath{\mathscr X}\xspace}$ has finite products that are preserved by the tangent functor $T:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$. In this case we denote by ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$ the category of commutative monoid objects in ${\ensuremath{\mathscr X}\xspace}$. For each object $X$ of ${\ensuremath{\mathscr X}\xspace}$, the functor ${\ensuremath{\mathscr X}\xspace}(X,-):{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{set}}}\xspace}$ preserves limits and so sends each commutative monoid $E$ in ${\ensuremath{\mathscr X}\xspace}$ to a commutative monoid ${\ensuremath{\mathscr X}\xspace}(X,E)$ in ${\ensuremath{\textnormal{\textbf{set}}}\xspace}$. When $X$ itself is a commutative monoid in ${\ensuremath{\mathscr X}\xspace}$, the hom-set ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})(X,E)$ is a submonoid of ${\ensuremath{\mathscr X}\xspace}(X,E)$. Composition in ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$ preserves this monoid structure in each variable separately, so we say that ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$ is an *additive category*, which, following previous papers on tangent categories, we take to mean a category enriched in commutative monoids. Moreover, we have the following:
\[prop:cmons\_in\_tngt\_cat\] Let $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ be a Cartesian tangent category. Then the tangent endofunctor $T:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ lifts to an endofunctor ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}(T):{\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace}) \rightarrow {\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$. Moreover, the endofunctor ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}(T)$ is additive in the sense that it preserves the commutative monoid structure on the hom-sets of ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$. Furthermore, if $\phi:T^i \Rightarrow T^j$ is a natural transformation between iterates of $T$ ($i,j \in {\ensuremath{\mathbb N}\xspace}$), then for each commutative monoid $E$ in ${\ensuremath{\mathscr X}\xspace}$ the morphism $\phi_E:T^iE \rightarrow T^j E$ is a homomorphism of commutative monoids in ${\ensuremath{\mathscr X}\xspace}$.
There is a 2-category ${\ensuremath{\textnormal{\textbf{Cart}}}
\xspace}$ whose objects are categories with finite products, wherein the 1-cells are functors preserving finite products and the 2-cells are arbitrary natural transformations. Letting ${\ensuremath{\mathscr C}\xspace}$ denote the Lawvere theory of commutative monoids, there is an equivalence of categories ${\ensuremath{\textnormal{\textbf{Cart}}}
\xspace}({\ensuremath{\mathscr C}\xspace},{\ensuremath{\mathscr D}\xspace}) \simeq {\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr D}\xspace})$ for every object ${\ensuremath{\mathscr D}\xspace}$ of ${\ensuremath{\textnormal{\textbf{Cart}}}
\xspace}$. But ${\ensuremath{\textnormal{\textbf{Cart}}}
\xspace}({\ensuremath{\mathscr C}\xspace},-):{\ensuremath{\textnormal{\textbf{Cart}}}
\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{Cat}}}
\xspace}$ is a 2-functor valued in the 2-category of categories, and it follows that the assignment ${\ensuremath{\mathscr D}\xspace}\mapsto {\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr D}\xspace})$ underlies a 2-functor ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}:{\ensuremath{\textnormal{\textbf{Cart}}}
\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{Cat}}}
\xspace}$. We can apply this 2-functor to the 1-cell $T$ and to the 2-cell $\phi$, thus proving two of the above claims. Lastly, ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$ is an additive category with finite products, which are therefore finite biproducts, and since $T$ preserves finite products it follows that ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}(T)$ preserves finite biproducts and hence is additive.
Differential objects {#sec:diff_objs}
--------------------
Before we define sector forms and singular forms, we need to consider the objects in which these forms will take their values. These will be *differential objects*, which are certain objects $E$ whose tangent bundle $TE$ is simply a product $E \times E$. This is formulated more precisely as follows.
\[def:diff\_objs\] Let $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ be a Cartesian tangent category. A **differential object** in $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ consists of a commutative monoid $(E,\mu,\eta)$ in ${\ensuremath{\mathscr X}\xspace}$ (equivalently, an additive bundle over $1$) together with a map $\lambda: E \to TE$ (known as the *lift*) such that
- $(\lambda,\eta)$ is an additive bundle morphism from $(E,!_E,\mu,\eta)$ to $(TE,p,+,0)$;
- $\lambda$ is a homomorphism of commutative monoids from $(E,\mu,\eta)$ to $(TE,T(\mu),T(\eta))$;
- the equation $\lambda{\ensuremath{\mathop{;}}}T(\lambda) = \lambda{\ensuremath{\mathop{;}}}\ell_E$ holds;
- the map $$E \times E \to^{\nu \::=\: {\langle}\pi_1{\ensuremath{\mathop{;}}}\lambda,\:\pi_2{\ensuremath{\mathop{;}}}0_E{\rangle}{\ensuremath{\mathop{;}}}T(\mu)} TE$$ is an isomorphism.
Say that $(E,\mu,\eta,\lambda)$ is a **subtractive** differential object if the commutative monoid $(E,\mu,\eta)$ is an abelian group. In this case we denote the inverse operation by $-: E \to E$.
\[diff-object-examples\] Here are some important examples of differential objects from [@diffBundles].
(i) In the category of smooth manifolds, each Cartesian space $\mathbb{R}^n$ is a differential object, where $\lambda:{\ensuremath{\mathbb R}\xspace}^n \rightarrow T{\ensuremath{\mathbb R}\xspace}^n = {\ensuremath{\mathbb R}\xspace}^n \times {\ensuremath{\mathbb R}\xspace}^n$ sends $x$ to $(x,0)$, construed as the tangent vector $x$ at the point $0$.
(ii) Similarly, in the category of convenient manifolds, each convenient vector space is a differential object.
(iii) In the category of affine schemes $\mathbf{cRing}^{\mbox{\scriptsize{op}}}$, polynomial rings $\mathbb{Z}[x_1, x_2, \ldots, x_n]$ are differential objects.
(iv) Differential objects in a tangent category associated to a model of SDG are precisely the Euclidean $R$-modules [@lavendhomme 1.1.4] (see [@diffBundles Theorem 3.9] for a proof of this).
All of the above examples are subtractive.
Remark. {#sec:princ_projn}
--------
By definition, if $E$ is a differential object, then $TE \cong E \times E$. Through the isomorphism $\nu$, one can show that the projection from the second component is $p_E: TE \to E$. We will write $\hat{p}:TE \rightarrow E$ for the projection to the first component, and refer to it as the **principal projection**. Differential objects can be alternatively axiomatized in terms of the principal projection $\hat{p}$. For example, this was how differential objects were originally presented [@sman3 Definition 4.8]. It is a relatively straightforward exercise to show the equivalence of the two definitions [@diffBundles Proposition 3.4].
We will make use of both the lift $\lambda: E \to TE$ and the principal projection $\hat{p}: TE \to E$ when investigating differential forms with values in $E$. In particular, the following results about these maps will be useful.
\[prop:diffObjs\] If $E$ is a differential object with lift $\lambda$ and principal projection $\hat{p}$, then
(i) $\hat{p}$ is a homomorphism of commutative monoids, where $TE$ has the commutative monoid structure as discussed in \[prop:cmons\_in\_tngt\_cat\].
(ii) $\lambda {\ensuremath{\mathop{;}}}\hat{p} = 1_E$.
(iii) $0_E {\ensuremath{\mathop{;}}}\hat{p} = \;!_E {\ensuremath{\mathop{;}}}\zeta$.
(iv) $\ell_E{\ensuremath{\mathop{;}}}T(\hat{p}){\ensuremath{\mathop{;}}}\hat{p} = \hat{p}$.
(v) $c_E {\ensuremath{\mathop{;}}}T(\hat{p}){\ensuremath{\mathop{;}}}\hat{p} = T(\hat{p}){\ensuremath{\mathop{;}}}\hat{p}$.
(vi) $\ell_E{\ensuremath{\mathop{;}}}T(\hat{p}) = \hat{p}{\ensuremath{\mathop{;}}}\lambda$.
(vii) $T(\lambda){\ensuremath{\mathop{;}}}c_E {\ensuremath{\mathop{;}}}T(\hat{p}) = \hat{p}{\ensuremath{\mathop{;}}}\lambda$.
The first five parts are established in [@diffBundles Propositions 3.4 and 3.6, Definition 3.1]. For (vi), we regard $TE$ as a product $E \times E$ with projections $(\hat{p},p_E)$. Then $\lambda$ is the morphism $\langle 1_E, 0 \rangle$ induced by the identity morphism on $E$ and the zero element $0$ of the commutative monoid ${\ensuremath{\mathscr X}\xspace}(E,E)$ (\[sec:cmons\_in\_tngnt\_cat\]). Hence $$\hat{p} {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}\hat{p} = \hat{p} {\ensuremath{\mathop{;}}}1_E = \hat{p} = \ell_E {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p}$$ by (iv). Also, using the naturality of $p$, the equation $\ell {\ensuremath{\mathop{;}}}c = \ell$, and the fact that $(\ell_E,0_E)$ and $(c_E,1_{TE})$ are bundle morphisms (\[defnTangentCategory\]) we compute that $$\begin{array}{ccccccc}
\hat{p} {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}p_E & = & \hat{p} {\ensuremath{\mathop{;}}}0 & = & 0 & = & p_E {\ensuremath{\mathop{;}}}0\\
& = & p_E {\ensuremath{\mathop{;}}}0_E {\ensuremath{\mathop{;}}}\hat{p} & = & \ell_E {\ensuremath{\mathop{;}}}Tp_E {\ensuremath{\mathop{;}}}\hat{p} & = & \ell_E {\ensuremath{\mathop{;}}}c_E {\ensuremath{\mathop{;}}}p_{TE} {\ensuremath{\mathop{;}}}\hat{p}\\
& = & \ell_E {\ensuremath{\mathop{;}}}p_{TE} {\ensuremath{\mathop{;}}}\hat{p} & = & \ell_E {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}p_E
\end{array}$$ where each unadorned $0$ denotes the zero element of the relevant hom-set (\[sec:cmons\_in\_tngnt\_cat\]) while $0_E = \langle 0,1_E\rangle:E \rightarrow TE \cong E \times E$ denotes the zero section, i.e. the component of $0:I \Rightarrow T$ at $E$. Hence $\hat{p} {\ensuremath{\mathop{;}}}\lambda = \ell_E {\ensuremath{\mathop{;}}}T\hat{p}$ as needed.
For (vii), we again use the fact that $TE$ is a product $E \times E$ with projections $(\hat{p},p_E)$. Appending the first projection $\hat{p}$ to the equation in question, we compute that $$T\lambda {\ensuremath{\mathop{;}}}c_E {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} = T\lambda {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} = T(\lambda {\ensuremath{\mathop{;}}}\hat{p}) {\ensuremath{\mathop{;}}}\hat{p} = \hat{p} = \hat{p} {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}\hat{p}$$ using (ii) and (v). Appending the second projection $p_E$, we compute that $$\begin{array}{lllllllll}
T\lambda {\ensuremath{\mathop{;}}}c_E {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}p_E & = & T\lambda {\ensuremath{\mathop{;}}}c_E {\ensuremath{\mathop{;}}}p_{TE} {\ensuremath{\mathop{;}}}\hat{p} & = & T\lambda {\ensuremath{\mathop{;}}}Tp_E {\ensuremath{\mathop{;}}}\hat{p} & = & T(\lambda {\ensuremath{\mathop{;}}}p_E) {\ensuremath{\mathop{;}}}\hat{p} & = & T(0) {\ensuremath{\mathop{;}}}\hat{p}\\
& = & 0 {\ensuremath{\mathop{;}}}\hat{p} & = & 0 & = & \hat{p} {\ensuremath{\mathop{;}}}0 & = & \hat{p} {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}p_E
\end{array}$$ using the naturality of $p$, the additivity of $T$ (\[prop:cmons\_in\_tngt\_cat\]), the fact that $\lambda {\ensuremath{\mathop{;}}}p_E = 0$, the fact that $\hat{p}$ is a homomorphism of commutative monoids, and the fact that $c_E {\ensuremath{\mathop{;}}}p_{TE} = Tp_E$ since $(c,1_{TE})$ is a bundle morphism (\[defnTangentCategory\]).
Overview of main results, with examples of sector forms {#sec:results}
=======================================================
To prove the main results of this paper we have used a variety of techniques and definitions. However, we feel it is important to present many of the main results in a single place so as to be easily locatable and so as to feature them prominently. In this section we also look at some examples of sector forms, as they are perhaps much less familiar to the general reader than differential forms.
Throughout this section, we work in a Cartesian tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ with a fixed object $M$ and a fixed differential object $(E,\sigma,\zeta,\lambda)$.
We first define a number of natural transformations between powers of $T$ which will appear in the definitions of certain types of forms and their derivatives. Much of the work of sections \[sec:monSemigroups\] and \[sec:symCosimp\] deals with how to interpret these natural transformations and handle them more efficiently.
Define $$\ell^n_i \;:=\; T^{i - 1}\ell T^{n - i}\;\;:\;\;T^n \rightarrow T^{n+1}\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n)$$ $$c^n_i \;:=\; T^{i - 1} c T^{n - i - 1}\;\;:\;\;T^n \rightarrow T^n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n - 1)$$ $$c^n_{(i)} \;:=\; c_{i-1}{\ensuremath{\mathop{;}}}c_{i-2} {\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}c_2 {\ensuremath{\mathop{;}}}c_1\;\;:\;\;T^n \rightarrow T^n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n)$$ $$a^n_{i} \;:=\; \ell_i^n {\ensuremath{\mathop{;}}}c_{(i)}^{n+1}\;\;:\;\;T^{n} \rightarrow T^{n+1}\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n)$$
The main object of study of this paper is the following notion of form, originally due to White [@white].
\[def:sector\_nform\] A **sector $n$-form** on $M$ with values in $E$ is a morphism $\omega: T^n M \rightarrow E$ such that for each $i \in \{1,...,n\}$, $\omega$ is *linear in the $i$th variable*; that is, the following diagram commutes[^5]: $$\bfig
\square<500,300>[T^nM`E`T^{n+1}M`TE;\omega`a^n_i M`\lambda`T(\omega)]
\efig$$ The set of sector $n$-forms on $M$ with values in $E$ will be denoted by $\Psi_n(M;E)$ (often abbreviated to $\Psi_n(M)$).
To help explore the similarities and differences between ordinary differential forms and sector forms, we will briefly look at sector 1- and 2-forms on $\mathbb{R}$ in the category of smooth manifolds.
\[ex1Form\] Let us first consider what sector 1-forms on $\mathbb{R}$ with values in $\mathbb{R}$ consist of. By definition, such a form consists of a map $\omega: T\mathbb{R} \to \mathbb{R}$ that satisfies the single linearity equation $$\bfig
\square<500,300>[T\mathbb{R}`\mathbb{R}`T^2\mathbb{R}`T\mathbb{R};\omega`\ell_{\mathbb{R}}`\lambda`T\omega]
\efig$$
Recall that $T\mathbb{R}$ is simply $\mathbb{R} \times \mathbb{R}$, via the two projections $p_{\mathbb{R}}: T\mathbb{R} \to \mathbb{R}$ and $\hat{p}: T\mathbb{R} \to \mathbb{R}$. Hence the commutativity of the above diagram is equivalent to its commutation when post-composed with $p_{{\ensuremath{\mathbb R}\xspace}}$ and $\hat{p}$, respectively. But $\omega {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}p_{\ensuremath{\mathbb R}\xspace}= \omega {\ensuremath{\mathop{;}}}\;!_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\zeta = \;!_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}\zeta$ by \[def:diff\_objs\], and $\ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}p_{\ensuremath{\mathbb R}\xspace}= \ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}p_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}\omega = \ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}c_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}p_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}\omega = \ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}Tp_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega = p_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega$ by the axioms for tangent categories (\[defnTangentCategory\]). Hence $\omega {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}p_{\ensuremath{\mathbb R}\xspace}= \ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}p_{\ensuremath{\mathbb R}\xspace}\Leftrightarrow \;!_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}\zeta = p_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega \Leftrightarrow p_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\;!_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\zeta = p_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega \Leftrightarrow \;!_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\zeta = 0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega$ since $p_{\ensuremath{\mathbb R}\xspace}$ is a retraction (of $0_{\ensuremath{\mathbb R}\xspace}$). On the other hand, $\ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = \ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}D\omega$ where $D \omega = T\omega {\ensuremath{\mathop{;}}}\hat{p}:T^2{\ensuremath{\mathbb R}\xspace}\rightarrow {\ensuremath{\mathbb R}\xspace}$ is the directional derivative of $\omega$, so since $\lambda {\ensuremath{\mathop{;}}}\hat{p} = 1$ by \[prop:diffObjs\] we find that the above linearity equation holds if and only if the following equations hold: $$\ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}D\omega = \omega\;\;\;\;\;\;\;\;!_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\zeta = 0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega\;.$$ But the first equation entails the second, since if the first holds then $0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega = 0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}D\omega = 0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = 0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}0_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = 0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega {\ensuremath{\mathop{;}}}0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\hat{p} =
0_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega {\ensuremath{\mathop{;}}}\;!_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\zeta = \;!_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\zeta$ by \[defnTangentCategory\] and \[prop:diffObjs\].
Hence the linearity equation is equivalent to the equation $\ell_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}D\omega = \omega$. In order to reformulate this equation more concretely, let us write $\omega:{\ensuremath{\mathbb R}\xspace}\times {\ensuremath{\mathbb R}\xspace}\rightarrow {\ensuremath{\mathbb R}\xspace}$ as a function $\omega(x,v)$ of two variables $x,v$, so that we may write its first and second partial derivatives briefly as $\frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}x}$ and $\frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}v}$. We can write $T^2{\ensuremath{\mathbb R}\xspace}= T({\ensuremath{\mathbb R}\xspace}\times {\ensuremath{\mathbb R}\xspace})$ as a product $T^2{\ensuremath{\mathbb R}\xspace}= ({\ensuremath{\mathbb R}\xspace}\times {\ensuremath{\mathbb R}\xspace}) \times ({\ensuremath{\mathbb R}\xspace}\times {\ensuremath{\mathbb R}\xspace})$, whereupon $D \omega: T^2\mathbb{R} \to \mathbb{R}$ is given by $${\langle}x_0,v_1,v_2,d{\rangle}\mapsto \frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}x}(x_0,v_1)\cdot v_2 + \frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}v}(x_0,v_1) \cdot d \;,$$ and $\ell: T\mathbb{R} \to T^2\mathbb{R}$ is given by $${\langle}x,v{\rangle}\mapsto {\langle}x,0,0,v{\rangle}\;.$$ Hence $\ell {\ensuremath{\mathop{;}}}D\omega$, evaluated at ${\langle}x,v{\rangle}$, is $$\frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}x}(x,0)\cdot 0 + \frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}v}(x,0)\cdot v = \frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}v}(x,0)\cdot v.$$ So the linearity equation says $$\omega(x,v) = \frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}v}(x,0)\cdot v.$$ Thus, if we set $f(x) = \frac{{\ensuremath{\partial}}\omega}{{\ensuremath{\partial}}v}(x,0)$, then $$\omega(x,v) = f(x) \cdot v.$$ It is easy to see that any map of this form is indeed a sector 1-form. So, in this case, sector 1-forms are precisely the same as ordinary differential 1-forms (more generally, this is true for any smooth manifold).
\[ex2Form\] Despite \[ex1Form\], sector $n$-forms for $n \geq 2$ are in general quite different from ordinary differential $n$-forms, even for a simple smooth manifold such as $\mathbb{R}$. For example, for $n=2$, a sector 2-form on $\mathbb{R}$ consists of a map $\omega: T^2\mathbb{R} \to \mathbb{R}$ such that $$\bfig
\square<500,300>[T^2\mathbb{R}`\mathbb{R}`T^3\mathbb{R}`T\mathbb{R};\omega`\ell_{T\mathbb{R}}`\lambda`T\omega]
\place(1000,150)[\mbox{and}]
\square(1750,0)<500,300>[T^2\mathbb{R}`\mathbb{R}`T^3\mathbb{R}`T\mathbb{R};\omega`T\ell_{\mathbb{R}}{\ensuremath{\mathop{;}}}c_{T\mathbb{R}}`\lambda`T\omega]
\efig$$ Using similar reasoning to the previous example, it is straightforward to show that any map of the form $$\omega({\langle}x,v_1,v_2,d{\rangle}) = f(x)\cdot v_1 \cdot v_2 + g(x) \cdot d$$ (where $f(x)$ and $g(x)$ are smooth functions from $\mathbb{R}$ to itself) is an example of a sector 2-form[^6] on $\mathbb{R}$. This is very different from the general description of ordinary differential 2-forms on ${\ensuremath{\mathbb R}\xspace}$: there is only one, namely the zero form.
One of the key points of this paper is that the sector $n$-forms have a rich variety of operations that can be performed on them. In particular, there are $n+1$ different *derivative* or *co-face* operations $\delta_i^n$ which take sector $n$-forms to sector $(n+1)$-forms: $$\delta_i^n(\omega) := c^{n+1}_{(i)}M {\ensuremath{\mathop{;}}}T\omega{\ensuremath{\mathop{;}}}\hat{p}\;\;\; \mbox{ (Theorem \ref{thm:sym_cosimpl_cmon_sector_forms})}$$ there are $n-1$ different *co-degeneracy* operations $\varepsilon_{i}^{n-1}$ which take sector $n$-forms to sector $(n-1)$-forms: $$\varepsilon_i^{n-1}(\omega) := \ell_i^{n-1}M{\ensuremath{\mathop{;}}}\omega\;\;\; \mbox{ (Proposition \ref{thm:cod_symm_sector_forms})}$$ and there are $n-1$ different *symmetry* operations $\sigma_i^n$ which take sector $n$-forms to sector $n$-forms: $$\sigma_i^n(\omega) := c_i^nM{\ensuremath{\mathop{;}}}\omega\;\;\; \mbox{ (Proposition \ref{thm:cod_symm_sector_forms}).}$$ Much of the work of this paper goes into proving the following result (Theorem \[thm:sym\_cosimpl\_cmon\_sector\_forms\]): $$\begin{minipage}{4.5in}
\textit{The operations $\delta$, $\varepsilon$, $\sigma$ together endow the set of sector forms on $M$ with the structure of an (augmented) symmetric cosimplicial commutative monoid, i.e. a functor from the category of finite cardinals to the category of commutative monoids.}
\end{minipage}$$ In particular, this means that for every function between finite cardinals $f: n \rightarrow m$ there is an associated monoid homomorphism $\Psi_f: \Psi_n(M) \to \Psi_m(M)$ (given as some composite of the above co-face, co-degeneracy, and symmetry operations), and this entire assignment is functorial.
Moreover, if $E$ is subtractive, then this structure forms a symmetric cosimplicial abelian group. In general, any cosimplicial abelian group has an associated cochain complex whose differential $\partial$ is given by taking an alternating sum of the co-face maps: $$\partial(\omega) := \sum_{i=1}^{n+1} (-1)^{i-1} \delta_i^n(\omega)$$ And so there is a complex of sector forms (\[def:compl\_sector\_forms\]), whose differential we call the *exterior derivative*. The fact that the sector forms constitute a cochain complex appears to be a new result in differential geometry.
Let us consider the first several groups of this complex for $M = E = \mathbb{R}$ in the category of smooth manifolds. By definition, a sector $0$-form on $\mathbb{R}$ is simply a smooth map $\omega: \mathbb{R} \to \mathbb{R}$. By Example \[ex1Form\], a sector 1-form on $\mathbb{R}$ is the same as a differential 1-form on $\mathbb{R}$; that is, a map $T\mathbb{R} \to \mathbb{R}$ of the form ${\langle}x,v{\rangle}\mapsto f(x) \cdot v$. By Example \[ex2Form\], a sector 2-form on $\mathbb{R}$ is a map $T\mathbb{R} \to \mathbb{R}$ of the form $${\langle}x,v_1,v_2,d{\rangle}\mapsto g(x) \cdot v_1 \cdot v_2 + h(x) \cdot d.$$
The exterior derivative of a 0-form $\omega$ is the same as for ordinary differential forms: $$\partial (\omega)({\langle}x,v{\rangle}) = \omega'(x) \cdot v.$$
For a sector 1-form $\omega: {\langle}x,v{\rangle}\mapsto f(x) \cdot v$ we have $T\omega {\ensuremath{\mathop{;}}}\hat{p} = D\omega$, the directional derivative of $\omega$, which in this case is $$D\omega: {\langle}x,v_1,v_2,d{\rangle}\mapsto f'(x) \cdot v_1 \cdot v_2 + f(x)\cdot d \;.$$ So then the exterior derivative of $\omega$ is ${\ensuremath{\partial}}(\omega) = D\omega \:-\: c_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}D\omega$ and hence is given by $${\ensuremath{\partial}}(\omega)(x,v_1,v_2,d) = (f'(x) \cdot v_1 \cdot v_2 + f(x)\cdot d) - (f'(x) \cdot v_2 \cdot v_1 + f(x)\cdot d) = 0,$$ since the effect of $c_{\ensuremath{\mathbb R}\xspace}$ is simply to switch the middle two co-ordinates ($v_1$ and $v_2$). So, in this case, every sector 1-form has exterior derivative $0$. (Note that this is also automatic since every 1-form is the exterior derivative of a 0-form and the sector forms constitute a complex. However, it is useful to see how this works explicitly).
However, the exterior derivative of a sector 2-form is not typically zero. For a sector 2-form $$\omega: {\langle}x,v_1,v_2,d{\rangle}\mapsto g(x) \cdot v_1 \cdot v_2 + h(x) \cdot d,$$ its directional derivative takes as input an 8-tuple ${\langle}x,v_1,v_2,d_1,v_3,d_2,d_3,t{\rangle}$, and maps it to $$g'(x)\cdot v_1 \cdot v_2 \cdot v_3 + h'(x) \cdot d_1 \cdot v_3 + g(x) \cdot v_2 \cdot d_2 + g(x) \cdot v_1 \cdot d_3 + h(x) \cdot t.$$ Now, the effect of $c_{T\mathbb{R}}$ is $${\langle}x,v_1,v_2,d_1,v_3,d_2,d_3,t{\rangle}\mapsto {\langle}x,v_1,v_3,d_2,v_2,d_1,d_3,t{\rangle}$$ and the effect of $Tc_\mathbb{R}$ is $${\langle}x,v_1,v_2,d_1,v_3,d_2,d_3,t{\rangle}\mapsto {\langle}x,v_2,v_1,d_1,v_3,d_3,d_2,t{\rangle}$$ Hence since $$\partial (\omega) \;=\; D\omega \:-\: c_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}D\omega \:+\: Tc_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}c_{T{\ensuremath{\mathbb R}\xspace}} {\ensuremath{\mathop{;}}}D\omega,$$ $\partial (\omega)$ maps ${\langle}x,v_1,v_2,d_1,v_3,d_2,d_3,t{\rangle}$ to $$g'(x) \cdot v_1 \cdot v_2 \cdot v_3 + h'(x)\cdot v_3 \cdot d_1 + (2g(x) - h'(x))\cdot v_2 \cdot d_2 + h'(x)\cdot v_1 \cdot d_3 + h(x) \cdot t.$$ Note that if this is identically zero, then by setting all variables except $t$ to 0, we get $h(x) = 0$, and then by setting all variables but $v_2$ and $d_2$ to 0, we also get $g(x) = 0$. Thus, a sector 2-form on ${\ensuremath{\mathbb R}\xspace}$ has exterior derivative $0$ if and only it is identically zero..
Taken together, these results tell us the first three sector cohomology groups of $\mathbb{R}$. The 0th cohomology is the same as ordinary de Rham cohomology (namely, $\mathbb{R}$, since the constant functions are those with derivative $0$). Similarly, the 1st cohomology is the same (namely, $0$), since every 1-form (sector or differential) is the image of a $0$-form. Finally, the second sector form cohomology group is also zero, but for a different reason than for de Rham cohomology. In de Rham cohomology, it is zero since there are no non-trivial 2-forms on $\mathbb{R}$. For sector forms, it is zero since by the above, the only closed sector 2-form is the zero form.
It is an open question whether sector form cohomology is always the same as de Rham cohomology; the basic examples given above, however, at least show that the complexes they form are quite different. We hope to explore the relationship between sector form and de Rham cohomology in a future paper.
It is also important to note that the individual ‘derivative’ operations $\delta_i^n$ on sector forms appear to have geometric significance: for more on this, see [@white Chapter 4].
Returning to our general setting, we shall consider the following further property possessed by some sector forms:
If $E$ is subtractive, a **singular $n$-form** on $M$ with values in $E$ is a sector $n$-form $\omega: T^n M \rightarrow E$ such that $\omega$ is *alternating*; that is, for each $1 \leqslant i \leqslant n-1$, $$\bfig
\square<500,300>[T^nM`E`T^{n}M`E;\omega`c^n_i M`-`\omega]
\efig$$
The exterior derivative operation $\partial$ defined above restricts to such singular forms, and so there is also a complex of singular forms (Proposition \[prop:complexOfSectorForms\]).
Let us consider which sector 2-forms on $\mathbb{R}$ are alternating. By the above, a sector 2-form on $\mathbb{R}$ takes the form $$\omega: {\langle}x,v_1,v_2,d{\rangle}\mapsto f(x)\cdot v_1 \cdot v_2 + g(x) \cdot d.$$ For 2-forms, the condition of being alternating amounts to a single equation $$c_{\ensuremath{\mathbb R}\xspace}{\ensuremath{\mathop{;}}}\omega = -\omega.$$ Since $c_{\ensuremath{\mathbb R}\xspace}$ swaps $v_1$ and $v_2$, $\omega$ is alternating if and only if for all $x,v_1,v_2,d$, $$f(x) \cdot v_2 \cdot v_1 + g(x) \cdot d = -(f(x)\cdot v_1 \cdot v_2 + g(x) \cdot d).$$ But this implies that $f(x) = g(x) = 0$, so $\omega$ is constantly zero. Hence the only singular 2-form on $\mathbb{R}$ is the zero form.
In fact, we can show much more generally that the complex of singular forms on any smooth manifold (with values in $\mathbb{R}$) is isomorphic to its de Rham complex. We shall prove this in \[thm:classical\_derham\] after first comparing the above singular forms to those studied in synthetic differential geometry [@Kock:DiffFormsSDG; @KRV; @kock]. Indeed, we shall show that in the tangent category determined by a model of SDG, the above complex of singular forms is isomorphic to its SDG counterpart (\[thm:singf\_vs\_sdgsingf\]), and in certain models of SDG the latter complex is known to be isomorphic to the ordinary de Rham complex of differential forms when $M$ is a smooth manifold [@reyes IV, Proposition 3.7].
Symmetric monoids, semigroups, and finite sets {#sec:monSemigroups}
==============================================
In working towards the symmetric cosimplicial structure on sector forms, we will make use of an algebraic structure carried by the tangent endofunctor $T$, namely the structure of a *symmetric semigroup* (\[def:symm\_semi\_mon\]). Many of the results and ideas in this section are due to previous authors [@Burr:HdWordPr; @Barr; @Gra:Symm; @Laf:Eq2d; @Laf:Bool], but the applications to tangent categories are new.
Monoids and semigroups
----------------------
Given a strict monoidal category ${\ensuremath{\mathscr V}\xspace}$, let us denote the unit object of ${\ensuremath{\mathscr V}\xspace}$ by $I$ and write the monoidal product in ${\ensuremath{\mathscr V}\xspace}$ as juxtaposition. By definition, a **semigroup** $(S,m)$ in ${\ensuremath{\mathscr V}\xspace}$ is an object $S$ of ${\ensuremath{\mathscr V}\xspace}$ equipped with a morphism $m:SS \rightarrow S$ (called a *multiplication*) that satisfies the following **associative law** $$\label{eq:assoc}Sm {\ensuremath{\mathop{;}}}m = mS {\ensuremath{\mathop{;}}}m\;.$$ Explicitly, this means that the composites $SSS \xrightarrow{Sm} SS \xrightarrow{m} S$ and $SSS \xrightarrow{mS} SS \xrightarrow{m} S$ are equal. A **monoid** $(S,m,e)$ is a semigroup $(S,m)$ equipped with an additional morphism $e:I \rightarrow S$ (called the *unit*) that satisfies the following **unit laws** $$\label{eq:unit}eS {\ensuremath{\mathop{;}}}m = 1_S\;\;\;\;\;\;Se {\ensuremath{\mathop{;}}}m = 1_S\;,$$ noting that $IS = S = SI$ since ${\ensuremath{\mathscr V}\xspace}$ is strict monoidal.
If the given strict monoidal category ${\ensuremath{\mathscr V}\xspace}$ underlies a *symmetric* monoidal category, then we say that a semigroup $(S,m)$ or monoid $(S,m,e)$ in ${\ensuremath{\mathscr V}\xspace}$ is **commutative** if it satisfies the following **commutative law** $$\label{eq:comm_law}
s {\ensuremath{\mathop{;}}}m = m$$ where $s:SS \rightarrow SS$ is the symmetry isomorphism carried by ${\ensuremath{\mathscr V}\xspace}$.
Example: The tangent functor as (co)semigroup {#exa:tng_func_sgrp}
---------------------------------------------
Given a tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$, the tangent functor $T:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ carries the structure of a semigroup in the monoidal category $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{{\mbox{\textnormal{\scriptsize op}}}}}$, i.e. the *opposite* of the category $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ of endofunctors on ${\ensuremath{\mathscr X}\xspace}$. Indeed the vertical lift $\ell:T \rightarrow TT$ serves as an associative multiplication in $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{{\mbox{\textnormal{\scriptsize op}}}}}$. (Note that in general no unit exists to make this semigroup into a monoid.)
Monoidal categories of finite cardinals {#sec:cats_fin_cards}
---------------------------------------
Writing ${\ensuremath{\textnormal{\textbf{set}}}\xspace}$ for the category of sets, let us denote by ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ the full subcategory of ${\ensuremath{\textnormal{\textbf{set}}}\xspace}$ whose objects are the finite cardinals, which we identify with their corresponding ordinals and also with the natural numbers $n \in {\ensuremath{\mathbb N}\xspace}$. The sum $n + m$ of a pair of finite cardinals carries the structure of a coproduct in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$, where the associated mappings $n \rightarrow n + m$ and $m \rightarrow n + m$ are order preserving and injective and send $n$ and $m$, respectively, onto initial and final segments of the ordinal $n + m$. In general, if a category ${\ensuremath{\mathscr C}\xspace}$ is equipped with designated binary coproducts and a designated initial object, then ${\ensuremath{\mathscr C}\xspace}$ carries an associated structure of symmetric monoidal category. In particular, ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ is therefore symmetric monoidal, with monoidal product $+$ and unit object $0$. Further, $({\ensuremath{\textnormal{\textbf{finCard}}}\xspace},+,0)$ is a *strict* monoidal category, but note that although $n + m = m + n$ as objects of ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$, the symmetry isomorphism $\sigma_{nm}:n + m \rightarrow m + n$ is not the identity map.
We shall consider several non-full subcategories of ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ with the same objects as ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ itself:
1. ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$, whose morphisms are *surjections*;
2. ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b$, whose morphisms are *bijections*, all of which are automorphisms;
3. ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}$, whose morphisms are *order preserving maps*;
4. ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s$, whose morphisms are *order preserving surjections*.
Each of these subcategories is closed under the monoidal product in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ and hence inherits the structure of a strict monoidal category. Note that ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ and ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b$ contain the symmetries $\sigma_{mn}$ and so are symmetric strict monoidal categories, whereas the other subcategories are merely strict monoidal categories.
Universal monoids and semigroups {#sec:univ_mon_sgrps}
--------------------------------
The cardinal $1$ carries the structure of a commutative monoid $(1,\mu,\eta)$ in the symmetric monoidal category ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$, where the associated multiplication $\mu$ and unit $\eta$ are the unique maps $$\mu:1 + 1 \rightarrow 1\;\;\;\;\;\;\;\;\eta:0 \rightarrow 1\;.$$ Since these maps are order preserving, $(1,\mu,\eta)$ is also a monoid in ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}$. These monoids and their underlying semigroups have the following universal properties:
\[thm:univ\_mon\_and\_comm\_mon\] Let ${\ensuremath{\mathscr V}\xspace}$ be a strict monoidal category.
1. Given a monoid $(S,m,e)$ in ${\ensuremath{\mathscr V}\xspace}$, there is a unique strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}\rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$, $S^\sharp(\mu) = m$, and $S^\sharp(\eta) = e$.
2. Given a semigroup $(S,m)$ in ${\ensuremath{\mathscr V}\xspace}$, there is a unique strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s \rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$ and $S^\sharp(\mu) = m$.
3. [(Burroni [@Burr:HdWordPr 2.2], Grandis [@Gra:Symm 4.1])]{.nodecor} If ${\ensuremath{\mathscr V}\xspace}$ is symmetric, then given a commutative monoid $(S,m,e)$ in ${\ensuremath{\mathscr V}\xspace}$, there is a unique symmetric strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$, $S^\sharp(\mu) = m$, and $S^\sharp(\eta) = e$.
4. [(Lafont [@Laf:Bool 2.3, p. 266])]{.nodecor} If ${\ensuremath{\mathscr V}\xspace}$ is symmetric, then given a commutative semigroup $(S,m)$ in ${\ensuremath{\mathscr V}\xspace}$, there is a unique symmetric strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$ and $S^\sharp(\mu) = m$.
\(i) and (ii) are well-known, e.g. see [@MacL VII.5, Proposition 1 and Exercise 3]. We will defer the proofs of (iii) and (iv) until \[rem:results\_in\_symm\_mon\_setting\_follow\] below, where we will see that they follow from more general results on the basis of the cited work of Burroni, Grandis, and Lafont.
Hence, up to a bijection, monoids (resp. semigroups) in strict monoidal categories are the same as strict monoidal functors on ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}$ (resp. ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s$), and analogous statements hold for the commutative variants of these notions. In the terminology of [@MacL:CatAlg], ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}$ is therefore the *PRO* that defines the notion of monoid, and ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ is the *PROP* that defines the notion of commutative monoid.
Symmetric monoids and semigroups
--------------------------------
One of the ramifications of \[thm:univ\_mon\_and\_comm\_mon\](iii) is that it provides a way to generalize the notion of commutative monoid to the context of *non-symmetric* monoidal categories. Indeed, work of Burroni [@Burr:HdWordPr 2.2] and of Grandis [@Gra:Symm §2] shows that a strict monoidal functor ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr V}\xspace}$ valued in a mere strict monoidal category ${\ensuremath{\mathscr V}\xspace}$ is equivalently given by a monoid in ${\ensuremath{\mathscr V}\xspace}$ equipped with a compatible *symmetry* isomorphism, per the following definition:
\[def:symm\_semi\_mon\] Let ${\ensuremath{\mathscr V}\xspace}$ be a strict monoidal category.
1. A **symmetry** on an object $S$ of ${\ensuremath{\mathscr V}\xspace}$ is a morphism $s:SS \rightarrow SS$ satisfying the following equations: $$\label{eq:symm}s {\ensuremath{\mathop{;}}}s = 1_{SS}\;\;\;\;\;\;\;\;Ss {\ensuremath{\mathop{;}}}sS {\ensuremath{\mathop{;}}}Ss = sS {\ensuremath{\mathop{;}}}Ss {\ensuremath{\mathop{;}}}sS\;.$$
2. A **symmetric semigroup** $(S,m,s)$ in ${\ensuremath{\mathscr V}\xspace}$ consists of a semigroup $(S,m)$ in ${\ensuremath{\mathscr V}\xspace}$ together with a symmetry $s$ on the object $S$ such that the following equation is satisfied $$\label{eq:compat_symm_mult}Sm {\ensuremath{\mathop{;}}}s = sS {\ensuremath{\mathop{;}}}Ss {\ensuremath{\mathop{;}}}mS$$ and the commutativity law is also satisfied.
3. (Grandis [@Gra:Symm §2]) A **symmetric monoid** $(S,m,e,s)$ in ${\ensuremath{\mathscr V}\xspace}$ consists of a monoid $(S,m,e)$ in ${\ensuremath{\mathscr V}\xspace}$ with a symmetry $s$ on $S$ such that $(S,m,s)$ is a symmetric semigroup and the following equation is satisfied: $$\label{eq:comp_symm_unit}eS {\ensuremath{\mathop{;}}}s = Se\;.$$
One can generalize each of the above notions to the setting of an arbitrary monoidal category ${\ensuremath{\mathscr V}\xspace}$ by inserting associativity and unit isomorphisms as needed.
\[rem:equivalent\_defs\_symm\_sgrp\_mnd\] It is readily verified that one obtains an equivalent definition of symmetric semigroup by replacing the equation with the equation $$\label{eq:alt_compat_symm_mult}mS {\ensuremath{\mathop{;}}}s = Ss {\ensuremath{\mathop{;}}}sS {\ensuremath{\mathop{;}}}Sm$$ which appears in [@Burr:HdWordPr 2.2], [@Laf:Bool p. 265], and [@Laf:Eq2d 3.3]. Similarly, we obtain an equivalent definition of symmetric monoid by replacing the equation with $$\label{eq:at_comp_symm_unit}Se {\ensuremath{\mathop{;}}}s = eS$$ which appears in [@Burr:HdWordPr 2.2] and [@Laf:Eq2d 3.3].
\[rem:comm\_sgrp\_yields\_symm\_sgrp\] Any commutative semigroup $(S,m)$ (resp. commutative monoid $(S,m,e)$) in a symmetric strict monoidal category ${\ensuremath{\mathscr V}\xspace}$ carries the structure of a symmetric semigroup $(S,m,s)$ (resp. symmetric monoid $(S,m,e,s)$) in ${\ensuremath{\mathscr V}\xspace}$ when we take $s$ to be the relevant component of the symmetry isomorphism carried by ${\ensuremath{\mathscr V}\xspace}$. In particular, the monoid $(1,\mu,\eta)$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ carries the structure of a symmetric monoid $(1,\mu,\eta,\sigma)$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$, and its underlying symmetric semigroup $(1,\mu,\sigma)$ is also a symmetric semigroup in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$.
\[thm:univ\_symm\_mon\_sgrp\] Let ${\ensuremath{\mathscr V}\xspace}$ be a strict monoidal category.
1. [(Burroni [@Burr:HdWordPr 2.2], Grandis [@Gra:Symm 4.1])]{.nodecor} Given a symmetric monoid $(S,m,e,s)$ in ${\ensuremath{\mathscr V}\xspace}$, there is a unique strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$, $S^\sharp(\mu) = m$, $S^\sharp(\eta) = e$, and $S^\sharp(\sigma) = s$.
2. [(Lafont [@Laf:Bool 2.3, p. 266])]{.nodecor} Given a symmetric semigroup $(S,m,s)$ in ${\ensuremath{\mathscr V}\xspace}$, there is a unique strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$, $S^\sharp(\mu) = m$, and $S^\sharp(\sigma) = s$.
3. [(Lafont [@Laf:Eq2d 3.2])]{.nodecor} The object $1$ carries a symmetry $\sigma$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b$ that is universal in the sense that if $S$ is an object of ${\ensuremath{\mathscr V}\xspace}$ and $s$ is a symmetry on $S$, then there is a unique strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \rightarrow {\ensuremath{\mathscr V}\xspace}$ with $S^\sharp(1) = S$ and $S^\sharp(\sigma) = s$.
\(i) is explicitly proved in the cited work of Grandis and also follows immediately from the cited earlier result of Burroni. (ii) follows immediately from the cited result of Lafont, which gives a presentation of the strict monoidal category ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ in terms of the generators $\mu,\sigma$ and the relations for a symmetric semigroup (\[def:symm\_semi\_mon\], \[rem:equivalent\_defs\_symm\_sgrp\_mnd\]). Similarly, (iii) follows from the cited result of Lafont, which presents the strict monoidal category ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b$ in terms of the generator $\sigma$ and the relations for a symmetry on an object (\[def:symm\_semi\_mon\]).
\[rem:results\_in\_symm\_mon\_setting\_follow\] We may apply the preceding theorem to *commutative* monoids in symmetric strict monoidal categories ${\ensuremath{\mathscr V}\xspace}$ by way of \[rem:comm\_sgrp\_yields\_symm\_sgrp\], yielding a proof of \[thm:univ\_mon\_and\_comm\_mon\](iii). Similarly we obtain a proof of \[thm:univ\_mon\_and\_comm\_mon\](iv).
Example: The tangent functor as symmetric (co)semigroup {#sec:tng_func_symm_cosgrp}
-------------------------------------------------------
Given a tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$, the tangent endofunctor $T:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ carries the structure of a symmetric semigroup in the opposite $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{{\mbox{\textnormal{\scriptsize op}}}}}$ of the endofunctor category $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$. Indeed, within the definition of tangent structure (\[defnTangentCategory\]), the axioms under the heading *coherence of $\ell$ and $c$* assert precisely that $(T,\ell,c)$ is a symmetric semigroup in $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{{\mbox{\textnormal{\scriptsize op}}}}}$, where $\ell:T \rightarrow TT$ is the vertical lift and $c:TT \rightarrow TT$ is the canonical flip.
Symmetric cosimplicial objects {#sec:symCosimp}
==============================
The category $\Delta$ of *positive* finite ordinals and order preserving maps admits a geometric interpretation that can be illustrated by way of a well-known functor from $\Delta$ to the category of topological spaces, sending $n$ to the *standard geometric $(n-1)$-simplex* $\Delta_{n - 1} \subseteq {\ensuremath{\mathbb R}\xspace}^n$, i.e. the convex hull of the standard basis vectors in ${\ensuremath{\mathbb R}\xspace}^n$. Consequently, presheaves on $\Delta$ abound in topology and are called *simplicial sets*. A similar geometric interpretation applies to each of the categories of finite cardinals that we have considered in \[sec:cats\_fin\_cards\], leading to several corresponding variants of the notion of simplicial set:
\[def:varns\_on\_simpl\_obj\] Let ${\ensuremath{\mathscr C}\xspace}$ be a category.
1. 1. A **degenerative object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
2. A **codegenerative object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s \rightarrow {\ensuremath{\mathscr C}\xspace}$.
2. 1. An **(augmented) simplicial object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
2. An **(augmented) cosimplicial object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
3. 1. A **permutative object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
2. A **copermutative object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \rightarrow {\ensuremath{\mathscr C}\xspace}$.
4. 1. A **symmetric degenerative object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
2. A **symmetric codegenerative object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr C}\xspace}$.
5. (Barr [@Barr], Grandis [@Gra:Symm])
1. An **(augmented) symmetric simplicial object**[^7] $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
2. An **(augmented) symmetric cosimplicial object** $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$.
For brevity, we will omit the modifier “augmented” when employing these terms within the present paper. The *category of degenerative objects* in ${\ensuremath{\mathscr C}\xspace}$ is defined as the functor category $[{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}_s,{\ensuremath{\mathscr C}\xspace}]$. Similarly, each of the listed notions determines an associated category in which the morphisms are arbitrary natural transformations.
Given a category ${\ensuremath{\mathscr C}\xspace}$, any functor $F:{\ensuremath{\mathscr A}\xspace}\rightarrow {\ensuremath{\mathscr B}\xspace}$ determines a functor $[F,{\ensuremath{\mathscr C}\xspace}]:[{\ensuremath{\mathscr B}\xspace},{\ensuremath{\mathscr C}\xspace}] \rightarrow [{\ensuremath{\mathscr A}\xspace},{\ensuremath{\mathscr C}\xspace}]$ between the associated categories of ${\ensuremath{\mathscr C}\xspace}$-valued functors. In particular, the inclusions $$\label{eq:incl_cats_fincards}
\xymatrix{
& {\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s \ar@{^{(}->}[d] \ar@{^{(}->}[r] & {\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}\ar@{^{(}->}[d]\\
{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \ar@{^{(}->}[r] & {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \ar@{^{(}->}[r] & {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}}$$ induce functors between the various functor categories defined in \[def:varns\_on\_simpl\_obj\]. For example, every symmetric degenerative object carries the structure of a permutative object.
\[rem:symm\_grp\_actions\] By definition, a **graded object** $C$ in a category ${\ensuremath{\mathscr C}\xspace}$ is a sequence of objects $C_n$ in ${\ensuremath{\mathscr C}\xspace}$ indexed by the finite cardinals $n$. Observe that a copermutative object $C$ in ${\ensuremath{\mathscr C}\xspace}$ is equivalently described as a graded object $C$ in ${\ensuremath{\mathscr C}\xspace}$ equipped with a sequence of group homomorphisms $S_n \rightarrow {\ensuremath{\textnormal{Aut}}}_{\ensuremath{\mathscr C}\xspace}(C_n)$ from the **symmetric groups** $S_n = {\ensuremath{\textnormal{Aut}}}_{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}(n)$ into the automorphism groups ${\ensuremath{\textnormal{Aut}}}_{\ensuremath{\mathscr C}\xspace}(C_n)$ of the objects $C_n$ of ${\ensuremath{\mathscr C}\xspace}$ (§\[sec:notn\]). Dually, a permutative object $C$ in ${\ensuremath{\mathscr C}\xspace}$ is a graded object equipped with group homomorphisms $S^{{\mbox{\textnormal{\scriptsize op}}}}_n \rightarrow {\ensuremath{\textnormal{Aut}}}_{\ensuremath{\mathscr C}\xspace}(C_n)$ where $S^{{\mbox{\textnormal{\scriptsize op}}}}_n$ is the opposite of the symmetric group. But every group $G$ is isomorphic to its opposite $G^{{\mbox{\textnormal{\scriptsize op}}}}$ via the map $(-)^{-1}:G \rightarrow G^{{\mbox{\textnormal{\scriptsize op}}}}$, so copermutative objects are in bijective correspondence with permutative objects. From another perspective, this bijective correspondence is induced by an identity-on-objects isomorphism of categories $${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \xrightarrow{(-)^{-1}} {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b^{{\mbox{\textnormal{\scriptsize op}}}}$$ given on arrows by $\xi \mapsto \xi^{-1}$.
Example: The degenerative object of iterated tangent functors {#exa:deg_obj_it_tgt}
-------------------------------------------------------------
Given a tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$, we saw in \[sec:tng\_func\_symm\_cosgrp\] that the tangent endofunctor $T:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ carries the structure of a symmetric semigroup $(T,\ell,c)$ in the opposite $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{\mbox{\textnormal{\scriptsize op}}}}$ of the category of endofunctors on ${\ensuremath{\mathscr X}\xspace}$. Hence by \[thm:univ\_symm\_mon\_sgrp\], this symmetric semigroup determines a corresponding strict monoidal functor $T^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{\mbox{\textnormal{\scriptsize op}}}}$ sending each finite cardinal $n$ to the $n$-th iterate $T^n$ of $T$. This functor is an example of a symmetric codegenerative object in $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{\mbox{\textnormal{\scriptsize op}}}}$, equivalently, a symmetric degenerative object ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ in the category of endofunctors on ${\ensuremath{\mathscr X}\xspace}$.
Symmetric simplicial objects by generators and relations
--------------------------------------------------------
It is well-known that the category of finite ordinals has a convenient presentation by generators and relations, leading to a familiar equivalent way of defining simplicial sets in terms of *face* and *degeneracy* maps; see, e.g. [@MacL VII.5]. Barr [@Barr] and Grandis [@Gra:Symm] gave an analogous presentation of the larger category of finite cardinals ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ in terms of the following larger collection of generators:
\[def:gens\]
1. We denote by $$\label{eq:epsilon}\varepsilon^n_i:n + 1 \rightarrow n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n)$$ the map $(i - 1) + \mu + (n - i):n + 1 \rightarrow n$ in the notation of \[sec:cats\_fin\_cards\], where $\mu:1 + 1 \rightarrow 1$ is the multiplication carried by $1$ (\[sec:univ\_mon\_sgrps\]). We call these **codegeneracy maps**.
2. We denote by $$\label{eq:delta}\delta^n_i:n \rightarrow n + 1\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n + 1)$$ the map $(i - 1) + \eta + (n - i + 1):n \rightarrow n + 1$ in the notation of \[sec:cats\_fin\_cards\], where $\eta:0 \rightarrow 1$ is the unit carried by $1$ (\[sec:univ\_mon\_sgrps\]). We call these **coface maps**.
3. We denote by $$\label{eq:sigma}\sigma^n_i:n \rightarrow n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n - 1)$$ the map $(i - 1) + \sigma + (n - i - 1):n \rightarrow n$ in the notation of \[sec:cats\_fin\_cards\], where $\sigma:1 + 1 \rightarrow 1 + 1$ is the symmetry carried by $1$ (\[rem:comm\_sgrp\_yields\_symm\_sgrp\]). We call these **symmetry maps**.
We shall omit the superscripts $n$ when they are clear from the context.
The following theorem is well-known; for example, a proof is given in [@MacL VII.5].
The category ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}$ of finite ordinals and order preserving maps can be presented by generators and relations (in the sense of [[@MacL II.8]]{.nodecor}) as follows:
1. Generators: The maps $\varepsilon^n_i, \delta^n_i$ of \[def:gens\].
2. Relations: The following [**pure codegeneracy relations**]{.nodecor}: $$\label{eq:codegen_rels}\varepsilon_i{\ensuremath{\mathop{;}}}\varepsilon_j = \varepsilon_{j+1}{\ensuremath{\mathop{;}}}\varepsilon_i\;\;\;\;\;\;(i \leqslant j)$$ together with the following [**pure coface relations**]{.nodecor}: $$\label{eq:coface_rels}\delta_j{\ensuremath{\mathop{;}}}\delta_i = \delta_i{\ensuremath{\mathop{;}}}\delta_{j+1}\;\;\;\;\;\;(i \leqslant j)$$ as well as the following [**coface-codegeneracy relations**]{.nodecor}: $$\label{eq:cofcod_rels}\delta_i{\ensuremath{\mathop{;}}}\varepsilon_j =
\begin{cases}
\varepsilon_{j-1}{\ensuremath{\mathop{;}}}\delta_i & (i < j)\\
1 & (i = j,\;i = j+1)\\
\varepsilon_j{\ensuremath{\mathop{;}}}\delta_{i-1} & (i > j+1).
\end{cases}$$
\[rem:finords\_gens\_rels\] One can also present the category ${\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}_s$ of finite ordinals and order preserving surjections by generators and relations, namely the codegeneracies $\varepsilon^n_i$ and the pure codegeneracy relations [@MacL VII.5, Exercise 3].
By adding the symmetry maps as additional generators, together with further relations, Barr and Grandis established the following variation on the preceding theorem:
Theorem [(Barr [@Barr], Grandis [@Gra:Symm 4.2])]{.nodecor} {#thm:fincard_gens_rels}
-----------------------------------------------------------
\[rem:moore\_pres\_symm\_grp\] Grandis [@Gra:Symm §3] notes that for a fixed finite cardinal $n$, the maps $\sigma^n_i$ generate the symmetric group $S_n$, and the Moore relations constitute a classical presentation of this group by generators and relations.
By discarding the coface maps and all the relations involving them, we shall now establish an analogous presentation of ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ in terms of the codegeneracy and symmetry maps:
\[thm:fincards\_gens\_rels\] The category ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ of finite cardinals and surjections can be presented by generators and relations as follows:
1. Generators: The maps $\varepsilon^n_i,\sigma^n_i$ of \[def:gens\].
2. Relations: The pure codegeneracy relations , the Moore relations , and the codegeneracy-symmetry relations .
Let ${\ensuremath{\mathscr C}\xspace}$ denote the category presented by the given (formal) generators and relations (per [@MacL II.8]), with objects all finite cardinals. We will not distinguish notationally between the morphisms $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ and the generators in ${\ensuremath{\mathscr C}\xspace}$ that bear the same names.
First we show that ${\ensuremath{\mathscr C}\xspace}$ carries the structure of a strict monoidal category. In order to define a functor $+:{\ensuremath{\mathscr C}\xspace}\times {\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$, given on objects by addition, it suffices to define functors $(-) + m:{\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ and $m + (-):{\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ for all $m \in {\ensuremath{\mathop{\textnormal{ob}}}}{\ensuremath{\mathscr C}\xspace}$ and check that they satisfy the compatibility condition in [@MacL II.3.1]. On generators, we define $$\varepsilon^n_i + m := \varepsilon^{n+m}_i,\;\;\sigma^n_i + m := \sigma^{n+m}_i,\;\;m + \varepsilon^n_i := \varepsilon^{m+n}_{m+i},\;\;m + \sigma^n_i := \sigma^{m+n}_{m+i},$$ and it follows immediately from the relations , , that these assignments respect these relations (in the sense of [@MacL II.8.1]) and so define functors as needed. We must prove that the compatibility condition $$(\alpha + m) {\ensuremath{\mathop{;}}}\: (n' + \beta) = (n + \beta) {\ensuremath{\mathop{;}}}\: (\alpha + m')$$ holds for arbitrary morphisms $\alpha:n \rightarrow n'$ and $\beta:m \rightarrow m'$ in ${\ensuremath{\mathscr C}\xspace}$, but it suffices to verify this equation in the cases where $\alpha,\beta$ are generators, and in each of these (four) cases the needed equation reduces to an instance of one of the relations in , , .
The resulting functor $+:{\ensuremath{\mathscr C}\xspace}\times {\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ clearly satisfies the associativity law on objects. The verification of the associativity law $$(\alpha + \beta) + \gamma = \alpha + (\beta + \gamma):n + m + k \rightarrow n' + m' + k'$$ for arrows $\alpha:n \rightarrow n'$, $\beta:m \rightarrow m'$, and $\gamma:k \rightarrow k'$ in ${\ensuremath{\mathscr C}\xspace}$ reduces to verification of the equations $$(\alpha + m) + k = \alpha + (m + k),\;\;(n' + \beta) + k = n' + (\beta + k),\;\;(n' + m') + \gamma = n' + (m' + \gamma).$$ It suffices to consider the cases where $\alpha,\beta,\gamma$ are generators, and then the equations are immediate from the definition of $+$. Verification of the unit laws for $({\ensuremath{\mathscr C}\xspace},+,0)$ reduces to showing that the functors $0 + (-),\;(-) + 0:{\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ are merely the identity functor, but this is trivially verified on generators.
We claim that the object $1$ of ${\ensuremath{\mathscr C}\xspace}$ carries the structure of a symmetric semigroup $S = (1,\bar{\mu},\bar{\sigma})$ with $\bar{\mu} := \varepsilon^1_1:2 \rightarrow 1$ and $\bar{\sigma} := \sigma^2_1:2 \rightarrow 2$. The associativity law $(1 + \bar{\mu}){\ensuremath{\mathop{;}}}\bar{\mu} = (\bar{\mu} + 1){\ensuremath{\mathop{;}}}\bar{\mu}$ is precisely the equation $\varepsilon^2_2 {\ensuremath{\mathop{;}}}\varepsilon^1_1 = \varepsilon^2_1 {\ensuremath{\mathop{;}}}\varepsilon^1_1$, which is one of the pure codegeneracy relations . The equations making $\bar{\sigma}$ a symmetry on $1$ are instances of the Moore relations . The equation relating $\bar{\mu}$ and $\bar{\sigma}$ is precisely the equation $\varepsilon^2_1{\ensuremath{\mathop{;}}}\sigma^2_1 = \sigma^3_2{\ensuremath{\mathop{;}}}\sigma^3_1{\ensuremath{\mathop{;}}}\varepsilon^2_2$, which is one of the codegeneracy-symmetry relations . The commutative law for $S$ is an instance of the last codegeneracy-symmetry relation .
Hence by \[thm:univ\_symm\_mon\_sgrp\](ii) there is a unique strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr C}\xspace}$ with $S^\sharp(1) = 1$, $S^\sharp(\mu) = \bar{\mu}$, and $S^\sharp(\sigma) = \bar{\sigma}$. Note that $S^\sharp$ is identity-on-objects and sends the morphisms $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ to the similarly named generators in ${\ensuremath{\mathscr C}\xspace}$. Indeed, the definition (\[def:gens\]) of the morphisms $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ entails that the strict monoidal functor $S^\sharp$ sends them to $$S^\sharp(\varepsilon^n_i) \;\;=\;\; (i - 1) + \varepsilon^1_1 + (n - i) \;\;=\;\; \varepsilon^{n}_i\;\;\;:\;\;\;n + 1 \rightarrow n$$ $$S^\sharp(\sigma^n_i) \;\;=\;\; (i - 1) + \sigma^2_1 + (n - i - 1) \;\;=\;\; \sigma^n_i\;\;\;:\;\;\;n \rightarrow n\;,$$ respectively (using the definitions of $\bar{\mu}$, $\bar{\sigma}$, and $+$).
Next we define an identity-on-objects functor $M:{\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ by sending the generators $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\mathscr C}\xspace}$ to the similarly named morphisms $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$. This assignment respects the relations defining ${\ensuremath{\mathscr C}\xspace}$, simply because the morphisms $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \hookrightarrow {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ satisfy these relations (by \[thm:fincard\_gens\_rels\]).
The composite functor $M {\ensuremath{\mathop{;}}}S^\sharp:{\ensuremath{\mathscr C}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ preserves the generators $\varepsilon^n_i$ and $\sigma^n_i$ and so (by the universal property of ${\ensuremath{\mathscr C}\xspace}$) must be the identity functor. Hence $M$ is faithful. We claim that $M$ is also full (and hence is an isomorphism). Firstly, every morphism in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ can be expressed as a composite $\tau {\ensuremath{\mathop{;}}}\alpha:n \rightarrow m$ where $\tau \in S_n$ is a permutation and $\alpha:n \rightarrow m$ is order preserving [@Gra:Symm §3], and then $\alpha$ is necessarily surjective. But by \[rem:moore\_pres\_symm\_grp\] we can express $\tau$ as a composite of symmetry maps $\sigma^n_i$, and by \[rem:finords\_gens\_rels\] we can express $\alpha$ as a composite of codegeneracy maps. Therefore the symmetries and codegeneracies $\sigma^n_i,\varepsilon^n_i$ generate ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$, so since they lie in the image of $M$ it follows that $M$ is full.
As corollaries to the above theorems, we obtain not only the classical description of cosimplicial objects in terms of coface and codegeneracy morphisms but also analogous descriptions of symmetric cosimplicial objects and symmetric codegenerative objects, as follows:
\[thm:pres\_cosimpl\_obs\_gens\_rels\] Let ${\ensuremath{\mathscr C}\xspace}$ be a category.
1. A cosimplicial object $C:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ is equivalently given by a graded object $C$ in ${\ensuremath{\mathscr C}\xspace}$ equipped with morphisms $$\label{eq:grob_cod}\varepsilon^n_i:C_{n + 1} \rightarrow C_n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n)$$ $$\label{eq:grob_cof}\delta^n_i:C_n \rightarrow C_{n + 1}\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n + 1)$$ in ${\ensuremath{\mathscr C}\xspace}$ that satisfy the equations , , .
2. [(Barr [@Barr], Grandis [@Gra:Symm])]{.nodecor} A symmetric cosimplicial object $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ is equivalently given by a graded object $C$ in ${\ensuremath{\mathscr C}\xspace}$ equipped with morphisms $\varepsilon^n_i$ and $\delta^n_i$ as in , as well as morphisms $$\label{eq:grob_symm}\sigma^n_i:C_n \rightarrow C_n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n - 1)$$ in ${\ensuremath{\mathscr C}\xspace}$ such that these morphisms satisfy the equations , , , , , .
3. A symmetric codegenerative object $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr C}\xspace}$ is equivalently given by a graded object $C$ in ${\ensuremath{\mathscr C}\xspace}$ equipped with morphisms $\varepsilon^n_i$, $\sigma^n_i$ as in , such that these morphisms satisfy the equations , , .
We call the structural morphisms $\varepsilon^n_i,\delta^n_i,\sigma^n_i$ in \[thm:pres\_cosimpl\_obs\_gens\_rels\] **codegeneracies**, **cofaces**, and **symmetries**, respectively, just like their similarly notated counterparts in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$. Dually, a symmetric *simplicial* object $C$ carries **degeneracy morphisms** $\varepsilon^n_i:C_n \rightarrow C_{n+1}$, **face morphisms** $\delta^n_i:C_{n + 1} \rightarrow C_n$, and **symmetries** $\sigma^n_i:C_n \rightarrow C_n$.
The codegenerative object determined by a symmetric semigroup {#sec:symm_cod_obj_det_symm_sgrp}
-------------------------------------------------------------
Given any symmetric semigroup $(S,m,s)$ in a strict monoidal category ${\ensuremath{\mathscr V}\xspace}$, the corresponding strict monoidal functor $S^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr V}\xspace}$ (\[thm:univ\_symm\_mon\_sgrp\]) is an example of a symmetric codegenerative object in ${\ensuremath{\mathscr V}\xspace}$. Its underlying graded object consists of the $n$-fold monoidal powers $S^n$ of $S$. Since $S^\sharp$ is strict monoidal, the definitions of the generators of ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ in \[def:gens\] entail that the codegeneracies and symmetries carried by $S^\sharp$ can be expressed as $$\varepsilon^n_i \;=\; S^{i - 1}m S^{n - i}\;\;:\;\;S^{n + 1} \rightarrow S^n$$ $$\sigma^n_i \;=\; S^{i - 1}sS^{n - i - 1}\;\;:\;\;S^n \rightarrow S^n\;.$$
Example: The symmetric degenerative iterated tangent functor {#exa:symm_deg_it_tang}
------------------------------------------------------------
As a special case of \[sec:symm\_cod\_obj\_det\_symm\_sgrp\], we saw in \[exa:deg\_obj\_it\_tgt\] that the tangent functor $T:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ on a tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ carries the structure of a symmetric semigroup $(T,\ell,c)$ in $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{\mbox{\textnormal{\scriptsize op}}}}$ and so determines a symmetric codegenerative object $T^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{\mbox{\textnormal{\scriptsize op}}}}$, or equivalently, a symmetric degenerative object $$\label{eq:symm_deg_it_tang}T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}],\;\;\;\;n \mapsto T^n\;.$$ By \[sec:symm\_cod\_obj\_det\_symm\_sgrp\], the associated degeneracy and symmetry morphisms are the natural transformations $$\label{eq:tng_func_deg}\ell^n_i \;=\; T^{i - 1}\ell T^{n - i}\;\;:\;\;T^n \rightarrow T^{n+1}\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n)$$ $$\label{eq:tng_func_symm}c^n_i \;=\; T^{i - 1} c T^{n - i - 1}\;\;:\;\;T^n \rightarrow T^n\;\;\;\;\;\;(n,i \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n - 1).$$
Example: Codegenerative sets induced by tangent structure {#exa:cod_sets_ind_tngt_str}
---------------------------------------------------------
Let $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ be a tangent category. By transposition, the functor $T^{(-)}$ of determines a functor $$\label{eq:func_symm_cod_objs_in_tng_cat}{\ensuremath{\mathscr X}\xspace}\rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}},{\ensuremath{\mathscr X}\xspace}],\;\;\;\;M \mapsto T^{(-)}M$$ valued in the category of symmetric degenerative objects in ${\ensuremath{\mathscr X}\xspace}$. Explicitly, this functor sends each object $M$ of ${\ensuremath{\mathscr X}\xspace}$ to a symmetric degenerative object $$T^{(-)}M:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr X}\xspace},\;\;\;\;n \mapsto T^nM$$ whose underlying graded object consists of the total spaces $T^nM$ of the iterated tangent bundles of $M$. The degeneracy and symmetry morphisms carried by $T^{(-)}M$ are just the components $\ell^n_i M$, $c^n_i M$ at $M$ of those carried by $T^{(-)}$ (\[eq:tng\_func\_deg\], \[eq:tng\_func\_symm\]).
Fixing an object $E$ of ${\ensuremath{\mathscr X}\xspace}$, the functor induces[^8] a functor $$\label{eq:func_tng_cat_to_symm_cod_sets}{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\textnormal{\textbf{set}}}\xspace}]\;,\;\;\;\;M \mapsto {\ensuremath{\mathscr X}\xspace}(T^{(-)}M,E)$$ valued in the category of symmetric codegenerative sets. Explicitly, this functor sends each object $M$ of ${\ensuremath{\mathscr X}\xspace}$ to a symmetric codegenerative set ${\ensuremath{\mathscr X}\xspace}(T^{(-)}M,E)$ whose underlying graded set consists of the hom-sets ${\ensuremath{\mathscr X}\xspace}(T^nM,E)$. The associated codegeneracies and symmetries are the mappings $$\varepsilon^n_i = {\ensuremath{\mathscr X}\xspace}(\ell^n_i M,E)\;\;:\;\;{\ensuremath{\mathscr X}\xspace}(T^{n+1}M,E) \rightarrow {\ensuremath{\mathscr X}\xspace}(T^nM,E),\;\;\;\;\omega \mapsto \ell^n_iM {\ensuremath{\mathop{;}}}\omega$$ $$\sigma^n_i = {\ensuremath{\mathscr X}\xspace}(c^n_i M,E)\;\;:\;\;{\ensuremath{\mathscr X}\xspace}(T^nM,E) \rightarrow {\ensuremath{\mathscr X}\xspace}(T^nM,E),\;\;\;\;\omega \mapsto c^n_i M {\ensuremath{\mathop{;}}}\omega$$ given by precomposing with the degeneracies $\ell^n_i M:T^n M \rightarrow T^{n+1}M$ and the symmetries $c^n_i M:T^n M \rightarrow T^n M$.
Note that if ${\ensuremath{\mathscr X}\xspace}$ is a Cartesian tangent catgory and $E$ carries the structure of a commutative monoid (resp. abelian group) object in ${\ensuremath{\mathscr X}\xspace}$, then the representable presheaf ${\ensuremath{\mathscr X}\xspace}(-,E):{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\textnormal{\textbf{set}}}\xspace}$ lifts to a presheaf valued in the category ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}$ of commutative monoids (resp. the category ${\ensuremath{\textnormal{\textbf{ab}}}\xspace}$ of abelian groups). Hence the functor lifts to a functor $${\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\textnormal{\textbf{cmon}}}\xspace}]\;,\;\;\;\;M \mapsto {\ensuremath{\mathscr X}\xspace}(T^{(-)}M,E)$$ valued in the category of symmetric codegenerative objects in ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}$ (resp. ${\ensuremath{\textnormal{\textbf{ab}}}\xspace}$).
Presenting symmetric cosimplicial objects by fundamental cofaces {#sec:presFundamentalCoface}
================================================================
The coface morphisms $\delta^n_i:C_n \rightarrow C_{n + 1}$ carried by a symmetric cosimplicial object $C$ can be expressed in terms of the **fundamental cofaces** $\delta^n_1$ by repeated application of the equation $\delta_{i+1} = \delta_i{\ensuremath{\mathop{;}}}\sigma_i$ of . This leads to the following new succinct equational presentation of symmetric cosimplicial objects, which will be useful in establishing the symmetric cosimplicial structure that engenders the de Rham complex:
\[thm:fund\_cof\_pres\_symm\_cosimpl\_obj\] A symmetric cosimplicial object $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ in a category ${\ensuremath{\mathscr C}\xspace}$ is equivalently given by a graded object $C$ in ${\ensuremath{\mathscr C}\xspace}$ equipped with morphisms $\varepsilon^n_i$, $\sigma^n_i$ as in , together with a sequence of morphisms $$\label{eq:fund_cof}\delta^n_1:C_n \rightarrow C_{n+1}\;\;\;\;\;\;(n \in {\ensuremath{\mathbb N}\xspace})$$ such that the equations , , are satisfied along with the following further equations: $$\label{eq:fund_cofcod_eqs}\delta_1{\ensuremath{\mathop{;}}}\varepsilon_1 = 1\;\;\;\;\;\;\;\;\delta_1{\ensuremath{\mathop{;}}}\varepsilon_{j+1} = \varepsilon_j{\ensuremath{\mathop{;}}}\delta_1$$ $$\label{eq:fund_cofsymm_eqs}\delta_1{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_1 = \delta_1{\ensuremath{\mathop{;}}}\delta_1\;\;\;\;\;\;\;\;\delta_1{\ensuremath{\mathop{;}}}\sigma_{i + 1} = \sigma_i{\ensuremath{\mathop{;}}}\delta_1\;.$$ Therefore, in view of \[thm:pres\_cosimpl\_obs\_gens\_rels\](iii), a symmetric cosimplicial object $C$ in ${\ensuremath{\mathscr C}\xspace}$ is equivalently given by a symmetric codegenerative object $C$ equipped with a sequence of morphisms satisfying the equations , .
Before proving this, let us adopt the following notational conventions.
\[def:notn\_perms\]
1. By abuse of notation[^9] we write the elements of each finite ordinal $n \in {\ensuremath{\mathbb N}\xspace}$ in ascending order as $1,2,3,...,n$ (rather than $0,1,...,n-1$).
2. For each $n \in {\ensuremath{\mathbb N}\xspace}$ and each $i \in n$, let $\sigma^n_{(i)}$ denote the permutation $$\sigma^n_{(i)} = \bigl(i (i-1) (i-2) ... 3 2 1\bigr) \in S_n$$ written in cycle notation (i.e., $i \mapsto i-1$, $i-1 \mapsto i-2$, etc.) on the elements $1,...,n$ of the ordinal $n$. We sometimes omit the superscript $n$, writing just $\sigma_{(i)}$.
Observe that $\sigma_{(i)}$ can be written as a composite $$\sigma_{(i)} = \sigma_1{\ensuremath{\mathop{;}}}\sigma_2 {\ensuremath{\mathop{;}}}... {\ensuremath{\mathop{;}}}\sigma_{i-1}$$ of the transpositions $\sigma_j = (j(j+1))$ defined in \[def:gens\]. When $i = 1$ we interpret the resulting empty composite as the identity map on $n$, so that $\sigma_{(1)} = 1$.
\[thm:main\_fund\_coface\_lemma\] Let $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr C}\xspace}$ be a symmetric codegenerative object (\[def:varns\_on\_simpl\_obj\]) equipped with a sequence of morphisms $\delta^n_1:C_n \rightarrow C_{n+1}$ $(n \in {\ensuremath{\mathbb N}\xspace})$ satisfying the equations , . Then $C$ extends uniquely to a symmetric cosimplicial object $C':{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\mathscr C}\xspace}$ whose fundamental cofaces are the given morphisms $\delta^n_1$. Explicitly, the cofaces of $C'$ can be expressed in terms of the fundamental cofaces as $$\label{eq:cof_via_fund_cof}\delta^n_i = \delta^n_1 {\ensuremath{\mathop{;}}}\sigma^{n+1}_{(i)}\;\;\;\;\;\;(n \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n + 1)$$ where we write $\sigma^{n+1}_{(i)}:C_{n+1} \rightarrow C_{n+1}$ to denote the image of the automorphism $\sigma^{n+1}_{(i)}:n+1 \rightarrow n+1$ (\[def:notn\_perms\]) under the functor $C:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr C}\xspace}$.
In view of \[thm:fincard\_gens\_rels\], repeated application of the equation $\delta^n_{i+1} = \delta^n_i{\ensuremath{\mathop{;}}}\sigma^{n+1}_i$ of shows that the maps $\delta^n_1,\varepsilon^n_i,\sigma^n_i$ generate ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$. Hence, in view of \[thm:fincards\_gens\_rels\], the uniqueness of $C'$ is immediate if $C'$ exists. Defining the coface morphisms $\delta^n_i$ for $C'$ by way of the equation , it suffices (by \[thm:pres\_cosimpl\_obs\_gens\_rels\]) to show that these satisfy the relations , , when taken together with the morphisms $\varepsilon^n_i,\sigma^n_i$ carried by $C$.
In order to verify the first coface-codegeneracy relation $\delta_i{\ensuremath{\mathop{;}}}\varepsilon_j = \varepsilon_{j-1}{\ensuremath{\mathop{;}}}\delta_i$ $(i < j)$, we compute as follows, applying the second equation in and then repeatedly applying the first equation in : $$\begin{array}{lllll}
\varepsilon_{j-1}{\ensuremath{\mathop{;}}}\delta_i & = & \varepsilon_{j-1}{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1} & = & \delta_1{\ensuremath{\mathop{;}}}\varepsilon_j{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1}\\
& = & \delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1}{\ensuremath{\mathop{;}}}\varepsilon_j & = & \delta_i{\ensuremath{\mathop{;}}}\varepsilon_j\;.
\end{array}$$
Next, we prove the second coface-codegeneracy relation $\delta^n_i{\ensuremath{\mathop{;}}}\varepsilon^n_i = 1$ by induction on $i$ (with $n$ fixed). In the base case where $i = 1$, this holds by assumption. For the inductive step, suppose that $\delta^n_i{\ensuremath{\mathop{;}}}\varepsilon^n_i = 1$ holds for a given index $i$. Then by applying one of the codegeneracy-symmetry relations and the fact that $(\sigma_{i+1})^{-1} = \sigma_{i+1}$ we compute that $$\begin{array}{lllll}
\delta_{i+1}{\ensuremath{\mathop{;}}}\varepsilon_{i+1} & = &\delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1}{\ensuremath{\mathop{;}}}\sigma_i{\ensuremath{\mathop{;}}}\varepsilon_{i+1} & & \\
& = & \delta_1{\ensuremath{\mathop{;}}}\sigma_{(i)}{\ensuremath{\mathop{;}}}\sigma_i{\ensuremath{\mathop{;}}}\varepsilon_{i+1} & = & \delta_1{\ensuremath{\mathop{;}}}\sigma_{(i)}{\ensuremath{\mathop{;}}}\sigma_{i+1}{\ensuremath{\mathop{;}}}\varepsilon_i{\ensuremath{\mathop{;}}}\sigma_i\;.\\
\end{array}$$ in ${\ensuremath{\mathscr C}\xspace}$. But $\sigma_{(i)}{\ensuremath{\mathop{;}}}\sigma_{i+1} = \sigma_{i+1}{\ensuremath{\mathop{;}}}\sigma_{(i)}$ in ${\ensuremath{\mathscr C}\xspace}$ since the the permutations $\sigma_{(i)} = (i(i-1)...321)$ and $\sigma_{i+1} = ((i+1)(i+2))$ are disjoint cycles. Hence we compute as follows, applying the inductive hypothesis and one of the relations in : $$\begin{array}{lllllll}
\delta_{i+1}{\ensuremath{\mathop{;}}}\varepsilon_{i+1} & = &\delta_1{\ensuremath{\mathop{;}}}\sigma_{i+1}{\ensuremath{\mathop{;}}}\sigma_{(i)}{\ensuremath{\mathop{;}}}\varepsilon_i{\ensuremath{\mathop{;}}}\sigma_i & = & \sigma_i{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_{(i)}{\ensuremath{\mathop{;}}}\varepsilon_i{\ensuremath{\mathop{;}}}\sigma_i & &\\
& = & \sigma_i{\ensuremath{\mathop{;}}}\delta_i{\ensuremath{\mathop{;}}}\varepsilon_i{\ensuremath{\mathop{;}}}\sigma_i & = & \sigma_i{\ensuremath{\mathop{;}}}\sigma_i & = & 1\;.\\
\end{array}$$
The third coface-codegeneracy relation $\delta_{j+1}{\ensuremath{\mathop{;}}}\varepsilon_j = 1$ now follows, using the last codegeneracy-symmetry relation : $\delta_{j+1}{\ensuremath{\mathop{;}}}\varepsilon_j = \delta_1{\ensuremath{\mathop{;}}}\sigma_{(j+1)}{\ensuremath{\mathop{;}}}\varepsilon_j = \delta_1{\ensuremath{\mathop{;}}}\sigma_{(j)}{\ensuremath{\mathop{;}}}\sigma_j{\ensuremath{\mathop{;}}}\varepsilon_j = \delta_1 {\ensuremath{\mathop{;}}}\sigma_{(j)}{\ensuremath{\mathop{;}}}\varepsilon_j = \delta_j {\ensuremath{\mathop{;}}}\varepsilon_j = 1$.
We next prove the last coface-codegeneracy relation $\delta^{n+1}_i{\ensuremath{\mathop{;}}}\varepsilon^{n+1}_j = \varepsilon^n_j{\ensuremath{\mathop{;}}}\delta^n_{i-1}$ $(i > j+1)$. By definition, the left-hand side is $\delta_1{\ensuremath{\mathop{;}}}\sigma_{(i)}{\ensuremath{\mathop{;}}}\varepsilon_j:C_{n+1} \rightarrow C_{n+1}$, whereas by applying the second equation in we can express the right-hand side as $\varepsilon_j{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_{(i-1)} = \delta_1{\ensuremath{\mathop{;}}}\varepsilon_{j+1}{\ensuremath{\mathop{;}}}\sigma_{(i-1)}$. Hence it suffices to show that $$\sigma_{(i)}{\ensuremath{\mathop{;}}}\varepsilon_j = \varepsilon_{j+1}{\ensuremath{\mathop{;}}}\sigma_{(i-1)}\;\;\;:\;\;\;n+2 \rightarrow n+1$$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$, but it is straightforward to verify that the left- and right-hand sides of this equation both denote the map $\phi$ given by $$\phi(x) = \begin{cases}
i-1 & (x = 1)\\
x-1 & (1 < x \leqslant j+1)\\
x-2 & (j+1 < x \leqslant i)\\
x-1 & (i < x).
\end{cases}$$
Next we verify the first coface-symmetry relation $\delta_j^n{\ensuremath{\mathop{;}}}\sigma^{n+1}_i = \sigma^n_i{\ensuremath{\mathop{;}}}\delta^n_j$ $(i < j-1)$. By definition, the left-hand side is $\delta_1{\ensuremath{\mathop{;}}}\sigma_{(j)}{\ensuremath{\mathop{;}}}\sigma_i$, whereas by applying the second equation in we can express the right-hand side as $\sigma_i{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_{(j)} = \delta_1{\ensuremath{\mathop{;}}}\sigma_{i+1}{\ensuremath{\mathop{;}}}\sigma_{(j)}$. Hence it suffices to show that $$\sigma_{(j)}{\ensuremath{\mathop{;}}}\sigma_i = \sigma_{i+1}{\ensuremath{\mathop{;}}}\sigma_{(j)}\;\;\;:\;\;\;n+1 \rightarrow n+1$$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$, but it is straightforward to verify that the left- and right-hand sides of this equation both denote the map $\phi$ given by $$\phi(x) = \begin{cases}
j & (x = 1)\\
i+1 & (x = i+1)\\
i & (x = i+2)\\
x-1 & (1 < x \leqslant j\;\text{and}\;x \not\in \{i+1,i+2\})\\
x & (x > j)
\end{cases}$$
The second coface-symmetry relation $\delta_i{\ensuremath{\mathop{;}}}\sigma_i = \delta_{i+1}$ is almost immediate from the way that we have defined the cofaces, since $\delta_i{\ensuremath{\mathop{;}}}\sigma_i = \delta_1{\ensuremath{\mathop{;}}}\sigma_{(i)}{\ensuremath{\mathop{;}}}\sigma_i = \delta_1{\ensuremath{\mathop{;}}}\sigma_{(i+1)} = \delta_{i+1}$.
Next we establish the third coface-symmetry relation $\delta^n_j{\ensuremath{\mathop{;}}}\sigma^{n+1}_i = \sigma^n_{i-1}{\ensuremath{\mathop{;}}}\delta^n_j$ $(i > j)$. By definition $\delta_j{\ensuremath{\mathop{;}}}\sigma_i = \delta_1{\ensuremath{\mathop{;}}}\sigma_{(j)}{\ensuremath{\mathop{;}}}\sigma_i$ in ${\ensuremath{\mathscr C}\xspace}$, but since $i > j$ the permutations $\sigma_{(j)} = (j(j-1)...321)$ and $\sigma_i = (i(i+1))$ are disjoint cycles and hence commute. Thus $\sigma_{(j)}{\ensuremath{\mathop{;}}}\sigma_i = \sigma_i{\ensuremath{\mathop{;}}}\sigma_{(j)}$ in ${\ensuremath{\mathscr C}\xspace}$. Hence we can compute as follows, applying the second equation in with the knowledge that $i > 1$ (since $i > j \geqslant 1$): $$\delta_j{\ensuremath{\mathop{;}}}\sigma_i = \delta_1{\ensuremath{\mathop{;}}}\sigma_i{\ensuremath{\mathop{;}}}\sigma_{(j)} = \sigma_{i-1}{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_{(j)} = \sigma_{i-1}{\ensuremath{\mathop{;}}}\delta_j\;.$$
Finally, let us verify the pure coface relations $\delta^n_j{\ensuremath{\mathop{;}}}\delta^{n+1}_i = \delta^n_i {\ensuremath{\mathop{;}}}\delta^{n+1}_{j+1}$, where $i \leqslant j$. By repeatedly applying the equations in we deduce that $$\begin{array}{lllll}
\delta_j{\ensuremath{\mathop{;}}}\delta_i & = & \delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}... {\ensuremath{\mathop{;}}}\sigma_{j - 1}{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1} & & \\
& = & \delta_1{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_2 {\ensuremath{\mathop{;}}}... {\ensuremath{\mathop{;}}}\sigma_j{\ensuremath{\mathop{;}}}\sigma_1 {\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1} & = & \delta_1{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_2{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_j{\ensuremath{\mathop{;}}}\sigma_{(i)}\\
\delta_i{\ensuremath{\mathop{;}}}\delta_{j+1} & = & \delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_{i-1}{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_j & = & \delta_1{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_2{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_i{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_j\\
& = & \delta_1{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}\sigma_2{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_i{\ensuremath{\mathop{;}}}\sigma_1{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_j & = & \delta_1{\ensuremath{\mathop{;}}}\delta_1{\ensuremath{\mathop{;}}}\sigma_{(i+1)}{\ensuremath{\mathop{;}}}\sigma_{(j+1)}
\end{array}$$ in ${\ensuremath{\mathscr C}\xspace}$, so it suffices to show that $$\label{eq:pure_cod_arg}\sigma_2{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_j{\ensuremath{\mathop{;}}}\sigma_{(i)} = \sigma_{(i+1)}{\ensuremath{\mathop{;}}}\sigma_{(j+1)}\;\;\;:\;\;\;n+2 \rightarrow n+2$$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$. Using the fact that $\sigma_2{\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}\sigma_j = ((j+1)j...32)$ in cycle notation on the elements $1,2,...n+2$ of the set $n+2$, it is straightforward to verify that the left- and right-hand sides of both denote the map $\phi$ given by $$\phi(x) = \begin{cases}
i & (x = 1)\\
j+1 & (x = 2)\\
x-2 & (3 \leqslant x \leqslant i + 1)\\
x-1 & (i+1 < x \leqslant j+1)\\
x & (x > j+1).
\end{cases}$$
Using the preceding lemma, we now establish the following equational presentation of the category of finite cardinals, from which Theorem \[thm:fund\_cof\_pres\_symm\_cosimpl\_obj\] then immediately follows:
\[thm:fund\_cof\_pres\_fincard\] The category ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ of finite cardinals and arbitrary maps can be presented by generators and relations as follows:
1. Generators: The codegeneracy and symmetry maps $\varepsilon^n_i, \sigma^n_i$ of \[def:gens\] together with the fundamental coface maps $\delta^n_1$ $(n \in {\ensuremath{\mathbb N}\xspace})$ of \[def:gens\].
2. Relations: , , , , .
By \[thm:fincard\_gens\_rels\], the morphisms $\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ satisfy the relations , , , and it is straightforward to verify that the morphisms $\delta^n_1,\varepsilon^n_i,\sigma^n_i$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ also satisfy the relations , . Indeed, it is easy to check that these relations follow from the coface-codegeneracy relations , the coface-symmetry relations , and the pure coface relations , all of which hold in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ (\[thm:fincard\_gens\_rels\]).
Hence it suffices to show that ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ has the relevant universal property [@MacL §8]. But in view of \[thm:pres\_cosimpl\_obs\_gens\_rels\](iii) this universal property is equivalent to the extension property established in Lemma \[thm:main\_fund\_coface\_lemma\].
The symmetric cosimplicial set of sector forms {#sec:cosimpOfSectorForms}
==============================================
Let $E$ be a differential object in a Cartesian tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$, and let $M$ be an object of ${\ensuremath{\mathscr X}\xspace}$. Recall that $\Psi_n(M)$ $(n \in {\ensuremath{\mathbb N}\xspace})$ denotes the set of all sector $n$-forms on $M$ with values in $E$. In the present section we show that the graded set $(\Psi_n(M))_{n \in {\ensuremath{\mathbb N}\xspace}}$ carries the structure of a symmetric cosimplicial commutative monoid.
Remarks on the definition of sector form {#sec:rms_def_sector_form}
----------------------------------------
Recall from Definition \[def:sector\_nform\] that a sector $n$-form on $M$ is a morphism $\omega:T^n M \rightarrow E$ such that for each $j \in \{1,...,n\}$ the equation $a^n_j M{\ensuremath{\mathop{;}}}T\omega = \omega {\ensuremath{\mathop{;}}}\lambda$ holds, where $\lambda:E \rightarrow TE$ is the lift morphism carried by $E$. Using the results of the previous sections, we can get a better understanding of this equation. In particular, $a^n_j$ is the composite transformation $$a^n_j = \left(T^n \overset{\ell_j}{\longrightarrow} T^{n+1} \overset{c_{(j)}}{\longrightarrow} T^{n+1}\right)$$ where $\ell_j = T^{j-1}\ell T^{n-j}$ is the degeneracy morphism carried by the symmetric degenerative object $T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ (\[exa:symm\_deg\_it\_tang\]). The morphism $c_{(j)}$ is the composite $c_{j-1}{\ensuremath{\mathop{;}}}c_{j-2} {\ensuremath{\mathop{;}}}...{\ensuremath{\mathop{;}}}c_2 {\ensuremath{\mathop{;}}}c_1$, where $c_i = T^{i-1}cT^{n - i}$ denotes the symmetry carried by $T^{(-)}$ (\[exa:symm\_deg\_it\_tang\]). Equivalently, $c_{(j)}$ is obtained by applying the functor $T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ to the permutation $\sigma^{n+1}_{(j)} = (j(j-1)...321):n+1 \rightarrow n+1$ (\[def:notn\_perms\]) .
Hence if we define $\alpha^n_j$ as the composite morphism $$\alpha^n_j := \left(n+1 \overset{\sigma^{n+1}_{(j)}}{\longrightarrow} n+1 \overset{\varepsilon^n_j}{\longrightarrow} n\right)$$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$, then $$\begin{minipage}{3.5in}\textit{$a^n_j:T^n \rightarrow T^{n+1}$ is the image of $\alpha^n_j:n+1 \rightarrow n$ under the functor $T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$.}\end{minipage}$$ Concretely, one can readily verify that $\alpha^n_j$ is the mapping given by $$\alpha^n_j(x) = \begin{cases}
j & (x = 1)\\
x-1 & (x \neq 1).\\
\end{cases}$$
\[thm:sector\_nforms\_submonoid\] For each $n \in {\ensuremath{\mathbb N}\xspace}$, the set $\Psi_n(M)$ of sector $n$-forms is a submonoid $$\Psi_n(M) \hookrightarrow {\ensuremath{\mathscr X}\xspace}(T^n M, E)$$ of the commutative monoid of all morphisms $\omega:T^n M \rightarrow E$ in ${\ensuremath{\mathscr X}\xspace}$ (\[sec:cmons\_in\_tngnt\_cat\]). If $E$ is subtractive, then $\Psi_n(M)$ is a subgroup of the abelian group ${\ensuremath{\mathscr X}\xspace}(T^nM,E)$.
Given $\omega,\tau \in \Psi_n(M)$, the sum $\omega + \tau$ in ${\ensuremath{\mathscr X}\xspace}(T^n M,E)$ is a sector $n$-form, since for each $j = 1,...,n$ we can compute as follows, using the fact that ${\ensuremath{\textnormal{\textbf{cmon}}}\xspace}(T):{\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace}) \rightarrow {\ensuremath{\textnormal{\textbf{cmon}}}\xspace}({\ensuremath{\mathscr X}\xspace})$ is an additive functor (\[prop:cmons\_in\_tngt\_cat\]) and the fact that $\lambda:E \rightarrow TE$ is a homomorphism of commutative monoids (\[def:diff\_objs\]): $$\begin{array}{lllll}
a_j M {\ensuremath{\mathop{;}}}T(\omega + \tau) & = & a_j M {\ensuremath{\mathop{;}}}(T(\omega) + T(\tau)) & = & (a_j M {\ensuremath{\mathop{;}}}T(\omega)) + (a_j M {\ensuremath{\mathop{;}}}T(\tau))\\
& = & (\omega {\ensuremath{\mathop{;}}}\lambda) + (\tau {\ensuremath{\mathop{;}}}\lambda) & = & (\omega + \tau) {\ensuremath{\mathop{;}}}\lambda\;.
\end{array}$$ Also, the zero element $0$ of the commutative monoid ${\ensuremath{\mathscr X}\xspace}(T^n M,E)$ is a sector $n$-form since we compute that $a_jM {\ensuremath{\mathop{;}}}T(0) = a_jM {\ensuremath{\mathop{;}}}0 = 0 = 0 {\ensuremath{\mathop{;}}}\lambda$, where each occurrence of $0$ denotes the zero element of the relevant hom-set, again using the additivity of $T$ and the fact that $\lambda$ is a monoid homomorphism. The remaining claim is verified similarly.
Recall that the graded commutative monoid $({\ensuremath{\mathscr X}\xspace}(T^n M,E))_{n \in {\ensuremath{\mathbb N}\xspace}}$ carries the structure of a symmetric codegenerative commutative monoid (\[exa:cod\_sets\_ind\_tngt\_str\]). We now show that this structure restricts to sector forms:
\[thm:cod\_symm\_sector\_forms\] The codegeneracy and symmetry maps $$\varepsilon^n_i = {\ensuremath{\mathscr X}\xspace}(\ell^n_i M,E)\;\;:\;\;{\ensuremath{\mathscr X}\xspace}(T^{n+1}M,E) \rightarrow {\ensuremath{\mathscr X}\xspace}(T^nM,E),\;\;\;\;\omega \mapsto \ell^n_iM {\ensuremath{\mathop{;}}}\omega$$ $$\sigma^n_i = {\ensuremath{\mathscr X}\xspace}(c^n_i M,E)\;\;:\;\;{\ensuremath{\mathscr X}\xspace}(T^nM,E) \rightarrow {\ensuremath{\mathscr X}\xspace}(T^nM,E),\;\;\;\;\omega \mapsto c^n_i M {\ensuremath{\mathop{;}}}\omega$$ carried by the symmetric codegenerative commutative monoid ${\ensuremath{\mathscr X}\xspace}(T^{(-)}M,E)$ (\[exa:cod\_sets\_ind\_tngt\_str\]) restrict to yield homomorphisms $$\varepsilon^n_i\;\;:\;\;\Psi_{n+1}(M) \rightarrow \Psi_n(M)\;\;\;\;\;\;\;\;\;\;\;\;\sigma^n_i\;\;:\;\;\Psi_n(M) \rightarrow \Psi_n(M)$$ between the commutative monoids of sector forms.
Letting $\omega \in \Psi_{n+1}(M)$, we shall show first that $\varepsilon^n_i(\omega) = \ell^n_i M {\ensuremath{\mathop{;}}}\omega$ is a sector $n$-form on $M$. It suffices to show that for each $j \in \{1,...,n\}$ there is some $k \in \{1,...,n+1\}$ such that the following diagram commutes. $$\xymatrix{
T^n M \ar[d]_{a_j M} \ar[r]^{\ell_i M} & T^{n+1} M \ar[d]_{a_k M} \ar[r]^\omega & E \ar[d]^\lambda\\
T^{n+1}M \ar[r]_{T\ell_i M} & T^{n+2}M \ar[r]_{T\omega} & TE
}$$ But the rightmost square commutes since $\omega$ is a sector $(n+1)$-form, so it suffices to obtain the commutativity of the following diagram. $$\xymatrix{
T^n \ar[r]^{\ell_i} \ar[d]_{a_j} & T^{n+1} \ar[d]^{a_k}\\
T^{n+1} \ar[r]_{T\ell_i} & T^{n+2}
}$$ In view of \[sec:rms\_def\_sector\_form\], \[exa:deg\_obj\_it\_tgt\], and \[exa:symm\_deg\_it\_tang\], this diagram is obtained by applying the strict monoidal functor $T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ to the following diagram in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ $$\label{eq:diag_in_fincards_for_cod_sforms}
\xymatrix{
n & & n+1 \ar[ll]_{\varepsilon^n_i}\\
n+1 \ar[u]^{\alpha_j} & & n+2 \ar[u]_{\alpha_k} \ar[ll]^(.45){1+\varepsilon_i\:=\:\varepsilon_{i+1}}
}$$ so it suffices to find $k$ such that this diagram commutes. In the case where $i < j$ we can take $k = j+1$, whereas in the case where $i \geqslant j$ we can take $k = j$, for in each case it is straightforward to verify that both composites in are then equal to the map $\phi$ given by $$\phi(x) = \begin{cases}
j & (x = 1)\\
x-1 & (1 < x \leqslant i + 1)\\
x-2 & (x > i+1).
\end{cases}$$
Next we prove that if $\omega$ is a sector $n$-form on $M$ then $\sigma^n_i(\omega) = c_i^n M {\ensuremath{\mathop{;}}}\omega$ is a sector $n$-form on $M$. Letting $j \in \{1,...,n\}$, it suffices to show that the following diagram commutes: $$\xymatrix{
T^nM \ar[d]_{a_j M} \ar[r]^{c_i M} & T^nM \ar[r]^\omega & E \ar[d]^\lambda\\
T^{n+1}M \ar[r]_{Tc_i M} & T^{n+1}M \ar[r]_{T\omega} & TE
}$$ Again since $\omega$ is a sector $n$-form it suffices to show that there is some $k$ such that the following diagram commutes: $$\xymatrix{
T^n \ar[d]_{a_j} \ar[r]^{c_i} & T^n \ar[d]^{a_k}\\
T^{n+1} \ar[r]_{Tc_i} & T^{n+1}
}$$ But as above, we reason that this diagram is obtained by applying $T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ to the following diagram in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ $$\label{eq:diag_in_fincards_for_symm_sforms}
\xymatrix{
n & & n \ar[ll]_{\sigma_i}\\
n+1 \ar[u]^{\alpha_j} & & n+1 \ar[ll]^{1 + \sigma_i = \sigma_{i+1}} \ar[u]_{\alpha_k}
}$$ and so it suffices to show that this diagram commutes for some $k$.
In the case where $j \notin \{i,i+1\}$ we can take $k = j$, whereas in the case where $j = i$ we can take $k = i + 1$, while in the case where $j = i+1$ we can take $k = i$, for in each of these three cases it is straightforward to verify that both composites in are then equal to the mapping $\phi$ given by $$\phi(x) = \begin{cases}
j & (x = 1)\\
i+1 & (x = i+1)\\
i & (x = i+2)\\
x-1 & (x \notin \{1,i+1,i+2\}).
\end{cases}$$
\[thm:symm\_cod\_struct\_on\_sector\_forms\] There is a symmetric codegenerative commutative monoid $$\Psi(M)\;:\;{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\textnormal{\textbf{cmon}}}\xspace},\;\;\;\;n \mapsto \Psi_n(M)$$ where $\Psi_n(M)$ is the commutative monoid of sector $n$-forms on $M$.
By \[thm:cod\_symm\_sector\_forms\], the graded commutative monoid $(\Psi_n(M))$ is equipped with codegeneracy and symmetry homomorphisms $\varepsilon^n_i$ and $\sigma^n_i$. These are restrictions of the codegeneracy and symmetry maps carried by the symmetric codegenerative set ${\ensuremath{\mathscr X}\xspace}(T^{(-)}M,E)$, so they satisfy the equations listed in \[thm:pres\_cosimpl\_obs\_gens\_rels\](iii). Hence an application of \[thm:pres\_cosimpl\_obs\_gens\_rels\](iii) yields the needed result.
By the results of section \[sec:presFundamentalCoface\], in order to show that the symmetric codegenerative structure on sector forms is part of a symmetric cosimplicial structure, it suffices to define the *fundamental coface* maps $\delta^n_1:\Psi_n(M) \rightarrow \Psi_{n+1}(M)$ and check that they satisfy certain equations (\[thm:main\_fund\_coface\_lemma\]). We now proceed to define these maps.
The fundamental derivative of a sector form {#sec:fund_deriv_sf}
-------------------------------------------
Given a sector $n$-form $\omega:T^n M \rightarrow E$ on $M$, we define the **fundamental derivative** $\delta_1(\omega)$ of $\omega$ as the following composite morphism $$\delta_1(\omega) = \left(T^{n+1}M \xrightarrow{T\omega} TE \xrightarrow{\hat{p}} E\right)$$ where $\hat{p}$ denotes the principal projection associated to the differential object $E$ (\[sec:princ\_projn\]).
\[thm:fund\_deriv\_sector\_form\] The fundamental derivative of a sector $n$-form $\omega$ on $M$ is a sector $(n+1)$-form $\delta_1(\omega)$ on $M$.
Letting $j \in \{1,...,n+1\}$, it suffices to show that the following diagram commutes. $$\label{eq:diag_fund_cof_sf_proof}
\xymatrix{
T^{n+1}M \ar[d]_{a_j M} \ar[r]^{T\omega} & TE \ar[r]^{\hat{p}} & E \ar[d]^{\lambda}\\
T^{n+2}M \ar[r]_{TT\omega} & TTE \ar[r]_{T\hat{p}} & TE
}$$ To this end, we proceed in two cases. First consider the case where $j = 1$. Then $a^{n+1}_j = a^{n+1}_1 = \ell^{n+1}_1 {\ensuremath{\mathop{;}}}c^{n+2}_{{(1)}} = \ell T^n$, so it suffices to show that the following diagram commutes. $$\xymatrix{
T^{n+1}M \ar[d]_{a_1 M\:=\:\ell T^nM} \ar[r]^{T\omega} & TE \ar[d]_{\ell E} \ar[r]^{\hat{p}} & E \ar[d]^\lambda\\
T^{n+2}M \ar[r]_{TT\omega} & TTE \ar[r]_{T\hat{p}} & TE
}$$ But the leftmost square commutes by the naturality of $\ell$, and the rightmost square commutes by properties of differential objects (Proposition \[prop:diffObjs\](v)).
Now consider the case where $j > 1$. We want to show that the periphery of the following diagram commutes. $$\label{eq:diag_case_i_neq_j}
\xymatrix{
T^{n+1}M \ar[d]_{a_jM} \ar[r]^{T\omega} & TE \ar[d]_{T\lambda {\ensuremath{\mathop{;}}}cE} \ar[r]^{\hat{p}} & E \ar[d]^\lambda\\
T^{n+2}M \ar[r]_{TT\omega} & TTE \ar[r]_{T\hat{p}} & TE
}$$ The rightmost square again commutes by properties of differential objects, namely \[prop:diffObjs\](vi).
In order to show that the leftmost square in also commutes, it suffices to show that the following diagram commutes. $$\xymatrix{
T^{n+1}M \ar[rr]^{T\omega} \ar[dd]_{a_j M} \ar[dr]^{Ta_{j-1} M} & & TE \ar[d]^{T\lambda}\\
& T^{n+2}M \ar[dl]_{cT^n M} \ar[r]^{TT\omega} & TTE \ar[d]^{cE}\\
T^{n+2}M \ar[rr]_{TT\omega} & & TTE
}$$ The upper-right cell commutes since $\omega$ is a sector $n$-form, and the lower-right cell commutes by the naturality of $c$. Hence it suffices to show that the following diagram commutes. $$\xymatrix{
T^{n+1} \ar[rr]^{Ta_{j-1}} \ar[drr]_{a_j} & & T^{n+2} \ar[d]^{cT^n\:=\:c^{n+2}_1}\\
& & T^{n+2}
}$$ In view of \[sec:rms\_def\_sector\_form\], \[exa:deg\_obj\_it\_tgt\], \[exa:symm\_deg\_it\_tang\], this diagram is obtained by applying the strict monoidal functor $T^{(-)}:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]$ to the following diagram in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ $$\xymatrix{
n+1 \ar@{<-}[rr]^{1+\alpha_{j-1}} \ar@{<-}[drr]_{\alpha_j} & & n+2 \ar@{<-}[d]^{\sigma_1}\\
& & n+2
}$$ which commutes, as one readily verifies directly, using the fact that $1+\alpha_{j-1}$ is given by $1 \mapsto 1$ and $1 + x \mapsto 1 + \alpha_{j-1}(x)$ for all $x \in n+1$.
\[thm:sym\_cosimpl\_cmon\_sector\_forms\] There is a symmetric cosimplicial commutative monoid $$\Psi(M)\;:\;{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{cmon}}}\xspace},\;\;\;\;n \mapsto \Psi_n(M)$$ where $\Psi_n(M)$ is the commutative monoid of sector $n$-forms on $M$. The codegeneracies and symmetries carried by $\Psi(M)$ are those obtained in \[thm:cod\_symm\_sector\_forms\], and the fundamental cofaces of $\Psi(M)$ are the maps $$\delta^n_1\;:\;\Psi_n(M) \rightarrow \Psi_{n+1}(M)\;\;\;\;\;\;(n \in {\ensuremath{\mathbb N}\xspace})$$ that send a sector $n$-form $\omega$ to its fundamental derivative $\delta_1(\omega) = T\omega {\ensuremath{\mathop{;}}}\hat{p}$ (\[thm:fund\_deriv\_sector\_form\]).
By \[thm:fund\_deriv\_sector\_form\], the maps $\delta^n_1$ are well-defined, and they are homomorphisms of commutative monoids since $T$ is additive (\[prop:cmons\_in\_tngt\_cat\]) and $\hat{p}:TE \rightarrow E$ is a homomorphism of commutative monoids (\[prop:diffObjs\]). By \[thm:symm\_cod\_struct\_on\_sector\_forms\] we have already defined a symmetric codegenerative object $\Psi(M):{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\textnormal{\textbf{cmon}}}\xspace}$, so by \[thm:main\_fund\_coface\_lemma\] it suffices to verify the equations and , which govern the interaction of the fundamental coface maps $\delta^n_1$ with the codegeneracies and symmetries.
The first equation $\delta_1 {\ensuremath{\mathop{;}}}\varepsilon_1 = 1$ in holds, since for each $\omega \in \Psi_n(M)$ we compute that $$\varepsilon^n_1(\delta^n_1(\omega)) = \ell^n_1 M {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = a^n_1 M {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = \omega {\ensuremath{\mathop{;}}}\lambda {\ensuremath{\mathop{;}}}\hat{p} = \omega$$ using the fact that $\omega$ is a sector form as well as the equations $a^n_1 = \ell^n_1$, $\lambda {\ensuremath{\mathop{;}}}\hat{p} = 1_E$.
The second equation $\delta_1 {\ensuremath{\mathop{;}}}\varepsilon_{j+1} = \varepsilon_j {\ensuremath{\mathop{;}}}\delta_1$ in holds, since for each $\omega \in \Psi_{n+1}(M)$ we compute that $$\delta^n_1(\varepsilon^n_j(\omega)) = T\ell^n_j M {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = \ell^{n+1}_{j+1}M {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = \varepsilon^{n+1}_{j+1}(\delta^{n+1}_1(\omega))$$ using the fact that $T\ell^n_j = T^j\ell T^{n-j} = \ell^{n+1}_{j+1}$.
The first equation $\delta_1 {\ensuremath{\mathop{;}}}\delta_1 {\ensuremath{\mathop{;}}}\sigma_1 = \delta_1 {\ensuremath{\mathop{;}}}\delta_1$ in holds, since for each $\omega \in \Psi_n(M)$ we compute that $$\begin{array}{lllllll}
\sigma^{n+2}_1(\delta^{n+1}_1(\delta^n_1(\omega))) & = & c_1^{n+2} M {\ensuremath{\mathop{;}}}TT\omega {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} & = & cT^n M {\ensuremath{\mathop{;}}}TT\omega {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} & & \\
& = & TT\omega {\ensuremath{\mathop{;}}}cE {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} & = & TT\omega {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} & = & \delta^{n+1}_1(\delta^n_1(\omega))
\end{array}$$ using the naturality of $c$ and the fact that $cE {\ensuremath{\mathop{;}}}T\hat{p} {\ensuremath{\mathop{;}}}\hat{p} = T\hat{p} {\ensuremath{\mathop{;}}}\hat{p}$ (\[prop:diffObjs\]).
The second equation $\delta_1 {\ensuremath{\mathop{;}}}\sigma_{i+1} = \sigma_i {\ensuremath{\mathop{;}}}\delta_1$ in holds, since for each $\omega \in \Psi_n(M)$ we compute that $$\delta^n_1(\sigma^n_i(\omega)) = Tc^n_i M {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = c^{n+1}_{i+1} M {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p} = \sigma^{n+1}_{i+1}(\delta^n_1(\omega))$$ using the fact that $Tc^n_i = T^ic T^{n-i-1} = c^{n+1}_{i+1}$.
\[rmk:der\_posi\] By \[thm:main\_fund\_coface\_lemma\], the cofaces carried by the symmetric cosimplicial set of sector forms $\Psi(M)$ are the maps $$\delta^n_i = \delta^n_1 {\ensuremath{\mathop{;}}}\sigma^{n+1}_{(i)}\;\;:\;\;\Psi_n(M) \rightarrow \Psi_{n+1}(M)\;\;\;\;(n \in {\ensuremath{\mathbb N}\xspace}, 1 \leqslant i \leqslant n+1)$$ that send a sector $n$-form to the sector $(n+1)$-form $$\delta_i(\omega) = \left(T^{n+1}M \xrightarrow{c_{(i)}} T^{n+1}M \xrightarrow{T\omega} TE \xrightarrow{\hat{p}} E\right)$$ which we call the **derivative of $\omega$ in position $i$**.
\[thm:symm\_cosimpl\_ab\_grp\_sector\_forms\] If $E$ is a subtractive differential object in ${\ensuremath{\mathscr X}\xspace}$, then there is a symmetric cosimplicial abelian group $$\Psi(M)\;:\;{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{ab}}}\xspace},\;\;\;\;n \mapsto \Psi_n(M)$$ where $\Psi_n(M)$ is the abelian group of sector $n$-forms on $M$ with values in $E$.
This follows from the preceding theorem and \[thm:sector\_nforms\_submonoid\].
\[thm:func\_sector\_forms\] Let $E$ be a differential object in a Cartesian tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$.
1. There is a functor $$\Psi\;:\;{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace},{\ensuremath{\textnormal{\textbf{cmon}}}\xspace}],\;\;\;\;\;\;M \mapsto \Psi(M)$$ that sends each object $M$ of ${\ensuremath{\mathscr X}\xspace}$ to the symmetric cosimplicial commutative monoid $\Psi(M)$ of sector forms on $M$ with values in $E$.
2. If $E$ is a subtractive differential object, then the functor $\Psi$ lifts to a functor valued in the category $[{\ensuremath{\textnormal{\textbf{finCard}}}\xspace},{\ensuremath{\textnormal{\textbf{ab}}}\xspace}]$ of symmetric cosimplicial abelian groups.
Recall from \[exa:cod\_sets\_ind\_tngt\_str\] that we have a functor ${\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\textnormal{\textbf{cmon}}}\xspace}]$ that sends each morphism $f:M \rightarrow N$ in ${\ensuremath{\mathscr X}\xspace}$ to the natural transformation ${\ensuremath{\mathscr X}\xspace}(T^{(-)}f,E):{\ensuremath{\mathscr X}\xspace}(T^{(-)}N,E) \Rightarrow {\ensuremath{\mathscr X}\xspace}(T^{(-)}M,E)$ whose components ${\ensuremath{\mathscr X}\xspace}(T^n f,E):{\ensuremath{\mathscr X}\xspace}(T^n N,E) \rightarrow {\ensuremath{\mathscr X}\xspace}(T^n M,E)$ are given by precomposition with $T^nf:T^nM \rightarrow T^nN$. It follows immediately from the naturality of the transformations $a^n_j:T^n \Rightarrow T^{n+1}$ that ${\ensuremath{\mathscr X}\xspace}(T^n f,E)$ restricts to yield a homomorphism $\Psi_n(f):\Psi_n(N) \rightarrow \Psi_n(M)$ between the submonoids consisting of sector $n$-forms. We claim that the homomorphisms $\Psi_n(f)$ constitute a natural transformation $\Psi(f):\Psi(N) \Rightarrow \Psi(M)$. It suffices to verify the naturality condition on the generators $\varepsilon^n_i,\sigma^n_i,\delta^n_1$ of ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$ (\[thm:fund\_cof\_pres\_fincard\]). But for the generators $\varepsilon^n_i$, $\sigma^n_i$ this naturality condition follows from the naturality of ${\ensuremath{\mathscr X}\xspace}(T^{(-)}f,E)$, so it suffices to show that $$\xymatrix{
\Psi_n(N) \ar[d]_{\delta^n_1} \ar[r]^{\Psi_n(f)} & \Psi_n(M) \ar[d]^{\delta^n_1}\\
\Psi_{n+1}(N) \ar[r]_{\Psi_{n+1}(f)} & \Psi_{n+1}(M)
}$$ commutes, but this follows immediately from the definitions.
Complexes of forms and the exterior derivative {#sec:complexes}
==============================================
Given an (augmented) cosimplicial abelian group $C:{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{ab}}}\xspace}$, it is well-known [@weibel Definition 8.2.1] that the underlying graded abelian group $(C_n)_{n \in {\ensuremath{\mathbb N}\xspace}}$ carries the structure of a (non-negatively graded) cochain complex $C_\bullet$ when we define the differential $\partial_n:C_n \rightarrow C_{n+1}$ by $$\partial_n(c) = \sum_{i = 1}^{n+1} (-1)^{i-1} \delta^n_i(c)\;\;\;\;\;\;\;\;(c \in C_n).$$ In particular, we therefore have that $\partial_n {\ensuremath{\mathop{;}}}\partial_{n+1} = 0$. We call $C_\bullet$ the **cochain complex associated to** $C$.
In the present section, we show that when $C$ is a *symmetric* cosimplicial abelian group one also obtains a subcomplex $C_\bullet^{\ensuremath{\textnormal{alt}}}\hookrightarrow C_\bullet$ consisting of the *alternating elements* of $C$. Applied to the symmetric cosimpicial object of sector forms (\[thm:symm\_cosimpl\_ab\_grp\_sector\_forms\]), we obtain complexes of sector forms ($C_\bullet$) and singular forms ($C_\bullet^{\ensuremath{\textnormal{alt}}}$) in tangent categories.
Sector forms and the exterior derivative
----------------------------------------
Let $E$ be a subtractive differential object in a Cartesian tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$, and let $M$ be an object of ${\ensuremath{\mathscr X}\xspace}$.
\[def:compl\_sector\_forms\]
1. The **complex of sector forms** on $M$ is defined as the cochain complex $\Psi_\bullet(M)$ associated to the cosimplicial abelian group $\Psi(M)$ of sector forms on $M$ with values in $E$ (\[thm:symm\_cosimpl\_ab\_grp\_sector\_forms\]).
2. Given a sector $n$-form $\omega:T^n M \rightarrow E$ on $M$, the **exterior derivative** of $\omega$ is defined as the sector $(n+1)$-form $$\partial\omega = \partial_n(\omega) = \sum_{i = 1}^{n+1} (-1)^{i-1} \delta^n_i(\omega)$$ where $\partial_n:\Psi_n(M) \rightarrow \Psi_{n+1}(M)$ is the differential carried by the complex $\Psi_\bullet(M)$, recalling that the sector $(n+1)$-form $\delta^n_i(\omega) = c^{n+1}_{(i)} {\ensuremath{\mathop{;}}}T\omega {\ensuremath{\mathop{;}}}\hat{p}$ is the derivative of $\omega$ in position $i$ (\[rmk:der\_posi\]).
The following theorem is now immediate, but it would be difficult to prove if we had just defined the exterior derivative directly without first proving Theorem \[thm:sym\_cosimpl\_cmon\_sector\_forms\]:
Let $\omega$ be a sector $n$-form on $M$. Then $\partial_{n+1}(\partial_n(\omega)) = 0$.
\[rmk:func\_alt\_cof\_map\_complex\] It is well known that the the assignment $C \mapsto C_\bullet$ extends to a functor $(-)_\bullet$ from the category of cosimplicial abelian groups to the category ${\ensuremath{\textnormal{\textbf{cochain}}}\xspace}_+$ of non-negatively graded cochain complexes[^10]. Hence by \[thm:func\_sector\_forms\] we can form the composite functor $$\Psi_\bullet = \left({\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\xrightarrow{\Psi} [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace},{\ensuremath{\textnormal{\textbf{ab}}}\xspace}] \rightarrow [{\ensuremath{\textnormal{\textbf{finOrd}}}\xspace},{\ensuremath{\textnormal{\textbf{ab}}}\xspace}] \xrightarrow{(-)_\bullet} {\ensuremath{\textnormal{\textbf{cochain}}}\xspace}_+\right)$$ whose middle factor is the evident forgetful functor. This functor $\Psi_\bullet$ sends each object $M$ of ${\ensuremath{\mathscr X}\xspace}$ to the complex of sector forms on $M$.
The complex of alternating elements
-----------------------------------
Given a symmetric cosimplicial abelian group $C$, we now define a certain subcomplex $C^{\ensuremath{\textnormal{alt}}}_\bullet$ of $C_\bullet$.
\[def:alt\_elt\] Let $n \in {\ensuremath{\mathbb N}\xspace}$.
1. We say that an element $c$ of $C_n$ is an **alternating element** of $C$ if $\sigma^n_i(c) = -c$ for all $i \in \{1,...,n-1\}$, recalling that $\sigma^n_i:C_n \rightarrow C_n$ is the symmetry map carried by $C$.
2. We denote by $C^{\ensuremath{\textnormal{alt}}}_n \subseteq C_n$ the subset consisting of all alternating elements.
Given a symmetric cosimplicial abelian group $C$, the alternating elements of $C$ constitute a subcomplex $C^{\ensuremath{\textnormal{alt}}}_\bullet$ of the cochain complex $C_\bullet$ associated to $C$.
$C_n^{\ensuremath{\textnormal{alt}}}\hookrightarrow C_n$ is an intersection of equalizers in ${\ensuremath{\textnormal{\textbf{ab}}}\xspace}$ and hence is a subgroup inclusion. Letting $c \in C^{\ensuremath{\textnormal{alt}}}_n \subseteq C_n$, it suffices to show that the associated element $\partial(c) = \partial_n(c)$ of $C_{n+1}$ is alternating. Letting $i \in \{1,...,n\}$ we must show that $\sigma^{n+1}_i(\partial(c)) = -\partial(c)$. Since $\sigma^{n+1}_i$ is a homomorphism of abelian groups we compute that $$\sigma^{n+1}_i(\partial(c)) = \sum_{j = 1}^{n+1}(-1)^{j-1}\sigma^{n+1}_i(\delta^n_j(c))\;.$$ Using the coface-symmetry relations and the fact that $\sigma^{n+1}_i$ is self-inverse, we compute that $$\sigma^{n+1}_i(\delta^n_j(c)) = \begin{cases}
\delta^n_j(\sigma^n_{i-1}(c)) = \delta^n_j(-c) = -\delta^n_j(c) & (j < i)\\
\delta^n_{i+1}(c) & (j = i)\\
\delta^n_i(c) & (j = i+1)\\
\delta^n_j(\sigma^n_i(c)) = \delta^n_j(-c) = -\delta^n_j(c) & (j > i+1)\\
\end{cases}$$ since $c$ is alternating. Hence, recalling that $\partial(c) = \sum_{j = 1}^{n+1}t_j$ where $t_j = (-1)^{j-1}\delta^n_j(c)$, we compute that $$\sigma^{n+1}_i(\partial(c)) = \left(\sum_{j < i}-t_j\right) - t_{i+1} - t_i + \left(\sum_{j > i + 1}-t_j\right) = -\partial(c)$$ since $i$ and $i+1$ are of opposite parity.
\[def:compl\_alt\_elts\] Given a symmetric cosimplicial abelian group $C$, we call the subcomplex $C^{\ensuremath{\textnormal{alt}}}_\bullet$ of $C_\bullet$ the **complex of alternating elements of $C$**.
\[thm:func\_compl\_alt\_elts\] There is a functor $$(-)_\bullet^{\ensuremath{\textnormal{alt}}}\;:\;[{\ensuremath{\textnormal{\textbf{finCard}}}\xspace},{\ensuremath{\textnormal{\textbf{ab}}}\xspace}] \rightarrow {\ensuremath{\textnormal{\textbf{cochain}}}\xspace}_+\;\;\;\;\;\;\;\;C \mapsto C_\bullet^{\ensuremath{\textnormal{alt}}}$$ from the category of symmetric cosimplicial abelian groups to the category of (non-negatively graded) cochain complexes, sending a symmetric cosimplicial abelian group $C$ to the complex of alternating elements $C^{\ensuremath{\textnormal{alt}}}_\bullet$ of $C$.
Given a morphism of symmetric cosimplicial abelian groups $f:C \rightarrow D$, we claim that the associated morphism of chain complexes $f_\bullet:C_\bullet \rightarrow D_\bullet$ restricts to a morphism $f^{\ensuremath{\textnormal{alt}}}_\bullet:C^{\ensuremath{\textnormal{alt}}}_\bullet \rightarrow D^{\ensuremath{\textnormal{alt}}}_\bullet$ between the subcomplexes of alternating elements. Indeed, given $c \in C^{\ensuremath{\textnormal{alt}}}_n \subseteq C_n$, the associated element $f_n(c)$ of $D_n$ is alternating, since for each $i \in \{1,...,n-1\}$, we compute that $\sigma^n_i(f_n(c)) = f_n(\sigma^n_i(c)) = f_n(-c) = -f_n(c)$ since $f$ is natural and $c$ is alternating. The result now follows from \[rmk:func\_alt\_cof\_map\_complex\].
The complex of singular forms
-----------------------------
Again let us fix a subtractive differential object $E$ in a Cartesian tangent category $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$. Letting $M$ be an object of ${\ensuremath{\mathscr X}\xspace}$, recall that $\Psi(M)$ denotes the symmetric cosimplicial abelian group of sector forms on $M$ (\[thm:symm\_cosimpl\_ab\_grp\_sector\_forms\]).
\[def:sing\_nform\_compl\]
1. We call a sector $n$-form $\omega$ on $M$ a **singular $n$-form** if $\omega$ is an alternating element of $\Psi(M)$. Equivalently, a morphism $\omega:T^nM \rightarrow E$ is a singular $n$-form iff $\omega$ is a sector $n$-form and $c^n_iM {\ensuremath{\mathop{;}}}\omega = -\omega$ for every $i \in \{1,...,n-1\}$, recalling that $c^n_i M = T^{i-1}cT^{n-i-1}M:T^n M \rightarrow T^n M$ is the symmetry carried by $T^{(-)}M$ (\[exa:cod\_sets\_ind\_tngt\_str\]).
2. We denote the complex of alternating elements of $\Psi(M)$ (\[def:compl\_alt\_elts\]) by $$\Omega_\bullet(M) := (\Psi(M))_\bullet^{\ensuremath{\textnormal{alt}}}$$ and call it the **complex of singular forms** on $M$ with values in $E$.
Since $\Omega_\bullet(M)$ is a subcomplex of $\Psi_\bullet(M)$, the following theorem is now immediate:
The exterior derivative $\partial\omega$ of a singular $n$-form on $M$ is a singular $(n+1)$-form.
\[prop:complexOfSectorForms\] There is a functor $$\Omega_\bullet\;:\;{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\textnormal{\textbf{cochain}}}\xspace}_+\;\;\;\;\;\;\;\;M \mapsto \Omega_\bullet(M)$$ sending each object $M$ of ${\ensuremath{\mathscr X}\xspace}$ to the complex of singular forms on $M$ with values in $E$.
This follows from \[thm:func\_sector\_forms\] and \[thm:func\_compl\_alt\_elts\].
Relationship to de Rham in synthetic and classical differential geometry {#sec:relFormsClSDG}
========================================================================
*Synthetic differential geometry* (SDG) is an approach to differential geometry in terms of infinitesimals that was initiated in a lecture of Lawvere in 1967 and developed by several authors, starting with work of Wraith and of Kock [@Kock:SimplAxDiff] in the 1970s. The reader is referred to the books [@kock] and [@lavendhomme] for a comprehensive introduction to SDG. An approach to differential forms in SDG was developed in [@Kock:DiffFormsSDG; @KRV; @MoeRey:CohThSDG] (see [@kock], [@lavendhomme]), and in the present section we compare this work to the development of differential forms given above, recovering the classical de Rham complex of a smooth manifold as a corollary (\[thm:classical\_derham\]). This comparison involves specializing our treatment of sector forms to the case in which the tangent structure is *representable* (\[def:rep\_tan\]), an exercise that is illuminating in its own right.
In the most prevalent formulation of SDG, one begins with a topos ${\ensuremath{\mathscr E}\xspace}$ and a commutative ring object $R$ in ${\ensuremath{\mathscr E}\xspace}$, and then one defines $D$ to be the subobject of $R$ described by the equation $x^2 = 0$, so that $D$ is the part of $R$ that consists of square-zero ‘infinitesimal elements’. Writing $[M,N]$ for the internal hom between objects $M$ and $N$ of ${\ensuremath{\mathscr E}\xspace}$, one construes the object $[D,M]$ as the space $TM$ of all tangent vectors on $M$. One postulates that $R$ should satisfy the *Kock-Lawvere* axiom (see section \[sec:compn\_forms\_sdg\]), and sometimes further axioms, on the basis of which one can develop much differential geometry in ${\ensuremath{\mathscr E}\xspace}$. One can define specific toposes ${\ensuremath{\mathscr E}\xspace}$ into which the category ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ of smooth manifolds embeds, via an embedding ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}\hookrightarrow {\ensuremath{\mathscr E}\xspace}$ that sends the real numbers ${\ensuremath{\mathbb R}\xspace}$ to $R$; see [@reyes] and §\[sec:classical\_derham\] below.
The approach of defining $D$ as the square-zero part of $R$ was put forward in Kock’s 1977 paper [@Kock:SimplAxDiff], wherein it is indicated that Lawvere’s 1967 lecture did not define $D$ in this way but rather postulated that an object $D$ of infinitesimals should exist and that $[D,M]$ for each object $M$ should have properties expected of the tangent bundle of $M$. Evidently tangent categories provide an axiomatics of such properties, and indeed Rosický’s 1984 paper [@rosicky] considers in particular those tangent categories $({\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathbb T}\xspace})$ for which there is an exponentiable object $D$ with $T \cong [D,-]:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$. This leads to an axiomatics for structure and properties that should be possessed by an object of infinitesimals $D$ [@rosicky §4], [@sman3 5.6]. The following definition was given in [@sman3] and is a variation on a similar definition given in [@rosicky]. Herein, we say that an endofunctor $F$ on a category ${\ensuremath{\mathscr X}\xspace}$ with finite products is **representable** if it is isomorphic to an endofunctor of the form $[X,-]:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ for some exponentiable object $X$ of ${\ensuremath{\mathscr X}\xspace}$, and we then say that $F$ is represented by $X$.
\[def:rep\_tan\] A category ${\ensuremath{\mathscr X}\xspace}$ carries **representable tangent structure** if ${\ensuremath{\mathscr X}\xspace}$ has finite products and carries a tangent structure ${\ensuremath{\mathbb T}\xspace}$ in which the endofunctors $T^n$ and $T_n$ $(n \in {\ensuremath{\mathbb N}\xspace})$ are representable.
It is proved in [@sman3 Prop. 5.7] that a category ${\ensuremath{\mathscr X}\xspace}$ with finite products carries representable tangent structure if and only if there is an exponentiable object $D$ of ${\ensuremath{\mathscr X}\xspace}$ that carries the structure of an **infinitesimal object** in the sense of [@sman3 Def. 5.6]. In this case the tangent endofunctor $T$ is represented by $D$, and its iterates $T^n$ are represented by the $n$-th powers $D^n$ of $D$. In particular, representable tangent structure is necessarily Cartesian in the sense of \[sec:cmons\_in\_tngnt\_cat\].
Let us now fix a category ${\ensuremath{\mathscr X}\xspace}$ with representable tangent structure ${\ensuremath{\mathbb T}\xspace}$, represented by an infinitesimal object $D$ in ${\ensuremath{\mathscr X}\xspace}$. Throughout, we shall assume without loss of generality that $T = [D,-]$ on the nose. We shall not assume however that $T^n = [D^n,-]$ for $n > 1$, but rather we now define specific isomorphisms $T^n \cong [D^n,-]$ for use in the sequel.
\[def:isos\_rep\_it\_tngnt\] Let us define isomorphisms $$\psi_n:[D^n,-] \overset{\sim}{\longrightarrow} T^n\;\;\;\;\;\;(n \in {\ensuremath{\mathbb N}\xspace})$$ by recursion on $n$, as follows. Firstly, $\psi_0$ is defined as the canonical isomorphism from $[1,-]$ to the identity functor on ${\ensuremath{\mathscr X}\xspace}$. Next, the components of $\psi_{n+1}$ are defined as the composites $$\psi_{n+1}M := \left([D^{n+1},M] \xrightarrow{\phi_{n+1}M} [D,[D^n,M]] \xrightarrow{[D,\psi_nM]} [D,T^nM] = T^{n+1}M\right)\;\;\;(M \in {\ensuremath{\mathop{\textnormal{ob}}}}{\ensuremath{\mathscr X}\xspace})$$ where $\phi_{n+1}M$ is the isomorphism whose transpose $[D^{n+1},M] \times D \rightarrow [D^n,M]$ is in turn defined as the transpose of the evaluation morphism $$[D^{n+1},M] \times D \times D^n = [D^{n+1},M] \times D^{n+1} \longrightarrow M.$$ Note that $\psi_1$ is therefore the identity transformation on $[D,-] = T$.
Simple type theory and lambda calculus
--------------------------------------
In synthetic differential geometry one often makes use of the *internal language* of a given topos ${\ensuremath{\mathscr E}\xspace}$ in order to define morphisms in ${\ensuremath{\mathscr E}\xspace}$ by means of ‘elementwise’ formulae, to show that diagrams commute just by chasing elements, and so on; see, e.g., [@kock Part II]. After all, the internal language of ${\ensuremath{\mathscr E}\xspace}$ is a restricted form of set theory, or rather *higher-order intuitionistic type theory* [@LS].
Even though our given tangent category ${\ensuremath{\mathscr X}\xspace}$ is not assumed to be a topos, it still possesses an internal language, albeit a rather restricted one, namely the **simple type theory** of ${\ensuremath{\mathscr X}\xspace}$ [@Jac Chapter 2], considered as a category with finite products. We now informally review some basic elements of this language and one of its extensions, the *simply typed lambda calculus*; readers who are familiar with the latter may safely skip this section.
Given any morphism $f:X_1 \times ... \times X_n \rightarrow Y$ in ${\ensuremath{\mathscr X}\xspace}$, we can form a *typing judgment* or *term-in-context* $$x_1:X_1,...,x_n:X_n \;\;\vdash\;\; f(x_1,...,x_n)\;:\;Y$$ in which each expression $x_i:X_i$ indicates that $x_i$ is a formal variable of type $X_i$. The typing judgment asserts that the expression $f(x_1,...,x_n)$ is a *term* of type $Y$. The part of the typing judgement to the left of the turnstile $\vdash$ is called the *context*. The simple type theory of ${\ensuremath{\mathscr X}\xspace}$ includes various *term formation rules* which allow us to construct new terms-in-context from others [@Jac 2.1]. For example given terms-in-context $x:X \vdash f(x):Y$ and $y:Y \vdash g(y):Z$ associated to morphisms $f:X \rightarrow Y$ and $g:Y \rightarrow Z$ in ${\ensuremath{\mathscr X}\xspace}$, we can form a term-in-context $x:X \vdash g(f(x)):Z$. Every term-in-context denotes an associated morphism in ${\ensuremath{\mathscr X}\xspace}$, and in particular, the latter term-in-context denotes the composite morphism $f {\ensuremath{\mathop{;}}}g:X \rightarrow Z$. The simple type theory of ${\ensuremath{\mathscr X}\xspace}$ carries also a calculus of *equations* $$x_1:X_1,...,x_n:X_n \;\;\vdash\;\; t_1 = t_2\;\;:\;\;Y$$ where $t_1, t_2$ are terms in the same context $\Gamma$, namely $x_1:X_1,...,x_n:X_n$, and we say that such an equation *holds* in ${\ensuremath{\mathscr X}\xspace}$ if the morphisms in ${\ensuremath{\mathscr X}\xspace}$ denoted by $\Gamma \vdash t_1:Y$ and $\Gamma \vdash t_2:Y$ are equal. We shall often omit typing indications “$y:Y$” within terms-in-context and equations when the intended typing is clear.
Having assumed that ${\ensuremath{\mathscr X}\xspace}$ has an infinitesimal object $D$, which is exponentiable, we would also like to employ an internal language in reasoning about exponential transposition of morphisms. In the case where ${\ensuremath{\mathscr X}\xspace}$ is Cartesian closed, we can employ the **simply typed lambda calculus** of ${\ensuremath{\mathscr X}\xspace}$ [@Jac 2.3][@LS], which extends the simply type theory of ${\ensuremath{\mathscr X}\xspace}$ by adding term-formation rules corresponding to exponential transposition, together with corresponding rules governing equality. For example given a morphism $f:X \times Y \rightarrow Z$ in ${\ensuremath{\mathscr X}\xspace}$, the associated transpose $X \rightarrow [Y,Z]$ is denoted by the term-in-context $$x:X \;\;\vdash\;\; {\ensuremath{\underline{\lambda}}}y:Y.f(x,y)\;\;:\;\;[Y,Z]$$ where the construct “${\ensuremath{\underline{\lambda}}}y:Y.$” serves to bind the variable $y$ within the scope of the expression ${\ensuremath{\underline{\lambda}}}y:Y.f(x,y)$. We will often write just ${\ensuremath{\underline{\lambda}}}y.f(x,y)$.
Although the given tangent category ${\ensuremath{\mathscr X}\xspace}$ is not assumed Cartesian closed, we can clearly[^11] still employ simply typed lambda calculus and its interpretation in ${\ensuremath{\mathscr X}\xspace}$ as long as the instances of exponential transposition and evaluation employed are those permitted by the exponentiable objects $D^n$. Given objects $X,Y,Z$ of ${\ensuremath{\mathscr X}\xspace}$ with $X,Y$ exponentiable, we shall write $f:[X,Y],\; g:[Y,Z] \vdash f {\ensuremath{\mathop{;}}}g$ to denote the composition morphism $[X,Y] \times [Y,Z] \rightarrow [X,Z]$ in ${\ensuremath{\mathscr X}\xspace}$.
As a first application of this type-theoretic notation, we record the following:
\[thm:isos\_rep\_it\_tngnt\] For each $n \geqslant 1$ and each object $M$ of ${\ensuremath{\mathscr X}\xspace}$, the isomorphism $\psi_n:[D^n,M] \rightarrow T^nM = [D,[D,...]]$ defined in \[def:isos\_rep\_it\_tngnt\] is characterized by the following equation: $$\tau:[D^n,M] \;\;\vdash\;\; \psi_n(\tau) = {\ensuremath{\underline{\lambda}}}d_1.{\ensuremath{\underline{\lambda}}}d_2.\:...\:{\ensuremath{\underline{\lambda}}}d_n.\tau(d_1,d_2,...,d_n)\;\;:\;\;T^nM$$
Symmetric degenerative structure induced by $D$
-----------------------------------------------
Again fixing a category ${\ensuremath{\mathscr X}\xspace}$ with representable tangent structure, the representing infinitesimal object $D$ carries a structural morphism $\odot:D \times D \rightarrow D$ that makes $D$ a commutative semigroup in ${\ensuremath{\mathscr X}\xspace}$ [@sman3 Def. 5.6]. Hence $D$ carries the structure of a symmetric semigroup $(D,\odot,s)$ in ${\ensuremath{\mathscr X}\xspace}$ (\[def:symm\_semi\_mon\], \[rem:comm\_sgrp\_yields\_symm\_sgrp\]), where $s:D^2 \rightarrow D^2$ is the symmetry in ${\ensuremath{\mathscr X}\xspace}$. Within the simple type theory of ${\ensuremath{\mathscr X}\xspace}$, we shall denote the multiplication $\odot$ by $d_1:D, d_2:D \vdash d_1d_2\::\:D$.
Recalling that the vertical lift $\ell:T \rightarrow T^2$ and the canonical flip $c:T^2 \rightarrow T^2$ together equip $T$ with the structure of a symmetric semigroup $(T,\ell,c)$ in $[{\ensuremath{\mathscr X}\xspace},{\ensuremath{\mathscr X}\xspace}]^{{\mbox{\textnormal{\scriptsize op}}}}$ (\[sec:tng\_func\_symm\_cosgrp\]), it is clear from the discussion in [@sman3 §5.2] that this symmetric semigroup structure is induced by the semigroup structure $(D,\odot,s)$ on $D$ in the sense that the following diagrams commute. $$\label{eq:sgrp_str_on_t_ind_by_that_on_d}
\xymatrix{
[D,-] \ar@{=}[d] \ar[r]^{[\odot,-]} & [D^2,-] \ar[d]^{\psi_2}_\wr & & [D^2,-] \ar[d]_{\psi_2}^\wr \ar[r]^{[s,-]} & [D^2,-] \ar[d]^{\psi_2}_\wr\\
T \ar[r]_\ell & T^2 & & T^2 \ar[r]_{c} & T^2
}$$
Whereas ${\ensuremath{\mathscr X}\xspace}$ is not a strict monoidal category, we can construe $(D,\odot,s)$ as a symmetric semigroup in a Cartesian strict monoidal category ${\ensuremath{\mathscr X}\xspace}_D$ that is defined as follows. Define ${\ensuremath{\mathop{\textnormal{ob}}}}{\ensuremath{\mathscr X}\xspace}_D = {\ensuremath{\mathbb N}\xspace}$ and ${\ensuremath{\mathscr X}\xspace}_D(n,m) = {\ensuremath{\mathscr X}\xspace}(D^n,D^m)$, with composition as in ${\ensuremath{\mathscr X}\xspace}$. Informally, we will write $D^n$ for the object $n$ of ${\ensuremath{\mathscr X}\xspace}_D$, noting that ${\ensuremath{\mathscr X}\xspace}_D$ is equivalent to the full subcategory of ${\ensuremath{\mathscr X}\xspace}$ on the objects $D^n$. One encounters no complication in defining a Cartesian strict monoidal structure on ${\ensuremath{\mathscr X}\xspace}_D$, and it is for this reason that we work with ${\ensuremath{\mathscr X}\xspace}_D$ rather than the latter full subcategory of ${\ensuremath{\mathscr X}\xspace}$.
By \[thm:univ\_symm\_mon\_sgrp\], the symmetric semigroup $(D,\odot,s)$ in ${\ensuremath{\mathscr X}\xspace}_D$ induces a strict monoidal functor $$D^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr X}\xspace}_D,\;\;\;\;\;\;n \mapsto D^n,$$ and we will also write $D^\sharp$ to denote the functor $D^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr X}\xspace}$ obtained by composing with the canonical fully faithful functor ${\ensuremath{\mathscr X}\xspace}_D \rightarrow {\ensuremath{\mathscr X}\xspace}$.
\[thm:dsharp\]
1. The codegeneracies and symmetries carried by the symmetric codegenerative object $D^\sharp$ in ${\ensuremath{\mathscr X}\xspace}$ are the following morphisms, respectively: $$\odot^n_i := D^{i-1}\times \odot \times D^{n-i}:D^{n+1} \rightarrow D^n,\;\;\;\;\;\;s^n_i := D^{i-1} \times s \times D^{n-i-1}:D^n \rightarrow D^n.$$
2. Writing $D^f:D^m \rightarrow D^n$ for the morphism in ${\ensuremath{\mathscr X}\xspace}$ induced by a mapping $f:n \rightarrow m$ between finite cardinals $n,m$, the functor $D^\sharp:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s \rightarrow {\ensuremath{\mathscr X}\xspace}$ sends each permutation $\xi:n \xrightarrow{\sim} n$ to the automorphism $D^{\xi^{-1}}:D^n \rightarrow D^n$ in ${\ensuremath{\mathscr X}\xspace}$.
\(i) is immediate from \[sec:symm\_cod\_obj\_det\_symm\_sgrp\]. For (ii) it suffices to show that the copermutative object ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \rightarrow {\ensuremath{\mathscr X}\xspace}_D$ underlying $D^\sharp$ is equal to the composite $${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \xrightarrow{(-)^{-1}} {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b^{{\mbox{\textnormal{\scriptsize op}}}}\xrightarrow{D^{(-)}} {\ensuremath{\mathscr X}\xspace}_D$$ whose first factor is the identity-on-objects isomorphism given in \[rem:symm\_grp\_actions\]. But these two functors ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_b \rightarrow {\ensuremath{\mathscr X}\xspace}_D$ are both strict monoidal, and they both send $1$ to $D$ and send the symmetry $\sigma:2 \rightarrow 2$ to the symmetry $s:D^2 \rightarrow D^2$ in ${\ensuremath{\mathscr X}\xspace}_D$, so by \[thm:univ\_symm\_mon\_sgrp\](iii) they are equal.
\[thm:symm\_deg\_ind\_by\_d\] For each object $M$ of ${\ensuremath{\mathscr X}\xspace}$, the isomorphisms $\psi_n:[D^n,M] \rightarrow T^nM$ constitute an isomorphism of symmetric degenerative objects $$[D^\sharp(-),M] \cong T^{(-)}M\;\;:\;\;{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr X}\xspace}\;.$$
Since the codegeneracies $\varepsilon^n_i$ and symmetries $\sigma^n_i$ generate the category ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ (\[thm:fincards\_gens\_rels\]), it suffices to show that the isomorphisms $\psi_n$ commute with the degeneracies and symmetries carried by $[D^\sharp(-),M]$ and $T^{(-)}M$. By \[thm:dsharp\], the degeneracies and symmetries carried by $[D^\sharp(-),M]$ are the morphisms $[\odot^n_i,M]:[D^n,M] \rightarrow [D^{n+1},M]$ and $[s^n_i,M]:[D^n,M] \rightarrow [D^n,M]$, and by \[exa:symm\_deg\_it\_tang\], \[exa:cod\_sets\_ind\_tngt\_str\] those carried by $T^{(-)}M$ are (the components at $M$ of) the transformations $\ell^n_i = T^{i-1}\ell T^{n-i}$ and $c^n_i = T^{i-1}c T^{n-i-1}$. We can now use and \[thm:isos\_rep\_it\_tngnt\] to compute that the following equations hold in ${\ensuremath{\mathscr X}\xspace}$: $$\tau:[D^n,M] \vdash \ell^n_i(\psi_n(\tau)) = {\ensuremath{\underline{\lambda}}}d_1.\:...\:{\ensuremath{\underline{\lambda}}}d_{n+1}.\tau(d_1,...,d_{i-1},d_id_{i+1},d_{i+2},...,d_{n+1}) = \psi_{n+1}(\odot^n_i {\ensuremath{\mathop{;}}}\tau)$$ $$\tau:[D^n,M] \vdash c^n_i(\psi_n(\tau)) = {\ensuremath{\underline{\lambda}}}d_1.{\ensuremath{\underline{\lambda}}}d_2.\:...\:{\ensuremath{\underline{\lambda}}}d_n.\tau(d_1,...,d_{i-1},d_{i+1},d_i,d_{i+2},...,d_n) = \psi_n(s^n_i {\ensuremath{\mathop{;}}}\tau)$$
Some infinitesimal left actions {#sec:infl_actions}
-------------------------------
For each $n \in {\ensuremath{\mathbb N}\xspace}$ and each $j = 1,...,n$ we have a morphism $\alpha^n_j = \sigma^{n+1}_{(j)} {\ensuremath{\mathop{;}}}\varepsilon^n_j:n+1 \rightarrow n$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ (\[sec:rms\_def\_sector\_form\]). Using \[thm:dsharp\] and the definition of $\sigma^{n+1}_{(j)} = (j(j-1)...321)$ (\[def:notn\_perms\]), we compute that the morphism $D^\sharp(\sigma^{n+1}_{(j)}):D^{n+1} \rightarrow D^{n+1}$ carried by the symmetric codegenerative object $D^\sharp$ is characterized by $$\label{eq:dsharp_sigmaparensj}(d_0,d_1,...,d_n):D^{n+1} \;\vdash\; (D^\sharp(\sigma^{n+1}_{(j)}))(d_0,...,d_n) = (d_1,...,d_{j-1},d_0,d_j,d_{j+1},...,d_n)\;.$$ Again applying \[thm:dsharp\] we therefore compute that the morphism $$D^\sharp(\alpha^n_j):D \times D^n = D^{n+1} \rightarrow D^n$$ is characterized as follows: $$(d_0,d_1,...,d_n):D^{n+1} \;\vdash\; (D^\sharp(\alpha^n_j))(d_0,...,d_n) = (d_1,...,d_{j-1},d_0d_j,d_{j+1},...,d_n)\;:\;D^n$$ In effect, $D^\sharp(\alpha^n_j)$ is the left action of $D$ on the $j$-th factor of $D^n$.
Now fixing a differential object $E$ in ${\ensuremath{\mathscr X}\xspace}$, the third axiom in \[def:diff\_objs\] entails that $E$ carries an associative action of $D$, namely the transpose $\bullet:D \times E \rightarrow E$ of the lift morphism $\lambda:E \rightarrow [D,E] = TE$ carried by $E$. As with the multiplication carried by $D$, we will denote this left action on $E$ by juxtaposition within the lambda calculus.
\[def:synth\_sf\] Let $M$ be an object of ${\ensuremath{\mathscr X}\xspace}$, and let $n \in {\ensuremath{\mathbb N}\xspace}$.
1. For each $j = 1,...,n$, let $\bullet^n_j:D \times [D^n,M] \rightarrow [D^n,M]$ denote the transpose of the composite $$[D^n,M] \xrightarrow{[D^\sharp(\alpha^n_j),M]} [D^{n+1},M] \xrightarrow{\phi_{n+1}} [D,[D^n,M]]$$ where $\phi_{n+1}$ is the isomorphism defined in \[def:isos\_rep\_it\_tngnt\]. Writing $\bullet^n_j$ in infix notation in the lambda calculus, this morphism $\bullet^n_j$ is characterized by the following equation: $$d:D, \tau:[D^n,M]\;\vdash\; d \bullet^n_j \tau = {\ensuremath{\underline{\lambda}}}(d_1,...,d_n).\tau(d_1,...,d_{j-1},dd_j,d_{j+1},...,d_n)$$
2. We say that a morphism $\nu:[D^n,M] \rightarrow E$ in ${\ensuremath{\mathscr X}\xspace}$ is a **synthetic sector $n$-form** on $M$ with values in $E$ if $$\label{eq:synth_sf_multilin_ax}d:D, \tau:[D^n,M] \;\vdash\; \nu(d \bullet^n_j \tau) = d\nu(\tau)\;\;:\;\;E$$ holds in ${\ensuremath{\mathscr X}\xspace}$ for each $j = 1,...,n$, i.e. if the following diagram commutes: $$\label{eq:diag_synth_sf}
\xymatrix{
D \times [D^n,M] \ar[d]_{\bullet^n_j} \ar[r]^(.55){D \times \nu} & D \times E \ar[d]^\bullet\\
[D^n,M] \ar[r]_\nu & E
}$$
\[thm:symm\_cosimpl\_cmon\_syn\_sf\] For each object $M$ of ${\ensuremath{\mathscr X}\xspace}$, there is a symmetric cosimplicial commutative monoid $\Psi^{\ensuremath{\textnormal{syn}}}(M)$ in which $\Psi^{\ensuremath{\textnormal{syn}}}_n(M)$ is the set of all synthetic sector $n$-forms on $M$ with values in $E$ $(n \in {\ensuremath{\mathbb N}\xspace})$. Further, there is an isomorphism of symmetric cosimplicial commutative monoids $$\Psi(M) \cong \Psi^{\ensuremath{\textnormal{syn}}}(M)$$ between sector forms on $M$ and synthetic sector forms on $M$. The codegeneracies and symmetries carried by $\Psi^{\ensuremath{\textnormal{syn}}}(M)$ are given by pre-composition with the degeneracies and symmetries carried by $[D^\sharp(-),M]$ (\[thm:symm\_deg\_ind\_by\_d\]).
For each $n \in {\ensuremath{\mathbb N}\xspace}$, the isomorphism $\psi_n:[D^n,M] \rightarrow T^nM$ induces a bijection between morphisms $\nu:[D^n,M] \rightarrow E$ and morphisms $\omega:T^nM \rightarrow E$. Given a pair of morphisms $\nu,\omega$ that correspond under this bijection, so that $\nu = \psi_n {\ensuremath{\mathop{;}}}\omega$, we claim that $\nu$ is a synthetic sector $n$-form if and only if $\omega$ is a sector $n$-form. To prove this, first observe that the following diagram commutes, by the inductive definition of the isomorphisms $\psi_n$ (\[def:isos\_rep\_it\_tngnt\]). $$\xymatrix{
[D^{n+1},M] \ar[dr]_{\psi_{n+1}} \ar[r]^{\phi_{n+1}}_\sim & [D,[D^n,M]] \ar[d]^{T\psi_n} \ar[dr]^{T\nu \:=\: [D,\nu]} & \\
& T^{n+1}M \ar[r]_(.45){T\omega} & {[D,E] = TE}
}$$ Next recall that $\omega$ is a sector $n$-form if and only if $a^n_j {\ensuremath{\mathop{;}}}T\omega = \omega {\ensuremath{\mathop{;}}}\lambda$ for all $j = 1,...,n$ (\[def:sector\_nform\]). Here $a^n_j:T^nM \rightarrow T^{n+1}M$ is obtained by applying the functor $T^{(-)}M:{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\mathscr X}\xspace}$ to the morphism $\alpha^n_j:n+1 \rightarrow n$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s$ (\[sec:rms\_def\_sector\_form\]), so the naturality of the isomorphism $\psi:[D^\sharp(-),M] \xrightarrow{\sim} T^{(-)}M$ (\[thm:symm\_deg\_ind\_by\_d\]) entails that $\psi_n {\ensuremath{\mathop{;}}}a^n_j = [D^\sharp(\alpha^n_j),M] {\ensuremath{\mathop{;}}}\psi_{n+1}:[D^n,M] \rightarrow T^{n+1}M$. Using these facts we readily deduce that $\omega$ is a sector $n$-form if and only if the following diagram commutes for each $j = 1,...,n$. $$\xymatrix{
[D^n,M] \ar[d]_{[D^\sharp(\alpha^n_j),M]} \ar[rr]^\nu & & E \ar[d]^\lambda\\
[D^{n+1},M] \ar[r]_{\phi_{n+1}}^\sim & [D,[D^n,M]] \ar[r]_(.55){[D,\nu]} & [D,E]
}$$ But the two composites in this diagram are the exponential transposes of the two composites in the diagram whose commutativity characterizes synthetic sector $n$-forms.
Hence we have a bijection $\Psi_n(M) \xrightarrow{\sim} \Psi^{\ensuremath{\textnormal{syn}}}_n(M)$ given by $\omega \mapsto \psi_n {\ensuremath{\mathop{;}}}\omega$, and since $\Psi_n(M)$ is a submonoid of ${\ensuremath{\mathscr X}\xspace}(T^nM,E)$ it follows that $\Psi^{\ensuremath{\textnormal{syn}}}_n(M)$ is a submonoid of ${\ensuremath{\mathscr X}\xspace}([D^n,M],E)$ and the given bijection is an isomorphism of commutative monoids. Hence in view of \[thm:sym\_cosimpl\_cmon\_sector\_forms\] there is a unique functor $\Psi^{\ensuremath{\textnormal{syn}}}(M):{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}\rightarrow {\ensuremath{\textnormal{\textbf{cmon}}}\xspace}$ given on objects by $n \mapsto \Psi^{\ensuremath{\textnormal{syn}}}_n(M)$ such that the given isomorphisms $\Psi_n(M) \cong \Psi^{\ensuremath{\textnormal{syn}}}_n(M)$ are natural in $n \in {\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$. The naturality of these isomorphisms together with the naturality of the isomorphism $\psi:[D^\sharp(-),M] \xrightarrow{\sim} T^{(-)}M$ (\[thm:symm\_deg\_ind\_by\_d\]) entails the remaining claim.
\[thm:eltwise\_formula\_ext\_deriv\] Given a synthetic sector $n$-form $\nu:[D^n,M] \rightarrow E$ and any $i = 1,...,n+1$, the $i$-th coface of $\nu$ is the synthetic sector $(n+1)$-form $\delta^n_i(\nu):[D^{n+1},M] \rightarrow E$ characterized by the following equation: $$\tau:[D^{n+1},M] \;\vdash\; (\delta^n_i(\nu))(\tau) = \hat{p}\bigl({\ensuremath{\underline{\lambda}}}d_0\colon D.\nu({\ensuremath{\underline{\lambda}}}(d_1,...,d_n)\colon D^n.\tau(d_1,...,d_{i-1},d_0,d_i,...,d_n))\bigr)$$ where $\hat{p}:[D,E] = TE \rightarrow E$ is the principal projection (\[sec:princ\_projn\]).
By \[thm:main\_fund\_coface\_lemma\] we know that $\delta^n_i(\nu) = \sigma^{n+1}_{(i)}(\delta^n_1(\nu))$ where $\sigma^{n+1}_{(i)}:\Psi^{\ensuremath{\textnormal{syn}}}_{n+1}(M) \rightarrow \Psi^{\ensuremath{\textnormal{syn}}}_{n+1}(M)$ is the automorphism induced by the permutation $\sigma^{n+1}_{(i)}:n+1 \rightarrow n+1$ in ${\ensuremath{\textnormal{\textbf{finCard}}}\xspace}$.
Given any synthetic sector $(n+1)$-form $\gamma:[D^{n+1},M] \rightarrow E$, we deduce by \[thm:symm\_cosimpl\_cmon\_syn\_sf\] that $\sigma^{n+1}_{(i)}(\gamma)$ is the composite $$[D^{n+1},M] \xrightarrow{[D^\sharp(\sigma^{n+1}_{(i)}),M]} [D^{n+1},M] \overset{\gamma}{\longrightarrow} E,$$ so by using we compute as follows: $$\label{eq:deriv_syn_sf1}\tau:[D^{n+1},M] \;\vdash\; (\sigma^{n+1}_{(i)}(\gamma))(\tau) = \gamma({\ensuremath{\underline{\lambda}}}(d_0,...,d_n).\tau(d_1,...,d_{i-1},d_0,d_i,...,d_n))$$
Let $\omega:T^nM \rightarrow E$ be the sector $n$-form corresponding to $\nu$, so that $$\nu = \left([D^n,M] \xrightarrow{\psi_n} T^nM \xrightarrow{\omega} E\right).$$ In view of the proof of \[thm:symm\_cosimpl\_cmon\_syn\_sf\], $\delta^n_1(\nu):[D^{n+1},M] \rightarrow E$ is the composite $$[D^{n+1},M] \xrightarrow{\psi_{n+1}} T^{n+1}M \xrightarrow{\delta^n_1(\omega)} E$$ where $\delta^n_1(\omega) = T\omega {\ensuremath{\mathop{;}}}\hat{p}$ is the fundamental derivative of $\omega$ (\[sec:fund\_deriv\_sf\], \[thm:sym\_cosimpl\_cmon\_sector\_forms\]). Hence since $\omega = \psi^{-1}_n {\ensuremath{\mathop{;}}}\nu$ we compute that $$\delta^n_1(\nu) \;=\; \psi_{n+1} {\ensuremath{\mathop{;}}}T(\psi_n^{-1}) {\ensuremath{\mathop{;}}}T\nu {\ensuremath{\mathop{;}}}\hat{p} \;=\; \phi_{n+1} {\ensuremath{\mathop{;}}}T\nu {\ensuremath{\mathop{;}}}\hat{p} \;=\; \phi_{n+1} {\ensuremath{\mathop{;}}}[D,\nu] {\ensuremath{\mathop{;}}}\hat{p}$$ by the inductive definition of $\psi$ (\[def:isos\_rep\_it\_tngnt\]). Hence we compute as follows: $$\label{eq:formula_for_fund_deriv_synsf}\tau:[D^{n+1},M] \;\vdash\; (\delta^n_1(\nu))(\tau) = \hat{p}(\phi_{n+1}(\tau) {\ensuremath{\mathop{;}}}\nu) = \hat{p}({\ensuremath{\underline{\lambda}}}d_0.\nu({\ensuremath{\underline{\lambda}}}(d_1,...,d_n).\tau(d_0,d_1,...,d_n)))$$ Applying this together with in the case where $\gamma = \delta^n_1(\nu)$, we obtain the needed result.
Now let us assume that $E$ is a subtractive differential object in ${\ensuremath{\mathscr X}\xspace}$, and again let $M$ be an object of ${\ensuremath{\mathscr X}\xspace}$.
\[def:synth\_singf\]
1. We call a synthetic sector $n$-form $\nu$ on $M$ a **synthetic singular $n$-form** if $\nu$ is an alternating element of the symmetric cosimplicial abelian group $\Psi^{\ensuremath{\textnormal{syn}}}(M)$ (\[def:alt\_elt\]).
2. We define the cochain complex $\Omega^{\ensuremath{\textnormal{syn}}}_\bullet(M)$ as the complex of alternating elements (\[def:compl\_alt\_elts\]) of $\Psi^{\ensuremath{\textnormal{syn}}}(M)$, and we call $\Omega^{\ensuremath{\textnormal{syn}}}_\bullet(M)$ the **complex of synthetic singular forms** on $M$ with values in $E$.
By \[thm:symm\_cosimpl\_cmon\_syn\_sf\] and \[thm:dsharp\], the symmetry $\sigma^n_i:\Psi^{\ensuremath{\textnormal{syn}}}_n(M) \rightarrow \Psi^{\ensuremath{\textnormal{syn}}}_n(M)$ $(i = 1,...,n-1)$ sends each synthetic sector $n$-form $\nu$ to the composite $$[D^n,M] \xrightarrow{[s^n_i,M]} [D^n,M] \overset{\nu}{\longrightarrow} E$$ in the notation of \[thm:dsharp\]. Hence we obtain the following:
A morphism $\nu:[D^n,M] \rightarrow E$ is a synthetic singular $n$-form iff $\nu$ is a synthetic sector $n$-form and $$\label{eq:synth_alt_ax}\tau:[D^n,M] \;\vdash\; \nu\bigl(\lambda(d_1,...,d_n).\tau(d_1,...,d_{i-1},d_{i+1},d_i,d_{i+2},...,d_n)\bigr) \:=\: -\nu(\tau)\;\;:\;\;E$$ holds for each $i = 1,...,n-1$, where $-:E \rightarrow E$ is the negation morphism.
\[thm:compl\_singf\_iso\_synsingf\] There is an isomorphism of cochain complexes $$\Omega_\bullet(M) \cong \Omega^{\ensuremath{\textnormal{syn}}}_\bullet(M)$$ between the complex of singular forms and the complex of synthetic singular forms.
This follows immediately from \[thm:symm\_cosimpl\_cmon\_syn\_sf\], \[thm:func\_compl\_alt\_elts\], and \[def:sing\_nform\_compl\].
Relationship to de Rham in SDG {#sec:compn_forms_sdg}
------------------------------
Let ${\ensuremath{\mathscr E}\xspace}$ be a topos equipped with a commutative ring object $R$. Writing the multiplication in $R$ as juxtaposition in the lambda calculus, let $(-)^2,0:R \rightarrow R$ denote the morphisms in ${\ensuremath{\mathscr E}\xspace}$ denoted by the terms-in-context $x:R \vdash xx:R$ and $x:R \vdash 0:R$, respectively. Let $D \hookrightarrow R$ denote the equalizer of these morphisms $(-)^2,0$.
An $R$-module object $E$ in ${\ensuremath{\mathscr E}\xspace}$ is said to be a **Kock-Lawvere $R$-module**, or satisfy the **Kock-Lawvere axiom**, if the morphism $$\label{eq:kl_compn_morph}E \times E \xrightarrow{\kappa} [D,E]$$ denoted by $$(e_1,e_0):E \times E \;\vdash\; {\ensuremath{\underline{\lambda}}}d:D\:.\:de_1 + e_0$$ is an isomorphism, where we have written the action of $R$ on $E$ as juxtaposition. $R$ is said to be a **ring of line type** if $R$ itself satisfies the Kock-Lawvere axiom when considered as an $R$-module.
Let us now assume that $R$ is a ring of line type. The books [@kock I.14], [@reyes IV], [@lavendhomme Ch. 4] treat differential forms valued in a Kock-Lawvere $R$-module $E$, following [@Kock:DiffFormsSDG; @KRV; @MoeRey:CohThSDG]. We now recall the definition of the notion of differential form employed in the cited sections of these books. We shall soon show that these are exactly the same as the synthetic singular forms defined in \[def:synth\_singf\] above.
\[def:sdg\_sing\_form\] Let $M$ be an object of ${\ensuremath{\mathscr E}\xspace}$, and let $n \in {\ensuremath{\mathbb N}\xspace}$. For each $j = 1,...,n$, define a morphism $*^n_j:R \times [D^n,M] \rightarrow [D^n,M]$ by $$r:R, \tau:[D^n,M]\;\vdash\; r *^n_j \tau = {\ensuremath{\underline{\lambda}}}(d_1,...,d_n).\tau(d_1,...,d_{j-1},rd_j,d_{j+1},...,d_n)$$ where we have written $*^n_j$ in infix notation. We say that a morphism $\nu:[D^n,M] \rightarrow E$ is an **SDG singular $n$-form** on $M$ with values in $E$ if the equation holds for all $i = 1,...,n-1$ and the following equation holds for all $j = 1,...,n$: $$\label{eq:sdg_sing_form_multilin_ax}r:R, \tau:[D^n,M] \;\vdash\; \nu(r *^n_j \tau) = r\nu(\tau)\;\;:\;\;E$$
Observe that the axioms for an SDG singular form are almost exactly the same as those for a synthetic singular form as defined in \[def:synth\_singf\] above, except that for SDG singular forms the axiom applies to arbitrary scalars $r:R$ rather than just $d:D$. We shall show that these notions of form are identical in a suitable setting. The key idea is as follows.
\[thm:pres\_dact\_implies\_pres\_ract\] If a morphism $h:E \rightarrow F$ in ${\ensuremath{\mathscr E}\xspace}$ between Kock-Lawvere $R$-modules $E,F$ satisfies the equation $d:D, e:E \vdash h(de) = dh(e)$, then $h$ satisfies the equation $r:R, e:E \vdash h(re) = rh(e)$.
Let $\rho,\tau:R \times E \rightarrow F$ be the morphisms denoted by $r:R, e:E \vdash h(re)$ and $r:R, e:E \vdash rh(e)$, respectively. Note that the multiplicative action $D \times R \rightarrow R$ factors through the inclusion $D \hookrightarrow R$ since $d:D, r:R \vdash (dr)^2 = d^2r^2 = 0r^2 = 0$ holds. Denoting the resulting morphism $D \times R \rightarrow D$ by $d:D, r:R \vdash dr:D$, we compute that $$d:D, r:R, e:E \vdash dh(re) = h(dre) = drh(e)\;:\;F$$ since $h$ preserves the actions by $D$. Letting $\lambda:F \rightarrow [D,F]$ be the morphism denoted by $f:F \vdash {\ensuremath{\underline{\lambda}}}d.df$, we compute that $$r:R, e:E \vdash \lambda(\rho(r,e)) = ({\ensuremath{\underline{\lambda}}}d.dh(re)) = ({\ensuremath{\underline{\lambda}}}d.drh(e)) = \lambda(\tau(r,e))\;:\;[D,F]$$ so $\rho {\ensuremath{\mathop{;}}}\lambda = \tau {\ensuremath{\mathop{;}}}\lambda$. But $\lambda$ is the composite $F \xrightarrow{(1_F,0)} F \times F \xrightarrow{\kappa} [D,F]$ in the notation of , so $\lambda$ is a split monomorphism and hence $\rho = \tau$.
The tangent category of microlinear objects {#sec:tngt_cat_microlin_objs}
-------------------------------------------
In order to be able to use certain results given in [@lavendhomme], we shall now assume that $R$ satisfies the **Kock-Weil axiom (K-W)** of [@lavendhomme 2.1.3]. This axiom entails the above Kock-Lawvere axiom, and it also entails that the object $R$ of ${\ensuremath{\mathscr E}\xspace}$ is **microlinear** [@lavendhomme 2.3.1] (that is, $R$ perceives finite quasi-colimits of infinitesimal objects as colimits).
Let $${\ensuremath{\mathscr X}\xspace}\hookrightarrow {\ensuremath{\mathscr E}\xspace}$$ denote the full subcategory of ${\ensuremath{\mathscr E}\xspace}$ consisting of the microlinear objects, and let ${\ensuremath{\mathscr E}\xspace}_{iv} \hookrightarrow {\ensuremath{\mathscr E}\xspace}$ denote the full subcategory consisting of those objects that are both *infinitesimally linear* and *vertically linear* in the sense used in [@sman3 5.2, 5.3]. Note that ${\ensuremath{\mathscr X}\xspace}$ is contained in ${\ensuremath{\mathscr E}\xspace}_{iv}$. By [@sman3 5.4] ${\ensuremath{\mathscr E}\xspace}_{iv}$ carries representable tangent structure, with representing object $D$. Both ${\ensuremath{\mathscr X}\xspace}$ and ${\ensuremath{\mathscr E}\xspace}_{iv}$ are closed under finite limits and exponentials in ${\ensuremath{\mathscr E}\xspace}$ ([@lavendhomme 2.3.1], [@sman3 5.4]). Hence the tangent structure on ${\ensuremath{\mathscr E}\xspace}_{iv}$ restricts to ${\ensuremath{\mathscr X}\xspace}$, and $$\begin{minipage}{3.6in}\textit{${\ensuremath{\mathscr X}\xspace}$ is a Cartesian closed category with representable tangent structure, represented by $D$.}\end{minipage}$$
Now let us fix an object $M$ of ${\ensuremath{\mathscr X}\xspace}$ and a Kock-Lawvere $R$-module $E$ that lies in ${\ensuremath{\mathscr X}\xspace}$. By [@diffBundles 3.9], $E$ carries the structure of a subtractive differential object in ${\ensuremath{\mathscr X}\xspace}$, where the associated lift morphism $\lambda:E \rightarrow TE = [D,E]$ is the transpose of the restricted action $D \times E \rightarrow E$, so that the latter is the morphism written as $\bullet$ in \[sec:infl\_actions\].
Given any object $X$ of ${\ensuremath{\mathscr E}\xspace}$, let $X^*:{\ensuremath{\mathscr E}\xspace}= {\ensuremath{\mathscr E}\xspace}\slash 1 \rightarrow {\ensuremath{\mathscr E}\xspace}\slash X$ denote the functor given by pullback along $!:X \rightarrow 1$. Since $X^*$ is a logical functor between toposes [@Joh:TT 1.42], $X^*$ sends $R$ to a ring of line type $X^*(R) = (\pi_2:R \times X \rightarrow X)$ in ${\ensuremath{\mathscr E}\xspace}\slash X$, and $X^*$ sends $E$ to a Kock-Lawvere $X^*(R)$-module $X^*(E)$ in ${\ensuremath{\mathscr E}\xspace}\slash X$.
\[thm:kl\_rmods\_structs\_it\_tan\]
1. Given any object $X$ of ${\ensuremath{\mathscr X}\xspace}$, the tangent bundle $(TX,p_X)$ carries the structure of a Kock-Lawvere $X^*(R)$-module in ${\ensuremath{\mathscr E}\xspace}\slash X$.
2. For each $n \in {\ensuremath{\mathbb N}\xspace}$ and each $j = 1,...,n+1$, the morphism $$p_j:[D^{n+1},M] \rightarrow [D^n,M]$$ given by $$\tau:[D^{n+1},M] \;\vdash\; p_j(\tau) = {\ensuremath{\underline{\lambda}}}(d_1,...,d_n).\tau(d_1,...,d_{j-1},0,d_j,...,d_n)$$ carries the structure of a Kock-Lawvere $([D^n,M])^*(R)$-module in ${\ensuremath{\mathscr E}\xspace}\slash [D^n,M]$. The associated action of $([D^n,M])^*(R)$ is given by the morphism $$*^{n+1}_j:R \times [D^{n+1},M] \rightarrow [D^{n+1},M]$$ of \[def:sdg\_sing\_form\].
\(i) is established in [@lavendhomme 3.1.2, Prop. 4], noting that Lavendhomme’s universal quantification over points of the base space is interpreted in an internal sense in ${\ensuremath{\mathscr E}\xspace}$, cf. [@kock II.6]. Letting $X = [D^n,M]$ and $Y = [D^{n+1},M]$, we have an isomorphism $\phi_{n+1}:Y \xrightarrow{\sim} [D,X] = TX$ (\[def:isos\_rep\_it\_tngnt\]) that commutes with the projections $p_1$ and $p_X$ to $X$. Hence by (i), $(Y,p_1)$ carries the structure of a Kock-Lawvere $X^*(R)$-module in ${\ensuremath{\mathscr E}\xspace}\slash X$, and in view of [@lavendhomme 3.1.1] the associated action of $X^*(R)$ is $*^{n+1}_1$. By we have an automorphism $\Sigma_{(j)} = [D^\sharp(\sigma^{n+1}_{(j)}),M]:Y \xrightarrow{\sim} Y$ in ${\ensuremath{\mathscr X}\xspace}\hookrightarrow {\ensuremath{\mathscr E}\xspace}$ characterized by $$\tau:[D^{n+1},M] \;\vdash\; \Sigma_{(j)}(\tau) = {\ensuremath{\underline{\lambda}}}(d_1,...,d_{n+1}).\tau(d_2,...,d_j,d_1,d_{j+1},...,d_{n+1})\;.$$ Observe that $\Sigma_{(j)}$ is in fact an isomorphism $\Sigma_{(j)}:(Y,p_j) \xrightarrow{\sim} (Y,p_1)$ in ${\ensuremath{\mathscr E}\xspace}\slash X$. Hence there is a unique $X^*(R)$-module structure on $(Y,p_j)$ such that $\Sigma_{(j)}$ is an isomorphism of $X^*(R)$-modules $(Y,p_j) \xrightarrow{\sim} (Y,p_1)$. Hence $(Y,p_j)$ is a Kock-Lawvere $X^*(R)$-module, and the result follows.
\[thm:sdgsing\_vs\_synthsing\] A morphism $\nu:[D^n,M] \rightarrow E$ is an SDG singular $n$-form if and only if $\nu$ is a synthetic singular $n$-form in the sense of \[def:synth\_singf\].
If $n = 0$ then the result is immediate, so we may assume that $n \geqslant 1$. Since the morphisms $\bullet^n_j$ of \[def:synth\_sf\] are restrictions of the morphisms $*^n_j$ of \[def:sdg\_sing\_form\], it suffices to assume that $\nu$ is a synthetic singular $n$-form and show that holds for each $j \in \{1,...,n\}$. Letting $Y = [D^n,M]$ and $X = [D^{n-1},M]$, we deduce by \[thm:kl\_rmods\_structs\_it\_tan\] that $(Y,p_j)$ is a Kock-Lawvere $X^*(R)$-module with action morphism $*^n_j:R \times Y \rightarrow Y$. Also $X^*(E) = (E \times X,\pi_2)$ is a Kock-Lawvere $X^*(R)$-module, and $\nu$ induces a morphism $\bar{\nu} := (\nu,p_j):(Y,p_j) \rightarrow X^*(E)$ in ${\ensuremath{\mathscr E}\xspace}\slash X$. Note that since $X^*$ preserves limits, $X^*$ sends $D \hookrightarrow R$ to the square zero part $X^*(D) \hookrightarrow X^*(R)$ of $X^*(R)$. But the axiom entails that $\bar{\nu}$ preserves the $X^*(D)$-actions carried by the Kock-Lawvere $X^*(R)$-modules $(Y,p_j)$ and $X^*(E)$, so $\bar{\nu}$ preserves the $X^*(R)$-actions by \[thm:pres\_dact\_implies\_pres\_ract\]. Hence holds.
\[thm:singf\_vs\_sdgsingf\] The complex of singular forms $\Omega_\bullet(M)$ is isomorphic to the complex $\Omega^{\textnormal{sdg}}_\bullet(M)$ of SDG singular forms on $M$ with values in $E$, as defined in [@Kock:DiffFormsSDG; @KRV], [@lavendhomme Ch. 4]. Further, $\Omega^{\textnormal{sdg}}_\bullet(M)$ is identical to the complex of synthetic singular forms $\Omega^{\ensuremath{\textnormal{syn}}}_\bullet(M)$.
By \[thm:compl\_singf\_iso\_synsingf\], it suffices to prove the second claim. By \[thm:sdgsing\_vs\_synthsing\], the complexes $\Omega^{\textnormal{sdg}}_\bullet(M)$ and $\Omega^{\ensuremath{\textnormal{syn}}}_\bullet(M)$ have the same underlying graded abelian group. The differential $\partial_n:\Omega^{\ensuremath{\textnormal{syn}}}_n(M) \rightarrow \Omega^{\ensuremath{\textnormal{syn}}}_{n+1}(M)$ is given by $\partial_n(\nu) = \sum_{i=1}^{n+1}(-1)^{n-1}\delta^n_i(\nu)$, and by consulting the formula for $\delta^n_i(\nu)$ in \[thm:eltwise\_formula\_ext\_deriv\] we find that $\partial_n(\nu)$ is precisely the exterior derivative of $\nu$ described in [@KRV §1, (1.3)] and [@lavendhomme 4.2.3, Prop. 4].
Relationship to de Rham for smooth manifolds {#sec:classical_derham}
--------------------------------------------
Let ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ denote the category of (Hausdorff, second-countable) smooth manifolds, and let $M$ be an object of ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$. Since ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ is a Cartesian tangent category and ${\ensuremath{\mathbb R}\xspace}$ is a subtractive differential object in ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$, we can consider the complex of singular forms $\Omega_\bullet(M)$ on $M$ with values in ${\ensuremath{\mathbb R}\xspace}$ (\[def:sing\_nform\_compl\]). In order to show that $\Omega_\bullet(M)$ is isomorphic to the classical de Rham complex of $M$[^12], we shall consider an embedding of ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ into a topos ${\ensuremath{\mathscr E}\xspace}$ modelling SDG, and then we shall invoke \[thm:singf\_vs\_sdgsingf\] and a result on differential forms within this specific topos [@reyes IV.3.7].
In particular, we shall take ${\ensuremath{\mathscr E}\xspace}$ to be *the Dubuc topos*, i.e., the topos denoted by $\mathcal{G}$ in [@reyes] and by $\widetilde{\mathcal{B}^{{\mbox{\textnormal{\scriptsize op}}}}}$ in [@kock]. Explicitly, ${\ensuremath{\mathscr E}\xspace}$ is the topos of sheaves on the opposite of the category of germ-determined, finitely generated $C^\infty$-rings, with respect to the open cover topology. But more to the point, ${\ensuremath{\mathscr E}\xspace}$ is a topos equipped with an embedding $$\iota:{\ensuremath{\textnormal{\textbf{mf}}}\xspace}\hookrightarrow {\ensuremath{\mathscr E}\xspace}$$ such that $R = \iota({\ensuremath{\mathbb R}\xspace})$ is a ring of line type satisfying the Kock-Weil axiom [@lavendhomme 8.3.3], and $\iota$ has several further pleasant properties making $({\ensuremath{\mathscr E}\xspace},\iota)$ a *well-adapted model* (see [@kock III.3, III.4, III.8.4]). We now show that any well-adapted model gives rise to an embedding of tangent categories:
\[thm:well\_ad\_model\_yields\_str\_mor\_tngt\_cats\] Let $\iota:{\ensuremath{\textnormal{\textbf{mf}}}\xspace}\hookrightarrow {\ensuremath{\mathscr E}\xspace}$ be a well-adapted model of SDG.
1. The embedding $\iota$ factors as ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}\overset{\iota'}{\hookrightarrow} {\ensuremath{\mathscr X}\xspace}\hookrightarrow {\ensuremath{\mathscr E}\xspace}$, where ${\ensuremath{\mathscr X}\xspace}$ is the full subcategory of ${\ensuremath{\mathscr E}\xspace}$ consisting of microlinear objects.
2. The embedding $\iota':{\ensuremath{\textnormal{\textbf{mf}}}\xspace}\hookrightarrow {\ensuremath{\mathscr X}\xspace}$ carries the structure of a strong morphism of Cartesian tangent categories (in the sense of [@sman3 2.7, 2.8]).
For any manifold $M \in {\ensuremath{\mathop{\textnormal{ob}}}}{\ensuremath{\textnormal{\textbf{mf}}}\xspace}$, the object $\iota(M)$ of ${\ensuremath{\mathscr E}\xspace}$ is a *formal manifold* [@kock I.17] (by [@kock III.3.4, III.4.C]), so by [@kock I, Notes 2006, Footnote 27; Appendix D]), $\iota(M)$ is microlinear and (i) is proved. Letting $T$ denote the tangent endofunctor on ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$, it is proved in [@kock III.4.1] that there is an isomorphism $\alpha_M:\iota(TM) \xrightarrow{\sim} [D,\iota(M)]$ natural in $M \in {\ensuremath{\textnormal{\textbf{mf}}}\xspace}$, recalling that $[D,-]:{\ensuremath{\mathscr X}\xspace}\rightarrow {\ensuremath{\mathscr X}\xspace}$ is the tangent endofunctor on ${\ensuremath{\mathscr X}\xspace}$ (\[sec:tngt\_cat\_microlin\_objs\]). Also, by [@kock III.3.A], $\iota$ preserves finite products as well as the (iterated) pullbacks that are given as part of the tangent structure on ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ (\[defnTangentCategory\]), since these are all *transversal pullbacks* (as each can be described as a pullback of a submersion; see [@diffBundles]). But ${\ensuremath{\mathscr X}\xspace}$ is closed under finite limits in ${\ensuremath{\mathscr E}\xspace}$ (\[sec:tngt\_cat\_microlin\_objs\]), so it remains only to show that $\alpha$ satisfies the diagrammatic axioms[^13] of [@sman3 2.7].
Letting ${\ensuremath{\textnormal{\textbf{cart}}}
\xspace}\hookrightarrow {\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ denote the full subcategory consisting of the *Cartesian spaces* ${\ensuremath{\mathbb R}\xspace}^n$, we know that the Cartesian tangent structure on ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$ restricts to a Cartesian tangent structure on ${\ensuremath{\textnormal{\textbf{cart}}}
\xspace}$, wherein $T({\ensuremath{\mathbb R}\xspace}^n) = {\ensuremath{\mathbb R}\xspace}^n \times {\ensuremath{\mathbb R}\xspace}^n$. Indeed, ${\ensuremath{\textnormal{\textbf{cart}}}
\xspace}$ is a particularly simple kind of tangent category, namely (the tangent category associated to) a *Cartesian differential category* [@sman3 4.1, 4.2]. The embedding $\iota$ sends the Cartesian spaces ${\ensuremath{\mathbb R}\xspace}^n$ to Kock-Lawvere $R$-modules $\iota({\ensuremath{\mathbb R}\xspace}^n) \cong R^n$ in ${\ensuremath{\mathscr E}\xspace}$, where $R = \iota({\ensuremath{\mathbb R}\xspace})$ is the line object in ${\ensuremath{\mathscr E}\xspace}$. The microlinear Kock-Lawvere $R$-modules in ${\ensuremath{\mathscr E}\xspace}$ are precisely the differential objects of ${\ensuremath{\mathscr X}\xspace}$ [@diffBundles 3.9] and so constitute a Cartesian differential category ${\ensuremath{\textnormal{\textbf{Diff}}}
\xspace}({\ensuremath{\mathscr X}\xspace})$ [@sman3 4.11], such that the inclusion ${\ensuremath{\textnormal{\textbf{Diff}}}
\xspace}({\ensuremath{\mathscr X}\xspace}) \hookrightarrow {\ensuremath{\mathscr X}\xspace}$ carries the structure of a strong morphism of Cartesian tangent categories [@sman3 4.12]. The restriction $\iota'':{\ensuremath{\textnormal{\textbf{cart}}}
\xspace}\hookrightarrow {\ensuremath{\textnormal{\textbf{Diff}}}
\xspace}({\ensuremath{\mathscr X}\xspace})$ of $\iota'$ preserves finite products, and $\iota''$ preserves the Jacobian derivative [@kock III.3.3] and therefore preserves Cartesian differential structure. Therefore $\iota''$ is a strong morphism of Cartesian tangent structure, so the composite ${\ensuremath{\textnormal{\textbf{cart}}}
\xspace}\hookrightarrow {\ensuremath{\textnormal{\textbf{Diff}}}
\xspace}({\ensuremath{\mathscr X}\xspace}) \hookrightarrow {\ensuremath{\mathscr X}\xspace}$ is a strong morphism of Cartesian tangent structure whose structural isomorphisms $$\iota({\ensuremath{\mathbb R}\xspace}^n \times {\ensuremath{\mathbb R}\xspace}^n) = \iota(T{\ensuremath{\mathbb R}\xspace}^n) \xrightarrow{\sim} [D,\iota({\ensuremath{\mathbb R}\xspace}^n)]$$ are the isomorphisms $\alpha_M$ with $M = {\ensuremath{\mathbb R}\xspace}^n$ as constructed in [@kock III.4.1].
Hence the restriction of $\iota':{\ensuremath{\textnormal{\textbf{mf}}}\xspace}\hookrightarrow {\ensuremath{\mathscr X}\xspace}$ to ${\ensuremath{\textnormal{\textbf{cart}}}
\xspace}$ is a strong morphism of Cartesian tangent structure. Now for an arbitrary manifold $M \in {\ensuremath{\mathop{\textnormal{ob}}}}{\ensuremath{\textnormal{\textbf{mf}}}\xspace}$, we can choose a covering by open embeddings $e_j:U_j \hookrightarrow M$ $(j \in J)$ where the $U_j$ are Cartesian spaces. It follows that the families $(Te_j)_{j \in J}$, $(T^2e_j)_{j \in J}$, and $(T_2e_i)_{j \in J}$ (in the notation of \[defnTangentCategory\]) are coverings by open embeddings. For each of the structural transformations $t \in \{p,0,+,\ell,c\}$ carried by ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$, with corresponding transformation $t'$ in ${\ensuremath{\mathscr X}\xspace}$, we can now use the fact that $\alpha_{U_j}$ commutes with $t_{U_j},t'_{U_j}$ for each $j \in J$ to show that $\alpha_M$ commutes with $t_M,t'_M$.
\[thm:classical\_derham\] The classical de Rham complex $\Omega^{\textnormal{dR}}_\bullet(M)$ of a smooth manifold $M$ is isomorphic to the complex $\Omega_\bullet(M)$ of singular forms on $M$ with values in ${\ensuremath{\mathbb R}\xspace}$ (\[def:sing\_nform\_compl\]), where $M$ is considered as an object of the tangent category ${\ensuremath{\textnormal{\textbf{mf}}}\xspace}$. In symbols, $$\Omega^{\textnormal{dR}}_\bullet(M) \;\cong\; \Omega_\bullet(M)\;.$$
Taking ${\ensuremath{\mathscr E}\xspace}$ to be the Dubuc topos, we deduce by \[thm:well\_ad\_model\_yields\_str\_mor\_tngt\_cats\] that the associated embedding $\iota':{\ensuremath{\textnormal{\textbf{mf}}}\xspace}\hookrightarrow {\ensuremath{\mathscr X}\xspace}$ is a strong morphism of Cartesian tangent structure, and we can invoke \[thm:singf\_vs\_sdgsingf\] and [@reyes IV.3.7] in order to compute that $$\Omega_\bullet(M) \cong \Omega_\bullet(\iota'(M)) \cong \Omega^{\textnormal{sdg}}_\bullet(\iota(M)) \cong \Omega^{\textnormal{dR}}_\bullet(M)$$ as complexes, where $\Omega_\bullet(\iota'(M))$ is the complex of singular forms on $\iota'(M)$ in ${\ensuremath{\mathscr X}\xspace}$ with values in $R = \iota'({\ensuremath{\mathbb R}\xspace})$.
Conclusions and future work {#sec:futureWork}
===========================
In this paper, we have shown that not only do tangent categories support a generalization of de Rham cohomology, but that they support a second cohomology, the cohomology of sector forms; furthermore, sector forms have a rich algebraic structure that goes beyond this cohomology. There are many possible extensions of this work.
- We have shown that tangent categories possess a cohomology of sector forms. Even in the canonical case of smooth manifolds, this may be distinct from the ordinary de Rham cohomology of classical differential forms; further investigation is required to compare these cohomologies.
- The relationship between classical differential forms and singular forms in an arbitrary tangent category needs to be better understood. In general, one would expect that any object $M$ which is “locally a differential object” would have the property that classical differential forms on $M$ and singular forms on $M$ would be in bijective correspondence, but this requires detailed work to check. Another possibility is that differential forms and singular forms may correspond if $M$ possesses a “symmetric $n$-connection” [@diffFormsLavendhommeNishimura], suitably defined in a tangent category.
- An important operation on differential forms is the wedge product. Since this involves multiplication in $\mathbb{R}$, in the setting of tangent categories, one would need the coefficient object $E$ to have ring structure. Once such a generalized wedge product is defined, one could consider how such an operation interacts with the co-face, symmetry, and co-degeneracy maps.
- It is a well-known result that the exterior derivative is the unique map from $n$ forms to $n+1$ forms satisfying certain algebraic properties [@spivakVol1 Proposition 7.11]. It would be interesting to determine for which tangent categories this uniqueness result holds.
- It is not clear to what cohomology theories the cohomologies found here correspond in algebraic geometry (for example, in the category of schemes). The cohomologies may recover an existing cohomology theory or represent a new one; further investigation is required.
Finally, as mentioned in the introduction, sector forms generalize covariant tensors, and because of this, White writes that “the *calculus* of \[sector forms\] can serve as a unified framework for the presentation of classical local Riemannian geometry, and that it can lead to new methods of analysis in modern differential geometry” [@white pg. x]. The results presented here on sector forms contribute to this calculus by means of a methodology which is applicable more generally.
[^1]: Note that while pre-composing with the projection or zero natural transformations gives an obvious way to define co-face and co-degeneracy operations, these are not the operations we are using: see section \[sec:results\] for a basic overview of the definitions of these operations.
[^2]: The first person to write on sector forms in standard differential geometry, J.E. White, describes the co-face and symmetry maps of sector forms, but does not describe the the co-degeneracy maps or any of the equations these various maps must satisfy in order to yield a symmetric cosimplicial object [@white Definition 3.8, 3.12]
[^3]: In the context of synthetic differential geometry, there is a different way of obtaining the de Rham complex from simplicial data, using the idea of *infinitesimal cochains* [@kock Section I.18] [@diffFormsasInfCochains]. However, this approach is different from the approach via sector forms that we pursue here.
[^4]: That is, unital rings without additive inverses, also known as unital semirings.
[^5]: White instead states the definition in terms of a criterion of fibrewise linearity.
[^6]: One can also show that *all* sector 2-forms on $\mathbb{R}$ are of this form. A general description of sector $n$-forms on ${\ensuremath{\mathbb R}\xspace}^m$ was given by White [@white].
[^7]: The terminology is due to Grandis [@Gra:Symm]; Barr [@Barr] employed the term *augmented FDP complex*.
[^8]: Explicitly, determines a functor ${\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}]$ which we can then compose with the functor $[{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\mathscr X}\xspace}(-,E)]:[{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}] \rightarrow [{\ensuremath{\textnormal{\textbf{finCard}}}\xspace}_s,{\ensuremath{\textnormal{\textbf{set}}}\xspace}]$ given by composition with the hom-functor ${\ensuremath{\mathscr X}\xspace}(-,E):{\ensuremath{\mathscr X}\xspace}^{{\mbox{\textnormal{\scriptsize op}}}}\rightarrow {\ensuremath{\textnormal{\textbf{set}}}\xspace}$.
[^9]: rather than a breach of the well-foundedness axiom of set theory
[^10]: Although it is well known, the authors are unable to find an explicit statement of precisely this fact in the literature. However, it can be verified almost immediately, and it can be seen also as a corollary to [@stacks-project [Tag 0194](http://stacks.math.columbia.edu/tag/0194)], in view of [@stacks-project [Tag 018F](http://stacks.math.columbia.edu/tag/018F)].
[^11]: One way of justifying this claim is to embed ${\ensuremath{\mathscr X}\xspace}$ into the Cartesian closed category of presheaves on ${\ensuremath{\mathscr X}\xspace}$ taking values in some universe with respect to which ${\ensuremath{\mathscr X}\xspace}$ is small. But clearly more elementary approaches are possible.
[^12]: The relationship between singular forms and classical differential forms is also briefly discussed in an exercise in White’s book; see [@white pg. 116, exercise 7].
[^13]: Note that composition of functors is written in diagrammatic order in the cited source.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we obtain local in time existence and (suitable) uniqueness and continuous dependence for the KP-I equation for small data in the intersection of the energy space and a natural weighted $L^{2}$ space.'
address:
- University of Toronto
- University of Chicago
- Massachusetts Institute of Technology
author:
- 'J. Colliander'
- 'C. Kenig'
- 'G. Staffilani'
title: 'Low Regularity Solutions for the Kadomtsev-Petviashvili I Equation'
---
[^1]
[^2]
[^3]
Introduction
============
We consider the KP initial value problem (IVP) $$\left\{ \begin{array}{l}
\partial_t u + \partial_x^3 u + \gamma\partial_x^{-1}
\partial_y^2u
+\beta\partial_x (u^2) = 0, \\
u(x,0) = u_0(x) \hspace{1.5cm}(x,y) \in {{{\mathbb }R}}^2,
\, t \in {{{\mathbb }R}},
\end{array}\right.
\label{ivp}$$ where $u=u(t,x,y)$ is a scalar unknown function, $\beta\ne 0$ and $\gamma\ne 0$ are real constant. If $\gamma<0$ the IVP (\[ivp\]) is called KP-I and if $\gamma>0$ it takes the name KP-II. These equations model [@KP] the propagation along the $x$-axis of nonlinear dispersive long waves on the surface of a fluid with a slow variation along the $y$-axis. They also arise as universal models in wave propagation and may be considered as two dimensional generalizations of the Korteweg-de Vries equation.
The first result regarding well-posedness for a KP type equation is due to Ukai [@U]. He uses a standard energy method that does not recognize the type I or II of the equation. His result provides local well-posedness for initial data and their antiderivatives in $ H^{s}, \, s\geq 3$. Faminskii [@F] observed a better smoothing effect in the KP-II evolution and used this to prove well-posedness results. Bourgain performed a Fourier analysis [@B] of the term $\partial_x (u^2)$ in the KP-II equation in which the derivative is recovered in a nonlinear way. The result obtained gave local well-posedness of KP-II for initial data in $L^2 $. Since the $L^2$ norm is conserved during the KP-II evolution, the $L^2$ local result may be iterated to prove global well-posedness. Takaoka [@Ta1] and Takaoka and Tzvetkov [@TT] improved Bourgain’s result by proving local well-posedness in an anisotropic Sobolev space[[^4]]{} $H^{-\frac{1}{3}+\epsilon,0}_{x,y}$. For the KP-I equation the situation is more delicate. There are several results on local and global existence of solutions, but not a satisfactory well-posedness theory for data with no more than two derivatives in $L^2$. Fokas and Sung [@FS], and Zhou [@Z], obtained global existence for small data via inverse scattering techniques. Schwarz [@Sw] proved existence of weak global periodic solutions with small $L^{2}$ data. The smallness condition was subsequently removed [@C]. Tom [@T] proved existence of global weak solutions for initial data in $H^{1}$ together with their antiderivative. For well-posedness results, we recall the work of Saut [@S], Isaza, Mejía and Stallbohom [@IMS] and finally the work of Iório and Nunes [@IN]. The last two authors use the quasi linear theory of Kato, together with parabolic regularization, to prove local well-posedness with data and their antiderivatives in $H^{s}, s>2$. The limitation $s>2$ is needed in order to insure that $\partial_{x}u\in L^{\infty}$, an essential assumption for the proof. Molinet, Saut and Tzvetkov [@MST1] also proved that if one is willing to assume more regularity for the initial data (at least three derivatives in the $x$ variable and two in the $y$ variable need to be in $L^{2}$), then global well-posedness holds. Recently we [@CKS] were able to obtain well-posedness for small data in a weighted Sobolev space with essentially $H^{2}$ regularity, we will return to this result later.
We recall a few known facts associated with the KP equations. If one defines the Fourier transform for a function $f(x,y)$ as $$\hat{f}(\xi,\mu)=\int_{{{{\mathbb }R}}^{2}}f(x,y)e^{-i(x\xi+y\mu)}dx dy,$$ then it is easy to see that the dispersive function associated to this equation is $${\omega}({\xi},{\mu})=\xi^{3} -\gamma \frac{\mu^{2}}{\xi}.
\label{disp}$$ The analysis of the KP initial value problem depends crucially on the $sign$ of $\gamma$. We describe [*[three differences]{}*]{} due to the choice of sign: the strength of the smoothing effect, the bilinear dispersive identity and (non)positivity of the top order terms in the energy.
A first example of the relevance of the sign of $\gamma$ comes from the following observation. If we compute the gradient of ${\omega}$, we have that for KP-I ($\gamma=-1$, for example) $$|\nabla {\omega}(\xi,\mu)|=|(3{\xi}^{2}-\frac{\mu^{2}}{{\xi}^{2}},2\frac{{\mu}}{{\xi}})|\geq
|\xi|,
\label{grad1}$$ and for KP-II ($\gamma=1$, for example) $$|\nabla {\omega}(\xi,\mu)|=|(3{\xi}^{2}+\frac{\mu^{2}}{{\xi}^{2}},-2\frac{{\mu}}{{\xi}})|\geq
|\xi|^{2}.
\label{grad2}$$ Then, following the argument of Kenig, Ponce and Vega in [@KPV1], we can claim that thanks to KP-II recovers a full derivative smoothness along the $x$ direction, while by KP-I recovers only ${\frac{1}{2}}$ derivative smoothness along that same direction. Because the nonlinear term in (\[ivp\]) presents a derivative along the $x$ direction, this explains, at least formally, why well-posedness questions for the KP-I IVP are much more difficult to answer than for the KP-II problem.
The “sign problem” illustrated above appears also if one approaches well-posedness questions using the method presented by Bourgain in [@B]. This method is based on the strength of various denominators which are controlled using the bilinear dispersive identity $$\begin{aligned}
\nonumber & &{\omega}(\xi_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2})-{\omega}({\xi}_{1},{\mu}_{1})
-{\omega}({\xi}_{2},{\mu}_{2})\\
&=&\frac{{\xi}_{1}{\xi}_{2}}{({\xi}_{1}+{\xi}_{2})}
\left(3({\xi}_{1}+{\xi}_{2})^{2}+\gamma\left(\frac{{\mu}_{1}}{{\xi}_{1}}-
\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}\right).
\label{den}\end{aligned}$$ Clearly if $\gamma<0$ ( KP-I) this quantity could be zero, while if $\gamma>0$ (KP-II) $$|{\omega}(\xi_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2})-{\omega}({\xi}_{1},{\mu}_{1})
-{\omega}({\xi}_{2},{\mu}_{2})|\geq C|{\xi}_{1}||{\xi}_{2}||{\xi}_{1}+{\xi}_{2}|.$$ This is enough to control the derivative in the nonlinear term and to obtain well-posedness results for very rough data (see also Takaoka [@Ta1], Takaoka-Tzvetkov [@TT]).
The IVP (\[ivp\]) has two conserved integrals, the $L^{2}$-norm and the Hamiltonian: $$\begin{aligned}
\label{l2}\|u\|_{L^{2}}&=&\|u_{0}\|_{L^{2}}\\
\label{hamiltonian} H(u)&=&\int_{{{{\mathbb }R}}^{2}}((\partial_{x}u)^{2}-\gamma
(\partial_{x}^{-1}\partial_{y}u)^{2}-\frac{\beta}{3}u^{3})dx dy
=H(u_{0}).\end{aligned}$$ This time, for KP-I, the sign is favorable. In fact one can prove [^5] that a combination of with when $\gamma=-1$, gives $$\|u(t)\|_{L^{2}}+\|\partial_{x}u(t)\|_{L^{2}}+
\|\partial_{x}^{-1}\partial_{y}u(t)\|_{L^{2}}\leq C
\left(\|u_{0}\|_{L^{2}}+\|\partial_{x}u_{0}\|_{L^{2}}+
\|\partial_{x}^{-1}\partial_{y}u_{0}\|_{L^{2}}\right),
\label{en}$$ for any sufficiently smooth solution $u$, uniformly in time. The Sobolev space defined by is naturally called the [*energy space*]{}. It is the natural space on which the Hamiltonian is defined, and thus it would be desirable to obtain a local well-posedness theory for KP-I in this space. (As we mentioned before, Tom [@T] proved the existence of global weak solutions, for data in the energy space, using and compactness arguments, but the uniqueness of these weak solutions remains an open problem). Moreover, if one could also prove that the time $T$ of existence in this (desired) local existence theorem depends only on the norms involved in , then a simple iteration argument, combined with , would yield global in time solutions for data in the energy space, and hence the Hamiltonian would be defined globally in time, for the natural space of initial data, providing a satisfactory “low-regularity” space in which KP-I is globally well-posed, and in which the Hamiltonian is naturally defined. We next remark that this desired dependence on $T$ above is validated by scaling considerations. In fact, if we fix $|\beta| = 1, \gamma = -1$ and $u(x,y,t)$ is a solution of , then $u_\lambda (x, y,t) = {\lambda}^2 u({\lambda}x ,
{\lambda}^2 y , {\lambda}^3 t)$ is also a solution of , with initial data $u_{{\lambda}, 0} (x, y) = {\lambda}^2 u_0 ({\lambda}x , {\lambda}^2 y).$ Note that ${{\| u_{{\lambda}, 0} \|}_{L^2}} = {\lambda}^{\frac{1}{2}}{{\| u_0\|}_{L^2}}$, ${{\|
\partial_x u_{{\lambda}, 0} \|}_{L^2}} = {\lambda}^{\frac{3}{2}} {{\|
\partial_x u_0 \|}_{L^2}}, {{\| \partial_x^{-1} \partial_y u_{{\lambda},
0} \|}_{L^2}} = {\lambda}^{\frac{3}{2}} {{\| \partial_x^{-1}
\partial_y u_{ 0} \|}_{L^2}}$, so that is “sub-critical” in the energy space and thus one expects the time of existence in a local well-posedness theorem, as the one discussed before, to depend only on the norm of the initial data in the energy space. (See also Remark \[rescaling\] for further discussion of the notion of “criticality”). Note also that if one is only interested in global existence of solutions of KP-I, with fairly regular initial data, the recent work [@MST1] provides a very satisfactory global existence theory, by combining the local well-posedness results of Iorio and Nunes [@IN], mentioned before, with higher order conservation laws for KP-I (suitably regularized by the use of Strichartz inequalities).
In the attempt to establish a local well-posedness theory for KP-I in the energy space, one is confronted by the following difficulty, which we have not been able to overcome: so far, in the many studies of local well-posedness for nonlinear dispersive equations, the only successful approach to the issue of “low regularity” data has been through the use of fixed point theorems based on Picard iteration. However, the recent counterexamples of Molinet, Saut and Tzvetkov [@MST], [@MST1], show that, for KP-I, we cannot prove local well-posedness in any type of anisotropic $L^2$-based Sobolev space $H^{s_1, s_2}_{x,y},$ or in the energy space, by using Picard fixed point methods for the integral equation formulation of the KP-I initial value problem. In light of this, to study the local well-posedness theory in the energy space, one must abandon Picard iteration, and proceed in a new way. Since weak solutions have been constructed in [@T], as we mentioned before, the key issue is uniqueness and one needs to establish this without relying on the classical Gronwall inequality, which seems to require too much regularity on the data. Possibly, recent works of Molinet and Ribaud on dissipative generalizations of KP [@MRKP] and KdV [@MRKdV] may prove useful in this direction, but we have not been able to establish the required uniqueness.
Given this unsatisfactory state of affairs, an alternative is to use spaces other than $H^{s_1, s_2}_{x,y}$, or the energy space, but with similar regularity properties, and for which Picard iteration might still work. For example, in our recent work [@CKS], we addressed the well-posedness question for KP-I, by restricting the space of initial data, which we took to consist (essentially) of functions, which together with two derivatives, belonged to the weighted space $L^2((1+|y|)^\alpha dx dy), \alpha > {\frac{1}{2}}$, and have small norm. Our proof relied on the so called “oscillatory integrals” method, which combines local smoothing effects and maximal function estimates.
Our goal in the present paper is to refine the local well-posedness result mentioned above, to reduce the number of the derivatives needed on the initial data to bring it to a space which is close to the energy space we discussed before. We use versions of the spaces and methods introduced by Bourgain [@B], extended to the context of weighted Besov spaces. The weights are used to exploit the fact that the region where is small, is a region of small measure. The estimates we present are sharp, in a sense that will be made clear later, and are obtained in Besov-type spaces involving derivatives of order $1-\epsilon$ and the weight $|y|$. We are able to remove any assumption on the initial data concerning small frequency, but due to the fact that in this case weighted spaces do not rescale well (see Remark \[rescaling\]), our well-posedness result again holds only for small data. From now on we will restrict ourselves to $|\beta| = 1$.
Let’s now define the energy space $E$ and the weighted space $P$ as $$\label{EPdefined}
E=\{f : f\in L^{2}, \partial_{x}f \in L^{2},
\partial_{y}\partial_{x}^{-1}f \in L^{2}\}, \, \mbox{ and }
P=\{f : yf \in L^{2}\}.$$
\[natep\] We consider the space $E \cap P$ natural in the context of KP-I. It was proved by Saut [@S], that for smooth solutions $u$ of KP-I, whose initial data $u_0$ is in $E\cap P$, then for any fixed time interval $[-T,T$, $u$ enjoys the a priori bound $${{\| u \|}_{L^\infty_{[-T, T]} (E \cap P)}} \leq C( T, {{\| u_0
\|}_{E\cap P}} ).$$
Let us denote now by $B_{\rho}$ the ball in $E\cap P$, centered at zero, and radius $\rho$. To state the main theorem we will also need the spaces $(E\cap P)_{1-\epsilon}$, and $Z_{1-\epsilon}$. The first space will be defined in , but for now, all the reader needs to know about it is that it roughly has $\epsilon$ fewer derivatives than the space $E\cap
P$. The space $Z_{1-\epsilon}$ is introduced in . It is a Bourgain type space (following the spaces introduced in [@B]), in which the contraction mapping theorem is applied.
\[energytheorem\] Assume that $\gamma = -1$ in and fix an interval of time $[0,T]$, and a small $\epsilon > 0$. Then, there exists $\delta > 0 $, $\delta = \delta (\epsilon, T)$ such that for any $u_0 \in E\cap P$ with ${{\| u_0 \|}_{E\cap P}} < \delta$, there exists a unique solution $$u \in L^\infty ([0,T]; E\cap P) \cap C([0,T]; (E\cap
P)_{1-\epsilon}) \cap Z_{1-\epsilon}$$ with ${{\| u \|}_{Z_{1-\epsilon}}} \leq C \delta$. Moreover, the map that associates the initial data in $E \cap P$ to the solution $u$ is smooth from the ball $B_\delta$ into the space $C([0,T]; (E\cap P)_{1-\epsilon})$.
This theorem is a consequence of a well-posedness result involving the Besov type spaces of initial data mentioned earlier, (see Theorem \[main0\] below). We start by giving a precise definition for these spaces.
Let $\theta_{0}(s)=\chi_{[-1,1]}(s), \, \, \theta_{m}(s)=
\chi_{[2^{m-1},2^{m}]}(|s|), \, \, m\in {{{\mathbb }N}}$. For $({\xi},{\mu})\in
{{{\mathbb }R}}^{2}$ let $\chi_{1}({\xi},{\mu})=\chi_{\{|{\xi}|\geq {\frac{1}{2}}\frac{|\mu|}{|{\xi}|}\}}$, and $\chi_{2}({\xi},{\mu})=\chi_{\{|{\xi}|< {\frac{1}{2}}{\frac{|\mu|}{|\xi|}}\}}$. We define the space $B^{2,1}_{s}$ of functions on ${{{\mathbb }R}}^{2}$ as the closure of the Schwartz functions, for which the norm below is finite, with respect to $$\begin{aligned}
\label{bs} \|f\|_{B^{2,1}_{s}}&=&\sum_{m\geq 0}\|(1+|{\xi}|+\gi)^{s}
\theta_{m}({\xi})\chi_{1}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}}\\
\nonumber &+&\sum_{n\geq 0}\|(1+|{\xi}|+\gi)^{s}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}}.
\end{aligned}$$ We also define a “weighted Besov space”, $P^{2,1}_{r}$ using the norm $$\begin{aligned}
\label{pr} \|f\|_{P^{2,1}_{r}}&=&\sum_{m\geq 0}\|(1+|{\xi}|+\gi)^{r}
\theta_{m}({\xi})\chi_{1}({\xi},{\mu})\partial_{{\mu}}\hat{f}({\xi},{\mu})\|_{L^{2}}\\
\nonumber &+&\sum_{n\geq 0}\|(1+|{\xi}|+\gi)^{r}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\partial_{{\mu}}\hat{f}({\xi},{\mu})\|_{L^{2}}.
\end{aligned}$$
\[diffregions\] Going back to the discussion of the smoothing effect involving and , one can see that the splitting into the two regions $R_{g}=\{|{\xi}|\geq {\frac{1}{2}}\frac{|\mu|}{|{\xi}|}\}$ and $R_{b}=\{|{\xi}|<{\frac{1}{2}}\frac{|\mu|}{|{\xi}|}\}$ is quite natural. In fact in the “good” region $R_{g}$ it’s easy to check that $|\nabla {\omega}(\xi,\mu)|\gtrsim |{\xi}|^{2}$, hence here one should expect a gain of a full derivative. On the other hand in the “bad” region $R_{b}$ one has $|\nabla {\omega}(\xi,\mu)|\gtrsim |{\xi}|$, and the gain should be only of half derivative, (see also Proposition \[regions\]).
\[embedding\] Because $l^{1}\subset l^{2}$, it follows that $B^{2,1}_{1}\cap
P^{2,1}_{0} \subseteq E\cap P$. Moreover, if $\epsilon>0$, then we also have $E\cap P\subseteq
B^{2,1}_{1-\epsilon}\cap P^{2,1}_{-\epsilon}$. To see this, first assume that $f\in E$. Then $$\begin{aligned}
&& \sum_{m\geq 0}\|(1+|{\xi}|+\gi)^{1-\epsilon}
\theta_{m}({\xi})\chi_{1}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim&
\sum_{m\geq 0}\|2^{-m\epsilon}(1+|{\xi}|)^{\epsilon}(1+|{\xi}|+
\gi)^{1-\epsilon}
\theta_{m}(\xi)\chi_{1}\hat{f}\|_{L^{2}}\\
&\lesssim&\sum_{m\geq 0}2^{-m\epsilon}\|(1+|{\xi}|+\gi)
\theta_{m}(\xi)\chi_{1}\hat{f}\|_{L^{2}},\end{aligned}$$ and Cauchy Schwarz concludes this part. On the other hand $$\begin{aligned}
&& \sum_{n\geq 0}\|(1+|{\xi}|+\gi)^{1-\epsilon}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim& \sum_{n\geq
0}\left\|\frac{(|{\xi}|\gi)^{{\frac{1}{2}}\epsilon}}
{(|{\xi}|\gi)^{{\frac{1}{2}}\epsilon}}(1+|{\xi}|+\gi)^{1-\epsilon}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\hat{f}({\xi},{\mu})\right\|_{L^{2}}\\
&\lesssim& \sum_{n\geq
0}2^{-\frac{n}{2}\epsilon}\|
(|{\xi}|^{\epsilon}+(\gi)^{\epsilon})(1+|{\xi}|+\gi)^{1-\epsilon}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}},\end{aligned}$$ and again Cauchy-Schwarz takes care of this term. Now assume that $f\in
P$. Then $$\begin{aligned}
&& \sum_{m\geq 0}\|(1+|{\xi}|+\gi)^{-\epsilon}
\theta_{m}({\xi})\chi_{1}({\xi},{\mu})\partial_{{\mu}}\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim& \sum_{m\geq 0}2^{-m\epsilon}\|
\theta_{m}({\xi})\chi_{1}({\xi},{\mu})\partial_{{\mu}}\hat{f}({\xi},{\mu})\|_{L^{2}},
\end{aligned}$$ and we use Cauchy-Schwarz. Finally, because $(1+|{\xi}|+\gi)\geq 1+(|{\xi}|\gi)^{{\frac{1}{2}}}$, we obtain $$\begin{aligned}
&&\sum_{n\geq 0}\|(1+|{\xi}|+\gi)^{-\epsilon}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\partial_{{\mu}}\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim& \sum_{n\geq 0}2^{-{\frac{n}{2}}\epsilon}\|
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\partial_{{\mu}}\hat{f}({\xi},{\mu})\|_{L^{2}},
\end{aligned}$$ and also this term is estimated. We are now ready to define the space $(E\cap P)_{1-\epsilon}$, introduced in the statement of Theorem \[energytheorem\], by setting $$\label{epsilon}
(E\cap P)_{1-\epsilon}=B^{2,1}_{1-\epsilon}\cap P^{2,1}_{-\epsilon},$$ for any $\epsilon \in {{{\mathbb }R}}$.
If one could prove well-posedness with initial data in $B^{2,1}_{1-\epsilon}\cap P_{-\epsilon}^{2,1}$, on $[-T,T]$, with $T$ depending only on the norm of the initial data in this space, for some $\epsilon>0$, then for data $u_{0}\in E\cap P$ we would obtain, in light of Remarks \[embedding\] and \[natep\], a unique global solution in $C([-T,T], B^{2,1}_{1-\epsilon}\cap P_{-\epsilon}^{2,1})\cap
L^{\infty}([-T,T], E\cap P)$, for each $T$, which would depend continuously on the initial data, in the $B^{2,1}_{1-\epsilon}\cap P_{-\epsilon}^{2,1}$ topology. However, as we will explain in Remark \[rescaling\], we show the required local well-posedness only for small data in $B^{2,1}_{1-\epsilon}\cap P_{-\epsilon}^{2,1}$, and our estimates barely miss giving the global result[^6].
\[cver12\] If in the definition of $B^{2,1}_{s}$, the constant ${\frac{1}{2}}$ appearing in $\chi_{1}$ and $\chi_{2}$ is replaced by $C$, we obtain the same space, with comparable norms. This holds also for $P^{2,1}_{r}$. Assume that $f\in B^{2,1}_{s}$ and $0<C<{\frac{1}{2}}$. We need to show that $$\label{C12}\sum_{m\geq 0}\|(1+|{\xi}|+\gi)^{s}
\theta_{m}({\xi})\chi_{\{C\gi<|{\xi}|<{\frac{1}{2}}\gi\}}({\xi},{\mu})
\hat{f}({\xi},{\mu})\|_{L^{2}}\lesssim \|f\|_{B^{2,1}_{s}}.$$ But if $C\gi<|{\xi}|<{\frac{1}{2}}\gi$, then $C|{\mu}|<|{\xi}|^{2}<{\frac{1}{2}}|{\mu}|$. If $m=0$, then $|{\xi}|^{2}\leq 1$ and $|{\mu}|\leq C^{-1}$. Then $$\begin{aligned}
&&\|(1+|{\xi}|+\gi)^{s}
\theta_{0}({\xi})\chi_{\{C\gi<|{\xi}|<{\frac{1}{2}}\gi\}}({\xi},{\mu})
\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim&\|(1+|{\xi}|+\gi)^{s}
\chi_{\{|{\mu}|\leq C^{-1}\}}\chi_{\{|{\xi}|<{\frac{1}{2}}\gi\}}({\xi},{\mu})
\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim&C\sum_{n\geq 0}\|(1+|{\xi}|+\gi)^{s}
\theta_{n}({\mu})\chi_{2}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}}.
\end{aligned}$$ If $m\geq 1, \, 2^{2m-1}\leq |{\mu}|\leq 2^{2m}/C$. Let $n_{0}$ be the smallest integer such that $2^{n_{0}}\geq 2^{2m}/C$. Then $n_{0}\leq
2m+C_{0}, C_{0}=C_{0}(C)$. Thus, $$\begin{aligned}
&&\sum_{m\geq 1}\|(1+|{\xi}|+\gi)^{s}
\theta_{m}({\xi})\chi_{\{C\gi<|{\xi}|<{\frac{1}{2}}\gi\}}({\xi},{\mu})
\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim&\sum_{m\geq 1}\sum_{2m-1\leq n_{0}\leq 2m+C_{0}}
\|(1+|{\xi}|+\gi)^{s}
\theta_{n_{0}}({\mu}) \chi_{2}({\xi},{\mu})
\hat{f}({\xi},{\mu})\|_{L^{2}}\\
&\lesssim&\sum_{n_{0}\geq 1}\|(1+|{\xi}|+\gi)^{s}
\theta_{n_{0}}({\mu})\chi_{2}({\xi},{\mu})\hat{f}({\xi},{\mu})\|_{L^{2}},
\end{aligned}$$ and follows. We also need to show that $$\label{C121}\sum_{n\geq 0}\|(1+|{\xi}|+\gi)^{s}
\theta_{n}({\mu})\chi_{\{C\gi<|{\xi}|<{\frac{1}{2}}\gi\}}({\xi},{\mu})
\hat{f}({\xi},{\mu})\|_{L^{2}}\lesssim \|f\|_{B^{2,1}_{s}}$$ and the argument is similar since $$2^{n}\leq |{\mu}|\leq 2^{n+1} \Longrightarrow C^{{\frac{1}{2}}}2^{\frac{n}{2}}\leq
|{\xi}|\leq ({\frac{1}{2}})^{{\frac{1}{2}}}2^{n/2+1}.$$ The case $C>1/2$ is proved in the same way, reversing the role of $C$ and $1/2$. A similar proof can be given for the space $P_{r}^{2,1}$. This remark will be used implicitly in our proofs.
We are now ready to introduce the Banach spaces in which we will perform a fixed point argument to obtain the solution for . Below, we use $\widehat{f}$ to denote the Fourier transform of a function of $(x,y,t)$, defined in a similar fashion as for functions of $(x,y)$. We hope that this will not cause confusion to the reader.
Let $\chi_{0}(s)=\chi_{\{|s|<1\}}(s), \chi_{j}(s)=
\chi_{\{2^{j-1}\leq |s|<2^{j}\}}(s), {\omega}({\xi},{\mu})={\xi}^{3}+\mu^{2}/{\xi}$ and $w({\xi},{\mu})=(1+|{\xi}|+\gi)$. We define the space $X_{s,b}$ through the following norm: $$\begin{aligned}
\label{xs}\|f\|_{X_{s,b}}&=&\sum_{j, m\geq 0}
2^{jb}\left(\int_{{{{\mathbb }R}}^{3}}
\chi_{j}(\tau-{\omega}(\xi,{\mu}))\chi_{1}(\xi,{\mu})\theta_{m}(\xi)
w^{2s}|\hat{f}|^{2}({\xi},{\mu},\tau)d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}}\\
\nonumber&+&\sum_{j, m \geq 0}
2^{jb}\left(\int_{{{{\mathbb }R}}^{3}}
\chi_{j}(\tau-{\omega}(\xi,{\mu}))\chi_{2}(\xi,{\mu})\theta_{n}({\mu})
w^{2s}|\hat{f}|^{2}({\xi},{\mu},\tau)d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}}.
\end{aligned}$$ We also define the space $$\label{yrs}Y_{s,r,b}=\{f : tf \in X_{s,b}, \mbox{ and } yf \in
X_{r,b}\},$$ and the spaces $$\label{zsb}Z_{s,b}=X_{s,b}\cap Y_{s,s-1,b}, \, \,
Z_{1-\epsilon}=Z_{1-\epsilon,{\frac{1}{2}}}.$$ \[spaces\]
A statement similar to Remark \[cver12\] holds for these spaces.
We are now ready to state the well-posedness result for initial data in Besov spaces introduced above.
Assume that $\gamma=-1$ in . For any $\epsilon_{0}<\frac{1}{16}$ and for any interval of time $[0,T]$, there exists $\delta=\delta(\epsilon_{0},T)>0$ such that for any $u_{0}\in B_{1-\epsilon_{0}}^{2,1}\cap
P_{-\epsilon_{0}}^{2,1}$ and $\|u_{0}\|_{B_{1-\epsilon_{0}}^{2,1}\cap P_{-\epsilon_{0}}^{2,1}}\leq
\delta$, there exists a unique solution $u$ for in $Z_{1-\epsilon_0}$. Moreover $u \in
B^{2,1}_{1-\epsilon_0} \cap P^{2,1}_{-\epsilon_0}$ and smoothness with respect to initial data holds in the appropriate topology. \[main0\]
From now on we assume that $\gamma=-1$. In the rest of the paper we often use the notation $A \lesssim B$ if there exists $C>0$ such that $A\leq CB$, and $A\thicksim B$ if $A \lesssim
B$ and $B \lesssim A$ with possibly different $C's.$
The paper is organized in the following way. In Section 2, we introduce some estimates for the solution of the linear KP-I initial value problem. In Section 3, we present two bilinear estimates (see Theorems \[xbil\] and \[ybil\] below) that are the heart of the matter for the proof of Theorem \[main0\]. The section concludes with a counterexample showing the optimality of our analysis. We finish with Section 4, in which we briefly present the proofs of Theorems \[energytheorem\] and \[main0\]. Section 4 also contains a scaling argument which reveals that the optimal analysis in Section 3 is “endpoint critical”.
[**[Acknowledgment.]{}**]{} The authors are very grateful to the referees for their extremely careful reading of the manuscript, and their many suggestions, that have greatly clarified our original version of the paper.
The Linear Estimates
====================
Consider the linear IVP $$\left\{ \begin{array}{l}
\partial_t u + \partial_x^3 u -\partial_x^{-1}\partial_y^2u= 0, \\
u(x,0) = u_0(x) \hspace{1.5cm}(x,y) \in {{{\mathbb }R}}^2, \, t \in {{{\mathbb }R}},
\end{array}\right.
\label{livp}$$ and let $S(t)u_{0}$ be the solution. By taking the Fourier transform of the first equation in and solving the ODE one can easily see that $$S(t)u_{0}(x,y)=\int_{{{{\mathbb }R}}^{2}}e^{i(t{\omega}({\xi},{\mu})+x{\xi}+y{\mu})}
\widehat{u_{0}}({\xi},{\mu}) d{\xi}d{\mu}.$$
We now show that the space $X_{s,{\frac{1}{2}}}\cap Y_{s,s-1,{\frac{1}{2}}}$ is well behaved with respect to the group operator $S(t)$.
([[linear homogeneous estimates]{}]{}) Assume $\psi \in C^{\infty}_{0}(|t|\leq 1), \psi =1$ on $ |t|<{\frac{1}{2}}$. Then $$\begin{aligned}
\label{xsgroup}\|\psi(t) S(t)u_{0}\|_{X_{s,{\frac{1}{2}}}}
&\lesssim&
\|u_{0}\|_{B_{s}^{2,1}},\\
\label{ysgroup}\|\psi(t) S(t)u_{0}\|_{Y_{s,s-1,{\frac{1}{2}}}}&
\lesssim&
\|u_{0}\|_{P^{2,1}_{s-1}}+\|u_{0}\|_{B^{2,1}_{s}}.\end{aligned}$$ \[linear\]
The proof follows the same arguments used in [@KPV3]. We observe that $$(\psi(t) S(t)u_{0})\sphat ({\xi},{\mu},\tau)=\hat{\psi}
(\tau-{\omega}({\xi},{\mu}))\widehat{u_{0}}({\xi},{\mu}).
\label{fpsi}$$ Here $\sphat$ denotes the Fourier transform of a function of 3 variables on the left side of and also to denote the transform of functions of 1 and 2 variables on the right side. Then to prove we need to estimate the two integral expressions: $$\begin{aligned}
\label{xsgroup1}& &\sum_{j\geq 0}2^{\frac{j}{2}}\sum_{m\geq 0}
\left(\int_{{{{\mathbb }R}}^{3}}w({\xi},{\mu})^{2s}\chi_{1}\theta_{m}({\xi})
\chi_{j}(\tau-{\omega})|\hat{\psi}
(\tau-{\omega})|^{2}|\widehat{u_{0}}|^{2}d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}}\\
\label{xsgroup2}& &\sum_{j\geq 0}2^{\frac{j}{2}}\sum_{n\geq 0}
\left(\int_{{{{\mathbb }R}}^{3}}w({\xi},{\mu})^{2s}\chi_{2}\theta_{n}({\mu})
\chi_{j}(\tau-{\omega})|\hat{\psi}
(\tau-{\omega})|^{2}|\widehat{u_{0}}|^{2}d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}},\end{aligned}$$ where $w({\xi},{\mu})=(1+|{\xi}|+\gi)$. We observe that for $j=0$ $$\int_{{{{\mathbb }R}}}|\hat{\psi}(\lambda)|^{2}\chi_{j}(\lambda)
d\lambda\lesssim \|\hat{\psi}\|_{L^{\infty}}^{2}
\label{lambdaj0}$$ and for $j\geq 1$ $$\int_{{{{\mathbb }R}}}|\hat{\psi}(\lambda)|^{2}\chi_{j}(\lambda)
d\lambda\lesssim 2^{j}\frac{1}{(1+2^{j})^{2N}}
\|(1+|s|)^{N}\hat{\psi}(s)\|_{L^{\infty}}^{2}
\label{lambdaj1}$$ for any $N\in {{{\mathbb }N}}$. When we insert and in we obtain the bound $$\|u_{0}\|_{B^{2,1}_{s}}\|\hat{\psi}\|_{L^{\infty}}+\sum_{j\geq 1}
\frac{
2^{j}}{(1+ 2^{j})^{N}}\|u_{0}\|_{B^{2,1}_{s}}\|
(1+|s|)^{N}\hat{\psi}(s)\|_{L^{\infty}}.
\label{pro21}$$ It is easy to see that for $N>1$, $\sum_{j\geq 1}\frac{
2^{j}}{(1+2^{j})^{N}}\leq C$, hence is proved for . A similar argument can be used to estimate .
To estimate we first observe that $$t\psi(t) S(t)u_{0}=\widetilde{\psi}
(t) S(t)u_{0},$$ where $\widetilde{\psi}(t)=t\psi(t)$. Hence by $$\|t\psi(t)
S(t)u_{0}\|_{X_{s,b}}\lesssim \|u_{0}\|_{B^{2,1}_{s}}.$$ We then turn to $y\psi(t)
S(t)u_{0}$. Using the fact that $(yh(y))\sphat=-i\partial_{{\mu}}\hat{h}({\mu})$ and , it is easy to see that $${\mbox{{\em F}}}(y\psi(t)
S(t)u_{0})({\xi},{\mu},\tau)=-2\widehat{\widetilde\psi}(\tau-{\omega})
\frac{{\mu}}{{\xi}} \widehat{u_{0}}({\xi},{\mu})-
\hat{\psi}(\tau-{\omega})\widehat{yu_{0}}({\xi},{\mu}).
\label{ytranformation}$$ Then we can use to conclude that $$\|y\psi(t)
S(t)u_{0}\|_{X_{s-1,b}}\lesssim \|u_{0}\|_{B^{2,1}_{s}}+
\|yu_{0}\|_{B^{2,1}_{s-1}}.$$
There is an inhomogeneous version of Proposition \[linear\].
(linear inhomogeneous estimates) Assume $\psi \in C^{\infty}_{0}(|t|\leq 1), \psi =1$ on $ |t|<{\frac{1}{2}}$. Then, $$\begin{aligned}
\label{xsnonhom}& &\|\psi(t) \int_{0}^{t}S(t-t')h(t')\,dt'
\|_{X_{s,{\frac{1}{2}}}}
\lesssim
\|h\|_{X_{s,-{\frac{1}{2}}}},\\
\label{ysnonhom}& &\|\psi(t)\int_{0}^{t}S(t-t')h(t')
dt'\|_{Y_{s,s-1,{\frac{1}{2}}}}\lesssim (\|h\|_{X_{s,-{\frac{1}{2}}}}+
\|th\|_{X_{s,-{\frac{1}{2}}}}+\|yh\|_{X_{s-1,-{\frac{1}{2}}}}).\end{aligned}$$ \[linearhom\]
Also in this case the proof follows closely the arguments used in [@KPV3]. We start with the following lemma.
(stability under time cutoff) There exists $C>0$ such that for any $s\in {{{\mathbb }R}}$, $$\|\psi(t)f\|_{X_{s,{\frac{1}{2}}}}\leq C
\|f\|_{X_{s,{\frac{1}{2}}}}.$$ \[xsepsilon\]
To prove the lemma we need the auxiliary space $\tilde{X}_{s,b}$ defined as the closure of the functions in $S$ for which the norm below is finite, with respect to the norm $$\|f\|_{\tilde{X}_{s,b}}=
\left(\sum_{j\geq 0}
2^{j2b}\int_{{{{\mathbb }R}}^{3}}
\chi_{j}(\tau-{\omega}({\xi},{\mu}))
(1+|{\xi}|+\gi)^{2s}|\hat{f}|^{2}({\xi},{\mu},\tau)d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}}.$$ Notice that while the space $X_{s,b}$ is defined using an $l^{1}$ summation with respect to $j$, the space $\tilde{X}_{s,b}$ is defined using an $l^{2}$ summation. We have the following lemma.
For any $b$ such that $0<b<{\frac{1}{2}}$ and any $s\in {{{\mathbb }R}}$ we have $$\label{tilex1}\|\psi(t)f\|_{\tilde{X}_{s,b}}\lesssim
\|f\|_{\tilde{X}_{s,b}}.$$ For any $b$ such that ${\frac{1}{2}}<b<1$ and any $s\in {{{\mathbb }R}}$ we have $$\label{tilex2}\|\psi(t)f\|_{\tilde{X}_{s,b}}\lesssim
\|f\|_{\tilde{X}_{s,b}}.$$ \[tilde\]
We start by proving . Note that $$\begin{aligned}
& &\left(\sum_{j\geq 0}
2^{j2b}\int_{{{{\mathbb }R}}^{3}}
\chi_{j}(\tau-{\omega}({\xi},{\mu}))w^{2s}|\hat{f}|^{2}({\xi},{\mu},\tau)d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}}\\
&\simeq &\left(
\int_{{{{\mathbb }R}}^{3}}(1+|\tau-{\omega}({\xi},{\mu})|)^{2b}
w^{2s}|\widehat{f}|^{2}({\xi},{\mu},\tau)d{\xi}d{\mu}d\tau\right)^{{\frac{1}{2}}}\end{aligned}$$ where $w({\xi},{\mu})=(1+|{\xi}|+\gi)$. Because $$(\psi(\cdot)f)\sphat ({\xi},{\mu},\tau)=\hat{f}*_{\tau}
(\hat\psi(\cdot)),$$ by following the arguments in [@KPV3], it is easy to see that the proof reduces to showing that for any $a\in {{{\mathbb }R}}$, (with $(D^b f)\sphat (\tau) = |\tau|^b \widehat{f} (\tau)$) $$\int_{{{{\mathbb }R}}}|D^{b}(e^{iat}f(t)\psi(t))|^{2}dt\lesssim
\int_{{{{\mathbb }R}}}|\hat{f}(l)|^{2}(1+|l-a|)^{2b}dl.
\label{reduction}$$ We use fractional derivatives (see appendix in [@KPV2]) to obtain $$\|D^{b}(e^{iat}f(t)\psi(t))-
D^{b}(e^{iat}f(t))\psi(t)-
e^{iat}f(t)D^{b}(\psi(t))\|_{L^{2}}\leq
\|D^{b}(e^{iat}f(t))\|_{L^{2}}\|\psi\|_{L^{\infty}}.$$ It follows that $$\|D^{b}(e^{iat}f(t)\psi(t))\|_{L^{2}}\leq
\|D^{b}(e^{iat}f(t))\|_{L^{2}}\|\psi\|_{L^{\infty}}+
\|e^{iat}f(t)\|_{L^{2r}}
\|D^{b}(\psi(t))\|_{L^{2r'}},$$ for $\frac{1}{r}+\frac{1}{r'}=1$. Because $b<{\frac{1}{2}}$, if $\frac{1}{r} = 1-2b$, by the Sobolev embedding theorem, we can continue with $$\|D^{b}(e^{iat}f(t)\psi(t))\|_{L^{2}}
\leq \left(\int_{{{{\mathbb }R}}}
|\hat{f}(l)|^{2}(1+|l-a|)^{2b}dl\right)^{{\frac{1}{2}}}
(\|\psi\|_{L^{\infty}}+\|D^{b}(\psi(t))\|_{L^{2r'}}).$$ To finish observe that $$\|D^{b}(\psi(t))\|_{L^{2r'}}<\infty,$$ because $D^{b}(\psi)\in L^{2}\cap L^{\infty}$.
The proof of follows by combining the above arguments with those in Lemma 3.2 of [@KPV3].
Note that also follows from Lemma 2.2 in [@GTV:ZS].
We recall that by real interpolation (see Theorem 5.6.1 in [@BL]), if $A$ is a Banach space and $$l^{q}_{b}(A)=\{(f_{j}) : f_{j}\in A, \, \,
\left(\sum_{j\geq 0}(2^{jb}\|f_{j}\|_{A})^{q}\right)^{\frac{1}{q}}<\infty\},$$ then for any $q_{i}\in [1,\infty]$, $$(l^{q_{0}}_{b_{0}}(A), l^{q_{1}}_{b_{1}}(A))_{\theta,q}=
l^{q}_{b_{\theta}}(A), (b_0 \neq b_1)$$ where $\theta \in [0,1]$, $1\leq q\leq \infty$ and $b_{\theta}=\theta
b_{0}+(1-\theta)b_{1}$. We then apply this fact to the spaces $\tilde{X}_{s,b_{i}}$ with $b_{0}<{\frac{1}{2}}, \, b_{1}>{\frac{1}{2}}, \, q=1$, we use Lemma \[tilde\] and then we sum with respect to $m$ and $n$ to obtain Lemma \[xsepsilon\].
We follow the proof of Lemma 3.3 in [@KPV3]. We write $$\psi(t)\int_{0}^{t}S(t-t')h(t')dt'=I+II,$$ where $$\begin{aligned}
I&=&
\psi(t)\int_{-\infty}^{\infty}\int_{{{{\mathbb }R}}^{2}}
e^{i(x{\xi}+y{\mu})}\hat{h}({\xi},{\mu},\tau)\psi(\tau-{\omega})
\frac{e^{it\tau}-e^{it{\omega}}}{\tau-{\omega}({\xi},{\mu})}d{\xi}d{\mu}d\tau\\
II&=&\psi(t)\int_{-\infty}^{\infty}\int_{{{{\mathbb }R}}^{2}}
e^{i(x{\xi}+y{\mu})}\hat{h}({\xi},{\mu},\tau)[1-\psi(\tau-{\omega})]
\frac{e^{it\tau}-e^{it{\omega}}}{\tau-{\omega}({\xi},{\mu})}d{\xi}d{\mu}d\tau.
\end{aligned}$$ By Taylor expansion we can rewrite $I$ as $$\label{I}
I=\sum_{k=1}^{\infty}\frac{i^{k}}{k!}t^{k}\psi(t)
\int_{{{{\mathbb }R}}^{2}}e^{i(x{\xi}+y{\mu}+t{\omega})}\left(\int_{-\infty}^{\infty}
\hat{h}({\xi},{\mu},\tau)(\tau-{\omega}({\xi},{\mu}))^{k-1}\psi(\tau-{\omega})d\tau\right)
d{\xi}d{\mu}.$$ For $k\geq 1$ let $$t^{k}\psi(t)=\psi_{k}(t),$$ and note that for any $k\geq 1$ and for any $s\in {{{\mathbb }R}}$, $$|\widehat{\psi_{k}}(s)|\leq C,$$ and for $|s|>1$, $$|\widehat{\psi_{k}}(s)|\leq C\frac{(1+k)^{2}}{(1+|s|)^{2}}.
\label{psik}$$ From it is easy to see that $$I=\sum_{k=1}^{\infty}\frac{i^{k}}{k!}
\psi_{k}(t)S(t)h_{k}(x,y),$$ where $$\widehat{h_{k}}({\xi},{\mu})=\int_{-\infty}^{\infty}
\hat{h}({\xi},{\mu},\tau)(\tau-{\omega}({\xi},{\mu}))^{k-1}\psi(\tau-{\omega})d\tau.$$ Then by Proposition \[linear\], in particular , and , we obtain $$\|I\|_{X_{s,{\frac{1}{2}}}}\lesssim \sum_{k\geq 1}\frac{(1+k)^{2}}{k!}
\|h_{k}\|_{B^{2,1}_{s}}.$$ On the other hand it is easy to check that $$\|h_{k}\|_{B^{2,1}_{s}}\lesssim \|h\|_{X_{s,-{\frac{1}{2}}}},$$ which inserted above gives . We now pass to $II$. We write $II=II_{1}+II_{2}$, where $$\begin{aligned}
II_{1}&=&\psi(t)\int_{-\infty}^{\infty}\int_{{{{\mathbb }R}}^{2}}
e^{i(x{\xi}+y{\mu})}\hat{h}({\xi},{\mu},\tau)[1-\psi(\tau-{\omega})]
\frac{e^{it\tau}}{\tau-{\omega}({\xi},{\mu})}d{\xi}d{\mu}d\tau,\\
II_{2}&=&-\psi(t)
\int_{{{{\mathbb }R}}^{2}}
e^{i(x{\xi}+y{\mu})}\int_{-\infty}^{\infty}\hat{h}({\xi},{\mu},\tau)
[1-\psi(\tau-{\omega})]
\frac{e^{it{\omega}}}{\tau-{\omega}({\xi},{\mu})}d\tau d{\xi}d{\mu}.\end{aligned}$$ By Lemma \[xsepsilon\] $$\|II_{1}\|_{X_{s,{\frac{1}{2}}}}\lesssim
\|h\|_{X_{s,-{\frac{1}{2}}}}.$$ On the other hand, by Proposition \[linear\] we have $$\|II_{2}\|_{X_{s,{\frac{1}{2}}}}\lesssim \|\tilde{h}\|_{B^{2,1}_{s}},$$ where $$\widehat{\tilde{h}} (\xi, \mu) =\int_{-\infty}^{\infty}
[1-\psi(\tau-{\omega})]
\frac{\hat{h}({\xi},{\mu},\tau)}{\tau-{\omega}({\xi},{\mu})}d\tau.$$ To finish the proof of one just needs to observe that $$\|\tilde{h}\|_{B^{2,1}_{s}}\lesssim \|h\|_{X_{s,-{\frac{1}{2}}}}.$$ To prove we first observe that by $$\|t\psi(t)\int_{0}^{t}S(t-t')h(t')dt'
\|_{X_{s,{\frac{1}{2}}}}\lesssim
\|h\|_{X_{s,-{\frac{1}{2}}}}.$$ We now use to write $$\begin{aligned}
& &\|y\psi(t)\int_{0}^{t}S(t-t')h(t')dt'
\|_{X_{s-1,{\frac{1}{2}}}}\lesssim
\|\tilde{\psi}(t)\int_{0}^{t}S(t-t')
\bar{h}(t')dt'\|_{X_{s-1,{\frac{1}{2}}}}\\
&+&\|\psi(t)\int_{0}^{t}S(t-t')t'\bar{h}(t')dt'
\|_{X_{s-1,{\frac{1}{2}}}}+\|\psi(t)\int_{0}^{t}S(t-t')yh(t')dt'
\|_{X_{s-1,{\frac{1}{2}}}},\end{aligned}$$ where $\tilde{\psi}(t)=t\psi(t)$ and $
\hat{\bar{h}}({\xi},{\mu})=2\mu/{\xi}\hat{h}({\xi},{\mu})$. We use again and we continue with $$\|y\psi(t)\int_{0}^{t}S(t-t')h(t')dt'
\|_{X_{s-1,{\frac{1}{2}}}}\lesssim\|h\|_{X_{s,-{\frac{1}{2}}}}
+\|th\|_{X_{s,-{\frac{1}{2}}}}
+\|yh\|_{X_{s-1,-{\frac{1}{2}}}}.$$ This concludes the proof of Proposition \[linearhom\].
In this second part of the section we prove some a priori estimates enjoyed by the solution $S(t)u_{0}$ of the linear problem . The first estimate we present is of Strichartz type and is due to Ben-Artzi and Saut [@BS]:
(linear homogeneous estimate) Assume $u_{0} \in L^{2}$, then $$\|S(t)u_{0}(x,y)\|_{L^{4}([0,T],L^{4}({{{\mathbb }R}}^{2}))}\lesssim
\|u_{0}\|_{L^{2}}.
\label{strichartz}$$
We would like to use to obtain an $L^{4}$ estimates for any generic function $f$, not necessarily a linear solution. This can be done by foliating the space ${{{\mathbb }R}}^{3}$ using dyadic level sets $\Lambda_{j}=\{({\xi},{\mu},\tau) / |\tau-{\omega}|\sim 2^{j}\}.$
Let $ \chi_{j}({\xi},{\mu},\tau)=\chi_{j}(\tau-{\omega}({\xi},{\mu}))$ as in Definition \[spaces\]. Then, for $\epsilon > 0$, and with $\spcheck$ denoting the inverse Fourier transform, $$\begin{aligned}
\label{st1}\|(\chi_{j}|\hat{f}|)\spcheck \|_{L^{4}}&\lesssim&
2^{\frac{j}{2}}\|\hat{f}\chi_{j}\|_{L^{2}}\\
\label{st2} \|f\|_{L^{4}}&\lesssim&\left(\sum_{j\geq 0}
2^{(1+\epsilon)j}
\|\hat{f}\chi_{j}\|_{L^{2}}^{2}\right)^{{\frac{1}{2}}}.\end{aligned}$$
Set $(\chi_{j}|\hat{f}|)\spcheck (x,y,t)=g_{j}(x,y,t)$. Then $$g_{j}(x,y,t)=\int_{{{{\mathbb }R}}^{3}}e^{i(x\xi+y{\mu}+t\tau)}|\hat{f}|\chi_{j}
({\xi},{\mu},\tau)
\,d{\xi}d{\mu}d\tau.$$ We now observe that $g_{j}$ can be written as an integral of linear solutions for with appropriate initial data. More precisely, by a simple change of variables one can write $$\begin{aligned}
g_{j}(x,y,t)&=&\int_{{{{\mathbb }R}}^{3}}e^{i(x\xi+y{\mu}+t(\lambda-{\omega}({\xi},{\mu}))}
|\hat{f}|({\xi},{\mu},\lambda-{\omega}({\xi},{\mu}))\chi_{j}(\lambda)\,d{\xi}d{\mu}d\lambda\\
&=&\int_{{{{\mathbb }R}}}e^{it\lambda}\chi_{j}(\lambda)
\left[\int_{{{{\mathbb }R}}^{2}}e^{i(x\xi+y{\mu}-t{\omega}({\xi},{\mu}))}
|\hat{f}|({\xi},{\mu},\lambda-{\omega}({\xi},{\mu}))\,d{\xi}d{\mu}\right] d\lambda\\
&=&\int_{{{{\mathbb }R}}}e^{it\lambda}\chi_{j}(\lambda)S(-t)g_{\lambda}(x,y)
\,d\lambda,
\end{aligned}$$ where $\widehat{g_{\lambda}}({\xi},{\mu})=|\hat{f}|({\xi},{\mu},\lambda-
{\omega}({\xi},{\mu}))$. Then, by $$\|g_{j}\|_{L^{4}}\lesssim \int_{{{{\mathbb }R}}}\chi_{j}(\lambda)
\|S(t)g_{\lambda}(x,y)\|_{L^{4}}\,d\lambda
\lesssim \int_{{{{\mathbb }R}}}\chi_{j}(\lambda)
\|\hat{f}({\xi},{\mu},\lambda-{\omega}({\xi},{\mu}))\|_{L^{2}_{{\xi},{\mu}}}\,d\lambda,$$ and after Cauchy-Schwarz in $\lambda$ is proved. To prove we first foliate the function $f$ over the dyadic levels $|\tau-{\omega}({\xi},{\mu})|\sim
2^{j}$, that is we write $f(x,y,t)=\sum_{j\geq 0}f_{j}$, where $$f_{j}(x,y,t)=\int_{{{{\mathbb }R}}^{3}}e^{i(x\xi+y{\mu}+t\tau)}\hat{f}
\chi_{j}({\xi},{\mu},\tau)
\,d{\xi}d{\mu}d\tau.$$ Then if we proceed as above and we use Minkowski’s inequality, we obtain $$\|f\|_{L^{4}}\lesssim \sum_{j\geq 0}\int_{{{{\mathbb }R}}}\chi_{j}(\lambda)
\|S(t)f_{\lambda}(x,y)\|_{L^{4}}\,d\lambda
\lesssim \sum_{j\geq 0}\int_{{{{\mathbb }R}}}\chi_{j}(\lambda)
\|\hat{f}({\xi},{\mu},\lambda-{\omega}({\xi},{\mu}))\|_{L^{2}_{{\xi},{\mu}}}\,d\lambda$$ where $\widehat{f_{\lambda}}({\xi},{\mu})=|\hat{f}|({\xi},{\mu},
\lambda-{\omega}({\xi},{\mu}))$. Using Cauchy-Schwarz, first in ${\lambda}$, and then in $j$, the proof is concluded.
We also need a smoothing effect estimate and a matching maximal function estimate. We start by defining the operators $P_{+}, P_{-}$ and $P_{0}$ such that $$\begin{aligned}
\widehat{P_{+}(f)}({\xi},{\mu})&=&
\hat{f}({\xi},{\mu})\chi_{\{|{\xi}|>>\frac{|{\mu}|}{|{\xi}|}\}}({\xi},{\mu}),\\
\widehat{P_{-}(f)}({\xi},{\mu})&=&
\hat{f}({\xi},{\mu})\chi_{\{|{\xi}|<<\frac{|{\mu}|}{|{\xi}|}\}}({\xi},{\mu}),\\
\widehat{P_{0}(f)}({\xi},{\mu})&=&
\hat{f}({\xi},{\mu})\chi_{\{|{\xi}|\sim\frac{|{\mu}|}{|{\xi}|}\}}({\xi},{\mu}).\end{aligned}$$ We also recall the operator $D^{s}, \, s\geq 0$ defined through the Fourier transform as $\widehat{D^{s}f}({\xi})=|{\xi}|^{s}\hat{f}$. We then have the following proposition.
\[regions\] (smoothing effect estimates) For any $u_{0}\in L^{2}({{{\mathbb }R}}^{2})$, $$\begin{aligned}
\label{smoothing+-}
\|\partial_{x}S(t)P_{\pm}u_{0}\|_{L^{\infty}_{x}L^{2}_{y}L^{2}_{t}}
&\leq& C\|u_{0}\|_{L^{2}},\\
\label{smoothing0} \|D_{x}^{{\frac{1}{2}}}S(t)P_{0}u_{0}
\|_{L^{\infty}_{y}L^{2}_{x}L^{2}_{t}}
&\leq& C\|u_{0}\|_{L^{2}}.
\end{aligned}$$
The proof follows the argument presented by Kenig, Ponce and Vega in the proof of the one-dimensional KdV smoothing effect in [@KPV2]. To prove we first define the regions of integration $$\begin{aligned}
A_{+}&=&\{|{\xi}|>>\frac{|{\mu}|}{|{\xi}|}\}\\
A_{-}&=&\{|{\xi}|<<\frac{|{\mu}|}{|{\xi}|}\}.\end{aligned}$$ Then we write $$\partial_xS(t)P_{\pm}u_0(x,y)= i\int_{A_{\pm}}
e^{i(x \xi + y \mu )}\xi e^{it{\omega}(\xi,{\mu})}\widehat{u_0}
(\xi,{\mu})\,d\xi \,d{\mu},$$ where ${\omega}({\xi},{\mu})= \xi^3+\frac{{\mu}^2}{\xi}$. We make the change of variables $(\zeta, {\mu}) = (\xi^3+{\mu}^2/\xi,
{\mu})$, and it is to check that if $J({\xi},{\mu})$ represents the jacobian, then in $A_{\pm}$, $|J(\xi,{\mu})|\gtrsim |\xi|^2$ holds. Now assume that ${\xi}=\theta(\zeta,{\mu})$, then the term above equals $$\int_{\R^2} e^{i{\mu}y+it\zeta}[e^{i\theta(\zeta,{\mu})x}
\theta(\zeta,{\mu})
\widehat{u_0}(\theta(\zeta,{\mu}),{\mu})
\chi_{\tilde{A}_{\pm}}|J|^{-1}]\,d\zeta d{\mu},$$ where $\tilde{A}_{\pm}$ is the transformation of $A_{\pm}$ under the given change of variables. Then by Plancherel’s theorem $$\begin{aligned}
\|\partial_xS(t)P_{\pm}u_0(x)\|_{L^2_{y,t}}&=&
\|e^{i\theta(\zeta,{\mu})x}
\theta(\zeta,{\mu})\widehat{u_0}(\theta(\zeta,{\mu}),{\mu})
\chi_{\tilde{A}_{\pm}}|J|^{-1}
\|_{L^2_{\zeta,{\mu}}}\\
&=&\left(\int_{\tilde{A}_{\pm}}
|\theta(\zeta,{\mu})|^2|\widehat{u_0}|^2|J|^{-2}
\,d\zeta d{\mu}\right)^{{\frac{1}{2}}}\\
&\lesssim&\left(\int|\xi|^2|\widehat{u_0}|^2|\xi|^{-2} d\xi d\mu \right)^{{\frac{1}{2}}}
=\|u_0\|_{L^2_{x,y}}.\end{aligned}$$ To prove we use a similar argument. We set $A_{0}=\{|{\xi}|\sim\frac{|{\mu}|}{|{\xi}|}\}$, and we write $$D_x^{1/2}S(t)P_{0}u_0(x,y)=
\int_{A_{0}}
e^{i(\xi,{\mu})(x,y)}|\xi|^{{\frac{1}{2}}} e^{it{\omega}(\xi,{\mu})}
\widehat{u_0}(\xi,{\mu})\,d\xi d{\mu}.$$ We make the change of variables $(\xi,\rho) = (\xi,\xi^3+{\mu}^2/\xi)$, and we observe that this time the estimate for the jacobian is $|J(\xi,{\mu})|\gtrsim |\xi|$. We set ${\mu}=\gamma(\xi,\rho)$ and we continue the chain of inequalities above with $$\int e^{i\xi x+it\rho}[e^{i\gamma(\xi,\rho)y}|\xi|^{{\frac{1}{2}}}
\widehat{u_0}(\xi,\gamma(\xi,\rho))\chi_{\tilde{A}_{0}}
|J|^{-1}]\,d\rho d\xi,$$ where $\tilde{A}_{0}$ is the transformation of $A_{0}$ under the above change of variables. Then by Plancherel’s theorem $$\begin{aligned}
\|D_x^{{\frac{1}{2}}}S(t)P_{0}u_0(y)\|_{L^2_{x,t}}&=&
\|e^{i\gamma(\xi,\rho)y}
|\xi|^{{\frac{1}{2}}}\widehat{u_0}(\xi,\gamma(\xi,\rho))
\chi_{\tilde{A}_{0}\}}|J|^{-1}\|_{L^2_{\xi,\rho}}\\
&=&\left(\int_{\R^2} |\xi||\widehat{u_0}|^2J^{-2}\,d\xi d\rho
\right)^{{\frac{1}{2}}}\\
&\lesssim&\left(\int_{\R^2} |\xi||\widehat{u_0}|^2|\xi|^{-1} d\xi d\mu \right)^{{\frac{1}{2}}}
=\|u_0\|_{L^2_{x,y}}.\end{aligned}$$
Using the argument of foliation with $\chi_{j}$ introduced to prove , one obtains
Let $ \chi_{j}({\xi},{\mu},\tau)=\chi_{j}(\tau-{\omega}({\xi},{\mu}))$ be as in Definition \[spaces\]. Then $$\begin{aligned}
\label{smoothj0}\|(|{\xi}|^{{\frac{1}{2}}}\chi_{\{|{\xi}|\sim \frac{|{\mu}|}{|{\xi}|}\}}
\hat f(\xi,{\mu},\tau)\chi_{j}
(\tau-{\omega}({\xi},{\mu})))\spcheck \|_{L^{\infty}_{y}L^{2}_{x}L^{2}_{t}}&\lesssim&
2^{\frac{j}{2}}\|\hat f\chi_{j}\|_{L^{2}_{{\xi},{\mu},\tau}}\\
\label{smoothj1}\|(|{\xi}|\chi_{\{|{\xi}|>>\frac{|{\mu}|}{|{\xi}|}\}}
\hat f(\xi,{\mu},\tau)\chi_{j}
(\tau-{\omega}({\xi},{\mu})))\spcheck \|_{L^{\infty}_{x}L^{2}_{y}L^{2}_{t}}&\lesssim&
2^{\frac{j}{2}}\|\hat f\chi_{j}\|_{L^{2}_{{\xi},{\mu},\tau}}\\
\label{smoothj2}\|(|{\xi}|\chi_{\{|{\xi}|<<\frac{|{\mu}|}{|{\xi}|}\}}
\hat f(\xi,{\mu},\tau)\chi_{j}
(\tau-{\omega}({\xi},{\mu})))\spcheck \|_{L^{\infty}_{x}L^{2}_{y}L^{2}_{t}}&\lesssim&
2^{\frac{j}{2}}\|\hat f\chi_{j}\|_{L^{2}_{{\xi},{\mu},\tau}}.\end{aligned}$$
On the other hand, using interpolation with the trivial $L^{2}$ norm estimate, we also obtain
Let $ \chi_{j}({\xi},{\mu},\tau)=\chi_{j}(\tau-{\omega}({\xi},{\mu}))$ be as in Definition \[spaces\]. Then $$\begin{aligned}
\label{smoothj5}\|(|{\xi}|^{\frac{1}{4}}\chi_{\{|{\xi}|\sim
\frac{|{\mu}|}{|{\xi}|}\}}
\hat f(\xi,{\mu},\tau)\chi_{j}
(\tau-{\omega}({\xi},{\mu})))\spcheck \|_{L^{4}_{y}L^{2}_{x}L^{2}_{t}}&\lesssim&
2^{\frac{j}{4}}\|\hat f\chi_{j}\|_{L^{2}_{{\xi},{\mu},\tau}}\\
\label{smoothj3}\|(|{\xi}|^{1/2}\chi_{\{|{\xi}|>>\frac{|{\mu}|}{|{\xi}|}\}}
\hat f(\xi,{\mu},\tau)\chi_{j}
(\tau-{\omega}({\xi},{\mu})))\spcheck \|_{L^{4}_{x}L^{2}_{y}L^{2}_{t}}&\lesssim&
2^{j/4}\|\hat f\chi_{j}\|_{L^{2}_{{\xi},{\mu},\tau}}\\
\label{smoothj4}\|(|{\xi}|^{1/2}\chi_{\{|{\xi}|<<\frac{|{\mu}|}{|{\xi}|}\}}
\hat f(\xi,{\mu},\tau)\chi_{j}
(\tau-{\omega}({\xi},{\mu})))\spcheck\|_{L^{4}_{x}L^{2}_{y}L^{2}_{t}}&\lesssim&
2^{j/4}\|\hat f\chi_{j}\|_{L^{2}_{{\xi},{\mu},\tau}}\end{aligned}$$
We finally introduce a maximal function estimate.
(maximal function estimate) Let $T_{m}$ be the operator such that $\widehat{T_{m}(f)}({\xi},{\mu},\tau)=
m({\mu},\tau)\hat{f}({\xi},{\mu},\tau)$. Then $$\|T_{m}(f)\|_{L^{2}_{x}L^{\infty}_{y,t}}\leq
C\|m\|_{L^{2}_{{\mu},\tau}}\|\hat{f}\|_{L^{2}_{{\xi},{\mu},\tau}}.
\label{maxfun}$$ Similarly, if $\widehat{T_{w}(f)}({\xi},{\mu},\tau)=
w({\xi},\tau)\hat{f}({\xi},{\mu},\tau)$. Then $$\|T_{w}(f)\|_{L^{2}_{y}L^{\infty}_{x,t}}\leq
C\|w\|_{L^{2}_{{\xi},\tau}}\|\hat{f}\|_{L^{2}_{{\xi},{\mu},\tau}}.
\label{maxfun1}$$
We only prove . We first write $$T_{m}f(x,y,t)=\int_{{{{\mathbb }R}}^{2}}\check{m}(y-y',t-t')f(x,y',t')
dy'dt'.$$ Then $$|T_{m}f(x,y,t)|\lesssim
\|\check{m}\|_{L^{2}}\|f(x,\cdot,\cdot)\|_{L^{2}}.$$ To end the proof one just has to take the $L^{2}$ norm in the $x$ variable.
It is also useful to observe that interpolating and with the trivial $L^{2}$ estimates, we obtain $$\begin{aligned}
\|T_{m}(f)\|_{L^{2}_{x}L^{4}_{y,t}}&\leq&
C\|m\|_{L^{4}_{{\mu},\tau}}\|\hat{f}\|_{L^{2}_{{\xi},{\mu},\tau}},
\label{maxfunint}\\
\|T_{w}(f)\|_{L^{2}_{y}L^{4}_{x,t}}&\leq&
C\|w\|_{L^{4}_{{\xi},\tau}}\|\hat{f}\|_{L^{2}_{{\xi},{\mu},\tau}}.
\label{maxfunint1}\end{aligned}$$
We end this section with a simple weighted Sobolev inequality that will be useful later.
(weighted Sobolev) Assume $w({\xi},{\mu})\gtrsim 1$ for any $({\xi},{\mu})\in {{{\mathbb }R}}^{2}$. Then for any $\epsilon_{0}\geq 0$, for any $p>2$ and $\theta=(p-2)/2p$, $$\|f\|_{L^{p}_{{\mu}}}\lesssim
\|w(\xi, \cdot)^{\epsilon_{0}}f\|_{L^{2}_{{\mu}}}^{(1-\theta)}
\|w(\xi, \cdot)^{-\epsilon_{0}}f'\|_{L^{2}_{{\mu}}}^{\theta}.
\label{fp}$$ \[sobolev\]
We write $$\begin{aligned}
{\frac{1}{2}}f^{2}({\mu})&=&{\frac{1}{2}}\int_{-\infty}^{{\mu}}
\frac{d}{d{\mu}'}(f^{2}({\mu}'))d{\mu}'\\
&=&
\int_{-\infty}^{{\mu}}f({\mu}')\frac{d}{d{\mu}'}(f({\mu}'))d{\mu}'=
\int_{-\infty}^{{\mu}}w({\xi},{\mu}')^{\epsilon_{0}}f({\mu}'
)w({\xi},{\mu}')^{-\epsilon_{0}}
\frac{d}{d{\mu}'}(f({\mu}'))d{\mu}'\\
&\leq&\|w({\xi},\cdot)^{\epsilon_{0}}f\|_{L^{2}_{{\mu}}}
\|w({\xi},\cdot)^{-\epsilon_{0}}f'\|_{L^{2}_{{\mu}}}.\end{aligned}$$ From here it follows that $$\|f\|_{L^{\infty}}\leq
\|w({\xi},\cdot)^{\epsilon_{0}}
f\|_{L^{2}_{{\mu}}}^{{\frac{1}{2}}}
\|w({\xi},\cdot)^{-\epsilon_{0}}f'\|_{L^{2}_{{\mu}}}^{{\frac{1}{2}}}.
\label{finfty}$$ If $p>2$ $$\int_\R |f|^{p}d{\mu}=\int_\R |f|^{p-2}|f|^{2}d{\mu}\leq\|f\|_{L^{\infty}}^{p-2}\|f\|_{L^{2}}^{2},$$ and by it follows that $$\|f\|_{L^{p}_{{\mu}}}\leq \|f\|_{L^{2}_{{\mu}}}^{2/p}
\|w({\xi},\cdot)^{\epsilon_{0}}
f\|_{L^{2}_{{\mu}}}^{(p-2)/2p}
\|w({\xi},\cdot)^{-\epsilon_{0}}f'\|_{L^{2}_{{\mu}}}^{(p-2)/2p}$$ and because $w({\xi},{\mu})\gtrsim 1$ $$\|f\|_{L^{p}_{{\mu}}}\leq \|w({\xi},\cdot)^{\epsilon_{0}}
f\|_{L^{2}_{{\mu}}}^{(p+2)/2p}
\|w({\xi},\cdot)^{-\epsilon_{0}}f'\|_{L^{2}_{{\mu}}}^{(p-2)/2p},$$ and the lemma is proved.
The Bilinear Estimates
======================
As announced at the end of Section 1, the core of the well-posedness result we present in this paper is contained in the following two theorems.
Assume $0 <\epsilon_{0}<\frac{1}{8}$. Then for any $\frac{1}{4}<\epsilon<1$, we have $$\begin{aligned}
\label{Xbilinear}
\|\partial_{x}(uv)\|_{X_{1-\epsilon_{0},-{\frac{1}{2}}}}&\leq& C
\|u\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}(\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}+
\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}^{1-\epsilon }
\|v\|_{Y_{1-\epsilon_{0},-\epsilon_{0}, {\frac{1}{2}}}}^{\epsilon })\\
&+&C
\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}(\|u\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}+
\|u\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}^{1-\epsilon }
\|u\|_{Y_{1-\epsilon_{0},-\epsilon_{0}, {\frac{1}{2}}}}^{\epsilon }).\end{aligned}$$ \[xbil\]
The companion of the above bilinear estimate is
Assume $0 <\epsilon_{0}<\frac{1}{16}$. Then for any $\frac{1}{4}
<\epsilon<1$, we have $$\begin{aligned}
\label{Ybilinear}
\|\partial_{x}(uv)\|_{Y_{1-\epsilon_{0},-\epsilon_{0},-{\frac{1}{2}}}}&\leq&
C
\|u\|_{Y_{1-\epsilon_{0},-\epsilon_{0}, {\frac{1}{2}}}}
(\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}+
\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}^{1-\epsilon }
\|v\|_{Y_{1-\epsilon_{0},-\epsilon_{0}, {\frac{1}{2}}}}^{\epsilon })\\
&+&C
\|v\|_{Y_{1-\epsilon_{0},-\epsilon_{0}, {\frac{1}{2}}}}
(\|u\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}+
\|u\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}^{1- \epsilon }
\|u\|_{Y_{1-\epsilon_{0},-\epsilon_{0}, {\frac{1}{2}}}}^{\epsilon }).\end{aligned}$$ \[ybil\]
We have not attempted to find the optimal value for $\epsilon_{0}$ for which our argument can be carried out.
An estimate for the bilinear expression $D_{x}^{1/2}(uv)$ in spaces not involving weights already appeared in [@MST].
To give an idea of how the proof will be conducted we write the left hand side of the bilinear inequality in Theorem \[xbil\] using duality.
We have to estimate, for $g_j \geq 0$, $$\begin{aligned}
& &\label{duality-1}\sum_{j\geq 0}2^{-j/2}\sum_{m\geq 0}
\int_{A*} g_{j}({\xi},{\mu},\tau)\theta_{m}({\xi})
\chi_{1}({\xi},{\mu})\chi_{j}(\tau-{\omega}({\xi},{\mu}))\\
\nonumber
& &\times |{\xi}|\max(1,|{\xi}|)^{1-\epsilon_{0}}
|\hat{u}|({\xi}_{1},{\mu}_{1},\tau_{1})
|\hat{v}|({\xi}_{2},{\mu}_{2},\tau_{2})d{\xi}_{1}d{\mu}_{1}
d\tau_1d{\xi}_2d{\mu}_{2}d\tau_{2}\end{aligned}$$ and $$\begin{aligned}
\label{duality0}& &\sum_{j\geq 0}2^{-j/2}\sum_{n\geq 0}
\int_{A*}
g_{j}({\xi},{\mu},\tau)\theta_{n}({\mu})
\chi_{2}({\xi},{\mu})\chi_{j}(\tau-{\omega}({\xi},{\mu}))\\
\nonumber
& &\times |{\xi}|\max(1, \gi)^{1-\epsilon_{0}}
|\hat{u}|({\xi}_{1},{\mu}_{1},\tau_{1})
|\hat{v}|({\xi}_{2},{\mu}_{2},\tau_{2})d{\xi}_{1}d{\mu}_{1}
d\tau_{1}d{\xi}_{2}d{\mu}_{2}d\tau_{2},\end{aligned}$$ where $ A*$ is the set $\{\xi_{1}+{\xi}_{2}={\xi},
{\mu}_{1}+{\mu}_{2}=\mu, \tau_{1}+\tau_{2}=\tau \}$ and $\|g_{j}\theta_{m}\chi_{1}\chi_{j}
\|_{L^{2}_{{\xi},{\mu},\tau}}\leq 1 $ and $\|g_{j}\theta_{n}\chi_{2}\chi_{j}
\|_{L^{2}_{{\xi},{\mu},\tau}}\leq 1 $. It is clear that by symmetry one can always assume that $|{\xi}_{1}|\geq |{\xi}_{2}|$. Based on Remark \[diffregions\] one can easily understand that many different cases need to be considered in view of the fact that there will be a combination of interactions between “good” regions of type $R_{g}$, bad regions of type $R_{b}$, regions with relatively small or large frequencies. The whole analysis is complicated further by the fact that the spaces we use are anisotropic. We start by subdividing $A*$ into six domains of integration $$\begin{aligned}
\label{a1}&& A_{1}= A*\cap \{|{\xi}_{1}|\geq |{\xi}_{2}|, |{\xi}_1|\leq
1\},\\
\label{a2}&& A_{2}=A*\cap \{|{\xi}_{1}|\geq |{\xi}_{2}|, |{\xi}_1|>1,
|{\xi}_{2}|\sim |{\xi}_{1}|\},\\
\label{a3}&& A_{3}= A*\cap \{|{\xi}_1|>1,
1<|{\xi}_{2}|\leq 10^{-10} |{\xi}_{1}|, |{\mu}_{2}|/|{\xi}_{2}|\gtrsim
\max{\{|{\xi}_{1}|, |{\mu}_{1}|/|{\xi}_{1}|\}}\},\\
\label{a4}&& A_{4}= A*\cap B\cap \{|{\xi}_1|>1, |{\xi}_{2}|\leq 1,
|{\xi}_{2}|\leq 10^{-10} |{\xi}_{1}|, |{\mu}_{2}|/|{\xi}_{2}|\gtrsim
\max{\{|{\xi}_{1}|, |{\mu}_{1}|/|{\xi}_{1}|\}}\},\\
\label{a5}&& A_{5}= A*\cap B^{c}\cap \{|{\xi}_1|>1, |{\xi}_{2}|\leq 1,
|{\xi}_{2}|\leq 10^{-10}|{\xi}_{1}|, |{\mu}_{2}|/|{\xi}_{2}|\gtrsim
\max{\{|{\xi}_{1}|, |{\mu}_{1}|/|{\xi}_{1}|\}}\},\\
\label{a6}&& A_{6}= A*\cap \{|{\xi}_1|>1,
|{\xi}_{2}|\leq 10^{-10} |{\xi}_{1}|, |{\mu}_{2}|/|{\xi}_{2}|\leq 10^{-10}
\max{\{ |{\xi}_{1}|, |{\mu}_{1}|/|{\xi}_{1}|\}}\},
\end{aligned}$$ where $B=\{|{\xi}_{1}|\geq \frac{1}{100}|{\mu}_{1}|/|{\xi}_{1}|\}$. We also use the auxiliary region $$\tilde{A}_{5}(\epsilon_{0})=
\{|{\mu}_{1}/{\xi}_{1}-{\mu}_{2}/{\xi}_{2}|^{2}<3/2
|{\xi}_{1}+{\xi}_{2}|^{2}, |\xi_{1}|^{1+\alpha\epsilon_{0}}
\leq |{\mu}_{1}/{\xi}_{1}|\},$$ where $\alpha$ will depend on $\epsilon_{0}$. The most delicate part of our estimate occurs in the region $A_{6}$ and it is only here that we need the weighted spaces and the Besov type norms. We start with a lemma.
If $\epsilon_{0}<\frac{1}{8}$, then $$\begin{aligned}
\label{lemma1}& &\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq
0}2^{-j/2}\int_{A}g_{j}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2},
\tau_{1}+\tau_{2})
\chi_{j}(\tau_{1}+\tau_{2}-{\omega}({\xi}_{1}+{\xi}_{2},
{\mu}_{1}+{\mu}_{2}))\\
\nonumber& &|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}
\Pi_{i=1,2}\phi_{i}({\xi}_{i},{\mu}_{i},\tau_{i})
\chi_{j_{i}}
(\tau_{i}-{\omega}({\xi}_{i},{\mu}_{i}))d{\xi}_{i}d{\mu}_{i}
d\tau_{i}\\
\nonumber&\lesssim&\sup_{j}\|g_{j}\|_{L^{2}}
\Pi_{i=1,2}\sum_{j_{i}\geq 0}
2^{j_{i}/2}\|\phi_{i}\chi_{j_{i}}\|_{L^{2}},
\end{aligned}$$ where $A=A_{k}, k=1,2,3,4$ or $A=A_{5}-
\widetilde{A}_{5}(\epsilon_{0})$. \[mainlemma\]
The proof that we present below gives a little more general result than the one stated. In particular we show that holds in sets larger than $A_{3}$ and $A_{4}$, namely in $$\tilde A_{3}=A*\cap \{|{\xi}_1|>1,
1<|{\xi}_{2}|\leq 10^{-10} |{\xi}_{1}|\},$$ and $$\tilde A_{4}=A*\cap B\cap \{|{\xi}_1|>1, |{\xi}_{2}|\leq 1,
|{\xi}_{2}|\leq 10^{-10} |{\xi}_{1}|\}.$$
For simplicity, for $i=1,2$, we set $$\phi_{i}({\xi}_{i},{\mu}_{i},\tau_{i})\chi_{j_{i}}
(\tau_{i}-{\omega}({\xi}_{i},{\mu}_{i}))=\phi_{i,j_{i}}({\xi}_{i},{\mu}_{i},
\tau_{i}).$$ Also, whenever we use a dyadic decomposition either with respect to $|{\xi}_{i}|\sim 2^{m_{i}}$ or $|{\mu}_{i}|\sim 2^{n_{i}}$, we write $$\begin{aligned}
\label{pmj}\phi_{i,j_{i},m_{i}}({\xi}_{i},{\mu}_{i},\tau_{i})&=&
\phi_{i,j_{i}}
\chi_{\{|{\xi}_{i}|\sim 2^{m_{i}}\}}({\xi}_{i},{\mu}_{i},\tau_{i})\\
\label{pnj}\phi_{i,j_{i},n_{i}}({\xi}_{i},{\mu}_{i},\tau_{i})
&=&\phi_{i,j_{i}}
\chi_{\{|{\mu}_{i}|\sim 2^{n_{i}}\}}({\xi}_{i},{\mu}_{i},\tau_{i}).\end{aligned}$$ We prove the theorem by analyzing the integral in on the different regions.
[**Region $A_{1}$:**]{} Here $|{\xi}_{1}+{\xi}_{2}|\lesssim 1$ and we can simply use the Strichartz inequality .
[**Region $A_{2}$:**]{}. Here we can assume also that $|{\xi}_{1}+{\xi}_{2}|>1$, otherwise we go back to the argument used in the region $A_{1}$. We dyadically decompose with respect to $|{\xi}_{1}|\sim 2^{m_{1}}$ (hence $|{\xi}_{2}|\sim 2^{m_{1}}$) and we rewrite the left hand side of as $$\sum_{m_{1}\geq 0}\sum_{j,j_{1},j_{2}}2^{-j/2}\int g_{j}\chi_{j}
|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\Pi_{i=1,2}\phi_{i,j_{i},m_{i}}
d\tau_{i}d{\xi}_{i}d{\mu}_{i}
\label{a2step1}$$ We now consider two cases:
[**Case A:**]{} $j>2\epsilon_{0}m_{1}$. We use the Strichartz inequality and can be bounded by $$\sum_{m_{1}\geq 0}\sum_{j>2\epsilon_{0}m_{1}}\sum_{j_{1},j_{2}\geq 0}
2^{-j/2+m_{1}\epsilon_{0}}
\|g_{j}\chi_{j}\|_{L^{2}}\Pi_{i=1,2}2^{j_{i}/2}
\|\phi_{i,j_{i},m_{i}}\|_{L^{2}}$$ and follows in this case.
[**Case B:**]{} $0\leq j\leq 2\epsilon_{0}m_{1}$. We change variable in $\tau_{1}$ and $\tau_{2}$ by setting $(\tau_{i}-{\omega}({\xi}_{i},{\mu}_{i}))=\theta_{i}$ and we write the left hand side of as $$\begin{aligned}
\label{a2step2}& &\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq
0}2^{-j/2}\int_{A}g_{j}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2},\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))\\
\nonumber & &\chi_{j}(\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})
-{\omega}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2}))\\
\nonumber& &|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}
\Pi_{i=1,2}\phi_{i}({\xi}_{i},{\mu}_{i},\theta_{i}+{\omega}({\xi}_{i},{\mu}_{i}))
\chi_{j_{i}}(\theta_{i})d{\xi}_{i}d{\mu}_{i}d\theta_{i}.
\end{aligned}$$ From we also have that $$\label{den1}1+ \left|\theta_{1}+\theta_{2}+\frac{{\xi}_{1}
{\xi}_{2}}{({\xi}_{1}+{\xi}_{2})}
\left(\left(\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}
{{\xi}_{2}}\right)^{2}
-3(\xi_{1}+\xi_{2})^{2}\right)\right|\sim 2^{j}.$$ [**Case B1:**]{} $\left|\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right|^{2}
\leq 3/2|\xi_{1}+\xi_{2}|^{2}.$\
Then from and it follows that $|{\xi}_{1}||{\xi}_{2}||{\xi}_{1}+{\xi}_{2}|
\lesssim
2^{\max{(j_{1},j_{2},j)}}$. Because now $|{\xi}_{1}+{\xi}_{2}|>1$, if $j=\max{(j_{1},j_{2},j)}$, then $2^{2m_{1}}\lesssim 2^{2\epsilon_{0}m_{1}}$, a contradiction if $\epsilon_{0}<1$ for $m_1$ large enough. So by symmetry we can assume that $|{\xi}_{1}+{\xi}_{2}|
\lesssim 2^{(j_{1}-2m_{1})}$, hence $|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\lesssim
2^{\epsilon_{0}(j_{1}-2m_{1})}$. We can then continue[^7] the estimate of by using Strichartz inequality with $$\sum_{m_{1}\geq 0}\sum_{j\leq 2\epsilon_{0}m_{1}}
\sum_{j_{1},j_{2}\geq 0}2^{-2\epsilon_{0}m_{1}}
2^{\epsilon_{0}j_{1}}
\|g_{j}\chi_{j}\|_{L^{2}}
\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}2^{j_{2}/2}
\|\phi_{2,j_{2},m_{1}}\|_{L^{2}}
\label{stg}$$ and after Cauchy-Schwarz in $m_{1}$ this gives provided $\epsilon_{0}<{\frac{1}{2}}$.
[**Case B2:**]{} $\left|\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right|^{2}
> 3/2|\xi_{1}+\xi_{2}|^{2}.$\
Here we consider the following change of variables $$\begin{array}{l}
u = {\xi}_{1}+{\xi}_{2}\\
v = {\mu}_{1}+{\mu}_{2}\\
w = {\omega}({\xi}_{1},{\mu}_{1})+{\omega}({\xi}_{2},{\mu}_{2})+\theta_{1}+\theta_{2}\\
{\mu}_{2}={\mu}_{2}.
\end{array}
\label{chanm}$$ The Jacobian associated to this change of variable is $$J_{{\mu}}=3({\xi}_{1}^{2}-{\xi}_{2}^{2})-
\left(\left(\frac{|{\mu}_{1}|}{|{\xi}_{1}|}\right)^{2}-
\left(\frac{|{\mu}_{2}|}{|{\xi}_{2}|}\right)^{2}\right).
\label{jm}$$ We observe that, for fixed $\theta_{1},
\theta_{2}, {\xi}_{1}, \xi_{2},
{\mu}_{1}$, the set where the free variable $\mu_{2}$ can range so that is verified is a union of two symmetric intervals and the length of these intervals is small. More precisely, if we denote with $\Delta_{\mu_{2}}$ this length, then $$\Delta_{{\mu}_{2}}\lesssim 2^{j-m_{1}}.
\label{deltam}$$ To see this we introduce the function $$f({\mu})=\theta_{1}+\theta_{2}+\frac{{\xi}_{1}{\xi}_{2}}{({\xi}_{1}+{\xi}_{2})}
\left(\left(\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}}{{\xi}_{2}}\right)^{2}
-3(\xi_{1}+\xi_{2})^{2}\right).$$ It’s easy to see that $|f'({\mu})|\gtrsim |{\xi}_{1}|$, hence follows. We now consider two subcases.
[**Case B2a:**]{} $\left|3({\xi}_{1}^{2}-\xi_{2}^{2})-\left(\left(\frac{{\mu}_{1}}
{{\xi}_{1}}\right)^{2}-
\left(\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}\right)\right|> 1$.\
We make the change of variable (now $|J_{{\mu}}|>1$). Denote with $H(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})$ the transformation of $$\Pi_{i=1,2}\phi_{i, j_i, m_1 }({\xi}_{i},{\mu}_{i},\theta_{i}+
{\omega}({\xi}_{i},{\mu}_{i}))
\chi_{j_{i}}(\theta_{i})$$ under the above change of variables. Then becomes $$\begin{aligned}
& &\sum_{m_{1},j_{1},j_{2}\geq 0}\sum_{0\leq j\leq
2\epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}\int g_{j}\chi_{j}(u,v,w)\\
& &|J_{{\mu}}|^{-1}H(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})dudvdw
d{\mu}_{2}d\theta_{1}d\theta_{2}\\
&\lesssim &\sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq 2\epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}2^{(j-m_{1})/2}\\
&&\int g_{j} \chi_{j}(u,v,w)\left(\int_{{\mu}_{2}}|J_{{\mu}}|^{-2}
H^{2}(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})d{\mu}_{2}
\right)^{{\frac{1}{2}}}dudvdw
d\theta_{1}d\theta_{2}.
\end{aligned}$$ Now we observe that by Cauchy-Schwarz and the inverse change of variable we have $$\begin{aligned}
& &\int g_{j} \chi_{j}(u,v,w)\left(\int_{{\mu}_{2}}|J_{{\mu}}|^{-2}
H^{2}(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})d{\mu}_{2}
\right)^{{\frac{1}{2}}}dudvdw
d\theta_{1}d\theta_{2}\\
&\leq& {{\| g_j \chi_j \|}_{L^2}} \left( \int \left( \int |J_\mu|^{-2}
H^2 ( u ,v,w,\mu_2, \theta_1, \theta_2 ) du dv dw d\mu_2
\right)^{\frac{1}{2}}d\theta_1 d\theta_2 \right) \\
&\lesssim & {{\| g_j \chi_j \|}_{L^2}} \left( \int \left( \int |J_\mu|^{-1}
H^2 ( u ,v,w,\mu_2, \theta_1, \theta_2 ) du dv dw d\mu_2
\right)^{\frac{1}{2}}d\theta_1 d\theta_2 \right) \\
&\lesssim & {{\| g_j \chi_j \|}_{L^2}}
\left( \int \left( \int \Pi_{i=1,2} \phi_{i, j_i, m_1}^2 (\xi_i ,
\mu_i , \theta_i + \omega(\xi_i , \mu_i ) ) \chi_{j_i} (\theta_i ) d
\xi_1 d\xi_2 d\mu_1 d\mu_2 \right)^{\frac{1}{2}}d\theta_1 d\theta_1 \right)
\\
&\lesssim & {{\| g_j \chi_j \|}_{L^2}} 2^{j_1 /2 } 2^{j_2 /2} \left(
\int \int \Pi_{i=1,2} \phi_{i, j_i , m_1}^2 ( \xi_i , \mu_i ,
\theta_i + \omega( \xi_i , \mu_i )) d\xi_1 d\xi_2 d\mu_1 d\mu_2
d\theta_1 d\theta_2 \right)^{\frac{1}{2}}\\
&\lesssim& \|g_{j}\chi_{j}\|_{L^{2}}
\Pi_{i=1,2}2^{j_{i}/2}\|\phi_{i,j_{i},m_{1}}\|_{L^{2}}
\end{aligned}$$ which inserted above, after a sum on $j$ gives $$\sum_{m_{1},j_{1},j_{2}\geq 0}
[1+ (2\epsilon_{0}m_{1})]
2^{\epsilon_{0}m_{1}}2^{-m_{1}/2}
\|g_{j}\chi_{j}\|_{L^{2}}
\Pi_{i=1,2}2^{j_{i}/2 }\|\phi_{i,j_{1},m_{i}}\|_{L^{2}}$$ and from here we obtain again for $\epsilon_{0}<{\frac{1}{2}}$.
[**Case B2b:**]{} $\left|3({\xi}_{1}^{2}-\xi_{2}^{2})-\left(\left(\frac{{\mu}_{1}}
{{\xi}_{1}}\right)^{2}-
\left(\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}\right)\right|\leq 1$.\
In this case the change of variables above cannot be used because the Jacobian may become zero. We consider instead the change of variables in which we leave ${\xi}_{1}$ free: $$\begin{array}{l}
u = {\xi}_{1}+{\xi}_{2}\\
v = {\mu}_{1}+{\mu}_{2}\\
w = {\omega}({\xi}_{1},{\mu}_{1})+{\omega}({\xi}_{2},{\mu}_{2})+\theta_{1}+\theta_{2}\\
{\xi}_{1}={\xi}_{1}.
\end{array}
\label{chanx}$$ In this case the Jacobian $J_{{\xi}}$ is given by $$J_{{\xi}}=
\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}},
\label{jx}$$ and because we are in Case B2, it follows that $$|J_{{\xi}}|\gtrsim |{\xi}_{1}+{\xi}_{2}|>1.$$ We observe that, for fixed $\theta_{1}, \theta_{2}, \xi_{2},
{\mu}_{1}, {\mu}_{2}$, the set where the free variable ${\xi}_{1}$ can range so that we remain in Case B is a union of two symmetric intervals and the length of these interval is small. More precisely, if we denote with $\Delta_{{\xi}_{1}}$ this length, then $$\Delta_{{\xi}_{1}}\lesssim 2^{-m_{1}}.
\label{deltax}$$ To see this we introduce the function $$h({\xi})=3({\xi}^{2}-\xi_{2}^{2})-\left(\left(\frac{{\mu}_{1}}
{{\xi}}\right)^{2}-
\left(\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}\right).$$ We compute $$h'({\xi})=6{\xi}-2({\mu}_{1}/{\xi})(-{\mu}_{1}/{\xi}^{2})=6{\xi}+
2({\mu}_{1}/{\xi})^{2}{\xi}^{-1}$$ and we notice that $h'({\xi})$ has the same sign as ${\xi}$, hence $|h'({\xi})|\gtrsim |{\xi}|$, and follows. Again denote with $H(u,v,w,{\xi}_{1},\theta_{1},\theta_{2})$ the transformation of[^8] $\Pi_{i=1,2}\phi_{i,j_{i}}({\xi}_{i},{\mu}_{i},\theta_{i})$ under the change of variables . Then becomes $$\begin{aligned}
& &\sum_{m_{1},j_{1},j_{2}\geq 0}\sum_{0\leq j\leq
2\epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}\int g_{j}\chi_{j}(u,v,w)\\
& &\times |J_{{\xi}}|^{-1}H(u,v,w,{\xi}_{1},\theta_{1},
\theta_{2})dudvdw
d{\xi}_{1}d\theta_{1}d\theta_{2}\\
&\lesssim &\sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq 2\epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}2^{-m_{1}/2}\\
&&\int g_{j} \chi_{j}(u,v,w)\left(\int_{{\xi}_{1}}|J_{{\xi}}|^{-2}
H^{2}(u,v,w,{\xi}_{1},\theta_{1},\theta_{2})d{\xi}_{1}
\right)^{{\frac{1}{2}}}dudvdw
d\theta_{1}d\theta_{2}\\
&\lesssim & \sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq 2\epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}2^{-m_{1}/2}\Pi_{i=1,2}2^{j_{i}/2}
\|\phi_{i,j_{i},m_{i}}\|_{L^{2}}\end{aligned}$$ and this again gives for $\epsilon_{0}<{\frac{1}{2}}$.
[**Region $A_{3}$:**]{} In this region, (see ), $|{\xi}_{1}+{\xi}_{2}|\sim|{\xi}_{1}|$. We dyadically decompose with respect to $|{\xi}_{1}|\sim 2^{m_{1}}$ (hence $|{\xi}_{1}+{\xi}_{2}|\sim 2^{m_{1}}$). The left hand side of now becomes $$\sum_{m_{1}\geq 0}\sum_{j,j_{1},j_{2}\geq 0}2^{-j/2}2^{m_{1}\epsilon_{0}}
\int g_{j,m_{1}}\chi_{j}\phi_{1,j_{1},m_{1}}\phi_{2,j_{1}} d\xi_i
d\mu_i d\tau_i
\label{a3step1}$$ where the arguments of the functions as in . We consider two subcases. For $0 < \delta \ll 1$, to be fixed, we have:\
[**Case A:**]{} $j>(2+2\delta)\epsilon_{0}m_{1}, 0<\delta<<1$. We use Strichartz inequality and we obtain $$\begin{aligned}
\eqref{a3step1}&\lesssim&
\sum_{m_{1}\geq 0}\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 2(1+\delta)\epsilon_{0}m_{1}}2^{-j/2}2^{m_{1}\epsilon_{0}}
\|g_{j,m_{1}}\chi_{j}\|_{L^{2}}2^{j_{1}/2}2^{j_{2}/2}
\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}\|\phi_{2,j_{2}}\|_{L^{2}}\\
&\lesssim&
\sum_{m_{1}\geq 0}\sum_{j_{1},j_{2}\geq 1}2^{-m_{1}\delta\epsilon_{0}}
\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}}2^{j_{1}/2}2^{j_{2}/2}
\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}\|\phi_{2,j_{2}}\|_{L^{2}}.\end{aligned}$$ [**Case B:**]{} $j\leq (2+2\delta)\epsilon_{0}m_{1}$.\
[**Case B1:**]{} $\left|\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right|^{2}
\leq 3/2|\xi_{1}+\xi_{2}|^{2}.$\
As in region $A_{2}$, it follows that $|{\xi}_{1}||{\xi}_{2}|
|{\xi}_{1}+{\xi}_{2}|\lesssim
2^{\max{(j_{1},j_{2},j)}}$ and since $|{\xi}_{2}|>1$ and $|{\xi}_{1}|\sim |{\xi}_{1}+{\xi}_{2}|$, we obtain that $|{\xi}_{1}|^{2}\lesssim
2^{\max{(j_{1},j_{2},j)}}$. If $j=\max{(j_{1},j_{2},j)}$, then $2^{2m_{1}}\lesssim 2^{(2+2\delta)\epsilon_{0}m_{1}}$, a contradiction if $\epsilon_{0}<\frac{1}{1+\delta}$ for $m_1$ large enough. Assume then that $j_{1}=
\max{(j_{1},j_{2},j)}$. It follows that $|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\lesssim 2^{\epsilon_{0}(j_{1}-m_{1})}$ and thanks to Strichartz[^9] inequality we can continue the chain of inequalities with $$\begin{aligned}
& &\sum_{m_{1}\geq 0}\sum_{j_{1},j_{2}\geq 0}
\sum_{j\leq 2(1+\delta)\epsilon_{0}m_{1}}2^{-j/2}
\|g_{j,m_{1}}\chi_{j}\|_{L^{2}}2^{j/2} 2^{\epsilon_{0}(j_{1}-m_{1})}
\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2} \|\phi_{2,j_{2}}\|_{L^{2}}\\
& \lesssim &\sum_{m_{1}\geq 0}\sum_{j_{1},j_{2}\geq 0}
2^{\epsilon_{0}(j_{1}-m_{1})}[(2+2\delta) \epsilon_0 m_1 +1] 2^{j_2 /2}
(\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}})
\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}},\end{aligned}$$ and this gives provided $\epsilon_{0}<{\frac{1}{2}}$. Clearly the case when $j_{2}=\max{(j_{1},j_{2},j)}$ is similar.
[**Case B2:**]{} $\left|\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}
{{\xi}_{2}}\right|^{2}
> 3/2|\xi_{1}+\xi_{2}|^{2}.$\
As we did for region $A_{2}$ here also we consider two subcases.
[**Case B2a:**]{} $\left|3({\xi}_{1}^{2}-\xi_{2}^{2})-\left(\left(\frac{{\mu}_{1}}
{{\xi}_{1}}\right)^{2}-
\left(\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}\right)\right|> 1$.\
We make the change of variable , for which now $|J_{{\mu}}|>1$ and we observe that also in this region holds. Then can be bounded by $$\begin{aligned}
& &\sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq (2+2\delta) \epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}2^{(j-m_{1})/2} \|g_{j}
\chi_{j}\|_{L^{2}}
2^{j_{1}/2}2^{j_{2}/2}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}\\
&\lesssim& \sum_{m_{1},j_{1},j_{2}\geq 0}
2^{(\epsilon_{0}-{\frac{1}{2}})m_{1}}((2+2\delta) \epsilon_{0}m_{1})
(\sup_{j} \| g_{j}\chi_{j}\|_{L^{2}})
2^{j_{1}/2}2^{j_{2}/2}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}
\end{aligned}$$ which again gives for $\epsilon_{0}<{\frac{1}{2}}$.
[**Case B2b:**]{} $\left|3({\xi}_{1}^{2}-\xi_{2}^{2})-\left(\left(\frac{{\mu}_{1}}
{{\xi}_{1}}\right)^{2}-
\left(\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}\right)\right|\leq 1$.\
We consider now the change of variables and we observe that $$\label{jacx}|J_{{\xi}}|=\left|
\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right|
\gtrsim |{\xi}_{1}+{\xi}_{2}|\sim 2^{m_{1}}.$$ We also remark that in this region holds too. Repeating the argument in Case B2b of region $A_{2}$ the left hand side of can be bounded by $$\sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq (2+2\delta) \epsilon_{0}m_{1}}
2^{-j/2}2^{\epsilon_{0}m_{1}}2^{-m_{1}/2}
\|g_{j}\chi_{j}\|_{L^{2}}
2^{j_{1}/2}2^{j_{2}/2}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}},$$ and this concludes the estimate in $A_{3}$ for $\epsilon_{0}<{\frac{1}{2}}$.
[**Region $A_{4}$**]{}. Notice that we only need to restrict the proof to the case when $0\leq j\leq (2+2\delta)\epsilon_{0}m_{1}$ and $\left|\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right|^{2}
\leq 3/2|\xi_{1}+\xi_{2}|^{2}$, since in the other situations (Case A in $A_3$ and case $B_2$ in $A_3$) and we didn’t use the assumption $|{\xi}_{2}|\geq 1$. Observe that by the above restriction we also have that $$\gib\lesssim \max(\gia, |{\xi}_{1}|)
\label{m2simx1}$$ We consider two cases.\
[**Case A:**]{} $|{\xi}_{2}||{\xi}_{1}|^{1+\alpha\epsilon_{0}}\geq 1$, for some $\alpha>0$ to be determined later.\
Going back to the argument presented in Case B1 in region $A_{3}$, we obtain $|{\xi}_{1}|^{1-\alpha\epsilon_{0}}\lesssim 2^{\max(j_{1},j_{2},j)}$. If we let $$\epsilon_{0}<\frac{1}{2+2\delta+\alpha},
\label{epsi}$$ then $\max(j_{1},j_{2},j)=\max(j_{1},j_{2})$. Let’s assume $\max(j_{1},j_{2})=j_{1}$ and $\theta>0$ and small. Then $2^{\epsilon_{0}m_{1}}\lesssim
2^{j_{1}\delta_{0}}2^{-\sigma_{0}m_{1}}$, where $$\delta_{0}=\frac{\epsilon_{0}}{(1-\alpha\epsilon_{0})(1-\theta)}, \,
\mbox{ and } \, \, \sigma_{0}=\frac{\epsilon_{0}\theta}{1-\theta}.$$ Notice that if $0<\theta<<1$, from it follows that $\delta_{0}\leq {\frac{1}{2}}$. We then use Strichartz and we bound the left hand side of with $$\begin{aligned}
&& \sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq (2+2\delta) \epsilon_{0}m_{1}}
2^{-j/2}2^{\delta_{0}j_{1}}2^{-\sigma_{0}m_{1}}
2^{j/2}(\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}})
2^{j_{2}/2}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}\\
&\lesssim&\sum_{m_{1},j_{1},j_{2}\geq 0}
2^{\delta_{0}j_{1}}2^{-\sigma_{0}m_{1}}
(2+2\delta) \epsilon_{0}m_{1}(\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}})
2^{j_{2}}/2\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}.\end{aligned}$$ The result is given by the Cauchy-Schwarz inequality in $m_{1}$. The argument when $\max(j_{1},j_{2})=j_{2}$ is similar.\
[**Case B:**]{} $|{\xi}_{2}||{\xi}_{1}|^{1+\alpha\epsilon_{0}}<1$.\
From and the definition of $A_{4}$ in , we also have $|{\mu}_{2}|\leq |{\xi}_{2}||{\xi}_{1}|\lesssim
|{\xi}_{2}||{\xi}_{1}|^{1+\alpha\epsilon_{0}}\lesssim 1$. We consider two subcases.\
[**Case B1:**]{} $\gia\leq \frac{1}{100} |{\xi}_{1}|$.\
Here we use smoothing effect and the maximal function inequalities. Let’s start by assuming that $|{\mu}_{2}|^{2}/|{\xi}_{2}|\lesssim 2^{j_{2}}$. Then $|{\omega}({\xi}_{2},{\mu}_{2})|, |\tau_{2}|, \lesssim 2^{j_{2}}, |{\mu}_{2}|
\lesssim 1$. We now set $m({\mu}_{2},\tau_{2})=\chi({\mu}_{2},\tau_{2})$, the characteristic function of the projection of the region of integration onto the ${\mu}_{2}-\tau_{2}$ plane. Then $$\|m\|_{L^{4}_{{\mu}_{2},\tau_{2}}}\sim
2^{j_{2}/4}.$$ We then use Plancherel, Hölder with the three spaces $L^{4}_{x,y,t}-L^{4}_{x}L^{2}_{y}L^{2}_{t}-L^{2}_{x}L^{4}_{y}L^{4}_{t}$ and inequalities and to bound with $$\begin{aligned}
&& \sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq (2+2\delta) \epsilon_{0}m_{1}}2^{-j/2}
2^{(\epsilon_{0}-{\frac{1}{2}})m_{1}}2^{j/2}\|g_{j}\chi_{j}\|_{L^{2}}
2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/4} \|\phi_{2,j_{2}}\|_{L^{2}}\\
&\lesssim&\sum_{m_{1},j_{1},j_{2}\geq 0}
2^{(\epsilon_{0}-{\frac{1}{2}})m_{1}}
[1+(2+2\delta)\epsilon_0 m_1] (\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}})
2^{j_{2}/4}2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}\end{aligned}$$ and if $\epsilon_{0}<{\frac{1}{2}}$ the lemma is proved also in this case. If $|{\mu}_{2}|^{2}/|{\xi}_{2}|>>2^{j_{2}}$, then $
|{\omega}({\xi}_{2},{\mu}_{2})|\sim |{\mu}_{2}|^{2}/|{\xi}_{2}|,
|\tau_{2}|\sim |{\mu}_{2}|^{2}/|{\xi}_{2}|$ and $|\tau_{2}-{\omega}({\xi}_{2},{\mu}_{2})|\sim 2^{j_{2}}$. From we have $$|{\mu}_{2}|\gib\lesssim |{\xi}_{2}||{\xi}_{1}|^{2}\lesssim |{\xi}_{2}|
|{\xi}_{1}|^{1+\alpha\epsilon_{0}}|{\xi}_{1}|^{1-\alpha\epsilon_{0}}\lesssim
1\cdot|{\xi}_{1}|^{1-\alpha\epsilon_{0}}.
\label{m2x}$$ In this case $$\|m\|_{L^{4}_{{\mu}_{2},\tau_{2}}}\sim
2^{m_{1}(1-\alpha\epsilon_{0})/4)}.$$ We then use again Plancherel, Hölder with the three spaces $L^{4}_{x,y,t}-L^{4}_{x}L^{2}_{y}L^{2}_{t}-L^{2}_{x}L^{4}_{y}L^{4}_{t}$ and inequalities and to bound the left hand side of with $$\begin{aligned}
&& \sum_{m_{1},j_{1},j_{2}\geq 0}
\sum_{0\leq j\leq (2+2\delta) \epsilon_{0}m_{1}}
2^{(\epsilon_{0}-{\frac{1}{2}})m_{1}}\|g_{j}\chi_{j}\|_{L^{2}}
2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{m_{1}(1-\alpha\epsilon_{0})/4}
\|\phi_{2,j_{2}}\|_{L^{2}}\\
&\lesssim&\sum_{m_{1},j_{1},j_{2}\geq 0}
[1+(2+2\delta) \epsilon_0 m_1] (\sup_{j}\|g_{j}
\chi_{j}\|_{L^{2}})
2^{(\epsilon_{0}-{\frac{1}{2}})m_{1}}2^{m_{1}
(1-\alpha\epsilon_{0})/4)}
2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}\end{aligned}$$ and if $\alpha>4$, for any $\epsilon_{0}$, the sum with respect to $m_{1}$ can be done and then is proved also in this case.\
[**Case B2:**]{} $\frac{1}{100}|{\xi}_{1}|\leq \gia \leq 100 |{\xi}_{1}|$.\
Here we use . Let’s start by assuming that $|{\mu}_{2}|^{2}/|{\xi}_{2}|\leq 2^{j_{2}}$. Clearly now $|\tau_{2}| \lesssim 2^{j_{2}}$ and $|{\xi}_{2}|\lesssim 1$. We use Plancherel, Hölder with the three spaces $L^{4}_{x,y,t}-L^{4}_{y}L^{2}_{x}L^{2}_{t}-L^{2}_{y}L^{4}_{x}L^{4}_{t}$ and inequalities and to bound the left hand side of with $$\sum_{m_{1},j_{1},j_{2}\geq 0}
[1+(2+2\delta) \epsilon_0 m_1 ](\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}})
2^{(\epsilon_{0}-\frac{1}{4})m_{1}}2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/4} \|\phi_{2,j_{2}}\|_{L^{2}}$$ and if $\epsilon_{0}<\frac{1}{4}$ then is proved also in this case. Assume now that $|{\mu}_{2}|^{2}/|{\xi}_{2}|>>2^{j_{2}}$. Then by $|\tau_{2}|\lesssim |{\mu}_{2}|^{2}/|{\xi}_{2}|\lesssim
|{\xi}_{1}|^{1-\alpha\epsilon_{0}}$, so that $\|w\|_{L^{2}_{{\xi}_{2},\tau_{2}}}\lesssim
2^{m_{1}(1-\alpha\epsilon_{0})/4}$. We use Plancherel, Hölder with the three spaces $L^{4}_{x,y,t}-L^{4}_{y}L^{2}_{x}L^{2}_{t}-L^{2}_{y}L^{4}_{x}L^{4}_{t}$ and inequalities and to bound the left hand side of with $$\sum_{m_{1},j_{1},j_{2}\geq 0}
[1+(2+2\delta) \epsilon_0 m_1] (\sup_{j}\|g_{j}\chi_{j}\|_{L^{2}})
2^{(\epsilon_{0}-\frac{1}{4})m_{1}}2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
(2^{j_{2}/4} +2^{m_{1}(1-\alpha\epsilon_{0})/4})
\|\phi_{2,j_{2}}\|_{L^{2}}$$ and if $\alpha>4$, then the lemma follows in this case too.
[**Region $A_{5}-\tilde{A}_{5}(\epsilon_{0})$**]{}. Also in this case we can assume that $0\leq j\leq (2+2\delta)\epsilon_{0}m_{1}$, and $\left|\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right|^{2}
\leq 3/2|\xi_{1}+\xi_{2}|^{2}$, since in the other situation we didn’t use the assumption $|{\xi}_{2}|\geq 1$. Also notice that here too holds, hence $\gia\sim \gib$. Because we are in $A_{5}-\tilde{A}_{5}(\epsilon_{0})$, to these restrictions we have to add $|{\xi}_{1}|^{1+\alpha\epsilon}>>\gia$, which in turn gives $$|{\mu}_{2}|<<|{\xi}_{1}|^{1+\alpha\epsilon_0}|{\xi}_{2}|.
\label{m21}$$ We consider two subcases.\
[**Case A:**]{} $|{\xi}_{1}|^{1+\alpha\epsilon_0}|{\xi}_{2}|\geq 1$.\
This case is identical to Case A in region $A_{4}$.\
[**Case B:**]{} $|{\xi}_{1}|^{1+\alpha\epsilon_0}|{\xi}_{2}|< 1$.\
By we now have that $|{\mu}_{2}|\lesssim 1$. If $|{\mu}_{2}|^{2}/|{\xi}_{2}|\leq
2^{j_{2}}$ we can use the same arguments presented in the first part of Case B1 in region $A_{4}$, by replacing with . If $|{\mu}_{2}|^{2}/|{\xi}_{2}|>>
2^{j_{2}}$, we have that $|\tau_{2}|\sim |{\mu}_{2}|^{2}/|{\xi}_{2}|$ and $|\tau_{2}-{\omega}({\xi}_{2},{\mu}_{2})|\sim 2^{j_{2}}$. We estimate $$|{\mu}_{2}||{\mu}_{2}|/|{\xi}_{2}|\lesssim |{\xi}_{2}|(|{\mu}_{1}|/|{\xi}_{1}|)^{2}
\lesssim |{\xi}_{2}||{\xi}_{1}|^{1+\alpha\epsilon_{0}}
|{\xi}_{1}|^{1+\alpha\epsilon_{0}}\lesssim 1\cdot
|{\xi}_{1}|^{1+\alpha\epsilon_{0}}$$ and from here on we can proceed like in the second part of Case B1 in region $A_{4}$, again by replacing with . We then obtain: $$\sum_{m_1, j_1, j_2 \geq 0} \sum_{0 \leq j \leq
(2+2\delta) \epsilon_0 m_1} ( \sup_j {{\| g_j \chi_j
\|}_{L^2}} ) 2^{(\epsilon_0 - {\frac{1}{2}}) m_1} 2^{m_1 (1 + \alpha
\epsilon_0) /4} 2^{j_1 /4} {{\| \phi_{1, j_1 , m_1 } \|}_{L^2}}
{{\| \phi_{2, j_2} \|}_{L^2}},$$ and if $(\epsilon_0 - {\frac{1}{2}}) +
(1+\alpha \epsilon_0 )/4 <0 $, or $\epsilon_0 (1+ \alpha /4 ) <
\frac{1}{4},$ we can sum in $m_1$. This is always possible for $\epsilon_0 < \frac{1}{8}$, with some $\alpha > 4$. The analysis of this case concludes the proof of the lemma.
We are now ready to prove Theorem \[xbil\]. To make the presentation more clear we summarize below the main cases considered in our analysis [^10].
- Region $A_{1}$
- Case A: $|{\xi}|\geq {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Case B: $|{\xi}|< {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Case B1: $|{\mu}_{1}|\leq |{\mu}_{2}|$
- Case B2: $|{\mu}_{1}|> |{\mu}_{2}|$
- Region $A_{2}$
- Case A: $|{\xi}|\geq {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Case B: $|{\xi}|< {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Case B1: $|{\mu}_{1}|\leq |{\mu}_{2}|$
- Case B2: $|{\mu}_{1}|> |{\mu}_{2}|$
- Region $A_{3}\cup A_{4}\cup (A_{5}-\tilde{A}_{5}(\epsilon_{0}))$
- Case A: $|{\xi}|\geq {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Case A1: $|{\xi}_{1}|\geq {\frac{1}{2}}\gia$
- Case A2: $|{\xi}_{1}|\leq {\frac{1}{2}}\gia$
- Case B: $|{\xi}|< 1/2|{\mu}|/|{\xi}|$
- Case B1: $|{\mu}_{1}|\leq |{\mu}_{2}|$
- Case B2: $|{\mu}_{1}|> |{\mu}_{2}|$
- Region $A_{5}\cap \tilde{A}_{5}(\epsilon_{0})$
- Case A: $|{\xi}|\geq {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Case B: $|{\xi}|< {\frac{1}{2}}|{\mu}|/|{\xi}|$
- Region $A_{6}$
- Case A: $|{\xi}_{1}|\geq 10^{2}\gia$
- Case B: $|{\xi}_{1}|\leq 10^{-2}\gia$
- Case C: $10^{-2}\gia \leq |{\xi}_{1}|\leq 10^{2}\gia.$
We reexpress the left hand side of in Theorem \[xbil\] using duality and we obtain and . We analyze these expressions on the regions described in -. For the estimates in the regions $A_{1}$ through $A_{5}$ we find it convenient to normalize the functions $u$ and $v$ so that the expression in the right hand side of the bilinear inequality involves only $L^{2}$ norms. So define $$\begin{aligned}
\label{normalizeu} \phi_{1,j_{1}}({\xi},{\mu},\tau)&=&
\max(1,|{\xi}|,|{\mu}|/|{\xi}|)^{1-\epsilon_{0}}|\hat{u}|
\chi_{j_{1}}({\xi},{\mu},\tau),\\
\label{normalizev}
\phi_{2,j_{2}}({\xi},{\mu},\tau)&=&\max(1,|{\xi}|,|{\mu}|/|{\xi}|)^{1-\epsilon_{0}}
|\hat{v}|
\chi_{j_{2}}({\xi},{\mu},\tau).\end{aligned}$$ If we use the identities ${\mu}={\mu}_{1}+{\mu}_{2}$ and ${\xi}={\xi}_{1}+{\xi}_{2}, \tau = \tau_1 + \tau_2$, and , we can rewrite the left hand side of as $$\begin{aligned}
\label{duality1}&&\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{m\geq 0}2^{-j/2}
\int_{A^*} g_j ( \xi , \mu , \tau) \chi_1 ( \xi_1 , \mu) \chi_j ( \tau -
\omega (\xi , \mu))
\theta_{m}({\xi}) |{\xi}_{1}+{\xi}_{2}|\\
\nonumber&&\max(1,
|{\xi}_{1}+{\xi}_{2}|)^{1-\epsilon_{0}}\frac{\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{\max(1, |{\xi}_{1}|,\gia)^{1-\epsilon_{0}}}
\frac{\phi_{2,j_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{\max(1, |{\xi}_{2}|,\gib)^{1-\epsilon_{0}}} d\xi_1 d\xi_2 d\mu_1 d\mu_2
d\tau_1 d\tau_2\end{aligned}$$ and the left hand side of as $$\begin{aligned}
\label{duality2}&&\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{n\geq 0}2^{-j/2}
\int_{A^*} g_j ( \xi , \mu , \tau) \chi_1 ( \xi_1 , \mu) \chi_j ( \tau -
\omega (\xi , \mu))
\theta_{n}({\mu}) |{\xi}_{1}+{\xi}_{2}|\\
\nonumber&&\max(1,
|{\mu}_{1}+{\mu}_{2}|/|{\xi}_{1}+{\xi}_{2}|)^{1-\epsilon_{0}}
\frac{\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{\max(1, |{\xi}_{1}|,\gia)^{1-\epsilon_{0}}}
\frac{\phi_{2,j_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{\max(1, |{\xi}_{2}|,\gib)^{1-\epsilon_{0}}} d\xi_1 d\xi_2 d\mu_1 d\mu_2 .\end{aligned}$$ Below, case A will alway correspond to the estimate , while case B will correspond to the estimate for . Note that in what follows, $\chi_1$ and $\chi_2$ will always denote the functions introduced in Definition 1, and $\chi_j$ the ones defined in Definition 2. .\
[**Case A:**]{} $|{\xi}|\geq {\frac{1}{2}}|{\mu}|/|{\xi}|$.\
Note that here $|{\xi}_{1}+{\xi}_{2}|\leq 2$ so the sum in $m$ is finite and we can simply use the Strichartz inequality and the fact that $l^{1}\subset l^{2}$, to obtain $$\eqref{duality1}\lesssim \sup_{m,j}\|g_{j}\chi_{1}\chi_{j}
\theta_{m}\|_{L^{2}}\sum_{j_{1},j_{2}\geq 0}\Pi_{i=1,2}2^{j_{i}/2}
\|\phi_{i,j_{i}}\|_{L^{2}}$$ and the theorem follows in this case.
[**Case B:**]{} $|{\xi}|< {\frac{1}{2}}|{\mu}|/|{\xi}|$.\
If $\gi\leq 1$, it follows that $|{\xi}|\leq {\frac{1}{2}}$ and hence $|{\mu}|\leq {\frac{1}{2}}$, that is the sum on $n$ in reduces to a finite sum and we proceed as above using the Strichartz inequality . So we can assume that $\gi>1$. In this case reduces to $$\begin{aligned}
\label{a1mstep1}
&&\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{n\geq 0}2^{-j/2}
\int_{A_{1}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n}({\mu}) |{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}
|{\mu}_{1}+{\mu}_{2}|^{1-\epsilon_{0}}\\
\nonumber
&&\frac{\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{\max(1, |{\xi}_{1}|,\gia)^{1-\epsilon_{0}}}
\frac{\phi_{2,j_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{\max(1, |{\xi}_{2}|,\gib)^{1-\epsilon_{0}}}.\end{aligned}$$ We then consider two subcases.
[**Case B1:**]{} $|{\mu}_{1}|\leq |{\mu}_{2}|$.\
If ${\frac{1}{2}}\gib\leq |{\xi}_{2}|$ then $|{\mu}_{2}|\leq 2, |{\mu}_{1}+{\mu}_{2}|\leq 4$. Again the sum on $n$ reduces to a finite sum and because still $|{\xi}_{1}+{\xi}_{2}|\leq 1$ we go back to the previous case. If ${\frac{1}{2}}\gib> |{\xi}_{2}|$ we introduce a dyadic decomposition with respect to ${\mu}_{2}$ and we set $|{\mu}_{2}|\sim 2^{n_{2}}$. Then $|{\mu}_{1}+{\mu}_{2}|\leq C|{\mu}_{2}|
\leq C2^{n_{2}}$ and we can write $1+ |{\mu}_{1}+{\mu}_{2}|\sim
2^{n_{2}+1-r}, \, 0\leq r\leq n_{2}$. We can bound with $$\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{n_{2}\geq 0}\sum_{0\leq r
\leq n_{2}}
2^{-j/2}\int_{A_{1}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{2}+1-r}2^{(n_{2}+1-r)(1-\epsilon_{0})}\phi_{1,j_{1}}({\xi}_
{1},{\mu}_{1},\tau_{1})
\frac{\chi_{2}\phi_{2,j_{2},n_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{(\gib)^{1-\epsilon_{0}}}.$$ We now use the fact that $|{\xi}_{2}|\leq 2$ and again the Strichartz inequality to continue with $$\sum_{j,j_{1},j_{2}\geq 0}\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}
2^{-j/2}\|g_{j}\chi_{2}\chi_{j}
\theta_{n_{2}+1-r}\|_{L^{2}}2^{-r(1-\epsilon_{0})}2^{j_{1}/2}
\|\phi_{1,j_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\chi_{2}\phi_{2,j_{2},n_{2}}\|_{L^{2}}$$ and this is enough to prove the theorem in this case.\
[**Case B2:**]{} $|{\mu}_{1}|\geq |{\mu}_{2}|$.\
This case can be treated like Case B1 by replacing the role of $({\xi}_{2},{\mu}_{2})$ by $({\xi}_{1},{\mu}_{1})$.
[**Region $A_{2}$**]{}.
[**Case A:**]{} $|{\xi}|\geq {\frac{1}{2}}|{\mu}|/|{\xi}|$.\
Here becomes $$\begin{aligned}
\label{a2x}
&&\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{m\geq 0}2^{-j/2}
\int_{A_{2}}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m}({\xi}) |{\xi}_{1}+{\xi}_{2}|^{2-\epsilon_{0}}\\
\nonumber
&&\frac{\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{\max(1, |{\xi}_{1}|,\gia)^{1-\epsilon_{0}}}
\frac{\phi_{2,j_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{\max(1, |{\xi}_{2}|,\gib)^{1-\epsilon_{0}}}.\end{aligned}$$ We dyadically decompose $|{\xi}_{1}|\sim|{\xi}_{2}|\sim 2^{m_{1}}$. Then $1+ |{\xi}|\sim 2^{m}$, with $m=m_{1}+1-r, 0\leq r\leq m_{1}$ and $$\begin{aligned}
\eqref{a2x}&\lesssim &\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{m_1 \geq 0}\sum_{0\leq r \leq m_{1}}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}+1-r}({\xi}) |{\xi}_{1}+{\xi}_{2}|^{2-\epsilon_{0}}\\
\nonumber
&&\frac{\phi_{1,j_{1},m_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{2^{m_{1}(1-\epsilon_{0})}}
\frac{\phi_{2,j_{2},m_{1}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{2^{m_{1}(1-\epsilon_{0})}}\\
\nonumber &\lesssim&\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{m_1 \geq 0}
\sum_{0\leq r \leq m_{1}}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}+1-r}({\xi}) 2^{-r(2-2\epsilon_{0})}\\
& &|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\Pi_{i=1,2}\phi_{i,j_{1},m_{i}}
({\xi}_{i},{\mu}_{i},\tau_{i}).
\end{aligned}$$ We apply Lemma \[mainlemma\] relative to the region $A_{2}$ and we continue with $$\begin{aligned}
&\lesssim&\sup_{j,m}\|g_{j}\chi_{1}\chi_{j}\theta_{m}\|_{L^{2}}
\sum_{m_1 \geq 0}\sum_{0\leq r \leq m_{1}}2^{-r(2-2\epsilon_{0})}
\Pi_{i=1,2}\sum_{j_{i}\geq 0}2^{j_{i}/2}\|\phi_{i,j_{i},m_{1}}\|_{L^{2}}\\
&\lesssim&\sum_{j_{1},j_{2}\geq 0}2^{j_{1}/2}2^{j_{2}/2}
\left(\sum_{m_{1}\geq 0}\|\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\|\phi_{2,j_{2},m_{1}}\|_{L^{2}}\right) \\
&\lesssim& \sum_{j_1 , j_2 \geq 0} 2^{j_1 /2 } 2^{j_2 /2} \left(
\sum_{m_1 \geq 0} {{\| \phi_{1, j_1 , m_1} \|}^2_{L^2}}
\right)^{\frac{1}{2}}\left(
\sum_{m_1 \geq 0} {{\| \phi_{2, j_2 , m_1} \|}^2_{L^2}}
\right)^{\frac{1}{2}}\\
&\lesssim &
\left(
\sum_{j_1 \geq 0} 2^{j_1 /2} {{\| \phi_{1, j_1 } \|}^2_{L^2}}
\right)^{\frac{1}{2}}\left(
\sum_{j_2 \geq 0} 2^{j_2 /2} {{\| \phi_{2, j_2 } \|}^2_{L^2}}
\right)^{\frac{1}{2}},\end{aligned}$$ which concludes the argument. [**Case B:**]{} $|{\xi}|< {\frac{1}{2}}|{\mu}|/|{\xi}|$.\
If $\gi\leq 1$ it follows that $|{\xi}|\leq {\frac{1}{2}}$ and hence $|{\mu}|\leq {\frac{1}{2}}$ and we go back to the same estimates presented for region $A_{1}$. So we can assume $\gi\geq 1$. We have to estimate where now the integral takes place in $A_{2}$. Again we consider two subcases.\
[**Case B1:**]{} $|{\mu}_{1}|\leq |{\mu}_{2}|$.\
If ${\frac{1}{2}}\gib\leq |{\xi}_{2}|$ then $|{\mu}_{2}|\leq 2|{\xi}_{2}|^{2}, |{\mu}_{1}+{\mu}_{2}|\leq 4|{\xi}_{2}|^{2}$. We dyadically decompose with respect to $|{\xi}_{1}|\sim |{\xi}_{2}|\sim 2^{m_{2}}$, so that $1+ |{\mu}_{1}+{\mu}_{2}|\sim 2^{-r+2m_{2}}, 0\leq r\leq 2m_{2}.$ Then in this case we bound with $$\begin{aligned}
&&\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{m_2 \geq 0}
\sum_{0\leq r \leq 2m_{2}}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{2m_{2}+1-r}({\mu}) 2^{-r(1-\epsilon_{0})}\\
& & |{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\Pi_{i=1,2}\phi_{i,j_{i},m_{2}}
({\xi}_{i},{\mu}_{i},\tau_{i}).
\end{aligned}$$ and one can use again Lemma \[mainlemma\] in region $A_{2}$ as above. If ${\frac{1}{2}}\gib> |{\xi}_{2}|$ we introduce a dyadic decomposition with respect to ${\mu}_{2}$ and we set $|{\mu}_{2}|\sim 2^{n_{2}}$. Then $|{\mu}_{1}+{\mu}_{2}|\leq C|{\mu}_{2}|
\leq C2^{n_{2}}$ and we can write $1+|{\mu}_{1}+{\mu}_{2}|\sim 2^{n_{2}+1-r},
\, \,
0\leq r\leq n_{2}$. We then bound with $$\begin{aligned}
& & \sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{2}+1-r}2^{(n_{2}+1-r)(1-\epsilon_{0})}\\
& &\times|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}
\frac{\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}{|{\xi}_{1}|^{1-\epsilon_{0}}}
\frac{\chi_{2}\phi_{2,j_{2},n_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{(\gib)^{1-\epsilon_{0}}}\\
&\lesssim&\sum_{j_{1},j_{2}\geq 0}
\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}\sum_{j\geq 0}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{2}+1-r}2^{-r(1-\epsilon_{0})}\\
& &\times
|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}
\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\chi_{2}\phi_{2,j_{2},n_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})\end{aligned}$$ We use again Lemma \[mainlemma\] and we continue with $$\sup_{j,n}\|g_{j}\chi_{2}\chi_{j}
\theta_{n}\|_{L^{2}}\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}
2^{-r(1-\epsilon_{0})}\left(\sum_{j_{1}\geq 0}2^{j_{1}/2}
\|\phi_{1,j_{1}}\|_{L^{2}}\right)
\left(\sum_{j_{2}\geq 0}2^{j_{2}/2}
\|\chi_{2}\phi_{2,n_{2},j_{2}}\|_{L^{2}}\right)$$ and this is enough to prove the theorem in this case.\
[**Case B2:**]{} $|{\mu}_{1}|\geq |{\mu}_{2}|$. One can use the same argument presented for Case B1 inverting the role of $({\xi}_{1},{\mu}_{1})$ and $({\xi}_{2},{\mu}_{2})$.
[**Region $A_{3}\cup A_{4} \cup (A_{5}-\tilde{A}_{5}(\epsilon_{0}))$**]{}.\
[**Case A:**]{} $ {\frac{1}{2}}\gi\leq|{\xi}|$.\
We consider two subcases.\
[**Case A1:**]{} $|\xi_{1}|\geq {\frac{1}{2}}\gia$. We dyadically decompose with respect to $|{\xi}|\sim|{\xi}_{1}|\sim 2^{m_{1}}$ and we bound , now integrated over the region $A=A_{3}\cup A_{4} \cup (A_{5}-\tilde{A}_{5} (\epsilon_0))$, with $$\begin{aligned}
&&\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{m_{1}\geq 0}2^{-j/2}
\int_{A}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}}({\xi})
|{\xi}_{1}+{\xi}_{2}|^{2-\epsilon_{0}}\\
&&\frac{\chi_{2}\phi_{1,j_{1},m_{1}}}{|{\xi}_{1}|^{1-\epsilon_{0}}}
({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}{(\gib)^{1-\epsilon_{0}}}({\xi}_{2},{\mu}_{2},\tau_{2}).\end{aligned}$$ But in this region $\gib\gtrsim |\xi_{1}|\sim |{\xi}_{1}+{\xi}_{2}|$, hence we can continue with $$\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{m_{1}\geq 0}2^{-j/2}
\int_{A}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}}({\xi})
|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\chi_{1}\phi_{1,j_{1},m_{1}}
\phi_{2,j_{2}}
\label{helpy1}$$ We then apply Lemma \[mainlemma\] and we obtain the desired result.\
[**Case A2:**]{} $|\xi_{1}|\leq {\frac{1}{2}}\gia$.\
We do a dyadic decomposition of $\phi_{2}$ in $({\mu}_{2}/{\xi}_{2})$, so that we write $\phi_{2}=\sum_{r_{2} \geq 0}\tilde\phi_{2,r_{2}}= \sum_{r_2 \geq 0} \sum_{j_2 \geq 0}
\widetilde{\phi}_{2, r_2, j_2} $, where $1+ |{\mu}_{2}|/|{\xi}_{2}|\sim 2^{r_{2}}$ and $1 + | \tau_2 - \omega( \xi_2 , \mu_2 ) | \thicksim 2^{j_2}$. Note that $|{\xi}_{1}|\leq C\gia \leq C\gib$, and that if $|{\xi}_{1}|\sim 2^{m_{1}}$, then $2^{r_{2}}\sim 2^{m_{1}+r}, \, r\leq -C$. We bound with $$\begin{aligned}
\nonumber&&\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{m_{1}\geq 0}
\sum_{r\geq -C}2^{-j/2}
\int_{A}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}}({\xi})\\
\nonumber & &|{\xi}_{1}+{\xi}_{2}|^{2-\epsilon_{0}}
\frac{\chi_{2}\phi_{1,j_{1},m_{1}}}{|{\xi}_{1}|^{1-\epsilon_{0}}}
({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\tilde\phi_{2,m_{1}+r,j_{2}}}
{(\gib)^{1-\epsilon_{0}}}({\xi}_{2},{\mu}_{2},\tau_{2})\\
\label{helpy2}&\lesssim&\sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{m_{1}\geq 0}\sum_{-C\leq
r}2^{-j/2}2^{-r(1-\epsilon_{0})}
\int_{A}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}}({\xi}) \\
\nonumber & & |{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\chi_{2}\phi_{1,j_{1},m_{1}}
\tilde\phi_{2,m_{1}+r,j_{2}}.\end{aligned}$$ If we use again Lemma \[mainlemma\] we can continue the chain of inequalities with $$\sup_{j,m_1 }\|g_{j}\chi_{1}\chi_{j}\theta_{m_{1}}\|_{L^{2}}
\sum_{m_{1}\geq 0}\sum_{-C \leq r}2^{-r(1-\epsilon_{0})}
\sum_{j_{1}\geq 0}2^{j_{1}/2}\|\chi_{2}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
\sum_{j_{2}\geq 0}2^{j_{2}/2}\|\tilde\phi_{2,m_{1}+r,j_{2}}\|_{L^{2}}.
\label{helpy3}$$ Now Cauchy-Schwarz in $m_{1}$ is enough to obtain Theorem \[xbil\] in this case.\
[**Case B:**]{} $ {\frac{1}{2}}\gi>|{\xi}|$. As we observed in the analysis of region $A_{i}, i=1,2$, without loss of generality we can assume that $\gi\geq 1$. We consider two subcases.\
[**Case B1:**]{} $|{\mu}_{1}|\leq |{\mu}_{2}|$.\
We recall that in this region we also have $\gib>>|{\xi}_{2}|$. We repeat the argument presented in the second part of Case B1 of region $A_{2}$.\
[**Case B2:**]{} $|{\mu}_{1}|\geq |{\mu}_{2}|$.\
If ${\frac{1}{2}}\gia\leq|{\xi}_{1}|$ then we use an argument similar to the one in the first part of Case B1 in region $A_{2}$, where the role of $({\xi}_{2},{\mu}_{2})$ is now played by $({\xi}_{1},{\mu}_{1})$. In particular, if $|\xi_1 | \thicksim 2^{m_1}$, since $|\mu_2 | / |\xi_2 | \gtrsim |\xi_1 |$ and $|\xi_1 + \xi_3| \thicksim |\xi_1|$, we can bound with $$\begin{aligned}
& & \sum_{j_1, j_2 \geq 0} \sum_{j \geq 0} \sum_{m_1 \geq 0} 2^{-j/2} \int_A g_j \chi_s \chi_j (\xi, \mu, \tau) |\xi_1 + \xi_2 |^{\epsilon_0} \\
& \times & \frac{|\mu_1 + \mu_2 |^{1 - \epsilon_0}}{2^{2(1-\epsilon_0) m_1}} \theta_{m_1} ( \xi_1)
\phi_{1,j_1} (\xi_1, \mu_1, \tau_1) \phi_{2,j_2} (\xi_2, \mu_2, \tau_2). \end{aligned}$$ But, $|\mu_1 + \mu_2 | \leq 2 |\mu_1 | \lesssim \frac{|\mu_1|}{|\xi_1|} |\xi_1 | \lesssim |\xi_1|^2$, so that $$(1 + |\mu_1 + \mu_2 |) \thicksim 2^{2 m_1 -r + 1}, 0 \leq r \leq 2m_1 ,$$ so we can continue with $$\begin{aligned}
& & \sum_{j_1, j_2 \geq 0} \sum_{m_1 \geq 0} \sum_{0 \leq r \leq 2 m_1} 2^{-j/2}
\int_A g_j \chi_2 \chi_j (\xi, \mu, \tau) \theta_{2m_1 + 1 -r } (\mu ) \\
& \times& 2^{-r (1-\epsilon_0 )} |\xi_1 + \xi_2 |^{\epsilon_0} \chi_1 \phi_{1,j_1, m_1} \phi_{2, j_2}.\end{aligned}$$ We then use Lemma \[mainlemma\]. If ${\frac{1}{2}}\gia \geq |{\xi}_{1}|$, we dyadically decompose so that $|{\mu}_{1}|\sim 2^{n_{1}}$. Then $|{\mu}_{1}+{\mu}_{2}|\sim 2^{n_{1}+1-r},
0\leq r\leq n_{1}$. We bound with $$\begin{aligned}
& & \sum_{j\geq 1}\sum_{n_{1}\geq 0}\sum_{0\leq r\leq n_{1}}
2^{-j/2}\int_{A}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{1}+1-r}2^{(n_{1}+1-r)(1-\epsilon_{0})}\\
& &\times|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}
\frac{\chi_{2}\phi_{1,n_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{(\gia)^{1-\epsilon_{0}}}
\frac{\phi_{2}({\xi}_{2},{\mu}_{2},\tau_{2})}
{(\gib)^{1-\epsilon_{0}}}\\
&\lesssim&\sum_{n_{1}\geq 0}\sum_{0\leq r\leq n_{1}}\sum_{j\geq 1}
2^{-j/2}\int_{A}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{1}+1-r}2^{-r(1-\epsilon_{0})}\\
& &\times
|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}2^{n_{1}(1-\epsilon_{0})}
\frac{\chi_{2}\phi_{1,n_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})}
{(\gia)^{1-\epsilon_{0}}}
\frac{\phi_{2}({\xi}_{2},{\mu}_{2},\tau_{2})}
{|{\xi}_{1}|^{1-\epsilon_{0}}}\\
&\lesssim&\sum_{n_{1}\geq 0}\sum_{0\leq r\leq n_{1}}\sum_{j\geq 1}
2^{-j/2}\int_{A}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{1}+1-r}2^{-r(1-\epsilon_{0})}
|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}\chi_{2}\phi_{1,n_{1}}
\phi_{2}\end{aligned}$$ Now again one uses Lemma \[mainlemma\] to conclude the argument.
[**Region $A_{5}\cap\tilde{A}_{5}(\epsilon_{0})$**]{}. We summarize the restrictions that occur in this region: for $\alpha>4$ $$\begin{array}{l}
|{\xi}_{2}|\leq 1, \, \, \gia\lesssim \gib, \, \,
|{\mu}_{1}/{\xi}_{1}-{\mu}_{2}/{\xi}_{2}|\lesssim |{\xi}_{1}+{\xi}_{2}|\\
|{\xi}_{1}|^{1+\alpha\epsilon_{0}}\leq \gia,\, \, \,
\gia\geq 100|{\xi}_{1}|.
\end{array}
\label{conditions}$$ In this case, $A = A^* \cap A_5 \cap {\widetilde{A_5}} ( \epsilon_0 ).$
[**Case A:**]{} $|{\xi}|\geq {\frac{1}{2}}\gi$.\
We dyadically decompose with respect to ${\xi}$, so that $|{\xi}|\sim|\xi_{1}|\sim 2^{m}$. We write $$\begin{aligned}
& &\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{m\geq 0}2^{-j/2}
\int_{A}g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m}({\xi})
|{\xi}_{1}+{\xi}_{2}|^{2-\epsilon_{0}}\\
& &\times
\frac{\phi_{1,j_{1},m}}{(\gia)^{1-\epsilon_{0}}}
({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}
{(\gib)^{1-\epsilon_{0}}}({\xi}_{2},{\mu}_{2},\tau_{2}).\end{aligned}$$ Now let’s consider the multiplier in the above integral. Using (with $\alpha>4$) we can write $$\frac{|{\xi}_{1}|^{2-\epsilon_{0}}}{(\gia)^{2-2\epsilon_{0}}}
\lesssim \frac{|{\xi}_{1}|^{1-2\epsilon_{0}}}{(\gia)^{1-2\epsilon_{0}}}
\lesssim |{\xi}_{1}|^{-\alpha\epsilon_{0}(1-2\epsilon_{0})}.
\label{newmult}$$ Using and Strichartz inequality we can continue the chain of inequalities with $$\begin{aligned}
&\lesssim&\sum_{j\geq 0}2^{-j/2}\sum_{m\geq 0}
\|g_{j}\theta_{m}\chi_{1}\chi_{j}\|_{L^{2}}
2^{-\alpha\epsilon_{0}(1-2\epsilon_{0})m}
\sum_{j_{1},j_{2}\geq 0}2^{j_{1}/2}\|\phi_{1,m,j_{1}}\|_{L^{2}}
2^{j_{2}/2} \|\phi_{2,j_{2}}\|_{L^{2}}\end{aligned}$$ and Cauchy-Schwarz in $m$ is enough to prove the theorem in this case.
[**Case B:**]{} $|{\xi}|<{\frac{1}{2}}\gi$.\
From we have that $$\gia\sim\gib,
\label{m1x1m2x2}$$ hence $|{\mu}_{2}|\sim |{\xi}_{2}|\gia<<|{\mu}_{1}|$, hence $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|$. We dyadically decompose with respect to $|{\mu}_{1}|\sim 2^{n_{1}}$. We have to estimate $$\begin{aligned}
& &\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{n_{1}\geq 0}2^{-j/2}
\int_{A}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{1}}({\mu})
|{\xi}_{1}+{\xi}_{2}|^{\epsilon_{0}}|{\mu}_{1}+{\mu}_{2}|^{1-\epsilon_{0}}\\
& &\times
\frac{\chi_{2}\phi_{1,n_{1},j_{1}}}{(\gia)^{1-\epsilon_{0}}}
({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}
{(\gib)^{1-\epsilon_{0}}}({\xi}_{2},{\mu}_{2},\tau_{2}).\end{aligned}$$ Now let’s consider the multiplier in this integral. Using we can write $$\frac{|{\xi}_{1}|(\gia)^{1-\epsilon_{0}}}{(\gia)^{2-2\epsilon_{0}}}
\lesssim \frac{(\gia)^{1/(1+\alpha\epsilon_{0})}}{(\gia)^{1-\epsilon_{0}}}
\lesssim 1,$$ provided $\epsilon_{0}<(\alpha-1)/\alpha$. Then using Strichartz we can continue the chain of inequality with $$\begin{aligned}
&\lesssim&\sum_{j\geq 0}2^{-j/2}\sum_{n_{1}\geq 0}
\|g_{j}\theta_{n_{1}}\chi_{2}\chi_{j}\|_{L^{2}}
\sum_{j_{1},j_{2}\geq 0}2^{j_{1}/2}\|\chi_{2}
\phi_{1,n_{1},j_{1}}\|_{L^{2}}
2^{j_{2}/2} \|\phi_{2,j_{2}}\|_{L^{2}},\end{aligned}$$ and the theorem is proved also in this case.
[**Region $A_{6}$**]{}. Of the whole theorem this is the region in which the estimates are the most delicate. We summarize the restrictions on this region: $$\begin{array}{l}
|{\xi}_{1}|\geq 1, \, \, \, |{\xi}_{2}|\leq 10^{-10}|{\xi}_{1}|\\
\gib\leq 10^{-10}\max(|{\xi}_{1}|, \gia).
\end{array}
\label{conditionsa6}$$ We observe that the multipliers appearing in and can be bounded in the following way: $$\frac{|{\xi}|\max(1,|{\xi}|,\gi)^{1-\epsilon_{0}}}
{\max(1,|{\xi}_{1}|,\gia)^{1-\epsilon_{0}}
\max(1,|{\xi}_{2}|,\gib)^{1-\epsilon_{0}}}\lesssim
\frac{|{\xi}_{1}|}{\max(1,|{\xi}_{2}|,\gib)^{1-\epsilon_{0}}}
\label{a6mult}$$ This is obvious when $|{\xi}|\geq {\frac{1}{2}}\gi$ or when $|{\xi}|\leq {\frac{1}{2}}\gi$ and $\gi\leq 1$. In the remaining region we estimate the numerator $$\begin{aligned}
|{\xi}|\max(1,|{\xi}|,\gi)^{1-\epsilon_{0}}&\sim& |{\xi}_{1}|^{\epsilon_{0}}
|{\mu}_{1}+{\mu}_{2}|^{1-\epsilon_{0}}\\
&\lesssim&|{\xi}_{1}|^{\epsilon_{0}}
|{\mu}_{1}|^{1-\epsilon_{0}}+
|{\xi}_{1}|^{\epsilon_{0}}|{\mu}_{2}|^{1-\epsilon_{0}}\\
&\lesssim &|{\xi}_{1}|^{\epsilon_{0}}|{\xi}_{1}|^{1-\epsilon_{0}}
(\gia)^{1-\epsilon_{0}}+|{\xi}_{1}|^{\epsilon_{0}}|{\xi}_{2}|^{1-\epsilon_{0}}
(\gib)^{1-\epsilon_{0}}\\
&\lesssim&|{\xi}_{1}|((\gia)^{1-\epsilon_{0}}+
\max(|{\xi}_{1}|, \gia)^{1-\epsilon_{0}}).\end{aligned}$$ [**Case A:**]{} $|{\xi}_{1}|\geq 10^{2}\gia$.\
Let’s show that in this case we also have $|{\xi}|\geq \gi$. $$\begin{aligned}
\gi&\leq &(1-10^{-10})^{-1}\frac{|{\mu}_{1}+{\mu}_{2}|}{|{\xi}_{1}|}
\leq (1-10^{-10})^{-1}\left(\gia+\gib \frac{|{\xi}_{2}|}{|{\xi}_{1}|}\right)
\\
&\leq&(1-10^{-10})^{-1}\left(\gia+10^{-20}|{\xi}_{1}|\right)\leq
(1-10^{-10})^{-1}(10^{-2}+10^{-20})|\xi_{1}|\\
&\leq&
(1-10^{-10})^{-2}(10^{-2}+10^{-20})|{\xi}_{1}+{\xi}_{2}|\leq |{\xi}|.\end{aligned}$$ We have to estimate and we use again the functions $\phi_{i,j_{i}}$. We change variables in $\tau_{1}$ and $\tau_{2}$ as in and we use to bound with $$\begin{aligned}
\label{a6step1}&&
\sum_{j\geq 0} \sum_{m\geq 0} \sum_{j_1 , j_2 \geq 0}
2^{-j/2}
\int \chi_{1}\theta_{m_{1}}g_{j}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2},\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))\\
\nonumber & &\chi_{j}(\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})
-{\omega}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2}))\chi_{1}(\xi_{1}, \mu_1 )\\
\nonumber& &\frac{|{\xi}_{1}|}{\max(1,|{\xi}_{2}|, \gib)^{1-\epsilon_{0}}}
\Pi_{i=1,2}\phi_{i}({\xi}_{i},{\mu}_{i},\theta_{i}+{\omega}({\xi}_{i},{\mu}_{i}))
\chi_{j_{i}}(\theta_{i})d{\xi}_{i}d{\mu}_{i}d\tau_{i}.
\end{aligned}$$ We change variables again and this time we use . From we deduce that $|J_{{\mu}}|\geq C|{\xi}_{1}|^{2}$. We also perform a dyadic decompositions by setting $|{\xi}_{i}|\sim 2^{m_{i}}$ (hence $|{\xi}|\sim 2^{m_{1}}$) and $|\mu_{2}|\sim 2^{n_{2}}$. Let $m^{*}_{2}=\max(n_{2}-m_{2},m_{2})$, (here $m_{2}, n_{2}\in {{{\mathbb }Z}}$).\
[**Case A1:**]{} $j\geq \max(0, m^{*}_{2}).$\
Denote with $H(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})$ the transformation of $$\chi_{1}(\xi_{1}, \mu_1)\phi_{1}({\xi}_{1},{\mu}_{1},\theta_{1}+{\omega}({\xi}_{1},{\mu}_{1}))
\chi_{j_{1}}(\theta_{1})\frac{\phi_{2}
({\xi}_{2},{\mu}_{2},\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))
\chi_{j_{2}}(\theta_{2})}{\max(1,|{\xi}_{2}|,\gib)^{1-\epsilon_{0}}}$$ under the change of variables in . Here we use the fact that $\Delta_{{\mu}_{2}}\sim 2^{n_{2}}$. Then we can rewrite as $$\begin{aligned}
& &\sum_{m_{1}\geq 0, m_{2}}\sum_{n_{2}}
\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq \max(0, m^{*}_{2})}
2^{-j/2}2^{m_{1}}\int g_{j}\theta_{m_{1}}\chi_{1}\chi_{j}(u,v,w)\\
& &|J_{{\mu}}|^{-1}H(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})dudvdw
d{\mu}_{2}d\theta_{1}d\theta_{2}\\
&\lesssim &\sum_{m_{1}\geq 0, m_{2}}\sum_{n_{2}}
\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq \max(0, m^{*}_{2})}
2^{-j/2}2^{m_{1}}2^{n_{2}/2}\\
&&\int g_{j}\theta_{m_{1}}\chi_{1} \chi_{j}(u,v,w)
\left(\int_{{\mu}_{2}}|J_{{\mu}}|^{-2}
H^{2}(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})d{\mu}_{2}\right)^{{\frac{1}{2}}}dudvdw
d\theta_{1}d\theta_{2}.
\end{aligned}$$ Now we observe that by Cauchy-Schwarz and the inverse change of variable we have $$\begin{aligned}
& &\int g_{j}\theta_{m_{1}}\chi_{1} \chi_{j}(u,v,w)
\left(\int_{{\mu}_{2}}|J_{{\mu}}|^{-2}
H^{2}(u,v,w,{\mu}_{2},\theta_{1},\theta_{2})d{\mu}_{2}
\right)^{{\frac{1}{2}}}dudvdw
d\theta_{1}d\theta_{2}\\
&\lesssim& 2^{-m_{1}}\|g_{j}\chi_{j} \theta_{m_{1}}
\chi_{1} \|_{L^{2}}
2^{j_{1}/2}\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\frac{\|\phi_{2,j_2, m_{2},n_{2}}\|_{L^{2}}}
{\max(1,2^{m^{*}_{2}})^{1-\epsilon_{0}}},
\end{aligned}$$ where $\phi_{2, j_2, m_2, n_2} = \phi_{2, j_2 , m_2 } \theta_{n_2},$ and inserting this above we can continue with $$\begin{aligned}
\label{a6step2}&\lesssim &\sum_{m_{1}
\geq 0, m_{2}}\sum_{n_{2}}
\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq \max(0, m^{*}_{2})}
2^{-j/2}2^{n_{2}/2}\\
\nonumber& &\|g_{j}\chi_{j} \theta_{m_{1}}\chi_{1} \|_{L^{2}}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\frac{\|\phi_{2,j_2 m_{2},n_{2}}\|_{L^{2}}}
{\max(1,2^{m^{*}_{2}})^{1-\epsilon_{0}}}.\end{aligned}$$ [**Case A1a:**]{} $m_{2}\leq 0, \, n_{2}-m_{2}>0$.\
Now $m^{*}_{2}=n_{2}-m_{2}$ and can be bounded by $$\begin{aligned}
& &\sum_{m_{1}
\geq 0, m_{2}\leq 0}\sum_{n_{2}\geq m_{2}}
\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq \max(0, m^{*}_{2})}
2^{-j/2}2^{n_{2}/2}2^{(-n_{2}+m_{2})(1-\epsilon_{0})}\\
& &\|g_{j}\chi_{j} \theta_{m_{1}}\chi_{1} \|_{L^{2}}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_2 , m_{2},n_{2}}\|_{L^{2}},\end{aligned}$$ and if we sum for $j\geq n_{2}-m_{2}$ we can continue our chain of inequalities with $$\lesssim \sum_{m_{1}
\geq 0}\sum_{n_{2}\geq m_{2}, m_{2}\leq 0}
\sum_{j_{1},j_{2}\geq 0}
2^{m_{2}/2}2^{(-n_{2}+m_{2})(1-\epsilon_{0})}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_2 , m_{2},n_{2}}\|_{L^{2}},$$ and by Cauchy-Schwarz on $n_{2}$ and $m_{2}\leq 0$ we obtain $$\eqref{duality1}\lesssim \sum_{m_{1}\geq 0}
\sum_{j_{1},j_{2}\geq 0}2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_{2}}\|_{L^{2}}
\label{a6step3}$$ and this proves the theorem in this case.\
[**Case A1b:**]{} $m_{2}\leq 0, \, n_{2}-m_{2}\leq 0$.\
In this case $\max(1,2^{m^{*}_{2}})=1$. We repeat the argument above and we bound with $$\begin{aligned}
& &\sum_{m_{1}
\geq 0}\sum_{n_{2}\leq m_{2}\leq 0}
\sum_{j_{1},j_{2}\geq 0}\sum_{0\leq j}
2^{-j/2}2^{n_{2}/2}\\
\nonumber& &\|g_{j}\chi_{j}\theta_{m_{1}}\chi_{1} \|_{L^{2}}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_2, m_{2},n_{2}}\|_{L^{2}}\end{aligned}$$ and one obtains again after summing in $j$ and using Cauchy-Schwarz first in $n_{2}$ and then in $m_{2}$.\
[**Case A1c:**]{} $m_{2}>0, \, n_{2}-m_{2}\leq 0$.\
In this case $m_{2}^{*}=m_{2}>0$. We obtain $$\begin{aligned}
\eqref{duality1}&\lesssim &\sum_{m_{1}
\geq 0}\sum_{n_{2}\leq m_{2}}\sum_{0\leq m_{2}}
\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq m_{2}}
2^{-j/2}2^{n_{2}/2}2^{-m_{2}(1-\epsilon_{0})}\\
\nonumber& &\|g_{j}\chi_{j}\theta_{m_{1}}\chi_{1} \|_{L^{2}}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_2 , m_{2},n_{2}}\|_{L^{2}}\end{aligned}$$ and summing over $j$ $$\lesssim \sum_{m_{1}
\geq 0}\sum_{n_{2}\leq m_{2}}\sum_{0\leq m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{-m_{2}/2}2^{n_{2}/2}2^{-m_{2}(1-\epsilon_{0})}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}$$ and by Cauchy-Schwarz in $n_{2}$ $$\lesssim \sum_{m_{1}
\geq 0}\sum_{0\leq m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{-m_{2}(1-\epsilon_{0})}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_{2},m_{2}}\|_{L^{2}}$$ and a final Cauchy-Schwarz in $m_{2}$ concludes the argument.\
[**Case A1d:**]{} $m_{2}>0, \, n_{2}-m_{2}\geq 0$.\
In this case it is easy to see that $$2^{n_{2}/2}2^{-m^{*}_{2}}\leq 1.
\label{restr}$$ Using the same type of estimates presented above, after summing on $j\geq m^{*}_{2}\geq 0$ we obtain $$\eqref{duality1}\lesssim \sum_{m_{1}
\geq 0}\sum_{n_{2}, m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{-m^{*}_{2}(3/2-\epsilon_{0})}2^{n_{2}/2}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}},$$ and after using we can continue with $$\lesssim \sum_{m_{1}
\geq 0}\sum_{n_{2}, m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{-m^{*}_{2}({\frac{1}{2}}-\epsilon_{0})}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.$$ If $n_{2}-m_{2}\geq m_{2}$ then $$\lesssim \sum_{m_{1}
\geq 0}\sum_{m_{2}>0}\sum_{n_{2}\geq 2m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{(-n_{2}+m_{2})({\frac{1}{2}}-\epsilon_{0})}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.$$ and we conclude by Cauchy-Schwarz first with respect to $n_{2}$ and then $m_{2}$, (here we assume $\epsilon_{0}<{\frac{1}{2}}$). If $0\leq n_{2}-m_{2}< m_{2}$, we have $m_{2}\leq n_{2}<2 m_{2}$, $$\lesssim \sum_{m_{1}
\geq 0}\sum_{m_{2}\leq n_{2}<2 m_{2}}\sum_{m_{2}\geq 0}
\sum_{j_{1},j_{2}\geq 0}
2^{-m_{2}({\frac{1}{2}}-\epsilon_{0})}
2^{j_{1}/2}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.$$ and we proceed as in the previous case. This concludes the analysis of Case A1.\
[**Case A2:**]{} $0\leq j\leq m_{2}^{*}$.\
[**Case A2a:**]{} $m_{2}\geq 0$ or $m_{2}<0$ and $n_{2}>0$.\
We claim that in this case $j\leq\max{(j_{1},j_{2}})$. Recall the fundamental identity . Then combining this with the restrictions of Case A, we conclude that $2^{2m_{1}+m_{2}}\lesssim 2^{\max(j,j_{1},j_{2})}$. If $j\geq \max(j_{1},j_{2})$, then $2^{2m_{1}+m_{2}}\lesssim 2^{j}$ and if $m_{2}\geq 0$ this implies that $2^{2m_{1}}\leq \max(2^{m_{2}}, 2^{n_{2}-m_{2}})$, a contradiction if one compares this with . When $m_{2}<0$ and $
n_{2}\geq 0$, then $n_{2}-m_{2}\geq 0$ and we obtain $2^{2m_{1}}\lesssim 2^{n_{2}-2m_{2}}\lesssim 2^{2(n_{2}-m_{2})}$, again a contradiction if one compares with . If we assume that $j_{1}=\max(j_{1},j_{2})$, then $2^{2m_{1}+m_{2}}\lesssim 2^{j_{1}}$, that is $$|{\xi}_{1}|\sim 2^{m_{1}}\lesssim 2^{j_{1}/2-m_{2}/2}.$$ We can then bound with $$\lesssim \sum_{m_{1}
\geq 0}\sum_{n_{2}, m_{2}}\sum_{j\leq m_{2}^{*}}
\sum_{j_{1},j_{2}\geq 0}
2^{-j/2}2^{j_{1}/2-m_{2}/2}\int \chi_{j}g_{j}\chi_{1}\theta_{m_{1}}
\chi_{1}\phi_{1,j_{1},m_{1}}
2^{-m_{2}^{*}(1-\epsilon_{0})}\phi_{2, j_2 , m_2 , n_2 }
\label{a6step5}$$ We use the Strichartz inequality for $\chi_j g_j \chi_1 \theta_{m_1}$ and for $\phi_{2, j_2 , m_2 , n_2}$ and Plancherel for $\phi_{1, j_1, m_1}$ and Hölder’s inequality to continue the chain of inequalities with $$\lesssim \sum_{m_{1}\geq 0}\sum_{n_{2}, m_{2}}
\sum_{j\leq m_{2}^{*}}\sum_{j_{1},j_{2}\geq 0}
2^{j_{1}/2-m_{2}/2}2^{-m_{2}^{*}(1-\epsilon_{0})}
\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.$$ We now sum over $j$ to get
$$\lesssim \sum_{m_{1}\geq 0}\sum_{n_{2}, m_{2}}
\sum_{j_{1},j_{2}\geq 0}
(1+ m_{2}^{*})2^{-m_{2}/2}2^{-m_{2}^{*}(1-\epsilon_{0})}
2^{j_{1}/2}\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.
\label{a6step4}$$
Now assume that $m_{2}\geq 0$. We split the $m_{2}$ sum in into $n_{2}-m_{2}> m_{2}$ and $n_{2}-m_{2}\leq m_{2}$. In the first case becomes $$\lesssim \sum_{m_{1}\geq 0}\sum_{n_{2}\geq 2m_{2}}
\sum_{j_{1},j_{2}\geq 0}
[1+(n_{2}-m_{2})]2^{-m_{2}/2}2^{-(n_{2}-m_{2})(1-\epsilon_{0})}
2^{j_{1}/2}\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.$$ We first use Cauchy-Schwarz on $n_{2}$ and then on $m_{2}$ to finish. In the second case $n_{2}\leq 2m_{2}$ and $m_{2}^{*}=m_{2}$. In this case we go back to and we sum with respect to $n_{2}$. Then we use Strichartz inequality in the order $L^{4}L^{2}L^{4}$ (as above) to get $$\lesssim \sum_{m_{1}\geq 0}\sum_{m_{2}\geq 0}\sum_{0\leq j\leq m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{-m_{2}/2}2^{-m_{2}(1-\epsilon_{0})}
2^{j_{1}/2}\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_{2},m_{2}}\|_{L^{2}}.$$ Summing in $j$ and then using Cauchy-Schwarz in $m_2$ will prove the theorem also in this case [^11]. Assume now that $m_{2}<0$ and $n_{2}>0$, hence $m_{2}^{*}=n_{2}-m_{2}\geq 0$. Then becomes $$\sum_{m_{1}\geq 0}\sum_{n_{2}\geq 0}\sum_{m_{2}\leq 0}
\sum_{j_{1},j_{2}\geq 0}
[1+ (n_{2}-m_{2})]2^{-m_{2}/2}2^{-(n_{2}-m_{2})(1-\epsilon_{0})}
2^{j_{1}/2}\|\chi_{1}\phi_{1,j_{1},m_{1}}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2, j_2 , m_2 , n_2 }\|_{L^{2}}.$$ Observe that $$2^{-m_{2}/2}(n_{2}-m_{2})2^{-(n_{2}-m_{2})(1-\epsilon_{0})}\lesssim
2^{-(1-\sigma)(1-\epsilon_{0})n_{2}}2^{m_{2}
(-{\frac{1}{2}}+(1-\sigma)(1-\epsilon_{0}))}$$ for some $0 < \sigma<<1$ and $\epsilon_{0}<{\frac{1}{2}}$. This is enough for Cauchy-Schwarz with respect to $n_{2}$ and $m_{2}$.\
[**Case A2b:**]{} $m_{2}, \, n_{2}\leq 0$\
In this case $|{\xi}_{2}|, |{\mu}_{2}|\leq 1$. We bound with $$\begin{aligned}
\label{a6step7}&&\sum_{m_{1}\geq 0}
\sum_{0\leq j}
\sum_{j_{1},j_{2}\geq 0}2^{-j/2}\int_{|{\xi}_{2}|,|{\mu}_{2}|\leq 1}
(|{\xi}_{1}+{\xi}_{2}|^{{\frac{1}{2}}}g_{j}\theta_{m_{1}}
\chi_{1}\chi_{j})\\
\nonumber & &(|{\xi}_{1}|^{{\frac{1}{2}}}\chi_{1}\phi_{1,j_{1},m_{1}})
\frac{\phi_{2,j_{2}}}
{\max{(1, |{\xi}_{2}|,|{\mu}_{2}|/|{\xi}_{2}|)^{1-\epsilon_{0}}}}
\end{aligned}$$ If $|{\mu}_{2}|^{2}/|{\xi}_{2}|\lesssim 2^{j_{2}}$ then $|{\omega}({\xi}_{2},{\mu}_{2})|\lesssim 2^{j_{2}}$ and because $|\tau_{2}-{\omega}({\xi}_{2},{\mu}_{2})|\sim 2^{j_{2}}$ we also have $|\tau_{2}|\lesssim 2^{j_{2}}$. We then apply Hölder’s inequalities in the order $L^{4}_{x}L^{2}_{y}L^{2}_{t}-L^{4}_{x}L^{2}_{y}L^{2}_{t}
-L^{2}_{x}L^{\infty}_{y}L^{\infty}_{t}$ combined with and to continue the chain of inequalities in with $$\lesssim \sum_{m_{1}\geq 0}\sum_{0\leq j}
\sum_{j_{1},j_{2}\geq 0}2^{-j/2}2^{j/4}\|g_{j} \theta_{m_1}
\chi_{1}\chi_{j}\|_{L^{2}}
2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\chi_{1}\|_{L^{2}}
2^{j_{2}/2}\|\phi_{2,j_{2}}\|_{L^{2}}$$ and in this case we are done. Assume now that $|{\mu}_{2}|^{2}/|{\xi}_{2}|>> 2^{j_{2}}$. Then $|\tau_2|\sim
|{\mu}_{2}|^{2}/|{\xi}_{2}|$ and we can rewrite as follows $$\begin{aligned}
& &\sum_{m_{1}\geq 0}\sum_{0\leq j}
\sum_{j_{1},j_{2}\geq 0}2^{-j/2}\int(|{\xi}_{1}+{\xi}_{2}|^{{\frac{1}{2}}}g_{j}
\theta_{m_{1}}
\chi_{1}\chi_{j})(|{\xi}_{1}|^{{\frac{1}{2}}}\phi_{1,j_{1},m_{1}}\chi_{1})
\frac{\phi_{2,j_{2}}|{\mu}_{2}|^{1-\epsilon_{0}}}
{(|{\mu}_{2}|^{2}/|\xi_{2}|)^{1-\epsilon_{0}}}\\
&\sim& \sum_{m_{1}\geq 0}\sum_{0\leq j}
\sum_{j_{1},j_{2}\geq 0}2^{-j/2}\int(|{\xi}_{1}+{\xi}_{2}|^{{\frac{1}{2}}}g_{j}
\theta_{m_{1}}
\chi_{1}\chi_{j})(|{\xi}_{1}|^{{\frac{1}{2}}}\phi_{1,j_{1},m_{1}}\chi_{1})
\frac{\phi_{2,j_{2}}|{\mu}_{2}|^{1-\epsilon_{0}}}
{|\tau_{2}|^{1-\epsilon_{0}}}\\
&\sim& \sum_{m_{1}\geq 0}\sum_{0\leq j}
\sum_{j_{1},j_{2}\geq 0}2^{-j/2+j/4}\|g_{j}\theta_{m_{1}}
\chi_{1}\|_{L^{2}}2^{j_{1}/4}\|\phi_{1,j_{1},m_{1}}\chi_{1}\|_{L^{2}}
\|\phi_{2,j_{2}}\|_{L^{2}}\end{aligned}$$ where in the last step we used again the fact that $|{\mu}_{2}|\leq 1$, $\epsilon_{0}<{\frac{1}{2}}$, Hölder’s inequality in the order $L^{4}_{x}L^{2}_{y}L^{2}_{t}-L^{4}_{x}L^{2}_{y}L^{2}_{t}
-L^{2}_{x}L^{\infty}_{y}L^{\infty}_{t}$ and and . This concludes the analysis of Case A.\
[**Case B:**]{} $|{\xi}_{1}|\leq 10^{-2}\gia$.\
Using one can prove that $|{\mu}_{1}+{\mu}_{2}|/|{\xi}_{1}+{\xi}_{2}|\geq
2|{\xi}_{1}+{\xi}_{2}|$, that $|{\mu}_{1}+{\mu}_{2}|/|{\xi}_{1}+{\xi}_{2}|\sim |{\mu}_{1}|/|{\xi}_{1}|$, and that $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|$. We dyadically decompose $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|\sim 2^{n_{1}}$ and we write as follows $$\label{a6step9}\sum_{0\leq j}\sum_{j_{1},j_{2}\geq 0}
\sum_{n_{1}\geq 0}
2^{-j/2}\int|{\xi}_{1}|g_{j}\theta_{n_{1}}({\mu}_{1}+{\mu}_{2})
\chi_{2}\chi_{j}\phi_{1,j_{1},n_{1}}\chi_{2}
\frac{\phi_{2,j_{2}}}
{\max{(1, |{\xi}_{2}|,|{\mu}_{2}|/|{\xi}_{2}|)^{1-\epsilon_{0}}}}.$$ We dyadically decompose also $|{\xi}_{2}|\sim 2^{m_{2}}$ and $|{\mu}_{2}|\sim 2^{n_{2}}$. As in Case A, we define $m^{*}_{2}=
\max(n_{2}-m_{2}, m_{2})$ and we analyze two subcases.\
[**Case B1:**]{} $j\geq \max(0, m^{*}_{2}).$\
We use the change of variable . Now $|J_{{\mu}}|\gtrsim (\gia)^{2}
\gtrsim |{\xi}_{1}|^{2}$. Then we proceed like in Case A1 above, where the sum in $m_{1}$ is replaced by a sum in $n_{1}$.\
[**Case B2:**]{} $0\leq j\leq m^{*}_{2}.$\
We again consider the two subcases $m_{2}\geq 0$ or $m_{2}<0$ and $n_{2}\geq 0$, and $m_{2}< 0$ and $ n_{2}<0$.\
[**Case B2a:**]{} $m_{2}\geq 0$ or $m_{2}<0$ and $n_{2}>0$.\
The fundamental identity now gives $$(\gia)^{2}|{\xi}_{2}|\leq ({\frac{1}{2}}-10^{-10})^{-1}2^{\max(j_{1},j_{2},j)}.$$ This again forces $j\leq \max(j_{1},j_{2})$. In fact after setting $C=({\frac{1}{2}}-10^{-10})^{-1}$, if $j>
\max(j_{1},j_{2})$ and $m_{2}\geq 0$, then $(\gia)^{2}\leq C
2^{\max(n_{2}-m_{2},m_{2})}$ and so $(\gia)^{2}\leq C
\max(\gib, |{\xi}_{2}|)$ which is a contradiction in this region. If $m_{2}< 0$ and $ n_{2}\geq 0$, then $(\gia)^{2}|{\xi}_{2}|
\leq C2^{n_{2}-m_{2}}$ or $(\gia)^{2}\leq C2^{n_{2}-2m_{2}}
\leq C(\gib)^{2}$, which is again a contradiction. Thus (if for example $j_{2}\leq j_{1}$) $$|{\xi}_{1}|\leq C\gia\leq 10^{-2}2^{j_{1}/2}2^{-m_{2}/2},$$ and can be bounded by $$\sum_{j\leq m^{*}_{2}}\sum_{j_{1},j_{2}\geq 0}
\sum_{n_{1}\geq 0}\sum_{n_{2},m_{2}}
2^{-j/2}2^{j_{1}/2-m_{2}/2}\int
\chi_{2}\chi_{j}g_{j,n_{1}}\theta_{n_{1}}
\phi_{1,j_{1},n_{1}}\chi_{2}
2^{-m_{2}^{*}(1-\epsilon_{0})}\phi_{2,j_{2},m_{2},n_{2}}.
\label{a6step10}$$ At this point we argue like in Case A2a by replacing with .\
[**Case B2b:**]{} $n_{2},m_{2}\leq 0$.\
This case can be treated like Case A2b by replacing with $$\begin{aligned}
\label{a6ste10}&&\sum_{n_{1}\geq 0}
\sum_{0\leq j}
\sum_{j_{1},j_{2}\geq 0}2^{-j/2}\int_{|{\xi}_{2}|,|{\mu}_{1}|\leq 1}
(|{\xi}_{1}+{\xi}_{2}|^{{\frac{1}{2}}}g_{j}\theta_{n_{1}}
\chi_{2}\chi_{j})\\
\nonumber & &(|{\xi}_{1}|^{{\frac{1}{2}}}\chi_{2}\phi_{1,j_{1},n_{1}})
\frac{\phi_{2,j_{2}}}
{\max{(1, |{\xi}_{2}|,|{\mu}_{2}|/|{\xi}_{2}|)^{1-\epsilon_{0}}}}.
\end{aligned}$$ [**Case C:**]{} $10^{-2}\gia\leq |{\xi}_{1}|\leq 10^{2}\gia$.\
It is easy to show that in this case $|{\mu}_{1}+{\mu}_{2}|/|{\xi}_{1}+{\xi}_{2}|\lesssim |{\xi}_{1}+{\xi}_{2}|$. We dyadically decompose with respect to $|{\xi}_{1}|\sim |{\xi}_{1}+{\xi}_{2}|\sim 2^{m_{1}}$. We go back to and we consider two subcases: when $j>m_{1}$ and when $j\leq m_{1}$.\
[**Case C1:**]{} $j>m_{1}$.\
In this case we use the change of variable , where now the free variable is ${\xi}_{2}$ instead of ${\xi}_{1}$. It is easy to check that also in this case holds true and in particular $|J_{{\xi}}|\gtrsim \gia \gtrsim |{\xi}_{1}|$. We perform a dyadic decomposition in ${\xi}_{2}$, but only for large frequencies, that is for $|{\xi}_{2}|\geq 1$. Then becomes $$\begin{aligned}
& &\sum_{m_{1}, m_{2}\geq 0}\sum_{j_{1},j_{2}\geq 0}
\sum_{j> m_{1}}
2^{-j/2}2^{m_{1}}\int \chi_{1}g_{j}\theta_{m_{1}}\chi_{j}(u,v,w)\\
& &\times |J_{{\xi}}|^{-1}H(u,v,w,{\xi}_{2},\theta_{1},\theta_{2})
dudvdw
d{\xi}_{2}d\theta_{1}d\theta_{2},\end{aligned}$$ where $H(u,v,w,{\xi}_{2},\theta_{1},\theta_{2})$ is the transformation of $$\chi_{1}\phi_{1,m_{1}}
({\xi}_{1},{\mu}_{1},\theta_{1}+{\omega}({\xi}_{1},{\mu}_{1}))
\chi_{j_{1}}\frac{\phi_{2, m_2 }
({\xi}_{2},{\mu}_{2},\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))
\chi_{2}\chi_{j_{2}}}
{\max(1,|{\xi}_{2}|,\gib)^{1-\epsilon_{0}}}$$ under the above change of variables. We first use Cauchy-Schwarz in $(u,v,w)$ to obtain $$\begin{aligned}
&\lesssim & \sum_{m_1, m_2 \geq 0} \sum_{j_1, j_2 \geq 0} \sum_{j> m_1 }
2^{-j/2} 2^{m_1} {{\| g_j \theta_{m_1} \chi_1 \|}_{L^2}} \\
&\times & \left( \int \left( |J_\xi |^{-2} H^2 ( u, v, w , \xi_2 , \theta_1 , \theta_2 )
du dv dw \right)^{\frac{1}{2}}d\xi_2 d\theta_1 d\theta_2 \right)\end{aligned}$$ We now use the lower bound $|J_\xi| \gtrsim 2^{m_1}$, and Cauchy-Schwarz in $\xi_2 , \theta_1
, \theta_2,$ to obtain $$\begin{aligned}
& \lesssim& \sum_{m_1, m_2 \geq 0} \sum_{j_1 , j_2 \geq 0} \sum_{j > m_1 }
2^{-j/2} 2^{m_1 /2} 2^{m_2 /2 } 2^{j_1 /2 } 2^{j_2 /2} \\
&\times & {{\| g_j \theta_{m_1} \chi_1 \|}_{L^2}} \left( \int H^2 ( u , v, w , \xi_2 , \theta_1 ,
\theta_2 ) |J_\xi |^{-1} du dv dw d\xi_2 d\theta_1 d\theta_2 \right)^{\frac{1}{2}},\end{aligned}$$ which, upon undoing the change of variables yields (since $\max ( 1 , |\xi_2 | , |\mu_2 |/ |\xi_2 |) \geq 2^{m_2}$) $$\begin{aligned}
&\lesssim & \sum_{m_1, m_2 \geq 0} \sum_{j_1 , j_2 \geq 0} \sum_{j > m_1 }
2^{-j/2} 2^{m_1 /2} 2^{m_2 /2 } 2^{j_1 /2 } 2^{j_2 /2} \\
&\times & {{\| g_j \theta_{m_1} \chi_1 \|}_{L^2}} {{\| \phi_{1, j_1 , m_1 } \chi_1 \|}_{L^2}}
{{\| \phi_{2, j_2 , m_2 } \|}_{L^2}} 2^{-m_2 (1 - \epsilon_0 )}.\end{aligned}$$ Then a sum in $j$ and Cauchy-Schwarz in $m_{2}$ gives the result.\
[**Case C2:**]{} $j\leq m_{1}$.\
Let’s introduce the region $$R=\{10^{-2}\gia\leq |{\xi}_{1}|\leq 10^{2}\gia\}\cap A_{6}$$ This is the region where we need to introduce the space $Y_{1-\epsilon_{0},\epsilon_{0},{\frac{1}{2}}}$. We go back to and this time we keep $|\hat v|$ and we only normalize $|\hat u|$. Then becomes $$\begin{aligned}
\label{a6step11}& &\sum_{m_{1}\geq 0}\sum_{0\leq j\leq
m_{1}}\sum_{j_{1},j_{2}\geq 0}2^{-j/2}
\int_{R} \chi_{1}\theta_{m_{1}}g_{j}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2},
\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))\\
\nonumber & &\chi_{1}\chi_{j}(\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})
-{\omega}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2}))
|{\xi}_{1}|\\
\nonumber &&\chi_{1}\chi_{j_{1}}\phi_{1,m_{1}}({\xi}_{1},{\mu}_{1},\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1}))\chi_{j_{2}}|\hat{v}|({\xi}_{2},{\mu}_{2},\theta_{2}+
{\omega}({\xi}_{2},{\mu}_{2})).
\end{aligned}$$ Define $D_{m_{1},m_{2}}$ to be the dyadic block such that $|{\xi}_{i}|\sim 2^{m_{i}}, i=1,2$. We observe that for fixed $({\xi}_{1},\xi_{2}, {\mu}_{1},
\theta_{1}, \theta_{2})$, the set of ${\mu}_{2}$ such that $({\xi}_{1},\xi_{2}, {\mu}_{1}, {\mu}_{2},\theta_{1}, \theta_{2})\in R\cap
D_{m_{1},m_{2}}$ and such that is true, is a union of two symmetric intervals with length satisfying . Similarly, for fixed $(\xi_{2}, {\mu}_{1}, {\mu}_{2},\theta_{1}, \theta_{2})$ the set of ${\xi}_{1}$ such that $({\xi}_{1},\xi_{2}, {\mu}_{1}, {\mu}_{2},\theta_{1}, \theta_{2})\in R\cap
D_{m_{1},m_{2}}$ and such that is true, is a union of two symmetric intervals with length satisfying $$|\Delta_{{\xi}_{1}}|\lesssim 2^{j-m_{1}-m_{2}}.
\label{deltam1}$$ To prove this it’s enough to use the mean value theorem, and estimate from below $|g'(\xi_1)|$, where $$g({\xi}_1 )=\theta_{1}+\theta_{2}+\frac{{\xi}_1 {\xi}_{2}}{({\xi}_1 +{\xi}_{2})}
\left(\left(\frac{{\mu}_{1}}{{\xi}_1 }-\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}
-3(\xi_1 +\xi_{2})^{2}\right).$$ After a short calculation one has $$g'({\xi}_1 )=\frac{{\xi}_{2}^{2}}{({\xi}_1 +{\xi}_{2})^{2}}
\left(\left(\frac{{\mu}_{1}}{{\xi}_1 }-\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}
-3(\xi_1 +\xi_{2})^{2}
\right) - \xi_{2}\left(\frac{2{\mu}_{1}}{\xi_1 ({\xi}_1 +{\xi}_{2})}
\left(\frac{\mu_{1}}{{\xi}_1 }-\frac{\mu_{2}}{{\xi}_{2}}\right)+6{\xi}_1 \right).$$ Note that in $R$ $$\left|\frac{{\xi}_{2}^{2}}{({\xi}_{1}+{\xi}_{2})^{2}}
\left(\left(\frac{{\mu}_{1}}{{\xi}_{1}}-\frac{{\mu}_{2}}{{\xi}_{2}}\right)^{2}
-3(\xi_{1}+\xi_{2})^{2}
\right)\right| \leq C\frac{|{\xi}_{2}|^{2}}{|{\xi}_{1}|^{2}}
\left(\frac{|{\mu}_{1}|}{|{\xi}_{1}|}\right)^{2}\leq C10^{-10}
|{\xi}_{2}|^{2}.$$ On the other hand it’s easy to check that ${\xi}_1$ and $\frac{2{\mu}_{1}}{\xi_1 ({\xi}_1 +{\xi}_{2})}
\left(\frac{\mu_{1}}{{\xi}_1 }-\frac{\mu_{2}}{{\xi}_{2}}\right)$ have the same sign, hence $$\left|\xi_{2}\left(\frac{2{\mu}_{1}}{\xi_1 ({\xi}_1 +{\xi}_{2})}
\left(\frac{\mu_{1}}{{\xi}_1 }-\frac{\mu_{2}}{{\xi}_{2}}\right)+6{\xi}_1 \right)
\right|\geq 6|{\xi}_{2}|{\xi}_1 |$$ and the claim follows.\
[**Case C2a:**]{} $m_{2}\geq 0$\
In this case we use the change of variables , and by , $|J_{{\xi}}|\gtrsim 2^{m_{1}}$. We can write as $$\sum_{m_{1}\geq 0}\sum_{0\leq j\leq m_{1}}\sum_{m_{2},j_{1},j_{2}\geq 0}
2^{-j/2}2^{m_{1}}\int \chi_{1}\chi_{j}g_{j,m_{1}}(u,v,w)|J_{{\xi}}|^{-1}
H(u,v,w, \xi_{1},\theta_{1},\theta_{2})dudvdwd\xi_{1}d\theta_{1}
d\theta_{2},$$ where $H(u,v,w,\xi_{1},\theta_{1},\theta_{2})$ is the transformation of $$\chi_{j}\chi_{1}\chi_{j_{1}}\phi_{1,m_{1}}({\xi}_{1},{\mu}_{1},\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1}))\chi_{j_{2}}|\hat{v}|({\xi}_{2},{\mu}_{2},
\theta_{2}+
{\omega}({\xi}_{2},{\mu}_{2})) \theta_{m_2} (\xi_2 )$$ under the above change of variables. We will define $$\widehat{v}_{j_2 , m_2 } (\xi_2 , \mu_2 , \tau_2 ) = \chi_{j_2} (\tau_2 -
\omega ( \xi_2 , \mu_2 )) \widehat{v} ( \xi_2 , \mu_2 , \tau_2 ) \theta_{m_2 } (\xi_2 ).$$ We can then continue the estimate with $$\begin{aligned}
&\lesssim&\sum_{m_{1}\geq 0}\sum_{0\leq j\leq m_{1}}
\sum_{m_{2},j_{1},j_{2}\geq 0}
2^{-j/2}2^{m_{1}/2}2^{(j-m_{1}-m_{2})/2}\int
\chi_{j}\chi_{1}g_{j,m_{1}}(u,v,w)\\
&&\left(\int H^{2}(u,v,w, \xi_{1},\theta_{1},\theta_{2})
|J_{{\xi}}|^{-1}d{\xi}_{1}\right)^{{\frac{1}{2}}}dudvdwd\theta_{1}
d\theta_{2}\\
&\lesssim&\sum_{0\leq j\leq m_{1}}\sum_{m_{2}, m_{1},j_{1},j_{2}\geq 0}
2^{-j/2}2^{m_{1}/2}2^{(j-m_{1}-m_{2})/2}
\|g_{j}\theta_{m_{1}}\chi_{j}\chi_{1}\|_{L^{2}}2^{j_{1}/2}2^{j_{2}/2}\\
& &\left(\int
|\chi_{1}\phi_{m_{1},j_{1}}
(\xi_{1},{\mu}_{1},\theta_{1}+{\omega}({\xi}_{1},{\mu}_{1}))|^{2}
|\hat{v}_{m_{2},j_{2}}
(\xi_{2},{\mu}_{2},\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))|^{2}\right.\\
&&\left.\chi_{j}(\theta_{1}+
{\omega}({\xi}_{1},{\mu}_{1})+\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})
-{\omega}({\xi}_{1}+{\xi}_{2},{\mu}_{1}+{\mu}_{2}))\right)^{{\frac{1}{2}}}.\end{aligned}$$ We now use Hölder inequality in ${\mu}_{2}$ and Lemma \[sobolev\]. More precisely set $w({\xi},{\mu})=(1+|{\xi}|+\gi)$, then we write, for $\theta=(p-2)/2p, \, p=2r, \, r>1, 1 = \frac{1}{r} + \frac{1}{r'}$, $$\int |\hat{v}_{m_{2},j_{2}}|^{2}\chi_{j}\leq
|\Delta_{{\mu}_{2}}|^{1/r'}\|\hat{v}_{m_{2},j_{2}}\|_{L^{2r}}^{2}
\lesssim 2^{(j-m_{1})/r'}\|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}
\|_{L^{2}}^{2(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}\hat{v}_{m_{2},j_{2}}
\|_{L^{2}}^{2\theta}.
\label{rr'}$$ We insert this in the chain of inequalities above and we continue with $$\begin{aligned}
&\lesssim&\sum_{m_{1}\geq 0}\sum_{0\leq j\leq m_{1}}\sum_{m_{2}\geq 0}
\sum_{j_{1},j_{2}\geq 0}
2^{-j/2}2^{m_{1}/2}2^{(j-m_{1}-m_{2})/2}2^{j_{1}/2}2^{j_{2}/2}
2^{(j-m_{1})/2r'}\|g_{j}\theta_{m_{1}}\chi_{1}\chi_{j}\|_{L^{2}}
\|\chi_{1}\phi_{m_{1},j_{1}}\|_{L^{2}}\\
&&\left(\int \|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}
(\xi_{2},{\mu}_{2},\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))
\|_{L^{2}_{{\mu}_{2}}}^{2(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}
\hat{v}_{m_{2},j_{2}}
(\xi_{2},{\mu}_{2},\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2}))
\|_{L^{2}_{{\mu}_{2}}}^{2\theta}
d\xi_{2}d\theta_{2}\right)^{{\frac{1}{2}}},\end{aligned}$$ and after applying Hölder’s inequality with respect to $\xi_{2}, \theta_{2}$ with exponents $\theta^{-1}$ and $(1-\theta)^{-1}$, $$\begin{aligned}
&\lesssim&\sum_{m_{1}\geq 0}\sum_{0\leq j\leq m_{1}}\sum_{m_{2}\geq 0}
\sum_{j_{1},j_{2}\geq 0}
2^{-j/2}2^{m_{1}/2}2^{(j-m_{1}-m_{2})/2}2^{j_{1}/2}2^{j_{2}/2}
2^{(j-m_{1})/2r'}\|g_{j}\theta_{m_{1}}\chi_{1}\chi_{j}\|_{L^{2}}\\
&&\|\chi_{1}\phi_{m_{1},j_{1}}\|_{L^{2}}
\|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}
\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{\theta}.\end{aligned}$$ We first sum on $0\leq j\leq m_{1}$, then on $m_{1}$ and $j_{1}$ so that the norm $\|u\|_{X_{1-\epsilon_{0}},{\frac{1}{2}}}$ appears. After Cauchy-Schwarz in $m_{2}$ we are left with the following term to estimate $$\sum_{j_{2}\geq 0}2^{j_{2}/2}\left(\sum_{m_{2}\geq 0}
\|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{2(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}
\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{2\theta}\right)^{{\frac{1}{2}}}.$$ We use a Hölder inequality with respect to the sum on $m_{2}$ to obtain $$\begin{aligned}
&\leq &\sum_{j_{2}\geq 0}2^{j_{2}/2}
\left(\sum_{m_{2}\geq 0}
\|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{2}
\right)^{(1-\theta)/2}
\left(\sum_{m_{2}\geq 0}\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}
\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{2}\right)^{\theta/2}\\
&=&\sum_{j_{2}\geq 0}2^{j_{2}/2}\|w^{\epsilon_{0}}\hat{v}_{j_{2}}
\|_{L^{2}}^{(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}\hat{v}_{j_{2}}\|_{L^{2}}^{\theta},\end{aligned}$$ where $$\widehat{v}_{j_2} = \sum_{m_2 \geq 0} \widehat{v}_{j_2 , m_2 } = \chi_{j_2}
( \tau - \omega ( \xi_2 , \mu_2 )) \widehat{v} ( \xi_2 , \mu_2 , \tau_2 ).$$ then a Hölder inequality in $j_{2}$ to finish with $$\left(\sum_{j_{2}\geq 0}2^{j_{2}/2}\|w^{\epsilon_{0}}
\hat{v}_{j_{2}}\|_{L^{2}}
\right)^{(1-\theta)}
\left(\sum_{j_{2}\geq 0}2^{j_{2}/2}\|w^{-\epsilon_{0}}
\partial_{{\mu}_{2}}
\hat{v}_{j_{2}}\|_{L^{2}}\right)^{\theta}.
\label{a6step12}$$ Clearly the first coefficient of is controlled by $\|v\|_{X_{\epsilon_{0},{\frac{1}{2}}}}^{1-\theta}$. For the second one we write $$\begin{aligned}
&&\partial_{{\mu}_{2}}(\hat{v}_{j_{2}}({\xi}_{2},{\mu}_{2},
\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})))=\partial_{{\mu}_{2}}
(\chi_{j_{2}}(\theta_{2})\hat{v}({\xi}_{2},{\mu}_{2},
\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})))\\
&=&\chi_{j_{2}}\partial_{{\mu}_{2}}
\hat{v}({\xi}_{2},{\mu}_{2},
\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})+
2{\mu}_{2}/{\xi}_{2}\chi_{j_{2}}\partial_{\tau_{2}}
\hat{v}({\xi}_{2},{\mu}_{2},
\theta_{2}+{\omega}({\xi}_{2},{\mu}_{2})
\end{aligned}$$ which shows that the second term is controlled by $\|v\|_{Y_{1-\epsilon_{0},-\epsilon_{0},{\frac{1}{2}}}}^{\theta}$, with $\theta=(p-2)/2p, \, p=2r, \, r>1$.\
[**Case C2b:**]{} $m_{2}< 0$\
If $j-m_{1}\leq 2m_{2}$ we proceed like in Case C2a and we obtain $$\begin{aligned}
&\lesssim&\sum_{m_{1}\geq 0}\sum_{m_{2}<0}\sum_{0\leq j\leq m_{1}+2m_{2}}
\sum_{j_{1},j_{2}\geq 0}
2^{-j/2}2^{m_{1}/2}2^{(j-m_{1}-m_{2})/2}2^{j_{1}/2}2^{j_{2}/2}
2^{(j-m_{1})/2r'}\|g_{j}\theta_{m_{1}}\chi_{1}\chi_{j}\|_{L^{2}}
\|\chi_{1}\phi_{m_{1},j_{1}}\|_{L^{2}}\\
&&\|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}
\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{\theta}.\end{aligned}$$ We sum on $j$, and then choose $r'<2$ so that we can use the fact that $m_{2}<0$ and Cauchy-Schwarz in $m_{2}$, to finish like in Case C2a. If $j-m_{1}> 2m_{2}$ we use the change of variables , where we leave the variable ${\xi}_{2}$ free. It is easy to check that we have $|J_{{\xi}}|\gtrsim |{\xi}_{1}|$. Arguing as in Case C2a we are led to $$\begin{aligned}
&\lesssim&\sum_{m_{1}\geq 0}\sum_{0\leq j\leq m_{1}}
\sum_{m_{2}<(j-m_{1})/2}\sum_{j_{1},j_{2}\geq 0}
2^{(m_{1}-j)/2}2^{j_{1}/2}2^{j_{2}/2}
2^{(j-m_{1})/2r'}2^{m_{2}/2}\|g_{j,m_{1}}\chi_{1}\chi_{j}\|_{L^{2}}
\|\chi_{1}\phi_{m_{1},j_{1}}\|_{L^{2}}\\
&&\|w^{\epsilon_{0}}\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{(1-\theta)}
\|w^{-\epsilon_{0}}\partial_{{\mu}_{2}}
\hat{v}_{m_{2},j_{2}}\|_{L^{2}}^{\theta}.\end{aligned}$$ We first apply Cauchy-Schwarz with respect to $m_{2}$ and we obtain an extra factor $2^{(-m_{1}+j)/4}$. If we now have that $\frac{1}{4}+ \frac{1}{2r'}> {\frac{1}{2}}$, we can then sum in $j$, and repeat the argument in Case C2a. This is again the restriction $r'<2 $, which when we go to gives $p=2r>4$, and hence $\theta>\frac{1}{4}$.
This concludes the proof of Theorem \[xbil\].
We first check the part of the statement involving the term $\partial_{\tau}(\widehat{\partial_{x}(uv)})$. We proceed by writing $$\begin{aligned}
& &\partial_{\tau}\left({\xi}\int
\hat{v}({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1})\hat{u}
({\xi}_{1},{\mu}_{1},\tau_{1})d{\xi}_{1}d{\mu}_{1}d\tau_{1}\right)\\
&=&\partial_{\tau}\left({\xi}\int_{|{\xi}-{\xi}_{1}|\leq |{\xi}_{1}|}
\hat{v}({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1})\hat{u}
({\xi}_{1},{\mu}_{1},\tau_{1})d{\xi}_{1}d{\mu}_{1}d\tau_{1}\right)\\
&+&\partial_{\tau}\left({\xi}\int_{|{\xi}-{\xi}_{1}|\geq |{\xi}_{1}|}
\hat{v}({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1})\hat{u}
({\xi}_{1},{\mu}_{1},\tau_{1})d{\xi}_{1}d{\mu}_{1}d\tau_{1}\right)\\
&=&{\xi}\int_{|{\xi}-{\xi}_{1}|\leq |{\xi}_{1}|}
\hat{v}({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1})\partial_{\tau_{1}}\hat{u}
({\xi}_{1},{\mu}_{1},\tau_{1})d{\xi}_{1}d{\mu}_{1}d\tau_{1}\\
&+&{\xi}\int_{|{\xi}-{\xi}_{1}|\geq |{\xi}_{1}|}
\partial_{\tau}\hat{v}({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1})\hat{u}
({\xi}_{1},{\mu}_{1},\tau_{1})d{\xi}_{1}d{\mu}_{1}d\tau_{1}.
\end{aligned}$$ Then the estimates for each one of these two terms follows from the proof of Theorem \[xbil\]. Next we check the part involving $\partial_{{\mu}}
\widehat{\partial_{x}(uv)}$. If we proceed as above, we only need to check $$I({\xi},{\mu},\tau)=\frac{|{\xi}|}{\max(1,|{\xi}|,|{\mu}|/|{\xi}|)^{\epsilon_{0}}}
\int_{|{\xi}-{\xi}_{1}|\leq |{\xi}_{1}|}
|\hat{v}|({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1})|\partial_{{\mu}_{1}}\hat{u}|
({\xi}_{1},{\mu}_{1},\tau_{1})d{\xi}_{1}d{\mu}_{1}d\tau_{1}.$$ We introduce the two functions $$\begin{aligned}
\label{normalizu1} \phi_{1, j_1}({\xi},{\mu},\tau)&=&
\max(1,|{\xi}|,|{\mu}|/|{\xi}|)^{-\epsilon_{0}}|\partial_{{\mu}}\hat{u}|
\chi_{j_1 }({\xi},{\mu},\tau),\\
\label{normalizv1}
\phi_{2, j_2}({\xi},{\mu},\tau)&=&\max(1,|{\xi}|,|{\mu}|/|{\xi}|)^{1-\epsilon_{0}}
|\hat{v}|\chi_{j_2 }({\xi},{\mu},\tau).\end{aligned}$$ corresponding to and in Theorem \[xbil\]. We now observe that in the region $|{\xi}-{\xi}_{1}|\leq
|{\xi}_{1}|$ we also have $\gia\leq |{\mu}-{\mu}_{1}|/|{\xi}-{\xi}_{1}| +
2\gi$, so $$I({\xi},{\mu},\tau)\lesssim J^{1}+J^{2},
\label{ybilstep1}$$ where $$\begin{aligned}
\label{j1}J^{1}&=&\frac{|{\xi}|}{\max(1,|{\xi}|,\gi)^{\epsilon_{0}}}
\int_{*}\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}
{\max(1,|{\xi}-{\xi}_{1}|,|{\mu}-{\mu}_{1}|/|{\xi}-{\xi}_{1}|)^{1-2\epsilon_{0}}}\\
\label{j2}J^{2}&=&|{\xi}|
\int_{*}\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}
{\max(1,|{\xi}-{\xi}_{1}|,|{\mu}-{\mu}_{1}|/|{\xi}-{\xi}_{1}|)^{1-\epsilon_{0}}}.\end{aligned}$$ where $\int_{*}$ is the integral over the region given by $|{\xi}-{\xi}_{1}|\leq |{\xi}_{1}|$. Also notice that if $$\tilde J^{2}=|{\xi}|
\int_{*}\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}
{\max(1,|{\xi}-{\xi}_{1}|,|{\mu}-{\mu}_{1}|/|{\xi}-{\xi}_{1}|)^{1-2\epsilon_{0}}},
\label{tildej2}$$ then $$J^{1}+J^{2}\leq \tilde J^{2}.
\label{j1j2}$$ By duality we need to estimate $$\begin{aligned}
\label{ybildualx}& &\sum_{i=1,2}\sum_{j\geq 0}\sum_{m\geq 0}
2^{-j/2} \int g_{j}({\xi},{\mu},\tau)J^{i}({\xi},{\mu},\tau)\theta_{m}({\xi})
\chi_{1}({\xi},{\mu})\chi_{j}(\tau-{\omega}({\xi},{\mu}))d{\xi}d{\mu}d\tau\\
\label{ybildualm}& &\sum_{i=1,2} \sum_{j\geq 0}
\sum_{n\geq 0}2^{-j/2} \int g_{j}({\xi},{\mu},\tau)J^{i}({\xi},{\mu},\tau) \theta_{n}({\mu})
\chi_{2}({\xi},{\mu})\chi_{j}(\tau-{\omega}({\xi},{\mu})) d{\xi}d{\mu}d\tau.\end{aligned}$$ We begin a case by case analysis for and . We introduce again the notation ${\mu}={\mu}_{1}+{\mu}_{2}$ and ${\xi}={\xi}_{1}+{\xi}_{2}$.\
[**Region $A_{1}$**]{}. In this region we use and we bound and by replacing $J^{i}, i=1,2$ with $\tilde J^{2}$ in .\
[**Case A:**]{} $|{\xi}|\geq {\frac{1}{2}}\gi$.\
We only have to estimate . We observe that $|{\xi}|\leq 2$ and hence, from , $$\tilde J^{2}\lesssim \int_{*}\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\phi_{2,j_{2}}({\xi}-{\xi}_{1},{\mu}-{\mu}_{1},\tau-\tau_{1}).$$ Then if we dyadically decompose for $1+ |{\xi}|\sim 2^{m}$, the sum in $m$ is a finite sum and we can use Strichartz inequality , applied to $\phi_{1, j_1 }$ and $\phi_{2, j_2}$, and Plancherel and Hölder, to prove the theorem in this case.\
[**Case B:**]{} $|{\xi}|< {\frac{1}{2}}\gi$.\
We have to estimate . If $\gi\leq 1$, it follows that $|{\mu}_{1}+{\mu}_{2}|\leq 2$, and we proceed like in Case A. If $\gi\geq 1$. We consider two subcases.\
[**Case B1:**]{} $|{\mu}_{1}|\lesssim |{\mu}_{2}|$.\
If ${\frac{1}{2}}\gib\leq |{\xi}_{2}|$, because $|{\xi}_{2}|\leq 1$, it follows that $|{\mu}_{1}|, |{\mu}_{2}|\leq 4$ and we go back to the previous case. If ${\frac{1}{2}}\gib> |{\xi}_{2}|$, we dyadically decompose with respect to ${\mu}_{2}$. Then $$|{\xi}_{1}+{\xi}_{2}|\leq |{\xi}_{1}+{\xi}_{2}|^{2\epsilon_{0}}
|{\mu}_{1}+{\mu}_{2}|^{1-2\epsilon_{0}}\leq 2|{\mu}_{1}+
{\mu}_{2}|^{1-2\epsilon_{0}}$$ this corresponds to Case B1 of region $A_{1}$ in the proof of Theorem \[xbil\]. In fact we have the following bound for : $$\begin{aligned}
& &\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{n_{2}\geq 0}
\sum_{0\leq r\leq n_{2}}
2^{-j/2}\int_{A_{1}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_2 + 1 -r}2^{(n_{2}+1-r)(1-2\epsilon_{0})}\\
&&
\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\chi_{2}\phi_{2,j_{2},n_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{(\gib)^{1-2\epsilon_{0}}},\end{aligned}$$ where $\phi_{2, j_2 , n_2} = \theta_{n_2} ( \mu_2 ) \phi_{2, j_2}$. (Note that $\phi_{1, j_1} , \phi_{2, j_2}$ are in this case as defined in , , and one concludes like in that case for $\epsilon_{0}<\frac{1}{4}$.\
[**Case B2:**]{} $|{\mu}_{1}|>>|{\mu}_{2}|$.\
Then $|{\mu}_{1}|\sim |{\mu}_{1}+{\mu}_{2}|$. If ${\frac{1}{2}}\gia\leq |{\xi}_{1}|$, it follows that $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|\leq 4$. So if we set $1+ |{\mu}_{1}+{\mu}_{2}|\sim
2^{n}$, then the sum on $n$ is finite and we use the Strichartz inequality . If ${\frac{1}{2}}\gia> |{\xi}_{1}|$, we dyadically decompose $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|\sim 2^{n_{1}}$. Then we use Strichartz inequality again.
[**Region $A_{2}$**]{}.\
[**Case A:**]{} $|{\xi}|\geq {\frac{1}{2}}\gi$.\
This case is similar to the corresponding case in Theorem \[xbil\]. If $|{\xi}_{1}|\sim |{\xi}_{2}|\sim 2^{m_{1}}$ and $|{\xi}_{1}+{\xi}_{2}|\sim 2^{m_{1}+1-r}, 0\leq r<m_{1}$, (and $|{\xi}_{1}+{\xi}_{2}|\leq 1$ when $r=m_{1}$), then one can bound with $$\sup_{j,m}\|g_{j}\chi_{1}\chi_{j}\theta_{m}\|_{L^{2}}
\sum_{m_1 \geq 0}\sum_{0\leq r \leq m_{1}}2^{-r(1-2\epsilon_{0})}
\Pi_{i=1,2}\sum_{j_{i}\geq 0}2^{j_{i}/2}\|\phi_{i,j_{i},m_{1}}\|_{L^{2}},$$ and this proves the estimate for $\epsilon_{0}<\frac{1}{4}$.\
[**Case B:**]{} $|{\xi}|< {\frac{1}{2}}\gi$.\
If $\gi\leq 1$, then $|{\xi}|, |{\mu}|\leq 4$ and we go back to the estimate in region $A_{1}$. So we assume that $\gi> 1$. We consider two subcases.\
[**Case B1:**]{} $|{\mu}_{1}| \lesssim |{\mu}_{2}|$\
If ${\frac{1}{2}}\gib\leq |{\xi}_{2}|$ we use the fact that $$|{\xi}_{1}+{\xi}_{2}|\leq
1/\sqrt{2}|{\mu}_{1}+{\mu}_{2}|^{{\frac{1}{2}}}\leq C|{\mu}_{2}|^{{\frac{1}{2}}}\leq C|{\xi}_{2}|,
\label{estimate}$$ and we dyadically decompose with respect to ${\xi}_{2}$. Then if $|{\xi}_{2}|\sim 2^{m_{2}}$, $1+ |{\mu}_{1}+{\mu}_{2}|\sim 2^{-r}2^{2m_{2}}, \, \,
0\leq r\leq 2m_{2}$. We proceed now like in the corresponding case for Theorem \[xbil\] where we replace $\epsilon_{0}$ by $2\epsilon_{0}$ and $r$ by $r/2$. If ${\frac{1}{2}}\gib> |{\xi}_{2}|$ we dyadically decompose with respect to ${\mu}_{2}$ by setting $|{\mu}_{2}|\sim 2^{n_{2}}$. Then $|{\mu}_{1}+{\mu}_{2}|\sim 2^{n_{2}-r+1}, \, \, 0\leq r\leq n_{2}$ and from also $|{\xi}_{1}+{\xi}_{2}|\lesssim 2^{n_{2}/2-r/2+1/2}$. We reduce our estimate to $$\begin{aligned}
& & \sum_{j_{1},j_{2}\geq 0}
\sum_{j\geq 0}\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_2 + 1 -r}2^{(n_{2}+1-r)(1-2\epsilon_{0})/2}\\
& &\times|{\xi}_{1}+{\xi}_{2}|^{2\epsilon_{0}}
\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\chi_{2}\phi_{2,j_{2},n_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})}
{(\gib)^{1-2\epsilon_{0}}}\\
&\lesssim&\sum_{j_{1},j_{2}\geq 0}
\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}\sum_{j\geq 0}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_2 + 1 -r}2^{-r(1-2\epsilon_{0})/2}\\
& &\times
|{\xi}_{1}+{\xi}_{2}|^{2\epsilon_{0}}
\phi_{1,j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{|{\xi}_{2}|^{1-2\epsilon_{0}}}{|{\mu}_{2}|^{(1-2\epsilon_{0})/2}}
\chi_{2}\phi_{2,j_{2},n_{2}}({\xi}_{2},{\mu}_{2},\tau_{2})\end{aligned}$$ We use again Lemma \[mainlemma\] and, in view of the fact that here $|{\xi}_{2}|^{2}/|{\mu}_{2}|\leq {\frac{1}{2}}$, we continue with $$\sup_{j,n}\|g_{j}\chi_{2}\chi_{j}
\theta_{n}\|_{L^{2}}\sum_{n_{2}\geq 0}\sum_{0\leq r\leq n_{2}}
2^{-r(1-2\epsilon_{0})/2}\left(\sum_{j_{1}\geq 0}2^{j_{1}/2}
\|\phi_{1,j_{1}}\|_{L^{2}}\right)
\left(\sum_{j_{2}\geq 0}2^{j_{2}/2}
\|\chi_{2}\phi_{2,n_{2},j_{2}}\|_{L^{2}}\right),$$ and we sum in $r$. This is enough to prove the estimate as long as $\epsilon_{0}<\frac{1}{4}$.\
[**Case B2:**]{} $|{\mu}_{1}|>>|{\mu}_{2}|$.\
If ${\frac{1}{2}}\gia\leq |{\xi}_{1}|$, then $$|{\xi}_{1}+{\xi}_{2}|\leq
1/\sqrt{2}|{\mu}_{1}+{\mu}_{2}|^{{\frac{1}{2}}}\leq C|{\mu}_{1}|^{{\frac{1}{2}}}\leq C|{\xi}_{1}|,$$ replaces and we can repeat the argument given in the first part of Case B1. If ${\frac{1}{2}}\gia> |{\xi}_{1}|$, since $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|$, we dyadically decompose with respect to $|{\mu}_{1}|\sim 2^{n_{1}}$. Because $|{\xi}_{1}+{\xi}_{2}|\lesssim |{\xi}_{2}|$, we reduce our estimate to $$\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{n_{1}\geq 0}
2^{-j/2}\int_{A_{2}}g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{1}}
|{\xi}_{1}+{\xi}_{2}|^{2\epsilon_{0}}
\chi_{2}\phi_{1,n_{1},j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\phi_{2,j_{2}}({\xi}_{2},{\mu}_{2},\tau_{2}),
\label{estimate1}$$ and at this point we can proceed by using Lemma \[mainlemma\].
[**Region $A_{3}\cup A_{4} \cup (A_{5}-\tilde{A}_{5}( 2 \epsilon_{0}))$**]{}.\
[**Case A:**]{} $ {\frac{1}{2}}\gi\leq|{\xi}|$.\
We consider two subcases.\
[**Case A1:**]{} $|\xi_{1}|\geq {\frac{1}{2}}\gia$.\
We dyadically decompose with respect to $|{\xi}|\sim|{\xi}_{1}|\sim 2^{m_{1}}$. Because in this region $\gib\gtrsim |\xi_{1}|\sim |{\xi}_{1}+{\xi}_{2}|$, we reduce our estimate to of the corresponding case in the proof of Theorem \[xbil\], where $\epsilon_{0}$ is replaced by $2\epsilon_{0}$.\
[**Case A2:**]{} $|\xi_{1}|\leq {\frac{1}{2}}\gia$.\
We use again a dyadic decomposition with respect to ${\xi}_{1}$. Observe that $(|{\xi}_{2}|<<)|{\xi}_{1}|\leq {\frac{1}{2}}\gia \leq C\gib$ hence $\gib \sim 2^{m_{1}+r}, r\geq
-C_{2}$. We then reduce our estimate to , again replacing $\epsilon_{0}$ with $2\epsilon_{0}$.\
[**Case B:**]{} $ {\frac{1}{2}}\gi>|{\xi}|$. As we observed in the analysis of region $A_{i}, i=1,2$, without loss of generality we can assume that $\gi\geq 1$. We consider two subcases.\
[**Case B1:**]{} $|{\mu}_{1}|\lesssim |{\mu}_{2}|$.\
We recall that in this region we also have $\gib>>|{\xi}_{2}|$ and we repeat the argument presented in the second part of Case B1 of region $A_{2}$.\
[**Case B2:**]{} $|{\mu}_{1}|>> |{\mu}_{2}|$.\
If ${\frac{1}{2}}\gia\leq|{\xi}_{1}|$ then we use an argument similar to the one used in the first part of Case B1 in region $A_{2}$, where the role of $({\xi}_{2},{\mu}_{2})$ is now played by $({\xi}_{1},{\mu}_{1})$. If ${\frac{1}{2}}\gia \geq |{\xi}_{1}|$, we dyadically decompose so that $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|\sim 2^{n_{1}}$. Then because $|{\xi}_{1}+{\xi}_{2}|\sim |{\xi}_{1}|\lesssim \gib$, we obtain the estimate and also this case is done.
[**Region $A_{5}\cap\tilde{A}_{5}(2\epsilon_{0})$**]{}. In this case we cannot use Lemma \[mainlemma\].\
[**Case A:**]{} $|{\xi}|\geq {\frac{1}{2}}\gi$.\
We observe that if [^12] $\epsilon_{0}<(\alpha -1)/2\alpha$, then $(1+\alpha 2\epsilon_{0})(1-2\epsilon_{0})>1$ and $$(\gia)^{1-2\epsilon_{0}}\geq |{\xi}_{1}|^{1+\delta}\sim
|{\xi}_{1}+{\xi}_{2}|^{1+\delta},
\label{estimate3}$$ for some $\delta$ such that $0<\delta<<1$. We dyadically decompose so that $|{\xi}_{1}|\sim
|{\xi}_{1}+{\xi}_{2}|\sim 2^{m_{1}}$ and using we reduce the estimate to $$\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{m_{1}\geq 0}
2^{-j/2}\int g_{j}\chi_{1}\chi_{j}({\xi},{\mu},\tau)
\theta_{m_{1}}2^{-m_{1}\delta}
\chi_{1}\phi_{1,j_{1},m_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\phi_{2,j_{2}}({\xi}_{2},{\mu}_{2},\tau_{2}),$$ and Strichartz inequality and Cauchy-Schwarz can be used to finish the proof also in this case.\
[**Case B:**]{} $|{\xi}|\leq {\frac{1}{2}}\gi$.\
Notice that from $|{\mu}_{2}|\sim |{\xi}_{2}|\gia<<|{\mu}_{1}|$, hence $|{\mu}_{1}+{\mu}_{2}|\sim |{\mu}_{1}|$. We dyadically decompose with respect to $|{\mu}_{1}|\sim 2^{n_{1}}$. Using we obtain the estimate $$\begin{aligned}
& &\sum_{j_{1},j_{2}\geq 0}\sum_{j\geq 0}\sum_{n_{1}\geq 0}2^{-j/2}
\int g_{j}\chi_{2}\chi_{j}({\xi},{\mu},\tau)
\theta_{n_{1}}({\mu})
|{\xi}_{1}|^{-\delta}(\gia)^{1-2\epsilon_{0}}\\
& &\times
\chi_{2}\phi_{1,n_{1},j_{1}}({\xi}_{1},{\mu}_{1},\tau_{1})
\frac{\phi_{2,j_{2}}}
{(\gib)^{1-2\epsilon_{0}}}({\xi}_{2},{\mu}_{2},\tau_{2}).\end{aligned}$$ Then Strichartz inequality is enough to conclude the proof also in this case.
[**Region $A_{6}$**]{}. As we did for the proof of Theorem \[xbil\], also in this case we consider three subcases: [**Case A:**]{} $|{\xi}_{1}|\geq 10^{2}\gia$; [**Case B:**]{} $|{\xi}_{1}|< 10^{-2}\gia$ and [**Case C:**]{} $10^{-2}\gia<|{\xi}_{1}|< 10^{2}\gia$. Since in the proof of Theorem \[xbil\] we used in cases A and B, we can treat these cases in the same way, with the understanding that now $\epsilon_{0}$ is replaced by $2\epsilon_{0}$. For Case C we go back to and and we show that in this region $$J^{1}\lesssim J^{2}.
\label{j1<j2}$$ If one assumes this for a moment, then it is easy to see that we can repeat exactly the argument we gave for Case C in the proof of Theorem \[xbil\]. To prove it’s enough to show that $$\max(|{\xi}|, \gi)\gtrsim \max(1,|{\xi}_{2}|, \gib).$$ To simplify the notation we set $\max(|{\xi}|, \gi)=M$ and $\max(1,|{\xi}_{2}|, \gib)=M_{2}$. Assume that $M=|{\xi}|$. If $M_{2}=1$, then $M_{2}=1\lesssim |{\xi}|=M$. If $M_{2}=|{\xi}_{2}|$, then $M\sim
|{\xi}_{1}|>>|{\xi}_{2}|=M_{2}$ and also this case is done. If $M_{2}=\gib$, then $M=|{\xi}|\sim |{\xi}_{1}|\sim \gia>>
\gib=M_{2}$. Now assume that $M=\gi$. If $M_{2}=1$, then $M_{2}=1\lesssim |{\xi}|\leq \gi=M$. If $M_{2}=|{\xi}_{2}|$, then $M_{2}<<
|{\xi}_{1}|\sim |{\xi}|\leq \gi=M_{2}$ and also this case is done. Finally If $M_{2}=\gib$, then $M_{2}<< \gia\sim |{\xi}_{1}|\sim |{\xi}|
\leq \gi=M$.
The proof of Theorem \[ybil\] is now complete.
We conclude this section with a counterexample that shows that if $\epsilon<\frac{1}{4}$ in Theorem \[xbil\], then the theorem does not hold. This counterexample is important because, as we will discuss below in Remark \[rescaling\], if we could have taken $\epsilon<\frac{1}{4}$, then we could have removed the smallness assumption in the initial data and at the same time we would have obtained a global result in the modified energy space $E\cap P$.
\[caunter\] The bilinear estimate $$\begin{aligned}
\label{bilinearcounter}\|\partial_{x}(uv)\|_{X_{1,-{\frac{1}{2}}}}&\leq& C
\|u\|_{X_{1,{\frac{1}{2}}}}(\|v\|_{X_{1,{\frac{1}{2}}}}+
\|v\|_{X_{1,{\frac{1}{2}}}}^{1-\epsilon }
\|v\|_{Y_{1,0, {\frac{1}{2}}}}^{\epsilon })\\
\nonumber&+&C
\|v\|_{X_{1,{\frac{1}{2}}}}(\|u\|_{X_{1, {\frac{1}{2}}}}+
\|u\|_{X_{1, {\frac{1}{2}}}}^{1-\epsilon }
\|u\|_{Y_{1,0, {\frac{1}{2}}}}^{\epsilon })\end{aligned}$$ fails for $\epsilon < \frac{1}{4}$.
The proof of the proposition is based on the example of Molinet, Saut and Tzvetkov from [@MST; @MST1]. We introduce the sets: $$E_1 = \{ (\xi , \mu , \tau): \xi \in [{\frac{1}{2}}\alpha, \alpha],~ \mu \in [
-6 \alpha^2 , 6 \alpha^2 ], ~|\tau - \omega ( \xi , \mu ) |
\lesssim 1
\}$$ $$E_2 = \{ (\xi , \mu , \tau): \xi \in [N , N +\alpha],~ \mu \in [
\sqrt{3} N^2 , \sqrt{3} N^2 + \alpha^2 ],
~|\tau - \omega ( \xi , \mu ) | \lesssim 1
\}.$$ where again $\omega (\xi , \mu)= \xi^3 + \frac{\mu^2}{\xi}$. We observe that $\partial_\xi \omega (\xi, \mu)
= 3 \xi^2 - \frac{\mu^{2}}{\xi^2}$, and $\partial_\mu \omega (\xi , \mu) =
\frac{2\mu}{\xi}.$ The reason for the $\sqrt{3}$ in the definition of $E_2$ may be seen by calculating $\partial_\xi \omega $ inside $E_2$. The $O(N^2)$ terms cancel and the next biggest term is $O(N \alpha )$. In the region $E_2$, the dispersive surface has slope $O(N \alpha)$ along the $\xi$ direction and slope $O(N)$ along $\mu$. An $\alpha \times \alpha^2$ piece of the tangent plane to the dispersive surface in $E_2$ stays within $N \alpha^2 $ of the surface. Therefore, we select $\alpha \thicksim N^{-{\frac{1}{2}}}$ so that $E_2$ is a $O(1)$ vertical thickening of this piece of tangent plane. These calculations then “explain” the choice of the $\alpha$, $\alpha^2$ scaling in the $\xi$ and $\mu$ directions. Note that $E_1$ is essentially an $\alpha \times \alpha^2 \times 1$ box.
Let $\widehat{u}= \alpha^{-\frac{3}{2}} \chi_{E_1}, ~ \widehat{v}
= N^{-1} \alpha^{-\frac{3}{2}} \chi_{E_2},$ where the functions $\chi_{E_i}$ are smoothed out characteristic functions.
We calculate $$| [\partial_x (u v)]\sphat(\xi, \mu, \tau) | \thicksim |\xi| |\widehat{u} *
\widehat{v} | ( \xi , \mu , \tau ) \thicksim |\xi | \alpha^{-3} N^{-1}
~\chi_{E_1}* \chi_{E_2} (\xi, \mu , \tau ).$$ We have $$\chi_{E_1}* \chi_{E_2} (\xi, \mu , \tau ) \thicksim \sup_{{trans}}
|{{trans}} (E_1) \cap E_2 | ~\chi_{E_1 + E_2 } ( \xi , \mu , \tau ),$$ where ${{trans}}$ denotes an arbitrary translation in the $(\xi, \mu , \tau)$ space. Geometric considerations similar to those discussed above show that $E_1 +E_2 $ contains an $\alpha \times \alpha^2 \times
N \alpha^2 $ box containing the point $(N+\alpha, \sqrt{3} N^2 + \alpha^2 , 4 N^3 )$. This point is at vertical distance $O(1)$ from the dispersive surface. Since $E_2$ has slope $N \alpha$ along $\xi$ and slope $N$ along $\mu$, we observe that the sup above is bounded by $(N \alpha)^{-1} \times N^{-1} \times 1$. Combining these remarks gives, $$| {\widehat{\partial_x u v}} | \thicksim
N ~ \alpha^{-3} N^{-1} ~ N^{-2} \alpha^{-1} \chi_{E_1 + E_2 } \thicksim
\alpha^{-4} N^{-2} \chi_{E_1 + E_2}.$$ Therefore, we have $$\|\partial_{x}(u v)\|_{X_{1,-{\frac{1}{2}}}} \gtrsim \alpha^{-4} N^{-2}~ N
(N \alpha^2 )^{-{\frac{1}{2}}}~ | E_1 + E_2 |^{{\frac{1}{2}}} \gtrsim \alpha^{-5}
N^{-1 - {\frac{1}{2}}} ~(\alpha^5 N )^{{\frac{1}{2}}}.$$ The choice of $\alpha = N^{-{\frac{1}{2}}}$ yields $$\|\partial_{x}(u v)\|_{X_{1,-{\frac{1}{2}}}} \gtrsim N^{\frac{1}{4}}.$$ as the size of the left-side of .
We now consider the right-side of . The functions $u,~
v$ are normalized to have size $O(1)$ in the various $X_{s,b}$-norms. The $Y_{1,0,{\frac{1}{2}}}$-norm has two pieces. The term arising from $\partial_\tau
u$ essentially reproduces $u$ since $E_1$ is of size $O(1)$ along the $\tau$ direction. The other term involves $\partial_\mu u$. Since $E_1$ has size $\alpha^2$ along $\mu$, we have that $|\partial_\mu
\chi_{E_2}| \thicksim \alpha^{-2} \chi_{E_2}$ so this part of the $Y_{1,0,{\frac{1}{2}}}$-norm is of size $O(\alpha^{-2}) = O( N )$. Upon taking this to the power $\epsilon$ and comparing with the size of the left-side, $N^{\frac{1}{4}}$, we see the failure of when $\epsilon <\frac{1}{4}.$
Proof of Theorem \[energytheorem\] and Theorem \[main0\]
=========================================================
We start with the proof of Theorem \[main0\], because Theorem \[energytheorem\] is a corollary of Theorem \[main0\]. The proof uses a classical fixed point theorem (see for example [@KPV3]). We will first carry out the proof when $T={\frac{1}{2}}$
We start by transforming into the integral equation $$\label{d} u=\psi(t)S(t)u_{0}-\psi(t)
\int_{0}^{t}
S(t-t')\partial_{x}(u^{2})(t') dt'.$$ (Here we have fixed $\beta = 1$.) Then it is clear that a solution for is a fixed point for the operator $$\label{oper}L(v)=\psi(t)S(t)u_{0}
-\psi(t)\int_{0}^{t}
S(t-t')\partial_{x}(v^{2})(t')dt'$$ To simplify the notation we set $Z_{1-\epsilon_{0}}=X_{1-\epsilon_{0},{\frac{1}{2}}}\cap
Y_{1-\epsilon_{0},- \epsilon_{0}, {\frac{1}{2}}}$. Then, for fixed $\sigma>0$, we assume that $\|u_{0}\|_{{B^{1,2}_{1-\epsilon_{0}}}\cap
P^{1,2}_{-\epsilon_{0}}}\leq \sigma$ and we set $a=4C\|u_{0}\|_{B^{1,2}_{1-\epsilon_{0}}\cap
P^{1,2}_{-\epsilon_{0}}}$, where $C$ is the constant in and . We show that, if $B_{a}$ is the ball centered at the origin and radius $a$ in $Z_{1-\epsilon_{0}}$, then $$\begin{aligned}
\label{ball-ball}& & L:B_{a}\longrightarrow B_{a},\\
\label{contr} & & \|L(u-v)\|_{Z_{1-\epsilon_{0}}}
\leq {\frac{1}{2}}\|u-v\|_{Z_{1-\epsilon_{0}}}
\end{aligned}$$ and this is enough to finish the proof of the theorem. To prove we use , , and to show that $$\|L(v)\|_{Z_{1-\epsilon_{0}}}\lesssim \frac{a}{4} +C_{1}
(\|\partial_{x}(v^{2})\|_{X_{1-\epsilon_{0},-{\frac{1}{2}}}}+
\|\partial_{x}(v^{2})\|_{Y_{1-\epsilon_{0},-\epsilon_{0},-{\frac{1}{2}}}})$$ and if we continue with Theorems \[xbil\] and \[ybil\] we obtain $$\begin{aligned}
\label{smallassumption}
& &\|L(v)\|_{Z_{1-\epsilon_{0}}}\lesssim \frac{a}{4}\\
\nonumber&&+C_{1}(\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}^{2}+
\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}
\|v\|_{Y_{1-\epsilon_{0},-\epsilon_{0},{\frac{1}{2}}}}+
\|v\|_{X_{1-\epsilon_{0},{\frac{1}{2}}}}^{1-\epsilon}
\|v\|_{Y_{1-\epsilon_{0},-\epsilon_{0},{\frac{1}{2}}}}^{1+\epsilon}),\end{aligned}$$ hence follows with our choice of $a$, for small $\sigma$.
We finish this section with a remark that should convince the reader that in a sense the fixed point method used above is performed in a critical regime. This criticality appears in an unusual way. This remark also shows how to obtain the case of general T in Theorem \[main0\], from the case $T={\frac{1}{2}}$. One simply chooses $T=\lambda^{3}/2$ below, $\lambda$ large, and then the norm small depending also on $\lambda$.
\[rescaling\]
If $u(x,y,t)$ is a solution of the IVP , then $$u_{\lambda}(x,y,t)=\lambda^{2}u(\lambda x,
\lambda^{2}y,\lambda^{3}t)
\label{rescsol}$$ is a solution for the IVP with initial data $u_{\lambda,0}(x,y)=\lambda^{2}u_{0}(\lambda x,
\lambda^{2}y)$. Following the directions in the literature we define the [*[critical Sobolev indices for the KP equation]{}*]{} as the couple of real numbers $(s_{c}^{1}, s_{c}^{2})$ such that the homogeneous Sobolev norm $$\|u_{\lambda,0}(x,y)\|_{\dot{H}^{s_{c}^{1}, s_{c}^{2}}_{x,y}}\sim C,$$ where $C$ is independent of $\lambda$. With a simple calculation we obtain $$s_{c}^{1}+2s_{c}^{2}=-{\frac{1}{2}}.
\label{critind}$$ While for the KP-II equation one can get a well-posedness theory for Sobolev spaces with indices satisfying a relationship pretty close to (see [@TT]), for the KP-I, due to the observations made in the introduction, we do not really expect to be able to reach near the critical indices in . The type of criticality that occurs in our case can be summarized in the following lemma:
In $u_{\lambda}$ is defined as in , $0<\lambda\leq 1$, then for any $s\geq 1$, $$\|u_{0,\lambda}\|_{B^{2,1}_{s}}^{1-\epsilon}
\|u_{0,\lambda}\|_{P^{2,1}_{s-1}}^{\epsilon}\lesssim
\lambda^{1-4\epsilon}.
\label{badrescal}$$
The proof of the lemma follows from simple changes of variables.
Observe that we can consider $\epsilon=\frac{1}{4}$ to be a critical exponent in this case. In fact, if we could take $\epsilon<\frac{1}{4}$ in Theorems \[xbil\] and \[ybil\], then we could use to remove the smallness assumption needed in the contraction argument presented for the proof of Theorem \[main0\]. But, like the counterexample in Proposition \[caunter\] shows, this is not possible.
We are now ready to sketch the proof of Theorem \[energytheorem\].
Let’s fix an interval of time $[-T, T]$ and a small $\epsilon > 0 ( \epsilon < \frac{1}{16})$. In light of Remark \[embedding\], we have that, if $u_0 \in E \cap P$, and $\| u_0 \|_{E\cap P} \leq
\delta_1 $, then $u_0 \in B^{2,1}_{1-\epsilon} \cap P^{2,1}_{-\epsilon}$, and we can choose $\delta_1 = \delta_1 ( \epsilon )$ so that ${{\| u_0 \|}_{B^{2,1}_{1-\epsilon} \cap P^{2,1}_{- \epsilon}}} \leq \delta$ where $\delta$ is chosen as in Theorem \[main0\], depending upon $\epsilon, \delta_1$. We can thus apply Theorem \[main0\] to obtain a unique solution $u \in Z_{1-\epsilon} = X_{1-\epsilon , {\frac{1}{2}}} \cap Y_{1-\epsilon, -\epsilon, {\frac{1}{2}}}.$ Now, by continuity with respect to the initial data, if $u_0^k$ is a smooth sequence that approximates $u_0$ in $B_{\delta_1} \subset E\cap P$, which also approximates it in $B^{2,1}_{1-\epsilon} \cap P^{2,1}_{-\epsilon}$, then the associated sequence of solutions $u_k$ is smooth, and it approximates $u$ in $C([-T,T]; B^{2,1}_{1-\epsilon} \cap
P^{2,1}_{-\epsilon})$. On the other hand, by Remark \[natep\], $${{\| u_k \|}_{L^\infty_{[-T,T]} (E \cap P )}} \leq C ( T , {{\| u_0 \|}_{E\cap P}}),$$ uniformly with respect to $k$. Thus, $u_k$ has a weak limit in $L^\infty_{[-T,T]} (E \cap P)$ that must coincide with $u$ for almost every $t$. Then, $${{\| u \|}_{L^\infty_{[-T,T]} (E \cap P )}} \leq C ( T , {{\| u_0 \|}_{E\cap P}}),$$ this concludes the proof.
[10]{}
M. Ben-Artzi, J.-C. Saut, *Uniform decay estimates for a class of oscillatory integrals and applications*, Diff. Int. Eq. **12** (1999), 137–145. J. Bergh, J. Löfström, *Interpolation Spaces*, Springer-Verlag, (1976).
J. Bourgain, *On the Cauchy problem for the Kadomtsev-Petviashvili equation*, Geometric and Funct. Anal. **3** (1993), 315–341.
J. Colliander, *Globalizing estimates for the periodic KPI equation*, Illinois J. Math. **40** (1996), no. 4, 692–698.
J. Colliander, C. Kenig, G. Staffilani, *Small solutions for the Kadomtsev-Petviashvili I equation*, Mosc. Math. J. **1** (2001), no. 4, 491–520.
A. V. Faminskii, *Cauchy problem for the generalized Kadomtsev- Petviashvili equation*, Sibirsk. Mat. Zh. **33** (1992), 160–172.
A. S. Fokas, L. Y. Sung, *On the solvability of the N-wave, Davey-Stewartson and Kadomtsev-Petviashvili equations*, Inverse Problems **8** (1992), 673-708.
J. Ginibre, Y. Tsutsumi, G. Velo, *On the Cauchy problem for the Zakharov system*, J. Funct. Anal. **151** (1997), no. 2, 384–436.
R. J. Iório, W. V. L. Nunes, *On equations of KP-type*, Proc. Royal Soc. Edin. **128A** (1998), 725–743.
P. J. Isaza, J. L. Mejía, V. Stallbohm, *Local solutions for the Kadomtsev- Petviashvili equation in ${{{\mathbb }R}}^{2}$*, J. Math. Anal. Appl. **196** (1995), 566–587.
P. J. Isaza, J. L. Mejía, V. Stallbohm, *Regularizing effects for the linearized Kadomtsev-Petviashvili (KP) equation*, Rev. Colombiana Mat. **31** (1997), 37–61.
B. B. Kadomtsev, V. I. Petviashvili, *On the stability of solitary waves in weakly dispersive media*, Soviet Phys. Dokl. **15** (1970), 539–541.
C. Kenig, G. Ponce, L. Vega , *The Cauchy problem for the Korteweg-de Vries equation in Sobolev spaces of negative indices*, Duke Math. J. **71**, No.1, (1993), 1–21.
C.E. Kenig, G. Ponce, L. Vega, *Oscillatory integrals and regularity of dispersive equations*, Indiana Univ. Math. J. **40** (1991), 33-69.
C.E. Kenig, G. Ponce, L. Vega, *Well-posedness and scattering results for generalized Korteweg-de Vries equation via the contraction principle*, Comm. Pure Appl. Math. **46** (1993), 527-620.
C.E. Kenig, G. Ponce, L. Vega, *The Cauchy problem for the Korteweg-de Vries equation in Sobolev spaces of negative indices*, Duke Math. J. **71** (1993), no. 1, 1-21.
C.E. Kenig, G. Ponce, L. Vega, *Small solutions to nonlinear Schrödinger equations*, Ann. Inst. H. PoincarÈ Anal. Non LinÈaire, **10** (1993), 255–288 .
C.E. Kenig, G. Ponce, L. Vega, *A bilinear estimate with applications to KdV equations*, J. AMS, **9** (1996), 573-603 .
L. Molinet, F. Ribaud, *The global Cauchy problem in Bourgain’s type spaces for a dispersive dissipative semilinear equation*, SIAM J. Math. Anal. **33** (2002), no. 6, 1269–1296.
L. Molinet, F. Ribaud, *On the low regularity of the Korteweg-de Vries-Burgers equation*, Int. Math. Res. Not. **2002**, no. 37, 1979–2005.
L. Molinet, J.-C. Saut, N. Tzvetkov, *Well-posedness and ill-posedness results for the Kadomtsev-Petviashvili-I equation*, Duke. Math. J. **115** (2002), no. 2, 353–384.
L. Molinet, J.-C. Saut, N. Tzvetkov, *Global well-posedness for the KP-I equation*, Math. Ann. **324** (2002), no. 2, 255–275.
J.-C. Saut, *Remarks on generalized Kadomtsev-Petviashvili equations*, Indiana Math. J. **42**, No. 3 (1993), 1011–1026. J.-C. Saut, N. Tzvetkov, *The Cauchy problem for higher-order KP equations*, J. Differential Equations, **153** (1999), no. 1, 196–222.
M. Schwarz, *Periodic solutions of Kadomtsev- Petviashvili equations*, Adv. Math. **66** (1987), 217–233.
H. Takaoka, *Time local well-posedness for the Kadomtsev-Petviashvili II equation*, Harmonic Analysis and nonlinear PDE **1102** (1999), 1–8.
H. Takaoka and N. Tzvetkov, *On the local regularity of Kadomtsev-Petviashvili II equation*, Internat. Math. Res. Notices **2** (2001), 77–114..
M. M. Tom, *On the generalized Kadomtsev- Petviashvili equation*, Contemp. Math. AMS **200** (1996), 193–210.
S. Ukai, *Local solutions of the Kadomtsev- Petviashvili equation*, J. Fac. Sci. Univ. Tokyo Sect. 1A Math. **36** (1989), 193-209.
X. Zhou, *Inverse scattering transform for the time dependent Schrödinger equations with application to the KP-I equations*, Comm. Math. Phys. **128** (1990), 551–564.
[^1]: J.E.C. was supported in part by N.S.F. Grant DMS 0100595 and N.S.E.R.C. Grant RGPIN 250233-03.
[^2]: C.K. was supported in part by N.S.F Grant DMS 9500725
[^3]: G.S. was supported in part by N.S.F. Grant DMS 0100375 and a grant by the Sloan Foundation.
[^4]: Here $H^{s_1, s_2}_{x,y} = \{ f
\in \mathcal{S}': \int \int | \langle \xi \rangle^{s_1} \langle
\mu \rangle^{s_2} \widehat{f} (\xi , \mu )|^2 d\xi d \mu < \infty \}$ where $\langle \cdot \rangle = (1 + | \cdot |^2 )^{\frac{1}{2}}$ and $\sphat$ denotes the Fourier transform from the spatial variables $(x,y)$ to their dual variables $(\xi, \mu)$.
[^5]: See for example [@C].
[^6]: Unfortunately though, our estimates are sharp as is shown in Proposition \[caunter\].
[^7]: Here one needs to take the inverse Fourier transform of $\phi_{2}$ and of $g$ in $L^{4}$ and that of $\phi_{1}$ in $L^{2}$.
[^8]: Here we are using the notation in .
[^9]: Here one needs to take the anti Fourier transform of $g$ and of $\phi_{2}$ in $L^{4}$ and that of $\phi_{1}$ in $L^{2}$.
[^10]: Recall that here ${\xi}={\xi}_{1}+{\xi}_{2}$ and ${\mu}={\mu}_{1}+{\mu}_{2}$.
[^11]: Observe that if $j_{2}=\max(j_{1},j_{2})$ one does the same analysis by applying Strichartz inequality in the order $L^{4}L^{4}L^{2}$.
[^12]: Based on the proof of Lemma \[mainlemma\], we assume that $\alpha>4$ and $2\epsilon_{0}< \frac{1}{8}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Combinatorial aspects of multivariate diagonal invariants of the symmetric group are studied. As a consequence it is proved the existence of a multivariate extension of the classical Robinson-Schensted correspondence. Further byproduct are a pure combinatorial algorithm to describe the irreducible decomposition of the tensor product of two irreducible representations of the symmetric group, and new symmetry results on permutation enumeration with respect to descent sets.'
author:
- Fabrizio Caselli
title: |
Diagonal invariants and\
the refined multimahonian distribution
---
[^1]
Introduction
============
The invariant theory of finite subgroups generated by reflections has attracted many mathematicians since their classification in the works of Chevalley [@C] and Shepard and Todd [@ST] with a particular attention on the combinatorial aspects of it. This is mainly due to the fact that the study of invariant and coinvariant algebras by means of generating functions leads naturally to nontrivial combinatorial properties of finite reflection groups. A crucial example in this context which is a link between the invariant theory and the combinatorics of the symmetric group is the Robinson-Schensted correspondence. This correspondence (see [@R; @S]) is a bijection between the symmetric group on $n$ elements and the set of ordered pairs of standard tableaux with $n$ boxes with the same shape. This is based on the row bumping algorithm and was originally introduced by Robinson to study the Littlewood-Richardson rule and by Schensted to study the lengths of increasing subsequences of a word. This algorithm has found applications in the representation theory of the symmetric group, in the theory of symmetric functions and the theory of the plactic monoid. Moreover, it is certainly fascinating from a combinatorial point of view and has inspired a considerable number of papers in the last decades. This correspondence has been generalized to other Weyl groups, by defining ad hoc tableaux, or to semistandard tableaux in the so-called RSK-correspondence, by considering permutations as special matrices with nonnegative integer entries.
The main goal of this work is to exploit further the relationship between the Robinson-Schensted correspondence and the theory of invariants of the symmetric group. By interpreting the Hilbert series with respect to a multipartition degree of certain (diagonal) invariant and coinvariant algebras in terms of (descents of) tableaux and permutations we deduce the existence of a multivariate extension of the Robinson-Schensted correspondence, which is based on the decomposition of tensor product of irreducible representations of the symmetric group. The idea of a relation between diagonal invariants and tensor product multiplicities for a finite subgroup of $GL(V)$ goes back to Solomon (see [@So Remark 5.14]) and pervades the results of Gessel [@Ge] on multipartite $P$-partitions. Although we can not define this correspondence explicitly, we can deduce from it an explicit combinatorial algorithm to describe the irreducible decomposition of the tensor product of two irreducible representations of the symmetric group. Finally, we show some further consequences in the theory of permutation enumeration.
Background
==========
Let $V$ be a finite dimensional $\mathbb C$-vector space and $W$ be a finite subgroup of the general linear group $GL(V)$ generated by reflections, i.e. elements that fix a hyperplane pointwise. We refer to such a group simply as a *reflection group*. The most significant example of such a group is the symmetric group acting by permuting a fixed linear basis of $V$. Other important examples are Weyl groups acting on the corresponding root space. In this paper we concentrate on the case of the symmetric groups (and some other related groups). Nevertheless, we preserve the symbol $W$ to denote the symmetric group $S_n$ on the $n$-element set $[n]{\stackrel{\rm def}{=}}\{1,2,\ldots,n\}$.
Given a permutation $\sigma\in W$ we denote by $${\mathrm{Des}}(\sigma) {\stackrel{\rm def}{=}}\{i\in[n-1] :\, \sigma(i)>\sigma(i+1)\}$$ the (right) *descent set* of $\sigma$ and its *major index* by $${\mathrm{maj}}(\sigma) {\stackrel{\rm def}{=}}\sum_{i\in {\mathrm{Des}}(\sigma)}i.$$ For example if $\sigma=35241$ we have ${\mathrm{Des}}(\sigma)=\{2,4\}$ and ${\mathrm{maj}}(\sigma)=6$. We recall the following equidistribution result due to MacMahon (see [@M]).
\[macmahon\] We have $$\begin{aligned}
W(q) & {\stackrel{\rm def}{=}}& \sum_{\sigma \in W}q^{{\mathrm{maj}}(\sigma)}=\sum_{\sigma \in W}q^{\mathrm{inv}(\sigma)}\\
& = & \prod_{i=1}^n(1+q+q^2+\cdots+q^i),\end{aligned}$$ where $\mathrm{inv}(\sigma)=|\{(i,j):\,i<j \textrm{ and } \sigma(i)>\sigma(j)\}|$ is the number of inversions of $\sigma$.
The dual action of a reflection group on $V^*$ can be extended to the symmetric algebra $S(V^*)$ of polynomial functions on $V$. If we fix a basis of $V$, the symmetric algebra is naturally identified with the algebra of polynomials $\mathbb C[X]$. Here and in what follows we use the symbol $X$ to denote an $n$-tuple of variables $X=(x_1,\ldots,x_n)$. The symmetric group $W$ acts on $\mathbb C[X]$ by permuting the variables. As customary, we denote by $\mathbb C[X]^W$ the ring of invariant polynomials (fixed points of the action of $W$). We also denote by $I^W_+$ the ideal of $\mathbb C[X]$ generated by homogeneous polynomials in $\mathbb C[X]^W$ of strictly positive degree. The *coinvariant algebra* associated to $W$ is defined as the corresponding quotient algebra
$$R^W{\stackrel{\rm def}{=}}\mathbb C[X]/ I^W_+.$$
The coinvariant algebra has important applications in the theory of representation since it is isomorphic to the group algebra of $W$ and in the topology of the flag variety since it is isomorphic to its cohomology ring.
If $R$ is a multigraded $\mathbb C$-vector space we can record the dimensions of its homogeneous components via its Hilbert series
$${\mathrm{Hilb}}(R)(q_1,\ldots,q_k){\stackrel{\rm def}{=}}\sum_{a_1,\ldots,a_k\in\mathbb N} \dim (R_{a_1,\ldots,a_k})q_1^{a_1}\cdots q_k^{a_k},$$ where $R_{a_1,\ldots,a_k}$ is the homogeneous subspace of $R$ of multidegree $(a_1,\ldots,a_k)$.\
We note that, since the ideal $I^W_+$ is generated by homogeneous polynomials (by total degree) the coinvariant algebra is graded in $\mathbb N$. It turns out that the polynomial $W(q)$ appearing in Theorem \[macmahon\] is the Hilbert series of the coinvariant algebra $R^W$: $$\label{wh}
W(q)={\mathrm{Hilb}}(R^W)(q).$$
This is a crucial example of interplay between the invariant theory of $W$ and the combinatorics of $W$ (by Theorem \[macmahon\]). All the other cases considered in this paper are algebraic and combinatorial variations and generalizations of this fundamental fact.
The coinvariant algebra affords also the structure of a multigraded vector space which refines the structure of graded algebra. This further decomposition can be described in terms of descents of permutations and descents of tableaux and was originally obtained in a work of Adin, Brenti and Roichman [@ABR] for Weyl groups of type $A$ and $B$ (see also [@BC1] for Weyl groups of type $D$ and [@BB] for other complex reflection groups).
If $M$ is a monomial in $\mathbb C[X]$ we denote by $\lambda(M)$ its *exponent partition*, i.e. the partition obtained by rearranging the exponents of $M$. We say that a polynomial is homogeneous of *partition degree* $\lambda$ if it is a linear combination of monomials whose exponent partition is $\lambda$. We note that the exponent partition is not well-defined in the coinvariant algebra. For example, for $n=3$ the monomials $x_1^2$ and $x_2 x_3$ are in the same class in the coinvariant algebra (since $x^2_1-x_2x_3=x_1(x_1+x_2+x_3)-(x_1x_2+x_1x_3+x_2x_3)$), though they have distinct exponent partitions. Nevertheless the exponent partition will be fundamental in defining a “partition degree” also in the coinvariant algebra.
We recall the definition of the *dominance order* in the set of partitions of $n$. We write $\mu\unlhd \lambda$, and we say that $\mu$ is smaller than or equal to $\lambda$ in the dominance order, if $\mu_1+\cdots +\mu_i\le \lambda_1+\cdots+\lambda_i$ for all $i$. We write $\mu\lhd \lambda$ if $\mu\unlhd \lambda$ and $\mu\neq \lambda$. We let $R^{(1)}_\lambda$ be the subspace of $R^W$ consisting of elements that can be represented as a linear combination of monomials with exponent partition smaller than or equal to $\lambda$ in dominance order. We also denote by $R^{(2)}_\lambda$ the subspace of $R^W$ consisting of elements that can be represented as a linear combination of monomials with exponent partition strictly smaller than $\lambda$ in dominance order. The subspaces $R^{(1)}_\lambda$ and $R^{(2)}_\lambda$ are also $W$-submodules of $R^W$ and we denote their quotient by $$R_\lambda {\stackrel{\rm def}{=}}R^{(1)}_\lambda /R^{(2)}_\lambda.$$ The $W$-modules $R_\lambda$ provide a further decomposition of the homogeneous components of the coinvariant algebra $R^W$.
There exists an isomorphism of $W$-modules $$\begin{diagram}
\varphi:R^W_k & \rTo^\cong & \bigoplus_{|\lambda|=k }R_\lambda,
\end{diagram}$$ such that $\varphi^{-1}(R_\lambda)$ can be represented by homogeneous polynomials of partition degree $\lambda$.
We can use this result to define a partition degree on the coinvariant algebra: we simply say that an element in $R_k^W$ is homogeneous of partition degree $\lambda$ if its image under the isomorphism $\varphi$ is in $R_\lambda$. We can therefore define the Hilbert polynomial of $R^W$ with respect to the partition degree by $${\mathrm{Hilb}}(R^W)(q_1,\ldots,q_n)=\sum_\lambda (\dim R_\lambda) q_1^{\lambda_{1}}\cdots q_n^{\lambda_{n}}.$$ The dimensions of the $W$-modules $R_\lambda$ can be easily described in terms of descents of permutations. Given $\sigma \in W$ we define a partition $\lambda(\sigma)$ be letting $$\lambda(\sigma)_i=|{\mathrm{Des}}(\sigma)\cap \{i,\ldots,n\}|.$$ Note that the knowledge of $\lambda(\sigma)$ is equivalent to the knowledge of ${\mathrm{Des}}(\sigma)$ and that ${\mathrm{maj}}(\sigma)=|\lambda(\sigma)|$. Then we have the following result which can be viewed as a refinement of Equation (see [@ABR Corollary 3.11]). $$\dim R_\lambda =|\{\sigma \in W: \lambda(\sigma)=\lambda\}|$$ Once we know the dimensions of the $W$-modules $R_\lambda$ we may wonder about their irreducible decomposition. For this reason we introduce the *refined fake degree polynomial* $f^\mu (q_1,\ldots, q_n)$ as the polynomial whose coefficient of $q_1^{\lambda_1}\cdots q_n^{\lambda_n}$ is the multiplicity of the representation $\mu$ in $R_\lambda$ if $\lambda$ is a partition, and zero otherwise, i.e. $$f^{\mu}(Q)=\sum_\lambda \langle \chi^{\mu},\chi^{R _\lambda} \rangle Q^\lambda,$$ where $Q^\lambda{\stackrel{\rm def}{=}}q_1^{\lambda_1}\cdots q_n^{\lambda_n}$. In this formula we denote by $\chi^{\rho}$ the character of a representation $\rho$ and by $\langle \cdot , \cdot \rangle$ the scalar product on the space of class functions on $W$ with respect to which the characters of the irreducible representations form an orthonormal basis. The polynomials $f^\mu (Q)$ have a very simple combinatorial interpretation based on standard tableaux that we are going to describe.\
Given a partition $\mu$ of $n$, the *Ferrers diagram of shape $\mu$* is a collection of boxes, arranged in left-justified rows, with $\mu_i$ boxes in row $i$. A *standard tableau* of shape $\mu$ is a filling of the Ferrers diagram of shape $\mu$ using the numbers from 1 to $n$, each occurring once, in such way that rows are increasing from left to right and columns are increasing from top to bottom. We denote by ${\mathcal {ST}}$ the set of standard tableaux with $n$ boxes. For example the following picture
(80,80) (0,80)[(1,0)[60]{}]{} (0,60)[(1,0)[60]{}]{} (0,40)[(1,0)[40]{}]{} (0,20)[(1,0)[20]{}]{} (0,0)[(1,0)[20]{}]{} (0,0)[(0,1)[80]{}]{} (20,0)[(0,1)[80]{}]{} (40,40)[(0,1)[40]{}]{} (60,60)[(0,1)[20]{}]{} (7,7)[7]{} (7,27)[4]{} (7,47)[2]{} (7,67)[1]{} (27,47)[6]{} (27,67)[3]{} (47,67)[5]{} (-30,40)[$T=$]{}
represents a standard tableau of shape $(3,2,1,1)$. We say that $i$ is a *descent* of a standard tableau $T$ if $i$ appears strictly above $i+1$ in $T$. We denote by ${\mathrm{Des}}(T)$ the set of descents of $T$ and we let ${\mathrm{maj}}(T)$ be the sum of its descents. Finally, we denote by $\mu(T)$ the shape of $T$. In the previous example we have ${\mathrm{Des}}(T)=\{1,3,5,6\}$ and so ${\mathrm{maj}}(T)=15$. As we did for permutations, given a tableau $T$ we define a partition $\lambda(T)$ by putting $$(\lambda(T))_i=|{\mathrm{Des}}(T)\cap \{i,\ldots,n\}|.$$
It is known that irreducible representations of the symmetric group $W$ are indexed by partitions of $n$. We therefore use the same symbol $\mu$ to denote a partition or the corresponding Specht module. The following result appearing in [@ABR Theorem 4.1] describes explicitly the decomposition into irreducibles of the $W$-modules $R_\lambda$ and refines a well-known result on the irreducible decomposition of the homogeneous components of $R^W$ attributed to Lusztig (unpublished) and to Kraśkiewicz and Weyman ([@KW]).
\[fmu\]The multiplicity of the representation $\mu$ in $R_\lambda$ is $$|\{T \textrm{ tableau}: \mu(T)=\mu \textrm{ and }\lambda(T)=\lambda\}|$$ and so $$f^{\mu}(q_1,\ldots,q_n)=\sum_{\{T:\mu(T)=\mu\}}Q^{\lambda(T)}.$$
Refined multimahonian distributions
===================================
We let $\mathbb C[{X_1,\ldots,X_k}]$ be the algebra of polynomials in the $nk$ variables $x_{i,j}$, with $i\in[k]$ and $j\in [n]$, i.e. we use the capital variable $X_i$ for the $n$-tuple of variables $x_{i,1},\ldots,x_{i,n}$. We consider the natural action of $W^k$ and of its diagonal subgroup $\Delta W$ on $\mathbb C[{X_1,\ldots,X_k}]$. By means of the above decomposition of the coinvariant algebra we can also decompose the algebra $$\frac{\mathbb C[{X_1,\ldots,X_k}]}{I^{W^k}_+}\cong \underbrace{R^W\otimes\cdots\otimes R^W}_k$$ in homogeneous components whose degrees are $k$-tuples of partitions with at most $n$ parts. In particular we say that an element in $\mathbb C[{X_1,\ldots,X_k}]/I^{W^k}_+$ is homogeneous of *multipartition degree* $({\lambda^{(1)},\ldots,\lambda^{(k)}})$ if it belongs to $R^W_{\lambda^{(1)}}\otimes \cdots \otimes R^W_{\lambda^{(k)}}$ by means of the above mentioned canonical isomorphism. We are mainly interested in the subalgebra $$\left( \frac {\mathbb C[X_1,\ldots,X_k]}{I^{W^k}_+}\right)^{\Delta W}\cong \frac{\mathbb C[X_1,\ldots,X_k]^{\Delta W}}{J^{W^k}_+}.$$ Here $J^{W^k}_+$ denotes the ideal generated by totally invariant polynomials with no constant term inside $\mathbb C[X_1,\ldots,X_k]^{\Delta W}$ and the isomorphism is due to the fact that the operator $$F\mapsto F^{\#}{\stackrel{\rm def}{=}}\frac{1}{|W|}\sum_{\sigma \in \Delta W}\sigma(F)$$ is an inverse of the natural projection $ \mathbb C[X_1,\ldots,X_k]^{\Delta W}/J^{W^k}_+ \rightarrow \left( \mathbb C[X_1,\ldots,X_k] / I^{W^k}_+\right)^{\Delta W} $. We can therefore consider the Hilbert polynomial $${\mathrm{Hilb}}\Big(\frac{\mathbb C[X_1,\ldots,X_k]^{\Delta W}}{J^{W^k}_+}\Big){\stackrel{\rm def}{=}}\sum_{{\lambda^{(1)},\ldots,\lambda^{(k)}}}\dim \Big(\frac{\mathbb C[X_1,\ldots,X_k]^{\Delta W}}{J^{W^k}_+}\Big)_{{\lambda^{(1)},\ldots,\lambda^{(k)}}} Q_1^{\lambda^{(1)}}\cdots Q_k^{\lambda^{(k)}}.$$ In this formula the symbol $Q_i$ stands for the $n$-tuple of variables $q_{i,1},\ldots,q_{i,n}$ and the sum is over all partitions ${\lambda^{(1)},\ldots,\lambda^{(k)}}$ with at most $n$ parts.
Our next target is to describe the previous Hilbert series. For this we need to introduce one further ingredient. We define the *Kronecker coefficients* of $W$ by $$\begin{aligned}
d_{{\mu^{(1)},\ldots,\mu^{(k)}}} & {\stackrel{\rm def}{=}}& \frac{1}{|W|} \sum_{\sigma\in W} \chi^{\mu^{(1)}}(\sigma) \cdots \chi^{\mu^{(k)}}(\sigma)\\
& = & \langle \chi^{\mu^{(1)}} \cdots \chi^{\mu^{(k-1)}},\chi^{\mu^{(k)}}\rangle_W,\end{aligned}$$ where ${\mu^{(1)},\ldots,\mu^{(k)}}$ are irreducible representations of $W$. In other words $d_{\mu^{(1)},\ldots,\mu^{(k)}}$ is the multiplicity of $\mu^{(k)}$ in the (reducible) representation $\mu^{(1)}\otimes\cdots \otimes \mu^{(k-1)}$. These numbers have been deeply studied in the literature (see, i.e. [@BK; @D; @Re; @Ro]) though they do not have an explicit description such as a combinatorial interpretation. A consequence of our main result is also a recursive combinatorial definition of the numbers $d_{\mu^{(1)},\ldots,\mu^{(k)}}$ which is independent of the character theory of $W$.\
Now we can state the following result which relates the Hilbert series above with Kronecker coefficients and the refined fake degree polynomials.
\[fcgen\]We have $$\begin{aligned}
{\mathrm{Hilb}}\Big(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W} / J^{W^k}_+\Big)(Q_1,\ldots,Q_k) & = & \sum_{{\mu^{(1)},\ldots,\mu^{(k)}}}d_{{\mu^{(1)},\ldots,\mu^{(k)}}}f^{\mu^{(1)}}(Q_1)\cdots f^{\mu^{(k)}}(Q_k)\\
& = & \sum _{T_1,\ldots,T_k}d_{\mu(T_1),\ldots,\mu(T_k)} Q_1^{\lambda(T_1)}\cdots Q_k^{\lambda(T_k)}\end{aligned}$$
The first equality is essentially due to Solomon (see [@So Theorem 5.11]). In fact, although Solomon’s result concerns only a simply graded G-module or a tensor power of a simply graded G-module (considered as a $G\wr S_n$ module), it can be easily generalized to the present context of a generic multigraded $G$-module.
The second equality follows directly from Theorem \[fmu\].
We recall that the algebra $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$, being Cohen-Macauley (see [@Sta1 Proposition 3.1]), is a free algebra over its subalgebra $ \mathbb C[{X_1,\ldots,X_k}]^{W^k}$. This implies directly that if we consider $\mathbb C[{X_1,\ldots,X_k}]$ as an algebra graded in $\mathbb N^k$ in the natural way, then, $$\label{pol}
{\mathrm{Hilb}}\Big(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}/I^{W^k}_+\Big)(q_1,\ldots,q_k) = \frac{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W})(q_1,\ldots,q_k)}{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{W^k})(q_1,\ldots,q_k)}.$$ Now, the algebra $\mathbb C[{X_1,\ldots,X_k}]$ is also multigraded by $k$-tuples of partitions with at most $n$ parts: we just say that a monomial $M$ is homogeneous of multipartition degree $(\lambda^{(1)},\ldots,\lambda^{(k)})$ if its exponent partition with respect to the variables $x_{i,1},\ldots x_{i,n}$ is $\lambda^{(i)}$ for all $i$. We write in this case $\lambda^{(i)}(M){\stackrel{\rm def}{=}}\lambda^{(i)}$ for all $i$ and $\Lambda(M){\stackrel{\rm def}{=}}({\lambda^{(1)},\ldots,\lambda^{(k)}})$. The refinement of Equation using the Hilbert series with respect to multipartition degree is no longer implied by the Cohen-Macauleyness of $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$. For this we need to use the existence of the representations $R_\lambda$ in a more subtle way. Given $\sigma \in W$ we define a monomial $$a_\sigma=\prod x_{\sigma(i)}^{\lambda(\sigma)_i}.$$ By definition we clearly have $\lambda(a_{\sigma})=\lambda(\sigma)$. In [@GS] and [@ABR] it is proved that the set of monomials $\{a_{\sigma}:\,\sigma \in W\}$ is a basis for the coinvariant algebra $R^W$. The proof in [@ABR] is based on a straightening law. For its description we need to introduce an ordering on the set of monomials of the same degree: for $m$ and $m'$ monomials of the same total degree in $\mathbb C[X]$ we let $m\prec m'$ if
1. $\lambda(m)\lhd \lambda(m')$; or
2. $\lambda(m)=\lambda(m')$ and $\textrm{inv}(\pi(m))>\textrm{inv} (\pi(m'))$,
where $\pi(m)$ is the permutation $\pi$ having a minimal number of inversions such that the exponent in $m$ of $x_{\pi(i)}$ is greater than or equal to the exponent in $m$ of $x_{\pi(i+1)}$ for all $i$. The straightening law is the following: let $m$ be a monomial in $\mathbb C[X]$. Then $\mu:=\lambda(m)-\lambda(\pi(m))$ is still a partition and $$\label{straight} m=m_\mu \cdot a_{\pi(m)}+\sum_{m'\prec m}c_{m,m'}m',$$ where $c_{m,m'}\in \mathbb C$ and $m_\mu$ is the monomial symmetric function. The straightening algorithm stated in [@ABR] uses elementary symmetric functions instead of monomial symmetric functions, but one can easily check that the two statements are equivalent. The fact that the set $\{a_{\sigma}:\,\sigma \in W\}$ is a basis of $R^W$ implies directly that the set of monomials $$a_{{\sigma_1,\ldots, \sigma_k}}{\stackrel{\rm def}{=}}a_{\sigma_1}(X_1)\cdots a_{\sigma_k}(X_k)$$ is a basis for the coinvariant algebra of $W^k$, i.e. the algebra $\mathbb C[{X_1,\ldots,X_k}]/I^{W^k}_+$. Now, the monomials $a_{{\sigma_1,\ldots, \sigma_k}}$ form a basis for the algebra $\mathbb C[X_1,\ldots X_k]$ as a free module over the subring $\mathbb C[{X_1,\ldots,X_k}]^{W^k}$ of $W^k$-invariants (being a basis of the coinvariant algebra $\mathbb C[{X_1,\ldots,X_k}]/I^{W^k}_+$), i.e. $$\mathbb C[{X_1,\ldots,X_k}]=\bigoplus_{{\sigma_1,\ldots, \sigma_k}\in W}\mathbb C[{X_1,\ldots,X_k}]^{W^k}a_{{\sigma_1,\ldots, \sigma_k}}.$$ The following result states a triangularity property of this basis. If $({\mu^{(1)},\ldots,\mu^{(k)}})$ and $({\lambda^{(1)},\ldots,\lambda^{(k)}})$ are two $k$-tuples of partitions we write $({\mu^{(1)},\ldots,\mu^{(k)}})\unlhd ({\lambda^{(1)},\ldots,\lambda^{(k)}})$ if $\mu^{(i)}\unlhd \lambda^{(i)}$ for all $i$ and we denote by $\mathbb C[{X_1,\ldots,X_k}]_{\unlhd({\lambda^{(1)},\ldots,\lambda^{(k)}})}$ the space of polynomials spanned by monomials with multipartition degree $\unlhd ({\lambda^{(1)},\ldots,\lambda^{(k)}})$. We similarly define $\mathbb C[{X_1,\ldots,X_k}]_{\lhd({\lambda^{(1)},\ldots,\lambda^{(k)}})}$.
\[trian\] Let $M\in \mathbb C[{X_1,\ldots,X_k}]$ be a monomial and let $$M=\sum_{{\sigma_1,\ldots, \sigma_k}\in W}f_{{\sigma_1,\ldots, \sigma_k}}a_{{\sigma_1,\ldots, \sigma_k}},$$ where $f_{{\sigma_1,\ldots, \sigma_k}}\in \mathbb C[{X_1,\ldots,X_k}]^{W^k}$. Then this sum is restricted to those ${\sigma_1,\ldots, \sigma_k}$ such that $\Lambda(M)-\Lambda(a_{{\sigma_1,\ldots, \sigma_k}})$ is a $k$-tuple of partitions and $$f_{{\sigma_1,\ldots, \sigma_k}}\in \mathbb C[{X_1,\ldots,X_k}]_{\unlhd \Lambda(M)-\Lambda(a_{{\sigma_1,\ldots, \sigma_k}})}.$$
Given two monomials $M=m_1(X_1)\cdots m_k(X_k)$ and $M'=m_1'(X_1)\cdots m_k'(X_k)$ we let $M\prec M'$ if $m_i\prec m'_i$ for all $i$. We proceed by a double induction on the total degree and on $\prec$ within the set of monomials of the same multidegree. If $M$ has total degree zero the result is trivial. Otherwise let $\Lambda(M)=({\lambda^{(1)},\ldots,\lambda^{(k)}})$. If $M$ is minimal with respect to the ordering $\prec$ then $\lambda^{(i)}$ is minimal with respect to the dominance order for all $i$. If there exists $i$ such that $|\lambda^{(i)}|\geq n$ then $M=(x_{i,1}\cdots x_{i,n})M'$ and the result follows by induction since the total degree of $M'$ is strictly smaller than the degree of $M$. If $|\lambda^{(i)}|< n$ for all $i$ then $\lambda^{(i)}=(1^{k_i})$ by the minimality condition. Then $ m_i=a_{\pi(m_i)}$ for all $i$ and the result follows. In the general case we can apply the straightening law (\[straight\]) to all the $m_i$’s getting $$M=f_M \cdot a_{\pi(m_1),\ldots,\pi(m_k)}+\sum_{M'\prec M}c_{M,M'}M',$$ where $f_M$ is a homogeneous $W^k$-invariant polynomial of multipartition degree $\Lambda(M)-\Lambda(a_{\pi(m_1),\ldots,\pi(m_k)})$. Then the result follows by induction.
Now recall the already mentioned sequence of isomorphisms of $W$-modules $$\begin{aligned}
\frac{\mathbb C [{X_1,\ldots,X_k}]^{\Delta W}}{J_+^{W^k}}&\cong &\left(\frac{\mathbb C [{X_1,\ldots,X_k}]}{I_+^{W^k}}\right)^{\Delta W}\cong (\underbrace{R^W\otimes \cdots \otimes R^W}_{k\textrm{ times}})^{\Delta W}\\& \cong &\bigoplus_{{\lambda^{(1)},\ldots,\lambda^{(k)}}}(R_{\lambda^{(1)}}\otimes\cdots \otimes R_{\lambda^{(k)}})^{\Delta W}.\end{aligned}$$ Consider a basis of $(R_{\lambda^{(1)}}\otimes\cdots \otimes R_{\lambda^{(k)}})^{\Delta W}$. Every element of such a basis can be represented by a homogeneous element in $\mathbb C [{X_1,\ldots,X_k}]$ of multipartition degree $({\lambda^{(1)},\ldots,\lambda^{(k)}})$ (by definition) which is invariant for the action of $\Delta W$. In fact, if a representative $F$ of a basis element is not $\Delta W$-invariant we can consider its symmetrization $F^{\#}$ since, clearly, $F$ and $F^{\#}$ represents the same class in $(R_{\lambda^{(1)}}\otimes\cdots \otimes R_{\lambda^{(k)}})^{\Delta W}$. We denote by $\mathcal B({\lambda^{(1)},\ldots,\lambda^{(k)}})$ this set of representatives, i.e. $\mathcal B({\lambda^{(1)},\ldots,\lambda^{(k)}})$ is a set of polynomials in $\mathbb C [{X_1,\ldots,X_k}]^{\Delta W}$ of multipartition degree $({\lambda^{(1)},\ldots,\lambda^{(k)}})$ whose corresponding classes form a basis of $(R_{\lambda^{(1)}}\otimes\cdots \otimes R_{\lambda^{(k)}})^{\Delta W}$. We denote by $\mathcal B$ the (disjoint) union of all $\mathcal B({\lambda^{(1)},\ldots,\lambda^{(k)}})$. By the Cohen-Macauleyness of $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$ we can deduce that the set $\mathcal B$ is a basis for $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$ as a free $\mathbb C[{X_1,\ldots,X_k}]^{W^k}$-module (see [@Sta1 Proposition 3.1]), i.e. $$\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}=\bigoplus_{b\in \mathcal B}\mathbb C[{X_1,\ldots,X_k}]^{W^k} \cdot b.$$ The following result implies a crucial triangularity property of the basis $\mathcal B$.
\[verytrian\] Let $F\in \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$ be homogeneous of multipartition degree $\Lambda(F)$. Then the unique expression $$F=\sum_{b\in \mathcal B}f_b b,$$ with $f_b\in \mathbb C[{X_1,\ldots,X_k}]^{W^k}$ for all $b\in \mathcal B$, is such that the sum is restricted to those $b\in\mathcal B$ for which $\Lambda(F)-\Lambda(b)$ is a $k$-tuple of partitions and $$f_b\in \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\unlhd \Lambda(F)-\Lambda(b)}.$$
Let $\prec$ be a total order on the set of $k$-tuples of partitions of length at most $n$ satysfying the following two conditions
- If $|\mu^{(1)}|+\cdots+|\mu^{(k)}|<|\lambda^{(1)}|+\cdots+|\lambda^{(k)}|$ then $({\mu^{(1)},\ldots,\mu^{(k)}})\prec({\lambda^{(1)},\ldots,\lambda^{(k)}})$;
- If $\mu^{(i)}\unlhd \lambda^{(i)}$ for all $i$, then $({\mu^{(1)},\ldots,\mu^{(k)}})\prec({\lambda^{(1)},\ldots,\lambda^{(k)}})$.
We proceed by induction on the multipartition degree of $F$ with respect to the total order $\prec$. If $F$ has degree zero then the result is trivial. Otherwise let $\Lambda(F)=({\lambda^{(1)},\ldots,\lambda^{(k)}})$ be the multipartite partition degree of $F$. Then $F$ represents an element in $(R_{\lambda^{(1)}}\otimes\cdots \otimes R_{\lambda^{(k)}})^{\Delta W}$. Therefore, $$F=\sum_{b\in \mathcal B({\lambda^{(1)},\ldots,\lambda^{(k)}})} c_b b$$ in $(R_{\lambda^{(1)}}\otimes\cdots \otimes R_{\lambda^{(k)}})^{\Delta W}$. This means that $$F=\sum_{b\in \mathcal B({\lambda^{(1)},\ldots,\lambda^{(k)}})}c_b b +G$$ in $(\underbrace{R^W\otimes \cdots \otimes R^W}_{k\textrm{ times}})^{\Delta W}$, where $G$ is a $\Delta W$-invariant polynomial such that $$G\in \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}_{\lhd \Lambda(F)}.$$ Finally we deduce from this that $$F=\sum_{b\in \mathcal B({\mu^{(1)},\ldots,\mu^{(k)}})}c_b b +G+H$$ in $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$, where $H$ belongs to $I_+^{W^k}$. We can clearly assume that $G$ and $H$ are homogeneous with the same total multidegree of $F$. The induction hypothesis applies directly to $G$. Regarding $H$, by Lemma \[trian\], we can express $H=\sum_{{\sigma_1,\ldots, \sigma_k}}f_{{\sigma_1,\ldots, \sigma_k}} a_{{\sigma_1,\ldots, \sigma_k}}$ with $f_{{\sigma_1,\ldots, \sigma_k}}\in \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\unlhd \Lambda(F)-\Lambda (a_{{\sigma_1,\ldots, \sigma_k}})}$ since $H$ is a sum of monomials of multipartition degree smaller than or equal to $\Lambda(F)$ in dominance order. Moreover, all the polynomials $f_{{\sigma_1,\ldots, \sigma_k}}$ have positive degree since $H\in I_+^{W^k}$. Now we can apply the operator $\#$ to this identity and we get $$H=\sum_{{\sigma_1,\ldots, \sigma_k}}f_{{\sigma_1,\ldots, \sigma_k}} a^{\#}_{{\sigma_1,\ldots, \sigma_k}}.$$ Finally we can apply our induction hypothesis to the polynomials $a^{\#}_{{\sigma_1,\ldots, \sigma_k}}$ since they have degree smaller than $F$ and the proof is completed by observing that, clearly, $$\mathbb C[{X_1,\ldots,X_k}]_{\unlhd \Lambda}\cdot \mathbb C[{X_1,\ldots,X_k}]_{\unlhd \Lambda'}\subseteq \mathbb C[{X_1,\ldots,X_k}]_{\unlhd \Lambda+\Lambda'}.$$
We observe that Lemma \[verytrian\] fails to be true for a generic homogeneous basis $\mathcal B$ of $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$ as a free $\mathbb C[{X_1,\ldots,X_k}]^{W^k}$-module. We refer the reader to [@Rei Section 4.2], [@A] and [@BL] for the explicit description of some bases of $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$ over $\mathbb C[{X_1,\ldots,X_k}]^{W^k}$, with particular attention to the case $k=2$.
For notational convenience, if $\Lambda=({\lambda^{(1)},\ldots,\lambda^{(k)}})$ is a multipartition, we denote by $$\mathcal Q^\Lambda{\stackrel{\rm def}{=}}Q_1^{\lambda^{(1)}}\cdots Q_k^{\lambda^{(k)}}=\prod_{i=1}^k\prod_{j=1}^n q_{i,j}^{\lambda^{(i)}_j}.$$
\[uou\] We have $${\mathrm{Hilb}}\Big(\frac{\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}}{J^{ W^k}_+}\Big)(Q_1,\ldots,Q_k)= \frac{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W})(Q_1,\ldots,Q_k)}{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{W^k})(Q_1,\ldots,Q_k)}=\sum_{b\in \mathcal B}\mathcal Q^{\Lambda(b)}$$
The fact that ${\mathrm{Hilb}}\Big(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W} / J^{ W^k}_+\Big)(Q_1,\ldots,Q_k)=\sum_{b\in \mathcal B}\mathcal Q^{\Lambda(b)}$ is clear from the definition of the multipartition degree on $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W} / J^{ W^k}_+$ and the definition of the set $\mathcal B$. Lemma \[verytrian\] implies that $$\dim \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}_{\unlhd \Lambda}= \sum_{b\in \mathcal B}\sum_{\{\Lambda':\Lambda'+\Lambda(b)\unlhd \Lambda\}}\dim \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\Lambda'},$$ and similarly with $\lhd$ instead of $\unlhd$. Therefore $$\begin{aligned}
\dim \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}_{\Lambda} &=& \dim \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}_{\unlhd \Lambda}-\dim \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}_{\lhd\Lambda}\\
& = & \sum_{b\in \mathcal B}\sum_{\{\Lambda':\Lambda'+\Lambda(b)=\Lambda\}}\dim \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\Lambda'}.\end{aligned}$$ Note that in the last sum there is only one summand corresponding to $\Lambda'=\Lambda-\Lambda(b)$ if this is a multipartition, and there are no summands otherwise. So we have $$\begin{aligned}
{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W})& = & \sum_{\Lambda}\dim \mathbb C[{X_1,\ldots,X_k}]^{\Delta W}_{\Lambda} \mathcal Q^\Lambda\\
& = & \sum_\Lambda \sum_{b\in \mathcal B}\sum_{\{\Lambda':\Lambda'+\Lambda(b)=\Lambda\}}\dim \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\Lambda'}\mathcal Q^{\Lambda}\\
&=&\sum_{b\in \mathcal B}\sum_{\Lambda'}\dim \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\Lambda'} \mathcal Q^{\Lambda'+\Lambda(b)}\\
&=&\sum_{b\in \mathcal B}\mathcal Q^{\Lambda(b)}\sum_{\Lambda'}\dim \mathbb C[{X_1,\ldots,X_k}]^{W^k}_{\Lambda'}Q^{\Lambda'}\\
&=&\sum_{b\in \mathcal B}\mathcal Q^{\Lambda(b)} {\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{W^k}).\end{aligned}$$
Now we need to study the two Hilbert series of the invariant algebras $\mathbb C[{X_1,\ldots,X_k}]^{\Delta W}$ and $\mathbb C[{X_1,\ldots,X_k}]^{W^k}$ with respect to the multipartition degree. Before stating our next result we need to recall a classical theorem that can be attributed to Gordon [@G] and Garsia and Gessel [@GG] on multipartite partitions. We say that a collection $(f^{(1)},\ldots,f^{(k)})$ of $k$ elements of $\mathbb N^n$ is a $k$-partite partition if $f^{(i)}_j\geq f^{(i)}_{j+1}$ whenever $f^{(h)}_j=f^{(h)}_{j+1}$ for all $h<i$. For notational convenience we denote by $W^{(k)}{\stackrel{\rm def}{=}}\{({\sigma_1,\ldots, \sigma_k})\in W^k:\,\sigma_1\cdots\sigma_k=1)\}$. The main property of $k$-partite partitions that we need is the following.
\[parpar\] There exists a bijection between the set of $k$-partite partitions and the set of $2k$-tuples $(\sigma_1,\ldots,\sigma_k, \mu^{(1)},\ldots, \mu^{(k)})$ such that
- $({\sigma_1,\ldots, \sigma_k})\in W^{(k)}$;
- $\mu^{(i)}$ is a partition with at most $n$ parts;
- $\mu^{(i)}_j>\mu^{(i)}_{j+1}$ whenever $j\in {\mathrm{Des}}(\sigma_i)$.
The bijection is such that $\mu^{(i)}$ is obtained by reordering the coefficients of $f^{(i)}$.
We can now prove the following formula for the quotient of the Hilbert polynomials with respect to the multipartition degree associated to the invariant algebras of $\Delta W$ and $W^k$.
\[GGgen\]We have $$\frac{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W})(Q_1,\ldots,Q_k)}{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{W^k})(Q_1,\ldots,Q_k)}=
\sum_{({\sigma_1,\ldots, \sigma_k})\in W^{(k)}}Q_1^{\lambda(\sigma_1)}\cdots Q_k^{\lambda(\sigma_k)}$$
We observe that the set of monomials $X_1^{f^{(1)}}\cdots X_k^{f^{(k)}}$ as $(f^{(1)},\ldots,f^{(k)})$ varies among all possible $k$-partite partitions is a set of representatives for the orbits of the action of $\Delta W$ in the set of monomials in $\mathbb C[{X_1,\ldots,X_k}]$. By means of Theorem \[parpar\] we can deduce that $${\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W})(Q_1,\ldots,Q_k)=\sum_{\substack{{\sigma_1,\ldots, \sigma_k},\\ {\mu^{(1)},\ldots,\mu^{(k)}}}}Q_1^{\mu^{(1)}}\cdots Q_k^{\mu^{(k)}},$$ where the indices in the previous sum are such that they satisfy the conditions stated in Theorem \[parpar\]. We now observe that we have an equivalence of conditions $$\mu^{(i)}_j>\mu^{(i)}_{j+1} \textrm{ whenever } j\in {\mathrm{Des}}(\sigma_i) \Longleftrightarrow \mu^{(i)}-\lambda(\sigma_i) \textrm{ is a partition}.$$ Therefore we can simplify the previous sum in the following way $$\sum_{\substack{{\sigma_1,\ldots, \sigma_k},\\ {\mu^{(1)},\ldots,\mu^{(k)}}}}Q_1^{\mu^{(1)}}\cdots Q_k^{\mu^{(k)}} = \sum_{({\sigma_1,\ldots, \sigma_k})\in W^{(k)}} \sum_{\nu^{(1)},\ldots,\nu^{(k)}} Q_1^{\nu^{(1)}+\lambda(\sigma_1)}\cdots Q_k^{\nu^{(k)}+\lambda(\sigma_k)},$$ where the last sum is on all possible $k$-tuples of partitions $\nu^{(1)},\ldots,\nu^{(k)}$ of length at most $n$. The result follows since, clearly, $${\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{W^k} = \sum_{\nu^{(1)},\ldots,\nu^{(k)}}Q_1^{\nu^{(1)}}\cdots Q_k^{\nu^{(k)}}.$$
Putting all these results together we obtain the following sequence of equivalent interpretations for what we may call the *refined multimahonian distribution*.
We have $$\begin{aligned}
W(Q_1,\ldots,Q_k)& {\stackrel{\rm def}{=}}& \sum_{T_1,\ldots,T_k}d_{\mu(T_1),\ldots,\mu(T_k)}Q_1^{\lambda(T_1)},\ldots,Q_k^{\lambda(T_k)}\\
& = & \sum_{{\mu^{(1)},\ldots,\mu^{(k)}}} d_{\mu^{(1)},\ldots,\mu^{(k)}} f^{\mu^{(1)}}(Q_1)\cdots f^{\mu^{(k)}}(Q_k)\\
& = & {\mathrm{Hilb}}\Big(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W} / J^{ W}_+)\Big)(Q_1,\ldots,Q_k)\\
& = & \frac{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{\Delta W})(Q_1,\ldots,Q_k)}{{\mathrm{Hilb}}(\mathbb C[{X_1,\ldots,X_k}]^{W^k})(Q_1,\ldots,Q_k)}\\
& = & \sum_{\sigma_1\cdots \sigma_k=1}Q_1^{\lambda(\sigma_1)}\cdots Q_k^{\lambda(\sigma_k)}\end{aligned}$$
The four identities are the contents of Theorem \[fcgen\], Corollary \[uou\] and Theorem \[GGgen\] respectively.
The cardinality of the set of $k$-tuples of permutations in $W^{(k)}$ having fixed descent sets was already studied by Gessel in [@Ge] and the idea to use Kronecker products (of quasi symmetric functions) is already present in his work. Nevertheless, Gessel’s approach by means of quasi symmetric functions and the present approach my means of diagonal invariants are certainly different, and we believe that the present results can not be deduced from Gessel’s work in a standard way.
The reason why we call the distribution $W(Q_1,\ldots,Q_k)$ refined is that one can consider its coarse version $W(q_1,\ldots q_k)$ obtained by putting $q_{i,j}=q_i$ for all $i$ and $j$. In this case one obtains the so-called *multimahonian distribution* $\sum_{({\sigma_1,\ldots, \sigma_k})\in W^{(k)}}q_1^{{\mathrm{maj}}(\sigma_1)}\cdots q_k^{{\mathrm{maj}}(\sigma_k)}$ which has been extensively studied in the literature (see, e.g., [@AR; @BRS; @BC; @FS; @GG]).
\[conj\] There exists a map $\mathcal T:W^{(k)}\longrightarrow {\mathcal {ST}}^k$ satisfying the following two conditions:
1. For every $k$-tuple of tableaux $(T_1,\ldots,T_k)$, $$|\mathcal T^{-1}(T_1,\ldots,T_k)|=d_{\mu(T_1),\ldots,\mu(T_k)}.$$ In particular it depends only on the shapes of the tableaux $T_1,\ldots,T_k$;
2. if $\mathcal T(\sigma_1,\ldots,\sigma_k)=(T_1,\ldots,T_k)$ then ${\mathrm{Des}}(T_i)={\mathrm{Des}}(\sigma_i)$ for all $i=1,\ldots,k$.
The classical Robinson-Schensted correspondence (see [@Sta §7.11] for a description of this correspondence) provides a bijective proof of this corollary in the case $k=2$.
We can also conjecture that the correspondence $\mathcal T$ of Corollary \[conj\] should be well-behaved with respect to cyclic permutations of the arguments in the sense that $$\mathcal T(\sigma_1,\ldots, \sigma_k)=(T_1,\ldots,T_k) \Longrightarrow \mathcal T(\sigma_2,\ldots, \sigma_k, \sigma_1)=(T_2,\ldots,T_k, T_1).$$
\[rsmulti\] Find the map $\mathcal T$ of Corollary \[conj\] explicitly.
We observe that the resolution of this Problem would provide also an explicit combinatorial interpretation for the coefficients $d_{{\mu^{(1)},\ldots,\mu^{(k)}}}$.
Combinatorial applications
==========================
In this final section we deduce some combinatorial results on Kronecker coefficients and permutation enumeration that follow from the results of the previous section. Our next goal is to show that we do not need to know the coefficients $d_{{\mu^{(1)},\ldots,\mu^{(k)}}}$ to solve Problem \[rsmulti\]. This is because Corollary \[conj\] uniquely determines the coefficients $d_{{\mu^{(1)},\ldots,\mu^{(k)}}}$ in the following sense.
\[duni\] Let $\mathcal T:W^{(k)}\rightarrow {\mathcal {ST}}^k$ be such that
1. $|\mathcal T^{-1}(T_1,\ldots,T_k)|$ depends uniquely on the shapes of $T_1,\ldots,T_k$;
2. if $\mathcal T(\sigma_1,\ldots,\sigma_k)=(T_1,\ldots,T_k)$ then ${\mathrm{Des}}(T_i)={\mathrm{Des}}(\sigma_i)$ for all $i=1,\ldots,k$.
Then $\mathcal T$ satisfies the conditions of Corollary \[conj\], i.e. $|\mathcal T^{-1}(T_1,\ldots,T_k)|=d_{\mu(T_1),\ldots,\mu(T_k)}$ for all $(T_1,\ldots,T_k) \in {\mathcal {ST}}^k$.
This proposition is an immediate consequence of the following results which provides also an explicit entirely combinatorial algorithm to compute the coefficients $d_{{\mu^{(1)},\ldots,\mu^{(k)}}}$.
Let $\mu=(\mu_1,\ldots,\mu_r)$ be a partition of $n$ (with $\mu_r>0$). Then
1. there exists a unique standard tableau $T_\mu$ of shape $\mu$ and descent set ${\mathrm{Des}}(T_\mu)=\{\mu_1,\mu_1+\mu_2,\ldots, \mu_1+\cdots+\mu_{r-1}\}$;
2. if $T$ is a standard tableau and ${\mathrm{Des}}(T)={\mathrm{Des}}(T_\mu)$, then $\mu(T)\unrhd\mu$.
\[uiop\]
The unique tableau $T_\mu$ satisfying these conditions is obtained by inserting the numbers form $1$ to $\mu_1$ in the first row of the Ferrers diagram of $\mu$, then the numbers from $\mu_1+1$ to $\mu_1+\mu_2$ in the second row and so on in the following rows. The uniqueness of $T_\mu$ is clear since the position of any single entry is forced by the descents conditions and the shape of the tableau.
We observe that the number of descents of a tableau is at least the number of rows minus 1. This is because if $i\neq 1$ appears in the first column of $T$ then $i-1$ is necessarily above it and so $i-1\in {\mathrm{Des}}(T)$. Suppose $T$ is a standard tableau with the same descents as $T_\mu$. Consider the subtableau $T_i$ of $T$ consisting of the boxes filled with entries bounded by $\mu_1+ \cdots +\mu_i$. Then $T_i$ has exactly $i-1$ descents. So, if $j$ is the number of rows of $T_i$, we have $i-1\geq j-1$, i.e. $j\leq i$, by the previous observation. This implies that the numbers from 1 to $\mu_1+ \cdots +\mu_i$ appear all in the first $i$ rows of $T$ and the claim follows.
The following corollary is a recursion satisfied by the coefficients $d_{{\mu^{(1)},\ldots,\mu^{(k)}}}$ and proves Proposition \[duni\].
\[algo\] Let ${\mu^{(1)},\ldots,\mu^{(k)}}$ be a multipartition and let $D_i={\mathrm{Des}}(T_{\mu^{(i)}})$. Then $$d_{{\mu^{(1)},\ldots,\mu^{(k)}}}=\sum_{\substack {({\sigma_1,\ldots, \sigma_k})\in W^{(k)}: \\ {\mathrm{Des}}(\sigma_i)=D_i}} 1- \sum_{\left\{\substack{T_1,\ldots,T_k\in {\mathcal {ST}}:\, {\mathrm{Des}}(T_i)=D_i \textrm{ and}\\ \left(\mu(T_1),\ldots, \mu(T_k)\right) \rhd({\mu^{(1)},\ldots,\mu^{(k)}}) }\right\}}d_{\mu(T_1),\ldots,\mu(T_k)}.$$
By Corollary \[conj\] we have $$\sum_{\substack{T_1,\ldots,T_k\in {\mathcal {ST}}:\\ {\mathrm{Des}}(T_i)=D_i}} d_{\mu(T_1),\ldots,\mu(T_k)}=\sum_{\substack {({\sigma_1,\ldots, \sigma_k})\in W^{(k)}: \\ {\mathrm{Des}}(\sigma_i)=D_i}} 1.$$ Now the claim follows by Lemma \[uiop\].
Let $n=4$ and $k=3$. We compute $d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,0){\line(1,0){6}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,0){\line(0,1){6}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}$ by means of Corollary \[algo\]. In this case we have $D_1=\{3\}, D_2=\{2\}, D_3=\{2,3\}$. We have to determine all tableaux having these descents sets. Now we observe that the unique tableau having descent set $D_1$, is $T_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}}}$ and the unique tableau having descent set $D_3$ is $T_{{\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}$. On the other hand there are two tableaux having descent set $D_2$ and these are $T_{{\begin{picture}(9,6)
\put(1,0){\line(1,0){6}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,0){\line(0,1){6}}
\end{picture}}}$ and $\begin{picture}(24,12)
\put(0,-4){\line(1,0){8}}
\put(0,4){\line(1,0){24}}
\put(0,12){\line(1,0){24}}
\put(0,-4){\line(0,1){16}}
\put(8,-4){\line(0,1){16}}
\put(16,4){\line(0,1){8}}
\put(24,4){\line(0,1){8}}
\put(2,-2){\tiny{$3$}}
\put(2,6){\tiny{$1$}}
\put(10,6){\tiny{$2$}}
\put(18,6){\tiny{$4$}}
\end{picture}$. So we have
$$d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,0){\line(1,0){6}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,0){\line(0,1){6}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}=\sum_{\substack {(\sigma_1,\sigma_2,\sigma_3)\in W^{(3)}: \\ {\mathrm{Des}}(\sigma_i)=D_i}} 1-d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}$$ The sum in the previous formula is $2$ since the only two triplets in the index set are $(1243,1423,1432)$ and $(2341,2413,2431)$. So we deduce that $
d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,0){\line(1,0){6}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,0){\line(0,1){6}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}=2-d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}$. Now we compute $d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}$ by applying again Corollary \[algo\]. In this case we have $D_1=D_2=\{3\}$ and $D_3=\{2,3\}$. By the previous observations on the tableaux having these descent sets we deduce that $$d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}},{\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}=\sum_{\substack {(\sigma_1,\sigma_2,\sigma_3)\in W^{(3)}: \\ {\mathrm{Des}}(\sigma_i)=D_i}} 1.$$ The last sum equals 1 since the only element in the index set is $(1342,1243,1432)$ and we can conclude that $d_{{\begin{picture}(12,6)
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){9}}
\put(1,6){\line(1,0){9}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,3){\line(0,1){3}}
\put(10,3){\line(0,1){3}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,0){\line(1,0){6}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,0){\line(0,1){6}}
\put(4,0){\line(0,1){6}}
\put(7,0){\line(0,1){6}}
\end{picture}}, {\begin{picture}(9,6)
\put(1,-3){\line(1,0){3}}
\put(1,0){\line(1,0){3}}
\put(1,3){\line(1,0){6}}
\put(1,6){\line(1,0){6}}
\put(1,-3){\line(0,1){9}}
\put(4,-3){\line(0,1){9}}
\put(7,3){\line(0,1){3}}
\end{picture}}}=1$.
To conclude this paper we prove some new results on permutation statistics that can be deduced from Corollary \[conj\]. The first observation is a direct consequence of the symmetry of Kronecker coefficients with respect to their arguments. Let $D_1,\ldots,D_k\subseteq [n-1]$ and let $\pi$ be any permutation on $\{1,\ldots,k\}$. Then $$\label{sym}
|\{(\sigma_1,\ldots,\sigma_k)\in W^{(k)}:\,{\mathrm{Des}}(\sigma_i)=
D_i\}|=|\{(\sigma_1,\ldots,\sigma_k)\in W^{(k)}:\,{\mathrm{Des}}(\sigma_i)=D_{\pi(i)}\}|.$$ This naturally leads to consider the following.
Find a combinatorial bijective proof for the identity .
It is plausible that the resolution of this problem could be a first step towards the resolution of Problem \[rsmulti\].
The classical Robinson-Schensted correspondence allows one to prove some results on permutation enumeration and in particular on the bimohonian distributions (see for example [@DF; @FS; @BRS]). So, it is natural to ask how we can generalize these properties using the existence of the multivariate Robinson-Schensted correspondence. If $X$ is either a permutation on $n$ elements or a tableau with $n$ entries we denote by $$\begin{aligned}
{\mathrm {Codes}}(X)&{\stackrel{\rm def}{=}}&\{i:n-i\in {\mathrm{Des}}(X)\};\\
{\mathrm {Asc}}(X) & {\stackrel{\rm def}{=}}& [n-1]\setminus {\mathrm{Des}}(X);\\
{\mathrm {Coasc}}(X)& {\stackrel{\rm def}{=}}& \{i:n-i\in {\mathrm {Asc}}(X)\}.\end{aligned}$$ The following is a multivariate generalization of a result of Foata and Schützenberger ([@FS Theorem 2])
\[descodes\] For all $I \subset \{1,...,k\}$ there exists an involution $ F_I:W^{(k)} \rightarrow W^{(k)}$ such that, if $ F_I({\sigma_1,\ldots, \sigma_k})=(\tau_1,...,\tau_k)$, then $${\mathrm{Des}}(\sigma_i)=\left\{\begin{array}{ll} {\mathrm {Codes}}(\tau_i) & \textrm{if }i\in I;\\ {\mathrm{Des}}(\tau_i) & \textrm{otherwise.}\end{array}\right.$$
We recall (see [@FS; @Schutzen]) that there exists a bijection $T\mapsto T^J$ on the set of standard tableaux of the same shape such that ${\mathrm{Des}}(T)={\mathrm {Codes}}(T^J)$. Then the result follows immediately from Corollary \[rsmulti\].
It is clear that we can substitute in Proposition \[descodes\] ${\mathrm{Des}}$ with ${\mathrm {Asc}}$ and ${\mathrm {Codes}}$ with ${\mathrm {Coasc}}$ obtaining an analogous result. We can unify and generalize Equation and Proposition \[descodes\] in the following statement.
\[dcac\] Fix $k$ subsets $D_1,\ldots,D_k \subseteq [n-1]$ arbitrarily. Then for any integer sequence $0\leq i_1\leq i_2 \leq i_3 \leq k$ and for any permutation $\pi $ on $[k]$ the cardinality $C(i_1,i_2,i_3;\pi)$ of the set $$\left\{({\sigma_1,\ldots, \sigma_k})\in W^{(k)}:\,\,D_i=\left\{ \begin{array}{ll}{\mathrm{Des}}(\sigma_{\pi(i)})& \textrm{if }0<i\leq i_1,\\ {\mathrm {Codes}}(\sigma_{\pi(i)})& \textrm{if }i_1 < i\leq i_2,\\ {\mathrm {Asc}}(\sigma_{\pi(i)})& \textrm{if }i_2< i \leq i_3, \\ {\mathrm {Coasc}}(\sigma_{\pi(i)})& \textrm{if }i_3<i\leq k, \end{array}
\right.\right\}$$ depends only on the parity of $i_2$, and in particular it does not depend on the indices $i_1, i_3$ and on the permutation $\pi$.
Since the knowledge of one of the three sets ${\mathrm {Codes}}(\sigma)$, ${\mathrm {Asc}}(\sigma)$, ${\mathrm {Coasc}}(\sigma)$ is equivalent to the knowledge of the set ${\mathrm{Des}}(\sigma)$ it is clear that, by , the number $C(i_1,i_2,i_3;\pi)$ does not depend on $\pi$. So we can assume that $\pi=Id$. Then, from Proposition \[descodes\], we deduce that $C(i_1,i_2,i_3;Id)$ depends only on $i_2$.
To prove that $C(i_1,i_2,i_3;\pi)$ depends only on the parity of $i_2$ we consider the following permutation on $W^{(k)}$ $$({\sigma_1,\ldots, \sigma_k})\mapsto (\sigma_1w_0,w_0\sigma_2, \ldots, \sigma_{2h-1}w_0,w_0\sigma_{2h}, \sigma_{2h+1},\ldots,\sigma_{k}),$$ where $h$ is an integer such that $2h\leq k$ and $w_0=(n, n-1, \ldots, 1)$ is the top element of $W$. Observing that ${\mathrm{Des}}(\sigma)={\mathrm {Coasc}}(\sigma w_0)={\mathrm {Asc}}(w_0\sigma)$ for all $\sigma\in W$, this map proves bijectively that $C(k,k,k,Id)=C(k-2h,k-2h,k-h,\pi)$, where $\pi=(2h+1,2h+2,\ldots,k,2,4,\ldots,2h, 1,3,\ldots, 2h-1)$ and $C(k-1,k-1,k-1,Id)=C(k-1-2h,k-1-2h,k-1-h),\pi')$, where $\pi'=(2h+1,2h+2,\ldots,k-1,2,4,\ldots,2h, 1,3,\ldots, 2h-1,k)$) and the proof is complete.
As an example on how we can obtain properties on the Kronecker coefficients by means of Proposition \[duni\] avoiding their definition in terms of characters of the symmetric group we show the following result. If $\lambda$ is a partition we denote by $\lambda'$ the partition conjugate to $\lambda$.
We have $$d_{{\lambda^{(1)},\ldots,\lambda^{(k)}}}=d_{ {\lambda^{(1)'}},\lambda^{(2)'},\lambda^{(3)},\ldots,\lambda^{(k)}}$$
It is clear that transposition $T \mapsto T^t$ is a bijection between standard tableaux of conjugate shapes such that ${\mathrm{Des}}(T)={\mathrm {Asc}}(T^t)$. Then, by Corollary \[rsmulti\] and Theorem \[dcac\] we have $$\begin{aligned}
\sum_{\substack{(T_1,\ldots,T_k)\in {\mathcal {ST}}^k:\\ {\mathrm{Des}}(T_i)=D_i}}d_{\lambda(T_1)',\lambda(T_2)',\lambda(T_3)\ldots\lambda(T_k)} & = & \sum_{\substack{(T_1,\ldots,T_k)\in {\mathcal {ST}}^k:\\ {\mathrm {Asc}}(T_i)=D_i,\,i=1,2\\ {\mathrm{Des}}(T_i)=D_i,\,i\geq3}}d_{\lambda(T_1),\ldots,\lambda(T_k)}
= \sum_{\substack{({\sigma_1,\ldots, \sigma_k})\in W^{(k)}:\\ {\mathrm {Asc}}(\sigma_i)=D_i,\,i=1,2\\ {\mathrm{Des}}(\sigma_i)=D_i,\,i\geq3}}1\\
& = & \sum_{\substack{({\sigma_1,\ldots, \sigma_k})\in W^{(k)}:\\ {\mathrm{Des}}(\sigma_i)=D_i}}1 = \sum_{\substack{(T_1,\ldots,T_k)\in {\mathcal {ST}}^k:\\ {\mathrm{Des}}(T_i)=D_i}}d_{\lambda(T_1),\ldots, \lambda(T_k)}.\end{aligned}$$ Now we can conclude recalling the uniqueness of the coefficients $d_{{\lambda^{(1)},\ldots,\lambda^{(k)}}}$ given by Proposition \[duni\].\
We note that this result can also be easily obtained from the definition of the Kronecker coefficients recalling that the irreducible representation indexed by $\lambda'$ is isomorphic to the tensor product of the irreducible representation indexed by $\lambda$ with the alternant representation.
We conclude this paper with the following result that we state without proof. The Robinson-Schensted correspondence can be naturally generalized to the so-called RSK-correspondence (where the “K” stands for Knuth) between multisets with support in $\mathbb N^2$ of cardinality $n$ and pairs of semistandard tableaux of the same shape. One can show that an analogous correspondence exists also in the multivariate case between multisets with support in $\mathbb N^k $ of cardinality $n$ and $k$-tuples of semistandard Young tableaux of size $n$ preserving “descents”. Here the descent sets ${\mathrm{Des}}_1(A),\ldots{\mathrm{Des}}_k(A)$ of a multiset $A$ are by definition the descent sets of the permutations appearing in Theorem \[parpar\] for the $k$-partite partition obtained reordering the elements of $A$. The descents of a semistandard tableau are the descents of its standardized tableau.
There exists a correspondence $\mathcal T$ between multisets with support in $\mathbb N^k $ of cardinality $n$ and $k$-tuples of semistandard Young tableaux of size $n$ such that,
1. $|\mathcal T^{-1}(T_1,\ldots,T_k)|=d_{\mu(T_1),\ldots,\mu(T_k)}$;
2. if $\mathcal T(A)=(T_1,\ldots,T_k)$ then the multiplicity of $i$ in $T_j$ is equal to the multiplicity of $i$ within the $j$-th coordinates of the elements of $A$ and
3. ${\mathrm{Des}}(T_i)={\mathrm{Des}}_i(A)$.
[**Acknowkedgements.**]{} A particular gratitude goes to Vic Reiner for his precious comments on a preliminary draft of this paper.
[1]{} , Descent representations and multivariate statistics, Trans. Amer. Math. Soc. 357 (2005), no. 8, 3051–3082 , The descent monomials and a basis for the diagonally symmetric polynomials, J. Algebraic Combin. 3 (1994), 5–16 , The flag major index and group actions on polynomial rings, Europ. J. Combin. 22 (2001), 431–446. , Colored-descent representations of complex reflection groups $G(r,p,n)$. Israel J. Math. 160 (2007), 317–347 , Bimahonian distributions, J. London Math. Soc. 77 (2008), 627–646. , Decomposition of the diagonal action of $S\sb n$ on the coinvariant space of $S\sb n\times S\sb n$. Sém. Lothar. Combin. 52 (2004/07), Art. B52e. , On Kronecker products of complex representations of the symmetric and alternating groups, Pacific J. Math. 190 (1999), no. 2, 201–223. , Invariant algebras and major indices for classical Weyl groups, Proc. London Math. Soc. (3) 88 (2004), no. 3, 603–631. , A descent basis for the coinvariant algebra of type $D$, J. Algebra 275 (2004), 517–539. , Invariants of finite groups generated by reflections, Amer. J. Math. 77 (1955), 778–782. , Fonctions symétriques et séries hypergéométriques basiques multivariées, Bull. Soc. Math. France 113 (1985), no. 1, 3–22. , On the Kronecker product of $S\sb n$ characters,1 J. Algebra 154 (1993), no. 1, 125–140. , Major index and inversion number of permutations, Math. Nachr. 83 (1978), 143–159. , Permutation statistics and partitions, Adv. Math. 31 (1979) 288–305. , Group actions of Stanley-Reisner rings and invariants of permutations groups, Adv. Math. 51 (1984), 107–201. , Two theorems on multipartite partitions, J. London Math. Soc. 38 (1963) 459–464. , Multipartite P-partitions and inner products of skew Schur functions, Combinatorics and algebra 289–301, Contemp. Math., 34, Amer. Math. Soc., Providence, RI, 1984. , Algebra of coinvariants and the action of a Coxeter element, Bayreuth. Math. Schr. No. 63 (2001), 265–284. , Combinatory analysis, vol. 1, Cambridge University Press, London, 1915. , On the height of the Kronecker product of $S\sb{n}$ characters, Israel J. Math. 42 (1982), no. 1-2, 60–64. , Quotients of Coxeter complexes and $P$-partitions. Mem. Amer. Math. Soc. 95 (1992), no. 460. , On the representations of the symmetric group, Amer. J. Math. 60 (1938), no. 3, 745–760. , The Kronecker product of Schur functions indexed by two-row shapes or hook shapes, J. Algebraic Combin. 14 (2001), 153–173. , Longest increasing and decreasing subsequences, Canad. J. Math. 13 1961 179–191. , Quelques remarques sur une construction de Schensted, Math. Scand. 12 (1963), 117–128. , Finite unitary reflection groups, Canadian J. Math. 6 (1954), 274–304. , Partition identities and invariants of finite groups, J. Combin. Theory Ser. A 23 (1977), 148–175. , Invariants of finite groups and their applications to combinatorics, Bull. Amer. Math. Soc. 1 (1979), 475–511. , Enumerative combinatorics, vol. 2, Cambridge Studies in Advanced Mathematics 62, Cambridge University Press, Cambridge, 1999.
[Dipartimento di matematica, Università di Bologna,\
Piazza di Porta San Donato 5,\
Bologna 40126, Italy]{}\
*E-mail address:* [[email protected]]{}
[^1]: *MSC:* 05E10, 05A19
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using the energy variational method of magneto-solid-mechanical theory of a perfectly conducting elastic medium threaded by magnetic field, the frequency spectrum of Lorentz-force-driven global torsional nodeless vibrations of a neutron star with Ferraro’s form of axisymmetric poloidal nonhomogeneous internal and dipole-like external magnetic field is obtained and compared with that for this toroidal Alfvén mode in a neutron star with homogeneous internal and dipolar external magnetic field. The relevance of considered asteroseismic models to quasi-periodic oscillations of the X-ray flux during the ultra powerful outbursts of SGR 1806-20 and SGR 1900+14 is discussed.'
author:
- 'S.I. Bastrukov, H.-K. Chang, I.V. Molodtsova, E.-H. Wu, G.-T. Chen, S.-H. Lan'
title: 'Frequency spectrum of toroidal Alfvén mode in a neutron star with Ferraro’s form of nonhomogeneous poloidal magnetic field'
---
[**Keywords**]{} Neutron Stars, Asteroseismology, Torsional Alfvén Oscillations
Introduction
============
In the context of recent discovery of quasi-periodic oscillations (QPOs) in the X-ray luminosity during the giant flare of SGR 1806-20 and SGR 1900+14 that were interpreted as being produced by torsional vibrations of quaking magnetars (Israel et al. 2005, Watts & Strohmayer 2006), in (Bastrukov et al. 2009a, 2009b) several scenarios of the post-quake vibrational relaxation of a neutron star model with uniform internal and dipolar external magnetic field $$\begin{aligned}
\label{e1.1}
&& B_r=B\cos\theta,\quad B_\theta=-B\sin\theta,\quad B_\phi=0,\quad r\leq R\\
\label{e1.2}
&& B_r=B\left(\frac{R}{r}\right)^3\cos\theta,\quad B_\theta=-\frac{B}{2}\left(\frac{R}{r}\right)^3\sin\theta,
\quad B_\phi=0,\quad r > R
\end{aligned}$$ have been studied on the basis of equations of Newtonian magneto-solid-mechanics $$\begin{aligned}
\label{e1.3}
&& \rho{\ddot {\bf u}}=\frac{1}{c}[\delta {\bf j}\times {\bf B}],\quad
\delta {\bf j}=\frac{c}{4\pi}[\nabla\times \delta {\bf B}], \quad \delta {\bf B}({\bf r},t)=\nabla\times [{\bf u}\times {\bf B}].
\end{aligned}$$ These equations describe the Lorentz-force-driven non-compressional ($\delta \rho=-\rho \nabla\cdot {\bf u}=0$) fluctuations of star matter about axis of above fossil magnetic field ${\bf B}$ in terms of fluctuating material displacements $\bf u$ (the basic dynamical variable of solid mechanics) and the magnetic field $\delta {\bf B}$. It is implied that elastic stellar material is of an extremely high electrical conductivity[^1]. The chief argument for interpreting QPOs during the outbursts of above mentioned SGRs (detected on descending branch of the giant flare light-curves) as caused by torsional vibrations of a solid star driven by restoring force of magnetic field stresses is that it is an ultra strong magnetic field frozen-in the entire volume of magnetars serves as the main energy source and promoter of their X-ray bursting seismic activity (e.g., Woods & Thompson 2006, Mereghetti 2008) and, also, bearing in mind the fact that the very notion of torsional vibrations has come into theoretical seismology from the solid-mechanical theory of shear vibrations of an elastic sphere (e.g. Lapwood, & Usami 1981, Lay & Wallace 1995, Aki & Richards 2002, Stein & Wyssesson 2003). It worthy noting that theoretical investigations of non-radial torsional Alfvén oscillations of a fluid star in its own homogeneous magnetic field have a long story that was started, to the best of our knowledge, in works of Jensen (1955) and Plumpton & Ferraro (1955). Remarkably that in the latter paper, by emphasizing the basic discovery of Alfvén that perfectly conducting fluid threaded by magnetic field behaves like anisotropic elastic medium capable of transmitting mechanical disturbance by transverse hydromagnetic waves (e.g. Fälthammar 2007), it is argued that eigenfrequency problem of such vibrations can be tackled on the basis of equation $$\begin{aligned}
\label{e1.4}
&& \rho{\ddot {\bf u}}=\frac{1}{4\pi}[\nabla\times[\nabla\times [{\bf u}\times {\bf B}]]]\times {\bf B},\quad \nabla\cdot {\bf u}=0,\quad \nabla\cdot {\bf B}=0
\end{aligned}$$ which, as is evident, follows form equations (\[e1.3\]). It is clear from this last equation that the frequencies of Alfvén oscillations must substantially depend on both geometrical configuration of internal equilibrium magnetic field ${\bf B}$ and analytic form of oscillating field of material displacements ${\bf u}$. With this obvious observation in mind, in (Bastrukov et al. 2009a, 2009b) focus was laid on non-investigated before regime of large lengthscale nodeless Alfvén oscillations, both global (in the entire spherical volume of star) and crustal (locked in the peripheral finite-depth spherical layer). The most conspicuous feature of this regime is that the radial dependence of oscillating material displacements field ${\bf u}$ has no nodes. In a star undergoing global nodeless torsional oscillations, which are of particular interest for our present discussion, the fluctuating material displacements are described by the toroidal field of the form (Bastrukov et al 2007a, 2007b, 2009a) $$\begin{aligned}
\label{e1.5}
&& {\bf u}({\bf r},t)=[\mbox{\boldmath $\phi$}({\bf r})\times {\bf r}]\,\alpha(t),\quad \alpha(t)=\alpha_0\cos\omega t,\\
\label{e1.6}
&& \delta {\bf v}({\bf r},t)={\dot {\bf u}}({\bf r},t)=[\delta \mbox{\boldmath $\omega$}({\bf r},t)\times {\bf r}], \quad
\delta\mbox{\boldmath $\omega$}({\bf r},t)=\mbox{\boldmath $\phi$}({\bf r})\,{\dot \alpha}(t),\\
\label{e1.7}
&& \mbox{\boldmath $\phi$}({\bf r})=\nabla\chi({\bf r}),\quad \nabla^2\chi({\bf r})=0,\quad \chi({\bf r})=
\frac{{\cal A}_\ell}{\ell+1}\,r^\ell\,P_\ell(\cos\theta)
\end{aligned}$$ where $P_\ell$ is the Legender polynomial of degree $\ell$ and the other symbols have their usual meaning. Fig.1 illustrates the nodeless character of displacements in the star undergoing global non-radial differentially rotational, torsional, shear vibrations about polar axis in quadrupole and octupole overtones.
{width="9.0cm"}
With the aid of the Rayleigh’s energy method which is expounded below it was found that discrete frequencies of such vibrations is given by the following one-parametric spectral formula (Bastrukov et al. 2009a) $$\begin{aligned}
\label{e1.8}
&& \omega(_0a^t_\ell)=\omega_A\left[(\ell^2-1)\frac{2\ell+3}{2\ell-1}\right]^{1/2},\,\,
\omega_A=\frac{v_A}{R}=\sqrt{\frac{RB^2}{3M}},\\
\label{e1.9}
&& \quad v_A=\frac{B}{\sqrt 4\pi \rho},\quad M=\frac{4\pi}{3}\rho\,R^3
\end{aligned}$$ the only parameter of which is the Alfvén frequency, $\omega_A$, of shear magneto-elastic oscillations of perfectly conducting stellar matter matter pervaded by magnetic field $B$ in the star of radius $R$ and mass M. It must be emphasized, however, that this theoretical spectrum does not properly match the QPOs in the X-ray flux from flaring SGR 1806-20 and SGR 1900+14. One of reasons of this discrepancy may be inadequate assumption about homogeneous configuration of internal magnetic field and perhaps the most efficient way to clarify this conjecture is to investigate a model with geometrically different configuration of axisymmetric internal magnetic field. Before so doing it seems worth noting that the model of a star with [*homogeneous*]{} internal and dipolar external magnetic field has come into focus in astrophysics after seminal work of Chandrasekhar and Fermi (1953) in which the effect of mechanical flattening of the star at the poles of such magnetic field has been disclosed. Shortly after, similar conclusion has been drawn in outstanding paper of Ferraro (1954), but on the basis of star model with substantially [*nonhomogeneous*]{} internal and dipole-like axisymmetric poloidal magnetic field $$\begin{aligned}
\label{e1.10}
&& B_r=\frac{1}{r^2\sin\theta}\frac{\partial U}{\partial \theta},\quad B_\theta=-\frac{1}{r\sin\theta}\frac{\partial U}{\partial r},\quad B_\phi=0,\\
\label{e1.11}
&& U=U_{in}=\frac{B}{4R^2}\,r^2(3r^2-5R^2)\sin^2\theta,\quad r\leq R,\\
\label{e1.12}
&& U=U_{ex}=\frac{B}{2r}\,R^3 \sin^2\theta,\quad r> R
\end{aligned}$$ where $B$ stands for the magnetic field intensity at the poles and $\nabla\cdot {\bf B}=0$ as should be the case[^2]. The meridian cross section of Ferraro’s model of the star is sketched in Fig.2.
{width="9.0cm"}
Over the years, the different aspects, both mathematical and astrophysical, of Ferraro’s model have been the subject of extensive investigations (e.g. Chandrasekhar & Prendergast 1955, Robetrs 1955, Chandrasekhar 1956, Mestel 1956, Ledoux & Walraven 1958, Monaghan 1965, Ledoux & Renson 1966, Sood & Trehan 1970, Goossens 1972, Goossens, Smeyers & Denis 1976). The effects of Ferraro’s configuration of magnetic field (and magnetic fields of similar geometrical configuration) on the equilibrium shape, vibrational behavior and electromagnetic activity of pulsars and magnetars are considered in (Roberts 1981, Ioka 2001, Braithwaite & Spruit 2006, Geppert & Rheinhardt 2006, Haskell et al 2008, Lee 2008; Broderick & Narayan 2008, see also references therein).
In this work we focus on the non-studied before regime of global torsional Alfvén nodeless vibrations of neutron star about axis of Ferraro’s magnetic field (\[e1.10\])-(\[e1.12\]). In Section 2, the frequency spectrum of this toroidal mode is derived and compared with the frequency spectrum (\[e1.8\]) of the neutron star model with homogeneous internal magnetic field. In Section 3, the obtained spectral formula for the frequency is analyzed numerically in juxtaposition with data on QPOs during the flare of SGR 1806-20 and SGR 1900+14. Section 4 briefly accounts for the net outcome of this work. Technical details of analytic computations can be found in Appendix.
Global Alfvén torsional nodeless oscillations of neutron star in its own poloidal magnetic field of Ferraro’s form
==================================================================================================================
In the model under consideration a neutron star is identified with a finite spherical mass of an elastic solid, regarded as an incompressible continuous medium of uniform density $\rho$ and an infinite electrical conductivity, whose vibrations under the action of Lorentz magnetic force are governed by equations of magneto-solid-mechanics (\[e1.2\]) which can conveniently be represented in the following equivalent tensor form (e.g. Mestel 1999) $$\begin{aligned}
\label{e2.1}
&&\rho{\ddot u}_i=\nabla_k\, \delta M_{ik},\quad \delta M_{ik} =\frac{1}{4\pi}[B_i\,\delta B_k+B_k\,\delta B_i-(B_j\,\delta B_j)\delta_{ik}], \\
&& \delta B_i({\bf r},t)=(B_k \nabla_k)u_i-(u_k\nabla_k) B_i,\quad \nabla_i u_i=0
\end{aligned}$$ where $\delta M_{ik}$ stands for the Maxwell’s tensor of magnetic field stresses. The energy balance in the process of vibrations is controlled by equation $$\begin{aligned}
\label{e2.2}
&& \frac{\partial }{\partial t}\int \frac{\rho{\dot u}^2}{2}\,d{\cal V}
=- \int \delta M_{ik}\,{\dot u}_{ik}\,d{\cal V}=-\frac{1}{8\pi}\int [B_i\,\delta B_k+B_k\,\delta B_i]\,
[\nabla_i {\dot u}_k + \nabla_k {\dot u}_i]\,d{\cal V},\\
\label{e2.3}
&& {\dot u}^2={\dot u}_i\,{\dot u}_i,\quad {\dot u}_{ik}=\frac{1}{2}[\nabla_i {\dot u}_k+\nabla_k {\dot u}_i],\quad
{\dot u}_{kk}=\nabla_k {\dot u}_k=0.
\end{aligned}$$ To compute the eigenfrequency of toroidal Alfvén mode in question we take advantage of the Rayleigh’s energy method which has been utilized in our previous above mentioned investigations. The key idea of this method consists in separable representation of fluctuating variables such as the vector field of material displacements $u_i({\bf r},t)$ and the tensor field of shear strains $u_{ik}({\bf r},t)$ $$\begin{aligned}
\label{e2.4}
u_i({\bf r},t)=a_i({\bf r})\,{\alpha}(t),\quad u_{ik}({\bf r},t)=a_{ik}({\bf r})\,\,{\alpha}(t),\quad a_{ik}({\bf r})=\frac{1}{2}[\nabla_i a_k({\bf r})+\nabla_k a_i({\bf r})].
\end{aligned}$$ With this form of $u_i$, the magnetic flux density $\delta B_i({\bf r},t)$ and the tensor field of fluctuating magnetic field stresses $\delta M_{ik}({\bf r},t)$ are represented in a similar manner $$\begin{aligned}
\label{e2.5}
&& \delta B_i({\bf r},t)=b_i({\bf r})\,\alpha(t),\quad b_i({\bf r})=(B_k \nabla_k)a_i-(a_k\nabla_k) B_i,\\
\label{e2.6}
&& \delta M_{ik}({\bf r},t)=[\tau_{ik}({\bf r})-\frac{1}{2}\tau_{jj}\delta_{ik}]\alpha(t),\quad \tau_{ik}({\bf r})=\frac{1}{4\pi}[B_i({\bf r})\,b_k({\bf r})+B_k({\bf r})\,b_i({\bf r})].
\end{aligned}$$ The gist of this multiplicative decomposition of fluctuating variables is that on substituting (\[e2.4\])-(\[e2.6\]) in (\[e2.2\]) this latter equation is reduced to equation for time-dependent amplitude $\alpha(t)$ having the well-familiar form $$\begin{aligned}
\label{e2.7}
&& {\cal M}{\ddot \alpha}(t)+{\cal K}_m\alpha(t)=0,\quad {\cal M}=\int \rho\, a_i\,a_i d{\cal V},\\
\label{e2.8}
&& {\cal K}_m=\int \tau_{ik}\,a_{ik}\,d{\cal V}=\frac{1}{8\pi}\int [B_i\,b_k+B_k\,b_i]\,[\nabla_i\,a_k+\nabla_k\,a_i]\,d{\cal V}.
\end{aligned}$$ Thus, from technical argument, the computation of frequency $\omega=[{\cal K}/{\cal M}]^{1/2}$ is reduced to calculation of integral parameters of inertia ${\cal M}$ and stiffness ${\cal K}_m$ with the toroidal field of instantaneous, time-independent, displacements $$\begin{aligned}
\label{e2.9}
&& {\bf a}_t=A_t\,\nabla \times [{\bf
r}\,r^\ell\,P_\ell(\zeta)]:\,\, a_r=0,\,\,\, a_\theta=0,\,\,\, a_\phi=A_t r^{\ell}(1-\zeta^2)^{1/2}\frac{dP_\ell(\zeta)}{d\zeta}
\end{aligned}$$ and the magnetic field of Ferraro’s form whose spherical components inside the star are $$\begin{aligned}
\label{e2.10}
&& B_r=\frac{3B}{2R^2}\,r^2\left(r^2-\frac{5}{3}R^2\right)\cos\theta,\quad B_\theta=-\frac{3B}{2R^2}\,\left(2r^2-\frac{5}{3}R^2\right)\sin\theta,\quad \quad B_\phi=0.
\end{aligned}$$
{width="12.0cm"}
The torsional inertia ${\cal M}$ as a function of multipole degree $\ell$ of nodeless differentially rotational vibrations in question is given by (Bastrukov et al, 2007, 2008) $$\begin{aligned}
\label{e2.11}
{\cal M}=4\pi\rho A_t^2 R^{2\ell+3}\,m_\ell,\quad m_\ell=\frac{\ell(\ell+1)}{(2\ell+1)(2\ell+3)}.
\end{aligned}$$ To avoid destructing attention from basic inferences of this work, we place all technical details of tedious but simple computations of integrals for ${\cal K}_m$ in Appendix A. The final expression for this coefficient can be represented in the form $$\begin{aligned}
\label{e2.12}
{\cal K}_m=B^2\,A_t^2\,R^{2\ell+1}\, k_\ell,\quad
k_\ell=\frac{\ell \left(\ell^2-1 \right) \left(5\ell^3+7\ell^2+59\ell+84 \right) }
{2\left(4\ell^2-1 \right)\left(2\ell+3 \right) \left(2\ell+5 \right) }.
\end{aligned}$$ And for the frequency spectrum of global nodeless torsional Alfvén vibrations nodeless of the neutron star with Ferraro’s form of nonhomogeneous internal magnetic field we obtain $$\begin{aligned}
\label{e2.14}
&& \nu(_0a^t_\ell) =\nu_A\left[\frac{\left(\ell-1 \right) \left( 5\ell^3+7\ell^2+59\ell+84\right) }{2\left(2\ell-1 \right) \left(2\ell+5 \right)}\right]^{1/2},\\
&& \nu=\frac{\omega}{2\pi},\quad \omega_A=\frac{v_A}{R},\quad v_A=\frac{B}{\sqrt{4\pi\rho}},\quad \omega_A=B\sqrt{\frac{R}{3M}},\quad M=\frac{4\pi}{3}\rho\,R^3.\end{aligned}$$ It follows that the lowest overtone of this toroidal Alfvén mode is of quadrupole degree, $\ell=2$. At $\ell=1$, the parameter of magneto-mechanical rigidity of neutron star matter cancels, ${\cal K}_m(_0a^t_1)=0$, and the mass parameter equals to the moment of inertia of rigid sphere, ${\cal M}(_0a^t_1)={\cal J}=(2/5)MR^2$. It follows from Hamiltonian of normal vibrations, ${\cal H}=(1/2){\cal M}{\dot \alpha}^2+(1/2){\cal K}{\alpha}^2$, that in this dipole case a star sets in rigid-body rotation, rather than vibrations, about axis of its dipole magnetic moment; this feature of the model under consideration is quite similar to that of the neutron star model with homogeneous internal magnetic field. In Fig 3., we plot the frequency $\nu(_0a^t_\ell)$ and the period $P(_0a^t_\ell)=\nu^{-1}(_0a^t_\ell)$ of the Alfvén toroidal mode as functions of multipole degree computed in both homogeneous and nonhomogeneous neutron star models with indicated parameters.
{width="8.0cm"}
QPOs in X-ray luminosity of flaring SGR 1806-20 and SGR 1900+14 from the viewpoint of considered model
======================================================================================================
As was mentioned, the one-parametric spectral formula (\[e1.8\]), computed in the neutron star model with homogeneous internal and dipolar external magnetic field, does not reproduce general trends in data on QPOs frequencies whose numerical values for SGR 1806-20 are given by $\nu_{\rm data}=$18, 26, 30, 92, 150, 625, 1840 and for the SGR 1900+14 these are $\nu_{\rm data}=$28, 54, 84, 155 (Watts & Strohmayer 2007). It is tempting, therefore, to consider these data from the view point of investigated model by identifying the observed QPOs with overtones of spectral formula (\[e2.14\]). The result of $\ell$-pole modal specification of the detected QPOs frequencies as overtones of torsional Alfvén seismic vibrations in question is presented in Fig.4 and Fig.5. Specifically, for SGR 1900+14 we obtain: $\nu(_0a^t_2)=28$ Hz; $\nu(_0a^t_4)=53$; Hz $\nu(_0a^t_6)=84$ Hz; $\nu(_0t_{13})=155$ Hz, and for the SGR 1806-20 we get $\nu(_0a^t_2)=26$ Hz; $\nu(_0a^t_3)=30$, $\nu(_0a^t_{10})=92$ Hz; $\nu(_0a^t_{16})=155$ Hz; $\nu(_0a^t_{65})=625$ Hz and $\nu(_0a^t_{180})=1840$ Hz. While the lowest of detected oscillations, with $\nu=18$ Hz, cannot be specified in terms of considered seismic vibrations, it is clearly seen that the obtained spectrum correctly reflects general trends in the detected QPO frequencies. This suggests, if the detected QPOs are really produced by Lorentz-force-driven global nodeless torsional seismic vibrations about the dipole magnetic moment of magnetars, their internal magnetic fields should be of substantially nonhomogeneous configuration.
{width="12.0cm"}
The result of $\ell$-pole modal specification of the detected QPOs frequencies as overtones of torsional Alfvén seismic vibrations in question is presented in Fig.4 and Fig.5. Specifically, for SGR 1900+14 we obtain: $\nu(_0a^t_2)=28$ Hz; $\nu(_0a^t_4)=53$; Hz $\nu(_0a^t_6)=84$ Hz; $\nu(_0t_{13})=155$ Hz, and for the SGR 1806-20 we get $\nu(_0a^t_2)=26$ Hz; $\nu(_0a^t_3)=30$, $\nu(_0a^t_{10})=92$ Hz; $\nu(_0a^t_{16})=155$ Hz; $\nu(_0a^t_{65})=625$ Hz and $\nu(_0a^t_{180})=1840$ Hz. While the lowest of detected oscillations, with $\nu=18$ Hz, cannot be specified in terms of considered seismic vibrations, it is clearly seen that the obtained spectrum correctly reflects general trends in the detected QPO frequencies. This suggests, if the detected QPOs are really produced by Lorentz-force-driven global nodeless torsional seismic vibrations about the dipole magnetic moment of magnetars, their internal magnetic fields should be of substantially nonhomogeneous configuration.
Concluding remarks
==================
Any attempt to predict the behavior of Alfvén vibrational modes in pulsars and magnetars, presuming of course that the star material possesses properties of perfectly conducting continuous medium pervaded by magnetic fields, is beset with uncertainties regarding geometrical configuration of fossil internal magnetic field. It seems, therefore, that progress can be best made by studying these modes within the framework of comprehensive models. Among these are the models with homogeneous and nonhomogeneous axisymmetric poloidal magnetic fields considered long ago in works of Chandrasekhar and Fermi (1953) and by Ferraro (1954), respectively, to show that such fields have the same effect as rigid rotation, that is, tend to produce a flattening of the star shape along the magnetic field axis. Following this line of argument and continuing investigations reported in (Bastrukov et al 2009a), we have computed here the frequency spectrum of axisymmetric torsional nodeless vibrations, in the neutron star model with Ferraro’s form of nonhomogeneous poloidal magnetic field which is presented in Fig.3 in juxtaposition in a neutron star model with homogeneous internal field.
The practical usefulness of the obtained one-parametric spectral formula has been demonstrated by its application to $\ell$-pole identification of QPOs frequencies during the X-ray giant outbursts of SGR 1900+14 and SGR 1806-20. The result of our analysis, summarized in Fig.4 and Fig.5, shows that the model adequately regains the overall trends in the detected QPOs frequencies and, thus, supports theoretical interpretation of these QPOs, advanced in works reporting this discovery (Israel et al 2005, Watts & Strohmayer 2006), as owing their origin to quake-induced torsional seismic vibrations of underlying magnetar. Together with this, in (Bastrukov et al 2009b) it has been shown that the same data on the QPO frequencies can be consistently interpreted from the view point of two-component, core-crust, model of quaking neutron star (Franco et al. 2000) with homogeneous internal magnetic field as being produced by axisymmetric differentially rotational Alfvén nodeless oscillations of crustal solid-state plasma about axis of magnetic field frozen in the immobile core. With all these in mind, we conclude that it is the Lorentz force of magnetic field stresses plays decisive part in post-quake vibrational relaxation of above magnetars and that the toroidal fields of quake-induced material displacements are of substantially nodeless character.
Authors are grateful to Dr. Dima Podgainy for helpful assistance. This work has been supported by NSC of Taiwan, grant numbers NSC-098-2811-M-007-009 and NSC-96-2628-M-007-012-MY3.
Torsional stiffness of global torisonal Alfvén nodeless oscillations of a neutron star about axis of Ferraro’s nonhomogeneous poloidal magnetic field
=====================================================================================================================================================
In computing stiffness of torsional Alfvén oscillations $$\begin{aligned}
\nonumber
&& {\cal K}_m=\int \tau_{ik}({\bf r})\,a_{ik}({\bf r})\,d{\cal V}, \quad a_{ik}({\bf r})=\frac{1}{2}[\nabla_i\,a_k({\bf r})+\nabla_k\,a_i({\bf r})],\\
\nonumber
&& \tau_{ik}({\bf r})=\frac{1}{4\pi}[B_i({\bf r})\,b_k({\bf r})+B_k\,({\bf r})b_i({\bf r})],\quad b_i({\bf r})=(B_k ({\bf r})\nabla_k)a_i({\bf r})-(a_k({\bf r})\nabla_k) B_i({\bf r})
\end{aligned}$$ it is convenient to represent strain tensor $$\begin{aligned}
\nonumber
a_{ik}=\frac{1}{2}(\nabla_i a_k + \nabla_k a_i)
\end{aligned}$$ in spherical polar coordinates with use of the angle variable ${\zeta}=\hbox{cos}\,{\theta}$. In terms of this variable, the components of these tensor are $$\begin{aligned}
\nonumber
&&a_{rr}=\frac{\partial a_r}{\partial r},\quad
\quad\quad
a_{\theta\theta}=-\frac{(1-\zeta^2)^{1/2}}{r}
\frac{\partial a_r}{\partial \zeta}+\frac{a_r}{r},\\
\nonumber
&& a_{\phi\phi}=\frac{1}{r}\frac{1}{(1-\zeta^2)^{1/2}}
\frac{\partial{a_\phi}}{\partial{\phi}}+\frac{a_r}{r}+
\frac{\zeta}{(1-\zeta^2)^{1/2}}\frac{a_\theta}{r},
\\
\nonumber
&&a_{r\theta}=\frac{1}{2}\left[-\frac{(1-\zeta^2)^{1/2}}{r}
\frac{\partial a_r}{\partial \zeta}-\frac{a_\theta}{r}+
\frac{\partial a_\theta}{\partial r}\right],
\\
\nonumber
&&a_{r\phi}=\frac{1}{2}\left[\frac{1}{r}\frac{1}{(1-\zeta^2)^{1/2}}
\frac{\partial a_r}{\partial \phi}-\frac{a_\phi}{r}+
\frac{\partial a_\phi}{\partial r}\right],
\\
\nonumber
&&a_{\theta\phi}=\frac{1}{2}\left[\frac{1}{r}\frac{1}{(1-\zeta^2)^{1/2}}
\frac{\partial a_\theta}{\partial \phi}-
\frac{\zeta}{(1-\zeta^2)^{1/2}}\frac{a_\phi}{r}-
\frac{(1-\zeta^2)^{1/2}}{r}\frac{\partial a_\phi}{\partial \zeta}\right].
\end{aligned}$$ In the torsional mode of nodeless vibrations the field of instantaneous displacements has solely one non-zero $\phi-th$ component $$\begin{aligned}
\nonumber
a_r=0\quad a_\theta=0\quad a_{\phi}=A_t\,r^\ell(1-\zeta^2)^{1/2}P'_\ell(\zeta),\quad
P'_\ell(\zeta)=\frac{dP_\ell(\zeta)}{d\zeta}.
\end{aligned}$$ In this case we have only two non-zero components of the strain tensor $$\begin{aligned}
\nonumber
&& a_{rr}=a_{\theta\theta}=a_{\phi\phi}=a_{r\theta}=0,\\
\nonumber
&& a_{r\phi}=\frac{A_t}{2}r^{\ell-1}
(\ell-1)(1-\zeta^2)^{1/2}P'_\ell,\quad a_{\theta\phi}=-\frac{A_t}{2}r^{\ell-1}
\left[2\zeta P'_\ell-\ell(\ell+1)P_\ell(\zeta)\right].
\end{aligned}$$ In spherical polar coordinates the components of vector field $b_i({\bf r})=(B_k \nabla_k)a_i-(a_k\nabla_k) B_i$ are given by $$\begin{aligned}
\nonumber
b_r&=&\left[B_r\frac{\partial }{\partial r}-\frac{B_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial }{\partial \zeta}+
\frac{B_\phi}{r}(1-\zeta^2)^{-1/2}\frac{\partial }{\partial \phi}\right]\,a_r-
\frac{B_\theta\,a_\theta+B_\phi\,a_\phi}{r}\\
\nonumber
&-&
\left[a_r\frac{\partial }{\partial r}-
\frac{a_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial }{\partial \zeta}+
\frac{a_\phi}{r}(1-\zeta^2)^{-1/2}\frac{\partial }{\partial \phi}\right]\,B_r+
\frac{a_\theta\,B_\theta+a_\phi\,B_\phi}{r},
\\[0.3cm]
\nonumber
b_\theta&=&\left[B_r\frac{\partial }{\partial r}-\frac{B_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial }{\partial \zeta}+
\frac{B_\phi}{r}(1-\zeta^2)^{-1/2}\frac{\partial }{\partial \phi}\right] \,a_\theta
+\frac{B_\theta\,a_r-B_\phi\,a_\phi\,\zeta(1-\zeta^2)^{-1/2}}{r}\\
\nonumber
&-&\left[a_r\frac{\partial }{\partial r}-
\frac{a_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial }{\partial \zeta}+
\frac{a_\phi}{r}(1-\zeta^2)^{-1/2}\frac{\partial }{\partial \phi}\right]
\,B_\theta
-\frac{a_\theta\,B_r-a_\phi\,B_\phi\,\zeta(1-\zeta^2)^{-1/2}}{r},
\\[0.3cm]
\nonumber
b_\phi&=&\left[B_r\frac{\partial }{\partial r}-\frac{B_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial }{\partial \zeta}+
\frac{B_\phi}{r}(1-\zeta^2)^{-1/2}\frac{\partial }{\partial \phi}\right]\,
a_\phi+
\frac{B_\phi\,a_r+B_\phi\,a_\theta\,\zeta(1-\zeta^2)^{-1/2}}{r}\\
\nonumber
&-&\left[a_r\frac{\partial }{\partial r}-
\frac{a_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial }{\partial \zeta}+
\frac{a_\phi}{r}(1-\zeta^2)^{-1/2}\frac{\partial }{\partial \phi}\right]\,
B_\phi-\frac{a_\phi\,B_r+a_\phi\,B_\theta\,\zeta(1-\zeta^2)^{-1/2}}{r}.
\end{aligned}$$ Taking into account that Ferraro’s field has only two non-zero components which can be conveniently represented in the form $$\begin{aligned}
\nonumber
B_r=\frac{2f}{r^2}\,\zeta,\quad B_\theta=-\frac{(1-\zeta)^{1/2}}{r}\,f',\quad B_\phi=0,\quad
f=\frac{B}{4R^2}\,r^2(3r^2-5R^2),\quad f'=\frac{df}{dr}\end{aligned}$$ for the components of $b_i$ we obtain $$\begin{aligned}
\nonumber
b_r=0,\quad b_\theta=0, \quad b_\phi=B_r\frac{\partial a_\phi}{\partial r}-
\frac{B_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial a_\phi}{\partial \zeta}
-\frac{a_\phi\,B_r}{r}-\frac{a_\phi\,B_\theta\,\zeta(1-\zeta^2)^{-1/2}}{r}
\end{aligned}$$ where $$\begin{aligned}
\nonumber
\frac{\partial a_\phi}{\partial r}&=&A_t\,\ell\,r^{\ell-1}(1-\zeta^2)^{1/2}P'_\ell,\quad\quad
\frac{\partial a_\phi}{\partial \zeta}=A_t\,r^\ell\,\left(1-\zeta^2 \right)^{-1/2}\left[\zeta\,P'_\ell-\ell\left(\ell+1 \right) \right].
\end{aligned}$$ The integrand of ${\cal K}_m$ reads $$\tau_{ik}a_{ik}=2\left( \tau_{r\phi}a_{r\phi}+\tau_{\theta\phi}a_{\theta\phi}\right)$$ so that relevant to computation of ${\cal K}_m$ components of tensor $\tau_{ik}=(1/4\pi)[B_i\,b_k+B_k\,b_i]$ are given by $$\begin{aligned}
\nonumber
\tau_{r\phi}&=&\frac{1}{4\pi}\left[B_rB_r\frac{\partial a_\phi}{\partial r}-
\frac{B_rB_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial a_\phi}{\partial \zeta}\,
-\frac{a_\phi\,B_rB_r}{r}-\frac{a_\phi\,B_rB_\theta\,\zeta(1-\zeta^2)^{-1/2}}{r}\right] \\
\nonumber
&=&\frac{A_t}{4\pi}\left\{4(\ell-1)f^2 r^{\ell-5} \zeta^2(1-\zeta^2)^{1/2}P'_\ell + 2f f'r^{\ell-4}\zeta(1-\zeta^2)^{1/2}[2\zeta P'_\ell-\ell(\ell+1)P_\ell]\right\},\\
\nonumber
\tau_{\theta\phi}&=&\frac{1}{4\pi}\left[B_\theta B_r\frac{\partial a_\phi}{\partial r}-
\frac{B_\theta B_\theta}{r}(1-\zeta^2)^{1/2}\frac{\partial a_\phi}{\partial \zeta}\,
-\frac{a_\phi\,B_\theta B_r}{r}-\frac{a_\phi\,B_\theta\, B_\theta\,\zeta(1-\zeta^2)^{-1/2}}{r}\right]\\
\nonumber
&=&\frac{A_t}{4\pi}\left[-2r^{\ell-4}\left(\ell-1 \right)f f'\zeta\left(1-\zeta^2 \right)P'_\ell -r^{\ell-3}f'^2\left(1-\zeta^2\right)\left[2\zeta P'_\ell-\ell\left( \ell+1\right)P_\ell \right]\right].
\end{aligned}$$ The integral for stiffness can be conveniently represented in the form $$\begin{aligned}
\nonumber
{\cal K}_m&=&2\int [ \tau_{r\phi}\,a_{r\phi}+\tau_{\theta\phi}\,a_{\theta\phi}]d{\cal V}=\frac{A_t^2}{2}\left\{4(\ell-1)^2R_{ff}I_1+4(\ell-1)R_{ff'}[2I_1-\ell(\ell+1)I_2] \right . \\
\nonumber
&+&
\left.
R_{f'f'}[4I_1-4\ell(\ell+1)I_2+\ell^2(\ell+1)^2I_3]\right\}
\end{aligned}$$ The integrals $I_i$ are computed with aid of standard recurrence relations between Legendre polynomials (e.g. Abramowitz & Stegan 1964) which yield $$\begin{aligned}
\nonumber
&&I_1=\int\limits_{-1}^{1}\zeta^2(1-\zeta^2) (P'_\ell)^2 d\zeta
=\frac{2\,\ell\,(\ell+1)(2\ell^2+2\ell-3)}{(4\ell^2-1)(2\ell+3)},\\ \nonumber
&&I_2=\int\limits_{-1}^{1}\zeta(1-\zeta^2)P_\ell\,P'_\ell d\zeta
=\frac{2\,\ell\,(\ell+1)}{(4\ell^2-1)(2\ell+3)},\\ \nonumber
&&I_3=\int\limits_{-1}^{1}(1-\zeta^2)P_\ell^2 d\zeta=\frac{4\,(\ell^2+\ell-1)}{(4\ell^2-1)(2\ell+3)},\quad I_4=\int\limits_{-1}^{1}\zeta^2(1-\zeta^2) P_\ell P'_\ell d\zeta
=\,0.\end{aligned}$$ For integrals with function $f=(B/4R^2)[r^2(3r^2-5R^2)]$ we obtain $$\begin{aligned}
\nonumber
&& R_{ff}=\int_0^R f^2(r) r^{2\ell-4} dr= \frac{B^2R^{2\ell+1}}{16}R_1,\quad R_1=\left[\frac{25}{2\ell+1}-\frac{30}{2\ell+3}+
\frac{9}{2\ell+5}\right],\\
\nonumber
&& R_{ff'}=\int_0^R f\left(\frac{df}{dr}\right) r^{2\ell-3} dr=\frac{B^2R^{2\ell+1}}{16}R_2,\quad
R_2=\left[\frac{50}{2\ell+1}
-\frac{90}{2\ell+3}+\frac{36}{2\ell+5}\right],\\
\nonumber
&& R_{f'f'}= \int_0^R \left(\frac{df}{dr}\right)^2 r^{2\ell-2} dr=\frac{B^2R^{2\ell+1}}{16}R_3,\quad R_3=\left[\frac{100}{2\ell+1}-\frac{240}{2\ell+3}+
\frac{144}{2\ell+5}\right].\end{aligned}$$ The resultant expression for the stiffness reads $${\cal K}_m=B^2\,A_t^2\,R^{2\ell+1}\, k_\ell$$ where $$\begin{aligned}
\nonumber
k_\ell&=&\frac{1}{32}\left\{4(\ell-1)^2R_1\,I_1+4(\ell-1)R_2[2I_1-\ell(\ell+1)I_2]+ R_3[4I_1-4\ell(\ell+1)I_2+\ell^2(\ell+1)^2I_3]\right\}\\ \nonumber
&=&\frac{\ell \left(\ell^2-1 \right) }
{2\left(4\ell^2-1 \right)\left(2\ell+3 \right) \left(2\ell+5 \right) }\left(5\ell^3+7\ell^2+59\ell+84 \right).
\end{aligned}$$
[100]{}
Abramowitz M. & Stegan I.A. 1964, Handbook of Mathematical Functions. Dover, New York
Aki K. & Richards P. G., 2002, Quantitative Seismology. University Science Books
Bastrukov, S. I., Chang, H.-K., Mişicu, Ş., Molodtsova, I. V. & Podgainy D. V. 2007a, Int. J. Mod. Phys. A, 22, 3261
Bastrukov, S. I., Chang, H.-K., Takata, J., Chen, G.-T. & Molodtsova I. V. 2007b, MNRAS, 382, 849
Bastrukov, S. I., Chang, H.-K., Chen, G.-T. & Molodtsova I. V. 2008, Mod. Phys. Lett. A., 23, 477
Bastrukov, S. I., Chen, G.-T., Chang H.-K., Molodtsova, I. V., Podgainy, D. V. 2009a, ApJ, 690, 998
Bastrukov, S. I., Chang, H.-K., Molodtsova I. V. & , Takata, J., 2009b, arXiv e-prints, 0812.4524
Bhattacharya, D. & van den Heuvel E. P. J. 1991, Phys. Rep., 2003, 1
Braithwaite, J. & Spruit, H. C. 2006, A&A, 450, 1097
Broderick, A. E. & Narayan, R. 2008, MNRAS, 383, 943
Chandrasekhar, S. & Fermi 1953, ApJ, 118, 116
Chandrasekhar, S. & Prendergast K. H. 1955, PNAS, 42, 5
Chandrasekhar, S. 1956, ApJ, 124, 232
Chandrasekhar, S. 1961, Hydromagnetic and Hydrodynamic Stability. Oxford University Press
Chanmugam, G. 1992, , 65, 301
Geppert, U. & Rheinhardt, M. 2006, A&A, 456, 639
Goldreich, P. & Reisenegger, A. 1992, ApJ, 395, 250
Goossens M., 1972, A&SS, 16, 386
Goossens M., Smeyers P. & Denis J., 1976, A&SS, 39, 257
Haskell, B., Samuelsson, L., Glampedakis, K. & Andersson, N. 2008, MNRAS, 385, 531
Jensen, E., 1955, ApJS, 2, 141
Fälthammar, C.-G., 2007, JASTP, 69, 1604
Ferraro, V.C.A., 1954, ApJ, 119, 407
Franco, L. M., Link B. & Epstein R. I. 2000, ApJ, 543, 987
Israel, G. L., Belloni, T., Stella, L., Rephaeli Y., Gruber, D. E., Casella, P., Dall’Osso, S., Rea, N., Persic, M. & Rothschild R. E. 2005, ApJ, 628, L53
Ioka, K. 2001, MNRAS, 327, 639
Lapwood, R. R. & Usami T., 1981, Free Oscillations of the Earth. Cambridge Univ. Press
Lay, T. & Wallace, T.C. 1995, Modern global seismology. Volume 58 of International geophysics series. Academic Press, San Diego
Ledoux, P. & Walraven, T. H. 1958, Handb. Der. Phys., 51, Ed. Flugge S., Springer, p.353
Ledoux, P. & Renson, P. 1966, ARA&A, 4, 293
Lee, U. 2008, MNRAS, 385, 2069
Mestel, L. 1956, MNRAS, 116, 324
Mestel, L. 1999, Stellar Magnetism. Clarendon Press, Oxford
Monaghan, J. J. 1965, MNRAS, 131, 105
Mereghetti, S. 2008, A&ARv, 15, 225
Plumpton, C. & Ferraro, V. C. A. 1955, ApJ, 121, 168
Roberts, P. H. 1955, ApJ, 112, 508
Roberts, P. H. 1981, AN, 302, 65
Sood N. K. & Trehan S. K., 1970, A&SS, 10, 393
Stein S. & Wyssesson M., 2003, An Introduction to Seismology, Earthquakes, and Earth structure. Blackwell, Oxford
Watts, A. L. & Strohmayer T. E. 2006, ApJ, 637, L117
Watts, A. L. & Strohmayer T. E. 2007, AdSpR, 40, 144
Woods, P. M. & Thompson, C. 2006, in Compact Stellar X-ray Sources, ed. Lewin, W. & van der Klis, M. (Cambridge: Cambridge University Press)
[^1]: The decay time of equilibrium magnetic field of the neutron stars is much longer than the time intervals between X-ray bursts and periods of their quiescent pulsed emission (e.g. Bhattacharya & van den Heuvel 1991, Chanmugham 1992, Goldreich and Reisenegger 1992), so that adopted approximation of an infinite electrical conductivity of the star matter is amply justified.
[^2]: It may be noteworthy that magnetic energy stored in the star volume, $W=(1/8\pi)\int B^2\,d{\cal V}$, with this nonhomogeneous (nh) internal magnetic field $W_{nh}=(69/252)B^2R^3\approx 0.24B^2R^3$ is somewhat larger than in the star with homogeneous (h) magnetic field $W_h=(1/6)B^2R^3\approx 0.17B^2R^3$.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
[Haocong Luo[^§^]{}[^$\dagger$^]{}]{}[Taha Shahroodi[^§^]{}]{}[Hasan Hassan[^§^]{}]{}[Minesh Patel[^§^]{}]{}\
[A. Giray Yağl[i]{}kç[i]{}[^§^]{}]{}[Jisung Park[^§^]{}]{}[Onur Mutlu[^§^]{}]{}\
\
[[^§^]{}*ETH Z[ü]{}rich* [^$\dagger$^]{}*ShanghaiTech University*]{}
bibliography:
- 'ref.bib'
title: |
[CLR-DRAM]{}: A Low-Cost DRAM Architecture\
Enabling Dynamic Capacity-Latency Trade-Off
---
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous ISCA 2020 reviewers for their feedback and the SAFARI group members for the stimulating intellectual environment they provide. We acknowledge the generous gifts provided by our industrial partners: Alibaba, Facebook, Google, Huawei, Intel, Microsoft, and VMware.
\[ no ligatures = [f]{}, outer kerning = [\*,\*]{} \] [ encoding = \* ]{} [ -40 ]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An equiangular tight frame (ETF) is a type of optimal packing of lines in a real or complex Hilbert space. In the complex case, the existence of an ETF of a given size remains an open problem in many cases. In this paper, we observe that many of the known constructions of ETFs are of one of two types. We further provide a new method for combining a given ETF of one of these two types with an appropriate group divisible design (GDD) in order to produce a larger ETF of the same type. By applying this method to known families of ETFs and GDDs, we obtain several new infinite families of ETFs. The real instances of these ETFs correspond to several new infinite families of strongly regular graphs. Our approach was inspired by a seminal paper of Davis and Jedwab which both unified and generalized McFarland and Spence difference sets. We provide combinatorial analogs of their algebraic results, unifying Steiner ETFs with hyperoval ETFs and Tremain ETFs.'
address:
- 'Department of Mathematics and Statistics, Air Force Institute of Technology, Wright-Patterson AFB, OH 45433'
- 'Department of Mathematics and Statistics, South Dakota State University, Brookings, SD 57007'
author:
- Matthew Fickus
- John Jasper
title: Equiangular tight frames from group divisible designs
---
equiangular tight frames ,group divisible designs 42C15
Introduction
============
Let $N\geq D$ be positive integers, let ${\mathbb{F}}$ be either ${\mathbb{R}}$ or ${\mathbb{C}}$, and let ${\langle{{\mathbf{x}}_1},{{\mathbf{x}}_2}\rangle}={\mathbf{x}}_1^*{\mathbf{x}}_2^{}$ be the dot product on ${\mathbb{F}}^D$. The *Welch bound* [@Welch74] states that any $N$ nonzero vectors ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ in ${\mathbb{F}}^D$ satisfy $$\label{eq.Welch bound}
\max_{n\neq n'}
\tfrac{{|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}}{{\|{{\boldsymbol{\varphi}}_n}\|}{\|{{\boldsymbol{\varphi}}_{n'}}\|}}
\geq{\bigl[{\tfrac{N-D}{D(N-1)}}\bigr]}^{\frac12}.$$ It is well-known [@StrohmerH03] that nonzero equal-norm vectors ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ in ${\mathbb{F}}^D$ achieve equality in if and only if they form an *equiangular tight frame* for ${\mathbb{F}}^D$, denoted an ${{\operatorname{ETF}}}(D,N)$, namely if there exists $A>0$ such that $A{\|{{\mathbf{x}}}\|}^2=\sum_{n=1}^{N}{|{{\langle{{\boldsymbol{\varphi}}_n},{{\mathbf{x}}}\rangle}}|}^2$ for all ${\mathbf{x}}\in{\mathbb{F}}^D$ (tightness), and the value of ${|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}$ is constant over all $n\neq n'$ (equiangularity). In particular, an ETF is a type of optimal packing in projective space, corresponding to a collection of lines whose minimum pairwise angle is as large as possible. ETFs arise in several applications, including waveform design for communications [@StrohmerH03], compressed sensing [@BajwaCM12; @BandeiraFMW13], quantum information theory [@Zauner99; @RenesBSC04] and algebraic coding theory [@JasperMF14].
In the general (possibly-complex) setting, the existence of an ${{\operatorname{ETF}}}(D,N)$ remains an open problem for many choices of $(D,N)$. See [@FickusM16] for a recent survey. Beyond orthonormal bases and regular simplices, all known infinite families of ETFs arise from combinatorial designs. Real ETFs in particular are equivalent to a class of *strongly regular graphs* (SRGs) [@vanLintS66; @Seidel76; @HolmesP04; @Waldron09], and such graphs have been actively studied for decades [@Brouwer07; @Brouwer17; @CorneilM91]. This equivalence has been partially generalized to the complex setting in various ways, including approaches that exploit properties of roots of unity [@BodmannPT09; @BodmannE10], abelian distance-regular covers of complete graphs (DRACKNs) [@CoutinkhoGSZ16], and association schemes [@IversonJM16]. Conference matrices, Hadamard matrices, Paley tournaments and quadratic residues are related, and lead to infinite families of ETFs whose *redundancy* $\frac ND$ is either nearly or exactly two [@StrohmerH03; @HolmesP04; @Renes07; @Strohmer08]. *Harmonic ETFs* and *Steiner ETFs* offer more flexibility in choosing $D$ and $N$. Harmonic ETFs are equivalent to *difference sets* in finite abelian groups [@Turyn65; @StrohmerH03; @XiaZG05; @DingF07], while Steiner ETFs arise from *balanced incomplete block designs* (BIBDs) [@GoethalsS70; @FickusMT12]. Recent generalizations of Steiner ETFs have led to new infinite families of ETFs arising from projective planes that contain hyperovals [@FickusMJ16] as well as from Steiner triple systems [@FickusJMP18], dubbed *hyperoval ETFs* and *Tremain ETFs*, respectively. Another new family arises by generalizing the SRG construction of [@Godsil92] to the complex setting, using generalized quadrangles to produce abelian DRACKNs [@FickusJMPW19].
Far less is known in terms of necessary conditions on the existence of complex ${{\operatorname{ETF}}}(D,N)$. The *Gerzon bound* implies that $N\leq\min{\{{D^2,(N-D)^2}\}}$ whenever a complex ${{\operatorname{ETF}}}(D,N)$ with $N>D>1$ exists [@LemmensS73; @HolmesP04; @Tropp05]. Beyond this, the only known nonexistence result in the complex case is that an ${{\operatorname{ETF}}}(3,8)$ does not exist [@Szollosi14], a result proven using computational techniques in algebraic geometry. In quantum information theory, ${{\operatorname{ETF}}}(D,D^2)$ are known as *symmetric informationally-complete positive operator-valued measures* (SIC-POVMs). It is famously conjectured that such Gerzon-bound-equality ETFs exist for any $D$ [@Zauner99; @Renes07; @FuchsHS17].
In this paper, we give a new method for constructing ETFs that yields several new infinite families of them. Our main result is Theorem \[thm.new ETF\], which shows how to combine a given initial ETF with a *group divisible design* (GDD) in order to produce another ETF. In that result, we require the initial ETF to be of one of the following types:
\[def.ETF types\] Given integers $D$ and $N$ with $1<D<N$, we say $(D,N)$ is *type $(K,L,S)$* if $$\label{eq.ETF param in terms of type param}
D
=\tfrac{S}{K}[S(K-1)+L]
=S^2-\tfrac{S(S-L)}{K},
\quad
N
=(S+L)[S(K-1)+L],$$ where $K$ and $S$ are integers and $L$ is either $1$ or $-1$. For a given $K$, we say $(D,N)$ is *$K$-positive* or *$K$-negative* when it is type $(K,1,S)$ or type $(K,-1,S)$ for some $S$, respectively. We simply say $(D,N)$ is *positive* or *negative* when it is $K$-positive or $K$-negative for some $K$, respectively. When we say that an ETF is one of these types, we mean its $(D,N)$ parameters are of that type.
It turns out that every known ${{\operatorname{ETF}}}(D,N)$ with $N>2D>2$ is either a harmonic ETF, a SIC-POVM, or is positive or negative. In particular, every Steiner ETF is positive, while every hyperoval ETF and Tremain ETF is negative. In this sense, the ideas and results of this paper are an attempt to unify and generalize several constructions that have been regarded as disparate. This is analogous to—and directly inspired by—a seminal paper of Davis and Jedwab [@DavisJ97], which unifies *McFarland* [@McFarland73] and *Spence* [@Spence77] difference sets under a single framework, and also generalizes them so as to produce difference sets whose corresponding harmonic ETFs have parameters $$\label{eq.Davis Jedwab parameters}
D=\tfrac13 2^{2J-1}(2^{2J+1}+1),
\quad
N=\tfrac13 2^{2J+2}(2^{2J}-1),$$ for some $J\geq 1$. It is quickly verified that such ETFs are type $(4,-1,S)$ where $S=\frac13(2^{2J+1}+1)$. As we shall see, combining our main result (Theorem \[thm.new ETF\]) with known ETFs and GDDs recovers the existence of ETFs with these parameters, and also provides several new infinite families, including:
\[thm.new neg ETS with K=4,5\] An ${{\operatorname{ETF}}}(D,N)$ of type $(K,-1,S)$ exists whenever:
1. $K=4$ and either $S\equiv 3\bmod 8$ or $S\equiv 7\bmod 60$;
2. $K=5$ and either $S\equiv 4\bmod 15$ or $S\equiv 5,309\bmod 380$ or $S\equiv 9\bmod 280$.
This result extends the $S$ for which an ETF of type $(4,-1,S)$ is known to exist from a geometric progression to a finite union of arithmetic progressions, with the smallest new ETF having $S=19$, namely $(D,N)=(266,1008)$, cf. [@FickusM16]. Meanwhile, the ETFs given by Theorem \[thm.new neg ETS with K=4,5\] in the $K=5$ case seem to be completely new except when $S=4,5,9$, with $(D,N)=(285,1350)$ being the smallest new example. Using similar techniques, we were also able to find new, explicit infinite families of $K$-negative ETFs for $K=6,7,10,12$. The description of these families is technical, and so is given in Theorem \[thm.new neg ETS with K>5\] as opposed to here. More generally, using asymptotic existence results for GDDs, we show that an infinite number of $K$-negative ETFs also exist whenever $K=Q+2$ where $Q$ is a prime power, $K=Q+1$ where $Q$ is an even prime power, or $K=8,20,30,42,56,342$.
In certain cases, the new ETFs constructed by these methods can be chosen to be real:
\[thm.new real ETFs\]
1. There are an infinite number of real Hadamard matrices of size $H\equiv 1\bmod 35$, and a real ETF of type $(5,-1,8H+1)$ exists for all such $H$.
2. There are an infinite number of real Hadamard matrices of size $H\equiv 1,8\bmod 21$, and a real ETF of type $(6,-1,2H+1)$ exists for all sufficiently large such $H$.
3. There are an infinite number of real Hadamard matrices of size $H\equiv 1,12\bmod 55$, and a real ETF of type $(10,-1,4H+1)$ exists for all sufficiently large such $H$.
4. There are an infinite number of real Hadamard matrices of size $H\equiv 1,277\bmod 345$, and a real ETF of type $(15,-1,4H+1)$ exists for all sufficiently large such $H$.
These correspond to four new infinite families of SRGs, with the smallest new example being a real ${{\operatorname{ETF}}}(66759,332640)$, which is obtained by letting $H=36$ in (a).
In the next section, we introduce known concepts from frame theory and combinatorial design that we need later on. In Section 3, we provide an alternative characterization of when an ETF is positive or negative (Theorem \[thm.parameter types\]), which we then use to help prove our main result (Theorem \[thm.new ETF\]). In the fourth section, we discuss how many known ETFs are either positive or negative, and then apply Theorem \[thm.new ETF\] to them along with known GDDs to obtain the new infinite families of negative ETFs described in Theorems \[thm.new neg ETS with K=4,5\] and \[thm.new neg ETS with K>5\]. We conclude in Section 5, using these facts as the basis for new conjectures on the existence of real and complex ETFs.
Preliminaries
=============
Equiangular tight frames
------------------------
For any positive integers $N$ and $D$, and any sequence ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ of vectors in ${\mathbb{F}}^D$, the corresponding *synthesis operator* is ${\boldsymbol{\Phi}}:{\mathbb{F}}^N\rightarrow{\mathbb{F}}^D$, ${\boldsymbol{\Phi}}{\mathbf{y}}:=\sum_{n=1}^{N}{\mathbf{y}}(n){\boldsymbol{\varphi}}_n$, namely the $D\times N$ matrix whose $n$th column is ${\boldsymbol{\varphi}}_n$. Its adjoint (conjugate transpose) is the *analysis operator* ${\boldsymbol{\Phi}}^*:{\mathbb{F}}^D\rightarrow{\mathbb{F}}^N$, which has $({\boldsymbol{\Phi}}^*{\mathbf{x}})(n)={\langle{{\boldsymbol{\varphi}}_n},{{\mathbf{x}}}\rangle}$ for all $n=1,\dotsc,N$. That is, ${\boldsymbol{\Phi}}^*$ is the $D\times N$ matrix whose $n$th row is ${\boldsymbol{\varphi}}_n^*$. Composing these two operators gives the $N\times N$ *Gram matrix* ${\boldsymbol{\Phi}}^*{\boldsymbol{\Phi}}$ whose $(n,n')$th entry is $({\boldsymbol{\Phi}}^*{\boldsymbol{\Phi}})(n,n')={\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}$, as well as the $D\times D$ *frame operator* ${\boldsymbol{\Phi}}{\boldsymbol{\Phi}}^*=\sum_{n=1}^{N}{\boldsymbol{\varphi}}_n^{}{\boldsymbol{\varphi}}_n^*$.
We say ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ is a *tight frame* for ${\mathbb{F}}^D$ if there exists $A>0$ such that ${\boldsymbol{\Phi}}{\boldsymbol{\Phi}}^*=A{\mathbf{I}}$, namely if the rows of ${\boldsymbol{\Phi}}$ are orthogonal and have an equal nontrivial norm. We say ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ is *equal norm* if there exists some $C$ such that ${\|{{\boldsymbol{\varphi}}_n}\|}^2=C$ for all $n$. The parameters of an equal norm tight frame are related according to $DA={\operatorname{Tr}}(A{\mathbf{I}})={\operatorname{Tr}}({\boldsymbol{\Phi}}{\boldsymbol{\Phi}}^*)={\operatorname{Tr}}({\boldsymbol{\Phi}}^*{\boldsymbol{\Phi}})=\sum_{n=1}^{N}{\|{{\boldsymbol{\varphi}}_n}\|}^2=NC$. We say ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ is *equiangular* if it is equal norm and the value of ${|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}$ is constant over all $n\neq n'$.
For any equal norm vectors ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ in ${\mathbb{F}}^D$, a direct calculation reveals $$0
\leq{\operatorname{Tr}}[(\tfrac1C{\boldsymbol{\Phi}}{\boldsymbol{\Phi}}^*-\tfrac{N}{D}{\mathbf{I}})^2]
=\sum_{n=1}^{N}\sum_{\substack{n'=1\\n'\neq n}}^{N}
\tfrac{{|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}^2}{C^2}
-\tfrac{N(N-D)}{D}
\leq N(N-1)\max_{n\neq n'}
\tfrac{{|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}^2}{C^2}
-\tfrac{N(N-D)}{D}.$$ Rearranging this inequality gives the Welch bound . Moreover, we see that achieving equality in is equivalent to having equality above throughout, which happens precisely when ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ is a tight frame for ${\mathbb{F}}^D$ that is also equiangular, namely when it is an ETF for ${\mathbb{F}}^D$.
If $N>D$ and ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ is a tight frame for ${\mathbb{F}}^D$ then completing the $D$ rows of ${\boldsymbol{\Phi}}$ to an equal-norm orthogonal basis for ${\mathbb{F}}^N$ is equivalent to taking a $(N-D)\times N$ matrix ${\boldsymbol{\Psi}}$ such that ${\boldsymbol{\Psi}}^*{\boldsymbol{\Psi}}=A{\mathbf{I}}$, ${\boldsymbol{\Phi}}{\boldsymbol{\Psi}}^*={\boldsymbol{0}}$ and ${\boldsymbol{\Phi}}^*{\boldsymbol{\Phi}}+{\boldsymbol{\Psi}}^*{\boldsymbol{\Psi}}=A{\mathbf{I}}$. The sequence ${\{{{\boldsymbol{\psi}}_n}\}}_{n=1}^{N}$ of columns of any such ${\boldsymbol{\Psi}}$ is called a *Naimark complement* of ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$. Since ${\boldsymbol{\Psi}}{\boldsymbol{\Psi}}^*=A{\mathbf{I}}$ and ${\boldsymbol{\Psi}}^*{\boldsymbol{\Psi}}=A{\mathbf{I}}-{\boldsymbol{\Phi}}^*{\boldsymbol{\Phi}}$, any Naimark complement of an ${{\operatorname{ETF}}}(D,N)$ is an ${{\operatorname{ETF}}}(N-D,N)$. Since any nontrivial scalar multiple of an ETF is another ETF, we will often assume without loss of generality that a given ${{\operatorname{ETF}}}(D,N)$ and its Naimark complements satisfy $$\label{eq.ETF scaling}
A=N{\bigl[{\tfrac{N-1}{D(N-D)}}\bigr]}^{\frac12},
\quad
{\|{{\boldsymbol{\varphi}}_n}\|}^2
={\bigl[{\tfrac{D(N-1)}{N-D}}\bigr]}^{\frac12},
\quad
{\|{{\boldsymbol{\psi}}_n}\|}^2
={\bigl[{\tfrac{(N-D)(N-1)}D}\bigr]}^{\frac12},
\quad
\forall n=1,\dotsc,N,$$ which equates to having ${|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}=1={|{{\langle{{\boldsymbol{\psi}}_n},{{\boldsymbol{\psi}}_{n'}}\rangle}}|}$ for all $n\neq n'$. For positive and negative ETFs in particular (Definition \[def.ETF types\]), we shall see that all of these quantities happen to be integers.
Any ${{\operatorname{ETF}}}(D,N)$ with $N=D+1$ is known as a *regular simplex*, and such ETFs are Naimark complements of ETFs for ${\mathbb{F}}^1$, namely sequences of scalars that have the same nontrivial modulus. In particular, a sequence of vectors ${\{{{\mathbf{f}}_n}\}}_{n=1}^{N}$ in ${\mathbb{F}}^{N-1}$ is a Naimark complement of the all-ones sequence in ${\mathbb{F}}^1$ if and only if $$\label{eq.regular simplex}
{\mathbf{F}}{\mathbf{F}}^*=N{\mathbf{I}},
\quad
\sum_{n=1}^{N}{\mathbf{f}}_n={\mathbf{F}}{\boldsymbol{1}}={\boldsymbol{0}},
\quad
{\mathbf{F}}^*{\mathbf{F}}=N{\mathbf{I}}-{\mathbf{J}},$$ where ${\boldsymbol{1}}$ and ${\mathbf{J}}$ denote an all-ones column vector and matrix, respectively. Equivalently, the vectors ${\{{1\oplus{\mathbf{f}}_n}\}}_{n=1}^{N}$ in ${\mathbb{F}}^N$ are equal norm and orthogonal. In particular, for any $N>1$, we can always take ${\{{1\oplus{\mathbf{f}}_n}\}}_{n=1}^{N}$ to be the columns of a possibly-complex Hadamard matrix of size $N$. In this case, ${\mathbf{F}}$ satisfies and is also *flat*, meaning every one of its entries has modulus one. As detailed below, flat regular simplices can be used to construct several families of ETFs, including Steiner ETFs as well as those we introduce in Theorem \[thm.new ETF\].
Harmonic ETFs are the best-known class of ETFs [@Strohmer08; @XiaZG05; @DingF07]. A harmonic ${{\operatorname{ETF}}}(D,N)$ is obtained by restricting the characters of an abelian group ${\mathcal{G}}$ of order $N$ to a *difference set* of cardinality $D$, namely a $D$-element subset ${\mathcal{D}}$ of ${\mathcal{G}}$ with the property that the cardinality of ${\{{(d,d')\in{\mathcal{D}}\times{\mathcal{D}}: g=d-d'}\}}$ is constant over all $g\in{\mathcal{G}}$, $g\neq 0$. The set complement ${\mathcal{G}}\backslash{\mathcal{D}}$ of any difference set in ${\mathcal{G}}$ is another difference set, and the two corresponding harmonic ETFs are Naimark complements. In particular, for any abelian group ${\mathcal{G}}$ of order $N$, the harmonic ETF arising from ${\mathcal{G}}\backslash{\{{0}\}}$ is a flat regular simplex that satisfies .
Group divisible designs
-----------------------
For a given integer $K\geq 2$, a $K$-GDD is a set ${\mathcal{V}}$ of $V>K$ vertices, along with collections ${\mathcal{G}}$ and ${\mathcal{B}}$ of subsets of ${\mathcal{V}}$, called *groups* and *blocks*, respectively, with the property that the groups partition ${\mathcal{V}}$, every block has cardinality $K$, and any two vertices are either contained in a common group or a common block, but not both. A $K$-GDD is *uniform* if its groups all have the same cardinality $M$, denoted in *exponential notation* as a “$K$-GDD of type $M^U$" where $V=UM$.
Letting $B$ be the number of blocks, a ${\{{0,1}\}}$-valued $B\times UM$ incidence matrix ${\mathbf{X}}$ of a $K$-GDD of type $M^U$ has the property that each row of ${\mathbf{X}}$ contains exactly $K$ ones. Moreover, for any $v=1,\dotsc,V=UM$, the $v$th column of ${\mathbf{X}}$ is orthogonal to $M-1$ other columns of ${\mathbf{X}}$, and has a dot product of $1$ with each of the remaining $(U-1)M$ columns. This implies that the *replication number* $R_v$ of blocks that contain the $v$th vertex satisfies $$(U-1)M
=\sum_{\substack{v'=1\\v'\neq v}}^V({\mathbf{X}}^*{\mathbf{X}})(v,v')
=\sum_{b=1}^{B}{\mathbf{X}}(b,v)\sum_{\substack{v'=1\\v'\neq v}}^V{\mathbf{X}}(b,v')
=\sum_{b=1}^{B}\left\{\begin{array}{cl}K-1,&{\mathbf{X}}(b,v)=1\\0,&{\mathbf{X}}(b,v)=0\end{array}\right\}
=R_v(K-1).$$ As such, this number $R_v=R$ is independent of $v$. At this point, summing all entries of ${\mathbf{X}}$ gives $BK=VR$ and so $B$ is also uniquely determined by $K$, $M$ and $U$. Because of this, the existence of a $K$-GDD of type $M^U$ is equivalent to that of a ${\{{0,1}\}}$-valued $B\times UM$ matrix ${\mathbf{X}}$ with $$\label{eq.GDD incidence matrix}
R=\tfrac{M(U-1)}{K-1},
\quad
B=\tfrac{MUR}{K}=\tfrac{M^2U(U-1)}{K(K-1)},
\quad
{\mathbf{X}}{\boldsymbol{1}}=K{\boldsymbol{1}},
\quad
{\mathbf{X}}^*{\mathbf{X}}=R\,{\mathbf{I}}+({\mathbf{J}}_U-{\mathbf{I}}_U)\otimes{\mathbf{J}}_M.$$ In the special case where $M=1$, a $K$-GDD of type $1^U$ is called a ${{\operatorname{BIBD}}}(U,K,1)$. In the special case where $U=K$, a $K$-GDD of type $M^K$ is called a *transversal design* ${{\operatorname{TD}}}(K,M)$, which is equivalent to a collection of $K-2$ *mutually orthogonal Latin squares* (MOLS) of size $M$.
In order for a $K$-GDD of type $M^U$ to exist, the expressions for $R$ and $B$ given in are necessarily integers. Beyond this, we necessarily have $U\geq K$ since we can partition any given block into its intersections with the groups, and the cardinality of these intersections is at most one. Altogether, the parameters of a $K$-GDD of type $M^U$ necessarily satisfy $$\label{eq.GDD necessary conditions}
U\geq K,
\quad
\tfrac{M(U-1)}{K-1}\in{\mathbb{Z}},
\quad
\tfrac{M^2U(U-1)}{K(K-1)}\in{\mathbb{Z}}.$$ Though these necessary conditions are not sufficient [@Ge07], they are asymptotically sufficient in two distinct ways: for any fixed $K\geq 2$ and $M\geq 1$, there exists $U_0=U_0(K,M)$ such that a $K$-GDD of type $M^U$ exists for all $U\geq U_0$ such that is satisfied [@Chang76; @LamkenW00]; for any fixed $U\geq K\geq 2$, there exists $M_0=M_0(K,U)$ such that a $K$-GDD of type $M^U$ exists for all $M\geq M_0$ such that is satisfied [@Mohacsy11]. In the $M=1$ and $U=K$ cases, these facts reduce to more classical asymptotic existence results for BIBDs and MOLS, respectively.
Many specific examples of GDDs are formed by combining smaller designs in clever ways. We in particular will make use of the following result, which is a special case of Wilson’s approach [@Wilson72]:
\[lem.Wilson\] If a $K$-GDD of type $M^U$ and a $U$-GDD of type $N^V$ exist, then a $K$-GDD of type $(MN)^V$ exists.
Let ${\mathbf{X}}$ and ${\mathbf{Y}}$ be incidence matrices of the form for the given $K$-GDD of type $M^U$ and $U$-GDD of type $N^V$, respectively. In particular, taking $R$ and $B$ as in , we can write ${\mathbf{X}}=\left[\begin{array}{ccc}{\mathbf{X}}_1&\cdots&{\mathbf{X}}_U\end{array}\right]$ where each ${\mathbf{X}}_u$ is a $B\times M$ matrix with ${\mathbf{X}}_u^*{\mathbf{X}}_u^{}=R{\mathbf{I}}$, and ${\mathbf{X}}_u^*{\mathbf{X}}_{u'}^{}={\mathbf{J}}$ for any $u\neq u'$. We now construct the incidence matrix ${\mathbf{Z}}$ of a $K$-GDD of type $(MN)^V$ in the following manner: in each row of ${\mathbf{Y}}$, replace each of the $U$ nonzero entries with a distinct matrix ${\mathbf{X}}_u$, and replace each of the zero entries with a $B\times M$ matrix of zeros.
This result generalizes MacNeish’s classical method for combining MOLS [@MacNeish22]: if a ${{\operatorname{TD}}}(K,M)$ and a ${{\operatorname{TD}}}(K,N)$ exist, then applying Lemma \[lem.Wilson\] to them produces a ${{\operatorname{TD}}}(K,MN)$. We will also use one GDD to “fill the holes" of another:
\[lem.filling holes\] If $K$-GDDs of type $M^U$ and $(MU)^V$ exist, then a $K$-GDD of type $M^{UV}$ exists.
Letting ${\mathbf{X}}$ and ${\mathbf{Y}}$ be incidence matrices of the form for the given $K$-GDDs of type $M^U$ and $(MU)^V$, respectively, it is straightforward to verify that $${\mathbf{Z}}=\left[\begin{array}{c}{\mathbf{I}}_V\otimes{\mathbf{X}}\\{\mathbf{Y}}\end{array}\right]$$ is the incidence matrix of a $K$-GDD of type $M^{UV}$.
Previously known constructions of ETFs involving BIBDs and MOLS
---------------------------------------------------------------
In the next section, we introduce a method for constructing ETFs that uses GDDs. This method makes use of a concept from [@FickusJMP18], which we now generalize from BIBDs to GDDs:
\[def.embeddings\] Take a $K$-GDD of type $M^U$ where $M\geq1$ and $U\geq K\geq 2$, and define $R$, $B$ and an incidence matrix ${\mathbf{X}}$ according to . Without loss of generality, write the columns of ${\mathbf{X}}$ as where, for each $u$, the vectors ${\{{{\mathbf{x}}_{u,m}}\}}_{m=1}^{M}$ have disjoint support. Then, for any $u$ and $m$, a corresponding *embedding operator* ${\mathbf{E}}_{u,m}$ is any ${\{{0,1}\}}$-valued $B\times R$ matrix whose columns are standard basis elements that sum to ${\mathbf{x}}_{u,m}$.
In the special case where $M=1$, this concept leads to an elegant formulation of Steiner ETFs [@FickusJMP18]: letting be the embedding operators of a ${{\operatorname{BIBD}}}(V,K,1)$, and letting ${\{{1\oplus{\mathbf{f}}_i}\}}_{i=0}^{R}$ be the columns of a possibly-complex Hadamard matrix of size $R+1=\frac{V-1}{K-1}+1$, the $V(R+1)$ vectors form an ETF for ${\mathbb{F}}^B$. In [@FickusJMP18], this fact is proven using several properties of embedding operators. We now show those properties generalize to the GDD setting; later on, we use these facts to prove our main result:
\[lem.embed\] If ${\{{{\mathbf{E}}_{u,m}}\}}_{u=1,}^{U},\,_{m=1}^{M}$ are the embedding operators arising from a $K$-GDD of type $M^U$, $${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{}
=\left\{\begin{array}{cl}
{\mathbf{I}},&\ u=u', m=m',\\
{\boldsymbol{0}},&\ u=u', m\neq m',\\
{\boldsymbol{\delta}}_{r}^{}{\boldsymbol{\delta}}_{\smash{r'}}^*,&\ u\neq u'.
\end{array}\right.$$ Here, for any $u\neq u'$ and $m,m'$, ${\boldsymbol{\delta}}_r$ and ${\boldsymbol{\delta}}_{\smash{r'}}$ are standard basis elements in ${\mathbb{F}}^R$ whose indices $r,r'$ depend on $u,u',m,m'$.
Each ${\mathbf{E}}_{u,m}$ is a matrix whose columns are standard basis elements that sum to ${\mathbf{x}}_{u,m}$, and so is an isometry, that is, ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u,m}^{}={\mathbf{I}}$. Moreover, for any $u$, $u'$, $m$, $m'$, ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{}$ is a matrix whose entries are nonnegative integers that sum to: $$\sum_{r=1}^{R}\sum_{r'=1}^{R}({\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{})(r,r')
={\boldsymbol{1}}^*{\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{}{\boldsymbol{1}}^{}
={\langle{{\mathbf{x}}_{u,m}},{{\mathbf{x}}_{u',m'}}\rangle}
=\left\{\begin{array}{cl}
R,&\ u=u', m=m',\\
0,&\ u=u', m\neq m',\\
1,&\ u'\neq u.
\end{array}\right.$$ When $u=u'$ and $m\neq m'$, this implies ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{}={\boldsymbol{0}}$. If instead $u\neq u'$ then this implies that ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{}$ has a single nonzero entry, and that this entry has value $1$. This means there exists some $r,r'=1,\dotsc,R$, $r=r(u,m,u',m')$, $r'=r'(u,m,u',m')$ such that ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{\smash{u,m'}}^{}={\boldsymbol{\delta}}_{r}^{}{\boldsymbol{\delta}}_{\smash{r'}}^*$.
Other Steiner-like constructions of ETFs include hyperoval ETFs [@FickusMJ16] and Tremain ETFs [@FickusJMPW19]. Beyond Steiner and Steiner-like techniques, there are at least two other methods for constructing ETFs that make direct use of the incidence matrix of some kind of GDD. One method leads to the *phased BIBD ETFs* of [@FickusJMPW19]: if ${\mathbf{X}}$ is the $B\times V$ incidence matrix of a ${{\operatorname{BIBD}}}(V,K,1)$, and ${\boldsymbol{\Phi}}$ is any matrix obtained by replacing each $1$-valued entry of ${\mathbf{X}}$ with any unimodular scalar, then the columns of ${\boldsymbol{\Phi}}$ are immediately equiangular, and the challenge is to design them so that they form a tight frame for their span. Another method constructs ETFs with $(D,N)=(\tfrac12M(M\pm1),M^2)$ from MOLS. To elaborate, a ${{\operatorname{TD}}}(K,M)$ is a $K$-GDD of type $M^K$, meaning by that it has an $M^2\times KM$ incidence matrix ${\mathbf{X}}$ that satisfies $$\label{eq.incidence matrix of TD 1}
{\mathbf{X}}{\boldsymbol{1}}=K{\boldsymbol{1}},
\quad
{\mathbf{X}}^*{\mathbf{X}}=M{\mathbf{I}}+({\mathbf{J}}_K-{\mathbf{I}}_K)\otimes{\mathbf{J}}_M.$$ Here, the columns of ${\mathbf{X}}$ have support $M$, and are arranged as $K$ groups of $M$ columns apiece, where columns in a common group have disjoint support. Together, these facts imply, in turn, that $$\label{eq.incidence matrix of TD 2}
{\mathbf{X}}({\mathbf{I}}_K\otimes{\boldsymbol{1}}_M)={\boldsymbol{1}}_{M^2}^{}{\boldsymbol{1}}_K^*,
\quad
({\mathbf{X}}{\mathbf{X}}^*)^2=M{\mathbf{X}}{\mathbf{X}}^*+K(K-1){\mathbf{J}}.$$ At this point, the traditional approach is to let ${\mathbf{A}}={\mathbf{X}}{\mathbf{X}}^*-K{\mathbf{I}}$ be the adjacency matrix of the TD’s *block graph*, and use and to show that this graph is strongly regular with parameters $(M^2,K(M-1),M+K(K-3),K(K-1))$. In the $M=2K$ case, applying Theorem 4.4 of [@FickusJMPW18] to this graph then produces a real ETF with $(D,N)=(\frac12M(M-1),M^2)$ whose vectors sum to zero, while applying this same result in the $M=2(K-1)$ case produces a real ETF with $(D,N)=(\frac12M(M+1),M^2)$ whose synthesis operator’s row space contains the all-ones vector.
That said, a careful read of the literature reveals that this construction can be made more explicit, and that doing so has repercussions for coding theory. To elaborate, in [@BrackenMW06], MOLS are used to produce quasi-symmetric designs (QSDs) which, via the techniques of [@McGuire97], yield self-complementary binary codes that achieve equality in the Grey-Rankin bound. In [@JasperMF14], such codes are shown to be equivalent to flat real ETFs. A method for directly converting the incidence matrices of certain QSDs into synthesis operators of ETFs was also recently introduced [@FickusJMP19]. Distilling these ideas leads to the following streamlined construction: let ${\mathbf{X}}$ be the incidence matrix of a ${{\operatorname{TD}}}(K,M)$, let ${\{{1\oplus{\mathbf{f}}_m}\}}_{m=1}^{M}$ be the columns of a possibly-complex Hadamard matrix of size $M$, let ${\mathbf{F}}$ be the $(M-1)\times M$ synthesis operator of ${\{{{\mathbf{f}}_m}\}}_{m=1}^{M}$, and consider the $K(M-1)\times M^2$ matrix $$\label{eq.first flat ETF from MOLS}
{\boldsymbol{\Phi}}=({\mathbf{I}}_K\otimes{\mathbf{F}}){\mathbf{X}}^*.$$ Using , , and along with the fact that ${\mathbf{F}}$ is flat, it is straightforward to show that ${\boldsymbol{\Phi}}$ is flat and satisfies ${\boldsymbol{\Phi}}{\boldsymbol{\Phi}}^*
=M^2{\mathbf{I}}$ and ${\boldsymbol{\Phi}}^*{\boldsymbol{\Phi}}=M{\mathbf{X}}{\mathbf{X}}^*-K{\mathbf{J}}$. As such, the columns of ${\boldsymbol{\Phi}}$ form a flat *two-distance tight frame* (TDTF) for ${\mathbb{F}}^{K(M-1)}$ [@BargGOY15]. Moreover, this TDTF is an ETF when $M=2K$. In particular, if there exists a ${{\operatorname{TD}}}(K,2K)$ and a real Hadamard matrix of size $2K$, then there exists a flat real ${{\operatorname{ETF}}}(K(2K-1),4K^2)$. Using the equivalence between flat real ETFs and Grey-Rankin-bound-equality codes given in [@JasperMF14], or alternatively the equivalence between such ETFs and certain QSDs given in [@FickusJMP19], this recovers Theorem 1 of [@BrackenMW06]. In the $K=6$ case, that result gives the only known proof to date of the existence of a flat real ${{\operatorname{ETF}}}(66,144)$.
For any TD and corresponding flat regular simplex, it is quickly verified that the corresponding TDTF is *centered* [@FickusJMPW18] in the sense that ${\boldsymbol{\Phi}}{\boldsymbol{1}}={\boldsymbol{0}}$, namely that the all-ones vector is orthogonal to the row space of ${\boldsymbol{\Phi}}$. This fact leads to an analogous reinterpretation of the second main result of [@BrackenMW06]: in lieu of , we instead consider the $[K(M-1)+1]\times M^2$ flat matrix $$\label{eq.second flat ETF from MOLS}
{\boldsymbol{\Psi}}=\left[\begin{array}{l}{\boldsymbol{1}}^*\\{\boldsymbol{\Phi}}\end{array}\right]
=\left[\begin{array}{c}{\boldsymbol{1}}^*\\({\mathbf{I}}_K\otimes{\mathbf{F}}){\mathbf{X}}^*\end{array}\right].$$ Here, the properties of ${\boldsymbol{\Phi}}$ immediately imply ${\boldsymbol{\Psi}}{\boldsymbol{\Psi}}^*=M^2{\mathbf{I}}$ and ${\boldsymbol{\Psi}}^*{\boldsymbol{\Psi}}=M{\mathbf{X}}{\mathbf{X}}^*-(K-1){\mathbf{J}}$, meaning the columns of ${\boldsymbol{\Psi}}$ form a flat TDTF. However, unlike , the columns of are equiangular precisely when $M=2(K-1)$. Replacing $K$ with $K+1$, this implies in particular that if there exists a ${{\operatorname{TD}}}(K+1,2K)$ and a real Hadamard matrix of size $2K$, then there exists a flat real ${{\operatorname{ETF}}}(K(2K+1),4K^2)$. This recovers Theorem 2 of [@BrackenMW06] via the equivalences of [@JasperMF14; @FickusJMP19], and gives the only known proof of the existence of a flat real ${{\operatorname{ETF}}}(78,144)$.
Simply put, if certain TDs exist, then certain ETFs exist. In the next section, we introduce a new method of constructing ETFs from TDs.
Constructing equiangular tight frames with group divisible designs
==================================================================
In [@DavisJ97], Davis and Jedwab unify McFarland [@McFarland73] and Spence [@Spence77] difference sets under a single framework, and also generalize them so as to produce difference sets with parameters . McFarland’s construction relies on nice algebro-combinatorial properties of the set of all hyperplanes in a finite-dimensional vector space over a finite field. Davis and Jedwab exploit these properties to form various types of *building sets*, which in some cases lead to difference sets.
In [@JasperMF14], it is shown that every harmonic ETF arising from a McFarland ETF is unitarily-equivalent to a Steiner ETF arising from an affine geometry. When we applied a similar analysis to the building sets of [@DavisJ97], we discovered that they have an underlying TD-like incidence structure. (We do not provide this analysis here since it is nontrivial and does not help us prove our results in their full generality.) This eventually led us to the ETF construction technique of Theorem \[thm.new ETF\] below. In short, our approach here is directly inspired by that of [@DavisJ97], though this is not apparent from our proof techniques. In particular, the fact that the $L$ parameter in Definition \[def.ETF types\] is either $1$ or $-1$ is a generalization of Davis and Jedwab’s notion of *extended building sets* with “$+$" and “$-$" parameters, respectively. To facilitate our arguments later on, we now consider these types of parameters in greater detail:
\[thm.parameter types\] If $1<D<N$ and $(D,N)$ is type $(K,L,S)$, see Definition \[def.ETF types\], then $$\label{eq.type param in terms of ETF param}
S={\bigl[{\tfrac{D(N-1)}{N-D}}\bigr]}^{\frac12},
\quad
K=\tfrac{NS}{D(S+L)},$$ where $S\geq 2$. Conversely, given $(D,N)$ such that $1<D<N$, and letting $L$ be either $1$ or $-1$, if the above expressions for $S$ and $K$ are integers then $(D,N)$ is type $(K,L,S)$.
Moreover, in the case that the equivalent conditions above hold, scaling an ETF ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ for ${\mathbb{F}}^D$ so that ${|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}=1$ for all $n\neq n'$ gives that it and its Naimark complements ${\{{{\boldsymbol{\psi}}_n}\}}_{n=1}^{N}$ have tight frame constant $A=K(S+L)$ and $$\label{eq.norms of pos or neg ETFs}
{\|{{\boldsymbol{\varphi}}_n}\|}^2
={\bigl[{\tfrac{D(N-1)}{N-D}}\bigr]}^{\frac12}
=S,
\quad
{\|{{\boldsymbol{\psi}}_n}\|}^2
={\bigl[{\tfrac{(N-D)(N-1)}{D}}\bigr]}^{\frac12}
=S(K-1)+KL,
\quad
\forall n=1,\dotsc,N.$$
Whenever $(D,N)$ is type $(K,L,S)$ we have that $L$ is either $1$ or $-1$ by assumption, at which point the fact that $L^2=1$ coupled with gives $$\begin{aligned}
\nonumber
N-1
&=(S+L)[S(K-1)+L]-1
=S^2(K-1)+SKL
=S[S(K-1)+KL],\\
\tfrac{N}{D}-1
&=\tfrac{K(S+L)[S(K-1)+L]}{S[S(K-1)+L]}-1
=\tfrac{K(S+L)}{S}-1
=\tfrac1{S}[S(K-1)+KL].\end{aligned}$$ Multiplying and dividing these expressions immediately implies that $$\label{eq.N in terms of D and K 0}
{\bigl[{\tfrac{D(N-1)}{N-D}}\bigr]}^{\frac12}=S,
\quad
{\bigl[{\tfrac{(N-D)(N-1)}{D}}\bigr]}^{\frac12}=S(K-1)+KL.$$ Here, $S$ is an integer by assumption, and is clearly positive. Moreover, if $S=1$ then implies $D=1$. Since $D>1$ by assumption, we thus have $S\geq 2$. Continuing, further implies $$\tfrac{NS}{D(S+L)}
=\tfrac{K(S+L)[S(K-1)+L]}{S[S(K-1)+L]}\tfrac{S}{S+L}
=K.$$
Conversely, now assume that $S$ and $K$ are defined by , where $L$ is either $1$ or $-1$, and that $S$ and $K$ are integers. As before, the fact that $D>1$ implies that $S\geq 2$ and so $K>0$. We solve for $N$ in terms of $D$, $K$, and $L$. Here, gives $\frac{N}{DK}=\frac{S+L}{S}=1+\frac{L}{S}$. Since $L^2=1$, this implies $$\label{eq.N in terms of D and K 1}
{\bigl[{\tfrac{N-D}{D(N-1)}}\bigr]}^{\frac12}
=\tfrac1{S}
=L{\bigl({\tfrac{N}{DK}-1}\bigr)}.$$ Squaring this equation and multiplying the result by $N-1$ gives $$\tfrac{N}{D}-1
=\tfrac{N-D}{D}
=(N-1){\bigl({\tfrac{N}{DK}-1}\bigr)}^2
=N{\bigl({\tfrac{N}{DK}-1}\bigr)}^2-\tfrac{N}{DK}{\bigl({\tfrac{N}{DK}-2}\bigr)}-1.$$ Adding $1$ to this equation and multiplying by $\tfrac{(DK)^2}{N}$ then leads to a quadratic in $N$: $$DK^2
=(N-DK)^2-(N-2DK)
=N^2-(2DK+1)N+DK(DK+2).$$ Applying the quadratic formula then gives $$\label{eq.N in terms of D and K 2}
N=DK+\tfrac12{\bigl\{{1\pm{\bigl[{4DK(K-1)+1}\bigr]}^{\frac12}}\bigr\}}.$$ As such, becomes $\tfrac1{S}
=L{\bigl({\tfrac{N}{DK}-1}\bigr)}
=\tfrac{L}{2DK}{\bigl\{{1\pm{\bigl[{4DK(K-1)+1}\bigr]}^{\frac12}}\bigr\}}$, implying “$+$" and “$-$" here correspond to $L=1$ and $L=-1$ respectively, that is, $$\label{eq.N in terms of D and K 3}
\tfrac1{S}
=\tfrac{L}{2DK}{\bigl\{{1+L{\bigl[{4DK(K-1)+1}\bigr]}^{\frac12}}\bigr\}}
=\tfrac1{2DK}{\bigl\{{L+{\bigl[{4DK(K-1)+1}\bigr]}^{\frac12}}\bigr\}}.$$ Moreover, since $N$ is an integer, implies that $4DK(K-1)+1$ is the square of an odd integer, that is, that $4DK(K-1)+1=(2J-1)^2=4J(J-1)+1$ or equivalently that $DK(K-1)=J(J-1)$ for some positive integer $J$. Writing $D=\frac{J(J-1)}{K(K-1)}$, and then become $$\begin{aligned}
\label{eq.N in terms of D and K 4}
N
&=DK+\tfrac12{\bigl\{{1+L{\bigl[{4DK(K-1)+1}\bigr]}^{\frac12}}\bigr\}}
=DK+\tfrac12[1+L(2J-1)],
\\
\nonumber
\tfrac1{S}
&=\tfrac1{2DK}{\bigl\{{L+{\bigl[{4DK(K-1)+1}\bigr]}^{\frac12}}\bigr\}}
=\tfrac{K-1}{2J(J-1)}[L+(2J-1)]
=\left\{\begin{array}{cl}
\tfrac{K-1}{J-1},&L=1\smallskip\\
\tfrac{K-1}{J},&L=-1
\end{array}\right\}
=\tfrac{2(K-1)}{2J-L-1}.\end{aligned}$$ That is, $J=S(K-1)+\frac12(L+1)$. Substituting this into $D=\frac{J(J-1)}{K(K-1)}$ and and again using the fact that $L^2=1$ then gives the expressions for $D$ and $N$ given in Definition \[def.ETF types\]: $$\begin{aligned}
D
&=\tfrac{[S(K-1)+\frac12(L+1)][S(K-1)+\frac12(L-1)]}{K(K-1))}
=\tfrac{S^2(K-1)^2+S(K-1)L}{K(K-1))}
=\tfrac{S}{K}[S(K-1)+L],\\
N
&=S[S(K-1)+L]+\tfrac12\{1+L[2S(K-1)+L]\}
=(S+L)[S(K-1)+L].\end{aligned}$$
Finally, in the case where $(D,N)$ is type $(K,L,S)$, if ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ is an ETF for ${\mathbb{F}}^D$, and is without loss of generality scaled so that ${|{{\langle{{\boldsymbol{\varphi}}_n},{{\boldsymbol{\varphi}}_{n'}}\rangle}}|}=1$ for all $n\neq n'$, then and immediately imply that ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ and any one of its Naimark complements ${\{{{\boldsymbol{\psi}}_n}\}}_{n=1}^{N}$ satisfy , and that both are tight frames with tight frame constant $A=\frac{NS}{D}=K(S+L)$.
Theorem \[thm.parameter types\] implies that the $(D,N)$ parameters of an ETF with $N>D>1$ are type $(K,L,S)$ with $K=1$ if and only if that ETF is a regular simplex, and moreover that this only occurs when $L=1$ and $S=D$. Indeed, for any $D>1$, the pair $(D,N)=(D,D+1)$ satisfies when $(K,L,S)=(1,1,D)$. Conversely, in light of , an ETF of type $(1,L,S)$ has a Naimark complement ${\{{{\boldsymbol{\psi}}_n}\}}_{n=1}^{N}$ with the property that ${\|{{\boldsymbol{\psi}}_n}\|}^2=L=1$ and ${|{{\langle{{\boldsymbol{\psi}}_n},{{\boldsymbol{\psi}}_{n'}}\rangle}}|}=1$ for all $n\neq n'$, namely a Naimark complement that is an ${{\operatorname{ETF}}}(1,N)$.
We also emphasize that it is sometimes possible for the parameters of a single ETF to be simultaneously positive and negative for different choices of $K$. In particular, if $D>1$ and $(D,D+1)$ is type $(K,-1,S)$, then gives $S=D$ and $K=\frac{NS}{D(S+L)}=\frac{(D+1)D}{D(D-1)}=\frac{D+1}{D-1}=1+\frac{2}{D-1}$. Since $K$ is an integer, this implies either $D=2$ or $D=3$. And, letting $(K,L,S)$ be $(3,-1,2)$ and $(2,-1,3)$ in indeed gives that $(D,N)$ is $(2,3)$ and $(3,4)$, respectively.
In the next section, we provide a much more thorough discussion of positive and negative ETFs, including some other examples of ETFs that are both. For now, we turn to our main result, which shows how to combine a given ${{\operatorname{ETF}}}(D,N)$ whose parameters are type $(K,L,S)$ with a certain $K$-GDD to produce a new ETF whose parameters are type $(K,L,S')$ for some $S'>S$. Here, as with any GDD, we require $K\geq 2$. In light of the above discussion, this is not a significant restriction since $(D,N)$ has type $(1,L,S)$ if and only if $L=1$ and $S=D$, and we already know that ETFs of type $(1,1,D)$ exist for all $D>1$, being regular simplices.
\[thm.new ETF\] Assume an ETF of type $(K,L,S)$ exists where $K\geq 2$, and let $M=S(K-1)+L$. The necessary conditions on the existence of a $K$-GDD of type $M^U$ reduce to having $$\label{eq.new ETF 1}
U\geq K,
\quad
\tfrac{U-1}{K-1}\in{\mathbb{Z}},
\quad
\tfrac{(S-L)U(U-1)}{K(K-1)}\in{\mathbb{Z}}.$$ Moreover, if such a GDD exists, and $U$ has the additional property that $$\label{eq.new ETF 2}
\tfrac{(K-2)(U-1)}{(S+L)(K-1)}\in{\mathbb{Z}},$$ then there exists an ETF of type $(K,L,S')$ where $S'=S+R=\tfrac{MU-L}{K-1}$, where $R=\tfrac{M(U-1)}{K-1}$.
In particular, under these hypotheses, , and without loss of generality writing the given ETF as where ${\|{{\boldsymbol{\varphi}}_{m,i}}\|}^2=S$ for all $m$ and $i$, letting ${\{{{\mathbf{E}}_{u,m}}\}}_{u=1,}^{U}\,_{m=1}^{M}$ be the embedding operators of the GDD (Definition \[def.embeddings\]), letting ${\{{{\boldsymbol{\delta}}_u}\}}_{u=1}^{U}$ be the standard basis for ${\mathbb{F}}^U$, and letting ${\{{{\mathbf{e}}_i}\}}_{i=1}^{S+L}$ and ${\{{1\oplus{\mathbf{f}}_j}\}}_{j=0}^{W}$ be the columns of possibly-complex Hadamard matrices of size $S+L$ and $W+1$, respectively, then the following vectors form an ETF of type $(K,L,S')$: $$\label{eq.new ETF 3}
{\{{{\boldsymbol{\psi}}_{u,m,i,j}}\}}_{u=1,}^{U}\,_{m=1,}^{M}\,_{i=1,}^{S+L}\,_{j=0}^{W},
\quad
{\boldsymbol{\psi}}_{u,m,i,j}:=({\boldsymbol{\delta}}_u\otimes{\boldsymbol{\varphi}}_{m,i})\oplus{\bigl({{\mathbf{E}}_{u,m}({\mathbf{e}}_i\otimes{\mathbf{f}}_j)}\bigr)}.$$
As special cases of this fact, an ETF of type $(K,L,S')$ exists whenever either:
1. $U$ is sufficiently large and satisfies and ;
2. there exists a $K$-GDD of type $M^U$, provided we also have that $S+L$ divides $K-2$.
Since $M=S(K-1)+L$, $$\label{eq.pf of new neg ETF 1}
\tfrac{M}{K-1}
=S+\tfrac{L}{K-1}
=(S+L)-L{\bigl({\tfrac{K-2}{K-1}}\bigr)},
\quad
\tfrac{M}{K}
=S-\tfrac{S-L}{K}.$$ As such, the replication number of any $K$-GDD of type $M^U$ is $R=\frac{M(U-1)}{K-1}=S(U-1)+L(\frac{U-1}{K-1})$. In particular, such a GDD can only exist when $K-1$ necessarily divides $U-1$. Moreover, multiplying the expressions in gives that the number of blocks in any such GDD is $$\begin{aligned}
B
&=\tfrac{M^2U(U-1)}{K(K-1)}\\
&={\bigl({S-\tfrac{S-L}{K}}\bigr)}{\bigl({S+\tfrac{L}{K-1}}\bigr)}U(U-1)\\
&=S^2U(U-1)+LSU\tfrac{U-1}{K-1}-\tfrac{S(S-L)}{K}U(U-1)-L\tfrac{(S-L)U(U-1)}{K(K-1)}.\end{aligned}$$ Here, since our initial ${{\operatorname{ETF}}}(D,N)$ is type $(K,L,S)$, $\frac{S(S-L)}{K}=S^2-D$ is an integer, and so the above expression for $B$ is an integer precisely when $K(K-1)$ divides $(S-L)U(U-1)$. To summarize, since $M=S(K-1)+L$ where $K$ divides $S(S-L)$, the necessary conditions on the existence of a $K$-GDD of type $M^U$ reduce to having . These necessary conditions are known to be asymptotically sufficient [@Chang76; @LamkenW00]: for this fixed $K$ and $M$, there exists $U_0$ such that a $K$-GDD of type $M^U$ exists for any $U\geq U_0$ that satisfies . Regardless, to apply our construction below with any given $K$-GDD of type $M^U$, we only need $U$ to satisfy the additional property that $$W
=\tfrac{R}{S+L}
=\tfrac{M(U-1)}{(S+L)(K-1)}
=\tfrac{U-1}{S+L}{\bigl[{(S+L)-L{\bigl({\tfrac{K-2}{K-1}}\bigr)}}\bigr]}
=(U-1)-L\tfrac{(K-2)(U-1)}{(S+L)(K-1)}$$ is an integer, namely to satisfy . Since $K-1$ necessarily divides $U-1$, this is automatically satisfied whenever $S+L$ happens to divide $K-2$, and some of the ETFs we will identify in the next section will have this nice property. Regardless, there are always an infinite number of values of $U$ which satisfy and , including, for example, all $U\equiv 1\bmod (S+L)K(K-1)$.
Turning to the construction itself, the fact that the given ${{\operatorname{ETF}}}(D,N)$ is type $(K,L,S)$ implies $$\label{eq.proof of new ETF 1}
D
=\tfrac{S}{K}[S(K-1)+L]
=\tfrac{SM}{K},
\quad
N
=(S+L)[S(K-1)+L]
=(S+L)M.$$ In particular, since $N=(S+L)M$, the vectors in our initial ${{\operatorname{ETF}}}(D,N)$ can indeed be indexed as . Moreover, $$\label{eq.proof of new ETF 2}
MU
=M(U-1)+M
=(K-1)R+[S(K-1)+L]
=(S+R)(K-1)+L.$$ As such, the number of vectors in the collection is $$\label{eq.proof of new ETF 3}
N'
=UM(S+L)(W+1)
=(S+L)(\tfrac{R}{S+L}+1)MU
=[(S+R)+L][(S+R)(K-1)+L].$$ Also, for each $i$ and $j$, ${\mathbf{e}}_i\otimes{\mathbf{f}}_j$ lies in a space of dimension ${\mathbb{F}}^{(S+L)W}={\mathbb{F}}^R$. And, for each $u$ and $m$, ${\mathbf{E}}_{u,m}$ is a $B\times R$ matrix. As such, for any $u$, $m$, $i$ and $j$, ${\boldsymbol{\psi}}_{u,m,i,j}$ is a well-defined vector in ${\mathbb{F}}^{D'}$ where, by combining , and , we have $$\label{eq.proof of new ETF 4}
D'
=UD+B
=U\tfrac{SM}{K}+\tfrac{MU}{K}R
=\tfrac{S+R}{K}MU
=\tfrac{S+R}{K}[(S+R)(K-1)+L].$$ Comparing and against , we see that $(D',N')$ is indeed type $(K,L,S')$ where $S'=S+R$. Here, further implies $S'=S+R=\tfrac{MU-L}{K-1}$.
Continuing, since $(D',N')$ is type $(K,L,S')$, Theorem \[thm.parameter types\] gives that the Welch bound for $N'$ vectors in ${\mathbb{F}}^{D'}$ is . As such, to show is an ETF for ${\mathbb{F}}^{D'}$, it suffices to prove that ${\|{{\boldsymbol{\psi}}_{u,m,i,j}}\|}^2=S'$ for all $u$, $m$, $i$, $j$, and that ${\langle{{\boldsymbol{\psi}}_{u,m,i,j}},{{\boldsymbol{\psi}}_{u',m',i',j'}}\rangle}$ is unimodular whenever $(u,m,i,j)\neq(u',m',i',j')$. Here, for any $u,u'=1,\dotsc,U$, $m,m'=1,\dotsc,M$, $i,i'=1,\dotsc,S+L$, $j,j'=0,\dotsc,W$, $$\begin{aligned}
{\langle{{\boldsymbol{\psi}}_{u,m,i,j}},{{\boldsymbol{\psi}}_{u',m',i',j'}}\rangle}
\nonumber
&={\langle{({\boldsymbol{\delta}}_u\otimes{\boldsymbol{\varphi}}_{m,i})\oplus{\bigl({{\mathbf{E}}_{u,m}({\mathbf{e}}_i\otimes{\mathbf{f}}_j)}\bigr)}},{({\boldsymbol{\delta}}_{u'}\otimes{\boldsymbol{\varphi}}_{m',i'})\oplus{\bigl({{\mathbf{E}}_{u',m'}({\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'})}\bigr)}}\rangle}\\
\nonumber
&={\langle{{\boldsymbol{\delta}}_u},{{\boldsymbol{\delta}}_{u'}}\rangle}{\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m',i'}}\rangle}
+{\langle{{\mathbf{E}}_{u,m}({\mathbf{e}}_i\otimes{\mathbf{f}}_j)},{{\mathbf{E}}_{u',m'}({\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'})}\rangle}\\
\label{eq.proof of new ETF 5}
&={\langle{{\boldsymbol{\delta}}_u},{{\boldsymbol{\delta}}_{u'}}\rangle}{\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m',i'}}\rangle}
+{\langle{{\mathbf{e}}_i\otimes{\mathbf{f}}_j},{{\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}({\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'})}\rangle}.\end{aligned}$$ When $u=u'$, $m=m'$, $i=i'$ and $j=j'$, indeed becomes $${\|{{\boldsymbol{\psi}}_{u,m,i,j}}\|}^2
={\|{{\boldsymbol{\delta}}_u}\|}^2{\|{{\boldsymbol{\varphi}}_{m,i}}\|}^2+{\|{{\mathbf{e}}_i}\|}^2{\|{{\mathbf{f}}_j}\|}^2
=1S+(S+L)W
=S+R
=S'.$$ As such, all that remains is to show that is unimodular in all other cases. For instance, if $u\neq u'$ then Lemma \[lem.embed\] gives that for any $m,m'$ there exists $r,r'=1,\dotsc,R$ such that ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u',m'}^{}={\boldsymbol{\delta}}_{r}^{}{\boldsymbol{\delta}}_{\smash{r'}}^*$ meaning in this case becomes $${\langle{{\boldsymbol{\psi}}_{u,m,i,j}},{{\boldsymbol{\psi}}_{u',m',i',j'}}\rangle}
=0{\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m',i'}}\rangle}
+{\langle{{\mathbf{e}}_i\otimes{\mathbf{f}}_j},{{\boldsymbol{\delta}}_{r}^{}{\boldsymbol{\delta}}_{\smash{r'}}^*({\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'})}\rangle}
=\overline{({\mathbf{e}}_i\otimes{\mathbf{f}}_j)(r)}({\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'})(r'),$$ which is unimodular, being a product of unimodular numbers. If we instead have $u=u'$ and $m\neq m'$ then Lemma \[lem.embed\] gives ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u,m'}^{}={\boldsymbol{0}}$ and so becomes $${\langle{{\boldsymbol{\psi}}_{u,m,i,j}},{{\boldsymbol{\psi}}_{u,m',i',j'}}\rangle}
={\|{{\boldsymbol{\delta}}_u}\|}^2{\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m',i'}}\rangle}
+{\langle{{\mathbf{e}}_i\otimes{\mathbf{f}}_j},{{\boldsymbol{0}}({\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'})}\rangle}
={\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m',i'}}\rangle},$$ which is unimodular since ${\{{{\boldsymbol{\varphi}}_{m,i}}\}}_{m=1,}^{M}\,_{i=1}^{S+L}$ is an ETF of type $(K,L,S)$, and has been scaled so that ${\|{{\boldsymbol{\varphi}}_{m,i}}\|}^2=S$ for all $m$ and $i$. Next, if we instead have $u=u'$ and $m=m'$ then Lemma \[lem.embed\] gives ${\mathbf{E}}_{u,m}^*{\mathbf{E}}_{u,m}^{}={\mathbf{I}}$ and so becomes $$\label{eq.proof of new ETF 6}
{\langle{{\boldsymbol{\psi}}_{u,m,i,j}},{{\boldsymbol{\psi}}_{u,m,i',j'}}\rangle}
={\|{{\boldsymbol{\delta}}_u}\|}^2{\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m,i'}}\rangle}+{\langle{{\mathbf{e}}_i\otimes{\mathbf{f}}_j},{{\mathbf{e}}_{i'}\otimes{\mathbf{f}}_{j'}}\rangle}
={\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m,i'}}\rangle}+{\langle{{\mathbf{e}}_i},{{\mathbf{e}}_{i'}}\rangle}{\langle{{\mathbf{f}}_j},{{\mathbf{f}}_{j'}}\rangle}.$$ In particular, when $u=u'$, $m=m'$ and $i\neq i'$, the fact that ${\{{{\mathbf{e}}_i}\}}_{i=1}^{S+L}$ is orthogonal implies that reduces to ${\langle{{\boldsymbol{\varphi}}_{m,i}},{{\boldsymbol{\varphi}}_{m,i'}}\rangle}$, which is unimodular for the same reason as the previous case. The final remaining case is the most interesting: when $u=u'$, $m=m'$, $i=i'$ but $j\neq j'$, we have $0={\langle{1\oplus{\mathbf{f}}_j},{1\oplus{\mathbf{f}}_{j'}}\rangle}=1+{\langle{{\mathbf{f}}_j},{{\mathbf{f}}_{j'}}\rangle}$ and so ${\langle{{\mathbf{f}}_j},{{\mathbf{f}}_{j'}}\rangle}=-1$; when combined with the fact that ${\|{{\boldsymbol{\varphi}}_{m,i}}\|}^2=S$ and ${\|{{\mathbf{e}}_i}\|}^2=S+L$, this implies that in this case becomes $${\langle{{\boldsymbol{\psi}}_{u,m,i,j}},{{\boldsymbol{\psi}}_{u,m,i,j'}}\rangle}
={\|{{\boldsymbol{\varphi}}_{m,i}}\|}^2+{\|{{\mathbf{e}}_i}\|}^2{\langle{{\mathbf{f}}_j},{{\mathbf{f}}_{j'}}\rangle}
=S+(S+L)(-1)
=-L,$$ where, in Definition \[def.ETF types\], we have assumed that $L$ is either $1$ or $-1$.
The construction of Theorem \[thm.new ETF\] leads to the concept of ETFs of type $(K,L,S)$. To clarify, the construction and the above proof of its equiangularity is valid for any initial ${{\operatorname{ETF}}}(D,N)$ and any $K$-GDD of type $M^U$, provided $M=\frac{N}{S+L}$ for some $L\in{\{{-1,1}\}}$, and $S+L$ divides . However, it turns out that the first $UD$ rows of the corresponding synthesis operator have squared-norm $(W+1)\frac{NS}{D}$, whereas the last $B$ rows have squared-norm $(W+1)K(S+L)$. As such, the equiangular vectors are only a tight frame when . Applying the techniques of the proof of Theorem \[thm.parameter types\] then leads to the expressions for $(D,N)$ in terms of $(K,L,S)$ given in . These facts are not explicitly discussed in our proof above since any equal-norm vectors that attain the Welch bound are automatically tight.
\[rem.recursive\] There is no apparent value to recursively applying Theorem \[thm.new ETF\]. To elaborate, given an ${{\operatorname{ETF}}}(D,N)$ of type $(K,L,S)$ and a $K$-GDD of type $M^U$ where $M=S(K-1)+L$ and $U$ satisfies , Theorem \[thm.new ETF\] yields an ${{\operatorname{ETF}}}(D',N')$ of type $(K,L,S')$ where $S'=\frac{MU-L}{K-1}$. The “$M$" parameter of this new ETF is thus $M'=S'(K-1)+L=(MU-L)+L=MU$, and we can apply Theorem \[thm.new ETF\] a second time provided we have a $K$-GDD of type $(MU)^{U'}$ where $U'$ satisfies the appropriate analog of , namely $$\label{eq.recursive 1}
\tfrac{(K-2)(U'-1)}{(S'+L)(K-1)}\in{\mathbb{Z}}.$$ Doing so yields an ETF of type $(K,-1,S'')$ where $S''=\tfrac{M'U'-L}{K-1}=\tfrac{MUU'-L}{K-1}$. However, under these hypotheses, there is a simpler way to construct an ETF of this same type. Indeed, using the first GDD to fill the holes of the second GDD via Lemma \[lem.filling holes\] produces a $K$-GDD of type $M^{UU'}$. Moreover, $UU'$ is a value of “$U$" that satisfies : since $S'=S+R$, $$\tfrac{S'+L}{S+L}=\tfrac{S'-S}{S+L}+1=\tfrac{R}{S+L}+1=W+1\in{\mathbb{Z}},$$ and this together with and imply $$\tfrac{(K-2)(UU'-1)}{(S+L)(K-1)}
=\tfrac{(K-2)[U(U'-1)+(U-1)]}{(S+L)(K-1)}
=U\tfrac{(S'+L)}{(S+L)}\tfrac{(K-2)(U'-1)}{(S'+L)(K-1)}+\tfrac{(K-2)(U-1)}{(S+L)(K-1)}
\in{\mathbb{Z}}.$$ As such, we can combine our original ETF of type $(K,L,S)$ with our $K$-GDD of type $M^{UU'}$ via Theorem \[thm.new ETF\] to directly produce an ETF of type $(K,L,\tfrac{MUU'-L}{K-1})$.
We also point out that, in a manner analogous to how every McFarland difference set can be viewed as a degenerate instance of a Davis-Jedwab difference set [@DavisJ97], every Steiner ETF can be regarded as a degenerate case of the construction of Theorem \[thm.new ETF\]. Here, in a manner consistent with , we regard $(D,N)=(0,1)$ as being type $(K,L,S)=(K,1,0)$ where $K\geq 1$ is arbitrary. Under this convention, $M=S(K-1)+L=1$, and $W=R$, meaning we need a $K$-GDD of type $1^U$, namely a ${{\operatorname{BIBD}}}(V,K,1)$ where $V=U$. When such a BIBD exists, Theorem \[thm.new ETF\] suggests we let ${\{{{\boldsymbol{\varphi}}_{1,1}}\}}$ be some fictitious ETF for the nonexistent space ${\mathbb{F}}^0$, scaled so that . It also suggests we let be the embedding operators of the BIBD, let ${\mathbf{e}}_1=1$, and let ${\{{1\oplus{\mathbf{f}}_j}\}}_{j=0}^{R}$ be the columns of a possibly-complex Hadamard matrix of size $R$. Under these conventions, reduces to a collection of $V(R+1)$ vectors , where for any $v=u$ and $j$, the fact that ${\boldsymbol{\varphi}}_{1,1}$ lies in a “zero-dimensional space" makes it reasonable to regard $${\boldsymbol{\psi}}_{v,j}
={\boldsymbol{\psi}}_{u,1,1,j}
=({\boldsymbol{\delta}}_v\otimes{\boldsymbol{\varphi}}_{1,1})\oplus{\bigl({{\mathbf{E}}_{u,1}({\mathbf{e}}_1\otimes{\mathbf{f}}_j)}\bigr)}
={\mathbf{E}}_{u,1}(1\otimes{\mathbf{f}}_j)
={\mathbf{E}}_{u}{\mathbf{f}}_j.$$ Here, Theorem \[thm.new ETF\] leads us to expect that ${\{{{\mathbf{E}}_v{\mathbf{f}}_j}\}}_{v=1,}^{V}\,_{j=0}^{R}$ is an ETF of type $(K,1,S+R)=(K,1,R)$.
This is indeed the case: as discussed in the previous section, ${\{{{\mathbf{E}}_v{\mathbf{f}}_j}\}}_{v=1,}^{V}\,_{j=0}^{R}$ is by definition a Steiner ETF for ${\mathbb{F}}^B$ where $B=\frac{VR}{K}$, and moreover letting $(K,L,S)=(K,1,R)$ in gives: $$\begin{aligned}
\tfrac{S}{K}[S(K-1)+L]
&=\tfrac{R}{K}[R(K-1)+1]
=\tfrac{R}{K}V
=B,\\
(S+L)[S(K-1)+L]
&=(R+1)[R(K-1)+1]
=V(R+1).\end{aligned}$$ In particular, every Steiner ETF is a positive ETF.
As a degenerate case of Remark \[rem.recursive\], we further have that when a Steiner ETF is regarded as being positive, applying Theorem \[thm.new ETF\] to it yields an ETF whose parameters match those of another Steiner ETF. Indeed, since a Steiner ETF arising from a ${{\operatorname{BIBD}}}(V,K,1)$ is type $(K,1,R)$ where $R=\frac{V-1}{K-1}$, we can only apply Theorem \[thm.new ETF\] to it whenever there exists a $K$-GDD of type $M^U$ where $M=R(K-1)+1=V$ and $U$ satisfies . In this case, the resulting ETF is type $(K,1,\frac{VU-1}{K-1})$. However, under these same hypotheses, we can more simply use the ${{\operatorname{BIBD}}}(V,K,1)$ to fill the holes of the $K$-GDD of type $V^U$ via Lemma \[lem.filling holes\] to obtain a ${{\operatorname{BIBD}}}(UV,K,1)$, and its corresponding Steiner ETF is type $(K,1,\frac{VU-1}{K-1})$.
Families of positive and negative equiangular tight frames
==========================================================
In this section, we use Theorems \[thm.parameter types\] and \[thm.new ETF\] to better our understanding of positive and negative ETFs, in particular proving the existence of the new ETFs given in Theorems \[thm.new neg ETS with K=4,5\] and \[thm.new neg ETS with K>5\]. Recall that by , any ${{\operatorname{ETF}}}(D,N)$ of type $(K,L,S)$ has $D=\tfrac{S}{K}[S(K-1)+L]=S^2-\tfrac{S(S-L)}{K}$, and so $K$ necessarily divides $S(S-L)$. As we shall see, it is reasonable to conjecture that such ETFs exist whenever this necessary condition is satisfied.
Positive equiangular tight frames
---------------------------------
In light of Definition \[def.ETF types\] and Theorem \[thm.parameter types\], an ${{\operatorname{ETF}}}(D,N)$ with $1<D<N$ is positive if and only if there exists integers $K\geq 1$ and $S\geq 2$ such that $$\label{eq.positive ETF param}
D=\tfrac{S}{K}[S(K-1)+1]=S^2-\tfrac{S(S-1)}{K},
\quad
N=(S+1)[S(K-1)+1],$$ or equivalently that and $K=\frac{NS}{D(S+1)}$ are integers.
As discussed in the previous section, every regular simplex is a $1$-positive ETF and vice versa, and moreover, every Steiner ETF arising from a ${{\operatorname{BIBD}}}(V,K,1)$ is type $(K,1,R)$ where $R=\frac{V-1}{K-1}$. Here, the fact that $K$ necessarily divides $S(S-1)=R(R-1)$ is equivalent to having that $D=B$ is necessarily an integer. *Fisher’s inequality* also states that a ${{\operatorname{BIBD}}}(V,K,1)$ can only exist when $K\leq R$. These necessary conditions on the existence of a ${{\operatorname{BIBD}}}(V,K,1)$ are known to also be sufficient when $K=2,3,4,5$, and also asymptotically sufficient in general: for any $K\geq 2$, there exists $V_0$ such that for all $V\geq V_0$ with the property that $R=\frac{V-1}{K-1}$ and $B=\frac{VR}{K}=\frac{V(V-1)}{K(K-1)}$, a ${{\operatorname{BIBD}}}(V,K,1)$ exists [@AbelG07]. Moreover, explicit infinite families of such BIBDs are known, including affine geometries, projective geometries, unitals and Denniston designs [@FickusMT12], and each thus gives rise to a corresponding infinite family of (positive) Steiner ETFs.
That said, not every positive ETF is a Steiner ETF. In particular, for any prime power $Q$, there is a phased BIBD ETF [@FickusJMPW19] whose Naimark complements are ${{\operatorname{ETF}}}(D,N)$ where $$D=\tfrac{Q^3+1}{Q+1},
\quad
N=Q^3+1,
\quad
S={\bigl({\tfrac{N-1}{\frac ND-1}}\bigr)}^{\frac12}={\bigl({\tfrac{Q^3}{Q}}\bigr)}^{\frac12}=Q,
\quad
K=\tfrac{NS}{D(S+1)}=\tfrac{(Q+1)Q}{Q+1}=Q,$$ namely an ETF of type $(Q,1,Q)$. These parameters match those of a Steiner ETF arising from a projective plane of order $Q-1$. However, such an ETF can exist even when no such projective plane exists. For example, such an ETF exists when $Q=7$ despite the fact that no projective plane of order $Q-1=6$ exists.
Other ETFs of type $(K,1,S)$ have $K>S$, and are thus not Steiner ETFs since the underlying BIBD would necessarily violate Fisher’s inequality. To elaborate, when $K$ is a prime power, the requirement that $K$ divides $S(S-1)$ where $S$ and $S-1$ are relatively prime implies that either $K$ divides $S$ or $S-1$, and in either case $K\leq S$. However, when $K$ is not a prime power, we can sometimes choose it to be a divisor of $S(S-1)$ that is larger than $S$.
For example, taking $K=S(S-1)$ in gives $D=S^2-1$ and $N=(S^2-1)^2=D^2$. Such an ETF thus corresponds to a SIC-POVM in a space whose dimension is one less than a perfect square. Such SIC-POVMs are known to exist when $S=2,\dotsc,7,18$, and are conjectured to exist for all $S$ [@FuchsHS17; @GrasslS17]. Similarly, taking in gives $D=S^2-2$ and . We refer to such $(D,N)$ as being of *real maximal type* since real-valued examples of such ETFs meet the real-variable version of the Gerzon bound, and are known to exist when $S=3,5$. Remarkably, it is known that a real ETF of this type does not exist when $S=7$ [@FickusM16]. We also caution that there is a single pair $(D,N)$ with $N=\binom{D+1}{2}$ that is neither positive or negative despite the fact that an ${{\operatorname{ETF}}}(D,N)$ exists, namely $(D,N)=(3,6)$.
We summarize these facts as follows:
\[thm.known positive ETFs\] An ETF of type $(K,1,S)$ exists whenever:
1. $K=1$ and $S\geq 2$, (regular simplices);
2. $K\geq 2$ and there exists a ${{\operatorname{BIBD}}}(S(K-1)+1,K,1)$ (Steiner ETFs [@FickusMT12]), including:
1. $K=2,3,4,5$ and $S\geq K$ has the property that $K$ divides $S(S-1)$;
2. $K=Q$ and $S=\frac{Q^J-1}{Q-1}$ where $Q$ is a prime power and $J\geq 2$ (affine geometries);
3. $K=Q+1$ and $S=\frac{Q^J-1}{Q-1}$ where $Q$ is a prime power and $J\geq 2$ (projective geometries);
4. $K=Q+1$ and $S=Q^2$ where $Q$ is a prime power (unitals);
5. $K=2^{J_1}$ and $S=2^{J_2}+1$ where $2\leq J_1<J_2$ (Denniston designs);
6. $K\geq2$ and $S$ is sufficiently large and has the property that $K$ divides $S(S-1)$;
3. $K=Q$ and $S=Q$ whenever $Q$ is a prime power [@FickusJMPW19];
4. $K=S(S-1)$ where $S=2,\dotsc,7,18$ (SIC-POVMs [@FuchsHS17; @GrasslS17]);
5. $K=\binom{S}{2}$ where $S=3,5$ (real maximal type [@FickusM16]).
Because so much is already known regarding the existence of positive ETFs, we could not find any examples where Theorem \[thm.new ETF\] makes a verifiable contribution. As we now discuss, much less is known about negative ETFs, and this gives Theorem \[thm.new ETF\] an opportunity to be useful.
Negative equiangular tight frames
---------------------------------
By Definition \[def.ETF types\] and Theorem \[thm.parameter types\], an ${{\operatorname{ETF}}}(D,N)$ with $1<D<N$ is negative if and only if there exists integers $K\geq 1$ and $S\geq 2$ such that $$\label{eq.negative ETF param}
D=\tfrac{S}{K}[S(K-1)-1]=S^2-\tfrac{S(S+1)}{K},
\quad
N=(S-1)[S(K-1)-1],$$ or equivalently that and $K=\frac{NS}{D(S-1)}$ are integers. Here, since $\frac{N}{D}>1$ and $\frac{S}{S-1}>1$, we actually necessarily have that $K\geq 2$.
When $K=2$, becomes $(D,N)=(\frac12S(S-1),(S-1)^2)$ and such ETFs exist for any $S\geq 3$, being the Naimark complements of ETFs of type $(2,1,S-1)$.
In the $K=3$ case, becomes $$D
=\tfrac{S(2S-1)}{3}
=S^2-\tfrac{S(S+1)}{3},
\quad
N=(S-1)(2S-1).$$ For $D$ to be an integer, we necessarily have $S\equiv0,2\bmod 3$. Moreover, for any $S\geq 2$ with $S\equiv 0,2\bmod 3$, an ETF of type $(3,-1,S)$ exists. Indeed, the recent paper [@FickusJMP18] gives a way to modify the Steiner ETF arising from a ${{\operatorname{BIBD}}}(V,3,1)$ to yield a Tremain ${{\operatorname{ETF}}}(D,N)$ with $$D=\tfrac16(V+2)(V+3),\quad
N=\tfrac12(V+1)(V+2),$$ for any $V\geq 3$ with $V\equiv1,3\bmod 6$. Such an ETF is type $(3,-1,S)$ with $S=\frac12(V+3)$. We also note that for every $J\geq1$, there is a harmonic ${{\operatorname{ETF}}}(D,N)$ arising from a Spence difference set [@Spence77] with $(D,N)=(\tfrac12 3^{J}(3^{J+1}+1),\tfrac12 3^{J+1}(3^{J+1}-1))$, and such an ETF is type $(3,-1,S)$ where $S=\frac12(3^{J+1}+1)$. It remains unclear whether any Spence ETFs are unitarily equivalent to special instances of Tremain ETFs.
In the $K=4$ case, for any positive integer $J$, Davis and Jedwab [@DavisJ97] give a difference set whose harmonic ETF has parameters and so is a type $(4,-1,S)$ ETF where $S=\frac13(2^{2J+1}+1)$. Beyond these examples, a few other infinite families of negative ETFs are known to exist. In particular, in order for the expression for $D$ given in to be an integer, $K$ necessarily divides $S(S+1)$, and so it is natural to consider the special cases where $S=K$ and $S=K-1$.
When $S=K-1$, gives that ETFs of type $(K,-1,K-1)$ have $$D=(K-1)(K-2),
\quad
N=K(K-2)^2.$$ Remarkably, these are the same $(D,N)$ parameters as those of an ETF of type $(K-2,1,K-1)$. In particular, every Steiner ETF arising from an affine plane of order $Q$ is both of (positive) type $(Q,1,Q+1)$ and (negative) type $(Q+2,-1,Q+1)$. More generally, a Steiner ETF arising from a ${{\operatorname{BIBD}}}(V,K,1)$ has $(D,N)=(B,V(R+1))$ where $R=\frac{V-1}{K-1}$ and $B=\frac{VR}{K}$, and so is only negative when $$\label{eq.when Steiner ETF is neg}
\tfrac{NS}{D(S-1)}
=\tfrac{V(R+1)R}{B(R-1)}
=\tfrac{K(R+1)}{R-1}$$ is an integer. When $R$ is even, $R-1$ and $R+1$ are relatively prime, and this can only occur when $R-1$ divides $K$. Here, since Fisher’s inequality gives $K\leq R$, this happens precisely when either $R=K=2$ or $R=K+1$, namely when the underlying BIBD is a ${{\operatorname{BIBD}}}(3,2,1)$ or is an affine plane of odd order $K$, respectively. Meanwhile, when $R$ is odd, $R-1$ and $R+1$ have exactly one prime factor in common, namely $2$, and is an integer precisely when $\frac12(R-1)$ divides $K$. Since $K\leq R$, this happens precisely when either $R=K=3$, $R=K+1$ or $R=2K+1$, namely when the underlying BIBD is the projective plane of order $2$, an affine plane of even order, or is a ${{\operatorname{BIBD}}}(V,K,1)$ where $V=(2K+1)(K-1)+1=K(2K-1)$ for some $K\geq 2$, respectively. With regard to the latter, it seems to be an open question whether a ${{\operatorname{BIBD}}}(K(2K-1),K,1)$ exists for every $K\geq 2$, though a Denniston design provides one whenever $K=2^J$ for some $J\geq1$, and they are also known to exist when $K=3,5,6,7$ [@MathonR07]. For such ETFs, becomes $K+1$, meaning they are type $(K',-1,2K'-1)$ where $K'=K+1$.
Meanwhile, when $S=K$, gives that ETFs of type $(K,-1,K)$ have $$\label{eq.S=K param}
D=K^2-K-1,
\quad
N=(K-1)(K^2-K-1).$$ The recently-discovered hyperoval ETFs of [@FickusMJ16] are instances of such ETFs whenever $K=2^J+1$ for some $J\geq 1$. In the $K=4$ case, becomes $(D,N)=(11,33)$, and this seems to be the smallest set of positive or negative parameters for which the existence of a corresponding ETF remains an open problem.
Apart from these examples, it seems that only a finite number of other negative ETFs are known to exist. For example, since $K$ necessarily divides $S(S+1)$, it is natural to also consider the cases where $K=S(S+1)$ and $K=\binom{S+1}{2}$. When $K=S(S+1)$, becomes $(D,N)=(S^2-1,(S^2-1)^2)$. In particular, every positive SIC-POVM is also negative. Similarly, when $K=\binom{S+1}{2}$, gives $N=\binom{D+1}{2}$ where $D=S^2-2$, meaning every positive $(D,N)$ of real maximal type is also negative.
The only other example of a negative ETF that we found in the literature was an ${{\operatorname{ETF}}}(22,176)$, which has type $(10,-1,5)$, and arises from a particular SRG. When searching tables of known ETFs such as [@FickusM16], it is helpful to note that most positive and negative ETFs have redundancy $\frac{N}{D}>2$, with the only exceptions being $1$-positive ETFs (regular simplices), $2$-negative ETFs (Naimark complements of $2$-positive ETFs), ${{\operatorname{ETF}}}(2,3)$ when regarded as type $(3,-1,2)$, and ${{\operatorname{ETF}}}(5,10)$, which are type $(3,-1,3)$. Indeed, when $K\geq2$, any $K$-positive ETF has $\tfrac{N}{D}=\frac{K(S+1)}{S}>K\geq2$. Meanwhile, when $K\geq 3$, any ETF of type $(K,-1,S)$ only has $\frac{K(S-1)}{S}=\tfrac{N}{D}\leq 2$ when $S\leq\frac{K}{K-2}$. Since $K$ necessarily divides $S(S+1)$ where $S\geq 2$, such an ETF can only exist when $K=3$ and $S=2,3$. We summarize these previously-known constructions of negative ETFs as follows:
\[thm.known negative ETFs\] An ETF of type $(K,-1,S)$ exists whenever:
1. $K=2$ and $S\geq 3$ (Naimark complements of $2$-positive ETFs);
2. $K=3$ and $S\geq 2$ with $S\equiv 0,2\bmod 3$ (Tremain ETFs [@FickusJMP18] and Spence harmonic ETFs [@Spence77]);
3. $K=4$ and $S=\frac13(2^{2J+1}+1)$ for some $J\geq1$ (Davis-Jedwab harmonic ETFs [@DavisJ97]);
4. $K=Q+2$ and $S=K-1$ where $Q$ is a prime power (Steiner ETFs from affine planes [@FickusMT12]);
5. $K=2^J+1$ and $S=2K-1$ where $J\geq 1$ (Steiner ETFs from Denniston designs [@FickusMT12]);
6. $K=2^J+1$ and $S=K$ where $J\geq 1$ (hyperoval ETFs [@FickusMJ16]);
7. $K=S(S+1)$ where $S=2,\dotsc,7,18$ (SIC-POVMs [@FuchsHS17]);
8. $K=\binom{S+1}{2}$ where $S=3,5$ (real maximal type [@FickusM16]);
9. $(K,S)=(4,7),(6,11),(7,13),(8,15),(10,5)$ (various other ETFs [@FickusM16]).
From this list, we see that for any $K\geq 5$, the existing literature provides at most a finite number of $K$-negative ETFs. Theorem \[thm.new ETF\](a) implies that many more negative ETFs exist: if an ETF of type $(K,-1,S)$ exists, then an ETF of type $(K,-1,\frac{MU+1}{K-1})$ exists for all sufficiently large $U$ that satisfy and . Combining this fact with Theorem \[thm.known negative ETFs\] immediately gives:
There exists an infinite number of $K$-negative ETFs whenever:
1. $K=Q+2$ where $Q$ is a prime power;
2. $K=Q+1$ where $Q$ is an even prime power;
3. $K=2,8,12,20,30,42,56,342$.
In particular, we now know that there are an infinite number of values of $K$ for which an infinite number of $K$-negative ETFs exist, with $K=14$ being the smallest open case. With these asymptotic existence results in hand, we now focus on applying Theorem \[thm.new ETF\] with explicit GDDs.
For example, the “Mercedes-Benz" regular simplex ${{\operatorname{ETF}}}(2,3)$ is type $(K,L,S)=(3,-1,2)$ and so has $M=S(K-1)+L=3$. Since $S+L=1$ divides $K-2=1$, Theorem \[thm.new ETF\] can be applied with any $3$-GDD of type $3^U$ so as to produce an ETF of type $(K,L,\tfrac{MU-L}{K-1})=(3,-1,\tfrac12(3U+1))$, and moreover that the necessary conditions on the existence of such a GDD reduce to , namely to having $U\geq 3$, $\tfrac12(U-1)\in{\mathbb{Z}}$ and $\tfrac12U(U-1)\in{\mathbb{Z}}$. In fact, such GDDs are known to exist whenever these necessary conditions are satisfied [@Ge07], namely when $U\geq 3$ is odd. (This also follows from the fact that such GDDs are equivalent to the incidence structures obtained by removing a parallel class from a resolvable Steiner triple system.) That is, writing $U=2J+1$ for some $J\geq 1$, we can apply Theorem \[thm.new ETF\] with an ${{\operatorname{ETF}}}(2,3)$ and a known $3$-GDD of type $3^{2J+1}$ to produce an ETF of type $(3,-1,\tfrac12(3U+1))=(3,-1,3J+2)$.
In summary, applying Theorem \[thm.new ETF\] to an ETF of type $(3,-1,2)$ produces ETFs of type $(3,-1,S)$ for any $S\equiv 2\bmod 3$, and so recovers the parameters of “half" of all possible $3$-negative ETFs, cf. Theorem \[thm.known negative ETFs\] and [@FickusJMP18], including the parameters of all harmonic ETFs arising from Spence difference sets. To instead recover some of the ETFs of type $(3,-1,S)$ with $S\equiv0\bmod 3$, one may, for example, apply Theorem \[thm.new ETF\] to the well-known ${{\operatorname{ETF}}}(5,10)$, which is type $(3,-1,3)$.
In order to obtain ETFs with verifiably new parameters, we turn our attention to applying Theorem \[thm.new ETF\] to known $K$-negative ETFs with $K\geq 4$. Here, the limiting factor seems to be a lack of knowledge regarding uniform $K$-GDDs: while the literature has much to say when $K=3,4,5$ [@Ge07], we are relegated to well-known simple constructions involving Lemmas \[lem.Wilson\] and \[lem.filling holes\] whenever $K>5$. As such, we consider the $K=4,5$ cases separately from those with $K>5$:
The ${{\operatorname{ETF}}}(6,16)$ is type $(4,-1,3)$, and so we can apply Theorem \[thm.new ETF\] whenever there exists a $4$-GDD of type $8^U$ where $U$ satisfies . By Theorem \[thm.new ETF\], the known necessary conditions on the existence of such GDDs reduce to : $$U\geq 4,
\quad
\tfrac{U-1}{3}\in{\mathbb{Z}},
\quad
\tfrac{U(U-1)}{3}=\tfrac{4U(U-1)}{4(3)}\in{\mathbb{Z}},$$ namely to having $U\geq 4$ and $U\equiv 1\bmod 3$. These necessary conditions on the existence of $4$-GDDs of type $8^U$ are known to be sufficient [@Ge07]. Moreover, for any such $U$, we have is automatically satisfied since . Altogether, for any $U\geq 4$ with $U\equiv 1\bmod 3$, we can apply Theorem \[thm.new ETF\] to the ETF of type $(4,-1,3)$ with a $4$-GDD of type $8^U$, and doing so produces an ETF of type $(4,-1,\frac13(8U+1))$. Here, letting $U=1$ recovers the parameters of the original ETF. Overall, writing $U=3J+1$ for some $J\geq0$, this means that an ETF of type $(4,-1,S)$ exists whenever $S=\frac13(8U+1)=\frac13[8(3J+1)+1]=8J+3$ for any $J\geq0$, namely whenever $S\equiv 3\bmod 8$. In particular, for any $J\geq 1$ we can take $U=4^{J-1}$ to obtain an ETF of type $(4,-1,S)$ where $S=\tfrac13(8U+1)=\tfrac13(2^{2J-1}+1)$. This means that applying Theorem \[thm.new ETF\] to the ETF of type $(4,-1,3)$ recovers the parameters of harmonic ETFs corresponding to Davis-Jedwab difference sets.
In light of Remark \[rem.recursive\], applying Theorem \[thm.new ETF\] to any ETF of type $(4,-1,S)$ with $S\equiv 3\bmod 8$ simply recovers a subset of the ETF types obtained by applying Theorem \[thm.new ETF\] to the ETF of type $(4,-1,3)$. As such, to obtain more $4$-negative ETFs via Theorem \[thm.new ETF\], we need to apply it to initial ETFs that lie outside of this family. By Theorem \[thm.known negative ETFs\], the existing literature gives one such set of parameters, namely ETFs of type $(4,-1,7)$, which have $(D,N)=(35,120)$. Here, $M=20$, and a $4$-GDD of type $20^U$ can only exist if $U$ satisfies : $$U\geq 4,
\quad
\tfrac{U-1}{3}\in{\mathbb{Z}},
\quad
\tfrac{U(U-1)}{3}=\tfrac{4U(U-1)}{4(3)}\in{\mathbb{Z}}.$$ Moreover, these necessary conditions are known to be sufficient [@Ge07], meaning Theorem \[thm.new ETF\] can be applied whenever $U$ also satisfies , namely . Thus, for any $U\equiv 1\bmod 9$, Theorem \[thm.new ETF\] yields an ETF of type $(4,-1,\frac13(20U+1))$. In particular, an ETF of type $(4,-1,S)$ exists for any $S\equiv 7\bmod 60$. ETFs of type $(4,-1,S)$ with $S\equiv 67\bmod 120$ thus arise from both constructions; in the statement of Theorem \[thm.new neg ETS with K=4,5\], we elect to not remove the overlapping values of $S$ from either family so as to not emphasize one family over the other, and possibly make it easier for future researchers to identify potential patterns.
In a similar manner, as summarized in Theorem \[thm.known negative ETFs\], the existing literature provides ETFs of type $(5,-1,S)$ for exactly three values of $S$, namely ${{\operatorname{ETF}}}(12,45)$, ${{\operatorname{ETF}}}(19,76)$ and ${{\operatorname{ETF}}}(63,280)$ which have $S=4,5,9$ and so $M=15,19,35$, respectively. For these particular values of $M$, the corresponding necessary conditions on the existence of $5$-GDDs of type $M^U$ are known to be sufficient [@Ge07], meaning we only need $U$ to satisfy and for the corresponding value of $S$, namely to satisfy $U\geq 5$, $\frac14(U-1)\in{\mathbb{Z}}$ and that $$\begin{aligned}
\tfrac{5U(U-1)}{5(4)}\in{\mathbb{Z}},
\quad
\tfrac{3(U-1)}{3(4)}\in{\mathbb{Z}},&\text{ when }S=4,\\
\tfrac{6U(U-1)}{5(4)}\in{\mathbb{Z}},
\quad
\tfrac{3(U-1)}{4(4)}\in{\mathbb{Z}},&\text{ when }S=5,\\
\tfrac{10U(U-1)}{5(4)}\in{\mathbb{Z}},
\quad
\tfrac{3(U-1)}{8(4)}\in{\mathbb{Z}},&\text{ when }S=9.\end{aligned}$$ An ETF of type $(5,-1,S)$ thus exists when $S=\frac14(15U+1)$ with $U\equiv 1\bmod 4$, or $S=\frac14(19U+1)$ with $U\equiv 1,65\bmod 80$, or $S=\frac14(35U+1)$ with $U\equiv 1\bmod 32$. That is, an ETF of type $(5,-1,S)$ exists when $S\equiv 4\bmod 15$, or $S\equiv 5,309\bmod 380$, or $S\equiv 9\bmod 280$.
In certain special cases, these techniques yield real ETFs:
The ETF constructed in Theorem \[thm.new ETF\] is clearly real when the initial ${{\operatorname{ETF}}}(D,N)$ ${\{{{\boldsymbol{\varphi}}_n}\}}_{n=1}^{N}$ and the Hadamard matrices of size $S+L$ and $W+1$ are real.
In particular, a real ${{\operatorname{ETF}}}(63,280)$ exists [@FickusM16], and such ETFs are type $(5,-1,9)$. Since a real Hadamard matrix of size $S+L=8$ exists, then letting $M=S(K-1)+L=35$, Theorem \[thm.new ETF\] yields a real ETF of type $(5,-1,\frac14(35U+1))$ whenever there exists a $5$-GDD of type $35^U$ where $U$ satisfies and there exists a real Hadamard matrix of size $$H
=W+1
=\tfrac{R}{S+L}+1
=\tfrac{M(U-1)}{(S+L)(K-1)}+1
=\tfrac{35(U-1)}{8(4)}+1
=\tfrac{35U-3}{32}.$$ Here, we recall from the proof of Theorem \[thm.new neg ETS with K=4,5\] that such GDDs exist whenever $U\equiv1\bmod 32$, namely whenever $H\equiv1\bmod 35$. Altogether, since $\frac14(35U+1)=\frac14[35(\frac{32H+3}{35})+1]=8H+1$, Theorem \[thm.new ETF\] yields a real ETF of type $(5,-1,8H+1)$ whenever there exists a Hadamard matrix of size $H$ when $H\equiv1\bmod 35$. An infinite number of such Hadamard matrices exist: since $17$ is relatively prime to $140$, Dirichlet’s theorem implies an infinite number of primes $Q\equiv 17\bmod 140$ exist, and each has the property that $Q\equiv 1\bmod 4$, meaning that Paley’s construction yields a real Hadamard matrix of size $2(Q+1)\equiv 36\bmod 280$, in particular of size $2(Q+1)\equiv 1\bmod 35$.
Similarly, a real ${{\operatorname{ETF}}}(7,28)$ exists [@FickusM16], is type $(6,-1,3)$, and there exists a real Hadamard matrix of size $S+L=2$. Letting $M=14$, Theorem \[thm.new ETF\] thus yields a real ETF of type $(6,-1,\frac15(14U+1))$ whenever there exists a $6$-GDD of type $14^U$ where $U$ satisfies and there exists a real Hadamard matrix of size $H=W+1=\frac15(7U-2)$. Here, the necessary conditions and on the existence of such a GDD reduce to having $U\equiv 1,6\bmod 15$, namely to having $H\equiv 1,8\bmod 21$. Since $K=6$, these necessary conditions are not known to be sufficient. Nevertheless, they are asymptotically sufficient: there exists $U_0$ such that for all $U\geq U_0$ with $U\equiv 1,6\bmod 15$, there exists a $6$-GDD of type $14^U$ where $U$ satisfies . As such, there exists $H_0$ such that for all $H\geq H_0$ with $H\equiv 1,8\bmod 21$, if there exists a real Hadamard matrix of size $H$, then there exists a real ETF of type $(6,-1,\tfrac15(14U+1))=(6,-1,2H+1)$. As above, an infinite number of such Hadamard matrices exist: since $\gcd(73,84)=1$, there are an infinite number of primes $Q\equiv 73\bmod 84$, and each has the property that $Q\equiv 1\bmod 4$, meaning Paley’s construction yields a real Hadamard matrix of size $2(Q+1)\equiv 148\bmod 168$, in particular of size $2(Q+1)\equiv 1\bmod 21$.
Applying these same techniques to real ${{\operatorname{ETF}}}(22,176)$ and ${{\operatorname{ETF}}}(23,276)$ yields the infinite families stated in (c) and (d) of the result.
We have now seen that the construction of Theorem \[thm.new ETF\] recovers all Steiner ETFs as a degenerate case, recovers the parameters of “half" of all Tremain ETFs including those of all harmonic ETFs arising from Spence difference sets, and also recovers the parameters of harmonic ETFs arising from the Davis-Jedwab difference sets. This is analogous to how the approach of [@DavisJ97] unifies McFarland, Spence and Davis-Jedwab difference sets with those with parameters . From this perspective, the value of the generalization of [@DavisJ97] given in Theorem \[thm.new ETF\] is that it permits weaker conclusions to be drawn from weaker assumptions: while [@DavisJ97] forms new difference sets (i.e., new harmonic ETFs) by combining given difference sets with building sets formed from a collection of hyperplanes, Theorem \[thm.new ETF\] forms new ETFs by combining given ETFs with GDDs.
In fact, a careful read of [@DavisJ97] indicates that the building sets used there to produce difference sets with parameters are related to $4$-GDDs of type obtained by recursively using ${{\operatorname{TD}}}(4,2^{2j+1})$ to fill the holes of ${{\operatorname{TD}}}(4,2^{2j+3})$ for every $j=1,\dotsc,J-1$ via Lemma \[lem.filling holes\]. Alternatively, such GDDs can be constructed by using Lemma \[lem.Wilson\] to combine a ${{\operatorname{TD}}}(4,8)$ with a ${{\operatorname{BIBD}}}(4^{J-1},4,1)$ arising from an affine geometry. We now generalize these approaches, using Lemmas \[lem.Wilson\] and \[lem.filling holes\] to produce the GDDs needed to apply Theorem \[thm.new ETF\] to several known $K$-negative ETFs with $K>5$:
\[thm.new neg ETS with K>5\] If an ETF of type $(K,-1,S)$ exists where $\frac{K-2}{S-1}\in{\mathbb{Z}}$, and a ${{\operatorname{TD}}}(K,M)$ exists where $M=S(K-1)-1$, then an ETF of type $(K,-1,\tfrac{MU+1}{K-1})$ exists when either:
1. $U=1$ or $U=K$;
2. a ${{\operatorname{BIBD}}}(U,K,1)$ exists;
3. $U=K^J$ for some $J\geq2$, provided a ${{\operatorname{TD}}}(K,MK^j)$ exists for all $j=1,\dotsc,J-1$.
As a consequence, an ETF of type $(K,-1,S)$ exists when either:
1. $K=6$, $S=\frac15(9U+1)$ where either $U=6^J$ for some $J\geq 0$ or a ${{\operatorname{BIBD}}}(U,6,1)$ exists;
2. $K=6$, $S=\frac15(24U+1)$ where either $U=6^J$ for some $J\geq 0$ or a ${{\operatorname{BIBD}}}(U,6,1)$ exists;
3. $K=7$, $S=\tfrac1{6}(35U+1)$ where either $U=7^J$ for some $J\geq 0$ or a ${{\operatorname{BIBD}}}(U,7,1)$ exists;
4. $K=10$, $S=\tfrac1{9}(80U+1)$ where either $U=10^J$ for some $J\geq 0$ or a ${{\operatorname{BIBD}}}(U,10,1)$ exists;
5. $K=12$, $S=\frac1{11}(32U+1)$ when either $U=12^J$ for some $J\geq 0$ or a ${{\operatorname{BIBD}}}(U,12,1)$ exists.
Since and $M=S(K-1)-1$, any $K$-GDD of type $M^U$ can be combined with the given ETF of type $(K,-1,S)$ via Theorem \[thm.new ETF\] in order to construct an ETF of type . (Since $\frac{M+1}{K-1}=S$, such an ETF also exists when $U=1$, namely the given initial ETF.) For instance, the given ${{\operatorname{TD}}}(K,M)$ is a $K$-GDD of type $M^K$, and so such an ETF exists when $U=K$. Other examples of such GDDs can be constructed by combining the given ${{\operatorname{TD}}}(K,M)$ with any ${{\operatorname{BIBD}}}(U,K,1)$—a $K$-GDD of type $1^U$—via Lemma \[lem.Wilson\]. In particular, when $K$ is a prime power, we can construct these BIBDs from affine geometries of order $K$ to produce examples of such GDDs with $U=K^J$ for any $J\geq 2$. GDDs with these parameters also sometimes exist even when $K$ is not a prime power: if a ${{\operatorname{TD}}}(K,MK^j)$ exists for all $j=1,\dotsc,J-1$, then recursively using ${{\operatorname{TD}}}(K,MK^{j-1})$ to fill the holes of ${{\operatorname{TD}}}(K,MK^j)$ for all $j=1,\dotsc,J-1$ via Lemma \[lem.Wilson\] gives a $K$-GDD of type . We now apply these ideas to some known $K$-negative ETFs with $K>5$, organized according to the families of Theorem \[thm.known negative ETFs\]. In particular, ETFs of type $(Q+2,-1,Q+1)$ exist whenever $Q$ is a prime power, and $S-1=Q$ divides $K-2=Q$. For such ETFs, $M=Q(Q+2)$ and a ${{\operatorname{TD}}}(Q+2,Q(Q+2))$ equates to $Q$ MOLS of size $Q(Q+2)$, which is known to occur when $Q=2,3,4,5,8$ [@AbelCD07]. (When $Q$ and $Q+2$ are both prime powers, the standard method [@MacNeish22] only produces $Q-1$ MOLS of size $Q(Q+2)$.) Taking $Q=2,3$ recovers a subset of the ETFs produced in Theorem \[thm.new neg ETS with K=4,5\], and so we focus on $Q=4,5,8$. In particular, for these values of $Q$, the above methods yield an ETF of type when either $U=1$, $U=Q+2$, or a ${{\operatorname{BIBD}}}(U,Q+2,1)$ exists. Moreover, even when $Q+2$ is not a prime power, such an ETF with $U=(Q+2)^J$ exists, provided a ${{\operatorname{TD}}}(Q+2,Q(Q+2)^{j+1})$ exists for all $j=1,\dotsc,J-1$. When $Q=4,8$, this occurs for any $J\geq 2$, since for any $j\geq 1$, the number of MOLS of size $(4)6^{j+1}=2^{j+3}3^{j+1}$ is at least $\min{\{{2^{j+3},3^{j+1}}\}}-1\geq 7$, while the number of MOLS of size $(8)10^{j+1}=2^{j+4}5^{j+1}$ is at least $\min{\{{2^{j+4},5^{j+1}}\}}-1\geq 23$.
Other new explicit infinite families of negative ETFs arise from SIC-POVMs, which by Theorem \[thm.known negative ETFs\] are ETFs of type $(K,-1,S)$ where $K=S(S+1)$. Such ETFs always have the property that $S-1$ divides $K-2=(S-1)(S+2)$, meaning we can apply the above ideas whenever there exists a ${{\operatorname{TD}}}(K,M)$ where $M=(S-1)(S+1)^2$. Such TDs exist when $S=2,3$, being ${{\operatorname{TD}}}(6,9)$ and ${{\operatorname{TD}}}(12,32)$, respectively. When $S=2$ in particular, the above methods yield an ETF of type $(6,-1,\frac15(9U+1))$ when either $U=1$, $U=6$, a ${{\operatorname{BIBD}}}(U,6,1)$ exists, or $U=6^J$ for some $J\geq 2$. This final family arises from the fact that a ${{\operatorname{TD}}}(6,(9)6^j)$ exists for all $j\geq 1$: there are at least $5$ MOLS of size $54=(9)6$, at least $8$ MOLS of size $324=9(6^2)$ (since there are $8$ MOLS of size $9$, and at least $8$ MOLS of size $36$) [@AbelCD07], while for any $j\geq 3$, the number of MOLS of size $(9)6^j=2^j 3^{j+2}$ is at least $2^j-1\geq 7$. Similarly, taking $S=3$ yields an ETF of type $(12,-1,\frac1{11}(32U+1))$ when either $U=1$, $U=12$, a ${{\operatorname{BIBD}}}(U,12,1)$ exists, or $U=12^J$ for some $J\geq 2$: the number of MOLS of size $32(12)=384$ and $32(12)^2=4608$ is at least $15$ [@AbelCD07], while for $j\geq 3$, the number of MOLS of size $32(12)^j=2^{2j+5}3^j$ is at least $\min{\{{2^{2j+5},3^j}\}}-1\geq 26$.
To be clear, we view the ETFs produced in Theorem \[thm.new neg ETS with K>5\] as a “proof of concept," and believe that GDD experts will be able to find many more examples of new ETFs using Theorems \[thm.new ETF\] and \[thm.known negative ETFs\]. For this reason, we have omitted technical cases where an ETF of type $(K,-1,S)$ and a ${{\operatorname{TD}}}(K,M)$ exist where $M=S(K-1)-1$, but $S-1$ does not divide $K-2$, such as when $S=K=9$, and when $S=2K-1$ where $K=6,7,8,9,17,65537$. In such cases, one can still combine the TD with a ${{\operatorname{BIBD}}}(U,K,1)$ via Lemma \[lem.Wilson\] to produce a $K$-GDD of type $M^U$, but is not automatically satisfied.
We also point out that though Theorem \[thm.new ETF\] is a generalization of [@DavisJ97], the new ETFs we have found here are disjoint from those produced by another known generalization of this same work. In particular, the parameters of Davis-Jedwab difference sets can be regarded as the $Q=2$ case of the more general family: $$\label{eq.Davis Jedwab Chen parameters}
D=Q^{2J-1}{({\tfrac{2Q^{2J}+Q-1}{Q+1}})},
\quad
N=4Q^{2J}{({\tfrac{Q^{2J}-1}{Q^2-1}})},$$ where $J\geq 1$ and $Q\geq 2$ are integers. In particular, in [@Chen97], Chen generalizes the theory of [@DavisJ97] in a way that produces difference sets with parameters for any $J\geq 1$ and any $Q$ that is either a power of $3$ or any even power of an odd prime. For this reason, difference sets with parameters of the form are said to be *Davis-Jedwab-Chen difference sets* [@JungnickelPS07]. The inverse Welch bounds for the corresponding harmonic ETFs and their Naimark complements are always integers: $$S
={\bigl[{\tfrac{D(N-1)}{N-D}}\bigr]}^{\frac12}
=\tfrac{2Q^{2J}+Q-1}{Q+1},
\quad
{\bigl[{\tfrac{(N-D)(N-1)}{D}}\bigr]}^{\frac12}
=\tfrac{2Q^{2J}-Q-1}{Q-1}.$$ However, such ETFs are seldom positive or negative. Indeed, a direct computation reveals $$\tfrac{NS}{D(S+1)}
=2+\tfrac{2Q(Q^{2J-2}-1)}{(Q-1)(Q^{2J-1}+1)},
\quad
\tfrac{NS}{D(S-1)}
=2+\tfrac{2}{Q-1}.$$ Thus, by Theorem \[thm.parameter types\], the only such ETFs that are positive or negative are of type $(2,1,2Q-1)$, $(3,-1,\tfrac12(9^J+1))$ or $(4,-1,\tfrac13(2^{2J+1}+1))$, corresponding to the special cases of where $J=1$, $Q=3$ and $Q=2$, respectively. That is, the only overlap between the $K$-negative ETFs constructed in Theorems \[thm.new neg ETS with K=4,5\] and \[thm.new neg ETS with K>5\] and the ETFs constructed in [@Chen97] are those with parameters .
Conclusions
===========
Theorems \[thm.new neg ETS with K=4,5\] and \[thm.new neg ETS with K>5\] make some incremental progress towards resolving the ETF existence problem. By comparing the existing ETF literature, as summarized in [@FickusM16] for example, against Theorems \[thm.known positive ETFs\] and \[thm.known negative ETFs\], we find that every known ETF is either an orthonormal basis, or has the property that either it or its Naimark complement is a regular simplex, has $N=2D$ or $N=2D\pm1$, is a SIC-POVM, arises from a difference set, or is either positive or negative. In particular, every known ${{\operatorname{ETF}}}(D,N)$ with $N>2D>2$ is either a SIC-POVM ($N=D^2$), or is a harmonic ETF, or has $N=2D+1$ where $D$ is odd, or is either positive or negative. This fact, along with the known nonexistence of the ${{\operatorname{ETF}}}(3,8)$ [@Szollosi14], and the available numerical evidence [@TroppDHS05], leads us to make the following conjecture:
\[con.complex\] If $N>2D>2$, an ${{\operatorname{ETF}}}(D,N)$ exists if and only if $N=D^2$ or $\frac{D(D-1)}{N-1}\in{\mathbb{Z}}$ or $(D,N)$ is positive or negative (Definition \[def.ETF types\]).
This is a substantial strengthening of an earlier conjecture made by one of the authors, namely that if $N>D>1$ and an ${{\operatorname{ETF}}}(D,N)$ exists, then one of the three numbers $D$, $N-D$ and $N-1$ necessarily divides the product of the other two. To help resolve Conjecture \[con.complex\], it would in particular be good to know whether an ${{\operatorname{ETF}}}(9,25)$ exists: this is the smallest value of $D$ for which there exists an $N$ such that $N-1$ divides $D(D-1)$ but no known ETF exists. For context, we note that there are numerous pairs $(D,N)$ for which $N-1$ divides $D(D-1)$ and an ${{\operatorname{ETF}}}(D,N)$ is known to exist, despite the fact that a difference set of that size does not exist [@Gordon18], including $$\begin{gathered}
(35,120), (40,105), (45,100), (63,280), (70,231), (77,210), (91,196), (99,540), (130,560),\\
(143,924), (176,561), (187,528), (208,1105), (231,484), (247,780), (260,741).\end{gathered}$$
It would also be good to know whether an ${{\operatorname{ETF}}}(11,33)$ exists, since such an ETF would be type $(4,-1,4)$: as detailed in the previous section, an ETF of type $(K,L,S)$ can only exist when $K$ divides $S(S-L)$, and moreover this necessary condition for existence is known to be sufficient when $L=1$ and $K=1,2,3,4,5$, as well as when $L=-1$ and $K=2,3$.
The evidence also supports an analogous conjecture in the real case. In fact, comparing the relevant literature [@Brouwer17; @FickusM16] against Theorems \[thm.known positive ETFs\] and \[thm.known negative ETFs\], we find that every known real ${{\operatorname{ETF}}}(D,N)$ with $N>2D>2$ is either positive or negative. Moreover, when $1<D<N-1$, $N\neq 2D$ and a real ${{\operatorname{ETF}}}(D,N)$ exists, then it and its Naimark complements’ Welch bounds are necessarily the reciprocals of odd integers [@SustikTDH07]. In particular, any such ETF automatically satisfies one of the two integrality conditions given in Theorem \[thm.parameter types\] that characterize positive and negative ETFs. Conversely, since $S(K-1)+KL$ is odd whenever $S$ is odd, Theorem \[thm.parameter types\] implies that any ETF of type $(K,L,S)$ satisfies the necessary conditions of [@SustikTDH07] when $S$ is odd. That said, real ${{\operatorname{ETF}}}(19,76)$, ${{\operatorname{ETF}}}(20,96)$ and ${{\operatorname{ETF}}}(47,1128)$ do not exist [@AzarijaM15; @AzarijaM16; @FickusM16], despite the fact that $(19,76)$ is type $(5,-1,5)$, $(20,96)$ is both type $(4,1,5)$ and type $(6,-1,5)$, and $(47,1128)$ is both type $(21,1,7)$ and type $(28,-1,7)$. These facts suggest the following analog of Conjecture \[con.complex\]:
If $N>2D>2$ and a real ${{\operatorname{ETF}}}(D,N)$ exists, then $(D,N)$ is positive or negative.
Acknowledgments {#acknowledgments .unnumbered}
===============
The views expressed in this article are those of the authors and do not reflect the official policy or position of the United States Air Force, Department of Defense, or the United States Government. This work was partially supported by the Summer Faculty Fellowship Program of the United States Air Force Research Laboratory.
[WW]{}
R. J. R. Abel, C. J. Colbourn, J. H. Dinitz, Mutually Orthogonal Latin Squares (MOLS), in: C.J. Colbourn, J.H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 160–193.
R. J. R. Abel, M. Greig, BIBDs with small block size, in: C.J. Colbourn, J.H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 72–79.
J. Azarija, T. Marc, There is no (75,32,10,16) strongly regular graph, arXiv:1509.05933.
J. Azarija, T. Marc, There is no (95,40,12,20) strongly regular graph, arXiv:1603.02032.
W. U. Bajwa, R. Calderbank, D. G. Mixon, Two are better than one: fundamental parameters of frame coherence, Appl. Comput. Harmon. Anal. 33 (2012) 58-–78.
A. S. Bandeira, M. Fickus, D. G. Mixon, P. Wong, The road to deterministic matrices with the Restricted Isometry Property, J. Fourier Anal. Appl. 19 (2013) 1123–1149.
A. Barg, A. Glazyrin, K. A. Okoudjou, W.-H. Yu, Finite two-distance tight frames, Linear Algebra Appl. 475 (2015) 163–175.
B. G. Bodmann, H. J. Elwood, Complex equiangular Parseval frames and Seidel matrices containing $p$th roots of unity, Proc. Amer. Math. Soc. 138 (2010) 4387–4404.
B. G. Bodmann, V. I. Paulsen, M. Tomforde, Equiangular tight frames from complex Seidel matrices containing cube roots of unity, Linear Algebra Appl. 430 (2009) 396–417.
C. Bracken, G. McGuire, H. Ward, New quasi-symmetric designs constructed using mutually orthogonal [L]{}atin squares and [H]{}adamard matrices, Des. Codes Cryptogr. 41 (2006) 195–198.
A. E. Brouwer, Strongly regular graphs, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 852-–868.
A. E. Brouwer, Parameters of Strongly Regular Graphs, http://www.win.tue.nl/$\sim$aeb/graphs/srg/
K. I. Chang, An existence theory for group divisible designs, Ph.D. Thesis, The Ohio State University, 1976.
Y. Q. Chen, On the existence of abelian Hadamard difference sets and a new family of difference sets, Finite Fields Appl. 3 (1997) 234–256.
D. Corneil, R. Mathon, eds., Geometry and combinatorics: Selected works of J. J. Seidel, Academic Press, 1991.
G. Coutinho, C. Godsil, H. Shirazi, H. Zhan, Equiangular lines and covers of the complete graph, Linear Algebra Appl. 488 (2016) 264–283.
J. A. Davis, J. Jedwab, A unifying construction for difference sets, J. Combin. Theory Ser. A 80 (1997) 13–78.
C. Ding, T. Feng, A generic construction of complex codebooks meeting the Welch bound, IEEE Trans. Inform. Theory 53 (2007) 4245–4250.
M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, Tremain equiangular tight frames, J. Combin. Theory Ser. A 153 (2018) 54–-66.
M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, Hadamard equiangular tight frames, submitted, arXiv:1703.05353.
M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, C. E. Watson, Equiangular tight frames with centroidal symmetry, to appear in Appl. Comput. Harmon. Anal. M. Fickus, J. Jasper, D. G. Mixon, J. D. Peterson, C. E. Watson, Polyphase equiangular tight frames and abelian generalized quadrangles, to appear in Appl. Comput. Harmon. Anal. M. Fickus, D. G. Mixon, Tables of the existence of equiangular tight frames, arXiv:1504.00253 (2016).
M. Fickus, D. G. Mixon, J. Jasper, Equiangular tight frames from hyperovals, IEEE Trans. Inform. Theory 62 (2016) 5225–5236.
M. Fickus, D. G. Mixon, J. C. Tremain, Steiner equiangular tight frames, Linear Algebra Appl. 436 (2012) 1014–1027.
C. A. Fuchs, M. C. Hoang, B. C. Stacey, The SIC question: history and state of play, Axioms 6 (2017) 21.
G. Ge, Group divisible designs, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 255–260.
C. D. Godsil, Krein covers of complete graphs, Australas. J. Combin. 6 (1992) 245–255.
J. M. Goethals, J. J. Seidel, Strongly regular graphs derived from combinatorial designs, Can. J. Math. 22 (1970) 597–614.
D. Gordon, La Jolla Covering Repository, https://www.ccrwest.org/diffsets.html.
M. Grassl, A. J. Scott, Fibonacci-Lucas SIC-POVMs, J. Math. Phys. 58 (2017) 122201.
R. B. Holmes, V. I. Paulsen, Optimal frames for erasures, Linear Algebra Appl. 377 (2004) 31–51.
J. W. Iverson, J. Jasper, D. G. Mixon, Optimal line packings from nonabelian groups, submitted, arXiv:1609.09836.
J. Jasper, D. G. Mixon, M. Fickus, Kirkman equiangular tight frames and codes, IEEE Trans. Inform. Theory 60 (2014) 170-–181.
D. Jungnickel, A. Pott, K. W. Smith, Difference sets, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 419–435.
J. H. van Lint, J. J. Seidel, Equilateral point sets in elliptic geometry, Indag. Math. 28 (1966) 335–348.
E. R. Lamken, R. M. Wilson, Decompositions of edge-colored complete graphs, J. Combin. Theory Ser. A 89 (2000) 149–200.
P. W. H. Lemmens, J. J. Seidel, Equiangular lines, J. Algebra 24 (1973) 494–512.
R. Mathon, A. Rosa, $2-(v,k,\lambda)$ designs of small order, in: C. J. Colbourn, J. H. Dinitz (Eds.), Handbook of Combinatorial Designs, Second Edition (2007) 25–58.
H. F. MacNeish, Euler Squares, Ann. of Math. 23 (1922) 221–227.
R. L. McFarland, A family of difference sets in non-cyclic groups, J. Combin. Theory Ser. A 15 (1973) 1–10.
G. McGuire, Quasi-symmetric designs and codes meeting the Grey-Rankin bound, J. Combin. Theory Ser. A 78 (1997) 280–291.
Hedvig Mohácsy, The asymptotic existence of group divisible designs of large order with index one, J. Combin. Theory Ser. A 118 (2011) 1915–1924.
J. M. Renes, Equiangular tight frames from Paley tournaments, Linear Algebra Appl. 426 (2007) 497–501.
J. M. Renes, R. Blume-Kohout, A. J. Scott, C. M. Caves, Symmetric informationally complete quantum measurements, J. Math. Phys. 45 (2004) 2171–2180.
J. J. Seidel, A survey of two-graphs, Coll. Int. Teorie Combin., Atti dei Convegni Lincei 17, Roma (1976) 481–511.
E. Spence, A family of difference sets, J. Combin. Theory Ser. A 22 (1977) 103–106.
T. Strohmer, A note on equiangular tight frames, Linear Algebra Appl. 429 (2008) 326–-330.
T. Strohmer, R. W. Heath, Grassmannian frames with applications to coding and communication, Appl. Comput. Harmon. Anal. 14 (2003) 257–275.
M. A. Sustik, J. A. Tropp, I. S. Dhillon, R. W. Heath, On the existence of equiangular tight frames, Linear Algebra Appl. 426 (2007) 619–635.
F. Szöllősi, All complex equiangular tight frames in dimension 3, arXiv:1402.6429.
J. A. Tropp, Complex equiangular tight frames, Proc. SPIE 5914 (2005) 591401/1–11.
J. A. Tropp, I. S. Dhillon, R. W. Heath, Jr., T. Strohmer, Designing structured tight frames via an alternating projection method, IEEE Trans. Inform. Theory 51 (2005) 188–209.
R. J. Turyn, Character sums and difference sets, Pacific J. Math. 15 (1965) 319–346.
S. Waldron, On the construction of equiangular frames from graphs, Linear Algebra Appl. 431 (2009) 2228–2242.
L. R. Welch, Lower bounds on the maximum cross correlation of signals, IEEE Trans. Inform. Theory 20 (1974) 397-–399.
R. M. Wilson, An existence theory for pairwise balanced designs I. Composition theorems and morphisms, J. Combin. Theory Ser. A 13 (1972) 220–245.
P. Xia, S. Zhou, G. B. Giannakis, Achieving the Welch bound with difference sets, IEEE Trans. Inform. Theory 51 (2005) 1900–1907.
G. Zauner, Quantum designs: Foundations of a noncommutative design theory, Ph.D. Thesis, University of Vienna, 1999.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The temperature dependence of the magnetic field distribution in the mixed superconducting phase of randomly oriented MgB$_2$ powder was probed by $%
^{11}$B NMR spectroscopy. Below the temperature of the second critical ($B_{%
{\rm c2}}$) field, $T_{{\rm c2}}\approx 27$K, our spectra reveal two NMR signal components, one mapping the magnetic field distribution in the mixed superconducting state and the other one arising from the normal state. The complementary use of bulk magnetization and NMR measurements reveals that MgB$_2$ is an anisotropic superconductor with a $B_{c2}^c<2.35$ Tesla anisotropy parameter $\gamma\approx 6$.
address: 'Institute of Materials Science, National Center for Scientific Research “Demokritos”, 153 10 Athens, Greece'
author:
- 'G. Papavassiliou, M. Pissas, M. Fardis, M. Karayanni, and C. Christides [@address]'
title: '$^{11}$B NMR detection of the magnetic field distribution in the mixed superconducting state of MgB$_2$. '
---
epsf
[2]{}
The recent discovery of superconductivity [@Akamitsu01] in MgB$_2$ has revived the excitement on this area of research because this alloy becomes superconducting at unexpectedly [@Geballe] high temperatures ($T_c=39$ K), for ”light” main group elements residing between Be and S in the periodic table. Subsequent studies have shown that MgB$_2$ is a type-II superconductor with B$_{{\rm c1}}$($0$)$\approx 0.26$ Tesla, B$_{{\rm c2}}$($%
0$)$\approx 14$ Tesla, with a small condensation energy (relative to Nb$_3$Sn and YBa$_2$Cu$_3$O$_7$), a $\xi _0$$\approx 4.9$ nm, and a $\lambda _0$$%
\approx 185$ nm [@Wang01]. However, opinions about the nature of the superconductivity mechanism are still contradictory. Band structure calculations [@Kortus01] suggest that MgB$_2$ is a BCS superconductor, where superconductivity results from in-plane electron-phonon coupling on the boron sublattice. The detected isotope effect [@Bud'ko01] and the BCS-type energy gap, that is obtained by tunneling spectroscopy [@Bollinger01] and $^{11}$B NMR [@Kotegawa01; @Gerashenko01; @Tou01], are in support of this model. However, specific heat measurements indicate that the superconducting gap is either anisotropic or two-band-like [@Wang01]. Further deviations from the $s$-wave model have been detected on the temperature dependence of $B_{{\rm c1}}$ and $\lambda $ [@Li01] as well.
Recently few works[@Lima01; @0105271; @0105545; @Simon01; @0106577; @Patnaik] presented convincing evidence that the MgB$_2$ is an anisotropic superconductor with an anisotropy parameter $\gamma=(B_{{\rm c2}}^{{\rm ab}}/B_{{\rm c2}}^c)$ taking values into interval $2\leq\gamma \leq 6$.
In principle, a strong anisotropy in the mixed superconducting state of powder MgB$_2$ - if present- should be detectable with $^{11}$B NMR spectroscopy. As known, for fields $B_{c1}<B_0<B_{c2}$ a vortex lattice is formed that gives rise to a characteristic magnetic field distribution with van Hove singularities at fields where $\nabla \cdot B=0$. For a perfect hexagonal vortex lattice the field distribution exhibits a peak at a value $%
B_s$, which corresponds to the saddle point located midway between two vortices, whereas two steps at the maximum ($B_{{\rm max}}$) and minimum ($%
B_{{\rm min}}$) fields are expected [@Fite66; @Brandt]. Such a magnetic field distribution should be mapped on the NMR line shape, as the Larmor frequency of the resonating nuclei depends linearly on the local magnetic field. Succesfull mapping of the magnetic field distribution has been already presented in a variety of superconducting materials, like vanadium [@Fite66], and rare earth nickel borocarbides YNi$_2$B$_2$C with $^{11}$B NMR [@Lee00]. In case of strong anisotropy and in applied field $B_{%
{\rm c2}}^c<B_0<B_{{\rm c2}}^{ab}$, a powder superconducting sample with randomly oriented grains is expected to give a superposition of magnetic field distributions, ranging in between the normal state and the Abrikosov lattice. It is thus of particular interest to testify the possibility of anisotropy in the mixed superconducting state of MgB$_2$, by performing $^{11}$B NMR experiments in magnetic fields fullfilling the above condition.
In this letter we report $^{11}$B NMR line shape measurements on powder MgB$%
_2$ samples in external magnetic fields $B_0=2.35$ and $4.7$ Tesla, which exhibit a strong asymmetric low frequency broadening at temperatures lower than the temperature of the second critical field ($T$$<$$T_{{\rm c2}}$). The low temperature spectra may be decomposed in two components: One component corresponding to the unshifted NMR signal of the normal state, which indicates that a portion of the sample volume remains in the normal state even at the lowest measured temperature of $5$ K. A second component, which maps the magnetic field distribution of the vortex lattice in the mixed state (vortex component). Our data provide clear evidence that for a part of the MgB$_2$ grains the B$_{{\rm c2}}$($0$)$<2.35$ Tesla, whereas for another part of the grains B$_{{\rm c2}}$($0$) is sufficiently higher, and according to magnetic and CESR measurements [@Simon01] B$_{{\rm c2}}^{ab}$(0)$\approx 14$ Tesla. Thus, our results provide a strong experimental evidence that MgB$_2$ is an anisotropic superconductor with an anisotropy parameter $\gamma\approx 6$.
High quality MgB$_2$ powder samples were prepared by liquid-vapor to solid reaction in an alumina crucible placed inside a vacuum sealed silica tube, using a 3% excess of Mg. Pure Mg and B powders were thoroughly mixed and subsequently slowly heated up to $910^{\rm o}$ C. At this temperature the sample annealed for two hours and then cooled slowly down to room temperature. Rietveld refinement of x-ray powder diffraction spectra revealed that the examined sample consists of 95% MgB$_2$, with cell constants equal to $a=b=3.0849(1)$ Å and $c=3.5213(1)$ Å, and a secondary phase of 5% MgO.
To investigate the mixed state and the magnetic irreversibility of MgB$_2$ we have employed thermomagnetic and isothermal magnetization measurements on a SQUID magnetometer. The upper panel of Figure \[fig1\]a shows the zero field (ZFC) and field cooling (FC) magnetization curves for a sample with randomly oriented grains, which was also used in the NMR measurements. A marked feature is the observed curvature near the onset of the transition, which has been recently attributed [@Simon01] to the anisotropy of MgB$_2$. Specifically, the onset of the diamagnetic signal occurs at the B$_{{\rm c2}}^{{\rm ab}}$ which varies with temperature as $B_{{\rm c2}}^{{\rm ab}}(T)=B_{{\rm c2}}^{{\rm ab}}(0)
(1-T/T_c)^{1.27}$. Since we examine a powder sample containing randomly oriented grains, we have applied the equation derived by Simon [*et al*]{} [@Simon01] for a uniaxial superconductor with the magnetization lying parallel to the external field, in order to fit the reversible part of the thermomagnetic curves. The solid line in Figure \[fig1\]a shows the successful reproduction of the experimental data by using an anisotropy ratio $\gamma $$\sim 6$ and an average penetration depth: $\lambda =(\lambda _{{\rm ab}}^2\lambda _c)^{1/3}\approx 170$ nm, in agreement with the values in Ref.[@Simon01]. The lower panel of Figure \[fig1\]a shows the phase diagram that is obtained from the magnetic measurements. This diagram includes the temperature variation of $B_{{\rm c2}}^c$ which is estimated from $B_{{\rm c2}}^c=B_{{\rm c2
}}^{{\rm ab}}/\gamma$. In this figure we also include the data of $H_{c2}^{ab}$ for $H>5.5$ T, from ref. . The irreversibility line is derived from the temperature where the ZFC and the FC branches are separated. Finally, Figure \[fig1\]b shows the half of the isothermal hysteresis loop at $T=5$ K, which indicates that the irreversible magnetization is comparable to the reversible one. This small irreversible magnetization, in the powder sample has been attributed to the surface barriers [@Pissas]. The two insets show details of the loop in the regions of 2.35 and 4.7 Tesla, which are used below to reproduce the magnetic field distribution in the vortex lattice.
$^{11}$B NMR line shape measurements were performed on two spectrometers operating in external magnetic fields with $B_0=2.35$ and $4.7$ Tesla. The spectra were obtained from the Fourier transformation of half of the echo, following a typical $\pi /2$-$\tau $-$\pi $ spin-echo pulse sequence. In both magnetic fields and in the normal state the length of the $\pi /2$ pulse was $t_p(\pi /2)\leq 2$ $\mu $sec, corresponding to a radiofrequency (rf) irradiation field $B_1\geq 44$ Gauss. At room temperature the spectra were found to exhibit the typical powder pattern for a nuclear spin $I=3/2$ in the presence of quadrupolar effects with an axially symmetric electric field gradient, in agreement with previous works [@Gerashenko01; @Jung01]. The separation of the symmetric satellite lines gives a quadrupolar frequency [@Cohen57]: $\nu _Q=2\Delta \nu ^{(1)}=e^2qQ/2h\approx 0.836$ MHz. In Figure \[fig2\] we demonstrate the line shape of the central transition (-$\frac 12\longrightarrow \frac 12$) as a function of temperature in a magnetic field of $4.7$ Tesla. The line shape in the normal state is temperature independent and consistent with previous studies [@Gerashenko01; @Tou01; @Jung01]. In addition, a second order quadrupolar split of $\approx 6$ kHz is clearly observed. Considering that for $I=3/2$ the second order quadrupolar split is given by [@Cohen57]: $\Delta \nu
^{(2)}=(25\nu _Q^2/144\nu _L)[I(I+1)-3/4]$, we obtain a $\nu _Q\approx 0.860$ MHz, in agreement with the value obtained from the separation of the satellite peaks.
Below $T_{{\rm c2}}$ ($\approx 27$ K in $4.7$ Tesla) the spectra start to broaden and under the irreversibility temperature an extra signal shows up as a pronounced shoulder in the low frequency part of the spectrum. It is worth noting that such an extra feature was not mentioned in previous NMR studies [@Gerashenko01; @Jung01]. By further decreasing temperature the intensity of this shoulder increases and its location shifts to lower frequencies. This behavior is explicit to the magnetic field distribution in the mixed state, as implied by the dotted lines.
A significant part of the signal remains unshifted at the frequency of the normal state NMR signal. The unshifted part of the line shape may be explained if we consider a distribution of $B_{{\rm c2}}$ caused predominately by anisotropy and not by inhomogeneities, as the observed superconducting transition is extremely sharp. Assuming that the upper critical field depends on the angle of the $c-
$axis in each crystallite then the angular dependence of the second critical field is given by $B_{{\rm c2}}(\theta )=B_{{\rm c2}}^{{\rm ab}}(1+(\gamma
^2-1)\cos ^2\theta )^{-1/2}$. This equation shows that only crystallites with $B_{{\rm c2}}(\theta )>B_{{\rm o}}$ would give a characteristic signal of a type-II superconductor in the mixed state. If for a part of the grains $%
B_{{\rm o}}>B_{{\rm c2}}(\theta )$ then a NMR line shape would be observed, which is the sum of spectra coming from crystallites in the normal state and in the mixed state. Indeed the spectra under $B_{{\rm o}}=4.7$ Tesla show an unshifted component which comes from the normal part of the sample, revealing the presence of strong anisotropy.
The two signal components are more clearly resolved in Figure \[fig3\], which shows $^{11}$B NMR line shapes of the central transition in a magnetic field $B_0=2.35$ Tesla. In this field the low frequency shoulder becomes broader and shifts to lower frequencies in comparison to the spectra taken in $4.7$ Tesla. This is expected if we consider that the magnetic field distribution in the vortex lattice becomes less dense and exhibits stronger field gradients in lower external magnetic fields. Since a significant part of the signal intensity remains unshifted, then it can be argued that for a part of the grains the $B_{c2}^c$ must be lower than $2.35$ Tesla. Considering that magnetization measurements give $B_{c2}\approx 14$ Tesla, and by assuming that $B_{c2}^{ab}>B_{c2}^c$ [@Lima01], [@Simon01] an anisotropy parameter $\gamma \geq 6$ is estimated.
Following Ref. , a rough estimation of the singular points (at London limit $\lambda>>\xi$) of the field distribution at $T=5$ K and $H=4.7$ Tesla, gives: $B_{\max }-\left\langle B\right\rangle\approx 1.6\times 0.551 \Phi_{\rm o}/4\pi\lambda^2 \approx 50$ Gauss, $B_{\min }-\left\langle B\right\rangle\approx -0.6\times 0.551 \Phi_{\rm o}/4\pi\lambda^2 \approx -18$ Gauss and $B_s-\left\langle B\right\rangle\approx -0.5\times 0.551 \Phi_{\rm o}/4\pi\lambda^2 \approx -15$ Gauss. In the above estimations we used the experimental value of the magnetic induction at $T=5$ K, $\left\langle B\right\rangle=4\pi M_{rev}+H_{{\rm o}}\approx -14+H_{\rm o}$ Gauss (see inset of Figure \[fig1\]b) and $\lambda=170$ nm.
The estimated values of $B_{\max }$, $B_{\min}$ and $B_s$ are in good agreement with the characteristic points (see arrows in Fig. \[fig2\]) of the NMR spectrum at $T=5$ K. [*These results indicate that the low frequency shoulder of the NMR spectra is produced by the magnetic field distribution of the vortex lattice*]{}.
The extension of the low field tail below the theoretical $B_{min}$, and the slightly shifted broad maximum in the experimental magnetic field distribution might originate from small random pertubations by FL pinning or by structural defects in the FLL, which lead to smearing of the ideal field distribution. Deviation from the ideal distribution may be also produced by shearing of the FLL, which leads to splitting of the singularity at $B_s$ and to the jump at $B_{\min }$, but leaves $B_{\max }$ approximately unchanged [@Brandt]. Smearing from random shearing would dominate, if long-wavelength compression (flux density gradient) is not present, in particular for small or large $B$ values, where the shear modulus $c_{66}$ $%
\longrightarrow 0$ for both $B/B_{c2}\longrightarrow 0$ and $%
B/B_{c2}\longrightarrow 1$ . On the other hand, the magnetic field distribution is much more sensitive to fluctuations of the FLL density (random compression) than to shear deformations for $\left\langle
B\right\rangle >>4\pi M$, if energetically favorable [@Comment]. In such a case, even a small homogeneous compression of the FLL will shift rigidly the magnetic field distribution density [@Brandt], thus leading to smearing of all van Hove singularities. Finally, it should also be noted that in anisotropic superconductors the field distribution deviates from that of an isotropic hexagonal vortex lattice [@Thie89].
In conclusion, $^{11}$B NMR line shape measurements on powder MgB$_2$ show up two NMR signal components for $T<T_{c2}$: one coming from the vortex lattice from grains in the mixed superconducting state, and another one from grains that are still in the normal state as their orientation in respect to the external magnetic field is such that $B_o>B_{{\rm c2}}(\theta )$. Our measurements suggest a high anisotropy for the upper critical field with $%
\gamma \geq 6$, in agreement with recent CESR measurements [@Simon01]. This experimental result changes the balance in favor of anisotropic superconductivity in MgB$_2$, and should be taken into consideration in theories trying to explain the superconducting state in this material.
Permanent address, Department of Engineering Sciences, School of Engineering, University of Patras, 26110 Patras, Greece. J. Nagamatsu [*et al.*]{}, Nature(London) [**410**]{}, 63 (2001).
T. H. Geballe, Science [**293**]{}, 223 (2001).
Y. Wang, T. Plackowski, and A. Junod, Physica (Amsterdam) [**355C**]{}, 179 (2001).
J. Kortus [*et al.*]{}, Phys. Rev. Lett. [**86**]{}, 4656 (2001).
S. L. Bud’ko [*et al*]{}, Phys. Rev. Lett. [**86**]{}, 1877 (2001); Phys. Rev. B [**63**]{}, 220503 (2001).
G. Rubio-Bollinger, H. Suderow, and S. Vieira, cond-mat/0102242 (2001); G. Karapetrov [*et al.*]{}, cond-mat/0102312 (2001).; A. Sharoni, I. Felner, and O. Millo, cond-mat/0102325 (2001).
H. Kotegawa [*et al.*]{}, cond-mat/0102334 (2001).
A. Gerashenko, K. Mikhalev, S. Verkhovskii, cond-mat/0102421 (2001).
H. Tou [*et al.*]{}, cond-mat/0103484 (2001).
S. L. Li [*et al.*]{}, cond-mat/0103032 (2001).
O. F. de Lima [*et al.*]{}, Phys. Rev. Lett. [**86**]{},5974 (2001). M. Xu, H. Kitazawa, Y. Takano, J. Ye, K. Nishida, H. Abe, A. Matsushita and G. Kido,, cond-mat/0105271 (2001). S. Lee, H. Mori, T Masui, Yu. Eltsev, A. Yamamoto and S. Tajima, cond-mat/0105545 (2001). F. Simon [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 047002 (2001). S. L. Bud’ko, V. G. Kogan and P. C. Canfield, cond-mat/0106577 (2001). S. Patnaik [*et al.*]{}, Supercond. Sci. Technol. [**14**]{}, 315 (2001). W. Fite, and A. G. Redfield, Phys. Rev. Lett. [**17**]{}, 381 (1966).
E. H. Brandt, Journal of Low Temp. Phys. [**73**]{}, 355 (1988); E. H. Brandt and A. Seeger, Advances in Physics [**35**]{}, 189 (1986); E. H. Brandt, Phys. Rev. Lett. [**66**]{}, 3213 (1991).
K.H. Lee [*et al.*]{}, Phys. Rev. B [**62**]{}, 123 (2000). M. Pissas [*et al.*]{}, unpublished. J. K. Jung [*et al.*]{}, Phys. Rev. B [**64**]{}, 012514 (2001).
M. H. Cohen and F. Reif, in Solid State Physics: Advances in Research and Applications, edited by F. Seitz and D. Turnbull (Academic Press, New York, 1957), Vol. 5. Although we need higher energy to compress the FLL than to shear it, the field distribution will be dominated by random compressions when shear and compressional deformations will be of equal energy [@Brandt]. S. L. Thiemann, Z. Radovic, and V. G. Kogan, Phys. Rev. B [**39**]{}, 11406 (1989).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We explicitly compute the evolution of the density operator of a two-mode electromagnetic field when the influence of the thermal fluctuation of the vacuum is common for both modes. From this result, we give an example in which the bundle of quantum noisy channels turns out to be noiseless for the special type of signal states due to the correlation.'
author:
- Tohya Hiroshima
- Osamu Hirota
title: 'Continuous variable noise-free states in correlated quantum noisy channels'
---
Computing the classical capacity of quantum channels is one of the most important issues in quantum information theory [@NC]. The quantum channels are represented by a stochastic or completely positive map that maps the density operator $\pi $ to the density operator $\sigma $; $\pi \mapsto \sigma =\Phi (\pi )$ [@Kraus]. If the channels $\Phi _{1}$ and $\Phi _{2}$ have no correlation, the bundle of them is simply given by $\Phi _{1}\otimes \Phi _{2}$. Even in this simple case, whether the classical capacity of quantum channels is additive or not is still a long-standing open question. On the other hand, the correlation or memory effect between channels leads to interesting phenomena. One such example was given in [@MP], where the entangled signals enhance the mutual information of the pair of depolarizing channels with memory. Without correlation, the depolarizing channels are proven to be additive [@King] and explicit calculations have shown that the input of entangled qubits cannot increase the classical capacity of depolarizing channels [@BFMP]. Similar results for channels with memory have been obtained in [@MPV; @BDB; @YS]. Asymptotic classical capacity of quantum channels with memory has been also analyzed [@BM].
In this paper, we give an example in which the bundle of quantum noisy channels turns out to be noiseless for the special type of signal states due to the correlation. This constitutes one example of the noiseless codes [@PSE; @KBLW] in a continuous variable system.
Consider a two-mode electromagnetic field with the same frequency $\omega _{0}$ interacting with the reservoir: the thermal fluctuation of the vacuum. If these two modes are spatially well separated, the reservoir field acts independently on each mode. The interaction Hamiltonian takes the following form. $$\label{eq:Interaction1}
V=\hbar \sum_{i=1}^{2}a_{i}^{\dagger }\sum_{j}g_{j}^{(i)}b_{j}^{(i)}+h.c..$$ Here, $a_{i}^{\dagger }$ denotes the creation operator of the electromagnetic field of mode $i$ and $b_{j}^{(i)}$ denotes the annihilation operator of the electromagnetic field of reservoir $i$ acting on mode $i$. They obey the commutation relation $\left[ a_{i},a_{j}^{\dagger }\right] _{-}=\delta _{ij}$ and $\left[ b_{k}^{(i)},b_{l}^{(j)\dagger }\right] _{-}=\delta _{ij}\delta _{kl}$. Other commutators are zero. $g_{j}^{(i)}$ is the coupling constant. This independent reservoir model has been well investigated [@JLK; @Hiroshima; @SW01; @WJK]. However, if the two modes are close enough spatially, the action of noise on each mode must be more or less correlated. More precisely, the $i$th reservoir field $b^{(i)}$ acts partially on mode $j(\neq i)$. Therefore, $b_{j}^{(i)}$ in Eq. (\[eq:Interaction1\]) should be replaced by $\sqrt{1-\mu }b_{j}^{(i)}+\sqrt{\mu }b_{j}$ with $\left[ b_{i},b_{j}^{\dagger }\right] _{-}=\delta _{ij}$. Here, $b_{j}$ is the annihilation operator of the common reservoir that is independent of reservoirs 1 and 2 and the parameter $\mu$ ($0\leq \mu \leq 1$) represents the degree of correlation. Under the extreme condition of $\mu \rightarrow 1$, the interaction Hamiltonian takes the form $$\label{eq:Interaction2}
V=\hbar \left( \sum_{i=1}^{2}a_{i}^{\dagger }\right) \sum_{j}g_{j}b_{i}+h.c.,$$ where $g_{j}=g_{j}^{(1)}=g_{j}^{(2)}$. Namely, the reservoir acts on both modes simultaneously. The same model has been applied to the two-mode squeezed vacuum states [@P-B]. This ideal model may be hard to achieve experimentally, but it provides the essential feature of these correlated channels of intermediate value of $\mu$ ($0 < \mu < 1$). Hereafter, we confine ourselves to this extreme case.
Following the standard procedure for deriving the master equation for the density operator under the Markov approximation in the weak coupling limit [@WM], we get the following equation of motion for the two-mode density operator in the interaction picture. $$\begin{aligned}
\label{eq:MasterEquation}
\frac{d}{dt}\rho (t) &=&\frac{\Gamma }{2}(N_{0}+1)\left[ 2(a_{1}+a_{2})\rho
(t)(a_{1}^{\dagger }+a_{2}^{\dagger })-(a_{1}^{\dagger }+a_{2}^{\dagger
})(a_{1}+a_{2})\rho (t)-\rho (t)(a_{1}^{\dagger }+a_{2}^{\dagger
})(a_{1}+a_{2})\right] \nonumber \\
&&+\frac{\Gamma }{2}N_{0}\left[ 2(a_{1}^{\dagger }+a_{2}^{\dagger })\rho
(t)(a_{1}+a_{2})-(a_{1}+a_{2})(a_{1}^{\dagger }+a_{2}^{\dagger })\rho
(t)-\rho (t)(a_{1}+a_{2})(a_{1}^{\dagger }+a_{2}^{\dagger })\right] ,\end{aligned}$$ where $N_{0}=1/\left( e^{\hbar \omega _{0}/kT}-1\right) $ is the mean number of reservoir quanta at frequency $\omega _{0} $. The initial ($t=0$) density operator can be expanded in terms of coherent states as follows. $$\label{eq:InitialRho}
\rho (0) =\int \cdots \int \prod_{i=1}^{2}d^{2}\alpha _{i}d^{2}\beta _{i}
P\left( \alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right)
\Lambda \left( \alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right),$$ where $$\int \cdots \int \prod_{i=1}^{2}d^{2}\alpha _{i}d^{2}\beta _{i}P\left(
\alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right) =1,$$ and $$\Lambda \left( \alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right) =
\frac{\left| \alpha _{1},\alpha _{2}\right\rangle
\left\langle \beta _{1},\beta _{2}\right| }{\left\langle \beta _{1}\right|
\left. \alpha _{1}\right\rangle \left\langle \beta _{2}\right| \left. \alpha _{2}\right\rangle }
=\prod\limits_{i=1}^{2}\exp \left( -\alpha _{i}\beta _{i}^{*}\right)
\exp(\alpha _{i}a_{i}^{\dagger })\left| 0\right\rangle \left\langle 0\right|
\exp (\beta _{i}^{*}a_{i})$$ is an analytic function of $\alpha _{1}$, $\alpha _{2}$, $\beta _{1}^{*}$, and $\beta_{2}^{*}$. $\rho (t)$ is the output state for the input state $\rho (0)$ in the channel with correlation considered here, $\Phi _{c}$; $\rho (t)=\Phi _{c}\left[ \rho (0)\right] $.
In order to solve Eq. (\[eq:MasterEquation\]) under the initial condition (\[eq:InitialRho\]), we first calculate the two-mode characteristic function defined as $$\label{eq:Characteristic}
\chi (\lambda _{1},\lambda _{2};t)={\rm Tr}\left[ \prod_{i=1}^{2}\exp
\left( \lambda _{i}a_{i}^{\dagger }-\lambda _{i}^{+}a_{i}\right) \rho
(t)\right].$$ Following the standard procedure [@WM], we get the equation of motion for $\chi (\lambda _{1},\lambda _{2};t)$ from Eqs. (\[eq:MasterEquation\]) and (\[eq:Characteristic\]) as $$\label{eq:CharacteristicEquation}
\frac{d}{dt}\chi (\lambda _{1},\lambda _{2};t)=
-\Gamma \left( N_{0}+\frac{1}{2}\right) \left| \lambda _{+}\right|^{2}
\chi (\lambda _{1},\lambda _{2};t)
-\Gamma \left( \lambda _{+}^{*}\frac{\partial }{\partial \lambda _{+}^{*}}
+\lambda _{+}\frac{\partial }{\partial \lambda _{+}}\right)
\chi (\lambda _{1},\lambda _{2};t),$$ where $\lambda _{\pm }=\lambda _{1}\pm \lambda _{2}$. From Eqs. (\[eq:InitialRho\]) and (\[eq:Characteristic\]), the initial value of $\chi (\lambda _{1},\lambda _{2};t)$ is calculated as $$\label{eq:CharacteristicInitial}
\chi (\lambda _{1},\lambda _{2};0) =\int \cdots \int
\prod_{i=1}^{2}d^{2}\alpha _{i}d^{2}\beta _{i}
P\left( \alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right)
\prod\limits_{i=1}^{2}\exp
\left( -\frac{\left| \lambda _{i}\right|^{2}}{2}-\lambda _{i}^{*}\alpha _{i}+\lambda _{i}\beta _{i}^{*}\right).$$ Following the operator method described in [@Collett], we can solve Eq. (\[eq:CharacteristicEquation\]) under the initial condition (\[eq:CharacteristicInitial\]). The straightforward calculations yield $$\begin{aligned}
&&\chi (\lambda _{1},\lambda _{2};t)=\int \cdots \int
\prod_{i=1}^{2}d^{2}\alpha _{i}d^{2}\beta _{i}P\left( \alpha _{1},\alpha
_{2};\beta _{1}^{*},\beta _{2}^{*}\right) \nonumber \\
&&\times \exp \left[ -\frac{1}{2}N(t)\left| \lambda _{+}\right| ^{2}-\frac{1%
}{4}\left| \lambda _{+}\right| ^{2}-\frac{1}{4}\left| \lambda _{-}\right|
^{2}-\frac{1}{2}e^{-\Gamma t}\left( \alpha _{+}\lambda _{+}^{*}-\beta
_{+}^{*}\lambda _{+}\right) -\frac{1}{2}\left( \alpha _{-}\lambda
_{-}^{*}-\beta _{-}^{*}\lambda _{-}\right) \right] ,\end{aligned}$$ with $\alpha _{\pm }=\alpha _{1}\pm \alpha _{2}$, $\beta _{\pm }=\beta _{1}\pm \beta _{2}$, and $N(t)=N_{0}(1-e^{-2\Gamma t})$. It is an easy task to check that $\chi (\lambda _{1},\lambda _{2};t)$ obtained above satisfies Eqs. (\[eq:CharacteristicEquation\]) and (\[eq:CharacteristicInitial\]).
In the following, we calculate the two-mode Q-function from $\chi (\lambda _{1},\lambda _{2};t)$ and then calculate the density operator from the two-mode Q-function, which is defined as the diagonal element of the density operator in a coherent state; $ Q(\delta _{1},\delta _{2};t)=
\left\langle \delta _{1},\delta _{2}\right| \rho (t)\left| \delta _{1},\delta _{2}\right\rangle
/ \pi ^{2}$. This is also calculated from the two-mode anti-normally ordered characteristic function $$\chi ^{A}(\lambda _{1},\lambda _{2};t)={\rm Tr}\left[ \prod_{i=1}^{2}\exp
(-\lambda _{i}^{*}a_{i})\exp (\lambda _{i}a_{i}^{\dagger })\rho (t)\right]$$ via the following integral transformation $$Q(\delta _{1},\delta _{2};t) =
\frac{1}{\pi ^{4}}\int \int \prod_{i=1}^{2}d^{2}\lambda _{i}
\exp \left( \delta _{i}\lambda _{i}^{*}-\delta _{i}^{*}\lambda _{i}\right)
\chi ^{A}(\lambda _{1},\lambda _{2};t).$$ Since $\chi ^{A}(\lambda _{1},\lambda _{2};t)=e^{-\left| \lambda _{1}\right|
^{2}/2-\left| \lambda _{2}\right| ^{2}/2}\chi (\lambda _{1},\lambda _{2};t)$, we can calculate the Q-function by using the identity [@WM; @Puri] $$\label{eq:Identity}
\frac{1}{\pi }\int d^{2}\eta \exp \left( -\lambda \left| \eta \right|
^{2}+\mu \eta +\nu \eta ^{*}\right) =\frac{1}{\lambda }\exp \left( \frac{\mu
\nu }{\lambda }\right),$$ which holds for $\mathrm{{Re}(\lambda ) > 0}$ and arbitrary $\mu$ and $\nu$. The result is $$\label{eq:Qfunction}
Q(\delta _{1},\delta _{2};t)=\frac{1}{\pi ^{2}}\frac{1}{1+N(t)}\int \cdots
\int \prod_{i=1}^{2}d^{2}\alpha _{i}d^{2}\beta _{i}P\left( \alpha
_{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right) \exp \left[ \frac{1}{2}%
f(\alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*};\delta _{1},\delta
_{2})\right] ,$$ where $$\begin{aligned}
\label{eq:QfunctionExponential}
&&f(\alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*};\delta _{1},\delta _{2}) \nonumber \\
&=&-\frac{2+N(t)}{1+N(t)}
\left( \left| \delta _{1}\right|^{2}+\left| \delta _{2}\right| ^{2}\right)
+\frac{N(t)}{1+N(t)}\left( \delta _{1}\delta _{2}^{*}+\delta _{1}^{*}\delta _{2}\right)
-\frac{1}{1+N(t)}\alpha _{+}\beta _{+}^{*}e^{-2\Gamma t}
-\alpha _{-}\beta _{-}^{*} \nonumber \\
&&+\left( \frac{\alpha _{+}e^{-\Gamma t}}{1+N(t)}+\alpha _{-}\right) \delta _{1}^{*}
+\left( \frac{\beta _{+}^{*}e^{-\Gamma t}}{1+N(t)}+\beta _{-}^{*}\right) \delta _{1}
+\left( \frac{\alpha _{+}e^{-\Gamma t}}{1+N(t)}-\alpha _{-}\right) \delta _{2}^{*}
+\left( \frac{\beta _{+}^{*}e^{-\Gamma t}}{1+N(t)}-\beta _{-}^{*}\right) \delta _{2}.\end{aligned}$$ Now, we get the density operator from the following formula. $$\label{eq:DensityMatrix}
\rho (t) =\left( \frac{1}{4\pi }\right) ^{2}\int \cdots \int
\prod_{i=1}^{2}d^{2}\delta _{i}d^{2}\gamma _{i}\exp \left( -\frac{1}{4}%
\left| \delta _{i}-\gamma _{i}\right| ^{2}\right)
\Lambda \left( \delta _{1},\delta _{2};\gamma _{1}^{*},\gamma
_{2}^{*}\right)
Q\left( \frac{1}{2}\left( \delta _{1}+\gamma _{1}\right) ,\frac{1}{2%
}\left( \delta _{2}+\gamma _{2}\right) ;t\right).$$ This is an obvious extension of the relation [@DG] $$\label{eq:Formula}
\rho =\frac{1}{\pi }\int \int d^{2}\gamma d^{2}\delta \Lambda (\gamma
+\delta ,\gamma -\delta )e^{-\left\vert \delta \right\vert ^{2}}Q(\gamma )$$ to the two-mode case. In Eq. (\[eq:Formula\]), $\Lambda (\alpha ,\beta )=\left| \alpha \right\rangle \left\langle \beta
^{*}\right| /\left\langle \beta ^{*}\right| \left. \alpha \right\rangle $ and $Q(\gamma )=\left\langle \gamma \right| \rho \left| \gamma \right\rangle /\pi $ is the single-mode Q-function.
At zero temperature, i.e., $N_{0}=0$, the integral of Eq. (\[eq:DensityMatrix\]), can be explicitly performed by using the identity $$\frac{1}{\pi }\int d^{2}\alpha \exp \left( -\left| \alpha \right| ^{2}+\beta
\alpha ^{*}\right) f(\alpha )=f(\beta )$$ for an arbitrary analytic function $f$. The result is $$\rho (t) =\int \cdots \int \prod_{i=1}^{2}d^{2}\alpha _{i}d^{2}\beta
_{i}P\left( \alpha _{1},\alpha _{2};\beta _{1}^{*},\beta _{2}^{*}\right)
\Lambda \left( \alpha _{+}(t),\alpha _{-}(t);\beta
_{+}^{*}(t),\beta _{-}^{*}(t)\right),$$ where $\alpha _{\pm }(t)=(\alpha _{+}e^{-\Gamma t}\pm \alpha _{-})/2$ and $\beta _{\pm }(t)=(\beta _{+}e^{-\Gamma t}\pm \beta _{-})/2$. It is quite easy to see that $\rho (t)$ thus obtained satisfies the master equation (\[eq:MasterEquation\]) with $N_{0}=0$. It should be noted that the initial pure state $\rho (0)=\left| \alpha _{1},\alpha _{2}\right\rangle \left\langle \alpha
_{1},\alpha _{2}\right| $ remains pure; it becomes $\rho (t)=\Phi _{c}\left[ \rho (0)\right]=\left| \alpha _{+}(t),\alpha _{-}(t)\right\rangle \left\langle \alpha _{+}(t),\alpha _{-}(t)\right| $ as the output of the channel $\Phi _{c}$. More interestingly, the state $\left| \alpha ,-\alpha \right\rangle \left\langle \alpha ,-\alpha
\right| $ is not affected by the noise of the channel $\Phi _{c}$. This is also true for the entangled coherent state [@Sanders; @vH] of the form $\left( \left| \alpha ,-\alpha \right\rangle \pm e^{i\phi }\left| -\alpha ,\alpha \right\rangle \right) /\sqrt{2}$. These perfectly decoherence-free states arise from the quantum interference coming from the indistinguishability of modes 1 and 2 for the common reservoir. For the input state $\rho (0)=\left| \alpha ,-\alpha \right\rangle \left\langle \alpha ,-\alpha
\right| $, the quantum interference is completely constructive while for $\rho (0)=\left| \alpha ,\alpha \right\rangle \left\langle \alpha ,\alpha \right| $, it is destructive.
At finite temperatures $T>0$, the input pure state does not keep its purity during the time evolution: the output state is mixed even if the input state is pure. The output purity ${\rm Tr}\rho ^{2}(t)$ is calculated from Eq. (\[eq:DensityMatrix\]) with Eqs. (\[eq:Qfunction\]) and (\[eq:QfunctionExponential\]) by using identity (\[eq:Identity\]). The explicit calculations yield $${\rm Tr}\rho ^{2}(t)=\frac{2\left( 1+N(t)\right) }{2+6N(t)+5N(t)^{2}}.$$ This is a monotonically decreasing function of $t$. Another measure for characterizing the channel is the fidelity that quantifies how close the output state $\rho (t)$ and the input state $\rho (0)=\left\vert \alpha _{1},\alpha _{2}\right\rangle \left\langle \alpha _{1},\alpha _{2}\right\vert $ are: $$F\left( \alpha _{1},\alpha _{2}\right) ={\rm Tr}\left[ \rho (0)\rho (t)%
\right].$$ This is also calculated from Eq. (\[eq:DensityMatrix\]) with Eqs. (\[eq:Qfunction\]) and (\[eq:QfunctionExponential\]) by using identity (\[eq:Identity\]). The fidelity for the initial pure state $\rho (0)=\left| \alpha ,-\alpha \right\rangle \left\langle \alpha ,-\alpha \right| $ is $$F\left( \alpha,-\alpha\right) =\frac{1}{1+N(t)},$$ that is independent of $\alpha$ and tends to 1 for $N_{0}\rightarrow 0$. This clearly contrasts with the fidelity $$F\left( \alpha,\alpha\right) =\frac{1}{1+N(t)}\exp
\left[ -\frac{2(1-e^{-\Gamma t})^{2}}{1+N(t)}\left| \alpha \right| ^{2}\right]$$ calculated for the initial state $\rho (0)=\left| \alpha ,\alpha \right\rangle \left\langle \alpha ,\alpha \right| $. The state $\left| \alpha ,-\alpha \right\rangle \left\langle \alpha ,-\alpha \right| $ is almost noise-free even for finite temperatures while the state $\left| \alpha ,\alpha \right\rangle \left\langle \alpha ,\alpha \right| $ is much more influenced by the noise.
In summary, we gave an example of the noise-free or decoherence-free states in quantum noisy channels with correlation by analyzing two electromagnetic fields coupled to a common thermal reservoir. The results suggest that we can construct a partially noise-free signal state in the bundle of electromagnetic field channels if the modes of signal field are spatially close enough to share partially the common thermal reservoir.
[99]{}
M. A. Nielsen and I. L. Chuang, [*Quantum Computation and Quantum Information*]{} (Cambridge University Press, Cambridge, 2000).
K. Kraus, [*States, Effects, and Operations: Fundamental Notions of Quantum Theory*]{} (Springer-Verlag, Berlin, 1983).
C. Macchiavello and G. M. Palma, Phys. Rev. A [**65**]{}, 050301(R) (2002).
C. King, IEEE Trans. Inf. Theory [**49**]{}, 221 (2003).
D. Bruss, L. Faoro, C. Macchiavello, and G. M. Palma, J. Mod. Opt. [**47**]{}, 325 (2000).
C. Macchiavello, G. M. Palma, and S. Virmani, Phys. Rev. A [**69**]{}, 010303(R) (2004).
J. Ball, A. Dragan, and K. Banaszek, quant/ph-0309148.
Y. Yeo and A. Skeen. Phys. Rev. A [**67**]{}, 064301 (2003).
G. Bowen and S. Mancini, Phys. Rev. A [**69**]{}, 012306 (2004).
G. M. Palma, K.-A. Suominen, and A. K. Ekert, Proc. R. Soc. London, Ser. A [**452**]{}, 567 (1996).
J. Kempe, D. Bacon, D. A. Lidar, and K. B. Whaley, Phys. Rev. A [**63**]{}, 042307 (2001).
H. Jeong, J. Lee, and M. S. Kim, Phys. Rev. A [**61**]{}, 052101 (2000).
T. Hiroshima, Phys. Rev. A [**63**]{}, 022305 (2001).
S. Scheel and D.-G. Welsch, Phys. Rev. A [**64**]{}, 063811 (2001).
D. Wilson, H. Jeong, and M. S. Kim, J. Mod. Opt. [**49**]{}, 851 (2002).
J. S. Prauzner-Bechcicki, J. Phys. A [**37**]{}, L173 (2004).
D. F. Walls and G. J. Milburn, [*Quantum Optics*]{} (Springer-Verlag, Berlin, 1994).
M. J. Collett, Phys. Rev. A [**38**]{}, 2233 (1988).
R. R. Puri, [*Mathematical Methods of Quantum Optics*]{} (Springer-Verlag, Berlin, 2001).
P. D. Drummond and C. W. Gardiner, J. Phys. A [**13**]{}, 2353 (1980).
B. C. Sanders, Phys. Rev. A [**45**]{}, 6811 (1992).
S. J. van Enk and O. Hirota, Phys. Rev. A [**64**]{}, 022313 (2001).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Optimal Morse matchings reveal essential structures of cell complexes which lead to powerful tools to study discrete geometrical objects, in particular discrete $3$-manifolds. However, such matchings are known to be NP-hard to compute on $3$-manifolds, through a reduction to the erasability problem.
Here, we refine the study of the complexity of problems related to discrete Morse theory in terms of parameterized complexity. On the one hand we prove that the erasability problem is $W[P]$-complete on the natural parameter. On the other hand we propose an algorithm for computing optimal Morse matchings on triangulations of $3$-manifolds which is fixed-parameter tractable in the treewidth of the bipartite graph representing the adjacency of the $1$- and $2$-simplexes. This algorithm also shows fixed parameter tractability for problems such as erasability and maximum alternating cycle-free matching. We further show that these results are also true when the treewidth of the dual graph of the triangulated $3$-manifold is bounded. Finally, we investigate the respective treewidths of simplicial and generalized triangulations of $3$-manifolds.
author:
- 'Benjamin A. Burton[^1]'
- 'Thomas Lewiner[^2]'
- 'Jo[ã]{}o Paix[ã]{}o[^3]'
- 'Jonathan Spreer[^4]'
title: Parameterized Complexity of Discrete Morse Theory
---
[**Keywords**]{}: discrete Morse theory, parameterized complexity, fixed parameter tractability, treewidth, $W[P]$-completeness, computational topology, collapsibility, alternating cycle-free matching
Introduction {#sec:introduction}
============
Classical Morse theory [@morse1931smoothmorse] relates the topology of a manifold to the critical points of scalar functions defined on it, providing efficient tools to understand essential structures on manifolds. Forman [@forman1998morse] recently extended this theory to arbitrary cell complexes. In this discrete version of Morse theory, alternating cycle-free matchings in the Hasse diagram of the cell complex, so-called *Morse matchings*, play the role of smooth functions on the manifold [@forman1998morse; @chari2000decomposition]. For example, similarly to the smooth case [@reeb1946points], a closed manifold admitting a Morse matching with only two unmatched (critical) elements is a sphere [@forman1998morse]. The construction of specific Morse matchings has proven to be a powerful tool to understand topological [@forman1998morse; @joswig2004computing; @joswig2006computing; @lewiner2003optimal; @lewiner2003toward], combinatorial [@chari2000decomposition; @jonsson2005graph; @lange2004topocombinatorics] and geometrical [@gyulassy2008visualization; @lewiner2005geometric; @Robins2011; @reininghaus2012computationalmorse] structures of discrete objects.
Morse matchings that minimize the number of critical elements are known as [*optimal matchings*]{} [@lewiner2003optimal]. Together with their number and type of critical elements, these are topological (more precisely simple homotopy) invariants of the cell complex, just like in the case of the sphere described above. Hence, computing optimal matchings can be used as a purely combinatorial technique in computational topology [@Dey1999comptopo]. Moreover, optimal Morse matchings are useful in practical applications such as volume encoding [@lewiner2004apps; @szymczak1999grow], or homology and persistence computation [@lewiner2005geometric; @guenther12].
However, constructing optimal matchings is known to be NP-hard on general $2$-complexes and on $3$-manifolds [@joswig2004computing; @joswig2006computing; @lewiner2003optimal]. This result follows from a reduction to this problem from the closely related *erasability problem*: how many faces must be deleted from a $2$-dimensional simplicial complex before it can be completely erased, where in each erasing step only [*external triangles*]{}, i.e. triangles with an edge not lying in the boundary of any other triangle of the complex, can be removed [@eugeciouglu1996computationally]? Despite this hardness result, large classes of inputs – for which worst case running times suggest the problem is intractable – allow the construction of optimal Morse matchings in a reasonable amount of time using simple heuristics [@lewiner2003toward]. Such behavior suggests that, while the problem is hard to solve for some instances, it might be much easier to solve for instances which occur in practice. As a consequence, this motivates us to ask what [*parameter*]{} of a problem instance is responsible for the intrinsic hardness of the optimal matching problem.
In this article, we study the complexity of Morse type problems in terms of parameterized complexity. Following Downey and Fellows [@downey1994parameterized], an NP-complete problem is called [*fixed-parameter tractable*]{} (FPT) with respect to a [*parameter*]{} $k \in \mathbb{N}$, if for every input with parameter less or equal to $k$, the problem can be solved in $O(f(k)\cdot n^{O(1)})$ time, where $f$ is an arbitrary function independent of the problem size $n$. For NP-complete but fixed-parameter tractable problems, we can look for classes of inputs for which fast algorithms exist, and identify which aspects of the problem make it difficult to solve. Note that the significance of an FPT result strongly depends on whether the parameter is (i) small for large classes of interesting problem instances and (ii) easy to compute.
In order to also classify fixed-parameter intractable NP-complete problems, Downey and Fellows [@downey1994parameterized] propose a family of complexity classes called the [*$W$-hierarchy*]{}: $FPT \subseteq W[1] \subseteq W[2] \subseteq \cdots \subseteq W[P] \subseteq XP$. The base problems in each class of the $W$-hierarchy are versions of satisfiability problems with increasing logical depth as parameter. Class $W[P]$ contains the satisfiability problems with unbounded logical depth. The rightmost complexity class $XP$ of the $W$-hierarchy contains all problems which can be solved in $O(n^k)$ time where $k$ is the parameter of the problem.
Here, we use the notion of the $W$-hierarchy in a geometric setting. More precisely, we determine the hardness of Morse type problems using the mathematically rigid framework of the $W$-hierarchy. Our first main result shows that the erasability problem is $W[P]$-complete (Theorems \[thm:inWP\] and \[thm:WPhard\]), where the parameter is the [*natural parameter*]{} – the number of cells that have to be removed. In other words, we prove that the erasability problem is fixed-parameter intractable in this parameter. From a discrete Morse theory point of view, this reflects the intuition that reaching optimality in Morse matchings requires a global (at least topological) context. In this way, we also show that the $W$-hierarchy as a purely complexity theoretical tool can be used in a very natural way to answer questions in the field of computational topology. Although there are many results about the computational complexity of topological problems [@agol06-knotgenus; @burton2012complexity; @eugeciouglu1996computationally; @malgouyres2008determining; @tancer2008d], to the authors’ knowledge, erasability is the first purely geometric problem shown to be $W[P]$-complete.
Our second main result refines the observation that simple heuristics allow us to compute optimal matchings efficiently. For general $2$-complexes (and $3$-manifolds), the problem reduces directly to finding a maximal alternating cycle-free matching on a spine, i.e., a bipartite graph representing the $1$- and $2$-cell adjacencies [@ayala2012perfect; @joswig2006computing; @lewiner2004apps] (Lemma \[lem:morse\_manifold\]). To solve this problem, we propose an explicit algorithm for computing maximal alternating cycle-free matchings which is fixed-parameter tractable in the treewidth of this bipartite graph (Theorem \[thm:treewidth\]). Furthermore, we show that finding optimal Morse matchings on triangulated $3$-manifolds is also fixed-parameter tractable in the treewidth of the dual graph of the triangulation (Theorem \[thm:fpg\]), which is a common parameter when working with triangulated $3$-manifolds [@burton2012complexity].
Finally, we use the classification of simplicial and generalized triangulations of $3$-manifolds to investigate the “typical” treewidth of the respective graphs for relevant instances of Morse type problems. In this way, we give further information on the relevance of the fixed parameter results. The experiments show that the average treewidths of the respective graphs of simplicial triangulations of $3$-manifolds are particularly small in the case of generalized triangulations. Furthermore, experimental data suggest a much more restrictive connection between the treewidth of the dual graph and the spine of triangulated $3$-manifolds than the one stated in Theorem \[thm:fpg\].
Preliminaries {#sec:prelims}
=============
#### Triangulations {#triangulations .unnumbered}
Throughout this paper we mostly consider simplicial complexes of dimensions $2$ and $3$, although most of our results hold for more general combinatorial structures. All $2$-dimensional simplicial complexes we consider are (i) pure, i.e., all maximal simplexes are triangles ($2$-simplexes) and (ii) [*strongly connected*]{}, i.e., each pair of triangles is connected by a path of triangles such that any two consecutive triangles are joined by an edge ($1$-simplex). All $3$-dimensional simplicial complexes we consider are triangulations of closed $3$-manifolds, that is, simplicial complexes whose underlying topological space is a closed $3$-manifold. In particular every $3$-manifold can be represented in this way [@moise1952]. We will refer to these objects as [*simplicial triangulations of $3$-manifolds*]{}.
In Section \[sec:experimental\] we briefly concentrate on a slightly more general notion of a *generalized triangulation of a $3$-manifold*, which is a collection of tetrahedra all of whose faces are affinely identified or “glued together” such that the underlying topological space is a $3$-manifold. Generalized triangulations use far fewer tetrahedra than simplicial complexes, which makes them important in computational $3$-manifold topology (where many algorithms are exponential time). Every simplicial triangulation is a generalized triangulation, and the second barycentric subdivision of a generalized triangulation is a simplicial triangulation [@moise1952], hence both objects are closely related.
For the remainder of this article, we will often consider $2$-dimensional simplicial complexes as part of a simplicial triangulation of a $3$-manifold.
#### Erasability of simplicial complexes {#erasability-of-simplicial-complexes .unnumbered}
Let $\Delta$ be a $2$-dimensional simplicial complex. A triangle $t \in \Delta$ is called *external* if $t$ has at least one edge which is not in the boundary of any other triangle in $\Delta$; otherwise $t$ is called *internal*. Given a $2$-dimensional simplicial complex $\Delta$ and a triangle $t \in \Delta$, the $2$-dimensional simplicial complex obtained by removing (or [*erasing*]{}) $t$ from $\Delta$ is denoted by $\Delta \setminus t$. In addition, if $\Delta'$ is obtained from $\Delta$ by iteratively erasing triangles such that in each step the erased triangle is external in the respective complex, we will write $\Delta \rightsquigarrow \Delta'$. We say that the complex $\Delta$ is *erasable* if $\Delta \rightsquigarrow \emptyset$, where in this context $\emptyset$ denotes a complex with no triangle. Finally, for every $2$-dimensional simplicial complex $\Delta$ we define $\operatorname{er}(\Delta)$ to be the size of the smallest subset $\Delta_0$ of triangles of $\Delta$ such that $\Delta \setminus \Delta_0 \rightsquigarrow \emptyset$. The elements of $\Delta_0$ are called [*critical*]{} triangles and hence $\operatorname{er}(\Delta)$ is sometimes also referred to as the [*minimum number of critical triangles*]{} of $\Delta$. Determining $\operatorname{er}(\Delta)$ is known as the erasability problem [@eugeciouglu1996computationally].
[A $2$-dimensional simplicial complex $\Delta$.]{} [A non-negative integer $k$.]{} [Is $\operatorname{er}(\Delta)\le k$?]{}
#### Hasse diagram and spine {#hasse-diagram-and-spine .unnumbered}
Given a simplicial complex $\Delta$, one defines its [*Hasse diagram*]{} $H$ to be a directed graph in which the set of nodes of $H$ is the set of simplexes of $\Delta$, and an arc goes from $\tau$ to $\sigma$ if and only if $\sigma$ is contained in $\tau$ and $ \dim (\sigma)+1 = \dim (\tau)$. Let $H_i \subseteq H$ be the bipartite subgraph spanned by all nodes of $H$ corresponding to $i$- and $i+1$-dimensional simplexes. In particular, $H_1$ describes the adjacency between the $2$-simplexes and $1$-simplexes of $\Delta$, and will be called the [*spine*]{} of the simplicial complex $\Delta$. The spine of a simplicial complex will be one of the main objects of study in this work.
#### Matchings {#matchings .unnumbered}
By a [*matching of a graph*]{} $G = (N,A)$ we mean a subset of arcs $M \subset A$ such that every node of $N$ is contained in at most one arc in $M$. Arcs in $M$ are called [*matched arcs*]{} and the nodes of the matched arcs are called [*matched nodes*]{}. Nodes and arcs which are not matched are referred to as [*unmatched*]{}. By the [*size*]{} of a matching $M$ we mean the number of matched arcs. A matching $M$ is called a [*maximum matching*]{} of a graph $G$ if there is no matching with a larger size than the size of $M$.
#### Morse matchings {#morse-matchings .unnumbered}
Let $H$ be the Hasse diagram of a simplicial complex $\Delta$ and $M$ be a matching on $H$. Let $H(M)$ be the directed graph obtained from the Hasse diagram by reversing the direction of each arc of the matching $M$. If $H(M)$ is a directed acyclic graph, i.e., $H(M)$ does not contain directed cycles, then $M$ is a [*Morse matching*]{} [@chari2000decomposition]. Furthermore, the number $c_i$ of unmatched vertices representing $i$-simplexes of $\Delta$ is called the number of *critical $i$-dimensional simplexes* and the sum $c(M) = \sum_i c_i$ is said to be the [*total number of critical simplexes*]{}.
The motivation to find optimal Morse matchings is given by the following fundamental theorem of discrete Morse theory due to Forman which deals with simple homotopy [@Cohen1973].
\[forman\_theorem\] Let $M$ be a Morse matching on a simplicial complex $\Delta$. Then $\Delta$ is simple homotopy equivalent to a $CW$-complex with exactly one $d$-cell for each critical $d$-simplex of $M$.
In other words, a Morse matching with the smallest number of critical simplexes gives us the most compact and succinct topological representation up to homotopy. For more information about the basic facts of Morse theory we refer the reader to Forman’s original work [@forman1998morse]. This motivates a fundamental problem in discrete Morse theory, *optimal Morse matching*, as a decision problem in the following form.
[A simplicial complex $\Delta$.]{} [A non-negative integer $k$.]{} [Is there a Morse matching $M$ with $c(M) \leq k$?]{}
Note that <span style="font-variant:small-caps;">Erasability</span> can be restated as a version of <span style="font-variant:small-caps;">Morse Matching</span> where only the number of unmatched $2$-simplexes (that is, $c_2 (M)$) is counted [@lewiner2003optimal].
#### Complexity of Morse matchings {#complexity-of-morse-matchings .unnumbered}
The compelexity of computing optimal Morse matchings is linear on 1-complexes (graphs) [@forman1998morse] and $2$-complexes that are manifolds [@lewiner2003optimal]. Joswig and Pfetsch [@joswig2006computing] prove that if you can solve <span style="font-variant:small-caps;">Erasability</span> in the spine of a $2$-simplicial complex in polynomial time, then you can solve <span style="font-variant:small-caps;">Morse matching</span> in the *entire* complex in polynomial time. The proof technique easily extends to $3$-manifolds, leading to the following lemma which has been mentioned in previous works [@lewiner2004apps; @ayala2012perfect].
\[lem:morse\_manifold\] Let $M$ be a Morse matching on a triangulated $3$-manifold $\Delta$. Then we can compute a Morse matching $M'$ in polynomial time which has exactly one critical $0$-simplex and one critical $3$-simplex, such that $c(M') \leq c(M)$.
In other words, answering <span style="font-variant:small-caps;">Erasability</span> on the spine is the only difficult part when solving <span style="font-variant:small-caps;">Morse Matching</span> on a $3$-manifold. In Section \[sec:treewidth\] we show that if a spine has bounded treewidth, then we can solve <span style="font-variant:small-caps;">Erasability</span> in linear time. Lemma \[lem:morse\_manifold\] therefore generalizes this result to <span style="font-variant:small-caps;">Morse Matching</span> on $3$-manifolds.
W\[P\]-Completeness of the erasability problem {#sec:WpCompleteness}
==============================================
In order to prove that <span style="font-variant:small-caps;">Erasability</span> is $W[P]$-complete in the natural parameter, we first have to take a closer look at what has to be considered when proving hardness results with respect to a particular parameter.
\[def:parameterized\_reduction\] A parameterized problem $L$ reduces to a parameterized problem $L'$, denoted by $L \leq_{FPT} L'$, if we can transform an instance $(x, k)$ of $L$ into an instance $(x',g(k))$ of $L'$ in time $f(k)|x|^{O(1)}$ (where $f$ and $g$ are arbitrary functions), such that $(x, k)$ is a yes-instance of $L$ if and only if $(x',g(k))$ is a yes-instance of $L'$.
As an example, E[ğ]{}ecio[ğ]{}lu and Gonzalez [@eugeciouglu1996computationally] reduce <span style="font-variant:small-caps;">Set Cover</span> to <span style="font-variant:small-caps;">Erasability</span> to show that <span style="font-variant:small-caps;">Erasability</span> is NP-complete. Since their reduction approach turns out to be a parameterized reduction, these results can be restated in the language of parameterized complexity as follows.
<span style="font-variant:small-caps;">Set Cover</span> $\leq_{FPT}$ <span style="font-variant:small-caps;">Erasability</span>, therefore <span style="font-variant:small-caps;">Erasability</span> is $W[2]$-hard.
This shows that, if the parameter $k$ is simultaneously bounded in both problems, <span style="font-variant:small-caps;">Erasability</span> is *at least as hard as* <span style="font-variant:small-caps;">Set Cover</span>. In this section we will determine exactly how much harder <span style="font-variant:small-caps;">Erasability</span> is than <span style="font-variant:small-caps;">Set Cover</span>, which is $W[2]$-complete. Namely, we will show that <span style="font-variant:small-caps;">Erasability</span> is $W[P]$-complete in the natural parameter $k$. This will be done by i) using a $W[P]$-complete problem as an oracle to solve an arbitrary instance of <span style="font-variant:small-caps;">Erasability</span> (Theorem \[thm:inWP\], which shows that <span style="font-variant:small-caps;">Erasability</span> is in $W[P]$), and ii) reducing an arbitrary instance of a suitable problem which is known to be $W[P]$-complete to an instance of <span style="font-variant:small-caps;">Erasability</span> (Theorem \[thm:WPhard\], which shows that <span style="font-variant:small-caps;">Erasability</span> is $W[P]$-hard).
There are only a few problems described in the literature which are known to be $W[P]$-complete [@downey1999parameterized p. 473]. Amongst these problems, the following is suitable for our purposes.
[A finite set $S$ of [*sentences*]{}, and an [*implication relation*]{} $R$ consisting of pairs $(U,s)$ where $U \subseteq S$ and $s \in S$.]{} [A positive integer $k$.]{} [Is there a set $S_0 \subseteq S$ (called an [*axiom set*]{}) with $|S_0| \leq k$ and a positive integer $n$, for which $S_n = S$, where we define $S_i$, $1 \leq i \leq n$, to consist of exactly those $s \in S$ for which either $s \in S_{i-1}$ or there exists a set $U \subseteq S_{i-1}$ such that $(U, s) \in R$?]{}
<span style="font-variant:small-caps;">Minimum Axiom Set</span> is $W[P]$-complete.
In this paper, we show that, preserving the natural parameter $k$, <span style="font-variant:small-caps;">Minimum Axiom Set</span> is both at least and at most as hard as <span style="font-variant:small-caps;">Erasability</span>.
\[thm:inWP\] <span style="font-variant:small-caps;">Erasability</span> $\leq_{FPT}$ <span style="font-variant:small-caps;">Minimum Axiom Set</span>, therefore <span style="font-variant:small-caps;">Erasability</span> is in $W[P]$.
Theorem \[thm:inWP\] shows that <span style="font-variant:small-caps;">Erasability</span> is at most as hard as the hardest problems in $W[P]$. Please refer to the full version of this paper for a detailed proof of Theorem \[thm:inWP\].
In order to show that it is in fact amongst the hardest problems in this class we first need to build some gadgets.
\[def:implication\_gadget\] Let $(S,R,k)$ be an instance of <span style="font-variant:small-caps;">Minimum Axiom Set</span>.
Let $s \in S$ be a sentence. By an [*$s$-gadget*]{} or [*sentence gadget*]{} we mean a triangulated $2$-dimensional sphere with $2n + m$ punctures as shown in , where $m$ is the number of relations $(U,s) \in R$ and $n$ is the number of relations $(U,u) \in R$ such that $s \in U$.
Let $(U,s) \in R$ be a relation. A [*$(U,s)$-gadget*]{} or [*implication gadget*]{} is a collection of $|U|+1$ sentence gadgets for each sentence of $U\cup\{s\}$ together with $2|U|$ nested tubes as shown in such that (i) two tubes are attached to two punctures of the $u$-gadget for each $u \in U$ and (ii) all $2|U|$ boundary components at the other side of the tubes are identified at a single puncture of the $s$-gadget.
Then, by construction the following holds for the $(U,s)$-gadget.
\[lem:implication\_gadget\] A $(U,s)$-gadget can be erased if and only if all sentence gadgets corresponding to sentences in $U$ are already erased.
Clearly, if all sentence gadgets corresponding to sentences in $U$ are erased, the whole gadget can be erased tube by tube. If, on the other hand, one of the sentence gadgets still exists, this gadget together with the two tubes connected to it build a complex without external triangles which thus cannot be erased.
With these tools in mind we can now prove the main theorem of this section.
\[thm:WPhard\] <span style="font-variant:small-caps;">Minimum Axiom Set</span> $\leq_{FPT}$ <span style="font-variant:small-caps;">Erasability</span>, even when the instance of <span style="font-variant:small-caps;">Erasability</span> is a strongly connected pure 2-dimensional simplicial complex $\Delta$ which is embeddable in $\mathbb{R}^3$. Therefore <span style="font-variant:small-caps;">Erasability</span> is $W[P]$-hard.
The simplicial complex $\Delta$ () constructed to prove $W[P]$-hardness of <span style="font-variant:small-caps;">Erasability</span> is in fact embeddable into $\mathbb{R}^3$. This means that, even in the relatively well behaved class of embeddable $2$-dimensional simplicial complexes, <span style="font-variant:small-caps;">Erasability</span> when bounding the number of critical simplexes is still likely to be inherently difficult. Please refer to the full version of this paper for a detailed proof of Theorem \[thm:WPhard\]. The $W[P]$-completeness result implies that if <span style="font-variant:small-caps;">Erasability</span> turns out to be fixed parameter tractable, then $W[P]=FPT$, i.e., every problem in $W[P]$ including the ones lower in the hierarchy would turn out to be fixed parameter tractable, an unlikely and unexpected collapse in parameterized complexity. Also, it would imply that the $n$-variable SAT problem can be solved in time $2^{o(n)}$, that is, better than in a brute force search [@abrahamson1995fixed]. With respect to this result, if we want to prove fixed parameter tractability of <span style="font-variant:small-caps;">Erasability</span>, the parameter must be different from the natural parameter.
Fixed parameter tractability in the treewidth {#sec:treewidth}
=============================================
In this section, we prove that there is still hope to find an efficient algorithm to solve <span style="font-variant:small-caps;">Morse Matching</span>. We give positive results for the field of discrete Morse theory by proving that <span style="font-variant:small-caps;">Erasability</span> and <span style="font-variant:small-caps;">Morse Matching</span> are fixed parameter tractable in the treewidth of the spine of the input simplicial complex, and also in the dual graph of the problem instance in case it is a simplicial triangulation of a $3$-manifold.
Treewidth {#sec:tw}
---------
\[def:treewidth\] A tree decomposition of a graph $G$ is a tree $T$ together with a collection of bags $\{X_i\}$, where $i$ is a node of $T$. Each bag $X_i$ is a subset of nodes of $G$, and we require that (i) every node of $G$ is contained in at least one bag $X_i$ ([*node coverage*]{}); (ii) for each arc of $G$, some bag $X_i$ contains both its endpoints ([*arc coverage*]{}); and for all bags $X_i$, $X_j$ and $X_k$ of $T$, if $X_j$ lies on the unique simple path from $X_i$ to $X_k$ in $T$, then $X_i \cap X_k \subseteq X_j$ ([*coherence*]{}).
The *width* of a tree decomposition is defined as $\max |X_i|-1$, and the *treewidth* of $G$ is the minimum width over all tree decompositions. We will denote the *treewidth* of $G$ by $\mathbf{tw}(G)$.
For bounded treewidth, computing a tree decomposition of a graph $G=(V,E)$ of width $\leq k$ has running time $O(f(k) |V|)$ [@bodlaender1993treedcomposition] due to an algorithm by Bodlaender. Regarding the size of $f(k)$: using the improved algorithm by Perković and Reed [@Perkovic99TreeDecAlgo], at most $O(k^2)$ recursive calls of Bodlaender’s improved linear time fixed-parameter tractable algorithm for bounded treewidth from [@Bodlaender96EfficientTreeDecAlgo] are needed. This latter algorithm in turn is said to have a constant factor $f(k)$ which is “at most singly exponential in $k$”. For further reading on the running times of tree decomposition algorithms see [@Bodlaender05DiscoveringTreewidth; @kloks1994treewidth].
\[def:nice\_treewidth\] A tree decomposition $({X_i \, | \, i \in I}, T)$ is called a [*nice tree decomposition*]{} if the following conditions are satisfied:
1. There is a fixed bag $X_r$ with $|X_r|=1$ acting as the root of $T$ (in this case $X_r$ is called the [*root bag*]{}).
2. If bag $X_j$ has no children, then $|X_j|=1$ (in this case $X_j$ is called a [*leaf bag*]{}).
3. Every bag of the tree $T$ has at most two children.
4. If a bag $X_i$ has two children $X_j$ and $X_k$, then $X_i = X_j = X_k$ (in this case $X_i$ is called a [*join bag*]{}).
5. If a bag $i$ has one child $j$, then either
1. $|X_i| = |X_j| + 1$ and $X_j \subset X_i$ (in this case $X_i$ is called an [*introduce bag*]{}), or
2. $|X_j| = |X_i| + 1$ and $X_i \subset X_j$ (in this case $X_i$ is called a [*forget bag*]{}).
A given tree decomposition can be transformed into a nice tree decomposition () in linear time:
Given a tree decomposition of a graph $G$ of width $w$ and $O(n)$ bags, where $n$ is the number of nodes of $G$, we can find a nice tree decomposition of $G$ that also has width $w$ and $O(n)$ bags in time $O(n)$.
Alternating cycle-free matchings {#ssec:acfm}
--------------------------------
Given a graph $G = (N,A)$ and a matching $M \subset A$ on $G$, an *alternating path* is a sequence of pairwise adjacent arcs such that each matched arc in the sequence is followed by an unmatched arc and conversely. An *alternating cycle* of $M$ is a closed alternating path. Matchings which do not have any such alternating cycle are called *alternating cycle-free matchings*. From the definition of Morse matching, we can state <span style="font-variant:small-caps;">Erasability</span> in the language of alternating cycle-free matchings as follows:
[A bipartite graph $G=(N_1 \cup N_2,A)$.]{} [A nonnegative integer $k$.]{} [Does $G$ has an alternating cycle-free matching $M$ with at most $k$ unmatched nodes in $N_1$?]{}
Specifically, if $G=H_1$ is the spine for some simplicial complex $\Delta$, then <span style="font-variant:small-caps;">Erasability</span> is equivalent to the <span style="font-variant:small-caps;">Alternating cycle-free matching</span> problem.
FPT algorithm for the [alternating cycle-free matching]{} problem
-----------------------------------------------------------------
Courcelle’s theorem [@courcelle1990monadic] can be used to show that <span style="font-variant:small-caps;">Alternating cycle-free matching</span> is fixed parameter tractable (please refer to the full version of this paper). However, this is a purely theoretical result, since the stated complexity contains towers of exponents in the parameter function. This is the reason why, for the remainder of this section, we focus on the construction of a linear time algorithm to solve <span style="font-variant:small-caps;">Alternating cycle-free matching</span> for inputs of bounded treewidth with a significantly faster running time.
\[thm:treewidth\] Let $G = (N_1 \cup N_2, A)$ be a simple bipartite graph with a given nice tree decomposition $({X_i \, | \, i \in I}, T)$. Then the size of a maximum alternating cycle-free matching of $G$ can be computed in $O( 4^{w^2+w} \cdot w^3 \cdot \log(w) \cdot n)$ time, where $n=|N|$ and $w$ denotes the width of the tree decomposition.
#### Algorithm overview {#algorithm-overview .unnumbered}
Our algorithm constructs alternating cycle-free matchings of $G$ along the nice tree decomposition $({X_i \, | \, i \in I}, T)$ of $G$, from the leaves up to the root, visiting each bag exactly once. In the following we will denote by $F_i$, the set of nodes which are already processed and forgotten by the time $X_i$ is reached; we call this the [*set of forgotten nodes*]{}. At each bag $X_i$ of the decomposition, we construct a set $\Mmatchset{i}$ representing [*all*]{} valid alternating cycle-free matchings in the graph induced by the nodes in $X_i \cup F_i$.
The leaf bags contain a single node of $G$, and the only matching is thus empty. At each introduce bag $X_i=X_j\cup\{x\}$, each matching $\Mmatch$ of $\Mmatchset{j}$ can be extended to several matchings as follows. The newly introduced node $x$ can be either left unmatched, or matched with one of its neighbors as long as it generates a valid and cycle-free matching with $\Mmatch$. At each join bag $X_i=X_j=X_k$, $\Mmatchset{i}$ is build from the valid combinations of pairs of matchings from $\Mmatchset{j}$ and $\Mmatchset{k}$. The final list of valid matchings is then evaluated at the root bag $r$.
However, this final list $\Mmatchset{r}$ contains an exponential number of matchings. Fortunately, the nice tree decomposition allows us to group together, at each step, all matchings $\Mmatch$ that coincide on the nodes of $X_i$. Indeed, the algorithm takes the same decisions for all the matchings of the group. We can thus store and process a much smaller list $\Mmatchset{j}$ of matchings containing only one representative $\Mmatchr$ of each group. In each group, we choose one with the smallest number of unmatched nodes so far. This grouping takes place at the forget and join bags. This makes the algorithm exponential in the bag size, not the input size. The algorithm is described step-by-step and illustrated in the full version of this paper.
#### Matching data structure {#matching-data-structure .unnumbered}
The structure storing an alternating cycle-free matching $\Mmatch$ in a set $\Mmatchset{i}$ must be suitable for checking the matching validity whenever a matching is extended at an introduce bag or a join bag. It must store which nodes are already matched in $\Mmatch$ to avoid matching a node of $G$ twice ([*matching condition*]{}). We use a binary vector $\mathbf{v}(\Mmatch)$, where the $x$-th coordinate is $1$ if node $x \in X_i$ is matched and $0$ otherwise. Checking the matching condition and updating when nodes are matched has thus a constant execution time $O(1)$.
Also, the structure must store which nodes are connected by an alternating path in $\Mmatch$ to avoid closing a cycle when extending or combining $\Mmatch$ ([*cycle-free condition*]{}). When matching two nodes $x$ and $y$, an alternating cycle is created if there exists an alternating path from a neighbor of $x$ to a neighbor of $y$. To test this, we use a structure [@Tarjan1975unionfind] $\mathbf{uf}(\Mmatch)$, storing for each matched node $x$ the index of a matched node $c(x)$ connected to $x$ by an alternating path in $\Mmatch$. For a subset of matched nodes which are all connected to each other, the component index $c$ is chosen to be the node with the lowest index. For each unmatched node, we store the ordered list of component indexes of neighbor matched nodes. The cycle-free condition check reduces to calls on the adjacent lists, and the update of the structure when increasing the matching size reduces to calls, both executing in near-constant time. All the matchings are stored in a hash structure to allow faster search for duplicates. Finally, we can return not only the maximal cycle-free matching size, but an actual maximal cycle-free matching by storing, along with each representative matching, a binary vector of size $|X_i\cup F_i|$ with all the matched nodes so far.
#### Grouping {#grouping .unnumbered}
Traversing the nice tree decomposition in a bottom-up fashion, each node appears in a set of bags that form a subtree of the tree decomposition (*coherence requirement*). This means that, whenever a node is forgotten, it is never introduced again in the bottom-up traversal.
A naïve version of the algorithm described above would build the complete list of valid alternating cycle-free matchings: the set $\Mmatchset{i}$ would contain all valid matchings in the graph induced by the nodes in $X_i \cup F_i$. In particular, for each matching $\Mmatch \in \Mmatchset{i}$ the algorithm would store the binary vector $\mathbf{v}(\Mmatch)$ and the structure $\mathbf{uf}(\Mmatch)$ on $X_i \cup F_i$. However, it is sufficient to store the essential information about each $\Mmatch$ by restricting the structure $\mathbf{uf}(\Mmatch)$ and the binary vector $\mathbf{v}(\Mmatch)$ [*only*]{} to the nodes in the bag $X_i$ (for any matched node $x \in X_i$, node $c(x)$ of the structure is then chosen inside $X_i$). More precisely, we define an equivalence relation $\sim_i$ on the matchings of $\Mmatchset{i}$ such that $\Mmatch\sim_i\Mmatch'$ if and only if $\mathbf{v}(\Mmatch)=\mathbf{v}(\Mmatch')$ and $\mathbf{uf}(\Mmatch)=\mathbf{uf}(\Mmatch')$ on the nodes of $X_i$. Since two equivalent matchings only differ on the forgotten nodes $F_i$, the validation of the matching and cycle-free conditions of any extension of $\Mmatch$ or $\Mmatch'$ (or any combination with a third equivalent matching $\Mmatch''$) will be equal from now on.
Since we are interested in the alternating cycle-free matching with the minimum number of unmatched nodes, for each equivalence class we will choose a matching $\Mmatchr$ with the minimum number $m(\Mmatchr)$ of unmatched forgotten nodes as class representative. This number $m(\Mmatchr)$ is stored together with $(\mathbf{v},\mathbf{uf})$ for each equivalence class of $\Mmatchsetr{i}=\Mmatchset{i}/\!\!\sim_i$. In addition, we can compute the alternating cycle-free matching of maximum size by storing the complete binary vector $\mathbf{v}$ along with $m(\Mmatchr)$ (since the matching is cycle-free, this is sufficient to recover the set of arcs defining the matching).
#### Execution time complexity {#execution-time-complexity .unnumbered}
To measure the running time we need to bound the number of equivalence classes of $\Mmatchsetr{i}$. Let $w_i$ be the number of nodes in $X_i$. The number of equivalence classes of $\Mmatchsetr{i}$ is then bounded above by the number of possible pairs $(\mathbf{v},\mathbf{uf})$ on $w_i$ nodes. The stores for each node $x$, either a component node $c(x)\in X_i$ or a list of at most $w_i$ component nodes, leading to at worst $2^{w_i}$ different lists per node, giving $2^{w_i^2}$ possible combinations of lists. Also there are $2^{w_i}$ possible binary vectors $\mathbf{v}$ of length $w_i$, therefore there are at worst $2^{w_i^2}2^{w_i}$ elements in $\Mmatchsetr{i}$ (this enumeration includes invalid matchings and incoherences between $\mathbf{v}$ and $\mathbf{uf}$).
The time complexity is dominated by the execution at the join bag where pairs of equivalences classes from $\Mmatchsetr{j}$ and $\Mmatchsetr{k}$ have to be combined. Therefore we must square the number of equivalence classes in each set: the complexity for a join bag is $O( 4^{w^2+w} \cdot w^3 \cdot \log(w))$ (please refer to the full version of this paper for details). Since there are $O(n)$ bags in a nice tree decomposition, the total execution time is in $O( 4^{w^2+w} \cdot w^3 \cdot \log(w)\cdot n)$. Finally, as already stated in Section \[sec:tw\], for bounded treewidth computing a tree decomposition and a nice tree decomposition is linear. So the whole process from the bipartite graph to the resulting maximal alternating cycle-free matching is fixed-parameter tractable in the treewidth. Note that neither the decomposition nor the algorithm use the fact that the graph is bipartite.
$n$ $\min w$ $\max w$ $\min w$ $\max w$
----- ----------- ------------ ---------- ------------ ---------- ---------- ------ ------ --------- ----------
$1$ $ 4$ $(3)$ $ 1$ $ 2$ $ 1.50$ $(1.67)$ $ 0$ $ 0$ $ 0.00$
$2$ $ 17$ $(12)$ $ 1$ $ 3$ $ 2.47$ $(2.42)$ $ 1$ $ 1$ $ 1.00$
$3$ $ 81$ $(63)$ $ 1$ $ 3$ $ 2.51$ $(2.49)$ $ 1$ $ 2$ $ 1.60$ $(1.52)$
$4$ $ 577$ $(433)$ $ 1$ $ 5(4)$ $ 2.77$ $(2.73)$ $ 1$ $ 3$ $ 1.91$ $(1.87)$
$5$ $ 5184$ $(3961)$ $ 1$ $ 6(5)$ $ 2.95$ $ $ $ 1$ $ 4$ $ 2.16$ $(2.18)$
$6$ $ 57753$ $(43584)$ $ 1$ $ 6$ $ 3.16$ $(3.19)$ $ 1$ $ 4$ $ 2.31$ $(2.35)$
$7$ $ 722765$ $(538409)$ $ 1$ $ 7$ $ 3.35$ $(3.40)$ $ 1$ $ 4$ $ 2.45$ $(2.50)$
Correctness of the Algorithm
----------------------------
We must check that the algorithm, without the grouping, considers every possible alternating cycle-free matching in $G$ and that the grouping occurring at the forget and join bags does not discard the maximal matching.
The *node coverage* and *arc coverage* properties of nice tree decompositions (Definition \[def:treewidth\]) ensure that each node is processed and each arc is considered for inclusion in the matching at one introduce node. Since the introduce node discards only matchings that violate either the matching or the cycle condition, and these violations cannot be legalized by further extensions or combinations of the matchings, all possible valid matchings are considered.
Now, consider two matchings $\Mmatch$ and $\Mmatch'$ that are grouped together and represented by $\Mmatchr$ at a forget or join bag $X_i$. In the further course of the algorithm, the representative $\Mmatchr$ is then extended or combined with other matchings to form new valid matchings $\Mmatchr'$. The *coherence* property of Definition \[def:treewidth\] assures that no neighbor of a newly introduced node can be a forgotten node, so the extension or combination only modifies matchings $\Mmatch$ and $\Mmatchr$ on nodes of $X_i$, which are represented in the structure of $\Mmatchr'$. Hence, the valid matchings $\Mmatchr'$ actually represent all the valid extensions and combinations of $\Mmatch$ and $\Mmatchr$. The grouping thus generates all valid and relevant representatives of matchings in order to find a maximal alternating cycle-free matching. Moreover, in case $\Mmatch$ and $\Mmatch'$ are equivalent and both with the lowest number of forgotten unmatched nodes, choosing $\Mmatch$ or $\Mmatch'$ as representative leads to the exact same extensions and combinations.
Finally, let $\Mmatch_m$ be the alternating cycle-free matching of maximum size of $G$. In each bag the corresponding matching must be one of the matchings with the lowest number of unmatched nodes within its equivalence class $\Mmatchr_m \in \Mmatchsetr{i}$. Otherwise, a matching in the same class $\Mmatchr_m$, extended and combined as $\Mmatch_m$ in the sequel of the algorithm would give rise to a matching with fewer unmatched nodes. Therefore, the choice of the representative at the forget and join bags never discards the future alternating cycle-free matching of maximum size.
Treewidth of the dual graph
---------------------------
Up to this point, we have been dealing primarily with simplicial complexes and their spines. We now turn our attention to simplicial triangulations of $3$-manifolds and a more natural parameter associated to them.
The *dual graph* of a simplicial triangulation of a $3$-manifold $\mathcal{T}$, denoted $\Gamma(\mathcal{T})$, is the graph whose nodes represent tetrahedra of $\mathcal{T}$, and whose arcs represent pairs of tetrahedron faces that are joined together.
We show that, if the treewidth of the dual graph is bounded, so is the treewidth of the spine, as stated by the following theorem (please refer to the full version of this paper for the proof).
\[thm:fpg\] Let $G$ be the spine of a simplicial triangulation of a $3$-manifold $\mathcal{T}$. If $\mathbf{tw}(\Gamma(\mathcal{T})) \leq k$, then $\mathbf{tw}(G) \leq 10k+9$.
$n$ $\#$ triangulations
------ --------------------- ------------ --------- ------------ --------- --------- ---------- --------- ------ --------- ------- --------- ---------
$5$ $ 1$ $ 6$ $ 6$ $ 6.00$ $ 4$ $ 4$ $ 4.00$
$6$ $ 2$ $ \leq$ $ 7$ $ \leq$ $ 8$ $ \leq$ $ 7.50$ $ 4$ $ 5$ $ 4.50$
$7$ $ 5$ $ \leq$ $ 8$ $ \leq$ $ 11$ $ \leq$ $ 9.40$ $ 4$ $ 6$ $ 5.00$
$8$ $ 39$ $ \leq$ $ 8$ $ \leq$ $ 14$ $ \leq$ $ 11.23$ $ 4$ $ 7$ $ 5.74$
$9$ $ 1297$ $ \leq$ $ 8$ $ \leq$ $ 18$ $ \leq$ $ 13.55$ $ 4$ $ 9$ $ 7.01$
$10$ $ 249015$ $ \leq$ $ 8$ $ \leq$ $ 22$ $ \leq$ $ 16.33$ $ \leq$ $ 4$ $ \leq$ $ 13$ $ \leq$ $ 8.99$
Experimental Results {#sec:experimental}
====================
In Section \[sec:WpCompleteness\] we have seen that the problem of finding optimal Morse matchings is hard to solve in general. In Section \[sec:treewidth\] on the other hand we proved that in the case of a small treewidth of the spine of a $2$-dimensional complex or, equivalently, in the case of a bounded treewidth of the dual graph of a simplicial triangulation of a $3$-manifold, finding an optimal Morse matching becomes easier. Up to a certain scaling factor, the results stated in Section \[sec:treewidth\] hold for generalized triangulations as well (also, note that the notion of a spine or the dual graph can be extended in a straightforward way to generalized triangulations).
Given this situation, a natural question to ask is the following: What is a [*typical*]{} value for the treewidth of the respective graphs of (i) small generic generalized triangulations of $3$-manifolds, and (ii) small generic simplicial triangulations of $3$-manifolds? In a series of computer experiments we computed the treewidth of the relevant graphs (i.e., the spine and the dual graph) of all closed generalized triangulations of $3$-manifolds up to $7$ tetrahedra [@burton11-genus], and all simplicial triangulations of $3$-manifolds up to $10$ vertices [@Lutz08ThreeMfldsWith10Vertices]. The computer experiments were done using [@Dijk06TreewidthDotCom] to compute the treewidth / upper bounds for the treewidth, with the help of the package [@simpcompISSAC; @simpcomp] and the $3$-manifold software [@burton04-regina; @regina]. We report the minimal, maximal and average treewidths of all triangulations with the same number of vertices in Tables \[tab:closed\] and \[tab:comb\].
Regarding the treewidth of generalized triangulations of $3$-manifolds, we observe that there is a large difference between the average treewidth and the maximal treewidth for both the dual graph and the spine. In particular, the average treewidth appears to be relatively small. Moreover, there is only a slight difference between the data for general closed triangulations and $1$-vertex triangulations. This fact is somehow in accordance with our intuition since the number of $0$-dimensional simplexes should neither directly affect the spine nor the dual graph of a generalized triangulation.
On the other hand, the gap between the maximum treewidth and the average treewidth in the case of simplicial triangulations of $3$-manifolds is relatively small compared to the data for generalized triangulations. In addition, the treewidth of the spines of some particularly interesting $2$-dimensional simplicial complexes (reported in the full version of this paper) is significantly smaller than the (upper bound of the) treewidth of simplicial triangulations of $3$-manifolds. At this point it is important to note that, while the data concerning the spines for simplicial complexes only consists of upper bounds, experiments applying the algorithm for the upper bound to smaller graphs and then computing their real treewidths suggest that these upper bounds (in average) are reasonably close to the exact treewidth.
Further analysis shows that the average treewidth of the spines for both generalized and simplicial triangulations of $3$-manifolds is mostly less than twice the treewidth of the dual graph, and hence much below the theoretical upper bound given by Theorem \[thm:fpg\]. Also, the ratio between these two numbers appears to be more or less stable for all values shown in Tables \[tab:closed\] and \[tab:comb\]. This can be seen as experimental evidence that for triangulated $3$-manifolds the treewidth of the dual graph is responsible for the inherent difficulty to solve <span style="font-variant:small-caps;">Erasability</span> and related problems.
Despite the small values of $n$ in our tables, there are theoretical reasons to believe that the patterns of small treewidth will continue for larger $n$. For instance, the conjectured minimal triangulations of Seifert fibered spaces over the sphere have dual graphs with $O(1)$ treewidth for arbitrary $n$. Moreover, following recent results of Gabai *et al.* [@Gabai09MinVolCuspedHyperb3Mflds] there are reasons to believe that large infinite classes of topological $3$-manifolds admit triangulations whose treewidths are below provable upper bounds. Investigating these upper bounds is work in progress.
Acknowledgments
===============
This work is partially financed by CNPq, FAPERJ, PUC-Rio, CAPES, and Australian Research Council Discovery Projects DP1094516 and DP110101104. We would also like to thank Michael Joswig for fruitful discussions.
[10]{}
K. A. Abrahamson, R. G. Downey, and M. R. Fellows. Fixed-parameter tractability and completeness [IV]{}: On completeness for [W\[P\]]{} and [PSPACE]{} analogues. , 73(3):235–276, 1995.
I. Agol, J. Hass, and W. Thurston. The computational complexity of knot genus and spanning area. , 358(9):3821–3850, 2006.
R. Ayala, D. Fern[á]{}ndez-Ternero, and J. Vilches. Perfect discrete morse functions on triangulated 3-manifolds. , pages 11–19, 2012.
H. Bodlaender. Discovering treewidth. In P. Vojt[á]{}[š]{}, M. Bielikov[á]{}, B. Charron-Bost, and O. S[ý]{}kora, editors, [*SOFSEM 2005: Theory and Practice of Computer Science*]{}, volume 3381 of [*Lecture Notes in Computer Science*]{}, pages 1–16. Springer Berlin Heidelberg, 2005.
H. L. Bodlaender. A linear time algorithm for finding tree-decompositions of small treewidth. In [*Symposium on Theory of Computing*]{}, pages 226–234. ACM, 1993.
H. L. Bodlaender and T. Kloks. Efficient and constructive algorithms for the pathwidth and treewidth of graphs. , 21(2):358–402, 1996.
B. A. Burton. Introducing [R]{}egina, the 3-manifold topology software. , 13(3):267–272, 2004.
B. A. Burton. Detecting genus in vertex links for the fast enumeration of 3-manifold triangulations. In [*International Symposium on Symbolic and Algebraic Computation*]{}, pages 59–66. ACM, 2011.
B. A. Burton, R. Budney, W. Pettersson, et al. Regina: Software for 3-manifold topology and normal surface theory. <http://regina.sourceforge.net/>, 1999–2012.
B. A. Burton and J. Spreer. The complexity of detecting taut angle structures on triangulations. In [*[SODA]{} ’13: Proceedings of the Twenty-Fourth Annual [ACM-SIAM]{} Symposium on Discrete Algorithms*]{}, pages 168–183. SIAM, 2013.
M. K. Chari. On discrete [M]{}orse functions and combinatorial decompositions. , 217:101–113, 2000.
M. M. Cohen. . Graduate text in Mathematics. Springer, New York, 1973.
B. Courcelle. The monadic second-order logic of graphs [I]{}: recognizable sets of finite graphs. , 85(1):12–75, 1990.
T. K. Dey, H. Edelsbrunner, and S. Guha. Computational topology. In [*Advances in Discrete and Computational Geometry*]{}, volume 223 of [*Contemporary mathematics*]{}, pages 109–143. AMS, 1999.
R. Downey, M. Fellows, B. Kapron, M. Hallett, and H. Wareham. The parameterized complexity of some problems in logic and linguistics. , pages 89–100, 1994.
R. G. Downey and M. R. Fellows. , volume 3. Springer New York, 1999.
F. Effenberger and J. Spreer. simpcomp - a [GAP]{} toolbox for simplicial complexes. , 44(4):186 – 189, 2010.
F. Effenberger and J. Spreer. simpcomp - a [GAP]{} package, [V]{}ersion 1.5.4. <http://code.google.com/p/simpcomp/>, 2011.
. E[ğ]{}ecio[ğ]{}lu and T. F. Gonzalez. A computationally intractable problem on simplicial complexes. , 6(2):85–98, 1996.
R. Forman. Morse theory for cell complexes. , 134(1):90–145, 1998.
D. Gabai, R. Meyerhoff, and P. Milley. Minimum volume cusped hyperbolic three-manifolds. , 22(4):1157–1215, 2009.
M. C. Golumbic, T. Hirst, and M. Lewenstein. Uniquely restricted matchings. , 31(2):139–154, 2001.
D. Günther, J. Reininghaus, H. Wagner, and I. Hotz. Efficient computation of 3[D]{} [M]{}orse-[S]{}male complexes and persistent homology using discrete [M]{}orse theory. , 28:959–969, 2012.
A. G. Gyulassy. . PhD thesis, UC Davis, 2008. Advised by Bernd Hamann.
J. Jonsson. . PhD thesis, KTH, 2005. Advised by Anders Bj[ö]{}rner.
M. Joswig and M. E. Pfetsch. Computing optimal discrete morse functions. , 17:191–195, 2004.
M. Joswig and M. E. Pfetsch. Computing optimal morse matchings. , 20(1):11–25, 2006.
T. Kloks. , volume 842. Springer, 1994.
C. Lange. . PhD thesis, Technische Universit[ä]{}t, Berlin, 2004. Advised by G[ü]{}nter M. Ziegler.
T. Lewiner. . PhD thesis, Mathematics, PUC-Rio, 2005.
T. Lewiner, H. Lopes, and G. Tavares. Optimal discrete morse functions for 2-manifolds. , 26(3):221–233, 2003.
T. Lewiner, H. Lopes, and G. Tavares. Toward optimality in discrete morse theory. , 12(3):271–285, 2003.
T. Lewiner, H. Lopes, and G. Tavares. Applications of [F]{}orman’s discrete [M]{}orse theory to topology visualization and mesh compression. , 10(5):499–508, 2004.
F. H. Lutz. Combinatorial 3-manifolds with 10 vertices. , 49(1):97–106, 2008.
R. Malgouyres and A. Franc[é]{}s. Determining whether a simplicial 3-complex collapses to a 1-complex is [NP]{}-complete. In [*Discrete Geometry for Computer Imagery*]{}, pages 177–188. Springer, 2008.
E. E. Moise. Affine structures in 3-manifolds [V]{}: The triangulation theorem and hauptvermutung. , 56(1):96–114, 1952.
M. Morse. The critical points of functions of n variables. , 33:77–91, 1931.
L. Perkovi[ć]{} and B. Reed. An improved algorithm for finding tree decompositions of small width. In [*Graph-theoretic concepts in computer science ([A]{}scona, 1999)*]{}, volume 1665 of [*Lecture Notes in Comput. Sci.*]{}, pages 148–154. Springer, Berlin, 1999.
G. Reeb. Sur les points singuliers d’une forme de [P]{}faff compl[è]{}tement int[é]{}grable ou d’une fonction num[é]{}rique. , 222:847–849, 1946.
J. Reininghaus. . PhD thesis, Freie Universit[ä]{}t, Berlin, 2012. Advised by Ingrid Hotz.
V. Robins, P. J. Wood, and A. P. Sheppard. Theory and algorithms for constructing discrete [M]{}orse complexes from grayscale digital images. , 33(8):1646–1658, 2011.
A. Szymczak and J. Rossignac. Grow & fold: Compression of tetrahedral meshes. In [*Symposium on Solid Modeling and Applications*]{}, pages 54–64. ACM, 1999.
M. Tancer. Strong [$d$]{}-collapsibility. , 6(2):32–35, 2011.
R. E. Tarjan. Efficiency of a good but not linear set union algorithm. , 22(2):215–225, 1975.
T. van Dijk, J.-P. van den Heuvel, and W. Slob. Computing treewidth with [LibTW]{}. <http://www.treewidth.com/>, 2006.
[**Appendix**]{}
Proof of Lemma \[lem:morse\_manifold\] {#app:morse_manifold}
======================================
The proof of Lemma \[lem:morse\_manifold\] actually follows directly from Joswig and Pfetsch’s proof of the following lemma.
Let $M$ be a Morse matching on a simplicial complex $\Delta$. Then we can compute a Morse matching $M'$ in polynomial time which has exactly one critical $0$-simplex, the same number of critical simplexes of dimension greater or equal $2$ as $M$, and $c(M') \leq c(M)$.
The proof builds a Morse matching from a spanning tree of the primal graph, i.e. the graph obtained considering only the vertices and edges of $\Delta$. For a $3$-manifold $\Delta$, the proof of the previous lemma can be applied exactly the same way on the dual graph of $\Delta$ to obtain the following result.
Let $M$ be a Morse matching on a closed triangulated $3$-manifold $\Delta$. Then we can compute a Morse matching $M'$ in polynomial time which has exactly one critical $3$-simplex, the same number of critical simplexes of dimension less or equal $1$, and $c(M') \leq c(M)$.
Since the proof works independently on the primal and dual graph, Lemma \[lem:morse\_manifold\] is a combination of these results. Here, we simply reproduce the proof of Joswig and Pfetsch [@joswig2006computing] verbatim applying it to $3$-manifold complexes, using Poincaré’s duality.
First consider a Morse matching $M$ for a connected $3$-manifold $\Delta$. Let $\gamma(M)$ be the graph obtained from the primal graph of $\Delta$ by removing all arcs (edges of $\Delta$) matched with triangles in $M$ and let $\gamma^\ast(M)$ be obtained from the dual graph of $\Delta$ by removing all the arcs (triangles in $\Delta$) where the corresponding triangles are matched with edges of $\Delta$ in $M$. Note that $\gamma(M)$ and $\gamma^\ast(M)$ contain all vertices and tetrahedra of $\Delta$, respectively.
The graph $\gamma(M)$ and dual graph $\gamma^\ast(M)$ are connected.
Suppose that $\gamma(M)$ is disconnected. Let $N$ be the set of nodes in a connected component of $\gamma(M)$, and let $C$ be the set of cut edges, that is, edges of $\Delta$ with one vertex in $N$ and one vertex in its complement. Since $\Delta$ is connected, $C$ is not empty. By definition of $\gamma(M)$, each edge in $C$ is matched to a unique $2$-simplex.
Consider the directed subgraph $D$ of the Hasse diagram consisting of the edges in $C$ and their matching $2$-simplexes. The standard direction of arcs in the Hasse diagram (from the higher to the lower dimensional simplexes) is reversed for each matching pair of $M$, i.e., $D$ is a subgraph of $H(M)$. We construct a directed path in $D$ as follows. Start with any node of $D$ corresponding to a cut edge $e_1$. Go to the node of $D$ determined by the unique $2$-simplex $\tau_1$ to which $e_1$ is matched to. Then $\tau_1$ contains at least one other cut edge $e_2$, otherwise $e_1$ cannot be a cut edge. Now iteratively go to $e_2$, then to its unique matching 2-simplex $\tau_2$, choose another cut edge $e_3$, and so on. We observe that we obtain a directed path $e_1,\tau_1,e_2,\tau_2, \cdots$ in $D$, i.e., the arcs are directed in the correct direction. Since we have a finite graph at some point the path must arrive at a node of $D$ which we have visited already. Hence, $D$ (and therefore also $H(M)$) contains a directed cycle, which is a contradiction since $M$ is a Morse matching.
To prove that $\gamma^\ast(M)$ is connected, we repeat the proof above on the dual graph.
Let $M$ be a Morse matching on a triangulated $3$-manifold $\Delta$. Then we can compute a Morse matching $M'$ in polynomial time which has exactly one critical $0$-simplex and one critical $3$-simplex, such that $c(M') \leq c(M)$.
Since $\gamma(M)$ and $\gamma^\ast(M)$ are connected, they both have spanning trees, and we will use them to build the Morse matching. First pick an arbitrary node $r_1$ and any spanning tree of $\gamma(M)$ and direct all arcs away from $r_1$. Then pick an arbitrary tetrahedron (a node in the dual graph) $r_2$ and any spanning tree of $\gamma^\ast (M)$ and direct all triangles (arcs in dual graph) away from $r_2$. This yields a maximum Morse matching on $\gamma(M)$ and $\gamma^\ast (M)$. It is easy to see that replacing the part of $M$ on $\gamma(M)$ and $\gamma^\ast (M)$ with this matching yields a Morse matching. This Morse matching has only one critical vertex (the root $r_1$) and one critical tetrahedron (the root $r_2$). Note that every Morse matching in a triangulated $3$-manifold contains at least one critical vertex and at least one critical tetrahedron; this can be seen from Theorem \[forman\_theorem\]. Furthermore, the total number of critical simplexes can only decrease, since we computed an optimal Morse matching on $\gamma(M)$ and $\gamma^\ast (M)$.
Proof of Theorem \[thm:inWP\] {#app:inWP}
=============================
<span style="font-variant:small-caps;">Erasability</span> $\leq_{FPT}$ <span style="font-variant:small-caps;">Minimum Axiom Set</span>, therefore <span style="font-variant:small-caps;">Erasability</span> is in $W[P]$.
We show membership of <span style="font-variant:small-caps;">Erasability</span> in $W[P]$ by reducing a given instance $(\Delta,k)$ of <span style="font-variant:small-caps;">Erasability</span> to an instance $(S,R,k)$ of <span style="font-variant:small-caps;">Minimum Axiom Set</span>.
W. l. o. g. we can assume that the $2$-dimensional simplicial complex $\Delta$ has no external edges (if $\Delta$ has external edges we first remove these edges until no external edge exists and reduce the remaining problem instance to an instance of <span style="font-variant:small-caps;">Minimum Axiom Set</span>). We now identify the set of triangles of $\Delta$ with the set of sentences $S$ in a one-to-one correspondence. For every edge $e \in \Delta$ we denote the set of all triangles containing $e$ by $\operatorname{star}_{\Delta} (e) \subset \Delta$, we write for the corresponding set of sentences $S_e \subset S$, and we define the set of implication relations $R$ by the relations $$(S_e \setminus \{s\}, s)$$ for each triangle $s \in S_e$ for all edges $e \in \Delta$. Note that $\Delta$ has no external edges and thus $S_e \setminus \{s\} \neq \emptyset$ for all $e$.
In a next step, we show that for all axiom sets $S_0 \subset S$ of size $k$ we have $ \Delta \setminus \Delta_0 \rightsquigarrow \emptyset$ for the associated subset of triangles $\Delta_0 \subset \Delta$ of size $k$. To see that this is true, note that for the augmenting sequence $S_0 \subset S_1 \subset \ldots \subset S_n = S$ of $S$, their corresponding subsets of triangles $\Delta_0 \subset \Delta_1 \subset \ldots \subset \Delta_n = \Delta$ and $i \in \{ 1 , \ldots, n \}$ fixed, all sentences $s \in S_i \setminus S_{i-1}$ have to occur in a relation $(S_e \setminus \{s\}, s)$ for some edge $e$ with $S_e \setminus \{ s \} \subset S_{i-1}$. For the triangle $t \in \Delta$ corresponding to $s$ this means that, $\operatorname{star}_{\Delta} (e) \setminus t \subset \Delta_{i-1}$. Thus, if we assume that all triangles in $\Delta_{i-1}$ are already erased, $t$ must be external and thus can be erased as well. The statement now follows by the fact that for $i=1$, all triangles in $\Delta_0$ are already erased in $\Delta \setminus \Delta_0$ and hence $ \Delta \setminus \Delta_0 \rightsquigarrow \emptyset$.
Conversely, let $\Delta_0 \subset \Delta$ be of size $k$ such that $ \Delta \setminus \Delta_0 \rightsquigarrow \emptyset$. Since $\Delta$ has no external triangles but $ \Delta \setminus \Delta_0 \rightsquigarrow \emptyset$, there must be external triangles $ t \in \Delta \setminus \Delta_0 $ and hence for $s\in S$ being the sentence corresponding to the triangle $t$ there is a relation $(S_e \setminus \{s\}, s)$ with $S_e \setminus \{s\} \subset S_0$, where $S_0$ is the set of sentences corresponding to the set of triangles $\Delta_0$. We then define $S_1$ to be the union of $S_0$ with all sentences $s$ of the type described above and iterating this step results in a sequence of subsets $S_0 \subset S_1 \subset \ldots \subset S_n = S$ for some $n$ what proves the result.
Proof of Theorem \[thm:WPhard\] {#app:WPhard}
===============================
<span style="font-variant:small-caps;">Minimum Axiom Set</span> $\leq_{FPT}$ <span style="font-variant:small-caps;">Erasability</span>, even when the instance of <span style="font-variant:small-caps;">Erasability</span> is a strongly connected pure 2-dimensional simplicial complex $\Delta$ which is embeddable in $\mathbb{R}^3$. Therefore <span style="font-variant:small-caps;">Erasability</span> is $W[P]$-hard.
To show $W[P]$-hardness of <span style="font-variant:small-caps;">Erasability</span>, we will reduce an arbitrary instance $(S,R,k)$ from <span style="font-variant:small-caps;">Minimum Axiom Set</span> to an instance $(\Delta,k)$ of <span style="font-variant:small-caps;">Erasability</span>. In order to do so, we will use a sentence gadget for each element of $S$ and an implication gadget for each relation $R$ (cf. Definition \[def:implication\_gadget\]) to construct a $2$-dimensional simplicial complex $\Delta$ with a polynomial number of triangles in the input size.
By construction, we can glue all sentence and implication gadgets together in order to obtain a simplicial complex $\Delta$ without any exterior triangles. Note that the only place where $\Delta$ is not a surface is at the former $m$ boundary components per sentence gadget corresponding to the $m$ relations in $R$ with the respective right hand side.
For any axiom set $S_0 \subset S$ of size $k$, let $\Delta_0$ be a set of $k$ triangles, one from each sentence gadget corresponding to a sentence in $S_0$. It follows by Lemma \[lem:implication\_gadget\], that $\Delta \setminus \Delta_0$ can be erased to a complex where all the sentence gadgets $s$ corresponding to relations $(U,s)$, $U \subset S_0$, have external triangles. Since $S_0$ is an axiom set, iterating this process erases the whole complex $\Delta$.
Conversely, let $\Delta_0$ be a set of $k$ triangles such that $\Delta \setminus \Delta_0 \rightsquigarrow \emptyset$. First, note that erasing a triangle of any tube of an implication gadget always allows us to remove the sentence gadget at the right end of this tube. Hence, w. l. o. g. we can assume that all $k$ triangles in $\Delta_0$ are triangles of some sentence gadget in $\Delta$. Now, if any sentence gadget contains more than one triangle of $\Delta_0$ we delete all additional triangles obtaining a set $\Delta_0'$ of $k'$ triangles, $k' \leq k$, such that $\Delta \setminus \Delta_0' \rightsquigarrow \emptyset$ and thus the corresponding set of sentences is an axiom set of size $k' \leq k$.
The result now follows by the observation that $\Delta$ can be realized by at most a quadratic number of triangles in the input size of $(S,R,k)$.
Fixed parameter tractability of <span style="font-variant:small-caps;">Alternating cycle-free matching</span> from Courcelle’s theorem {#app:courcelle}
======================================================================================================================================
Courcelle’s celebrated theorem [@courcelle1990monadic] states that all graph properties that can be defined in [*Monadic Second-Order Logic*]{} (MSOL) can be decided in linear time when the graph has bounded treewidth. Here, we want to use Courcelle’s theorem to show that problems in discrete Morse theory are fixed parameter tractable in the treewidth of some graph associated to the problem. However, it is not obvious how to [*directly*]{} state <span style="font-variant:small-caps;">Erasability</span> and <span style="font-variant:small-caps;">Morse Matching</span> in MSOL. Instead, we will apply Courcelle’s theorem to <span style="font-variant:small-caps;">Alternating cycle-free matching</span> which by the comment made in Section \[ssec:acfm\] is a graph theoretical problem equivalent to <span style="font-variant:small-caps;">Erasability</span>.
\[thm:courcelle\] Let $w \geq 1$. Given a bipartite graph with $\mathbf{tw}(G) \leq w$, <span style="font-variant:small-caps;">Alternating cycle-free matching</span> can be solved in linear time.
We will write a MSOL formulation of <span style="font-variant:small-caps;">Alternating cycle-free matching</span> based on the fact that $M$ is an alternating cycle-free matching if and only if $M$ is a matching and every induced $M$-subgraph contains a node of degree $1$ [@golumbic2001uniquely]: $$\begin{gathered}
\max M:
\forall x \in N [\neg\exists a_1,a_2 \in M (a_1 \neq a_2 \land inc(x,a_1) \land inc(x,a_2))] \\
\phantom{Mm} \land \forall M' \subseteq M (\exists a \in M', \exists x \in N[inc(x,a) \land (\forall x_1 \in N( \neg \exists a_1 \in M'( x \neq x' \land adj(x,x_1) \land inc(x_1,a_1))))])\end{gathered}$$ where $inc(x,a)$ is the incidence predicate between node $x$ and arc $a$ and $adj(x,x')$ is the adjacency predicate between node $x$ and node $x'$. The above statement can be translated to plain English as follows: “Find the largest matching $M$ of $G$, where each node is incident to at most one arc, such that in every subset $M'$ of the matching $M$ there exists a matched node $x$ in $M'$ such that its only neighbor matched in $M'$ is the other endpoint of the unique matched arc incident to $x$.
Algorithm for <span style="font-variant:small-caps;">Alternating cycle-free matching</span>: step by step {#app:algo}
=========================================================================================================
The algorithm visits the bags of the nice tree decomposition bottom-up from the leaves to the root evaluating the corresponding mappings in each step according to the following rules ().
#### Leaf bag {#leaf-bag .unnumbered}
The set of matchings $\Mmatchsetr{i}$ of a leaf bag $X_i=\{x\}$ is trivial with a unique empty matching $\Mmatchr$ represented by $\mathbf{v}(\Mmatchr) = [0]$, and $\textbf{uf}(\Mmatchr)(x)$ defined by as an empty list, associated with $m(\Mmatchr)=0$.
#### Introduce bag {#introduce-bag .unnumbered}
Let $X_i=X_j\cup\{x\}$ be an introduce bag with child bag $X_j$. The set of valid matchings $\Mmatchsetr{i}$ is built from $\Mmatchsetr{j}$ by introducing $x$ in each matching $\Mmatchr \in \Mmatchsetr{j}$, generating several possible matchings $\Mmatchr'$. We can always introduce $x$ as an unmatched node, then $\Mmatchr$ is extended on $x$ by setting $\mathbf{v}(\Mmatchr')_{|x} = 0$ and updating $\textbf{uf}(\Mmatchr')$ with the ordered list of components for each matched neighbor of $x$. In addition, for each unmatched neighbor $y \in X_j$, we can introduce $x$ as a matched node in the following way. We match both $x$ and $y$ in $\Mmatchr$ and set $\mathbf{v}(\Mmatchr')_{|x} = 1$ and $\mathbf{v}(\Mmatchr')_{|y} = 1$. If the intersection of the list of neighbor components of $x$ and $y$ is empty, then the matching of $x$ and $y$ does not create a cycle. In this case $\Mmatchr'$ is a valid extension of $\Mmatchr$. The update of the structure must then reflect the extensions of all alternating paths through arc $\{x,y\}$. We perform in $\textbf{uf}(\Mmatchr')$ a ${\ifmmode \mathsf{union}\else\textsf{union}\fi}$ operation for $x$ and all its matched neighbors (including $y$), and for $y$ and all its matched neighbors. We also add the merged component index $c(x)$ to the list of neighbor components of each unmatched neighbor of $x$ and $y$. Then we include all valid extensions $\Mmatchr'$ to $\Mmatchsetr{i}$, reducing $\mathbf{v}(\Mmatchr')$ by calling for each node and neighbor component list entry, and we set $m(\Mmatchr)=m(\Mmatchr')$ for all extensions $\Mmatchr'$ of $\Mmatchr$.
#### Running time {#running-time .unnumbered}
There are at most $2^{w_i^2+w_i}$ extended matchings $\Mmatchr'$ for bag $X_i$ (including all invalid ones), where $w_i=|X_i|=|X_j|+1$ (a new possible matching can be generated only once). Each new matching is validated by a direct lookup at $\mathbf{v}(\Mmatchr')$ and ordered list comparison, leading to a linear time $w_i$. The update of each structure requires constant time for each matched neighbor of $x$ and almost linear time $O(w_i)$ plus the sorted insertion $O(w_i\cdot \log(w_i))$ for each unmatched neighbor, and there are at most $w_i$ neighbors in the bag. Thus, the total running time of an introduce bag is in $O(2^{w_i^2+w_i} \cdot w_i^2 \cdot \log(w_i) )$.
#### Forget bag {#forget-bag .unnumbered}
Let $X_i=X_j\setminus\{x\}$ be a forget bag with child bag $X_j \ni x$. While the set of all possible matchings on $X_i \cup F_i$ does not change ($\Mmatchset{j}=\Mmatchset{i}$), the equivalence relation $\sim_i$ possibly identifies more matchings than $\sim_j$. For each matching $\Mmatchr \in \Mmatchsetr{j}$, a new matching $\Mmatch'$ is obtained by deleting coordinate $x$ of $\mathbf{v}(\Mmatchr)$. If $c(x)=x$, $\textbf{uf}(\Mmatchr)$ needs to be updated. To do so, the set of nodes $X_i$ is traversed twice, once to look for node $y\neq x$ of minimal index such that $c(y)=c(x)$ (eventually, $y$ is empty), and a second time to replace $x$ by $y$ each time $x$ is used as a component index. If $x$ was unmatched in $\Mmatchr$ (i.e., $\mathbf{v}(\Mmatchr)_{|x}=0$), then we set $m(\Mmatch')=m(\Mmatchr)+1$, otherwise we set $m(\Mmatch')=m(\Mmatchr)$. Once the set $\Mmatchset{j}'$ of all the newly generated $\Mmatch'$ is computed, $\Mmatchset{i}$ is obtained as the quotient of $\Mmatchset{j}'$ by $\sim_i$, the equivalence relation on $X_i$. More precisely, each pair $(\Mmatch',\Mmatch'')\in \Mmatchset{j}'^2$ is tested for equality on both $\mathbf{v}$ and $\mathbf{uf}$. If they are equal, the one with the lowest $m$ is defined to be the new representative in $\Mmatchset{i}$.
#### Running time {#running-time-1 .unnumbered}
Each new matching $\Mmatch'$ is obtained from a single element of $\Mmatchset{j}$ in worst-case time $O(w_i^2\cdot \log(w_i))$. Equivalent matchings are detected on-the-fly when filling the hash structure of $\Mmatchset{i}$, and each equivalence test is linear in $w_i^2$. The complexity is thus in $O( 2^{w_j^2+w_j} \cdot w_j^2 \cdot \log(w_j) )$.
#### Join bag {#join-bag .unnumbered}
Let $X_i=X_j=X_k$ be a join bag with child bags $X_j$ and $X_k$. The matchings of $\Mmatchset{i}$ are generated by combining all the pairs of matchings $(\Mmatch,\Mmatch')\in \Mmatchset{j} \times \Mmatchset{k}$. A combination is valid if and only if it satisfies both the matching and cycle-free conditions. The matching condition says that a node cannot be matched in both $\Mmatch$ and $\Mmatch'$, which is checked by a logical $AND$ operation ($\mathbf{v}(\Mmatch)\ AND\ \mathbf{v}(\Mmatch')$). The cycle-free condition is checked with the structures $\Mmatch$ and $\Mmatch'$: the combination is valid if no node of the component of a matched node in $\mathbf{uf}(\Mmatch)$ is a neighbor of the same component in $\mathbf{uf}(\Mmatch')$ and vice versa, each test requiring $O(w_i^2)$ per component. If a combination is valid, its structure $\Mmatch''$ is defined by $\mathbf{v}(\Mmatch'') = \mathbf{v}(\Mmatch)\ OR\ \mathbf{v}(\Mmatch')$. The structure is initialized from $\mathbf{uf}(\Mmatch)$, and updated as the introduce bag for each matched nodes of $\Mmatch'$. Finally, $m(\Mmatch'') = m(\Mmatch)+m(\Mmatch')$. As in the forget bag, two combinations may result in equivalent matchings, and we must compare them pairwise and choose the representative with the lowest number of unmatched forgotten bags. Note that the sets of forgotten nodes of $X_j$ and forgotten nodes of $X_k$ have to be disjoint by the *coherence* of Definition \[def:treewidth\] and hence no forgotten node can be matched twice in this setting. Furthermore, all possible combinations of matched and unmatched nodes are enumerated in $\Mmatchset{j}$ and $\Mmatchset{k}$ and hence no possible matching is overseen.
#### Running time {#running-time-2 .unnumbered}
Each pair of matchings is validated and updated in time $O(w_i\cdot w_i^2\cdot\log(w_i))$. The comparison and the choice of representative is done on-the-fly when filling the hash structure of $\Mmatchset{i}$. There are at worst $(2^{w_i^2+w_i})^2$ pairs. Thus, the complexity of the join bag dominates all other running times. Therefore, the complexity of the algorithm is in $O( 4^{w_i^2+w_i} \cdot w_i^3 \cdot \log(w_i) )$ per bag.
#### Root bag {#root-bag .unnumbered}
Let $X_r=\{x\}$ be the root of $T$. $\Mmatchset{r}$ consists of at most two matchings $\mathbf{v}(\Mmatchr) = [0]$ or $\mathbf{v}(\Mmatchr') = [1]$, where $\textbf{uf}(\Mmatchr)$ is an empty list and $\textbf{uf}(\Mmatchr')$ is defined by $c(x)=x$. It follows that the minimum number of unmatched nodes for any alternating cycle-free matching of $G$ is given by $m = \min \{ m(\Mmatchr) + 1 , m(\Mmatchr') \}$, and the maximum size of an alternating cycle-free matching is given by $(n - m)/2 $ where $n = |N|$ denotes the number of nodes of $G$.
#### Total Running Time {#total-running-time .unnumbered}
The total time complexity of the algorithm per bag is bounded above by the running time of the join bag. Since there is a linear number of bags, and since for every bag $X_i$ we have $|X_i| \leq tw(G) + 1 = w + 1$, the total time complexity of the algorithm described above is $$O( 4^{w^2+w} \cdot w^3 \cdot \log(w) \cdot n) .$$
Proof of Theorem \[thm:fpg\] {#app:fpg}
============================
Let $G$ be the spine of a simplicial triangulation of a $3$-manifold $\mathcal{T}$. If $\mathbf{tw}(\Gamma(\mathcal{T})) \leq k$, then $\mathbf{tw}(G) \leq 10k+9$.
Let $T$ be a tree decomposition of the dual graph, where each bag $X_i$ contains less or equal $k+1$ tetrahedra. We show how to construct a tree decomposition $T'$ of the spine of $\mathcal{T}$, modeled on the same underlying tree as $T$, in which each bag $X'_i$ contains less or equal $10(k+1)$ edges and triangles.
For each bag $X_i$ of $T$, we simply define the bag $X'_i$ to contain all edges and triangles of all tetrahedra in $X_i$. It remains to verify the three properties of a tree decomposition (Definition \[def:treewidth\]).
#### Node coverage {#node-coverage .unnumbered}
It is clear that every edge or triangle in the spine belongs to some bag $X'_i$, since every edge or triangle is contained in some tetrahedron $\delta$, which belongs to some bag $X_i$.
#### Arc coverage {#arc-coverage .unnumbered}
Consider some arc in the spine. This must join a triangle $t$ to an edge e that contains it. Let $\delta$ be some tetrahedron containing $t$; then $\delta$ contains both $t$ and e, and so if $X_i$ is a bag containing $\delta$ then the corresponding bag $X'_i$ contains the chosen arc in the spine (joining $t$ with e).
#### Coherence {#coherence .unnumbered}
Here we treat edges and triangles separately.
Let $t$ be some triangle in the simplicial complex. We must show that the bags containing $t$ correspond to a connected subgraph of the underlying tree. If $t$ is a boundary triangle, then $t$ belongs to a unique tetrahedron $\delta$, and the bags $X'_i$ that contain $t$ correspond precisely to the bags $X_i$ that contain $\delta$. Since the tree decomposition T satisfies the connectivity property, these bags correspond to a connected subgraph of the underlying tree. If $t$ is an internal triangle, then $t$ belongs to two tetrahedra $\delta_1$ and $\delta_2$, and the bags $X'_i$ that contain $t$ correspond to the bags $X_i$ that contain *either* $\delta_1$ or $\delta_2$. By the connectivity property of the original tree, the bags containing $\delta_1$ describe a connected subgraph of the tree, and so do the bags containing $\delta_2$. Furthermore, there is an arc in the dual graph from $\delta_1$ to $\delta_2$, and so some bag $X_i$ contains [*both*]{} $\delta_1$ and $\delta_2$. Thus the union of these two connected subgraphs is another connected subgraph, and we have established the connectivity property for $t$.
Now let $e$ be some edge of the simplicial complex. Again, we must show that the bags containing e correspond to a connected subgraph of the underlying tree. This is simply an extension of the previous argument. Suppose that e belongs to the tetrahedra $\delta_1,\ldots,\delta_m$ (ordered cyclically around e). Then for each $\delta_j$, the bags $X_i$ that contain $\delta_j$ describe a connected subgraph of the underlying tree, and the bags $X'_j$ containing e describe the union of these subgraphs, which we need to show is again connected. This follows because there is a sequence of arcs in the dual graph $(\delta_1,\delta_2)$, $(\delta_2,\delta_3)$ and so on; from the tree decomposition $T$ it follows that the subgraph for $\delta_1$ meets the subgraph for $\delta_2$, the subgraph for $\delta_2$ meets the subgraph for $\delta_3$, and so on. Therefore the union of these subgraphs is itself connected.
[^1]: School of Mathematics and Physics, The University of Queensland, Brisbane, Australia, `[email protected]`
[^2]: Department of Mathematics, Pontif[í]{}cia Universidade Cat[ó]{}lica, Rio de Janeiro, Brazil, `[email protected]`
[^3]: Department of Mathematics, Pontif[í]{}cia Universidade Cat[ó]{}lica, Rio de Janeiro, Brazil, `[email protected]`
[^4]: School of Mathematics and Physics, The University of Queensland, Brisbane, Australia, `[email protected]`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'It is often stated that homogeneity and isotropy of the Universe are assumptions of the almost Friedmann-Lemaître (FL) model (the hot big bang model), inspired from the Copernican Principle. However, only [*local*]{} homogeneity and isotropy are required by the model: multiply connected almost FL models are locally homogeneous and isotropic, but they can be [*globally*]{} anisotropic and/or [*globally*]{} inhomogeneous. Toy models are used here to show how global anisotropy and/or global inhomogeneity of an almost FL model could be shown directly in observations. This approach may avoid having to make any assumptions regarding global anisotropy and inhomogeneity.'
address:
- 'Inter-University Centre for Astronomy & Astrophysics, Post Bag 4, Ganeshkhind, Pune 411007, India'
- |
DARC, Observatoire de Paris-Meudon, 5, place Jules Janssen, Meudon Cedex, France\
[ email: [email protected]]{}
author:
- 'B. F. Roukema'
title: |
\
[ ]{}\
HOW TO AVOID THE AMBIGUITY IN APPLYING THE COPERNICAN PRINCIPLE FOR COSMIC TOPOLOGY: TAKE THE OBSERVATIONAL APPROACH
---
INTRODUCTION {#introduction .unnumbered}
============
The Copernican principle [@Coper43], defined for the present paper as the principle according to which the observer should not happen to be located in any special location in the Universe, is generally [*applied*]{} in one of three ways in twentieth century cosmology:
[()]{}
by assuming the Cosmological Principle, in which spatial sections of the Universe at constant cosmological time are close to homogeneous and isotropic when averaged on “large enough” length scales;
by assuming the Perfect Cosmological Principle, in which all of space-time is close to homogeneous and isotropic, on average; or
by assuming that the Universe is very chaotic, very far from homogeneity, but that by the Weak Anthropic Principle, we happen to live in a tiny bubble which is inflated to a locally close to flat Friedmann-Lemaître model;
or it is [*rejected*]{}
[()]{}
by taking the “observational” point of view that the observer lives at the centre of a close to isotropic but radially very inhomogeneous three-dimensional space-time cone with a spherical ($S^2$) boundary defined by the age of the local patch of the Universe, so that no space outside of this sphere exists as far as science is concerned, and which happens to coincide with what one expects in case (i) above.
Case (i) underlies what is generally accepted as the standard hot big bang model, which can be more formally referred to as the almost Friedmann-Lemaître model (e.g. @Schw00 [@Schw98; @deSitt17; @Fried23; @Fried97; @Lemait58; @Wein72]). This model may or may not have zero curvature, and it may or may not have trivial topology. Numerous observations now knit together to provide a very solid foundation for the almost Friedmann-Lemaître model, within the spatial, temporal and density limits of its validity.
Case (ii), that of a steady state or of a quasi-steady-state model [(@HBN93)]{}, is only defined in its general outline, and would require many more close links with recent observations in order to have predictive power. For example, the case (i) interpretation of quasars’ redshifts as primarily due to expansion of the Universe, when applied to observations, implies that quasars and Lyman break galaxies at high redshift trace large scale structure at $L\sim130\pm10$ (comoving) and that they function as a standard ruler which implies a low matter density $\Omm \sim 0.3$ [@RM00a; @RM00b; @BJ99]. This would be very difficult to explain if quasar (and Lyman break galaxy) redshifts were non-cosmological.
Case (iii) corresponds to the presently very popular chaotic inflation scenarios. Whether or not the Weak Anthropic Principle is consistent with the Copernican Principle is out of the scope of this paper.
Case (iii) is generally considered an extrapolation (towards short cosmological times and large spatial distances) of the almost Friedmann-Lemaître (FL) model. Case (ii) could also conceivably be defined as an extension of the FL model.
The anti-Copernican case (iv) is essentially identical to case (i) as far as observational analysis is concerned — as long as the topology of the Universe remains unmeasured. However, theoretical calculation is somewhat difficult if case (i) is not resorted to.
If the topology of space were measured by observation (@LR99, and see below), then the size of the Universe would be finite and less than the horizon diameter in at least one direction, i.e. finite. Models of case (ii) would then require serious revision and would no longer satisfy the Perfect Cosmological Principle, and case (iii) models would presumably require individual “universes” to be born as separate space-times and might have difficulty surviving Occam’s Razor.
Case (iv) would also be partly or totally avoidable: only the [*apparent*]{} space would be a sphere centred on the observer. At least some parts of the “edge of the Universe” of the case (iv) interpretation would no longer be edges.
On the other hand, whether or not case (i) would remain valid if the Universe is found to be multiply connected would depend on the resolution of the ambiguity in the meanings of the words “homogeneous” and “isotropic”. The solution of the Einstein-Hilbert equations is only found as a local solution, so only [*local*]{} homogeneity and isotropy are required. In general, although Friedmann-Lemaître models satisfy [*local*]{} homogeneity and isotropy, they need not satisfy [*global*]{} homogeneity and isotropy. Observational support for local homogeneity and isotropy does not imply global homogeneity and isotropy.
In this paper, the relevance of different interpretations of these two words in the context of the global geometry of space (curvature and topology) is explained and illustrated by simple toy models in two dimensions. By attempting to measure the more subtle elements of inhomogeneity and anisotropy which would reveal the global shape of the Universe, some of the ambiguity in deciding how to apply the Copernican principle to observational cosmology could be avoidable: empirical results sometimes help resolve theoretical dilemmas.
Proper distance (@Wein72, Eq. 14.2.21) in comoving units is used throughout this paper unless otherwise stated. The Hubble constant is parametrised here as $h\equiv H_0/100\,$km s$^{-1}$ Mpc$^{-1}.$ Values of the density parameter, $\Omm$, the dimensionless cosmological constant, $\Omega_\lambda$, the dimensionless curvature, $\Omega_\kappa\equiv \Omm + \Omega_\Lambda -1$, and the curvature radius $R_C \equiv (c/H_0) \Omega_\kappa^{-1/2},$ are indicated where used.
COSMIC TOPOLOGY, HOMOGENEITY AND ANISOTROPY {#cosmic-topology-homogeneity-and-anisotropy .unnumbered}
===========================================
Local Homogeneity and Isotropy {#local-homogeneity-and-isotropy .unnumbered}
------------------------------
When justifying the assumptions of (i) homogeneity and (ii) isotropy, these are generally interpreted to mean that
[()]{}
the comoving spatial density of astrophysical objects is approximately constant, and
the solid angular distribution of distant objects (or microwave background fluctuations) is approximately constant, after correction for the very anisotropic biases due to living in a dusty disk galaxy and due to our peculiar velocity with respect to the comoving reference frame,
respectively.
Both are reasonably consistent with numerous observations, and the latter has been further reinforced recently by the only cosmological class of objects known to be isotropic [*in spite*]{} of galactic absorption and our peculiar velocity: the gamma ray bursts. These may in fact reveal the births of black holes, i.e. of regions of extreme curvature (inhomogeneity) on very small scales, that are briefly visible before being engulfed by event horizons [@JDM00].
These two properties (a), (b) are generally referred to as “local homogeneity and isotropy”. Solutions to the Einstein-Hilbert equations are normally only found as local solutions of differential equations, and so only local assumptions of homogeneity and isotropy have any effect on Friedmann-Lemaître models.
One might also refer to these as first order homogeneity and isotropy, since only one-point statistics of density are considered.
Cosmic Topology and Global Anisotropy {#cosmic-topology-and-global-anisotropy .unnumbered}
-------------------------------------
For the earliest known article on cosmic topology, see @Schw00 [@Schw98]. A recent review is that of @LR99. The multiply connected models which have been compared with observations are all FL models, i.e. they are [*locally homogeneous and isotropic*]{}.
A simple way to illustrate multiply connected 3-manifolds which may correspond to the comoving 3-space we live in is to start with the 3-torus, $T^3$. This can be thought of
[()]{}
[*physically*]{}, as a cube of which opposite faces are identified so that space is continuous, with no boundaries, but finite,
[*observationally*]{}, as a tiling of Euclidean space $R^3$ by multiple copies of the single [*physical*]{} “cube”, or
as the analogy one dimension higher of a 2-torus embedded in $R^3$ but endowed with an intrinsically flat metric, i.e. a metric different to the normal metric of $R^3$.
For a demonstration of basic properties of locally homogeneous and isotropic manifolds, it is useful to subtract one dimension, i.e. to consider 2-manifolds. This is the approach adopted here for pedagogical purposes. For the physical Universe, the results extrapolate to the three-dimensional case.
The cube (i) is referred to as a [*fundamental domain*]{}, and the space (ii) is referred to as the [*universal covering space*]{} or in a new coinage, [*apparent space*]{}. Both are equivalent ways of thinking about the same manifold. The former is useful for conventional physical reasoning, according to which any particle or object only exists at one point in a space, but difficult to use for interpreting observations. The latter is useful for observational analyses, but surprising and sometimes confusing for physical reasoning, since objects “exist” many times in it.
In two dimensions, the corresponding space is the flat 2-torus $T^2$, of which the fundamental domain is a square (in general a parallelogram, but a square is used for the present paper), and the universal covering space is $R^2$.
Figure \[f-ptstor\] shows what such a universe would look like. For reference, the reader should remember that the horizon size for likely values of the curvature parameters $(\Omm\approx 0.3, \Omega_\Lambda\approx 0.7)$ is $R_H \approx 10$, and that objects like quasars or “Lyman break galaxies” are seen up to typical redshifts of $z\sim2-4$, which correspond to roughly half this distance. Clusters of galaxies and galaxies are mostly only seen to a few $100$.
Present constraints on the size of the Universe (defined here as the injectivity diameter $2\rinj$, which is the length of the shortest spatial geodesic linking an object to an image of itself) are $2\rinj \gtapprox 1$, whether the Universe is hyperbolic, flat or spherical. So, in order to see multiple images of a single object, catalogues of objects which are seen to typically a few tenths of $R_H$ would be required. This implies that effects of aging of the objects, or of viewing angles, or of movement relative to the comoving reference frame, would make multiple images of a single object appear quite different, so that proving the identity of two images would be difficult.
So, the identity between multiple copies of symbols in Fig. \[f-ptstor\] would not be obvious in any category of presently known objects.
The model is [*locally homogeneous and isotropic*]{}: any study in a small region around a point is independent of the orientation of the axes of a chosen coordinate system and of the origin of that system.
[*The model of Fig. \[f-ptstor\] is globally homogeneous, but globally anisotropic.*]{}
The anisotropy is shown by Fig. \[f-septor\], which shows vectors separating all pairs of images in the simulated universe. Separations between multiple images of single objects correspond to multiples of the generators \[the vectors $(0,1)$ and $(1,0)$\] and clearly cause clustering around these points. As pointed out by @LLL96, a one-dimensional pair separation histogram $\Delta N ( \Delta r )$ should show sharp spikes corresponding to these cluster points. (If no offsets were added between multiple images, then the spikes would be Dirac $\delta$ functions.)
The pairs of images contributing to these spikes can be referred to as [*generator pairs*]{} or as [*Type II pairs*]{} [(@ULL99)]{}, which induce spikes in flat multiply connected spaces, but not in curved multiply connected spaces [@Gomero99a; @ULL99].
If the $x-y$ axes are chosen at a different angle, then the diagram would be rotated: the pattern of spikes is not invariant (i.e. varies) under rotation. This proves global anisotropy.
However, the diagram would be statistically identical under a shift (without rotation) of the origin: the space is homogeneous.
Cosmic Topology and Global Inhomogeneity {#cosmic-topology-and-global-inhomogeneity .unnumbered}
----------------------------------------
### Use of “generator pairs” {#use-of-generator-pairs .unnumbered}
Many 3-manifolds are globally inhomogeneous, but the effect is observationally quite subtle.
The effect is illustrated here for the Klein bottle. This is like $T^2$, except that a twist is applied before identification of one pair of edges (here, the “vertical” edges). The distribution of images is shown in Fig. \[f-ptskle\]. Again, this model is [*locally homogeneous and isotropic*]{}.
Fig. \[f-sepkle\] shows that, as for the $T^2$ model, the Klein bottle is [*globally anisotropic*]{}: rotation of the coordinate axes would modify the pattern of crytallographic spikes.
In order to show global inhomogeneity using “generator pairs”,
[()]{}
knowledge of which pairs of images are due to multiple images of single objects is required, and
this knowledge is required [*despite the fact that the crystallographic spikes corresponding to “inhomogeneous generator pairs” are smeared out making them very difficult to detect.*]{}
Both (i) and (ii) are, of course, easy for a simulated universe: in the observed Universe they are difficult.
Comparison of Figs \[f-septor\] and \[f-sepkle\] shows that some crystallographic spikes seem to disappear due to the twist in identifying the vertical edges of the fundamental domain.
Figures \[f-sepklea\] and \[f-sepkleb\] show where these spikes have “disappeared” to, using property (i). Different choices of the $y$ origin (see figure captions) yield different patterns. Together, the two figures show how the spikes have been smeared out.
Hence, the Klein bottle is [*globally inhomogeneous*]{}.
One way to apply this technique observationally would be
[()]{}
to use properties of galaxies or quasars which uniquely label and identify individual objects, despite the large fractions of the age of the Universe over which the objects age.
This seems difficult given present observational limitations and known techniques, but not impossible.
An alternative method would be
[()]{}
firstly to detect the crystallographic spikes of generator pairs which [*do*]{} appear, and which would imply a fundamental domain of size $2L=2$ in the horizontal direction, and secondly to hypothesise that the real fundamental domain is just half this length.
This would make predictions of which pairs of objects should correspond to those seen separately in Figs \[f-sepklea\] and \[f-sepkleb\]. If the predicted multiply imaged pairs really were identical, then proving this by observation would be much easier than the “needle in a haystack” requirement of finding the corresponding pairs blindly as in the previous suggestion (a).
### The Copernican principle in an inhomogeneous universe {#the-copernican-principle-in-an-inhomogeneous-universe .unnumbered}
Figs \[f-sepklea\] and \[f-sepkleb\] illustrate one way in which the Copernican principle could be applied directly. One would not expect the observer to be at any special value of $y_0$: the value $y_0=0.00\pm0.01$ for the toy model shown here would be surprising.
### Use of “local pairs” or “local $n$-tuplets” {#use-of-local-pairs-or-local-n-tuplets .unnumbered}
Inspection of Fig. \[f-ptstor\] and a little reflection show that generator pairs are not the only pairs of images which occur repeatedly in a separation vector plot (or separation distance histogram).
Close pairs of images of [*non*]{}-identical objects within a single fundamental domain also define separation vectors which are repeated in the apparent space, though in a different way to generator pairs. These are [*local pairs*]{} or [*Type I*]{} pairs (@Rouk96; Section 2 of @ULL99). Since there are nine copies of the fundamental domain shown in Fig. \[f-ptstor\], there are nine local pairs clustered around each of the $20\ttimes 19/2=190$ separation vectors of modulus smaller than $L=1$. This is visible as a “graininess” in Figs \[f-septor\] and \[f-sepkle\].
In a flat space, the same effect also occurs for pairs across many copies of the fundamental domain, but in general does not occur in curved spaces. Thus the term “local”: it is most useful just to consider relatively close pairs of images.
@FagG99a noticed the effect of the local pairs in pair separation histograms, and @ULL99 showed how to collect these together and detect multiple connectedness. If these led to detection of multiple connectedness for a model known (mathematically) to be inhomogeneous, then this would enable (a) above to be applied: the generator pairs could be used to observationally illustrate inhomogeneity. However, local pairs in themselves would not show inhomogeneity: they are local.
Similarly, local $n$-tuplets [@Rouk96] would not directly show inhomogeneity. However, since non-orientable manifolds are inhomogeneous, the discovery of significant numbers of matching $n$-tuplets, where some of these required orientation reversals, would imply inhomogeneity.
For more details of global homogeneity and isotropy questions of 2-manifolds, and more particularly of 3-manifolds, see @Thur97.
CONCLUSION {#conclusion .unnumbered}
==========
The Copernican Principle, when used to motivate [*local*]{} homogeneity and isotropy of the Universe, does not necessarily imply [*global*]{} homogeneity and isotropy. If the topology of the Universe is observable, then this may reveal
[()]{}
that the Universe is globally anisotropic, though locally isotropic and both globally and locally homogeneous, as in the case of a $T^2$ model in two dimensions, or
that the Universe is globally both anisotropic and inhomogeneous, even though locally isotropic and homogeneous, as in the case of a Klein bottle model in two dimensions.
Most techniques for measuring cosmic topology should easily lead to direct geometrical illustrations of global anisotropy if multiple connectedness is detected (unless the 3-manifold we live is globally isotropic: the projective space, $S^3/Z_2$, which can be thought of by identifying all opposite points of the hypersphere $S^3$, is globally both isotropic and homogeneous, even though multiply connected).
By contrast, global inhomogeneity in a flat multiply connected manifold would be difficult to detect directly. However, prior knowledge of multiple connectedness would help in obtaining direct illustrations of global inhomogeneity.
Present observations imply that the observable Universe is close to flat [@Boom00a; @Maxima00a; @RM00a] and that multiply connected flat models of size $2\rinj \sim 2R_H/10$ are consistent with present large angle microwave background data [@Rouk00c]. The observational approach to global geometry might just possibly provide an empirical solution to an otherwise philosophical question, by showing whether not both local [*and*]{} global homogeneity and/or isotropy are valid assumptions..
, [Constraints on cosmological parameters from MAXIMA-1]{}, [*[Lett.]{}*]{}, , [arXiv:astro-ph/0005124]{}, 2000.
Broadhurst T., Jaffe A. H., Using the comoving maximum of the galaxy power spectrum to measure cosmological curvature. in the press, [ *[Lett.]{}, arXiv:astro-ph/9904348*]{}, 2000.
Copernicus, N., De revolvtionibvs orbium coelestium, libri VI., [*Norimbergae, apud Ioh. Petreium,*]{} 1543.
[de Sitter W., [Einstein’s theory of gravitation and its astronomical consequences. Third paper]{}, 3, 1917.]{}
Fagundes H. & Gausmann E., Cosmic Crystallography in Compact Hyperbolic Universes, in , CD-ROM version, arXiv:astro-ph/9811368, 1999.
Friedmann, A., , (Leningrad: Akademiya), 1923. Friedmann, A., L’Univers comme espace et temps (translation of Friedmann 1923), in , (Paris: Le Seuil/Sources du Savoir) 1997.
[Gomero G. I., Teixeira A. F. F., Rebouças M. J., Bernui A., Spikes in Cosmic Crystallography, , 1999.]{} [Hoyle F., Burbidge G., Narlikar J.V., [A quasi-steady state cosmological model with creation of matter]{}, 437, 1993.]{}
[Joshi P. S., Dadhich N. K., Maartens R., Gamma-ray bursts as the birth cries of black holes, , 2000.]{}
[Lange A. E. et al., First estimations of cosmological parameters from BOOMERANG, , 2000.]{} [Lehoucq R., Luminet J.-P., Lachièze-Rey M., [Cosmic crystallography]{}, 339, 1996.]{}
Lemaître G., La Structure et l’Evolution de l’Univers, , ed. Stoops R., (Brussels: Stoops), 1958. Luminet J.–P., Roukema B. F., Topology of the Universe: Theory and Observations, in , ed. Lachièze-Rey M., Dordrecht: Kluwer, p117, arXiv:astro-ph/9901364, 1999.
[Roukema B. F., On Determining the Topology of the Observable Universe via 3-D Quasar Positions, 1147, 1996.]{}
[Roukema B. F., A Counterexample to Claimed COBE Constraints on Compact Toroidal Universe Models, 3951, arXiv:astro-ph/0007140, 2000.]{}
[Roukema B. F., Mamon G. A., Tangential Large Scale Structure as a Standard Ruler: Curvature Parameters from Quasars, 395, arXiv:astro-ph/9911413, 2000.]{}
[Roukema B. F., Mamon G. A., Lifting cosmic degeneracy within a single data set, , arXiv:astro-ph/0010511, 2001.]{}
[Schwarzschild, K., [On the permissible curvature of space]{}, 337, 1900.]{} [Schwarzschild, K., [‘ “On the permissible curvature of space” by K Schwarzschild’]{}, 2539, 1998.]{} Thurston, W. P., , ed. Levy, S., Princeton, U.S.A.: Princeton University Press, 1997.
[Uzan J.–Ph., Lehoucq R. & Luminet J–P., A new crystallographic method for detecting space topology, 766, arXiv:astro-ph/9903155, 1999.]{}
Weinberg, S., , New York, U.S.A.: Wiley, 1972.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a simple yet effective approach for linking entities in queries. The key idea is to search sentences similar to a query from Wikipedia articles and directly use the human-annotated entities in the similar sentences as candidate entities for the query. Then, we employ a rich set of features, such as link-probability, context-matching, word embeddings, and relatedness among candidate entities as well as their related entities, to rank the candidates under a regression based framework. The advantages of our approach lie in two aspects, which contribute to the ranking process and final linking result. First, it can greatly reduce the number of candidate entities by filtering out irrelevant entities with the words in the query. Second, we can obtain the query sensitive prior probability in addition to the static link-probability derived from all Wikipedia articles. We conduct experiments on two benchmark datasets on entity linking for queries, namely the ERD14 dataset and the GERDAQ dataset. Experimental results show that our method outperforms state-of-the-art systems and yields 75.0% in F1 on the ERD14 dataset and 56.9% on the GERDAQ dataset.'
author:
- |
Chuanqi Tan$^\dag$ Furu Wei$^\ddag$ Pengjie Ren$^+$ Weifeng Lv$^\dag$ Ming Zhou$^\ddag$\
$^\dag$State Key Laboratory of Software Development Environment, Beihang University, China\
$^\ddag$Microsoft Research Asia $^+$Shandong University\
$^\dag$[[email protected]]{} $^+$[[email protected]]{}\
$^\ddag$[{fuwei, mingzhou}@microsoft.com]{} $^\dag$[[email protected]]{}
bibliography:
- 'emnlp2017.bib'
title: Entity Linking for Queries by Searching Wikipedia Sentences
---
Introduction
============
Query understanding has been an important research area in information retrieval and natural language processing [@croft2010query]. A key part of this problem is entity linking, which aims to annotate the entities in the query and link them to a knowledge base such as Freebase and Wikipedia. This problem has been extensively studied over the recent years [@Q15-1023; @usbeck2015gerbil; @cornolti2016piggyback].
The mainstream methods of entity linking for queries can be summed up in three steps: mention detection, candidate generation, and entity disambiguation. The first step is to recognize candidate mentions in the query. The most common method to detect mentions is to search a dictionary collected by the entity alias in a knowledge base and the human-maintained information in Wikipedia (such as anchors, titles and redirects) [@laclavik2014search]. The second step is to generate candidates by mapping mentions to entities. It usually uses all possible senses of detected mentions as candidates. Hereafter, we refer to these two steps of generating candidate entities as entity search. Finally, they disambiguate and prune candidate entities, which is usually implemented with a ranking framework.
There are two main issues in entity search. First, a mention may be linked to many entities. The methods using entity search usually leverage little context information in the query. Therefore it may generate many completely irrelevant entities for the query, which brings challenges to the ranking phase. For example, the mention “Austin” usually represents the capital of Texas in the United States. However, it can also be linked to “Austin, Western Australia”, “Austin, Quebec”, “Austin (name)”, “Austin College”, “Austin (song)” and 31 other entities in the Wikipedia page of “Austin (disambiguation)”. For the query “blake shelton austin lyrics”, Blake Shelton is a singer and made his debut with the song “Austin”. The entity search method detects the mention “austin” using the dictionary. However, while “Austin (song)” is most related to the context “blake shelton” and “lyrics”, the mention “austin” may be linked to all the above entities as candidates. Therefore candidate generation with entity search generates too many candidates especially for a common anchor text with a large number of corresponding entities. Second, it is hard to recognize entities with common surface names. The common methods usually define a feature called “link-probability” as the probability that a mention is annotated in all documents. There is an issue with this probability being static whatever the query is. We show an example with the query “her film”. “Her (film)” is a film while its surface name is usually used as a possessive pronoun. Since the static link-probability of “her” from all Wikipedia articles is very low, “her” is usually not treated as a mention linked to the entity “Her (film)”.
In this paper, we propose a novel approach to generating candidates by searching sentences from Wikipedia articles and directly using the human-annotated entities as the candidates. Our approach can greatly reduce the number of candidate entities and obtain the query sensitive prior probability. We take the query “blake shelton austin lyrics” as an example. Below we show a sentence in the Wikipedia page of “Austin (song)”.
**\[\[Austin (song)$\vert$Austin\]\]** is the title of a debut song written by David Kent and Kirsti Manna, and performed by American country music artist \[\[Blake Shelton\]\].
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: A sentence in the page “Austin (song)”.[]{data-label="e1"}
In the above sentence, the mentions “Austin” and “Blake Shelton” in square brackets are annotated to the entity “Austin (song)” and “Blake Shelton”, respectively. We generate candidates by searching sentences and thus obtain “Blake Shelton” as well as “Austin (song)” from this example. We reduce the number of candidates because many irrelevant entities linked by “austin” do not occur in returned sentences. In addition, as previous methods generate candidates by searching entities without the query information, “austin” can be linked to “Austin, Texas” with much higher static link-probability than all other senses of “austin”. However, the number of returned sentences that contain “Austin, Texas” is close to the number of sentences that contain “Austin (song)” in our system. We show another example with the query “her film” in Table \[e2\]. In this sentence, “Her”, “romantic”, “science fiction”, “comedy-drama” and “Spike Jonze” are annotated to corresponding entities. As “Her” is annotated to “Her (film)” by humans in this example, we have strong evidence to annotate it even if it is usually used as a possessive pronoun with very low static link-probability.
**\[\[Her (film)$\vert$Her\]\]** is a 2013 American \[\[romantic\]\] \[\[science fiction\]\] \[\[comedy-drama\]\] film written, directed, and produced by \[\[Spike Jonze\]\].
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: A sentence in the page “Her (film)”.[]{data-label="e2"}
We obtain the anchors as well as corresponding entities and map them to the query after searching similar sentences. Then we build a regression based framework to rank the candidates. We use a rich set of features, such as link-probability, context-matching, word embeddings, and relatedness among candidate entities as well as their related entities. We evaluate our method on the ERD14 and GERDAQ datasets. Experimental results show that our method outperforms state-of-the-art systems and yields 75.0% and 56.9% in terms of F1 metric on the ERD14 dataset and the GERDAQ dataset respectively.
Related Work
============
Recognizing entity mentions in text and linking them to the corresponding entries helps to understand documents and queries. Most work uses the knowledge base including Freebase [@chiu2014ntunlp], YAGO [@yosef2011aida] and Dbpedia [@olieman2014entity]. Wikify [@mihalcea2007wikify] is the very early work on linking anchor texts to Wikipedia pages. It extracts all n-grams that match Wikipedia concepts such as anchors and titles as candidates. They implement a voting scheme based on the knowledge-based and data-driven method to disambiguate candidates. @cucerzan2007large uses four recourses to generate candidates, namely entity pages, redirecting pages, disambiguation pages, and list pages. Then they disambiguate candidates by calculating the similarity between the contextual information and the document as well as category tags on Wikipedia pages. @milne2008learning generate candidates by gathering all n-grams in the document, and retaining those whose probability exceeds a low threshold. Then they define commonness and relatedness on the hyper-link structure of Wikipedia to disambiguate candidates.
The work on linking entities in queries has been extensively studied in recent years. TagME [@ferragina2010tagme] is a very early work on entity linking in queries. It generates candidates by searching Wikipedia page titles, anchors and redirects. Then disambiguation exploits the structure of the Wikipedia graph, according to a voting scheme based on a relatedness measure inspired by @milne2008learning. The improved version of TagME, named WAT [@piccinno2014tagme], uses Jaccard-similarity between two pages’ in-links as a measure of relatedness and uses PageRank to rank the candidate entities. Moreover, Meij [-@meij2012adding] proposes a two step approach for linking tweets to Wikipedia articles. They first extract candidate concepts for each n-gram, and then use a supervised learning algorithm to classify relevant concepts.
Unlike the work which revolves around ranking entities for query spans, the Entity Recognition and Disambiguation (ERD) Challenge [@carmel2014erd] views entity linking in queries as the problem of finding multiple query interpretations. The SMAPH system [@cornolti2014smaph] which wins the short-text track works in three phases: fetching, candidate-entity generation and pruning. First, they fetch the snippets returned by a commercial search engine. Next, snippets are parsed to identify candidate entities by looking at the boldfaced parts of the search snippets. Finally, they implement a binary classifier using a set of features such as the coherence and robustness of the annotation process and the ranking as well as composition of snippets. They further extend SMAPH-1 to SMAPH-2 [@cornolti2016piggyback]. They use the annotator WAT to annotate the snippets of search results to generate candidates and joint the additionally link-back step as well as the pruning step in the ranking phase, which gets the state-of-the-art results on the ERD14 dataset and their released dataset GERDAQ. There is another work closed to SMAPH that uses information of query logs and anchor texts [@blanco2015fast], which gives a ranked list of entities and is evaluated by means of typical ranking metrics.
Our work is different from using search engines to generate candidates. We firstly propose to search Wikipedia sentences and take advantage of human annotations to generate candidates. The previous work, such as SMAPH, employs search engine for candidate generation, which puts queries in a larger context in which it is easier to make sense of them. However, it uses WAT, an entity search based tool, to pre-annotate the snippets for candidate generation, which falls back the issues of entity search.
Our Approach
============
As shown in Figure 1, we introduce our approach with the query “blake shelton austin lyrics”. Our approach consists of three main phases: sentence search, candidate generation, and candidate ranking. First, we search the query in all Wikipedia articles to obtain the similar sentences. Second, we extract human-annotated entities from these sentences. We keep the entities whose corresponding anchor texts occur in the query as candidates, and treat others as related entities. Specifically, we obtain three candidates in this example, namely “Blake Shelton”, “Austin, Texas”, and “Austin (song)”. Finally, we use a regression based model to rank the candidate entities. We get the final annotations of “Blake Shelton” and “Austin (song)” whose scores are higher than the threshold selected on the development set. In the following sections, we describe these three phases in detail.
Sentence Search
---------------
Sentences in Wikipedia articles usually contain anchors linking to entities. We are therefore motivated to generate the candidate entities based on the sentence search instead of the common method using entity search. There are some issues in the original annotations because of the annotation regulation. First, entities in their own pages are usually not annotated. Thus we annotate these entities with matching between the text and the page title. Second, entities are usually annotated only in their first appearance. We annotate these entities if they are annotated in previous sentences in the page. Moreover, pronouns are widely used in Wikipedia sentences and are usually not annotated. We use the Stanford CoreNLP toolkit [@manning-EtAl:2014:P14-5] to do the coreference resolution. In addition, we use the content in the disambiguation page and the infobox. Although these two kinds of information may have incomplete grammatical structure, it contains enough context information for the sentence search in our task.
We use the Wikipedia snapshot of May 1, 2016, which contains 4.45 million pages and 120 million sentences. We extract sentences that contain at least one anchor in the Wikipedia articles, and extract human-annotated anchors as well as corresponding entities in the sentences. The original annotation contains 82.6 million anchors. We obtain 110 million annotated anchors in 48.4 million sentences after the incrementally annotation. All of above annotations are indexed by Lucene[^1] by building documents consisting of two fields: the first one contains the sentence and the second one contains all anchors with their corresponding entities. For each query, we search it with Lucene using its default ranker based on the vector space model and tf-idf to obtain the top K sentences (K is selected on the development set). We extract all entities as the related entities and use these sentences as their support sentences.
Candidate Generation
--------------------
We back-map anchors and corresponding entities extracted in sentences to generate candidates. We use $(a, e)$ to denote the pair of the anchor text and corresponding entity and use $w(a, e)$ to denote the number of sentences containing the pair $(a, e)$. Then, we prune the candidate pairs according to following rules.
First, we only keep the pair whose corresponding anchor text $a$ occurs in the query as a candidate, which has been used in previous work [@ferragina2010tagme]. Second, we follow the long-string match strategy. If we have two pairs $(a_1,e_1)$ and $(a_2,e_2)$ while $a_1$ is a substring of $a_2$, we drop $(a_1,e_1)$ if $w(a_1,e_1) < w(a_2,e_2)$. This is because $a_2$ is typically less ambiguous than $a_1$. For example, for the query “mesa community college football”, we can obtain the anchor “mesa”, “college”, “community college”, and “mesa community college”. We only keep “mesa community college” because it is longest and occurs most times in returned sentences. However, if $w(a_1,e_1) > w(a_2,e_2)$, we keep both candidate pairs because $a_1$ is more common in the query.
In addition, we keep the entity whose surface form is the same with the anchor text and prune others. If we have two pairs $(a,e_1)$ and $(a,e_2)$ with the same anchor, and only $e_2$ occurs in the query, we drop the pair $(a,e_1)$ if $w(a,e_1) < w(a,e_2)$. For example, for the query “business day south africa”, the anchor “south africa” can be linked to “south africa”, “union of south africa”, and “south africa cricket team”. We only keep the entity “south africa”.
Candidate Ranking
-----------------
We build a regression based framework to rank the candidate entities. In the training phase, we treat the candidates that are equal to the ground truth as the positive samples and the others as negative samples. The regression object of the positive sample is set to the score 1.0. The negative sample is set to the maximum score of overlapping ratio of tokens between its text and each gold answer. The regression object of the negative sample is not simply set to 0 in order to give a small score if the candidate is very closed to the ground truth. We find it benefits the final results. We use LIBLINEAR [@fan2008liblinear] with L2-regularized L2-loss support vector regression to train the regression model. The object function is to minimize $$w^Tw/2 + \mathnormal{C} \sum \max(0, |y_i-w^Tx_i|-eps)^2$$ where $x_i$ is the feature set, $y_i$ is the object score and $w$ is the parameter to be learned. We follow the default setting that $C$ is set to 1 and $eps$ is set to 0.1.
In the test phase, each candidate gets a score of $w^Tx_i$ and then we only output the candidate whose score is higher than the threshold selected on the development set.
We employ four different feature sets to capture the quality of a candidate from different aspects. All features are shown in Table \[table:Feature\].
Context-Independent Features
: This feature set measures each annotation pair $(a,e)$ without context information. Feature 1-4 catch the syntactic properties of the candidate. Feature 5 is the number of returned sentences that contain $(a,e)$. Feature 6 is the maximum search score (returned by Lucene) in its support sentences. Moreover, inspired by TagME [@ferragina2010tagme], we denote $freq(a)$ as the number of times the text $a$ occurs in Wikipedia. We use $link(a)$ to denote the number of times the text $a$ occurs as an anchor. We use $lp(a) = link(a)/freq(a)$ to denote the static link-probability that an occurrence of $a$ has been set as an anchor. We use $freq(a,e)$ to denote the number of times that the anchor text $a$ links to the entity $e$, and use $pr(e|a)$ = $freq(a,e)/link(a)$ to denote the static prior-probability that the anchor text $a$ links to $e$. Features 7 and 8 are these two probabilities.
Context-Matching Features
:
ID Name Description
---- --------------- ------------------------------------------------------------------------------------
1 $in\_query$ 1 if $e$ is in the query, 0 otherwise
2 $is\_pt$ 1 if $e$ contains parenthesis, 0 otherwise
3 $is\_cm$ 1 if $e$ contains comma, 0 otherwise
4 $len$ len($e$) by tokens
5 $w(a,e)$ number of support sentences
6 $sc(a,e)$ maximum search score of support sentences
7 $lp(a)$ static link-probability that $a$ is an anchor
8 $pr(a,e)$ static prior-probability that $a$ links to $e$
9 $cm\_sc$ context matching score to the support sentences
10 $cm\_fs$ context matching score to the first sentence of $e$’s page
11 $cm\_dd$ context matching score to the description in $e$’s disambiguation page
12 $embed\_sc$ maximum embedding similarity of the query and each support sentence
13 $embed\_fs$ embedding similarity of the query and the first sentence of $e$’s page
14 $embed\_dd$ embedding similarity of the query and the description in $e$’s disambiguation page
15 $rel\_cd\_sc$ number of candidates that occur in the support sentences
16 $rel\_cd\_sp$ number of candidates that occur in the same Wikipedia page
17 $rel\_re\_sc$ number of related entities that occur in the support sentences
18 $rel\_re\_sp$ number of related entities that occur in the same Wikipedia page
: Feature Set for Candidate Ranking[]{data-label="table:Feature"}
We treat the other words except for the anchor text as the context. This feature set measures the context matching to the query. Feature 9 is the context matching score calculated by tokens. We denote $c$ as the set of context words. For each $c_i$ in $c$, the $cm\_sc(c_i)$ is the ratio of times that $c_i$ occurs in the support sentences, and $cm\_sc(c) = \frac{1}{N}\sum{cm\_sc(c_i)}$. Features 10 and 11 are the ratio of context words occurring in the first sentence in the entity page and the description of entity’s disambiguation page (if existed), respectively. Moreover, we train a 300-dimensional word embeddings on all Wikipedia articles by word2vec [@mikolov2013efficient] and use the average embedding of each word as the sentence representation. Feature 12 is the maximum cosine score between the query and each support sentence. Features 13 and 14 are calculated with the first sentence in the entity’s page and the description in the disambiguation page.
Relatedness Features of Candidate Entities
: This set of features measures how much an entity is supported by other candidates. Feature 15 is the number of other candidate entities occurring in the support sentences. Feature 16 is the number of candidate entities occurring in the same Wikipedia page with the current entity.
Relatedness Features to Related Entities
: This set of features measures the relatedness between candidates and related entities outside of queries. Related entities can provide useful signals for disambiguating the candidates. Features 17 and 18 are analogous features with features 15 and 16, which are calculated by the related entities.
Experiment
==========
We conduct experiments on the ERD14 and GERDAQ datasets. We compare with several baseline annotators and experimental results show that our method outperforms the baseline on these two datasets. We also report the parameter selection on each dataset and analyze the quality of the candidates using different methods.
Dataset
-------
ERD14[^2]
: is a benchmark dataset in the ERD Challenge [@carmel2014erd], which contains both long-text track and short-text track. In this paper we only focus on the short-text track. It contains 500 queries as the development set and 500 queries as the test set. Due to the lack of training set, we use the development set to do the model training and tuning. This dataset can be evaluated by both Freebase and Wikipedia as the ERD Challenge Organizers provide the Freebase Wikipedia Mapping with one-to-one correspondence of entities between two knowledge bases. We use Wikipedia to evaluate our results.
GERDAQ[^3]
: is a benchmark dataset to annotate entities to Wikipedia built by @cornolti2016piggyback. It contains 500 queries for training, 250 for development, and 250 for test. The query in this dataset is sampled from the KDD-Cup 2005 and then annotated manually. Both name entities and common concepts are annotated in this dataset.
Evaluation Metric
-----------------
We use average F1 designed by ERD Challenge [@carmel2014erd] as the evaluation metrics. Specifically, given a query q, with labeled entities $\hat{A}=\{\hat{E_1},\dots,\hat{E_n}\}$. We define the F-measure of a set of hypothesized interpretations $A = \{E_1,\dots,E_m\}$ as follows: $$\begin{aligned}
&Precision = \frac{\vert\hat{A} \cap A\vert}{\vert A \vert}, Recall = \frac{\vert\hat{A} \cap A\vert}{\vert \hat{A} \vert} \\
&F_1 = \frac{2 \times Precision \times Recall}{Precision + Recall}
\end{aligned}$$ The average F1 of the evaluation set is the average of the F1 for each query: $$Average F_1 = \frac{1}{N}\sum_{i=1}^{N}F_1(q_i)$$ Following the evaluation guideline in ERD14 and GERDAQ, we define recall to be 1.0 if the gold binding of a query is empty and define precision to be 1.0 if the hypothesized interpretation is empty.
Baseline Methods
----------------
We compare with several baselines and use the results reported by the ERD organizer and @cornolti2016piggyback.
AIDA
: [@hoffart2011robust] searches the mention using Stanford NER Tagger based on YAGO2. We select AIDA as a representative system aiming to entity linking for documents following the work in @cornolti2016piggyback.
WAT
: [@piccinno2014tagme] is the improved version of TagME [@ferragina2010tagme].
Magnetic IISAS
: [@laclavik2014search] retrieves the index extracted from Wikipedia, Freebase and Dbpedia. Then it exploits Wikipedia link graph to assess the similarity of candidate entities for disambiguation and filtering.
Seznam
: [@eckhardt2014entity] uses Wikipedia and DBpedia to generate candidates. The disambiguation step is based on PageRank over the graph.
NTUNLP
: [@chiu2014ntunlp] searches the query to match Freebase surface forms. The disambiguation step is built on top of TagME and Wikipedia.
SMAPH-1
: [@cornolti2014smaph] is the winner in the short-text track in the ERD14 Challenge.
SMAPH-2
: [@cornolti2016piggyback] is the improved version of SMAPH-1. It generates candidates from the snippets of search results returned by the Bing search engine.
System $F1_{avg}$
--------------------------- ------------
AIDA 22.1
WAT 58.6
Magnetic IISAS 65.6
Seznam 66.9
NTUNLP 68.0
SMAPH-1 68.8
SMAPH-2 70.8
Our work **75.0\***
w/o Spell Check 74.0
w/o Additional Annotation 74.4
w/o Context Feature 72.6
w/o Relatedness Feature 74.5
: Results on the GERDAQ dataset. Results of the baseline systems are taken from Table 10 in @cornolti2016piggyback.[]{data-label="table:GERDAQ_Reslut"}
System $P_{avg}$ $R_{avg}$ $F1_{avg}$
--------------------------- ----------- ----------- ------------
AIDA 94.0 12.2 12.6
TagME 60.4 51.2 44.7
WAT 49.6 57.0 46.0
SMAPH-1 77.4 54.3 52.1
SMAPH-2 72.1 55.3 54.4
Our work 71.5 58.5 **56.9**
w/o Spell Check 75.4 48.6 49.3
w/o Additional Annotation 70.3 58.2 55.8
w/o Context Feature 69.2 56.4 55.5
w/o Relatedness Feature 73.3 57.4 56.7
: Results on the GERDAQ dataset. Results of the baseline systems are taken from Table 10 in @cornolti2016piggyback.[]{data-label="table:GERDAQ_Reslut"}
Result
------
We report results on the ERD datset and GERDAQ dataset in Table \[table:ERD\_Reslut\] and Table \[table:GERDAQ\_Reslut\], respectively. On the ERD14 dataset, WAT is superior to AIDA but it is still up to $10\%$ than SMAPH-1 that wins the ERD Challenge. SMAPH-2 improves $2\%$ than SMAPH-1. Our system significantly outperforms the state-of-the-art annotator SMAPH-2 by $4.2\%$. On the GERDAQ dataset, our system is $2.5\%$ superior to the state-of-the-art annotator SMAPH-2. The F1 score in this dataset is much lower than the ERD dataset because common concepts such as “Week” and “Game” that are not annotated in the ERD dataset are annotated in the GERDAQ dataset.
Spell checking has been widely used in the baseline annotators as it is not uncommon in queries [@laclavik2014search]. The SMAPH system that generates candidates by search results implicitly leverages the spell-checking embedded in search engines. In our experiments, spell checking improves 1.0% on the ERD dataset and 7.6% on the GERDAQ dataset. Furthermore, only $6.9\%$ of queries in the ERD14 dataset have spelling mistakes, whereas the number in the GERDAQ dataset is $23.0\%$. Thus spell-checking is more important in the GERDAQ dataset.
The result decreases 0.6% on the ERD dataset and 1.1% on the GERDAQ dataset without the additional annotation. Furthermore, while the F1 score decreases 2.4% on the ERD dataset and 1.4% on the GERDAQ dataset without the context features, the score only decreases 0.5% on the ERD dataset and 0.2% on the GERDAQ dataset without the relatedness features. Unlike the work on entity linking for documents [@eckhardt2014entity; @witten2008effective] that features derived from entity relations get promising results, the context features play a more important role than the relatedness features on entity linking for queries as search queries are short and contain fewer entities than documents.
![F1 scores with different search numbers and thresholds on the GERDAQ development set[]{data-label="fig:para_gerdaq"}](parameter_erd.pdf){width="2.94in"}
![F1 scores with different search numbers and thresholds on the GERDAQ development set[]{data-label="fig:para_gerdaq"}](parameter_gerdaq.pdf){width="2.94in"}
Parameter Selection
-------------------
There are two parameters in our framework, namely the number of search sentences and the threshold for final output. We select these two parameters on the development set. We show the F1 score with different numbers of search sentences and thresholds in Figure \[fig:para\_erd\] and Figure \[fig:para\_gerdaq\]. On the ERD development set, better results occur in the search number between 600 and 800 as well as the threshold 0.55 and 0.6. On the GERDAQ development set, better results occur in the search number between 700 and 1000 as well as the threshold between 0.45 and 0.5. In our experiment, we set the number of sentences to 700 and the threshold to 0.56 on the ERD dataset as well as 800 and 0.48 on the GERDAQ dataset according to the F1 scores on the development set.
----------------- ----------- ------------ -------
Number of Number of
anchors candidates
Entity Search 66.46
ES + RF 69.00
Sentence Search 73.81
SS + RF 75.01
----------------- ----------- ------------ -------
: Comparison with different candidate generation methods on the ERD dataset. +RF: integrating ranking features extracted by Sentence Search.[]{data-label="table:number_of_candidates"}
Method $C_{avg}$ $P_{avg}$ $R_{avg}$
----------------- ----------- ----------- -----------
Entity Search 78.87 77.56 66.04
Sentence Search 74.42 89.61 69.08
: Results for the 398 queries which have at least one labeled entity on the ERD dataset using different candidate generation methods. $C_{avg}$ is the average recall of candidates per query. $P_{avg}$ and $R_{avg}$ are calculated on the final results.[]{data-label="table:coverage_threshold"}
Model Analysis
--------------
The main difference between our method and most previous work is that we generate candidates by searching Wikipedia sentences instead of searching entities. For generating candidates with entity search, we build a dictionary containing all anchors, titles, and redirects in Wikipedia. Then we query the dictionary to get the mention and obtain corresponding entities as candidates. We use the same pruning rules and ranking framework in our experiments, but exclude the features from support sentences because the entity search method does not contain the information. The F1 score is shown in Table \[table:number\_of\_candidates\]. We achieve similar results in our implementation of the method using entity search on the ERD dataset as Magnetic IISAS [@laclavik2014search] which uses a similar method and ranks 4th with the F1 of 65.57 in the ERD14 Challenge.
We compare the two candidate generation methods in several aspects. First, we show the overall results in Table \[table:number\_of\_candidates\]. The average number of candidates from our method is much smaller. It is noted that the anchors from sentence search can also be found in entity search. However, we only extract the entities in the returned sentences while the methods by entity search use all entities linked by the anchors. In addition, features such as the number of sentences containing the entity from sentence search which provide query sensitive prior probability contribute to the ranking process. It improves the F1 score from 73.81 to 75.01 for sentence search and from 66.46 to 69.00 for entity search. More important, the result of “ES+RF” is still significantly worse than the result of both small candidate set and Wikipedia related features that prunes irrelevant candidates at the beginning, which proves that the high-quality candidate set is very important since the larger candidate set brings in lots of noise in training a ranking model. Moreover, there are 102 queries (20.4%) without labeled entities in the ERD dataset. We only give 7 incorrect annotations in these queries while the number is 13 from entity search. Furthermore, as shown in Table \[table:coverage\_threshold\], the coverage of our method is lower in queries with at least one entity, but we obtain better results on precision, recall and F1 in the final stage.
![F1 scores with number of candidates using different methods on the ERD dataset. The number of queries is shown in the parentheses.[]{data-label="fig:analysis"}](cropped_ana_1.pdf){width="3in"}
Figure \[fig:analysis\] illustrates the F1 score grouped by the number of candidates using entity search. In almost all columns the F1 score of our method is better than the baseline. In left columns (the number of candidates is less than 10), both methods generate few candidates. The F1 score of our method is higher, which proves that we train a better ranking model because of our small but quality candidate set. Moreover, the right columns (the number of candidates is more than 10) show that the F1 score using entity search gradually decreases with the incremental candidates. However, our method based on sentence search takes advantage of context information to keep a small set of candidates, which keeps a consistent result and outperforms the baseline.
Conclusion
==========
In this paper we address the problem of entity linking for open-domain queries. We introduce a novel approach to generating candidate entities by searching sentences in the Wikipedia to the query, then we extract the human-annotated entities as the candidates. We implement a regression model to rank these candidates for the final output. Two experiments on the ERD dataset and the GERDAQ dataset show that our approach outperforms the baseline systems. In this work we directly use the default ranker in Lucene for similar sentences, which can be improved in future work.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Ming-Wei Chang for sharing the ERD14 dataset. The first author and the fourth author are supported by the National Natural Science Foundation of China (Grant No. 61421003).
[^1]: <http://lucene.apache.org>
[^2]: <http://web-ngram.research.microsoft.com/erd2014/Datasets.aspx>
[^3]: <http://acube.di.unipi.it/datasets>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The effect of a newly born star cluster inside a giant molecular cloud (GMC) is to produce a hot bubble and a thin, dense shell of interstellar gas and dust swept up by the H II expansion, strong stellar winds, and repeated supernova explosions. Lying at the inner side of the shell is the photodissociation region (PDR), the origin of much of the far-infrared/sub-millimeter/millimeter (FIR/sub-mm/mm) radiation from the interstellar medium (ISM). We present a model for the expanding shell at different stages of its expansion which predict mm/sub-mm and far-IR emission line intensities from a series of key molecular and atomic constituents in the shell. The kinematic properties of the swept-up shell predicted by our model are in very good agreement with the measurements of the supershell detected in the nearby starburst galaxy M 82. We compare the modeling results with the ratio-ratio plots of the FIR/sub-mm/mm line emission in the central 1.0 kpc region to investigate the mechanism of star forming activity in M 82. Our model has yielded appropriate gas densities, temperatures, and structure scales compared to those measured in M 82, and the total H$_2$ content is compatible with the observations. This implies that the neutral ISM of the central star-forming region is a product of fragments of the evolving shells.'
author:
- Lihong Yao
- 'T. A. Bell, S. Viti, J. A. Yates'
- 'E. R. Seaquist'
title: 'Starburst Models For FIR/sub-mm/mm Line Emission. I. An Expanding Supershell Surrounding A Massive Star Cluster'
---
Introduction {#intro}
============
Starburst is a phenomenon when the star formation rate (SFR) cannot be sustained for the lifetime of the galaxy. It is now clear that active star formation or starburst activity is common throughout the universe [@mad98]. The bursts of massive star formation can dramatically alter the structure and evolution of their host galaxies by injecting large amounts of energy and mass into the ISM via strong stellar winds and repeated supernova explosions. The evolution of the superbubbles and supershells that have sizes ranging from several tens to hundreds of parsec plays an important role in understanding the amount and distribution of warm gas in the ISM. Furthermore, understanding the characteristics of starbursts and their relationship with the ISM, as well as to be able to parametrize the star formation history are crucial in understanding the galaxy evolution.
In the past, several models have been used to interpret the infrared, sub-millimeter, millimeter line observations of neutral gas in the central regions of nearby starburst galaxies [e.g. @mao00; @saf00; @wil92 and references therein]. These include the large velocity gradient (LVG) model [@gak74], the steady-state PDR model [@tah85], and the inhomogeneous radiative transfer model taking into account non local thermodynamic equilibrium (non-LTE) [@wil92]. These have revealed that the physical conditions (such as gas density, FUV flux, and gas kinetic temperature) are enhanced in starburst regions. However, none of these models are able to relate the observed line emission properties of the neutral gas in a starburst galaxy to its age and star formation history. In this paper, we introduce an evolving starburst model for FIR/sub-mm/mm line emission in gas media that allows us to ultimately achieve this goal.
Our model consists of a standard dynamical model of the bubble/shell structure around a young star cluster (see Fig. \[shell\]), which has been described in many publications [e.g. @cmw75; @wea77; @mak87; @fra90; @kam92], a time-dependent stellar population synthesis model [@lei99], a fully time-dependent PDR chemistry model [@bel05], and a one-dimensional non-LTE line radiative transfer model [@ray01]. In this paper, we conduct a preliminary study using this set of models. We first describe the methodology of our model (Section 2). We then follow the evolution of a GMC and a swept-up shell induced by massive star formation at the center, and calculate the dynamics, thermal structure, and the line radiative transfer of the selected molecular and atomic species in the expanding shell (Section 3). We compare our modeling results with the observations of the expanding supershell and average gas properties in the central 1.0 kpc region of the nearby starburst galaxy M 82 (Section 4). Finally, we present the conclusions of this study (Section 5).
The basic assumptions for our evolving starburst model are (1) star formation occurs primarily within the dense optically thick spherical cloud [e.g. @gao01], and that all stars form instantaneously in a compact spherical cluster located at the center of the cloud (the star cluster is therefore treated as a point source), and (2) the starlight produced by the central cluster is completely absorbed and reprocessed by the dust in the expanding shell [@efs00]. A summary of our evolving starburst model is presented in Table \[tbl-1\].
Starburst Models For Gas Media {#model}
==============================
The evolution of a giant molecular cloud is determined by H II expansion in the very early stage ($t$ $<$ 10$^5$ yr), when a hot bubble surrounded by a thin dense shell structure is created. The later evolution is driven by the strong stellar winds and repeated supernova explosions. We assume that repeated supernova explosions behave like a steady isotropic stellar wind injected to the bubble. The hot bubble will eventually cool, and the swept-up shell will stall after a few times 10$^7$ yr. The stars in the young cluster located at the center of the GMC are assumed to have masses between 0.1 M$_{\odot}$ and 120 M$_{\odot}$. The Salpeter initial mass function $dN/dm_{\ast}$ $\propto$ $m^{-2.35}_{\ast}$ [IMF; @sal55] is adopted in this study. A top-heavy IMF, which has an excess of stars in the mass range 10 - 20 M$_{\odot}$ over stars of 5 M$_{\odot}$ or less for starburst galaxies [e.g. @rie80], will be investigated in future work.
Shell Dynamics {#dyn}
--------------
The radius and velocity of the [*H II Expansion*]{} due to ionization can be written as [@spi78; @fra90],
$$\begin{aligned}
R_{HII}(t) & = & R_S\Big(1 + \frac{7}{4}\frac{c_i t}{R_S}\Big)^{\frac{4}{7}}, \\
V_{HII}(t) & = & c_i \Big(1 + \frac{7}{4}\frac{c_i t}{R_S}\Big)^{-\frac{3}{7}}\end{aligned}$$
where $R_S$ is the initial Strömgren radius in pc, and $c_i$ $\simeq$ 11.5 km s$^{-1}$ is the sound speed in the ionized gas with an equilibrium temperature of $\sim$ 10$^4$ K.
Almost as soon as the initial Strömgren sphere is formed, the strong winds start to impart large amounts of mechanical energy into the bubble. About 96% of the total wind energy is generated by stars with masses $>$ 30 $M_\odot$ [@mak87]. The size of the hot bubble is assumed to be much larger than the thickness of the swept-up shell, therefore the radius and velocity of the shell in the [*Winds*]{} phase can be written as [@mak87],
$$\begin{aligned}
R_w(t) & = & 269.0 \Big(\frac{L_{38}}{n}\Big)^{\frac{1}{5}} (t_7)^{\frac{3}{5}}, \\
V_w(t) & = & 16.1 \Big(\frac{L_{38}}{n}\Big)^{\frac{1}{5}} (t_7)^{-\frac{2}{5}} \end{aligned}$$
where $L_{38}$ = $L_w$/(10$^{38}$ ergs s$^{-1})$, $L_w$ is the wind mechanical luminosity, $L_w$ = $\int_{m_{1}}^{m_{2}} C_w C_m m^{\gamma - 2.35}_{\ast}
dm_{\ast}$, $t_7$ = $t$/(10$^7$ yr), $n$ is the ambient gas density in cm$^{-3}$, $m_{1}$ = 0.1 M$_{\odot}$, $m_{2}$ = 120 M$_{\odot}$, $C_w$ = 1.0 $\times$ 10$^{29}$, $C_m$ = 429.0, and $\gamma$ = 3.7 (derived from Abbott 1982). The main-sequence lifetime of the most massive star (120 M$_{\odot}$) in the star cluster is about 7.0 $\times$ 10$^5$ yr [@mam88]. After this time, we assume that the winds-equivalent energy produced by the first supernova and the subsequent ones drives the further expansion of the swept-up shell. The radius and velocity of the shell in the [*Supernova*]{} phase can be written as,
$$\begin{aligned}
R_{SN}(t) & = & 97.0 \Big(\frac{N_{\ast} E_{51}}{n}\Big)^{\frac{1}{5}} \Big[(t_7)^{\frac{3}{5}} - \Big(\frac{t_{1stSN}}{10^7}\Big)^{\frac{3}{5}}\Big] + R_w(t_{1stSN}), \\
V_{SN}(t) & = & 5.7 \Big(\frac{N_{\ast} E_{51}}{n}\Big)^{\frac{1}{5}} (t_7)^{-\frac{2}{5}}\end{aligned}$$
where $N_{\ast}$ is the number of stars with masses $\ge$ 8 M$_{\odot}$ in the cluster, $E_{51} = E_{SN}$/(10$^{51}$ ergs s$^{-1})$, $E_{SN}$ is the energy produced by each supernova explosion, $t_{1stSN}$ is the time when the first supernova occurs in the star cluster, and $R_w$($t_{1stSN}$) is the shell radius at $t_{1stSN}$ calculated from Equation (2a). The average rate of supernova explosions is $\sim$ 6.3 $\times$ 10$^{35}$ $N_{\ast} E_{51}$ ergs s$^{-1}$ [@mak87]. When the energy produced by the stellar winds and/or supernova explosions is much greater than the radiative losses, the bubble is adiabatic. This [*Adiabatic*]{} phase persists until the radiative cooling becomes important for the hot bubble at $t_c$,
$$t_{c} = 4 \times 10^6 \mathcal{Z}^{-1.5} (N_{\ast} E_{51})^\frac{3}{10} n^{-\frac{7}{10}}$$
where $\mathcal{Z}$ is the metallicity with respect to the solar. After $t_c$, the expansion of the bubble is no longer energy-driven, but momentum-driven. This momentum-driven phase is characterized as the snow-plow (SP) phase. For simplicity, we ignore the momentum deposition in the shell by SN ejecta [@mak87]. Hence, the radius and velocity of the shell in the [*Snow-plow*]{} phase can be written as,
$$\begin{aligned}
R_{SP}(t) & = & R_{c} \Big(\frac{t}{t_{c}}\Big)^\frac{1}{4}, \\
V_{SP}(t) & = & \frac{R_{c}}{4 t_{c}} \Big(\frac{t}{t_{c}}\Big)^{-\frac{3}{4}}\end{aligned}$$
where $R_{c}$ is the radius of the bubble at cooling time $t_{c}$. The snow-plow phase ends when the shell expansion velocity is close to the thermal sound speed of gas in the ISM (typically $\sim$ 10 km s$^{-1}$). The shell will stall and disperse due to the Rayleigh-Taylor instability [@mac99].
Our one-dimensional shell dynamical model may overestimate the winds and supernova mechanical luminosities as argued recently by Dopita et al. (2005), because the mixing and dynamical instabilities will occur in two dimensions, and the ISM is intrinsically inhomogeneous. Dopita et al. (2005) also suggested that the higher ISM pressure in starburst regions causes the expanding shell to stall at a smaller radius. Another argument is that the gravitational instability may induce new star formation inside the shells. These concerns may indicate that the conventional bubble/shell dynamics [@wea77; @mak87] may need to be modified.
Physical Conditions of The Swept-up Gas {#cond}
---------------------------------------
The PDRs that lie at the inner sides of the clouds or shells centrally illuminated by massive star formation are the origin of much of the FIR/sub-mm/mm radiation from the ISM. Physical conditions of the swept-up gas in these PDRs are very different from those of the cold gas components in the ISM. The gas temperature and density of the swept-up shells are a few orders of magnitude higher due to the strong FUV radiation and shock compression. The FUV radiation (6 eV $<$ $h\nu$ $<$ 13.6 eV) produced by newborn stars plays an important role in the heating and chemistry of PDRs, especially during the early evolution. Other sources that may contribute to the shell heating are the mechanical energy input by winds and SN explosions [@mck99], shocks caused by the accretion of gas at the outer surface of the shell [@mah80], and cosmic rays [@suc93; @bra03]. Cosmic rays may play an important role in the heating of swept-up gas after the stars with masses $\ge$ 8 M$_{\odot}$ have terminated as supernovae. Heating sources due to cloud-cloud collisions [@mns79] or shell-shell interaction [@saj99] are not considered in this study.
The total FUV flux is calculated by integrating the flux of the stellar population spectrum between 912 Å$ $ and 2055 Å$ $ for each time step using Starburst99, a time-dependent stellar population synthesis model developed by Leitherer et al. (1999). We consider an instantaneous burst for the star formation law, where the star formation occurs all at the same time (i.e. at age zero). The FUV field strength $G_0$ incident on the inner surface of the shell (visual extinction $A_v$ = 0) is then calculated by taking the ratio of the total FUV flux to the surface area 4$\pi$$r^2_s(t)$ of the expanding shell at each time-step. We use the same input parameters and assumptions for Starburst99 as those used in the shell dynamics calculation (see Table \[tbl-1\]).
The swept-up shell itself is supported by thermal gas pressure and non-thermal pressure due to micro-turbulence. The gas temperature decreases toward the outer surface of the shell, and the total gas density is assumed uniform. Therefore the pressure is lower at the outer surface. The shell density $n_s$ refers to the total H$_2$ density $n$(H$_2$) in this study. The shell density at each time-step is derived from balancing the pressure at the outer surface of the shell with the ram pressure,
$$n_{s}(t) = \frac{n_a v^2_s(t)}{kT_{s}(t)/\mu + \delta v^2_D}$$
where $n_a$ is the ambient molecular gas density, $v_s$($t$) is the expansion velocity, $T_s$($t$) is the gas temperature at the outer surface of the shell, $\mu$ is the mean molecular weight, $\mu$ = 0.62 $m_H$, $m_H$ is the mass of the hydrogen atom, and $\delta$$v_D$ is the micro-turbulent velocity inside the shell. The calculation of the gas temperature profile across the shell will be described in the following section. The thickness of the shell $d_s$ at each time-step is in turn calculated using the continuity equation (or mass conservation law),
$$d_{s}(t) = \frac{n_a r_{s}(t)}{3 n_{s}(t)}$$
The Time-dependent PDR Model {#pdr}
----------------------------
The gas temperature and chemical abundances of the swept-up shell are calculated self-consistently at each depth- and time-step using the time-dependent PDR model developed at UCL (called UCL\_PDR). A fully time-dependent treatment of the chemistry is employed in UCL\_PDR which includes 128 species involved in a network of over 1700 reactions [@bel05 and references therein]. The polycylic aromatic hydrocarbons (PAHs) chemistry is not included. The reaction rates are taken from the UMIST chemical database [@let00]. Detailed chemical modeling, heating and cooling mechanisms and the thermal balance between them are described in the literature [e.g. @tay93; @pap02 and references therein]. Heating due to shocks is not included. The UCL\_PDR code has been modified for the purpose of this study to include a pressure balance check at the outer surface of the shell, as well as the evolution information of shell density, thickness, and FUV radiation strength.
The UCL\_PDR code assumes a plane-parallel geometry and models the PDR as a semi-infinite slab of homogeneous density at a given time-step. The pressure is thus not in equilibrium across the PDR region. The FUV radiation field illuminates the shell from one side, and it becomes attenuated with increasing visual extinction $A_v$ into the shell at a given time-step as $G$ = $G_0$$e^{-2.4 k A_v}$, where $G_0$ is the FUV strength at $A_v$ = 0 calculated by the Starburst99 model. The coefficient $2.4 k$ in front of the $A_v$ in the exponent takes into account the difference in opacity from the visible to the UV and the influence of grain scattering. The timescale for gas in PDRs to reach chemical equilibrium depends on the gas density, temperature, degree of ionization, and species involved [@hat97; @vab98]. In our study, this timescale varies from 10$^5$ to 10$^7$ yr for the swept-up gas. Our comparative tests using a single time-step model fail to reproduce important chemical structure features predicted by the fully time-dependent model for ages up to 10 Myr. The use of a full time-dependent PDR code in which temperature and density changes with time is therefore justified in modeling the shell evolution over these timescales.
The Non-LTE Line Radiative Transfer Model {#lrt}
-----------------------------------------
The line radiative transfer properties are calculated using the Spherical Multi-Mol (SMMOL) code. The SMMOL model was also developed at UCL, implementing an accelerated $\Lambda$-iteration (ALI) method to solve multi-level non-LTE radiative transfer problems of gas inflow and outflow. The code computes the total radiation field and the level populations self-consistently. At each radial point, SMMOL generates the level populations and the line source functions. Our model assumes that the gas emission originates in the unresolved, homogeneous, spherical expanding shell and that all gas and dust in the H II region have been swept up into the shell, i.e. a dustless H II region. The background radiation field is assumed to be the cosmic background continuum of 2.73 K. A detailed description of the SMMOL radiative transfer model and its implementation can be found in the appendix of Rawlings & Yates (2001). The benchmarking and comparison with other line radiative transfer models are presented in van Zadelhoff et al. (2002).
Several programs were developed to separate and extract the gas temperature and fractional abundances for molecular and atomic species calculated by the UCL\_PDR code. These extracted gas temperature and abundances, along with the shell density, thickness, radius, and expansion velocity computed by the dynamical code, are re-gridded for a spherical geometry and used as input parameters for the SMMOL code to compute the total line intensity or flux. Einstein A and collisional rate coefficients for the molecular and atomic lines are taken from the Leiden Atomic and Molecular Database [@sch05]. The lowest 10 energy levels are calculated for all molecules, 3 for atomic \[C I\] and \[O I\], and 8 for atomic \[C II\].
Simulation of An Expanding Shell {#sim}
================================
Observational studies have shown that molecular clouds in the Milky Way have a distinct mass spectrum $M^{\alpha}_{GMC}$, with $\alpha$ = -1.5 $\pm$ 0.1 [@san85; @sol87] for cloud masses ranging between 10$^2$ and 10$^7$ M$_{\odot}$. Therefore, about 70% of the molecular mass in the Galaxy is contained in the GMCs with masses $>$ 10$^{6}$ M$_{\odot}$. These giant molecular clouds are known to be associated with active formation of massive stars. If we assume that the cloud mass distribution in a starburst galaxy follows a similar index to the galactic one, we would expect much of the luminosity of the starburst to arise from the GMCs with a fairly narrow range of masses. We adopt a value for the cloud mass of 10$^7$ M$_{\odot}$ for the GMCs in this study. We assume the average gas consumption rate or star-formation efficiency $\eta$ in starburst galaxies per 10$^8$ yr to be 0.25 [@ken98]. Therefore, the total stellar mass $M_{\ast}$ for the star cluster in the center of the GMC is 2.5 $\times$ 10$^6$ M$_{\odot}$, and the number of stars $N_{\ast}$ with masses $m_{\ast}$ $\ge$ 8 M$_{\odot}$ is about 2.2 $\times$ 10$^4$. The radius of the GMC is about 50 pc with an average cloud density $n_0$ = 300 cm$^{-3}$ and a cloud core density $n_c$ = 2 $\times$ 10$^3$ cm$^{-3}$ [@plu92; @efs00]. The ambient density $n_{ism}$ is assumed to be 30 cm$^{-3}$ [@cat94]. Here we present an idealized case study with this particular set of input parameters.
Kinematics of The Swept-up Gas {#kin}
------------------------------
The size of the H II region increases slowly with time. The Strömgren radius is about 4.9 pc assuming the number of Lyman continuum photons is 5 $\times$ 10$^{52}$ s$^{-1}$. The wind bubble catches up with the H II ionization front in a time less than 10$^5$ yr. The strong stellar winds cause the bubble to expand quickly into the cloud and sweep up more gas into the shell. The total wind power is estimated as $L_w$ $\simeq$ 1.4 $\times$ 10$^{40}$ erg s$^{-1}$ for the star cluster used in the model. When the most massive star in the center cluster (120 M$_{\odot}$) terminates as a supernova at $\sim$ 0.7 Myr, the thin shell caused by the H II region expansion and the stellar winds is still expanding at a speed of $\sim$ 40 km s$^{-1}$. At this time, the shell has swept up much of the mass of its parent cloud, and is propelled into the ISM with a uniform density. The mechanical energy produced by the first supernova and the subsequent ones re-energizes the shell. A supernova cut-off mass of 8 M$_{\odot}$ is assumed. The total energy generated by supernova explosions is $\sim$ 2.0 $\times$ 10$^{55}$ ergs over 40 Myr. At $\sim$ 7.5 Myr, the hot bubble starts to cool and loses its internal pressure, at which time the adiabatic phase ends. We adopt 1.0 for the metallicity $\mathcal{Z}$ with respect to the solar throughout this study. The effect of lower metallicity, which is suspected to be present in starburst galaxies, will be discussed in a future paper. The radius and velocity of the shell at the end of the adiabatic phase are about 270 pc and 24 km s$^{-1}$, respectively. At $\sim$ 50 Myr, the expansion velocity of the shell decreases to $\sim$ 10 km s$^{-1}$, the shell stalls and becomes thicker and less dense.
Fig. \[G0\] shows the FUV radiation strength $G_0$ incident on the inner surface of the shell ($A_v$ = 0) as a function of time. The $G_0$ value is in units of the Habing field, that is 1.6 $\times$ 10$^{-3}$ ergs cm$^{-2}$-s$^{-1}$ throughout this study. The FUV strength decreases from about 10$^8$ to 10$^5$ from the onset of star formation to about 5 Myr when most of the massive O stars ($>$ 30 M$_{\odot}$) have terminated as supernovae. It then decreases twice as fast to a value of 40 at 100 Myr.
PDRs are the origin of much of the FIR/sub-mm/mm line emission in a starburst galaxy. The surface layer ($A_v$ $\sim$ 1) contains atomic H, C, C$^{+}$ and O; the transition from atomic to molecular hydrogen occurs at the center layer ($A_v$ $\sim$ 1 - 2), whilst C$^{+}$ is converted into C and then CO over the region $A_v$ $\sim$ 2 - 4. H$_2$ and CO then extend to higher $A_v$ region and for $A_v$ $>$ 10 atomic O begins to be transformed into molecular O$_2$. The H$_2$ molecule provides effective self-shielding from the FUV radiation field. The CO layer also shows a degree of self-shielding, and therefore extends deeper into the shell. Small grains play an important role in the photoelectric heating of PDRs. Gas heating is dominated by collisional deexcitation of FUV-pumped H$_2$ and vibrationally excited H$_2$ at the PDR surface. The thermal energy radiated by the dust is important for the gas heating at larger optical depth [@hol91]. The gas heating/chemistry at later evolutionary stages is no longer dominated by stellar radiation but by other sources, such as cosmic-rays and X-rays. The PDR cooling is dominated by fine structure line emission, such the \[C II\] 158$\mu$m and \[O I\] 63 $\mu$m transitions, whose critical densities are 3 $\times$ 10$^3$ cm$^{-3}$ and 5 $\times$ 10$^5$ cm$^{-3}$, respectively. At greater depths, molecular line emission (CO, OH, H$_2$O), ro-vibrational transitions of H$_2$, and gas-dust collisions contribute to the PDR cooling.
Table \[tbl-2\] summarizes the input parameters for our fully time-dependent PDR model. The initial abundance of H$_2$ is set to $n$(H$_2$)/$n_H$ = 0.5 [@har03]. At the first time-step ($t$ = 0 yr) all depth-steps take as their initial abundances the values produced by a single-point dense dark-cloud model. The input parameters for the dark-cloud modeling are $n_H$ = 4 $\times$ 10$^5$ cm$^{-3}$, $T_{GMC}$ = 10 K, $G_0$ = 1, and the gas-phase abundances relative to H nuclei $x_{He}$ = 7.5 $\times$ 10$^{-2}$, $x_C$ = 1.8 $\times$ 10$^{-4}$, $x_O$ = 4.4 $\times$ 10$^{-4}$, and $x_{Mg}$ = 5.1 $\times$ 10$^{-6}$. For subsequent time-steps, the input abundances are re-set to the output abundances of the previous time-step generated by the UCL\_PDR code. The gas temperature and chemical abundances at each depth- and time-step are calculated by balancing the heating and cooling. The cosmic-rays ionization rate is enhanced by a factor of 1.5 at later times ($t$ $>$ 10 Myr) to artificially include the soft X-rays heating effect on the gas of the shell. We assume that the gas-to-dust mass ratio is 100. Fig. \[nsds\] shows the shell density $n_s$ (or $n$(H$_2$)) and thickness $d_s$ as a function of time, as calculated by the shell dynamical code and the UCL\_PDR code, under the condition that the gas pressure at the outer surface of the shell differs from the ambient gas pressure by $\le$ 10%. The shell density varies between 10$^3$ and 10$^6$ cm$^{-3}$, and the thickness of the shell changes from 10$^{-3}$ to 10 pc over the 100 Myr. We adopt a fixed micro-turbulent velocity $\delta$$v_D$ = 1.5 km s$^{-1}$ for the shell. The evolution of the shell density and thickness is constrained by the expansion velocity $v_s$, the shell temperature $T_s$, and the ambient density $n_a$ (See Equation (6) & (7)). Changes in $v_s$ and $T_s$ are relatively small during the H II expansion ($n_a$ = $n_0$ or 300 cm$^{-3}$), as a result we see the first plateau as shown in Fig. 3. The jump seen at $t$ $\sim$ 2 $\times$ 10$^4$ yr is caused by the change from the H II expansion to the Winds phase. During the early Winds phase and before the shell sweeps up all the material of its parent GMC ($t$ $<$ 0.8 Myr), the effect due to the shell deceleration is compensated for the effect due to the cooling in the shell. This produces a second plateau. After this time, the shell expands into a less dense ambient ISM, i.e. $n_a$ = $n_{ism}$ or 30 cm$^{-3}$. Less ambient pressure causes a decrease in the shell density or a increase in the shell thickness. Fig. \[T2Av\] and Fig. \[x2Av\] show the profiles of the gas temperature and chemical abundances as a function of visual extinction $A_v$ for an expanding shell at several characteristic ages. The size of the PDR changes with time indicated by different maximum values of $A_v$ in both Fig. \[T2Av\] and Fig. \[x2Av\]. At $\sim$ 0.7 Myr, all mass in the GMC has been swept into the shell.
Molecular and Atomic Line Emission {#lrad}
----------------------------------
The flux and intensity of FIR/sub-mm/mm line emission is calculated for several molecular and atomic species (CO, HCN, HCO$^+$, C, C$^+$, and O). The total flux or intensity of each line is the sum of the emission from the entire shell. For the initial 0.7 Myr, the emission from the parent GMC is also taken into account in the total line emission calculation. In this section, we present predictions of the line ratios for CO, \[C I\], and \[C II\] for an expanding shell. More simulations will be presented and discussed when we illustrate the model by a comparison with the observations of M 82 in § \[m82\].
Molecular CO is known as a good tracer for the diffuse components and total molecular gas content in a galaxy, but it is relatively poor tracer of the dense gas directly involved in massive star formation. Fig. \[r2t\_co\] shows our modeling results for line ratios of high-$J$ transitions to the 1-0 transition of the bright and highly abundant $^{12}$CO molecule as a function of the starburst age of the shell.
$$\begin{aligned}
r_{21} & = & I_{21} / I_{10} , \\
\vdots & & \vdots \\
r_{71} & = & I_{76} / I_{10} \end{aligned}$$
where $I_{J,J-1}$ is the line intensity, $r_{J+1,1}$ is the line intensity ratio, and $J$ = 1,$\ldots$,7.
For the adiabatic phase ($t$ $<$ 7.5 Myr), strong winds and supernova explosions compress the gas in the fast expanding shell to a high density $n$(H$_2$) $>$ 10$^4$ cm$^{-3}$ (see Fig. \[nsds\]), and the strong FUV radiation $G_0$ $>$ 10$^4$ heats up the gas and dust of the shell to a temperature $>$ 100 K (see Fig. \[T2Av\]). A significant amount of highly excited CO line emission is generated from the shell and its parent cloud, and therefore the line ratios of $r_{21}$ through $r_{71}$ are $\ge$ 1.0. At around 10 Myr, all line ratios (1 $\le$ $J$ $\le$ 7) have dropped below 1.0, the shell has entered the snow-plow phase, the corresponding FUV field $G_0$ is $\le$ 10$^4$, the shell density $n$(H$_2$) is $<$ 3.0 $\times$ 10$^3$ cm$^{-3}$, and the gas temperature in the shell $T_{gas}$ is between 20 and 230 K.
The far-infrared fine structure lines are the most important cooling lines of the ISM in a galaxy. Fig. \[r2t\_atom\] shows the modeling results of the line intensity ratio of \[C II\]158$\mu$m to \[C I\]610$\mu$m and the line flux ratio of \[C II\]158$\mu$m to CO(1-0). About 75% of the \[C II\]158$\mu$m emission comes from PDRs and 25% from the H II region [@col99]. The latter is not taken into account in our calculations. The \[C II\]158$\mu$m/CO(1-0) line flux ratio rises from about 10 to 10$^4$ after 1 Myr, and then slowly decreases to $\sim$ 10$^3$ at 80 Myr. It is clear that the cooling of the swept-up gas in the expanding shell is dominated by C$^+$, the contribution of the CO cooling is a small fraction to the total gas cooling in a massive star forming environment.
Application to The Nearby Starburst Galaxy M 82 {#m82}
===============================================
In section 4.1, we compare our modeling results with the observations of an expanding supershell in the nearby starburst galaxy M 82. In section 4.2, we compare our modeling results with the average gas properties in the central 1.0 kpc region of this galaxy.
M 82 is an irregular starburst galaxy located at a distance of about 3.25 Mpc. This galaxy has been observed over a wide range of wavelengths. The starburst activity in M 82 was likely triggered by tidal interaction with its companion M 81 beginning about 10$^8$ yr ago in the nucleus, and is currently propagating into the molecular rings. The infrared luminosity of M 82 is about 4 $\times$ 10$^{10}$ L$_{\odot}$ arising mostly from the central $\sim$ 400 pc region, which has a stellar bar structure and currently has a high supernova rate of $\sim$ 0.05 - 0.1 yr$^{-1}$ [@mux94]. The evolutionary scheme in M 82 remains under debate. The most common suggested ages of the M 82 starburst in the central regions are 3 - 7 Myr predicted by Colbert et al. (1999) using one instantaneous burst model in dusty media with a 100 M$_{\odot}$ cut off, and 10 - 30 Myr predicted by Efstathiou et al. (2000) using two instantaneous bursts model in dusty media with a 125 M$_{\odot}$ cut off. Recently, Föster-Schreiber et al. (2003) presented a more complete evolutionary scheme of the global starburst activity in M 82, and suggested that there are two bursts, one occurred at $\sim$ 5 Myr ago and another one at $\sim$ 10 Myr ago also using instantaneous bursts model in dusty media with a 100 M$_{\odot}$ cut off.
The Supershell Surrounding The SNR 41.9+58 {#superb}
------------------------------------------
Observations have detected an expanding supershell centered around the bright SNR 41.9+58 in both molecular line and radio continuum [@wei99; @wil99]. This supershell has a diameter of $\sim$ 130 pc, an expansion velocity of $\sim$ 45 km s$^{-1}$, and a mass of $\sim$ 8 $\times$ 10$^6$ M$_{\odot}$. Using the set of initial cloud conditions selected for our simulation (see § \[sim\]), i.e. a cloud mass $M_{GMC}$ = 10$^7$ M$_{\odot}$, cloud density $n_0$ = 300 cm$^{-3}$, ambient ISM density $n_{ism}$ = 30 cm$^{-3}$, and star formation efficiency $\eta$ = 0.25, we derive a swept-up shell that has very similar characteristics to the observed one. At the observed radius of $\sim$ 65 pc, our model indicates an age of 1 Myr, an expansion velocity of $\sim$ 45 km s$^{-1}$, and a swept up H$_2$ mass of $\sim$ 7.6 $\times$ 10$^6$ M$_{\odot}$. The kinetic energy of the observed supershell is estimated to be about 1.6 $\times$ 10$^{53}$ ergs [@wei99]. Our model predicts a kinetic energy of $\sim$ 1.5 $\times$ 10$^{53}$ ergs for the expanding shell at the age of 1 Myr. The total mechanical energy needed for the creation of this supershell is $\sim$ 1.7 $\times$ 10$^{54}$ ergs, which is contributed by winds and supernovae associated with $\sim$ 1700 O stars ($\ge$ 40 M$_{\odot}$). Therefore, about 10% of the total energy is present in the form of kinetic energy of the expanding shell. The remarkably good agreement between our model results and the observations implies that this supershell may be created by strong winds and supernova explosions from a star cluster with a total mass of 2.5 $\times$ 10$^6$ M$_{\odot}$ which occurred in the center about 1 Myr ago. The comparison is summarized in Table \[tbl-3\].
Furthermore, our model predictions of the CO, \[C I\], and \[C II\] line ratios for this expanding supershell can be used as a comparison with future observations, and also to constrain the physical conditions of the gas in the shell (see Fig. \[r2t\_co\] and Fig. \[r2t\_atom\] presented in § \[lrad\]). The model line ratios which are greater than 1.0 at $t$ $<$ 8 Myr imply that the molecular CO is optically thin in the expanding supershell. Therefore, it is better to look at the high CO transitions ($J$ $>$ 3) in this supershell.
The Central Starburst Region
----------------------------
Besides the known expanding supershell centered around SNR 41.9+58, there are other undetected shells with sizes ranging from several tens of parsec to more than 1 kiloparsec, and kinetic energies ranging from $\sim$ 10$^{50}$ ergs to more than 10$^{54}$ ergs. These shells would likely be present as partial arcs, or fragments, or cloud-like clumps due to strong winds and supernova explosions or due to shell-shell and shell-cloud interactions; only a few are visible as full circular arcs. The very good agreement between our model and the supershell observations indicates that the set of models we have put forward in this paper can be used to interpret other shells in a starburst galaxy like M 82. In this section, we illustrate the possibilities by comparing our model calculations with the observed FIR/sub-mm/mm properties of molecular and atomic line emission in the central starburst regions.
First of all, Fig. \[rr2t\_dens\] shows the model ratio-ratio plots of HCN(4-3)/(3-2) versus HCN(3-2)/(1-0) and HCO$^+$(4-3)/(3-2) versus HCO$^+$(3-2)/(1-0), and a comparison with the observations of dense gas in the central 300 pc region [@saf00]. The dense gas tracers HCN and HCO$^+$ are better indicators of active star formation than CO, but poor tracers of the total molecular gas content. For both plots, the best agreement is at a starburst age of $\sim$ 3 Myr, implying that the expanding shell size is about 300 pc. Secondly, Fig. \[rr2t\_co\] shows the model ratio-ratio plots for $^{12}$CO and $^{13}$CO, and a comparison with the observations of the CO line emission from three lobes (north-east, center, and south-west) in the central 300 to 600 pc regions [@mao00]. The isotope abundance ratio \[$^{12}$CO\]/\[$^{13}$CO\] is adopted to be 75 for the simulation. The best agreement shown in plots (a), (b), and (d) is at a starburst age of $\sim$ 6 Myr, and $\sim$ 3 Myr for plot (c), corresponding to shell sizes between 300 and 560 pc. The physical conditions for the gas of the shell at age 3 - 6 Myr are $G_0$ $\sim$ 2 $\times$ 10$^4$ - 15 $\times$ 10$^4$, $n$(H$_2$) $\sim$ 1.0 $\times$ 10$^4$ - 2.0 $\times$ 10$^4$ cm$^{-3}$, $T_{gas}$ $\sim$ 50 - 250 K, and total molecular gas mass $M_{mol}$ $\sim$ 0.3 $\times$ 10$^8$ - 2.1 $\times$ 10$^8$ M$_{\odot}$. Finally, Fig. \[rr2t\_atom\] shows the model ratio-ratio diagram of \[O I\]63$\mu$m/\[C II\]158$\mu$m versus \[O I\]63$\mu$m/\[O I\]145$\mu$m, and a comparison with the observations of these atomic lines from the central 1.1 kpc region [@neg01]. The model \[C II\]158$\mu$m line is underestimated by a factor of about 1.3, since about 25% of the total line emission coming from the H II region is not included in the calculation. Therefore, the best agreement between the model and the observation is achieved at an age of $\sim$ 25 Myr old. The atomic line data are based on a 80$^{\prime\prime}$ beam whereas the molecular line data pertain only to the the lobes and nuclear sources at a 22$^{\prime\prime}$ beam. Thus, part of the reason for the discordant age in the atomic data may be the different regions sampled, since they may have a different starburst history. Our predicted gas conditions for the shell at this age are $G_0$ $\sim$ 500, $n$(H$_2$) $\sim$ 1.8 $\times$ 10$^3$ cm$^{-3}$, $T_{gas}$ $>$ 15 K, and $M_{mol}$ $\sim$ 6.0 $\times$ 10$^8$ M$_{\odot}$. These conditions are consistent with the PDR model fits to the observations by Colbert et al. (1999). But the age inferred by Colbert et al. (1999) is 3 - 7 Myr. The large age discrepancy between the two different modeling results from the fact that our model includes a more massive cluster which then yields the same gas conditions (FUV flux and gas density) at a larger distance and hence, in the context of an expanding shell, an older age. It is clear that the starburst age of the whole central region is model dependent. More simulations with a variety of input cloud conditions and a comparison with data taken at a wider range of wavelengths are needed in order to identify the ages of starbursts accurately.
Although different stages of development are applicable to different central regions of M 82, the shell sizes and the physical conditions of the gas within the rings (diameter $\sim$ 300 - 600 pc) predicted by our model are similar to what is expected from models involving expanding shells from a central starburst such as those proposed by Carlstrom & Kronberg (1991). Therefore, it is possible that the molecular rings in M 82 are a product of gas that was swept-up by the nuclear starburst activity which has evolved for about 10$^8$ yr. Their hypothesis is supported by the observations of CO line emission and continuum emission, as well as the discovery of supershells that have not yet had time to break out of the galactic plane. However, it is important to realize the foregoing interpretation of the lobes as a ring or torus is not unique. A number of authors have argued that the molecular rings may be a product of Linblad resonance instabilities associated with the gravitational effects of the bar [e.g. @sal96; @wil00]. In future work, we will carry out more simulations to test the hypothesis suggested by Carlstrom & Kronberg (1991).
Conclusions {#sum}
===========
We have presented a set of starburst models that can be used to relate the observed FIR/sub-mm/mm properties of molecular and atomic gas in a starburst galaxy to its age and star formation history. As a preliminary approach, we have illustrated our model by a comparison with the observations of the expanding supershell centered around SNR 41.9+58. The very good agreement implies that the expanding supershell is created by strong stellar winds and SN explosions from a young star cluster ($\sim$ 2.5 $\times$ 10$^6$ M$_{\odot}$) located in its center.
Our model predictions of CO, HCN, and HCO$^+$ line ratios agree with the molecular data for the central lobes (300 - 600 pc) for a shell with an age in the 3 - 7 Myr range. This implies that the molecular rings are possibly a consequence of swept-up or compressed gas caused by massive star formation originating in the nucleus of M 82. More simulations in future work may be able to justify this hypothesis. The atomic line ratios calculated by our model do not fit the observed data as well as the molecular data, but suggest a much older shell, because the atomic line emission comes from a much larger region ($>$ 1 kpc). A variety of modeling parameters need to be considered to yield more accurate starburst ages.
Our model also yields appropriate values for the gas density, temperature, and structure scales compared to those measured in the center of M 82 [e.g. @las63; @rie89; @stu97; @saf00; @mao00; @neg01; @war03], and the total H$_2$ content within the inner 600 pc ($\sim$ 2.0 $\times$ 10$^8$ M$_{\odot}$) is compatible with the observations [e.g. @wil92]. Therefore, the neutral ISM in the central star-forming region of M 82 may be viewed as a product of evolving shells, and is now presenting itself in the form of fragments, small cloud clumps, sheets, or even full circular arcs.
LY would like to thank Dr. Andreas Efstathiou, Dr. Claus Leitherer, and Dr. Peter van Hoof. LY is grateful to Dr. Chris Loken and Mr. Hugh Merz at the Canadian Institute for Theoretical Astrophysics for their generosity in allowing her to run the PDR simulations on their fastest PC machines. LY is grateful for use of the PPARC funded Miracle Computing Facility, located at UCL, to perform her RT calculations. We thank the referee for helpful comments and suggests. This research was supported by a research grant from the Natural Sciences and Engineering Research Council of Canada to ERS, and a Reinhardt Graduate Student Travel Fellowship from the Department of Astronomy and Astrophysics at the University of Toronto. TAB is supported by a PPARC studentship. SV acknowledges individual financial support from a PPARC advanced Fellowship.
Abbott, D. C. 1982, , 263, 723 Bell, T. A., Viti, S., Williams, D. A., Crawford, I. A., & Price, R. J. 2005, , 357, 961 Bradford, C. M. et al., 2003, , 586, 891 Carlstrom, J. E., & Kronberg, P. P. 1991, , 366, 422 Castor, J., McCray, R., & Weaver, R. 1975, , 200, L107 Colbert, J. W. et al., 1999, , 511, 721 Comeron, F., & Torra, J. 1994, , 423, 652 Dopita, M. A. et al., 2005, , 619, 755 Efstathiou, A., Rowan-Robinson, M., & Siebenmorgen, R. 2000, , 313, 734 Förster-Schreiber, N. M., Genzel, R., Lutz, D., & Sternberg, A. 2003, , 299, 193 Franco, J., Tenorio-Tagle, G., & Bodenheimer, P. 1990, , 349, 126 Gao, Y., Lo, K. Y., Lee, S.-W., & Lee, T.-H. 2001, , 548, 172 Goldreich, P. & Kwan, J. 1974, , 189, 441 Hartquist, T. W., Falle, S. A. E. G., Williams, D. A. 2003, Ap&SS, 288, 369 Hollenbach, D. J., Takahashi, T., Tielens, A. G. G. M. 1991, , 377, 192 Hollenbach, D. J. & Tielens, A. G. G. M. 1997, , 35, 179 Kennicutt, R. C., Jr. 1998, , 498, 541 Koo, B. & McKee, C. F. 1992, , 388, 93 Leitherer, C. et al., 1999, , 123, 3 Le Teuff, Y. H., Millar, T. J., & Markwick, A. J. 2000, A&AS, 146, 157 Lynds, C. R., & Sandage, A. R. 1963, , 137, 1005 Mac Low, M. M. 1999, Variable and Non-spherical Stellar Winds in Luminous Hot Stars, Proceedings of the IAU Colloquium No. 169 Held in Heidelberg, Germany, 15-19 June 1998. Edited by B. Wolf, O. Stahl,and A. W. Fullerton. Springer-Verlag Berlin Heidelberg New York. Also Lecture Notes in Physics, volume 523, p.391 Mac Low, M. M., & McCray, R. 1988, , 324, 776 Madau, P. et al., 1998, , 498, 106M Martin, H. M., Hills, R. E., & Sanders, D. B. 1984, , 208, 35 Mao, R. Q., Henkel, C., Schulz, A., Zielinsky, M., Mauersberger, R., Störzer, H., Wilson, T. L., Gensheimer, P. 2000, , 358, 433 McCray, R. & Kafatos, M. 1987, , 317, 190 McCray, R.; Snow, T. P., Jr. 1979, , 17, 213 McKee, C. F. 1999, The Origin of Stars and Planetary Systems, Charles J. Lada and Nikolaos D. Kylafis. Kluwer Academic Publishers, p.29 McKee, C. F., Hollenbach, D. J. 1980, , 18, 219 Muxlow, T. W. B., Pedlar, A., Wilkinson, P. N., Axon, D. J., Sanders, E. M., & de Bruyn, A. G. 1994, , 266, 455 Negishi, T., Onaka, T., Chan, K.-W., & Roellig, T. L. 2001, , 375, 566 Papadopoulos, P. P., Thi, W.-F., & Viti, S. 2002, , 579, 270 Plume, R., Jaffe, D. T., & Evans, Neal J., II 1992, , 78, 505 Rawlings, J. M. C., & Yates, J. A. 2001, , 326, 1423 Rieke, G. H., Lebofsky, M. J., Thompson, R. I., Low, F. J., & Tokunaga, A. T. 1980, , 238, 24 Rieu, N. Q., Nakai, N., Jackson, J. M. 1989, , 220, 57 Salpeter, E. E. 1955, , 121, 161 Sanders, D. B., Scoville, N. Z., & Solomon, P. M. 1985, , 289, 373 Scalo, J., & Chappell, D. 1999, , 310, 1 Schöier, F. L., van der Tak, F. F. S., van Dishoeck, E. F., & Black, J. H. 1005, , 432, 369 Shen, Jianjun, & Lo, K. Y. 1996, IAUS, 170, 101 Seaquist, E. R., & Frayer, D. T. 2000, , 540, 765 Solomon, P. M., Rivolo, A. R., Barrett, J., & Yahil, A. 1987, , 319, 730 Spitzer, L. 1978, Physical Processes in the Interstellar Medium, Wiley Interscience, New York Stutzki, J. 1997, , 477, 33 Suchkov, A., Allen, R. J., Heckman, T. M. 1993, , 413, 542 Taylor, S. D., Hartquist, T. W., & Williams, D. A. 1993, , 264, 929 Tielens, A. G. G. M., & Hollenbach, D. 1985, , 291, 722 van Dishoeck, E. F., Blake, G. A. 1998, , 36, 317 van Zadelhoff, G. J. et al., 2002, , 395, 373 Ward, J. S., Zmuidzinas, J., Harris, A. I., & Isaak, K. G. 2003, , 587, 171 Weaver, R., McCray, R., Castor, J., Shapiro, P., & Morre, R. 1977, , 218, 377 Weiss, A., Walter, F., Neininger, N., Klein, U. 1999, , 345, 23 Wild, W. et al., 1992, , 265, 447 Wills, K. A., Redman, M. P., Muxlow, T. W. B., & Pedlar, A. 1999, , 309, 395 Wills, K. A., Das, M., Pedlar, A., Muxlow, T. W. B., & Robinson, T. G. 2000, , 316, 33
[ll]{} Assumptions: & - spherical geometry, non-magnetized (GMCs and shells)\
& - no interactions between shells, or shell and cloud\
& - dustless H II regions\
& - uniform densities of GMCs and ambient media\
& - all stars form instantaneously, no stars form inside the shells\
& - stellar mass 0.1 - 120 M$_{\odot}$ with Salpeter IMF $dN/dM_{\ast}$ $\propto$ $M^{-2.35}_{\ast}$\
& - PDRs exist primarily within the expanding shells\
Input Parameters: & - GMC mass $M_{GMC}$ = 10$^7$ M$_{\odot}$\
& - average cloud density $n_0$ = 300 cm$^{-3}$, cloud core density $n_c$ = 2000 cm$^{-3}$\
& - ambient ISM density $n_{ism}$ = 30 cm$^{-3}$\
& - star formation efficiency $\eta$ = 0.25\
& - metallicity $\mathcal{Z}$ = 1.0 Z$_{\odot}$\
& - gas-to-dust ratio = 100\
Output Parameters: & - radius, velocity, temperature, density, and thickness of the shell\
& - chemical abundances of different molecules and atoms in the shell\
& - integrated line intensity/flux, line ratios\
Observational & - line intensities/fluxes and line ratios for molecules (e.g. $^{12}$CO, $^{13}$CO, HCN, HCO$^+$),\
& $ $ $ $ and atoms (e.g. \[C I\], \[C II\], \[O I\])\
[lll]{} Starburst age (yr) & $t$ & 0 $\le$ $t$ $<$ 10$^8$\
Incident FUV flux (Habing field) & $G_0$ & 10 $<$ $G_0$ $\le$ 10$^8$\
Turbulent (microturbulence) velocity (km s$^{-1}$) & $\delta$$v_D$ & 1.5\
PDR surface density ($A_v$ = 0 mag) & $n_H$ & 10$^3$ $\le$ $n_H$ $<$ 10$^7$\
Initial gas-phase abundances relative to H & &\
PAH abundance & $x_{PAH}$ & 4.0 $\times$ 10$^{-7}$\
Dust visual absorption cross section (cm$^{-2}$) & $\sigma_v$ & 3.1 $\times$ 10$^{-10}$\
H$_2$ formation rate on dust at $A_v$ = 0 (cm$^3$ s$^{-1}$) & $\eta_{H_2}$ & 3.0 $\times$ 10$^{-18}$\
Cosmic-rays ionization rate (s$^{-1}$) & $\zeta$ & 1.3 $\times$ 10$^{-17}$\
[lll]{} Radius (pc) & 65.0 & 65.0\
Age (Myr) & 1.0 & 1.0\
Expansion velocity (km s$^{-1}$) & 45 & 45\
Total H$_2$ molecular gas mass ($\times$ 10$^6$ M$_{\odot}$) & 8.0 & 7.6\
Kinetic Energy ($\times$ 10$^{53}$ ergs) & 1.6 & 1.5\
Total stellar mass in the center cluster ($\times$ 10$^6$ M$_{\odot}$) & $\ldots$ & 2.5\
Total number of O stars ($\ge$ 40 M$_{\odot}$) & $\ldots$ & 1700\
Total Mechanical Energy ($\times$ 10$^{54}$ ergs) & $\ldots$ & 1.7\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Given a solution $Y$ to a rough differential equation (RDE), a recent result [@cl2018] extends the classical Itö-Stratonovich formula and provides a closed-form expression for $\int Y \circ \mathrm{d} \mathbf{X} - \int Y \, \mathrm{d} X$, i.e. the difference between the rough and Skorohod integrals of $Y$ with respect to $X$, where $X$ is a Gaussian process with finite $p$-variation less than 3. In this paper, we extend this result to Gaussian processes with finite $p$-variation such that $3 \leq p < 4$. The constraint this time is that we restrict ourselves to Volterra Gaussian processes with kernels satisfying a natural condition, which however still allows the result to encompass many standard examples, including fractional Brownian motion with $H > \frac{1}{4}$. Analogously to [@cl2018], we first show that the Riemann-sum approximants of the Skorohod integral converge in $L^2(\Omega)$ by adopting a suitable characterization of the Cameron-Martin norm, before appending the approximants with higher-level compensation terms without altering the limit. Lastly, the formula is obtained after a re-balancing of terms, and we also show how to recover the standard Itö formulas in the case where the vector fields of the RDE governing $Y$ are commutative.'
author:
- Thomas Cass and Nengli Lim
title: '[A Stratonovich-Skorohod formula for Volterra Gaussian rough paths]{}'
---
Introduction
============
Lyons’ rough path theory is a framework for giving a path-wise interpretation to stochastic differential equations of the form $$\begin{aligned}
\label{basic SDE}
\mathrm{d}Y_{t}=V(Y_{t})\circ\mathrm{d}X_{t},\quad
Y_{0}=y_{0},\end{aligned}$$ in particular for a broad class of continuous, vector-valued Gaussian processes $X$. A fundamental contribution of Lyons [@lyons98; @lcl2006] was to realize that this needs $X$ to be enriched to a *rough path* $\mathbf{X}$ whose components comprise not only $X$, but also the higher-order iterated integrals up to some finite degree. The model then ought to be rewritten as $$\begin{aligned}
\label{basic SDE1}
\mathrm{d}Y_{t}=V(Y_{t})\circ\mathrm{d}\mathbf{X}_{t},\quad
Y_{0}=y_{0},\end{aligned}$$ to reflect the dependence of the solution on the enriched rough path. The existence of $\mathbf{X}$ for a given stochastic process cannot be taken for granted. Nevertheless, if $X$ is Gaussian then $\mathbf{X}$ can be constructed in a canonical way as an appropriate limit of iterated integrals of smooth (i.e. bounded variation) approximations to $X$ such as piecewise linear approximations, see [@fv2010a; @fv2010b]. For this, conditions are needed on the covariance function, but these are flexible enough to encompass a range of examples including fractional Brownian motion with Hurst index $H > \frac{1}{4}$.
In this paper we will assume that $X=\left( X^{1},...,X^{d}\right) $ has i.i.d components, each centered with covariance function $R$, and that $X$ is defined on a probability space $\left( \Omega,\mathcal{F},\mathbb{P}\right)
$. For simplicity we assume that $\mathcal{F}$ is generated by $X.$ The process $X$ then gives rise an isonormal Gaussian process w.r.t. the Hilbert $\mathcal{H}_{1}^{d}=\oplus_{i=1}^{d}\mathcal{H}_{1}^{\left( i\right) }$ where, for all $i=1,...,d,$ $\mathcal{H}_{1}^{\left( i\right) } = \mathcal{H}_{1}$ and $\mathcal{H}_{1}$ is the completion of the real vector space $$\begin{aligned}
\text{span}\left\{ \mathds{1}_{[0,t)}(\cdot):\bigg\vert\;t\in[0,T]\right\}\end{aligned}$$ endowed with the inner-product $\left\langle \mathds{1}_{[0,t)}(\cdot), \mathds{1}_{[0,s)}(\cdot) \right\rangle_{\mathcal{H}_1} = R(t,s)$. The solution $Y$ to (\[basic SDE1\]) can also be viewed as a Wiener functional on $\left(\Omega,\mathcal{F},\mathbb{P}\right) $, and its properties can then be studied using Malliavin calculus. A number of recent works have opened up the interplay between Lyons’ and Malliavin’s calculi, see e.g. [@cfv2009], [@cf2011], [@hp2013] and [@chlt2015]. In particular, in a recent paper [@cl2018] the authors have proven a conversion formula for the difference between the rough path integral of $Y$ w.r.t $\mathbf{X}$ and the Skorohod integral $\delta^{X}$ of $Y$ (i.e. the $L^{2}\left( \Omega\right)$ adjoint of the Malliavin derivative operator). In more detail this result shows, for the case where $Y$ and $X$ are both $\mathbb{R}^{d}$-valued, the following almost sure identity $$\begin{aligned}
\label{corr form}
\begin{split}
\int_{0}^{T}\left\langle Y_{t}\circ\mathrm{d}\mathbf{X}_{t}\right\rangle - \delta^{X}\left( Y\right)
&= \frac{1}{2}\int_{0}^{T}\mathrm{tr}\left[V(Y_{t})\right] \,\mathrm{d}R(t) \\
&\qquad+ \int_{[0,T]^{2}}\mathds{1}_{[0,t)}(s) \mathrm{tr}\left[ J_{t}^{\mathbf{X}}\left( J_{s}^{\mathbf{X}}\right)^{-1} V(Y_{s})-V(Y_{t})\right] \,\mathrm{d}R(s,t).
\end{split}\end{aligned}$$ Here, $J_{t}^{\mathbf{X}}$ denotes the Jacobian of the flow map $y_{0} \rightarrow Y_{t},$ and the second part of the correction term is a proper 2D Young-Stieltjes integral (see [@fv2010a; @fv2010b]) with respect to the covariance function of $X.$ When $X$ is standard Brownian motion, this last term vanishes since the integrand is zero on the diagonal and $\mathrm{d}R(s,t) = \delta_{\{s=t\}}\,\mathrm{d}s\,\mathrm{d}t$. This, together with the fact that $R(t)=t$, allows us to recover the classical Itô-Stratonovich conversion formula. This a rather basic tool in stochastic analysis and generalizations are likely to be important especially given the now widespread adoption of Gaussian models, e.g. most recently in mathematical finance [@bfg2016].
In [@cl2018], conditions need to be imposed in the proof of the formula (\[corr form\]) which limit the range of applications. An important assumption, for instance, is that the covariance function of $X$ has finite (two-parameter) $\rho$-variation for $\rho \in \left[ 1, \frac{3}{2} \right)$. This implies that the sample paths of $X$ will have finite $p$-variation, for some $p\in [2, 3)$, and this excludes interesting examples such as fractional Brownian motion with $H \in \left( \frac{1}{4}, \frac{1}{3} \right]$.
The purpose of the present paper is to extend the correction formula to these less regular cases. To do so we will assume that the Gaussian process $X$ is a Volterra process; that is, the covariance function $R$ of each component can be written as $$\begin{aligned}
R(s,t)=\int_{0}^{t\wedge s}K(t,r)K(s,r)\,\mathrm{d}r,\end{aligned}$$ for some kernel $K$, a square-integrable function $K: \left[ 0, T \right]^2 \rightarrow\mathbb{R}$ with $K(t,s)=0,\;\forall s\geq t.$ We will present conditions on $K$ that allow us to generalize . In doing so, we need to overcome a number of serious obstacles. We highlight here the three most salient of these, outline the contribution of the present work and, at the same time, provide a road-map for the paper:
(i) We need to prove that the solution $Y$ belongs to the domain of the Skorohod integral $\delta^{X}.$ In fact we prove the stronger statement that $Y$ belongs to the Malliavin Sobolev space $\mathbb{D}^{1,2} \left(\mathcal{H}_{1} \right) \subset \mathrm{Dom}(\delta^{X})$ (for definitions see Section \[MDsection\]). To show that $Y$, a path-valued random variable, can be understood as a random variable in the Hilbert space $\mathcal{H}_{1}$, we need to identify a class of functions with a subset of $\mathcal{H}_{1}$. This was proved in [@cl2018], by taking advantage of the assumption that $\rho \in \left[1, \frac{3}{2} \right)$, but the less regular cases need a new argument that exploits the structure of the Volterra kernel. To handle the Malliavin derivative $\mathcal{D}Y$, we need a similar result that identifies a class of two-parameter functions as a subset of $\mathcal{H}_{1}\otimes\mathcal{H}_{1}$.
(ii) For the examples considered in this paper, the Gaussian rough path $\mathbf{X}$ will consist of iterated integrals up to degree three; i.e. $\mathbf{X=}\left( 1,X,\mathbf{X}^{2},\mathbf{X}^{3}\right) .$ This contrasts with the result in [@cl2018], where only the case $\mathbf{X=} \left( 1,X,\mathbf{X}^{2}\right)$ needs to be considered. This increases the complexity of the the arguments significantly; indeed, the rough integral in the left side of (\[corr form\]) is now well approximated locally by terms up to third-order $$\begin{aligned}
\int_{s}^{t}\left\langle Y_{t}\circ\mathrm{d}\mathbf{X}_{t} \right\rangle
\simeq\left\langle Y_{s},X_{s,t}\right\rangle +V\left( Y_{s}\right) \mathbf{X}_{s,t}^{2} + V^2(Y_{s}) \left(
\mathbf{X}_{s,t}^{3}\right).\end{aligned}$$ A key step in [@cl2018] is the proof that the second-order terms in this approximation satisfy $$\begin{aligned}
\lim_{\left\Vert \pi(n)\right\Vert \rightarrow 0}\left\Vert \sum_{i:\pi(n)=\left\{ t_{i}^{n}\right\}}V(Y_{t_{i}^{n}})\left( \mathbf{X}_{t_{i}^{n}, t_{i+1}^{n}}^{2}-\frac{1}{2}\sigma^{2}\left( t_{i}^{n},t_{i+1}^{n}\right) \mathcal{I}_{d}\right) \right\Vert_{L^{2} (\Omega)} = 0.\end{aligned}$$ For the present work we need to address the same problem for the third order terms, namely the existence of an $L^{2}(\Omega)$-limit for sums of terms of the form $$\begin{aligned}
V^2(Y_{t_{i}^{n}}) \left( \mathbf{X}_{t_{i}^{n},t_{i+1}^{n}}^{3}\right),\end{aligned}$$ possibly after rebalancing, over a sequence of partitions with mesh tending to zero. An important discovery of this paper is the somewhat surprising conclusion that this these terms have vanishing $L^{2}(\Omega)-$limit, without the the need to subtract any rebalancing terms. This is the concluding result of Section \[Sko aug\].
(iii) The proof of point (ii) relies on a rather intricate interplay between estimates from Malliavin’s calculus and rough path analysis. From the latter theory, we need estimates on the directional derivatives of RDE solutions. It is well known that an RDE solution of the form can be differentiated in a direction $h\in C^{q-\text{var}}\left( \left[ 0,T\right], \mathbb{R}^{d}\right)$ by considering the perturbed RDE solution driven by the translated rough path $T_{\epsilon h}\mathbf{X}$ and then evaluating the derivative in $\epsilon$ at zero. For $T_{\epsilon h}\mathbf{X}$ to make sense, $\mathbf{X}$ and $h$ must have Young-complementary regularity, i.e. $\frac{1}{p} + \frac{1}{q} > 1$, in which case Duhamel’s formula gives $$\begin{aligned}
\label{first o}
D_{g}Y_{t}=\int_{0}^{t}J_{t}^{\mathbf{X}}\left( J_{s}^{\mathbf{X}}\right)^{-1}V\left( Y_{s}\right) \,\mathrm{d}g(s),\end{aligned}$$ a well-defined Young integral. In Malliavin calculus, $g$ will typically be an element of the Cameron-Martin space (written as $\mathcal{H}^{d}$ in this paper), and this has spurred interest in results that prove that $\mathcal{H}^{d}$ can be continuously embedded into $q$-variation spaces, see e.g. [@cfv2009], [@fggr2016]. By combining these results with Young’s inequality, one can then say e.g. that $$\begin{aligned}
\label{embed}
\left\vert D_{g}Y_{t}\right\vert \lesssim\left\vert g\right\vert_{q\text{-var}}\lesssim\left\vert g\right\vert _{\mathcal{H}^{d}},\end{aligned}$$ and these arguments can be generalized to higher order directional derivatives, allowing one control over the Hilbert-Schmidt norm of the Malliavin derivative; see [@inahama2014]. Note however, that quality is lost in by use of the embedding. For the proof in (ii) we need subtler estimates on the higher order derivatives of the form $$\begin{aligned}
\label{high o}
\left\vert D_{g_{i},....,g_{n}}^{n}Y_{t}\right\vert \leq C_{n}\left(\mathbf{X}\right)
{\textstyle\prod\limits_{j=1}^{n}} \left\vert g_{j}\right\vert _{q\text{-var}}.\end{aligned}$$ High order derivatives here complicate matters since derivatives of order $2$ and higher are no longer representable as Young integrals as in ; instead genuine rough integrals appear. Much of the work underpinning point (ii) goes into deriving closed-form expressions for these high-order derivatives and then estimating them so as to arrive at . We must also pay careful attention to the random variable $C_{n}\left( \mathbf{X}\right)$ in which, for our application, must have finite positive moments of all orders. The first half of Section \[Sko aug\] is devoted to this material.
The culmination of the these arguments is presented in Section \[corr formula\], where we give a set of conditions under which a conversion formula holds for $\int_{0}^{T}\left\langle Y_{t}\circ\mathrm{d}\mathbf{X}_{t}\right\rangle -\delta^{X}\left( Y\right)$. This formula is reminiscent of the one obtained for the case of second-order rough paths, but there are interesting differences too. Most notably the second term in , $$\begin{aligned}
\label{non commut}
\int_{\lbrack0,T]^{2}}\mathds{1}_{[0,t)}(s)\mathrm{tr}\left[ J_{t}^{\mathbf{X}}\left( J_{s}^{\mathbf{X}}\right) ^{-1}V(Y_{s})-V(Y_{t})\right] \,\mathrm{d}R(s,t),\end{aligned}$$ which exists for $2 \leq p < 3$ as a well-defined 2D Young-Stieltjes integral, can only be identified as an $L^{2}$-limit of a sequence of approximating sums. The difference between the two cases stems from the lack of complementary Young regularity of the integrand and $R.$ Interestingly the integrand, while being continuous on $\left[ 0,T\right]^{2}$, is not Hölder bi-continuous and so we cannot even appeal to the relaxed criteria discussed in point (1) above. It is unknown at present whether the limit is interpretable as a 2D Young-Stieltjes integral. We discuss in detail two important corollaries of our result. The first is where $X$ is a fractional Brownian motion with $H$ in $\left( \frac{1}{4}, \frac{1}{3} \right]$, and the second is the case where the vector fields defining commute. In this latter case, we show that the second term (\[non commut\]) in the correction formula disappears and, as a special case, we can recover Itô-type formulas for Gaussian processes, thus connecting our work to a substantial recent corpus e.g. [@np1998], [@amn2001], [@ccm2003], [@no2011] and [@hjt2013].
The bulk of the content in this paper can be found in the second-named author’s doctoral dissertation [@lim2016].
Preliminaries
=============
Rough path concepts and notation
--------------------------------
We briefly review the basic notation of rough paths theory used in this article; the standard references in this area [@lyons98], [@lq2003], [@fh2014] and [@fv2010b] can be consulted for more detail. We let $T^{n}\left( \mathbb{R}^{d}\right) $ denote the degree $n$ truncated tensor algebra $T^{n}\left( \mathbb{R}^{d}\right) :=\mathbb{R}\oplus\mathbb{R}^{d}\oplus\cdots\oplus\left( \mathbb{R}^{d}\right)^{\otimes n}$ equipped with addition and scalar multiplication as defined in the usual fashion. The truncated tensor product of $a=\left( a^{0},a^{1},\ldots,a^{n}\right) $ (alternatively written as $a^{0}+a^{1}+\ldots+a^{n}$) and $b=\left( b^{0},b^{1},\ldots,b^{n}\right)\in T^{n}\left( \mathbb{R}^{d}\right)$ is defined by $$\begin{aligned}
a\otimes b:=\left( c^{0},c^{1},\ldots,c^{n}\right), \qquad c^{k}=\sum
_{i=0}^{k}a^{i}\otimes b^{k-i},\quad\forall\,0\leq k\leq n,\end{aligned}$$ where here we abuse the notation by re-using the same symbol for the tensor product in $\mathbb{R}^{d}$. The unit element is $e=(1,0\ldots,0)$ and the tangent space of $T^{n}(\mathbb{R}^{d})$ at $e$ is denoted $A_{T}^{n}\left(\mathbb{R}^{d}\right)$. The exponential and logarithm maps, $\exp:A_{T}^{n}(\mathbb{R}^{d})\rightarrow T^{n}(\mathbb{R}^{d})$ and $\log: T^{n}(\mathbb{R}^{d})\rightarrow A_{T}^{n}(\mathbb{R}^{d}),$ are mutually inverse and defined by $$\begin{aligned}
\exp(a):=\sum_{i=0}^{n}\frac{a^{\otimes i}}{i!}, \qquad \log(a)=\sum_{i=1}^{n}(-1)^{i+1}\frac{(a-e)^{\otimes i}}{i}.\end{aligned}$$ The step-$n$ nilpotent group (with $d$ generators), denoted by $G^{n}\left(\mathbb{R}^{d}\right) $, is the subgroup of $T^{n}\left( \mathbb{R}^{d}\right)$ corresponding to the sub-Lie algebra of $A_{T}^{n}\left(\mathbb{R}^{d}\right)$ generated by the Lie bracket $[a,b]=a\otimes b-b\otimes a$. We equip it with any symmetric, sub-additive homogeneous norm $\left\Vert \cdot\right\Vert $ (cf. [@fv2010a]) which induces the left-invariant metric $d\left( a,b\right) = {\left\lVerta^{-1} \otimes b\right\rVert}$. Given $\mathbf{x}\in\mathcal{C}\left( [0,T];G^{n}\left( \mathbb{R}^{d}\right) \right)$, a continuous $G^{n}\left( \mathbb{R}^{d}\right)$-valued path, we define the increment over $\left[ s,t\right] \subseteq\left[ 0,T\right]$ by $\mathbf{x}_{s,t}:=\mathbf{x}_{s}^{-1}\otimes\mathbf{x}_{t}$, and then the $p$-variation distance is $$\begin{aligned}
\label{pvarDist}
d_{p-var;[0,T]}(\mathbf{x},\mathbf{y}):=\sup_{\pi}\left( \sum_{i} d(\mathbf{x}_{t_{i},t_{i+1}},\mathbf{y}_{t_{i},t_{i+1}})^{p}\right)^{\frac{1}{p}},\end{aligned}$$ where the supremum runs over all partitions $\pi=\{t_{i}\}$ of $[0, T]$. We let $\left\Vert \mathbf{x}\right\Vert _{p-var;[0,T]}:=d_{p-var;[0,T]} (\mathbf{x},0)$ $,$where $0$ denotes the constant path $\mathbf{y}_{t}=e$ for all $t \in [0, T]$.
For $p \geq 1$, the *weakly geometric p-rough paths*, which we will denote by\
$\mathcal{C}^{p-var}\left([0,T];G^{\lfloor p\rfloor} \left( \mathbb{R}^{d}\right) \right) $, is the set of continuous functions $\mathbf{x}$ from $[0,T]$ onto $G^{\lfloor
p\rfloor} \left( \mathbb{R}^{d} \right)$ such that ${\left\lVert\mathbf{x}\right\rVert}_{p-var; [0, T]} < \infty$.
The simplest example of a weakly geometric $p$-rough path is as follows, given a bounded-variation path $x$ in $\mathbb{R}^{d}$, we can compute the signature of $x$ in $G^{\lfloor p\rfloor} \left( \mathbb{R}^{d} \right)$: $$\begin{aligned}
S_{\lfloor p\rfloor }(x)_{s,t}=\left( 1,\mathbf{x}_{s,t}^{1},\mathbf{x}_{s,t}^{2},\ldots ,\mathbf{x}_{s,t}^{\lfloor p\rfloor }\right),\end{aligned}$$ where $\mathbf{x}_{s,t}^{k}$ is the conventional $k$-th iterated integral of the path $x$ over the interval $[s,t]$: $$\begin{aligned}
\mathbf{x}_{s,t}^{k} = \sum_{j_1, \ldots, j_k=1}^d \left( \int_{s<r_{1}<\cdots <r_{k}<t}\,\mathrm{d}x^{(j_1)}_{r_{1}}\otimes \cdots \otimes \,\mathrm{d}x^{(j_k)}_{r_{k}} \right) e_{j_1} \otimes \cdots \otimes e_{j_k}.\end{aligned}$$
For $p \geq 1$, the space of *geometric p-rough paths*, which we will denote by\
$\mathcal{C}^{0,p-var}\left( [0,T]; G^{\lfloor p \rfloor}\left( \mathbb{R}^{d}\right) \right)$, is defined to be the closure of $$\begin{aligned}
\mathcal{C}^{\infty} \left( [0, T]; G^{\lfloor p \rfloor} \left(\mathbb{R}^d\right) \right)
:= \left\{ f \in \mathcal{C}^{p-var} \left([0, T]; G^{\lfloor p \rfloor} \left(\mathbb{R}^d\right)\right) \: \Big| \: f \, \mathrm{smooth} \right\}\end{aligned}$$ with respect to the topology given by the $p$-variation distance .
The notion of finite $p$-variation extends to (non-necessarily continuous) paths\
$f:[0,T]\rightarrow E$ taking values in a metric space $(E,d)$ via $$\begin{aligned}
\left\Vert f\right\Vert _{p-var;[0,T]}:=\sup_{\pi}\left( \sum_{i}d(f_{t_{i}}, f_{t_{i+1}})_{E}^{p}\right)^{\frac{1}{p}}.\end{aligned}$$ Letting $V^{p-var}\left( [0,T]; E \right) := \{f \:|\: \left\Vert f\right\Vert_{p-var;[0,T]}<\infty\}$, we use $\mathcal{C}^{p-var}\left( [0,T];E\right)$ and\
$\mathcal{C}_{pw}^{p-var}\left( [0, T]; E \right) $ to denote the subsets of $V^{p-var}\left( [0,T]; E\right) $ consisting of functions which are also, respectively, continuous functions and piecewise continuous. Given $f$ in\
$\mathcal{C}^{p-var}\left( [0,T];E\right)$, the function $\omega(s,t):=\left\Vert f\right\Vert _{p-var;[s,t]}^{p}$ defined on the simplex $\left\{ \left( s, t \right) \:|\: 0 \leq s\leq t\leq T\right\} $ is a *control*, by which we mean it is a continuous, non-negative, function that is super-additive and satisfies $\omega(t,t) = 0$ for all $t \in [0,T]$. When $\left( E,\left\Vert \cdot\right\Vert _{E}\right) $ is a Banach space we can define a norm on $\mathcal{C}^{p-var}\left( [0,T];E\right) $ via $$\begin{aligned}
\left\Vert f\right\Vert _{\mathcal{V}^{p};[0,T]}:=\left\Vert f\right\Vert_{p-var;[0,T]}+\sup_{t \in [0, T]}\left\Vert f_{t}\right\Vert_{E}.\end{aligned}$$
We will also need the notion of $p$-variation for two-parameter functions. Thus, given $f:[0,T]^{2} \rightarrow E$, we say $f$ is of finite 2D $p$-variation if $$\begin{aligned}
\left\Vert f\right\Vert _{p-var;[0,T]^{2}}:=\sup_{\pi}\left( \sum_{i,j} \left\Vert f
\begin{pmatrix}
u_{i},u_{i+1}\\
v_{j},v_{j+1}
\end{pmatrix} \right\Vert_{E}^{p}\right) ^{\frac{1}{p}}<\infty,\end{aligned}$$ where $\pi=\left\{ \left( u_{i},v_{j}\right) \right\} $ is a partition of $[0,T]^{2}$, and the rectangular increment is given by $$\begin{aligned}
\label{rectInc}
f \begin{pmatrix}
u_{i},u_{i+1}\\
v_{j},v_{j+1}
\end{pmatrix}
:=f(u_{i},v_{j})+f(u_{i+1},v_{j+1})-f(u_{i},v_{j+1})-f(u_{i+1},v_{j}).\end{aligned}$$ On occasion, we will use the notation $$\begin{aligned}
f(\Delta_{i},v):=f(u_{i+1},v)-f(u_{i},v)\text{ and }f(u,\Delta_{j} ) := f(u,v_{j+1})-f(u,v_{j}).\end{aligned}$$
Let $f$ and $g$ be functions defined on $[0,T]^{2}.$ We say that the 2D Young-Stieltjes integral of $f$ with respect to $g$ exists if there exists a scalar $I(f,g)\in\mathbb{R}$ such that $$\begin{aligned}
\label{quant1}
\lim_{\left\Vert \pi\right\Vert \rightarrow0}\left\vert \sum_{i,j}f\left(
u_{i},v_{j}\right) g
\begin{pmatrix}
u_{i} & u_{i+1}\\
v_{j} & v_{j+1}\end{pmatrix}
-I(f,g)\right\vert \rightarrow0,\end{aligned}$$ i.e. for each $\varepsilon > 0$, there exists a $\delta>0$ such that for all partitions $\pi=\{(u_{i},v_{j})\}$ of $[0,T]^{2}$ with $\left\Vert \pi\right\Vert <\delta$, the quantity on the left of is less than $\varepsilon$. In this case, we use $\int_{[0,T]^{2}}f\,\mathrm{d}g$ to denote $I(f,g)$, or $\int_{[s,t]\times\lbrack u,v]}f\,\mathrm{d}g$ whenever we restrict ourselves to any particular subset $[s,t]\times\lbrack u,v]$ of $[0,T]^{2}$.
We say that $f\in\mathcal{C}^{p-var}([s,t]\times\lbrack u,v])$ and $g\in\mathcal{C}^{q-var}([s,t]\times\lbrack u,v])$ have complementary regularity if $p^{-1}+q^{-1}>1$.
The significance of this definition lies in the following theorem (see [@towghi2002], [@fv2010a]), which gives the existence of the Young-Stieltjes integral and Young’s inequality in two dimensions; see [@lcl2006], [@fh2014], [@fv2010b] for the one-dimensional version.
\[2Dintegral\] Let $f \in V^{p-var}([s, t] \times[u, v])$ and $g \in V^{q-var}([s, t] \times[u, v])$ have complementary regularity. We also assume that $f(s, \cdot)$ and $f(\cdot, u)$ have finite p-variation, and that $f$ and $g$ have no common discontinuities. Then the 2D Young-Stieltjes integral exists and the following Young’s inequality holds; $$\begin{aligned}
\label{2DYoungIneq}
\begin{split}
\left| \int_{[s,t] \times[u,v]} f \, \mathrm{d} g\right| \quad\leq C_{p, q}
\, {\left\vert \kern-0.25ex\left\vert \kern-0.25ex\left\vert f \right\vert
\kern-0.25ex\right\vert \kern-0.25ex\right\vert } \left\| g\right\| _{q-var,
[s,t] \times[u,v]},
\end{split}\end{aligned}$$ where $$\begin{aligned}
{\left\vert \kern-0.25ex\left\vert \kern-0.25ex\left\vert f \right\vert
\kern-0.25ex\right\vert \kern-0.25ex\right\vert } = \left| f(s, u)\right| +
\left\| f(s, \cdot)\right\| _{p-var; [u, v]} + \left\| f(\cdot, u)\right\|
_{p-var; [s, t]} + \left\| f\right\| _{p-var, [s,t] \times[u,v]}.\end{aligned}$$
Gaussian rough paths
--------------------
We will work with a stochastic process $$\begin{aligned}
X_{t} = \left( X_{t}^{(1)},\ldots, X_{t}^{(d)}\right) , \quad t \in[0, T],\end{aligned}$$ which denotes a centered (i.e. zero-mean), continuous Gaussian process in $\mathbb{R}^{d}$ with i.i.d. components.
This process is defined on the canonical probability space $\left(\Omega,\mathcal{F},\mathbb{P}\right) $, where $\Omega=\mathcal{C}\left(
[0,T];\mathbb{R}^{d}\right) $, the space of continuous $\mathbb{R}^d$-valued paths equipped with the supremum topology, $\mathcal{F}$ is the completion of the Borel $\sigma$-algebra generated by $X$, and $\mathbb{P}$ is the unique Borel measure under which $X \left( \omega\right) = \left( \omega_{t}\right)_t \in \left[ 0, T\right] $ has the specified Gaussian distribution. We will use $$\begin{aligned}
R(s,t):=\mathbb{E}\left[ X_{s}^{(1)}X_{t}^{(1)}\right]\end{aligned}$$ to denote the covariance function common to the components. The variance $R\left( t, t\right) $ will be denoted simply by $R(t)$, and we will also use the notation $$\begin{aligned}
\sigma^{2}(s,t):=R
\begin{pmatrix}
s & t\\
s & t
\end{pmatrix}
=\mathbb{E}\left[ \left( X_{s,t}^{(1)}\right)^{2}\right];\end{aligned}$$ recall the definition of the rectangular increment in .
The triple $\left( \Omega,\mathcal{H}^{d},\mathbb{P} \right) $ denotes the abstract Wiener space associated to $X$, where $\mathcal{H}^{d}=\bigoplus_{i=1}^{d}\mathcal{H}$ is the Cameron-Martin space (or reproducing kernel Hilbert space). The Cameron-Martin space, which is densely and continuously embedded in $\Omega\text{,}$ is the completion of the linear span of the functions $$\begin{aligned}
\left\{ R(t, \cdot)^{(u)} := R(t,\cdot) \, e_{u} \; \bigg\vert \; t\in[0,T],
\; u = 1, \ldots, d \right\}\end{aligned}$$ under the inner-product $$\begin{aligned}
\left\langle R(t, \cdot)^{(u)}, R(s,\cdot)^{(v)} \right\rangle _{\mathcal{H}^{d}} = \delta_{uv} \, R(t,s), \quad u, v = 1, \ldots, d.\end{aligned}$$ By definition, $\mathcal{H}^{d}$ satisfies the following reproducing property; for any $f=\left( f^{(1)},\ldots, f^{(d)}\right) \in\mathcal{H}^{d}$, $$\begin{aligned}
\left\langle f_{\cdot}, R(t,\cdot)^{(u)} \right\rangle_{\mathcal{H}^{d}}
= f_{t}^{(u)}, \quad t \in[0,T].\end{aligned}$$
We assume that there exists $\rho < 2$ such that $R$ has finite 2D $\rho$-variation. The following theorem in [@fv2010a] (see also [@cq2002] in the case of fractional Brownian motion) then shows that one can canonically lift $X$ via its piecewise linear approximants $X^{\pi}$ to a geometric $p$-rough path for $p > 2 \rho$.
\[gaussianRP\] Assume $X$ is a centered continuous $\mathbb{R}^d$-valued Gaussian process with i.i.d. components. Let $\rho \in [1, 2)$ and assume that the covariance function has finite 2D $\rho$-variation.
(i) (Existence) There exists a random variable $\mathbf{X}=\left( 1, \mathbf{X}^{1}, \mathbf{X}^2, \mathbf{X}^3\right) $ on $\left( \Omega, \mathcal{F},\mathbb{P}\right) $ which takes values in $\mathcal{C}^{0, p-var} \left( [0, T]; G^3(\mathbb{R}^d) \right)$ for $p > 2\rho$ almost surely, and is hence a geometric $p-$rough path for $p \in (2\rho, 4)$. Moreover, $\mathbf{X}$ lifts the Gaussian process $X$ in the sense that $\mathbf{X}_{s,t}^{1}=X_{t}-X_{s}$ almost surely for all $s,t\in [0, T]$.
(ii) (Uniqueness and consistency) The lift $\mathbf{X}$ is unique in the sense that it is the $d_{p -v a r}$-limit in $L^{q} (\Omega)$, $q \in [1,\infty )$, of any sequence $S_{^{\left \lfloor p\right \rfloor }} (X^{\pi}) $ with $\left\| \pi\right\| \rightarrow 0$. Furthermore, if $X$ has a.s. sample paths of finite $[1 ,2)$-variation, $\mathbf{X}$ coincides with the signature of $X$.
Moreover, Proposition 17 in [@fv2010a] shows that for all $h\in\mathcal{H}^{d}$, $$\begin{aligned}
\label{CMembedding}
\left\| h\right\| _{\rho-var; [0, T]} \leq\left\| h\right\| _{\mathcal{H}^{d}} \sqrt{\left\| R\right\| _{\rho-var; [0, T]^{2}}},\end{aligned}$$ which implies that $\mathcal{H}^{d} \hookrightarrow\mathcal{C}^{\rho-var}([0 ,T] ;\mathbb{R}^{d})$ whenever $R$ has finite 2D $\rho$-variation. Thus if $\rho\in\left[ 1, \frac{3}{2}\right) $, corresponding to $2 \leq p < 3$, we have complementary regularity between $X$ and any path in the Cameron-Martin space.
In the case $\rho\in\left[ \frac{3}{2}, 2 \right)$, there exists a recent result for complementary regularity between the Cameron-Martin paths and $X$ which requires the following definition.
We say that a function $f \in V^{(p, q)-var}([0, T]^{2})$ is of mixed (left) $(p, q)$-variation if $$\begin{aligned}
\sup_{\pi} \left( \sum_{i} \left( \sum_{j} \left| f
\begin{pmatrix}
s_{i} & s_{i+1}\\
t_{j} & t_{j+1}\end{pmatrix}
\right| ^{p} \right) ^{\frac{q}{p}} \right) ^{\frac{1}{q}} < \infty,\end{aligned}$$ where the supremum runs over all partitions $\pi:= \left\{ \left( s_{i}, t_{j} \right) \right\} $ of $[0, T]^{2}$.
Theorem 1 in [@fggr2016] states that if $R$ is of mixed $(1, \rho)$-variation, then $$\begin{aligned}
\left\| h\right\| _{q-var; [0, T]} \leq\left\| h \right\| _{\mathcal{H}^{d}}
\sqrt{\left\| R\right\| _{(1, \rho)-var; [0, T]^{2}}},\end{aligned}$$ where $q = \frac{2\rho}{\rho+ 1}$. One can easily verify that this gives us $\frac{1}{p} + \frac{1}{q} > 1$ as long as $p < 4$.
The following condition collects the assumptions we impose on $X$, or equivalently $\mathbf{X}$.
\[newCond1\] Let $X$ be a continuous, centered Gaussian process in $\mathbb{R}^{d}$ with i.i.d. components, and assume that the covariance function satisfies
(a) $\left\| R\right\| _{\rho-var; [0, T]^{2}} < \infty$ for some $\rho
\in\left[ 1, 2 \right) $.
For $p \in[1, 4)$, let $\mathbf{X} \in\mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^{d}\right) \right) $ denote the geometric rough path constructed from the limit of the piecewise-linear approximations of $X$. Furthermore, assume that $\mathcal{H}^{d} \hookrightarrow\mathcal{C}^{q-var} \left( [0, T]; \mathbb{R}^{d} \right) $, where $q$ satisfies $\frac{1}{p} + \frac{1}{q} > 1$, i.e. for all $h \in\mathcal{H}^{d}$,
(b) $\left\| h\right\| _{q-var; [0, T]} \leq C \left\| h\right\|
_{\mathcal{H}^{d}}$.
Later on, we will need to impose further conditions on the covariance function. For all $s, t \in[0 ,T]$, we will assume there exists $C < \infty$ such that $$\begin{aligned}
\label{R1DBound}
\left\| R(t, \cdot) - R (s, \cdot) \right\| _{q-var; [0, T]} \leq C \left| t - s\right| ^{\frac{1}{\rho}}.\end{aligned}$$ This bound will be used to control the $L^{2} (\Omega)$ norm of the iterated integrals. An immediate consequence of the bound is illustrated in the following lemma (Lemma 2.14 in [@cl2018]).
\[RdotRhoVar\] Let $X$ be a continuous, centered Gaussian process in $\mathbb{R}$ and assume its covariance function satisfies $$\begin{aligned}
\left\| R (t , \cdot) - R (s , \cdot)\right\| _{q-var; [0, T]} \leq C \left|
t -s\right| ^{\frac{1}{\rho}}, \quad\forall s < t \in[0, T],\end{aligned}$$ for some $q, \rho\geq1$. Then
(i) $R(t) := R(t, t)$ is of bounded $\rho$-variation.
(ii) For $p > 2 \rho$, $X$ has a $\frac{1}{p}$-Hölder continuous modification.
Volterra processes and fractional Brownian motion
-------------------------------------------------
A *Volterra kernel* $K$ is a square-integrable function $K:\left[ 0,T\right]^2 \rightarrow \mathbb{R}$ such that $K(t,s) = 0 \;\forall s\geq t$. Associated with any Volterra kernel is a lower triangular, Hilbert-Schmidt operator $\mathbb{K}: L^{2}\left( \left[ 0,T \right] \right) \rightarrow L^{2}\left( \left[ 0,T\right] \right) $ given by $$\begin{aligned}
\mathbb{K} \left( f\right) \left( \cdot\right) =\int_{0}^{T}K\left( \cdot,s\right) f\left( s\right) \text{ for all }f\in L^{2}\left( \left[ 0,T\right] \right) .\end{aligned}$$ Given a standard Brownian motion $B$ and a Volterra kernel $K,$we define a Volterra process $X = \left( X_{t} \right)_{t\in\left[ 0,T\right] }$ as the Itô integral $$\begin{aligned}
\label{voltrep}
X_{t}=\int_{0}^{T}K(t,s)\,\mathrm{d}B_{s};\end{aligned}$$ this is a centered Gaussian process with covariance function $$\begin{aligned}
R(s,t) = \int_{0}^{t\wedge s}K(t,r)K(s,r)\,\mathrm{d}r.\end{aligned}$$
(i) Standard fractional Brownian motion $B^{H}$, with Hurst parameter $H\in(0,1)$, is the centered Gaussian process with covariance function $$\begin{aligned}
\label{covfbm}
R\left( s,t\right) =\frac{1}{2}\left( s^{2H}+t^{2H}-\left\vert t-s \right\vert^{2H}\right).\end{aligned}$$ It is well-known that $B^{H}$ has a Volterra representation of the form where the kernel can be expressed as a particular hypergeometric function; cf. [@du1999].
(ii) The Riemann-Liouville process with Hurst parameter $H\in\left( 0, 1 \right) $ is determined by the kernel $K(t,s) := C_{H}(t-s)^{H-\frac{1}{2}} \mathds{1}_{[0, t)}(s)$. Like the fractional Brownian motion, it is a self-similar process with variance $t^{2H}$; however, it does not have stationary increments.
We will need the following condition on the kernel $K$.
\[amnCond\] There exists constants $C < \infty$ and $\alpha \in \left[ 0, \frac{1}{4} \right)$ such that
(i) $\left| K(t,s) \right| \leq C s^{-\alpha} (t - s)^{-\alpha}$ for all $0 < s < t \leq T$.
(ii) $\frac{\partial K(t,s)}{\partial t}$ exists for all $0 < s < t \leq T$ and satisfies $\left| \frac{\partial K(t, s)}{\partial t} \right| \leq C \left( t-s \right)^{-(\alpha + 1)}$.
The following proposition summarizes the properties of fractional Brownian motion which will be relevant in the sequel.
\[fBMCond\] Let $B^{H}$ be standard fractional Brownian motion with Hurst index $H\in\left( \frac{1}{4},1\right) $, and let $K$ be the square-integrable kernel associated with it. We have:
(i) For any $p>\frac{1}{H}$ the sample paths of $B^{H}$ are almost surely $\frac{1}{p}$-Hölder continuous. Furthermore, there exists a geometric rough path $\mathbf{X}\in\mathcal{C}^{0,p-var}\left( [0,T];G^{\lfloor p\rfloor}\left( \mathbb{R}^{d}\right) \right) $ which is the $d_{p\text{-var }}$-limit of the paths $S_{\left\lfloor p\right\rfloor }\left( X^{\pi}\right) $ as $\left\vert \left\vert \pi\right\vert \right\vert \rightarrow 0$.
(ii) $B^{H}$ satisfies Condition \[newCond1\] with $\rho=\frac{1}{2H}$ and $$\begin{aligned}
q = \begin{cases}
\: \frac{2\rho}{\rho+1} \quad \text{if }\frac{1}{4}<H\leq\frac{1}{3},\\
\: \frac{1}{\rho}\wedge1 \quad \text{if }\frac{1}{3} < H \leq 1.
\end{cases}\end{aligned}$$
(iii) If $\frac{1}{4} < H \leq \frac{1}{2}$, then the kernel $K$ satisfies Condition \[amnCond\] with $\alpha = \frac{1}{2} - H$.
(iv) The covariance function (\[covfbm\]) of $B^{H}$ satisfies:
1. $\left\Vert R(t,\cdot)-R(s,\cdot)\right\Vert _{q-var;[0,T]}
\leq C \left\vert t-s\right\vert ^{\frac{1}{\rho}}$, if $\frac{1}{4}<H\leq\frac{1}{2}$,
2. $R(t)=t^{2H}$ is of bounded variation and thus of finite q-variation for any $q \geq 1$.
The sample paths of fractional Brownian motion have $H - \varepsilon$-Hölder regularity for any $\varepsilon > 0$ by Kolmogorov’s continuity theorem, and thus have $\frac{1}{p}$-Hölder regularity for any $p > 2\rho = \frac{1}{H}$. The proof that it has finite 2D $\rho$-variation can be found in [@fv2010a]; see also [@fv2011].
In the case $1 \leq p < 2$, or $H > \frac{1}{2}$, the geometric rough path is simply $\left( 1, B^H_t \right)$, and for $H \leq \frac{1}{2}$, one can invoke Theorem \[gaussianRP\] to construct the geometric rough path. Now Condition \[newCond1\] is satisfied since we have $\mathcal{H}^d \hookrightarrow \mathcal{C}^{q-var} \left([0, T]; \mathbb{R}^d\right)$ from Proposition 17 in [@fv2010a] and Theorem 1 in [@fggr2016]. Note that the second case, which applies when $\frac{1}{4} < H \leq \frac{1}{3}$, follows from the fact that the covariance function has finite mixed $(1, \rho)$-variation (cf. [@fggr2016]).
For any $H \in (0, 1)$, we have (see Theorem 3.2 in [@du1999]) $$\begin{aligned}
\label{kBound1}
|K (t,s)| \leq C_{1, H} s^{-\left| H - \frac{1}{2} \right|} (t - s)^{-\left( \frac{1}{2} - H \right)},\end{aligned}$$ for all $0 < s < t \leq T$, and we also have $$\begin{aligned}
\label{kBound2}
\frac{\partial K(t, s)}{\partial t} = C_{2, H} \left( \frac{t}{s} \right)^{H - \frac{1}{2}} (t - s)^{- \left(\frac{3}{2} - H \right)};\end{aligned}$$ see [@ccm2003] and [@nualart2006]. Thus, Condition \[amnCond\] is satisfied by the kernel as a consequence of the bounds above.
Now to prove that $$\begin{aligned}
{\left\lVertR(t, \cdot) - R (s, \cdot) \right\rVert}_{q-var; [0, T]} \leq C \left| t - s\right|^{\frac{1}{\rho}},\end{aligned}$$ when $\frac{1}{4} < H \leq \frac{1}{2}$, we will adopt the method in [@fv2011], and find bounds for $$\begin{aligned}
{\left\lVert{\mathbb{E} \left[ B^H_{s,t} B^H_{\cdot}\right]}\right\rVert}_{q-var; [s, t]}, \quad
{\left\lVert{\mathbb{E} \left[ B^H_{s,t} B^H_{\cdot}\right]}\right\rVert}_{q-var; [0, s]}, \quad \mathrm{and} \;
{\left\lVert{\mathbb{E} \left[ B^H_{s,t} B^H_{\cdot}\right]}\right\rVert}_{q-var; [t, T]},\end{aligned}$$ for all $s, t$ in $[0, T]$. For the first quantity in the preceding line, we use the fact that when $H > \frac{1}{4}$, we have (see [@fggr2016]) $$\begin{aligned}
{\left\lVertR\right\rVert}_{1, \rho-var; [s,t]^2} \leq C {\left\lvertt - s\right\rvert}^{2H},\end{aligned}$$ and $$\begin{aligned}
{\left\lVert{\mathbb{E} \left[ B^H_{s,t} B^H_{\cdot}\right]}\right\rVert}_{q-var; [s, t]}
&\leq {\left\lVertR(t, \cdot) - R(s, \cdot)\right\rVert}_{\mathcal{H}} \sqrt{{\left\lVertR\right\rVert}_{1, \rho-var; [s,t]^2}} \\
&\leq C {\left\lvertt - s\right\rvert}^H {\left\lvertt - s\right\rvert}^H = C {\left\lvertt-s\right\rvert}^{2H}.\end{aligned}$$ Now let $\{r_i\}$ be a partition of $[0, s]$. Then $$\begin{aligned}
\sum_i \left( {\mathbb{E} \left[ B^H_{s,t} B^H_{r_i, r_{i+1}}\right]} \right)^q
&\leq \left( \sum_i {\left\lvert{\mathbb{E} \left[ B^H_{s,t} B^H_{r_i, r_{i+1}}\right]}\right\rvert} \right)^q \\
&= {\left\lvert{\mathbb{E} \left[ B^H_{s,t}B^H_s\right]} \right\rvert}^q\end{aligned}$$ since the disjoint increments of fractional Brownian motion have non-positive correlation when $H \leq \frac{1}{2}$. Taking the supremum over all partitions of $[0, s]$, we have $$\begin{aligned}
{\left\lVert{\mathbb{E} \left[ B^H_{s,t} B^H_{\cdot}\right]}\right\rVert}_{q-var; [0, s]}
\leq {\left\lvert{\mathbb{E} \left[ B^H_{s,t} B^H_s\right]}\right\rvert}
\leq {\left\lvertt - s\right\rvert}^{2H},\end{aligned}$$ where we note that ${\left\lvert{\mathbb{E} \left[ B^H_{s,t} B^H_u\right]}\right\rvert} < {\left\lvertt - s\right\rvert}^{2H}$ for all $u$ if $H \leq \frac{1}{2}$, cf. Lemma 5 in [@nnt2010]. The bound for ${\left\lVert{\mathbb{E} \left[ B^H_{s,t} B^H_{\cdot}\right]}\right\rVert}_{q-var; [t, T]}$ is shown in the same manner when we again exploit the fact that the disjoint increments have the same sign.
Given a Banach space $E$ and a kernel $K$ satisfying Condition \[amnCond\] for some $\alpha\in\left[ 0,\frac{1}{4}\right) $, we introduce the linear operator $\mathcal{K}^{\ast}$ (see [@amn2001], [@dfond2005]) $$\begin{aligned}
\label{kStarDefn}
\left( \mathcal{K}^{\ast}\phi\right) (s):=\phi\left( s\right)
K(T,s)+\int_{s}^{T}\left[ \phi\left( r\right) -\phi\left( s\right)
\right] K(\mathrm{d}r,s),\end{aligned}$$ where the signed measure $K(\mathrm{d}r,s):=\frac{\partial K(r,s)}{\partial r}\,\mathrm{d}r$. The domain $D\left( \mathcal{K}^* \right)$ of $\mathcal{K}^*$ consists of measurable functions $\phi:\left[0, T \right] \rightarrow E$ for which the integral on the right-hand side exists for all $s$ in $\left[ 0,T\right] $. In particular, if $\phi$ is a $\lambda$-Hölder continuous function in the norm of $E$ for some $\lambda > \alpha$, then $\phi \in D\left( \mathcal{K}^* \right) $ and $\mathcal{K}^* \phi$ is in $L^{2}([0,T];E)$. Note also that for any $a$ in $[0,T]$, $\phi\mathds{1}_{[0,a)}$ is in $D\left( \mathcal{K}^* \right) $ whenever $\phi$ is, and we have the identity $$\begin{aligned}
\label{truncatedKStar}
\mathcal{K}^* \left( \phi\mathds{1}_{[0,a)}\right) (s)=\mathds{1}_{[0,a)}(s)\left( \phi(s)K(a,s)+\int_{s}^{a}\left[ \phi(r)-\phi(s)\right] K(\mathrm{d}r,s)\right).\end{aligned}$$
Malliavin calculus {#MDsection}
------------------
We will primarily work with the following Hilbert space which is isomorphic to $\mathcal{H}^{d}$.
Let $\mathcal{H}_{1}^{d}$ denote the completion of the linear span of $$\begin{aligned}
\left\{ \mathds{1}_{[0,t)}^{(u)}(\cdot):=\mathds{1}_{[0,t)}(\cdot
)\,e_{u}\;\bigg\vert\;t\in\lbrack0,T],\;u=1,\ldots,d\right\}\end{aligned}$$ (cf. [@amn2001], [@nualart2006]) with respect to the inner-product given by $$\begin{aligned}
\left\langle \mathds{1}_{[0,t)}^{(u)}(\cdot),\mathds{1}_{[0,s)}^{(v)}
(\cdot)\right\rangle _{\mathcal{H}_{1}^{d}}=\delta_{uv}\,R(t,s),\end{aligned}$$ where $\delta$ denotes the Kronecker delta. Furthermore, let $\Phi:\mathcal{H}_{1}^{d}\rightarrow\mathcal{H}^{d}$ denote the Hilbert space isomorphism obtained from extending the map $\mathds{1}_{[0,t)}^{(u)}(\cdot) \mapsto R(t,\cdot)^{(u)}$, $t \in [0, T],\;u=1,\ldots,d$.
We record some basic properties about the Malliavin calculus. For simplicity, we assume here that $d=1;$ the case of $d \in \mathbb{N}$ case needs only minor modifications. First we recall that the map $\mathds{1}_{\left[ 0, t\right) } (\cdot) \mapsto X_{t}$ extends to a unique linear isometry $I$ from $\mathcal{H}_{1}$ to $L^{2}\left( \Omega\right) $. It follows that $I\left( h\right) $ is a mean-zero Gaussian random variable with variance $\left\Vert h\right\Vert_{\mathcal{H}_{1}}^{2}$.
Given a smooth function $f: \mathbb{R}^{n}\rightarrow\mathbb{R}$ of at most polynomial growth, the Malliavin derivative $\mathcal{D}F$ of the functional $F=f\left( I\left( h_{1}\right) ,\ldots,I\left( h_{n}\right) \right) $ is the $\mathcal{H}_{1}$-valued random variable given by $$\begin{aligned}
\mathcal{D}F:=\sum_{i=1}^{n}\frac{\partial f}{\partial x_{i}}\left( I\left(
h_{1}\right) ,\ldots,I\left( h_{n}\right) \right) h_{i}.\end{aligned}$$ We let $\mathbb{D}^{1,2}$ denote the Hilbert space that arises from completing this subspace of cylinder functionals with respect to $$\begin{aligned}
\left\Vert F\right\Vert _{1,2}^{2}:=\left\Vert F\right\Vert _{L^{2}\left(
\Omega\right) }^{2}+\left\Vert \mathcal{D}F\right\Vert _{L^{2}\left(
\Omega;\mathcal{H}_{1}\right) }^{2},\end{aligned}$$ whereupon $D$ extends to a bounded linear operator from $\mathbb{D}^{1,2}$ to $L^{2}\left( \Omega;\mathcal{H}_{1}\right) .$ More generally, Banach spaces $\mathbb{D}^{1,p}$ $\ $for $p>1$ can be defined by replacing $2$ above with $p.$ For any $F$ in $\mathbb{D}^{1,2}$, we let $\mathcal{D}_{h}F:=\left\langle \mathcal{D}F,h\right\rangle _{\mathcal{H}_{1}}$. The divergence operator $\delta^{X}$ is the $L^{2}\left( \Omega\right) $-adjoint of $D$, and its domain Dom$\left( \delta^{X}\right) $ consists of those $G$ in $L^{2}\left( \Omega;\mathcal{H}_{1}\right) $ for which there exists a $c>0$ with $$\begin{aligned}
\left\vert \mathbb{E}\left[ \left\langle DF,G\right\rangle _{\mathcal{H}_{1}}\right] \right\vert \leq c\left\Vert F\right\Vert _{L^{2}\left(
\Omega\right) }, \quad \forall F\in\mathbb{D}^{1,2}.\end{aligned}$$ For such $G$, the Riesz representation theorem then provides that $\delta^{X}\left( G\right) $ is characterized by $$\begin{aligned}
\mathbb{E}\left[ \left\langle DF,G\right\rangle_{\mathcal{H}_{1}}\right]
= \mathbb{E}\left[ F\delta^{X}\left( G\right) \right].\end{aligned}$$ Analogous definitions for $DF$ apply when $F$ is an $E-$valued random variable, for any separable Hilbert space $E$, in which case we will denote the Sobolev spaces by $\mathbb{D}^{1,p}\left( E\right) $. Higher order derivatives $D^{n}F$ and their corresponding Sobolev spaces $\mathbb{D}^{n,p}$ can then be defined iteratively, and the operator $\delta_{n}^{X}:$ Dom$\left( \delta_{n}^{X}\right) \subset L^{2}\left( \Omega; \mathcal{H}_{1}^{\otimes n}\otimes E\right) $ $\rightarrow L^{2}\left( \Omega\right) $ can be defined as the adjoint of $D^{n}F$ as above. We will use the notation $\delta^{X}(h)$ and $\int_{0}^{T}h_{s}\,\mathrm{d}X_{s}$ interchangeably to denote the divergence operator. It is well-known that the domain of $\delta^{X} $ contains $\mathbb{D}^{1,2}\left( \mathcal{H}_{1}\right) $, see e.g. Proposition 1.3.1 in [@nualart2006].
For deterministic $h\in\mathcal{H}_{1}$ we notice that $\delta^{X}\left( h\right) =I\left( h\right) $ as introduced above. More generally, by fixing a multi-index $a=(a_{1},\ldots,a_{M})$ where $\left\vert a\right\vert :=\sum_{i=1}^{M}a_{i}=n$, we can define $I_{n}:\mathcal{H}_{1}^{\otimes n}\rightarrow L^{2}\left( \Omega\right) $ as follows: $$\begin{aligned}
I_{n}\left( h_{1}^{\otimes a_{1}}\otimes\cdots\otimes h_{M}^{\otimes a_{M}}\right) =a!\prod_{i=1}^{M}H_{a_{i}}(\delta^{X}(h_{i})),\end{aligned}$$ where $a!:=\prod_{i=1}^{M}a_{i}!$ and $H_{m}(x)$ denotes the $m^{th}$ Hermite polynomial. Again for deterministic $h\in\mathcal{H}_{1}^{\otimes n} $ we have that $I_{n}\left( h\right) =\delta_{n}^{X}\left( h\right) $. In particular, we have the duality formula: $$\begin{aligned}
\label{dualityFormula}
\mathbb{E}\left[ FI_{n}(h)\right] =\mathbb{E}\left[ \left\langle D^{n}F,h\right\rangle _{\mathcal{H}_{1}^{\otimes n}}\right] .\end{aligned}$$ For $f\in\mathcal{H}_{1}^{\otimes n}$, $g\in\mathcal{H}_{1}^{\otimes m}$, both $f$ and $g$ symmetric, we also have the following product formula (cf. Proposition 1.1.3 in [@nualart2006]) $$\begin{aligned}
\label{productFormula}
I_{n}(f)I_{m}(g)=\sum_{r=0}^{n\wedge m}r!
\begin{pmatrix}
n\\
r
\end{pmatrix}
\begin{pmatrix}
m\\
r
\end{pmatrix}
I_{n+m-2r}\left( f\tilde{\otimes}_{r}g\right) .\end{aligned}$$ Here $f\tilde{\otimes}_{r}g$ denotes the symmetrization of the tensor $f\otimes_{r}g$, which in turn denotes the contraction of $f$ and $g$ of order $r$ [@nnt2010]; i.e. given any orthonormal basis $\{h_{m}\}$ of $\mathcal{H}_{1}$, $$\begin{aligned}
f\otimes_{r}g:=\sum_{k_{1},\ldots,k_{r}=1}^{\infty}\left\langle f,h_{k_{1}}\otimes\cdots\otimes h_{k_{r}}\right\rangle _{\mathcal{H}_{1}^{\otimes r}}\otimes\left\langle g,h_{k_{1}}\otimes\cdots\otimes h_{k_{r}}\right\rangle
_{\mathcal{H}_{1}^{\otimes r}}\in\mathcal{H}_{1}^{\otimes(n+m-2r)}.\end{aligned}$$
Rough differential equations
----------------------------
In this paper we will focus on the RDEs with time-homogeneous vector fields driven by a Gaussian geometric rough path $$\begin{aligned}
\label{RDE}
\mathrm{d}Y_{t}=V(Y_{t})\circ\mathrm{d}\mathbf{X}_{t},\quad Y_{0} = y_{0} \in \mathbb{R}^{e},\end{aligned}$$ where $\left\{ V_{1}, \ldots, V_{d} \right\} $ denotes a collection of $\mathbb{R}^e$-valued vector fields which will always be at least continuously differentiable. Recall also from Theorem 2.25 in [@cl2018] that for all $s, t \in [0, t]$, ${\left\lVertY\right\rVert}_{p-var; [s, t]}$ is in $L^q(\Omega)$ for all $q > 0$.
For $h_{1}, \ldots, h_{n} \in\mathcal{H}_{1}^{d}$, we can take the directional derivatives of $Y_{t}$ in the directions $\Phi(h_{1}), \ldots, \Phi(h_{n})$ in $\mathcal{H}^{d} \subset\mathcal{C}^{q-var} \left( \left[
0,T \right] , \mathbb{R}^{d}\right) $ by setting $$\begin{aligned}
\label{nderiv}
\mathcal{D}_{h_{1}, \ldots, h_{n}}^{n} Y_{t}
:= \frac{\partial^{n}}{\partial\varepsilon_{1} \ldots\partial\varepsilon_{n}}
Y_{t}^{\varepsilon_{1}, \ldots, \varepsilon_{n}} \bigg\vert_{\varepsilon_{1}
= \ldots = \varepsilon_{n}=0},\end{aligned}$$ where $Y_{t}^{\varepsilon_{1}, \ldots, \varepsilon_{n}}$ solves $$\begin{aligned}
\mathrm{d} Y^{\varepsilon_{1}, \ldots, \varepsilon_{n}}_{t} = V(Y^{\varepsilon
_{1}, \ldots, \varepsilon_{n}}_{t}) \circ\mathrm{d} \left( T_{\varepsilon_{1}
\Phi(h_{1}) + \cdots+ \varepsilon_{n} \Phi(h_{n})} \mathbf{X} \right) _{t}.\end{aligned}$$ Here $T_{\Phi(h)} \mathbf{X}$ denotes the rough path translation of $\mathbf{X}$ by $\Phi(h)$ (see [@cf2011]), which is well-defined via Young-Stieltjes integration since $\frac{1}{p} + \frac{1}{q} > 1$. The path again has finite $p$-variation and can be written as a sum of rough integrals and/or Young-Stieltjes integrals; e.g. when $n=1$ the first-order derivative is given by (cf. [@fv2010b], [@cf2011]) $$\begin{aligned}
\label{dd}\mathcal{D}_{h} Y_{t} = \int_{0}^{t} J_{t}^{\mathbf{X}} \left(
J_{s}^{\mathbf{X}}\right) ^{-1} V \left( Y_{s} \right) \, \mathrm{d}
\Phi(h)(s),\end{aligned}$$ and explicit formulas in the cases $n \geq2$ were derived in [@cl2018]. Here $J^{\mathbf{X}}_{t}$ denotes the Jacobian of the flow map $y_{0} \rightarrow Y_{t}$ and satisfies $$\begin{aligned}
\label{jacobianRDE}
\mathrm{d} J^{\mathbf{X}}_{t} = \nabla V (Y_{t}) \left( \circ\mathrm{d} \mathbf{X}_{t} \right) J^{\mathbf{X}}_{t}, \quad
J^{\mathbf{X}}_{0} = \mathcal{I}_{e}.\end{aligned}$$
To bound the Jacobian, we will define $$\begin{aligned}
\label{NDefn}
N^{\mathbf{X}}_{\beta;[s,t]} := \sup\left\{ n\in\mathbb{N} \cup\{0\} \: \big\vert \: \tau_{n}(\beta)<t\right\}, \quad s, t \in [0, T], \, \beta > 0,\end{aligned}$$ where $\{\tau_{i}(\beta)\}$ is the “greedy sequence” (see [@cll2013]) given by $$\begin{aligned}
& \tau_{0}(\beta)=s,\\
& \tau_{i+1}(\beta)=\inf\left\{ u\in(\tau_{i},t] \: \big\vert \: \left\Vert
\mathbf{X}\right\Vert _{p-var;[\tau_{i},u]}^{p}\geq\beta\right\} \wedge t.\end{aligned}$$ We then have the following theorem (Theorem 2.27 in [@cl2018]).
\[JBoundThm\] Let $p \geq 1$. Then for all $s < t \in[0, T]$, $\left\| J^{\mathbf{X}}\right\|_{p-var; [s, t]}$ is in $L^{q}(\Omega)$ for all $q > 0$.
Using the fact that $N_{1; [s, t]}^{\mathbf{X}}$ has Gaussian tails (see Theorem 6.3 in [@cll2013]), we see that ${\mathbb{E} \left[ \exp \left( C_2 q N_{1; [s, t]}^{\mathbf{X}} \right)\right]} < \infty$ for all $q > 0$, $s < t \in [0, T]$. Now from equation (4.10) in [@cll2013], we have the bound $$\begin{aligned}
\label{JBound}
{\left\lVertJ^{\mathbf{X}}\right\rVert}_{p-var; [s, t]} \leq C_1 {\left\lVert\mathbf{X}\right\rVert}_{p-var; [s, t]} \exp \left( C_2 N^{\mathbf{X}}_{1; [s, t]} \right).\end{aligned}$$ The statement of the theorem then follows immediately using Cauchy-Schwarz since\
${\left\lVert\mathbf{X}\right\rVert}_{p-var; [s, t]}$ also has moments of all orders.
Note that as in [@cl2018], we will abuse the notation and write $\mathcal{D}_{h_{1}, \ldots, h_{n}}^{n} Y_{t}$ and $\mathcal{D}_{\Phi(h_{1}), \ldots, \Phi(h_{n})}^{n} Y_{t}$ interchangeably.
Convergence in $\mathbb{D}^{1, 2} \left( \mathcal{H}_1^d\right)$ {#various}
================================================================
In this section, we will discuss the various isomorphisms and subspaces of the Cameron-Martin space and its tensor product. The motivation is as follows: let $Y$ be a solution to RDE and given a partition $\pi = \{ r_i \}$ of $[0, T]$, denote $$\begin{aligned}
Y^{\pi} (t) := \sum_i Y_{r_i} \mathds{1}_{\left[ r_i, r_{i+1} \right)} (t).\end{aligned}$$ Now recall the following inequality from Proposition 1.3.1 in [@nualart2006] $$\begin{aligned}
\label{itoskoro}
{\mathbb{E} \left[ \delta^X\left(Y^{\pi} - Y \right)^2\right]}
\leq {\mathbb{E} \left[ {\left\lVert Y^{\pi} - Y \right\rVert}^2_{\mathcal{H}_1^d}\right]} +
{\mathbb{E} \left[ {\left\lVert \mathcal{D} Y^{\pi} - \mathcal{D} Y \right\rVert}^2_{\mathcal{H}_1^d \otimes \mathcal{H}_1^d}\right]},\end{aligned}$$ which in particular implies that Dom$\left( \delta^{X}\right) \supseteq \mathbb{D}^{1,2}(\mathcal{H}_{1}^{d})$.
Thus if we can show that almost surely, $Y$ and $DY$ can be identified as elements of $\mathcal{H}_1^d$ and $\mathcal{H}_1^d \otimes \mathcal{H}_1^d$ respectively, and furthermore ${\left\lVertY^{\pi} - Y\right\rVert}_{\mathcal{H}_1^d} $ and ${\left\lVert\mathcal{D}Y^{\pi} - \mathcal{D}Y\right\rVert}_{\mathcal{H}_1^d \otimes \mathcal{H}_1^d}$ both vanish as ${\left\lVert\pi\right\rVert} \rightarrow 0$, then with further integrability assumptions one can use and dominated convergence to show that $\delta^X (Y^{\pi})$ converges to $\delta^X (Y)$ in $L^2(\Omega)$.
Convergence in $\mathcal{H}_{1}^{d}$
------------------------------------
This main aim of this subsection is to investigate the (almost sure) regularity required of $Y$ to identify it as an element of $\mathcal{H}_1^d$, and to have ${\left\lVertY^{\pi} - Y\right\rVert}_{\mathcal{H}_1^d} \rightarrow 0$. For Volterra processes, the first issue is to find criteria ensuring that the step-function approximations to a given Hölder continuous function converge in $\mathcal{H}_{1}^{d}$. We recall the following result from [@lim2018] (see also [@amn2000]).
\[nualartProp\] Let $\left( E,\left\Vert \cdot\right\Vert _{E}\right)$ be a Banach space and $K:\left[ 0,T\right] ^{2}\rightarrow\mathbb{R}$ be a kernel satisfying Condition \[amnCond\] for some $\alpha\in\left[ 0,\frac{1}{4}\right) $. Let $\phi:[0,T]\rightarrow E$ be $\lambda$-Hölder continuous, i.e. there exists $C<\infty$ such that $$\begin{aligned}
\left\Vert \phi(t_{1})-\phi(t_{2})\right\Vert _{E}\leq C\,|t_{1}
-t_{2}|^{\lambda},\quad\forall t_{1},t_{2} \in [0, T],\end{aligned}$$ and for any partition $\pi=\{s_{i}\}$ of $[0,T]$, let $\phi^{\pi
}:[0,T]\rightarrow E$ denote $$\begin{aligned}
\phi^{\pi}(t)=\sum_{i}\phi(s_{i})\mathds{1}_{[s_{i},s_{i+1})}(t).\end{aligned}$$ Then if $\lambda>\alpha$ we have $$\begin{aligned}
\lim_{\left\Vert \pi\right\Vert \rightarrow0}\int_{0}^{T}\left\Vert
\mathcal{K}^{\ast}\left( \phi^{\pi}-\phi\right) (t)\right\Vert _{E}
^{2}\,\mathrm{d}t=0,\end{aligned}$$ where $\mathcal{K}^*$ is defined as in .
Rather than dealing with the Hilbert space $\mathcal{H}_{1}^{d}$ as an abstract completion, it will be useful to realize it as a closed subspace of an $L^{2}$ space. To this end, we define $\mathcal{H}_{2}^{d}$ to be the closure in $L^{2}([0,T];\mathbb{R}^{d})$ of the linear subspace generated by $$\begin{aligned}
\left\{ K(t,\cdot)^{(u)}:=K(t,\cdot)\,e_{u}\;\bigg\vert\;t\in\lbrack
0,T],\;u=1,\ldots,d\right\}.\end{aligned}$$ The inner-product is the usual one in $L^{2}([0,T];\mathbb{R}^{d})$, namely $\left\langle f, g \right\rangle_{\mathcal{H}_{2}^{d}} = \int_{0}^{T} \left\langle f_{s}, g_{s} \right\rangle {\, \mathrm{d} s}$ where $\left\langle \cdot, \cdot \right\rangle$ denotes the Euclidean inner product in $\mathbb{R}^{d}$. The following proposition is more or less immediate.
\[h1 and h2\] $\mathcal{H}_{1}^{d}$ and $\mathcal{H}_{2}^{d}$ are isomorphic as Hilbert spaces
Since $$\begin{aligned}
R(s, t) = \int_0^{s \wedge t} K(s, r) K(t, r) {\, \mathrm{d} r},\end{aligned}$$ we have $$\begin{aligned}
\left\langle K(t, \cdot)^{(u)}, K(s, \cdot)^{(v)} \right\rangle_{L^2([0, T];
\mathbb{R}^d)} = \delta_{uv} R(s, t) = \left\langle \mathds{1}_{[0, t)}^{(u)} (\cdot), \mathds{1}_{[0, s)}^{(v)} (\cdot) \right\rangle_{\mathcal{H}_1^d}\end{aligned}$$ for all $s, t \in [0, T], u, v = 1, \ldots, d$. From the definition of $\mathcal{K}^*$ in , we have $$\begin{aligned}
\mathcal{K}^* \left( \mathds{1}_{[0, t)}^{(u)} (\cdot) \right) (s) = K(t, s)^{(u)}, \quad u = 1, \ldots, d,\end{aligned}$$ which means that $\mathcal{K}^*$ extends uniquely to a linear isometry from $\mathcal{H}_{1}^d$ onto $\mathcal{H}_{2}^d$ so that $$\begin{aligned}
\label{iso3}
\left\langle f, g\right\rangle _{\mathcal{H}_{1}^d} = \left\langle \mathcal{K}^* f, \mathcal{K}^* g \right\rangle_{\mathcal{H}_2^d} \quad \forall f, g\in\mathcal{H}_{1}^d.\end{aligned}$$
In the case of standard Brownian motion the isomorphism $\mathcal{K}^{\ast}$ is the identity operator and $\mathcal{H}_{1}^{d}=\mathcal{H}_{2}^{d} = L^{2}([0,T];\mathbb{R}^{d})$.
Since the RDE solutions we work with are path-valued, it will be convenient to find subspaces of $\mathcal{H}_{1}^{d}$ whose elements are actual paths. We let $$\begin{aligned}
\Lambda_{\alpha}^{d}\ :=\bigcup_{\lambda>\alpha}\mathcal{C}_{pw}
^{\lambda-H\ddot{o}l}\left( [0,T];\mathbb{R}^{d}\right),\end{aligned}$$ where $\mathcal{C}_{pw}^{\lambda-H\ddot{o}l}\left( [0,T];\mathbb{R}^{d}\right) $ denotes the space of piecewise $\lambda$-Hölder continuous functions. By equipping $\Lambda_{\alpha}^{d}$ with the inner product $$\begin{aligned}
\left\langle f,g\right\rangle _{\Lambda_{\alpha}^{d}}:=\left\langle
\mathcal{K}^{\ast}(f),\,\mathcal{K}^*(g)\right\rangle_{L^{2}\left(
[0,T];\mathbb{R}^{d}\right) },\end{aligned}$$ whilst suppressing its dependence on $K$ in the notation, the following proposition shows that we can regard $\Lambda_{\alpha}^{d}$ as a dense subspace of $\mathcal{H}_{1}^{d}$.
\[A1eqH1\] Suppose $K$ is a kernel satisfying Condition \[amnCond\] for some $\alpha \in \left[0, \frac{1}{4} \right)$. Then $\Lambda_{\alpha}^{d}$ is a dense subspace of $\mathcal{H}_{1}^{d}$, and the inclusion map $i:(\Lambda_{\alpha
}^{d},\,\left\langle \cdot,\cdot\right\rangle _{\Lambda_{\alpha}^{d}})\rightarrow(\mathcal{H}_{1}^{d},\,\left\langle \cdot,\cdot\right\rangle_{\mathcal{H}_{1}^{d}})$ is an isometry.
Let $f \in \Lambda^d_{\alpha}$ and let $\pi(n) = \left\{ r_i^{(n)}\right\}$ be a sequence of partitions whose mesh vanishes as $n \rightarrow \infty$. We denote $$\begin{aligned}
f^{\pi(n)} (t) := \sum_i f\left( r^{(n)}_i \right) \mathds{1}_{\left[ r_i^{n}, r_{i+1}^{(n)} \right)} (t),\end{aligned}$$ and note that for each $n$, $f^{\pi(n)}$ is in $\Lambda_{\alpha}^d \cap \mathcal{H}_1^d$. Moreover, Proposition \[nualartProp\] tells us that ${\left\lVert\mathcal{K}^* \left( f^{\pi(n)} - f \right)\right\rVert}_{L^2 \left([0, T]; \mathbb{R}^d\right)} \rightarrow 0$. Hence, using , we see that $f^{\pi(n)}$ is Cauchy in $\mathcal{H}^d_1$. We again identify $f$ with the limit of the sequence, and under this identification we have $$\begin{aligned}
\label{LH1norm}
{\left\lVertf\right\rVert}_{\mathcal{H}^d_1}^2 = {\left\lVert\mathcal{K}^*(f)\right\rVert}_{L^2\left([0, T]; \mathbb{R}^d\right)}.\end{aligned}$$ $\Lambda^d_{\alpha}$ contains all the generating functions $\left\{ \mathds{1}^{(u)}_{[0, t)} (\cdot) \right\}$ of $\mathcal{H}^d_1$, and so its closure is $\mathcal{H}^d_1$.
We recall from [@cl2018] a similar result in terms of $p$-variation. In that paper, $\mathcal{H}_{1}^{d}$ was derived from a Gaussian covariance function $R$ which was assumed to be of finite 2D $\rho$-variation, $\rho \in [1, 2)$. It was shown that $$\begin{aligned}
\label{Wd}
\mathcal{W}_{\rho}^{d}
:= \bigcup_{q < \frac{\rho}{\rho - 1}} \mathcal{C}_{pw}^{q-var}\left( [0,T];\mathbb{R}^{d}\right),\end{aligned}$$ when equipped with the inner product $$\begin{aligned}
\left\langle f,g\right\rangle _{\mathcal{W}_{\rho}^{d}}:=\int_{[0,T]^{2}}\left\langle f_{s},\,g_{t}\right\rangle _{\mathbb{R}^{d}}\,\mathrm{d}R(s,t),\end{aligned}$$ is a dense subspace of $\mathcal{H}_{1}^{d}$ with the inclusion map again being an isometry. In the case when $\lambda > \alpha \wedge \left( 1 - \frac{1}{\rho} \right) $, any $f$ and $g$ belonging to $\mathcal{C}_{pw}^{\lambda-H\ddot{o}l} \left( [0, T]; \mathbb{R}^{d}\right) $ also belong to $\mathcal{W}_{\rho}^{d}\cap\Lambda_{\alpha}^{d}$, and we have $$\begin{aligned}
\left\langle f,g\right\rangle _{\mathcal{W}_{\rho}^{d}}=\int_{[0,T]^{2}
}\left\langle f_{s},g_{t}\right\rangle _{\mathbb{R}^{d}}\,\mathrm{d}R\left(
s,t\right) =\int_{0}^{T}\left\langle \mathcal{K}^{\ast}f(r),\mathcal{K}
^{\ast}g(r)\right\rangle _{\mathbb{R}^{d}}\,\mathrm{d}r=\left\langle
f,g\right\rangle _{\Lambda_{\alpha}^{d}}\end{aligned}$$
The follow figure depicts schematically the relationship between the various subspaces in the case of a Volterra process satisfying Condition \[amnCond\]. For convenience we have assumed we are in the scalar-valued case $d=1$.
\[xscale = 0.95, yscale = 0.6\] (0, 0) circle (2.6) (-0.8, 2) node [$\mathcal{W}_{\rho}$]{}; (1.5, 0) circle (2.6) (2.3, 2) node [$\Lambda_{\alpha}$]{}; (0.75, 0) circle (3.5) (0.75,3) node [$\mathcal{H}_1$]{} (0.75,0) node [$\left\langle \mathds{1}_{[0, t)}(\cdot), \mathds{1}_{[0, s)} (\cdot) \right\rangle_{\mathcal{H}_1}$]{}; (5, 7.5) circle (3.5) (5,10) node [$\mathcal{H}$]{} (5, 7.5) node [$\left\langle R(t, \cdot), R(s, \cdot) \right\rangle_{\mathcal{H}}$]{}; (9.25, 0) circle (3.5) (9.25,3) node [$\mathcal{H}_2$]{} (9.25,0) node [$\left\langle K(t, \cdot), K(s, \cdot) \right\rangle_{\mathcal{H}_2}$]{}; (9.25, 0.25) circle (4.1) (9.25, 3.85) node [$L^2\left([0, T]; \mathbb{R} \right)$]{}; (5, 2.5) node [$R(t, s)$]{}; (\[shift=[(5, 2.5)]{}\] 246.5:6) arc\[radius=6, start angle=246.5, end angle = 293.5\]; (\[shift=[(5, 2.5)]{}\] 6:6) arc\[radius=6, start angle = 6, end angle = 54\]; (\[shift=[(5, 2.5)]{}\] 174:6) arc\[radius=6, start angle = 174, end angle = 126\]; (5, -4 ) node [$\mathcal{K}^*$]{} (10.6, 5.75) node [$K$]{} (-0.6, 5.75) node [$\Phi$]{}; (4.85, 3) – (4.85, 7) (5.15, 3) – (5.15, 7); (5.45, 2) – (7.45, 0.3) (5.6, 2.3) – (7.6, 0.6); (4.5, 2) – (2.3, 0.3) (4.35, 2.3) – (2.15, 0.6);
Note that $K$ gives an isomorphism from $\mathcal{H}_{2}^{d}$ onto $\mathcal{H}^{d} $ because $R(t, \cdot) = \int_{0}^{T} K(\cdot, r) K(t, r) \,\mathrm{d} r$.
Convergence in $\mathcal{H}_{1}^{d}\otimes\mathcal{H}_{1}^{d}$ {#noCompRegularity}
--------------------------------------------------------------
The results of the previous subsection allow us to interpret RDE solutions (paths) as $\mathcal{H}_{1}^{d}$-valued random variables. The Malliavin derivatives of these random variables, when they exist, will take values in $\mathcal{H}_{1}^{d}\otimes\mathcal{H}_{1}^{d}$, and we therefore need similar results which identify suitable function spaces which are subspaces of this tensor product space. We will develop this point in the current subsection.
Throughout, $E$ will denote a general Banach space with norm $\left\Vert \cdot \right\Vert _{E}$. The following operator was defined in [@lim2018].
\[kStarTensorOp\] Let $\mathcal{K}^{\ast}\otimes\mathcal{K}^{\ast}$ denote the operator $$\begin{aligned}
(\mathcal{K}^{\ast}\otimes\mathcal{K}^{\ast})\psi(u,v)
&:= \psi (u,v)K(T,v)K(T,u) + K(T,v)A^{K}\big(\psi(\cdot,v)\big)(u) \\
&\qquad+ K(T,u) A^{K} \big(\psi(u,\cdot)\big)(v) + B^{K}(\psi)(u,v),\end{aligned}$$ where $$\begin{aligned}
& A^{K}(\phi)(s):=\int_{s}^{T}\left[ \phi(r)-\phi(s)\right] K(\mathrm{d}r,s)\\
& B^{K}(\psi)(u,v):=\int_{v}^{T}\int_{u}^{T}\psi\begin{pmatrix}
u & r_{1}\\
v & r_{2}\end{pmatrix}
K(\mathrm{d}r_{1},u)K(\mathrm{d}r_{2},v),\end{aligned}$$ which is defined for any measurable function $\psi:[0,T]^{2}\rightarrow E$ for which the integrals on the right side exist.
Using Proposition \[h1 and h2\] and the fact that $$\begin{aligned}
\label{kStarProduct}
\left( \mathcal{K}^{\ast}\otimes\mathcal{K}^{\ast}\right) \psi(s,t)=\left(
\mathcal{K}^{\ast}\psi_{1}\right) (s)\otimes\left( \mathcal{K}^{\ast}\psi_{2}\right) (t)\end{aligned}$$ when $\psi(s, t) = \psi_1(s) \psi_2(t)$, it is also clear that $\mathcal{K}^{\ast}\otimes\mathcal{K}^{\ast}$ maps $\mathcal{H}_{1}^{d}\otimes\mathcal{H}_{1}^{d}$ isometrically onto $\mathcal{H}_{2}^{d} \otimes\mathcal{H}_{2}^{d}$, which is a closed subspace of $L^{2}\left([0,T];\mathbb{R}^{d}\right) \otimes L^{2}\left( [0,T];\mathbb{R}^{d}\right)
\cong L^{2}\left( [0,T]^{2};\mathbb{R}^{d}\otimes\mathbb{R}^{d}\right)$.
To go beyond product functions in the domain of $\mathcal{K}^* \otimes\mathcal{K}^{\ast}$, we also recall the class of strongly Hölder bi-continuous functions from [@lim2018].
\[biHolderDef\] Let $0 < \lambda \leq 1$. We say that a function $\phi:[0,T]^{2}\rightarrow E$ is strongly $\lambda$-Hölder bi-continuous in the norm of $E$ (or simply strongly $\lambda $-Hölder bi-continuous in the case where $E$ is finite-dimensional), if for all $u_{1},u_{2},v_{1},v_{2}\in\lbrack0,T]$ we have $$\begin{aligned}
\sup_{v\in\left[ 0,T\right] }\left\Vert \phi(u_{2},v)-\phi(u_{2}
,v)\right\Vert _{E}\leq C\,\left\vert u_{2}-u_{1}\right\vert ^{\lambda}
,\quad\sup_{u\in\left[ 0,T\right] }\left\Vert \phi(u,v_{2})-\phi
(u,v_{1})\right\Vert _{E}\leq C\,\left\vert v_{2}-v_{1}\right\vert ^{\lambda},\end{aligned}$$ and $$\begin{aligned}
\label{biHolder2}
\left\Vert \phi
\begin{pmatrix}
u_{1} & u_{2}\\
v_{1} & v_{2}
\end{pmatrix}
\right\Vert _{E}\leq C\,\left\vert u_{2}-u_{1}\right\vert ^{\lambda}\left\vert v_{2}-v_{1}\right\vert ^{\lambda}.\end{aligned}$$
The following proposition is one of the main results of [@lim2018].
\[nualartPropNew\] Let $\psi:[0,T]^{2}\rightarrow E$ be a function which is strongly $\lambda$-Hölder bi-continuous in the norm of $E$. For any partition $\{(u_{i},v_{j})\}$ of $[0,T]^{2}$, let $\psi^{\pi}:[0,T]^{2}\rightarrow E$ denote $$\begin{aligned}
\psi^{\pi}(u,v):=\sum_{i,j}\psi(u_{i},v_{j})\mathds{1}_{[u_{i},u_{i+1})}(u)\mathds{1}_{[v_{j},v_{j+1})}(v).\end{aligned}$$ In addition, let $\mathcal{K}^{\ast}\otimes\mathcal{K}^*$ denote the operator in Definition \[kStarTensorOp\], where the Volterra kernel $K$ satisfies Condition \[amnCond\] for some $\alpha \in \left[ 0, \frac{1}{4} \right) $. Then if $\lambda>\alpha$, we have $$\begin{aligned}
\lim_{\left\Vert \pi\right\Vert \rightarrow0}\int_{[0,T]^{2}}\left\Vert
\left( \mathcal{K}^* \otimes\mathcal{K}^{\ast}\left( \psi^{\pi} - \psi\right) \right) (u,v)\right\Vert _{E}^{2}\,\mathrm{d}u\,\mathrm{d}v,\end{aligned}$$ and $$\begin{aligned}
\lim_{\left\Vert \pi\right\Vert \rightarrow0}\int_{0}^{T}\left\Vert \left(
\mathcal{K}^{\ast}\otimes\mathcal{K}^{\ast}\left( \psi^{\pi}-\psi\right)
\right) (r,r)\right\Vert _{E}\,\mathrm{d}r=0.\end{aligned}$$
For this paper, the result above, coupled with the fact that $\mathcal{H}_1^d \otimes \mathcal{H}_1^d$ is isomorphic to $\mathcal{H}_2^d \otimes \mathcal{H}_2^d$, shows that the strongly $\lambda$-Hölder bi-continuous functions are contained in $\mathcal{H}_{1}^{d} \otimes \mathcal{H}_{1}^{d}$ for the class of Volterra kernels we are considering. For orientation here, contrast this to Proposition \[nualartProp\], which showed a similar inclusion in $\mathcal{H}_{1}^{d}$ for the class of $\lambda$-Hölder continuous functions.
The Malliavin derivative and convergence in the tensor norm {#mall conv}
-----------------------------------------------------------
Here, we will apply the results of the last subsection to the Malliavin derivatives of RDE solutions. When $\mathbf{X} \in \mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^d\right) \right)$ satisfies Condition \[newCond1\], for all $h \in \mathcal{H}_1^d$, $\Phi(h)$ can be embedded in $\mathcal{C}^{q-var}\left( \left[ 0,T \right]; \mathbb{R}^{d}\right)$ where $\frac{1}{p} + \frac{1}{q} > 1$. Furthermore, the Malliavin derivative of $Y$ satisfying $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ is given by $$\begin{aligned}
\mathcal{D}_h Y_t = \int_{0}^{t} J_{t}^{\mathbf{X}} \left( J_{s}^{\mathbf{X}}\right)^{-1} V\left( Y_{s} \right) \, \mathrm{d} \Phi(h)(s)
= \int_{0}^{T} \mathds{1}_{[0,t)} \left( s\right) J_{t}^{\mathbf{X}} \left( J_{s}^{\mathbf{X}}\right)^{-1} V\left( Y_{s} \right) \, \mathrm{d} \Phi(h)(s).\end{aligned}$$ Denoting $$\begin{aligned}
\label{D as path}
\mathcal{D}_{s} Y_{t} = \mathds{1}_{[0, t)} (s) J_{t}^{\mathbf{X}} \left(J_{s}^{\mathbf{X}} \right)^{-1} V\left( Y_{s} \right)\end{aligned}$$ with respect to any partition $\pi = \{ r_i \}$ of $[0, T]$, we will write $$\begin{aligned}
\mathcal{D}_s Y^{\pi}_t
&= \sum_i \mathcal{D}_s Y_{r_i} \mathds{1}_{\left[ r_i, r_{i+1} \right)} (t).\end{aligned}$$ We will proceed to show that
(i) $\mathcal{D} Y^{\pi}$ lies in $\mathcal{H}_1^d \otimes \mathcal{H}_1^d$ almost surely,
and under suitable regularity assumptions on $\mathcal{D} Y$, we have
(ii) ${\left\lVert\mathcal{D} Y^{\pi} - \mathcal{D}Y\right\rVert}_{\mathcal{H}_1^d \otimes \mathcal{H}_1^d} \rightarrow 0$ as ${\left\lVert\pi\right\rVert} \rightarrow 0$.
Coupled with the results in the previous subsections, this will mean that $Y^{\pi}$ converges to $Y$ in $\mathbb{D}^{1,2} \left(\mathcal{H}_1^d\right)$, and $\delta^X(Y)$ is then the $L^2(\Omega)$ limit of $\delta^X \left( Y^{\pi} \right)$.
A potential problem with is the discontinuity at the diagonal $\{s = t\}$, which prevents it from being Hölder bi-continuous. The next two propositions show how to handle discontinuities of this form.
Given a Banach space $\left( E,\left\Vert \cdot\right\Vert _{E}\right) $, let $\psi:[0,T]^{2}\rightarrow E$ be of the form $$\begin{aligned}
\psi(u,v)=\mathds{1}_{[0,v)}(u)\tilde{\psi}(u,v),\end{aligned}$$ where $\tilde{\psi}:[0,T]^{2}\rightarrow E$ is strongly $\lambda$-Hölder bi-continuous in the norm of $E$. Assume that $K$ is a Volterra kernel which satisfies Condition \[amnCond\] for some $\alpha \in \left[ 0, \frac{1}{4} \right) $ and let $\mathcal{K}^{\ast}\otimes\mathcal{K}^{\ast}$ be the operator given in Definition \[kStarTensorOp\]. Then if $\lambda>\alpha$, $\left( \mathcal{K}^{\ast} \otimes\mathcal{K}^{\ast}\right) \psi$ is in $L^{2}([0,T]^{2};E)$.
We will investigate the integrability of $$\begin{aligned}
\label{kStarTensorOp2}
\begin{split}
\mathcal{K}^* \otimes \mathcal{K}^* \psi (u, v)
&= \psi(u, v) K(T, u) K(T, v) + K(T, v) A^K(\psi(\cdot, v)) (u) \\
&\qquad \qquad+ K(T, u) A^K(\psi(u, \cdot)) (v) + B^K(\psi) (u, v)
\end{split}\end{aligned}$$ in the regions $\{ u < v\}$ and $\{ v < u \}$ separately (ignoring the diagonal as it has zero Lebesgue measure).
\(i) $u < v$:
For the first term on the right of we have $$\begin{aligned}
\psi(u, v) K(T, u) K(T, v) = \tilde{\psi} (u, v) K(T, u) K(T, v) \in L^2([0, T]^2; E),\end{aligned}$$ and for the second term, we have $$\begin{aligned}
&{\left\lVertK(T, v) A^K \big( \psi(\cdot, v) \big) (u)\right\rVert}_E \\
&\qquad \qquad= {\left\lVertK(T, v) \left( \int_u^v \left[ \tilde{\psi}(r, v) -\tilde{\psi}(u, v) \right] K(\mathrm{d}r, u) - \int_v^T \tilde{\psi}(u, v) K(\mathrm{d}r, u) \right) \right\rVert}_E \\
&\qquad \qquad \leq C {\left\lvertK(T, v)\right\rvert} \left( (v - u)^{\lambda - \alpha} + \left( \frac{1}{(v- u)^{\alpha}} - \frac{1}{(T - u)^{\alpha}} \right) \right) \in L^2([0, T]^2).\end{aligned}$$ The third term satisfies $$\begin{aligned}
{\left\lVert K(T, u) A^K \big( \psi (u, \cdot) \big) (v) \right\rVert}_E
&= {\left\lVert K(T, u) \int_v^T \left[ \tilde{\psi}(u, r) - \tilde{\psi}(u, v) \right] K(\mathrm{d}r, v) \right\rVert}_E \\
&\leq C {\left\lvertK(T, u)\right\rvert} (T - v)^{\lambda - \alpha} \in L^2([0, T]^2).\end{aligned}$$ For the fourth term, given $r_1 \in (u, T]$, we have $$\begin{aligned}
(u, T] \times (v, T] = \big\{ (u, v] \times (v, T] \big\} \bigsqcup \big\{ (v, T] \times (v, r_1] \big\} \bigsqcup \big\{ (v, T] \times (r_1, T] \big\},\end{aligned}$$ and thus $$\begin{aligned}
{\left\lVertB^K(\psi) (u, v)\right\rVert}_E
&= \left\| \int_u^v \left( \int_v^T \tilde{\psi} \begin{pmatrix}
u & r_1 \\
v & r_2
\end{pmatrix} K(\mathrm{d}r_2, v) \right) K(\mathrm{d}r_1, u) \right. \\
&\qquad \left. + \int_v^T \left( \int_{r_1}^T \left[ \tilde{\psi}(r_1, r_2) + \tilde{\psi}(u, v) - \tilde{\psi}(u, r_2) \right] K(\mathrm{d}r_2, v) \right) K(\mathrm{d}r_1, u) \right. \\
&\qquad \left. + \int_v^T \left( \int_v^{r_1} \left[ \tilde{\psi}(u, v) - \tilde{\psi}(u, r_2) \right] K(\mathrm{d}r_2, v) \right) K(\mathrm{d}r_1, u) \right\|_E.\end{aligned}$$ This expression is bounded above by $$\begin{aligned}
C \left( (v - u)^{\lambda - \alpha} (T - v)^{\lambda - \alpha} + \int_v^T \frac{1}{(r_1 - v)^{\alpha} (r_1 - u)^{\alpha + 1}} {\, \mathrm{d} r_1} + \left( \frac{1}{(v - u)^{\alpha}} - \frac{1}{(T - u)^{\alpha}} \right) \right).\end{aligned}$$ Since $$\begin{aligned}
\label{estL2}
\begin{split}
\int_v^T \frac{1}{ (r_1 - v)^{\alpha} (r_1 - u)^{\alpha + 1}} {\, \mathrm{d} r_1}
&= \int_v^T \frac{1}{(r_1 - v)^{\alpha} (r_1 - u)^{\alpha + \frac{1}{4}} (r_1 - u)^{\frac{3}{4}}} {\, \mathrm{d} r_1} \\
&\leq \frac{1}{(v - u)^{\alpha + \frac{1}{4}}} \int_v^T \frac{1}{(r_1 - v)^{\alpha + \frac{3}{4}}} {\, \mathrm{d} r_1},
\end{split}\end{aligned}$$ and $\alpha < \frac{1}{4}$, the fourth term is also in $L^2([0, T]^2; E)$.
\(ii) $v < u$:
The first two terms on the right of vanish, and the third term obeys the estimate $$\begin{aligned}
{\left\lVertK(T, u) A^K \big( \psi (u, \cdot) \big) (v) \right\rVert}_E
&= \left\| K(T, u) \int_u^T \tilde{\psi} (u, r) K(\mathrm{d}r, v) \right\|_E \quad (\psi(u,r) = 0 \; \mathrm{when} \; v < r < u) \\
&\leq C {\left\lvertK(T, u)\right\rvert} \left( \frac{1}{(u - v)^{\alpha}} - \frac{1}{(T - v)^{\alpha}} \right),\end{aligned}$$ and hence it is in $L^2([0, T]^2; E)$. For the fourth term, note that $$\begin{aligned}
\psi \begin{pmatrix}
u & r_1 \\
v & r_2
\end{pmatrix} = 0 \; \mathrm{when} \; v < r_2 < u,\end{aligned}$$ and thus we have $$\begin{aligned}
{\left\lVertB^K(\psi) (u, v)\right\rVert}_E
&\leq {\left\lVert\int_u^T \left( \int_u^{r_2} \left[ \tilde{\psi}(r_1, r_2) - \tilde{\psi} (u, r_2) \right] K(\mathrm{d}r_1, u) \right) K(\mathrm{d}r_2, v) \right\rVert}_E \\
&\qquad+ {\left\lVert\int_u^T \left( \int_{r_2}^T \tilde{\psi} (u, r_2) K(\mathrm{d}r_1, u) \right) K(\mathrm{d}r_2, v)\right\rVert}_E \\
&\leq C \left( \left( \frac{1}{(u - v)^{\alpha}} - \frac{1}{(T - v)^{\alpha}} \right) + \int_u^T \frac{1}{(r_2 - u)^{\alpha} (r_2 - v)^{\alpha + 1}} {\, \mathrm{d} r_2} \right).\end{aligned}$$ Utilizing again, we see that the fourth term is also in $L^2([0, T]^2; E)$.
The following proposition then deals with the issue of convergence along discrete approximations.
\[HSprop\] Let $F$ denote either $\mathbb{R}^{e}$ or $L^{2}(\Omega; \mathbb{R}^{e})$, and let $\psi: [0, T]^{2} \rightarrow F$ be a function of the form $\psi(u, v) = \mathds{1}_{[0, v)}(u) \tilde{\psi}(u,v)$, where $\tilde{\psi}$ is strongly $\lambda$-Hölder bi-continuous in the norm of $F$. Given a partition $\pi= \{ r_{i} \}$ of $[0, T]$, denote $$\begin{aligned}
\label{psiApprox}
\psi^{\pi} (s, t) := \sum_{j} \psi(s, r_{j}) \mathds{1}_{[r_{j}, r_{j+1})}(t).\end{aligned}$$ Moreover, let $\mathcal{K}^{*} \otimes\mathcal{K}^{*}$ be the operator given in Definition \[kStarTensorOp\], where the Volterra kernel $K$ satisfies Condition \[amnCond\] for some $\alpha\in\left[ 0, \frac{1}{4} \right) $. Then if $\lambda> \alpha$, we have $$\begin{aligned}
\int_{[0, T]^{2}} \left\| \mathcal{K}^{*} \otimes\mathcal{K}^{*} \left(\psi^{\pi} - \psi\right) (s, t) \right\| ^{2}_{F} \, \mathrm{d} s \, \mathrm{d} t \rightarrow 0.\end{aligned}$$
We define $$\begin{aligned}
h(u, v)
:= \int_0^T \left\langle \mathcal{K}^* \left( \psi (\cdot, u) \right) (s), \, \mathcal{K}^* \left( \psi (\cdot, v) \right) (s) \right\rangle_F {\, \mathrm{d} s},\end{aligned}$$ and correspondingly, $$\begin{aligned}
h^{\pi} (u, v)
&:= \int_0^T \left\langle \mathcal{K}^* \left( \psi^{\pi} (\cdot, u) \right) (s), \, \mathcal{K}^* \left( \psi^{\pi} (\cdot, v) \right) (s) \right\rangle_F {\, \mathrm{d} s} \\
&= \sum_{i, j} \left( \int_0^T \left\langle \mathcal{K}^* \left( \psi^{\pi} (\cdot, r_i) \right) (s), \, \mathcal{K}^* \left( \psi^{\pi} (\cdot, r_j) \right) (s) \right\rangle_F {\, \mathrm{d} s} \right) \mathds{1}_{[r_i, r_{i+1})} (u) \mathds{1}_{[r_j, r_{j+1})} (v) \\
&= \sum_{i, j} h(r_i, r_j) \mathds{1}_{[r_i, r_{i+1})} (u) \mathds{1}_{[r_j, r_{j+1})} (v).\end{aligned}$$ Let $\lambda' := \frac{1}{4} \wedge \lambda$. Since $\alpha < \frac{1}{4}$, $\lambda'$ is greater than $\alpha$, and note that any strongly $\lambda$-Hölder bi-continuous function is also strongly $\lambda'$-Hölder bi-continuous. We will begin by first showing that $h(u, v)$ is strongly $\lambda'$-Hölder bi-continuous.
For all $u, v, u_1, u_2, v_1, v_2 \in [0,T]$, we have $$\begin{aligned}
&{\left\lverth(u_1, v) - h(u_2, v)\right\rvert}
\leq \left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, u_1) - \psi(\cdot, u_2) \right) (s) \right\|^2_F {\, \mathrm{d} s} \right)^{\frac{1}{2}} \left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, v) \right) (s) \right\|^2_F {\, \mathrm{d} s} \right)^{\frac{1}{2}}, \\
&{\left\lverth(u, v_1) - h(u, v_2)\right\rvert}
\leq \left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, v_1) - \psi(\cdot, v_2) \right) (s) \right\|^2_F {\, \mathrm{d} s} \right)^{\frac{1}{2}} \left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, u) \right) (s) \right\|^2_F {\, \mathrm{d} s} \right)^{\frac{1}{2}},\end{aligned}$$ and ${\left\lverth \begin{pmatrix}
u_1 & u_2 \\
v_1 & v_2
\end{pmatrix}\right\rvert}$ is bounded above by $$\begin{aligned}
\left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, u_1) - \psi(\cdot, u_2) \right) (s) \right\|^2_F {\, \mathrm{d} s} \right)^{\frac{1}{2}} \left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, v_1) - \psi(\cdot, v_2) \right) (s) \right\|^2_F {\, \mathrm{d} s} \right)^{\frac{1}{2}}.\end{aligned}$$ Note that for $p \geq 1$, using and fixing $w \in [0, T]$, we have $$\begin{aligned}
\label{estimateE}
\begin{split}
\left\| \mathcal{K}^* \psi (\cdot, w) (s)\right\|_F^p
&= {\left\lVert\tilde{\psi} (s, w) K(w, s) + \int_s^w \left[ \tilde{\psi}(r, w) - \tilde{\psi} (s, w) \right] K(\mathrm{d}r, s) \right\rVert}_F^p \\
&\leq C \, 2^{p-1} \left( \frac{1}{s^{p\alpha} (w - s)^{p\alpha}} + (w - s)^{p (\lambda' - \alpha)} \right).
\end{split}\end{aligned}$$ Since $\alpha < \frac{1}{4}$, $\int_0^T \left\| \mathcal{K}^* \psi (\cdot, w) (s)\right\|_F^p {\, \mathrm{d} s}$ is finite as long as $p \leq 4$.
Now, all we have to do is show that $$\begin{aligned}
\label{estimateA}
\int_0^T \left\| \mathcal{K}^* \psi(\cdot, w_2) - \psi(\cdot, w_1) (s) \right\|_F^2 {\, \mathrm{d} s} \leq C \, {\left\lvertw_2 - w_1\right\rvert}^{2\lambda'},\end{aligned}$$ for all $w_1, w_2 \in [0, T]$, where without loss of generality, we let $w_1 < w_2$. Observe that $$\begin{aligned}
\label{origExp}
\begin{split}
&\int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, w_2) - \psi(\cdot, w_1) \right) (s) \right\|^2_F {\, \mathrm{d} s} \\
&\qquad \qquad = \int_0^{w_1} \left\| \mathcal{K}^* \left( \psi(\cdot, w_2) - \psi(\cdot, w_1) \right) (s) \right\|^2_F {\, \mathrm{d} s} + \int_{w_1}^{w_2} \left\| \mathcal{K}^* \left( \psi(\cdot, w_2) - \psi(\cdot, w_1) \right) (s) \right\|^2_F {\, \mathrm{d} s}.
\end{split}\end{aligned}$$ For the first term above, for $s \in [0, w_1)$, we have (using ) $$\begin{aligned}
\label{expansion1}
\begin{split}
\mathcal{K}^* \left( \psi(\cdot, w_2) - \psi(\cdot, w_1) \right) (s)
&= \left(\psi(s, w_2) - \psi (s, w_1) \right) K(w_2, s) \\
&\qquad+ \int_s^{w_2} \left[ \psi (r, w_2) - \psi (s, w_2) - \psi (r, w_1) + \psi (s, w_1) \right] K(\mathrm{d}r, s) \\
&= \left(\tilde{\psi}(s, w_2) - \tilde{\psi} (s, w_1) \right) K(w_2, s) + \int_s^{w_1}
\tilde{\psi} \begin{pmatrix}
s & r \\
w_1 & w_2
\end{pmatrix} K(\mathrm{d}r, s) \\
&\qquad+ \int_{w_1}^{w_2} \left[ \tilde{\psi} (r, w_2) - \tilde{\psi} (s, w_2) + \tilde{\psi} (s, w_1) \right] K(\mathrm{d}r, s).
\end{split}\end{aligned}$$ Since $\tilde{\psi}$ is strongly $\lambda'$-Hölder bi-continuous, we have $$\begin{aligned}
\label{est1}
{\left\lVert\left( \tilde{\psi}(s, w_2) - \tilde{\psi} (s, w_1) \right) K(w_2, s)\right\rVert}_F
\leq C {\left\lvertw_2 - w_1\right\rvert}^{\lambda'} s^{-\alpha} (w_2 - s)^{-\alpha},\end{aligned}$$ and $$\begin{aligned}
\label{est2}
{\left\lVert\int_s^{w_1}
\tilde{\psi} \begin{pmatrix}
s & r \\
w_1 & w_2
\end{pmatrix} K(\mathrm{d}r, s)\right\rVert}_F
\leq C {\left\lvertw_2 - w_1\right\rvert}^{\lambda'} (w_1 - s)^{\lambda' - \alpha}.\end{aligned}$$ For the last integral in , we let $q_1$ denote $\frac{1}{1 - \lambda'}$ and use Hölder’s inequality to derive $$\begin{aligned}
\label{est3}
\begin{split}
{\left\lVert\int_{w_1}^{w_2} \left[ \tilde{\psi} (r, w_2) - \tilde{\psi} (s, w_2) + \tilde{\psi} (s, w_1) \right] K(\mathrm{d}r, s) \right\rVert}_F
&\leq C {\left\lvertw_2 - w_1\right\rvert}^{\lambda'} \left( \int_{w_1}^{w_2} {\left\lvert{\frac{\partial K(r, s)}{\partial r}}\right\rvert}^{q_1} {\, \mathrm{d} r} \right)^{\frac{1}{q_1}} \\
&\leq C {\left\lvertw_2 - w_1\right\rvert}^{\lambda'} \left( \int_{w_1}^{w_2} \frac{1}{(r - s)^{q_1 (\alpha + 1)}} {\, \mathrm{d} r} \right)^{\frac{1}{q_1}} \\
&\leq C {\left\lvertw_2 - w_1\right\rvert}^{\lambda'} (w_1 - s)^{-(\alpha + \lambda' )}.
\end{split}\end{aligned}$$ Putting estimates , and together, when $s < w_1$ we have $$\begin{aligned}
\label{diffEst}
{\left\lVert\mathcal{K}^* \left( \psi(\cdot, w_2) - \psi(\cdot, w_1) \right) (s) \right\rVert}_F
\leq C {\left\lvertw_2 - w_1\right\rvert}^{\lambda'} f(s),\end{aligned}$$ for some $f(s) \in L^2\left([0, T]\right)$ since $\lambda' > \alpha$ and $2(\alpha + \lambda' ) < 1$. This gives $$\begin{aligned}
\int_0^{w_1} {\left\lVert\mathcal{K}^* \left( \psi(\cdot, w_2) - \psi(\cdot, w_1) \right) (s) \right\rVert}_F^2 {\, \mathrm{d} s}
\leq C {\left\lvertw_2 - w_1\right\rvert}^{2 \lambda'}.\end{aligned}$$ Returning to the second term in , we let $q_2$ denote $\frac{1}{1 - 2\lambda'}$ and use Hölder’s inequality again to obtain $$\begin{aligned}
\int_{w_1}^{w_2} \left\| \mathcal{K}^* \left( \psi(\cdot, w_1) - \psi(\cdot, w_2) \right) (s) \right\|^2_F {\, \mathrm{d} s}
\leq |w_2 - w_1|^{2\lambda'} \left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, w_1) - \psi(\cdot, w_2) \right) (s) \right\|^{2q_2}_F {\, \mathrm{d} s} \right)^{\frac{1}{q_2}}.\end{aligned}$$ Since $\lambda' < \frac{1}{4}$, we have $2q_2 \leq 4$ and this gives $\left( \int_0^T \left\| \mathcal{K}^* \left( \psi(\cdot, w_1) - \psi(\cdot, w_2) \right) (s) \right\|^{2q_2}_F {\, \mathrm{d} s} \right)^{\frac{1}{q_2}} < \infty$ from . Now that we have shown that $h$ is strongly $\lambda'$-Hölder bi-continuous, we will show that $$\begin{aligned}
\int_{[0, T]^2} \left\| \mathcal{K}^* \otimes \mathcal{K}^* \left( \psi^{\pi} - \psi \right) (s, t) \right\|^2_F {\, \mathrm{d} s} {\, \mathrm{d} t}
= \int_0^T \left( \mathcal{K}^* \otimes \mathcal{K}^* \left( h^{\pi} - h \right) \right) (t, t) {\, \mathrm{d} t},\end{aligned}$$ and then invoke Proposition \[nualartPropNew\] to complete the proof.
Let $g(s, t)$ denote $\mathcal{K}^* \left(\psi (\cdot, t) \right) (s)$, and note that $g(s, t) = 0$ when $s \geq t$. We first compute $$\begin{aligned}
\label{pullOutS}
\begin{split}
\mathcal{K}^* \otimes \mathcal{K}^* h (t, t)
&= h(t, t) K(T, t)^2 + K(T, t) A^K\big( h(\cdot, t) \big) (t) + K(T, t) A^K \big( h(t, \cdot) \big) (t) + B^K(h)(t, t) \\
&= \int_0^T \left\langle g(s, t), g(s, t) \right\rangle_F K(T, t)^2 {\, \mathrm{d} s} \\
&\quad+ 2 K(T, t) \int_t^T \left( \int_0^T \left\langle g(s, r) - g(s, t), g(s, t) \right\rangle_F {\, \mathrm{d} s} \right) K(\mathrm{d}r, t) \\
&\quad+ \int_t^T \int_t^T \left( \int_0^T \left\langle g(s, r_1) - g(s, t), g(s, r_2) - g(s, t) \right\rangle_F {\, \mathrm{d} s} \right) K(\mathrm{d}r_1, t) K(\mathrm{d}r_2, t).
\end{split}\end{aligned}$$ The second term on the right vanishes when $s \geq t$, and when $s < t$, using and gives us $$\begin{aligned}
{\left\lvert\left\langle g(s, r) - g(s, t), g(s, t) \right\rangle_F\right\rvert} {\left\lvert{\frac{\partial K(r, t)}{\partial r}}\right\rvert}
\leq C {\left\lvertr - t\right\rvert}^{\lambda' - \alpha - 1} \tilde{f}(s)\end{aligned}$$ for some $\tilde{f}(s) \in L^1 ([0, T])$, and thus we can swap the integral with respect to $s$ outside the integral with respect to $r$. Similarly, the third term on the right of is bounded by $$\begin{aligned}
C \left( \int_s^T \frac{1}{(r - t)^{\alpha + 1}} {\, \mathrm{d} r} \right)^2\end{aligned}$$ when $s > t$ since the integrand vanishes when $r_1 \leq s$ or $r_2 \leq s$. Furthermore, when $s < t$, its integrand is bounded by $$\begin{aligned}
C {\left\lvertr_1 - t\right\rvert}^{\lambda' - \alpha - 1} {\left\lvertr_2 - t\right\rvert}^{\lambda' - \alpha - 1} f^2(s).\end{aligned}$$ Hence, we can also pull out the integral with respect to $s$, and we get $$\begin{aligned}
\mathcal{K}^* \otimes \mathcal{K}^* h (t, t)
= \int_0^T \mathcal{K}^* \otimes \mathcal{K}^* \left( \left\langle g(s, \cdot), g(s, \cdot) \right\rangle_F \right) (t, t) {\, \mathrm{d} s}.\end{aligned}$$ Observe that $$\begin{aligned}
\mathcal{K}^* \otimes \mathcal{K}^* \left( \left\langle g(s, \cdot), g(s, \cdot) \right\rangle_F \right) (t,t)
&= \left\langle \mathcal{K}^* \left( g(s, \cdot) \right) (t), \mathcal{K}^* \left( g(s, \cdot) \right) (t) \right\rangle_F \\
&= {\left\lVert\mathcal{K}^* (g(s, \cdot))(t)\right\rVert}^2_F,\end{aligned}$$ where here we use , and Fubini’s theorem in the case when $F = L^2 (\Omega; \mathbb{R}^e)$.
Fixing $s$, note that for all $t > s$, $g(s, \cdot)$ is $\lambda'$-Hölder continuous on $[t, T]$ (with the Hölder norm depending on $t$) from , and . Thus, $\mathcal{K}^* (g(s, \cdot))(t)$ is well defined for all $t > s$, vanishes when $t < s$, and we can apply Lemma 3.2 in [@lim2018] to obtain $$\begin{aligned}
\mathcal{K}^* \left( g(s, \cdot) \right) (t) = \mathcal{K}^* \otimes \mathcal{K}^* \psi (s, t)\end{aligned}$$ for all $s \neq t$. This concludes the proof.
It follows in particular that $\mathcal{H}_{1}^{d}\otimes\mathcal{H}_{1}^{d}$ contains functions $\psi: \left[0, T \right]^{2} \rightarrow \mathbb{R}^{d} \otimes \mathbb{R}^{d}$ of the form $\psi(u,v) = \mathds{1}_{[0,v)}(u)\tilde{\psi}(u,v)$ whenever $\tilde{\psi}$ is strongly $\lambda$-Hölder bi-continuous.
\[H1tensorH1equiv2\] Let $\psi: [0, T]^2 \rightarrow \mathbb{R}^d$ be of the form $\psi(u, v) = \mathds{1}_{[0, v)}(u) \tilde{\psi}(u,v)$, where $\tilde{\psi}$ is strongly $\lambda$-Hölder bi-continuous, and let $\mathcal{K}^* \otimes \mathcal{K}^*$ be defined as in Definition \[kStarTensorOp\], where the Volterra kernel $K$ satisfies Condition \[amnCond\] for some $\alpha \in \left[ 0, \frac{1}{4} \right)$.
Then if $\lambda > \alpha$, $\psi$ is an element of $\mathcal{H}_1^d \otimes \mathcal{H}_1^d$, with norm given by $$\begin{aligned}
\label{tensorNorm2}
{\left\lVert\psi\right\rVert}_{\mathcal{H}_1^d \otimes \mathcal{H}_1^d}
= \int_{[0, T]^2} {\left\lvert \mathcal{K}^* \otimes \mathcal{K}^* \psi (s, t) \right\rvert}^2_{\mathbb{R}^d \otimes \mathbb{R}^d} {\, \mathrm{d} s} {\, \mathrm{d} t},\end{aligned}$$ and with $\psi^{\pi}$ defined as in , we have $$\begin{aligned}
\label{tensorVanishing2}
{\left\lVert\psi^{\pi} - \psi\right\rVert}_{\mathcal{H}_1^d \otimes \mathcal{H}_1^d} \rightarrow 0\end{aligned}$$ as ${\left\lVert\pi\right\rVert} \rightarrow 0$.
Using the canonical identification $$\begin{aligned}
\label{canonID}
A(s) \mathds{1}_{[a, b)} (t) \simeq \sum_{j=1}^d \sum_{k=1}^d a^{(k)}_j (s) e_k \otimes \mathds{1}^{(j)}_{[a, b)} (t), \quad a, b \in [0, T],\end{aligned}$$ it is clear that $\psi^{\pi}$ is a member of $\Lambda^d_{\alpha} \otimes \mathcal{H}_1^d$, and thus lies in $\mathcal{H}_1^d \otimes \mathcal{H}_1^d$ by Proposition \[A1eqH1\]. Furthermore, ${\left\lVert\psi^{\pi}\right\rVert}^2_{\mathcal{H}_1^d \otimes \mathcal{H}_1^d}$ is equal to $$\begin{aligned}
&\sum_{k, l} \int_0^T \sum_{j=1}^d \left\langle \mathcal{K}^* \left( \psi_j (\cdot, r_k) \right) (s), \mathcal{K}^* \left( \psi_j(\cdot, r_l) \right) (s) \right\rangle_{\mathbb{R}^d} {\, \mathrm{d} s} \int_0^T \mathcal{K}^* \left( \mathds{1}_{\left[ r_k, r_{k+1} \right)} \right) (t) \mathcal{K}^* \left( \mathds{1}_{\left[ r_l, r_{l+1} \right)} \right) (t) {\, \mathrm{d} t} \\
&\qquad = \sum_{k,l} \int_{[0, T]^2} \left\langle \mathcal{K}^* \otimes \mathcal{K}^* \left( \psi (\cdot, r_k) \mathds{1}_{\Delta_k } (\cdot) \right) (s, t), \mathcal{K}^* \otimes \mathcal{K}^* \left( \psi (\cdot, r_l) \mathds{1}_{\Delta_l } (\cdot) \right) (s, t) \right\rangle_{\mathbb{R}^d \otimes \mathbb{R}^d} {\, \mathrm{d} s} {\, \mathrm{d} t}, \\
&\qquad = \int_{[0, T]^2} {\left\lvert\mathcal{K}^* \otimes \mathcal{K}^* \psi^{\pi} (s, t)\right\rvert}_{\mathbb{R}^d \otimes \mathbb{R}^d}^2 {\, \mathrm{d} s} {\, \mathrm{d} t},\end{aligned}$$ which we know is Cauchy as ${\left\lVert\pi\right\rVert} \rightarrow 0$ by Proposition \[HSprop\]. We now take any sequence of partitions $\pi(n)$ with vanishing mesh and identify $\psi$ with the limit of $\psi^{\pi(n)}$ in $\mathcal{H}_1^d \otimes \mathcal{H}_1^d$. Invoking Proposition \[HSprop\] again then gives us and .
The Itô-Skorohod isometry revisited {#itoSkorohodIsometry}
-----------------------------------
We now give another formulation for the Itô-Skorohod isometry for Volterra processes (see also [@ev2003], where an isometry formula in the specific case of fractional Brownian motion is provided).
\[isometry2\] Let $X$ be a Volterra process which satisfies Condition \[newCond1\] for some $\rho \in \left[ 1, 2 \right)$, and assume that its kernel satisfies Condition \[amnCond\] for $\alpha = \frac{1}{2} - \frac{1}{2\rho}$. Given $\lambda > \alpha$, let $Y$ be a random variable which satisfies, almost surely,
(i) $Y \in \mathcal{C}^{\lambda-H\ddot{o}l}_{pw} \left([0, T]; \mathbb{R}^d\right)$,
(ii) $\mathcal{D}Y: [0, T]^2 \rightarrow \mathbb{R}^d \otimes \mathbb{R}^d$ is a function of the form $\mathds{1}_{[0, t)}(s) g(s, t)$, where $g$ is strongly $\lambda$-Hölder bi-continuous.
Then $\lim_{{\left\lVert\pi\right\rVert} \rightarrow 0} Y^{\pi} = Y$ in $\mathbb{D}^{1, 2} (\mathcal{H}_1^d)$ if and only if $$\begin{aligned}
\lim_{{\left\lVert\pi\right\rVert} \rightarrow 0} {\mathbb{E} \left[ \int_0^T {\left\lvert\mathcal{K}^* \left( Y^{\pi} - Y \right) (t) \right\rvert}_{\mathbb{R}^d}^2 {\, \mathrm{d} t}\right]} = 0,\end{aligned}$$ and $$\begin{aligned}
\lim_{{\left\lVert\pi\right\rVert} \rightarrow 0} {\mathbb{E} \left[ \int_{[0, T]^2} {\left\lvert\mathcal{K}^* \otimes \mathcal{K}^* \left( \mathcal{D} Y^{\pi} - \mathcal{D} Y \right) (s, t) \right\rvert}_{\mathbb{R}^d \otimes \mathbb{R}^d}^2 {\, \mathrm{d} s} {\, \mathrm{d} t}\right]} = 0,\end{aligned}$$ in which case $\lim_{{\left\lVert\pi\right\rVert} \rightarrow 0} {\mathbb{E} \left[ \delta^X\left(Y^{\pi} - Y \right)^2\right]} = 0$ and ${\mathbb{E} \left[ \delta^X\left( Y \right)^2\right]} $ is equal to $$\begin{aligned}
{\mathbb{E} \left[ \int_0^T {\left\lvert \mathcal{K}^* Y (t) \right\rvert}_{\mathbb{R}^d}^2 {\, \mathrm{d} t}\right]} + {\mathbb{E} \left[ \int_{[0, T]^2} \mathrm{tr} \left( \mathcal{K}^* \otimes \mathcal{K}^* \mathcal{D} Y (s, t) \, \mathcal{K}^* \otimes \mathcal{K}^* \mathcal{D} Y (t, s) \right) {\, \mathrm{d} s} {\, \mathrm{d} t}\right]}.\end{aligned}$$
From the computation of the trace term in Theorem 4.5 of [@cl2018], we know that ${\mathbb{E} \left[ \delta^X\left( Y \right)^2\right]}$ is equal to $$\begin{aligned}
\lim_{{\left\lVert\pi\right\rVert} \rightarrow 0} {\mathbb{E} \left[ \int_0^T {\left\lvert \mathcal{K}^* Y^{\pi} (t) \right\rvert}_{\mathbb{R}^d}^2 {\, \mathrm{d} t}\right]}
+ \lim_{{\left\lVert\pi\right\rVert} \rightarrow 0} {\mathbb{E} \left[ \sum_{i, j} \sum_{k,l=1}^d \left\langle \mathcal{D}_{\cdot}^{(k)} Y^{(l)}_{t_j}, \mathds{1}_{\Delta_i} (\cdot) \right\rangle_{\mathcal{H}_1} \left\langle \mathcal{D}^{(l)}_{\cdot} Y^{(k)}_{t_i}, \mathds{1}_{\Delta_j} (\cdot) \right\rangle_{\mathcal{H}_1} \right]}.\end{aligned}$$ The first term converges to ${\mathbb{E} \left[ \int_0^T {\left\lvert\mathcal{K}^* Y (t) \right\rvert}_{\mathbb{R}^d}^2 {\, \mathrm{d} t}\right]} $ and for the second term we have $$\begin{aligned}
&{\mathbb{E} \left[ \sum_{i, j} \sum_{k,l=1}^d \left\langle \mathcal{D}_{\cdot}^{(k)} Y^{(l)}_{t_j}, \mathds{1}_{\Delta_i} (\cdot) \right\rangle_{\mathcal{H}_1} \left\langle \mathcal{D}^{(l)}_{\cdot} Y^{(k)}_{t_i}, \mathds{1}_{\Delta_j} (\cdot) \right\rangle_{\mathcal{H}_1} \right]} \\
&\qquad \qquad= {\mathbb{E} \left[ \sum_{i, j} \sum_{k,l=1}^d \int_0^T \mathcal{K}^* \left( \mathcal{D}_{\cdot}^{(k)} Y^{(l)}_{t_j} \right) (s) K(\Delta_i, s) {\, \mathrm{d} s} \int_0^T \mathcal{K}^* \left( \mathcal{D}_{\cdot}^{(l)} Y^{(k)}_{t_i} \right) (t) K(\Delta_j, t) {\, \mathrm{d} t} \right]}.\end{aligned}$$ Using Lemma 3.2 in [@lim2018], this expression is equal to $$\begin{aligned}
{\mathbb{E} \left[ \int_{[0, T]^2} \mathrm{tr} \left( \mathcal{K}^* \otimes \mathcal{K}^* \mathcal{D} Y^{\pi} (s, t) \, \mathcal{K}^* \otimes \mathcal{K}^* \mathcal{D} Y^{\pi} (t, s) \right) {\, \mathrm{d} s} {\, \mathrm{d} t}\right]},\end{aligned}$$ which converges as ${\left\lVert\pi\right\rVert} \rightarrow 0$ to $$\begin{aligned}
{\mathbb{E} \left[ \int_{[0, T]^2} \mathrm{tr} \left( \mathcal{K}^* \otimes \mathcal{K}^* \mathcal{D} Y (s, t) \, \mathcal{K}^* \otimes \mathcal{K}^* \mathcal{D} Y (t, s) \right) {\, \mathrm{d} s} {\, \mathrm{d} t}\right]}.\end{aligned}$$
In the case of Brownian motion both $\mathcal{K}^{\ast}$ and $\mathcal{K}^* \otimes\mathcal{K}^{\ast}$ are identity operators and Theorem \[isometry2\] recovers the usual Itô-Skorohod isometry: $$\begin{aligned}
\mathbb{E}\left[ \delta^{X}(Y)^{2}\right] =\mathbb{E}\left[ \int_{0}^{T}\left\vert Y_{t}\right\vert ^{2}\,\mathrm{d}t\right] +\mathbb{E}\left[\int_{[0,T]^{2}}\mathrm{tr}\left( \mathcal{D}_{t}Y_{s}\,\mathcal{D}_{s}
Y_{t}\right) \,\mathrm{d}s\,\mathrm{d}t\right].\end{aligned}$$
Approximation of the Skorohod integral {#Sko conv}
======================================
We will now put together the results of the previous section to show that the Skorohod integral of the discrete approximations to the solution of an RDE converge. Before we proceed, we will introduce additional notation.
Let $Y \in \mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^d; \mathbb{R}^m) \right)$ denote the path-level solution to $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ where $V \in \mathcal{C}^{\lfloor p \rfloor + 1}_b \left( \mathbb{R}^{dm}; \mathbb{R}^{dm} \otimes \mathbb{R}^d \right)$.
Recall that $$\begin{aligned}
\mathcal{D}_t Y_s = \mathds{1}_{[0, t)} (s) J^{\mathbf{X}}_{t \leftarrow s} V(Y_s), \quad s, t \in [0, T],\end{aligned}$$ where here and henceforth, we will use the shorthand $$\begin{aligned}
J^{\mathbf{X}}_{t \leftarrow s} := J^{\mathbf{X}}_t \left(J^{\mathbf{X}}_s\right)^{-1}, \quad 0 \leq s < t \leq T.\end{aligned}$$ Given a Hilbert space $H$, we will denote an element of $y$ of $\mathbb{R}^m \otimes H$ as $$\begin{aligned}
\label{bracketNotation}
y = \sum_{j=1}^m e_j \otimes [y]_j,\end{aligned}$$ where $[y]_j \in H$ for $j = 1, \ldots, m$. (Note that there may be several ways to perform the decomposition.)
Now fix $0 \leq s < t \leq T$. Since $V(Y_s) \in \mathbb{R}^{md} \otimes \mathbb{R}^d \simeq \mathbb{R}^m \otimes \mathbb{R}^d \otimes \mathbb{R}^d$, we will decompose $V(Y_s)$ as $$\begin{aligned}
V(Y_s) = \sum_{j=1}^m e_j \otimes \left[ V(Y_s) \right]_j,\end{aligned}$$ where $$\begin{aligned}
[V(Y_s)]_j := \sum_{i, k=1}^d V^{(d(j-1) + i)}_k (Y_s) \, e_i \otimes e_k.\end{aligned}$$ If we canonically identify $\mathbb{R}^{md} \otimes \mathbb{R}^d$ with the space of $md$-by-$d$ matrices, then $\left[ V(Y_s) \right]_j$ simply denotes the $d$-by-$d$ sub-matrix of $V(Y_s)$ which starts at the $(d(j - 1) + 1)^{th}$ row and ends at the $dj^{th}$ row. Contrast this with $V_j (Y_s)$, which denotes the $j^{th}$ column of $V(Y_s)$.
We will do the same with $Y_s \in \mathbb{R}^{md} \simeq \mathbb{R}^m \otimes \mathbb{R}^d$, and write $$\begin{aligned}
Y_s = \sum_{j=1}^m e_j \otimes \left[ Y_s \right]_j, \quad [Y_s]_j := \sum_{i=1}^d Y_s^{(d(j-1) + i)} e_i \in \mathbb{R}^d,\end{aligned}$$ and for $\displaystyle J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) = \sum_{i=1}^{md} \sum_{k=1}^d a_{i, k} \, e_i \otimes e_k \in \mathbb{R}^{md} \otimes \mathbb{R}^d \simeq \mathbb{R}^m \otimes \mathbb{R}^d \otimes \mathbb{R}^d$, we have $$\begin{aligned}
&J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) = \sum_{j=1}^m e_j \otimes \left[ J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) \right]_j, \\
&\left[ J^{\mathbf{X}}_{t \leftarrow s} V(Y_s)\right]_j := \sum_{i, k = 1}^d a_{d(j-1) + i, k} \, e_i \otimes e_k \in \mathbb{R}^d \otimes \mathbb{R}^d.\end{aligned}$$
\[skorohodLimit3rdLevel\] Let $\mathbf{X} \in \mathcal{C}^{0, p-var} \left([0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^d\right) \right)$, $1 \leq p < 4$, be a Volterra rough path which satisfies Condition \[newCond1\], and assume that its kernel satisfies Condition \[amnCond\] with $\alpha < \frac{1}{p}$. Let $Y \in \mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^d; \mathbb{R}^m) \right)$ denote the path-level solution to $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ where $V \in \mathcal{C}^{\lfloor p \rfloor + 1}_b \left( \mathbb{R}^{dm}; \mathbb{R}^{dm} \otimes \mathbb{R}^d \right)$. Then $Y \in \mathbb{D}^{1,2} (\mathbb{R}^m \otimes \mathcal{H}_1^d)$ and $$\begin{aligned}
\int_0^T Y_r \, \mathrm{d} X_r
&= \lim_{{\left\lVert\pi = \{ r_i \} \right\rVert} \rightarrow 0} \sum_i \left[ Y_{r_i} \left( X_{r_i, r_{i+1}} \right) - \sum_{j=1}^m \left( \int_0^{r_i} \mathrm{tr} \, \left[ J^{\mathbf{X}}_{r_i \leftarrow s} V(Y_s) \right]_j \, R(\Delta_i, {\, \mathrm{d} s}) \right) e_j \right],\end{aligned}$$ where the limit is taken in $L^2\left(\Omega\right)$.
We first use integration-by-parts to obtain $$\begin{aligned}
\left\langle \delta^X(Y^{\pi}), \, e_j \right\rangle_{\mathbb{R}^m}
&= \sum_i \left[ \left\langle \left[ Y_{r_i} \right]_j, \, X_{r_i, r_{i+1}} \right\rangle_{\mathbb{R}^d} - \int_0^{r_i} \mathrm{tr} \, \left[ J^{\mathbf{X}}_{r_i \leftarrow s} V(Y_s) \right]_j \, R(\Delta_i, {\, \mathrm{d} s}) \right],\end{aligned}$$ for all $j = 1, \ldots, m$. Next, we invoke Theorem \[isometry2\], which requires us to prove that $$\begin{aligned}
\label{conv1}
{\mathbb{E} \left[ \int_0^T {\left\lvert \mathcal{K}^* \left( Y^{\pi} - Y \right) (t)\right\rvert}^2_{\mathbb{R}^{md}} {\, \mathrm{d} t}\right]} \rightarrow 0,\end{aligned}$$ and $$\begin{aligned}
\label{conv2}
{\mathbb{E} \left[ \int_{[0, T]^2} {\left\lvert \mathcal{K}^* \otimes \mathcal{K}^* (\mathcal{D}_s Y^{\pi}_t - \mathcal{D}_s Y_t) (s, t)\right\rvert}_{\mathbb{R}^{md} \otimes \mathbb{R}^d}^2 {\, \mathrm{d} s} {\, \mathrm{d} t} \right]} \rightarrow 0.\end{aligned}$$ We will show that $Y$ is $\frac{1}{p}$-Hölder continuous in $L^2 \left( \Omega; \mathbb{R}^{md} \right)$, and then invoke Proposition \[nualartProp\] to obtain . We have $$\begin{aligned}
\label{YHolder}
\begin{split}
{\left\lvertY_{s,t}\right\rvert}
&\leq C \left( {\left\lVert\mathbf{X}\right\rVert}_{p-var; [s, t]} \vee {\left\lVert\mathbf{X}\right\rVert}^p_{p-var; [s, t]} \right) \\
&\leq C {\left\lVert\mathbf{X}\right\rVert}_{\frac{1}{p}-H\ddot{o}l; [s, t]} \left( (t - s) \vee (t - s)^{\frac{1}{p}} \right) \\
&\leq C {\left\lVert\mathbf{X}\right\rVert}_{\frac{1}{p}-H\ddot{o}l; [0, T]} \left( T^{1 - \frac{1}{p}} \vee 1 \right) (t - s)^{\frac{1}{p}}
\end{split}\end{aligned}$$ almost surely, and thus $$\begin{aligned}
\sqrt{{\mathbb{E} \left[ {\left\lvertY_{s,t}\right\rvert}\right]}} \leq C_1 {\left\lvertt - s\right\rvert}^{\frac{1}{p}}\end{aligned}$$ since ${\left\lVert\mathbf{X}\right\rVert}_{\frac{1}{p}-H\ddot{o}l; [0, T]}$ has moments of all orders.
To show , we will apply Proposition \[HSprop\] with $\psi(s, t) = \mathcal{D}_s Y_t = \mathds{1}_{[0, t)} (s) J_t^{\mathbf{X}} \left( J_s^{\mathbf{X}} \right)^{-1} V(Y_s)$. To do so, we have to show that $\tilde{\psi}(s, t) := J_t^{\mathbf{X}} \left( J_s^{\mathbf{X}} \right)^{-1} V(Y_s)$ is strongly $\frac{1}{p}$-Hölder bi-continuous in $L^2 \left(\Omega; \mathbb{R}^{md} \otimes \mathbb{R}^d \right)$. By Lemma 3.6 in [@lim2018], this is equivalent to showing that $J_{\cdot}^{\mathbf{X}}$ and $\left(J_{\cdot}^{\mathbf{X}} \right)^{-1} Y_{\cdot}$ are both $\frac{1}{p}$-Hölder continuous.
Using , we have $$\begin{aligned}
{\left\lvert J^{\mathbf{X}}_{s,t}\right\rvert}
&\leq C_1 {\left\lVert\mathbf{X}\right\rVert}_{p-var; [s,t]} \exp \left( C_2 N^{\mathbf{X}}_{1; [s, t]} \right) \\
&\leq C_1 (t - s)^{\frac{1}{p}} {\left\lVert\mathbf{X}\right\rVert}_{\frac{1}{p}-H\ddot{o}l;[0, T]} \exp \left( C_2 N^{\mathbf{X}}_{1; [0, T]} \right),\end{aligned}$$ which yields $\frac{1}{p}$-Hölder continuity for $J_{\cdot}^{\mathbf{X}}$ as the expression to the right of $(t-s)^\frac{1}{p}$ is in $L^q (\Omega)$ for all $q > 0$. The same is true for $\left( J^{\mathbf{X}}_{\cdot} \right)^{-1}$ since the inverse obeys the same bound.
Finally, $\left( J^{\mathbf{X}}_{\cdot} \right)^{-1} V(Y_{\cdot})$ is also $\frac{1}{p}$-Hölder continuous, since $V$ is $\mathcal{C}^1$ smooth and both $Y$ and $\left( J^{\mathbf{X}}_{\cdot} \right)^{-1}$ are $\frac{1}{p}$-Hölder continuous.
Augmenting the Skorohod integral with higher-level terms {#Sko aug}
========================================================
The main purpose of this section is to show that the usual Riemann-sum approximation to the Skorohod integral can be augmented with (suitably corrected) second-level and third-level rough path terms which vanish in $L^{2}(\Omega)$ as the mesh of the partition goes to zero.
Before we do so, we will extend the theory of controlled rough paths to the case $3 \leq p < 4$, and give bounds on the higher-directional derivatives of a controlled rough path satisfying an RDE.
Estimates for controlled rough paths of lower regularity {#controlledLR}
--------------------------------------------------------
To construct the rough integral of controlled rough paths for $3 \leq p < 4$, we need the following new definition.
Let $\mathbf{x}=\left( 1, x, \mathbf{x}^{2}, \mathbf{x}^{3} \right)
\in\mathcal{C}^{p-var}\left( [0, T]; G^{3}(\mathbb{R}^{d})\right) $, where $3
\leq p < 4$, and let $q$ be such that $\frac{1}{p} + \frac{1}{q} > 1$. Let $\left( \phi, \phi^{\prime}, \phi^{\prime\prime}\right) $ satisfy $$\begin{aligned}
& \phi\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{U} \right) ,\\
& \phi^{\prime}\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^{d}; \mathcal{U}) \right) , \quad\mathrm{and}\\
& \phi^{\prime\prime}\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^{d} \otimes\mathbb{R}^{d}; \mathcal{U}) \right) .\end{aligned}$$ Then we say that $\left( \phi,\phi^{\prime}, \phi^{\prime\prime}\right) $ (or $\phi$) is controlled by $\mathbf{x}$ if for all $s,t \in[0, T]$ we have $$\begin{aligned}
\label{controlled2}
\begin{split}
& \phi_{s,t} = \phi^{\prime}_{s} x_{s,t} + \phi^{\prime\prime}_{s}
\mathbf{x}^{2}_{s, t} + R_{s,t}^{\phi},\\
& \phi^{\prime}_{s,t} = \phi^{\prime\prime}_{s} x_{s,t} + R^{\phi^{\prime}}_{s,t},
\end{split}\end{aligned}$$ where the remainder terms satisfy $$\begin{aligned}
R^{\phi} \in\mathcal{C}^{q-var} \left( \left[ 0,T\right] ; \mathcal{U} \right) , \quad R^{\phi^{\prime}} \in\mathcal{C}^{\frac{p}{2}-var} \left(\left[ 0,T\right] ; \mathcal{U} \right) .\end{aligned}$$
Thus, $\phi$ is controlled by $\mathbf{x}$ if $\left\| \phi\right\|_{p,q-cvar} < \infty$, where the controlled variation norm is defined as $$\begin{aligned}
\left\| \phi\right\| _{p,q-cvar} := \left\| \phi\right\| _{\mathcal{V}^{p};
[0, T]} + \left\| \phi^{\prime}\right\| _{\mathcal{V}^{p}; [0, T]} + \left\|
\phi^{\prime\prime}\right\| _{\mathcal{V}^{p}; [0, T]} + \left\| R^{\phi
}\right\| _{q-var; \left[ 0,T\right] } + \left\| R^{\phi^{\prime}}\right\|
_{\frac{p}{2}-var; \left[ 0,T\right] }.\end{aligned}$$ Before we continue, note that $3 \leq p < 4$ implies that $\frac{p}{3} < \frac{p}{p-1} \leq\frac{p}{2}$. Since $q$-variation decreases with increasing $q$, we can always, if necessary, increase $q$ such that $$\begin{aligned}
\label{qBounds}\frac{p}{3} \leq q < \frac{p}{p-1} \leq\frac{p}{2}\end{aligned}$$ when we are working with $p$ in the interval $[3, 4)$.
The following theorem and the next two propositions are the lower-regularity analogues of Theorem 2.20, Proposition 2.22 and Proposition 2.21 respectively from [@cl2018].
\[controlledThm2\] Let $\mathbf{x}=\left( 1, x,\mathbf{x}^{2}, \mathbf{x}^{3} \right) \in\mathcal{C}^{p-var}\left( [0, T]; G^{3} (\mathbb{R}^{d})\right) $, where $3 \leq p < 4$, and let $q$ be such that $\frac{1}{p} + \frac{1}{q} > 1 $. Let $\left( \phi, \phi^{\prime}, \phi^{\prime\prime}\right) $ satisfy $$\begin{aligned}
& \phi\in\mathcal{C}^{p-var}\left( [0,T]; \mathcal{L}(\mathbb{R}^{d};\mathbb{R}^{e})\right) ,\\
& \phi^{\prime}\in\mathcal{C}^{p-var} \left( [0,T]; \mathcal{L} (\mathbb{R}^{d}; \mathcal{L}(\mathbb{R}^{d};\mathbb{R}^{e}))\right) ,
\quad\mathrm{and} \\
& \phi^{\prime\prime}\in\mathcal{C}^{p-var} \left( [0,T]; \mathcal{L} \left(
\mathbb{R}^{d} \otimes\mathbb{R}^{d}; \mathcal{L} (\mathbb{R}^{d}; \mathbb{R}^{e}) \right) \right) .\end{aligned}$$ If $\left( \phi, \phi^{\prime}, \phi^{\prime\prime}\right) $ is controlled by $\mathbf{x}$ with remainder terms $R^{\phi}$ and $R^{\phi^{\prime}}$ of bounded $q$-variation and $\frac{p}{2}$-variation respectively, we can define the rough integral $$\begin{aligned}
\label{controlledRPdefn2}
\int_{0}^{t} \phi_{r}\circ\mathrm{d}\mathbf{x}_{r}:=\underset{\left\Vert \pi\right\Vert \rightarrow 0,\pi=\left\{
0=r_{0}<\ldots<r_{n}=t\right\} }{\lim}\sum_{i=0}^{n-1}\left( \phi_{r_{i}} x_{r_{i},r_{i+1}}+\phi_{r_{i}}^{\prime} \mathbf{x}_{r_{i},r_{i+1}}^{2} + \phi^{\prime\prime}_{r_{i}} \mathbf{x}^{3}_{r_{i}, r_{i+1}} \right) ,\end{aligned}$$ where we have made use of the canonical identification $\mathcal{L} (\mathbb{R}^{d}; \mathcal{L}(\mathbb{R}^{d};\mathbb{R}^{e}))\simeq \mathcal{L}(\mathbb{R}^{d}\otimes\mathbb{R}^{d}; \mathbb{R}^{e})$ and $\mathcal{L} \left( \mathbb{R}^{d} \otimes\mathbb{R}^{d}; \mathcal{L} (\mathbb{R}^{d};\mathbb{R}^{e}) \right) \simeq\mathcal{L}(\mathbb{R}^{d} \otimes\mathbb{R}^{d}\otimes\mathbb{R}^{d}; \mathbb{R}^{e})$. Furthermore, if $q \geq\frac{p}{3}$, then denoting $$\begin{aligned}
z_{t}:=\int_{0}^{t}\phi_{r}\circ\mathrm{d}\mathbf{x}_{r},\quad z_{t}^{\prime}
:=\phi_{t}, \quad z^{\prime\prime}_{t} := \phi^{\prime}_{t},\end{aligned}$$ $(z, z^{\prime}, z^{\prime\prime})$ is again controlled by $\mathbf{x}$, and we have $$\begin{aligned}
\label{controlledBound2}
\left\| z\right\| _{p,q} \leq C_{p, q} \left\|
\phi\right\| _{p,q-cvar} \left( 1 + \left\| x\right\| _{p-var; [0, T]} +
\left\| \mathbf{x}^{2}\right\| _{\frac{p}{2}-var;[0, T]} + \left\|
\mathbf{x}^{3}\right\| _{\frac{p}{3}-var; [0, T]} \right) .\end{aligned}$$
Let $s < u < t$ and define $$\begin{aligned}
\Xi_{s,t} := \phi_s x_{s,t} + \phi'_s \mathbf{x}^2_{s,t} + \phi''_s \mathbf{x}^3_{s,t}.\end{aligned}$$ Then we have $$\begin{aligned}
\Xi_{s, u} + \Xi_{u, t} - \Xi_{s, t}
&= \left( \phi_s x_{s,u} + \phi'_s \mathbf{x}^2_{s,u} + \phi''_s \mathbf{x}^3_{s,u} \right)
+ \left( \phi_u x_{u,t} + \phi'_u \mathbf{x}^2_{u,t} + \phi''_u \mathbf{x}^3_{u,t} \right) - \phi_s \left( x_{s,u} + x_{u,t} \right) \\
&\qquad- \phi'_s \left( \mathbf{x}^2_{s,u} + \mathbf{x}^2_{u,t} + x_{s,u} \otimes x_{u,t} \right) - \phi''_s \left( \mathbf{x}^3_{s,u} + \mathbf{x}^3_{u,t} + \mathbf{x}^2_{s,u} \otimes x_{u,t} + x_{s,u} \otimes \mathbf{x}^2_{u,t} \right) \\
&= \phi_{s,u} x_{u,t} + \phi'_{s,u} \mathbf{x}^2_{u,t} + \phi''_{s,u} \mathbf{x}^3_{u,t} - \phi'_s \left( x_{s,u} \otimes x_{u,t} \right)
-\phi''_s \left( \mathbf{x}^2_{s,u} \otimes x_{u,t} + x_{s,u} \otimes \mathbf{x}^2_{u,t} \right) \\
&= \left( \phi'_s x_{s,u} + \phi''_s \mathbf{x}^2_{s,u} + R^{\phi}_{s,u} \right) x_{u,t} + \left(\phi''_s x_{s,u} + R^{\phi'}_{s,u} \right) \mathbf{x}^2_{u,t} + \phi''_{s,u} \mathbf{x}^3_{u,t} - \phi'_s \left( x_{s,u} \otimes x_{u,t} \right) \\
&\qquad- \phi''_s \left( \mathbf{x}^2_{s,u} \otimes x_{u,t} + x_{s,u} \otimes \mathbf{x}^2_{u,t} \right) \\
&= R^{\phi}_{s,u} x_{u,t} + R^{\phi'}_{s,u} \mathbf{x}^2_{u,t} + \phi''_{s,u} \mathbf{x}^3_{u,t}.\end{aligned}$$ Now let $\theta := \min \left( \frac{1}{p} + \frac{1}{q}, \frac{4}{p} \right)$, and let $\omega(s,t)$ denote the function $$\begin{aligned}
{\left\lVertR^{\phi}\right\rVert}^{\frac{1}{\theta}}_{q-var; [s,t]} {\left\lVertx\right\rVert}^{\frac{1}{\theta}}_{p-var; [s,t]} + {\left\lVertR^{\phi'}\right\rVert}^{\frac{1}{\theta}}_{\frac{p}{2}-var; [s,t]} {\left\lVert\mathbf{x}^2\right\rVert}^{\frac{1}{\theta}}_{\frac{p}{2}-var; [s,t]} + {\left\lVert\phi''\right\rVert}^{\frac{1}{\theta}}_{p-var; [s,t]} {\left\lVert\mathbf{x}^3\right\rVert}^{\frac{1}{\theta}}_{\frac{p}{3}-var; [s,t]}.\end{aligned}$$ This is a control as $\theta \leq \frac{4}{p}$ gives $\frac{1}{\theta} \left( \frac{4}{p} \right) \geq 1$ [@fv2010a]. Following Theorem 3.3 in [@cl2018], for any partition $\pi = \{r_i\}$ of $[s, t]$ with $k$ sub-intervals, there necessarily exists some $r_j \in \pi$ such that $$\begin{aligned}
{\left\lvert\Xi_{r_{j-1}, r_j} + \Xi_{r_j, r_{j+1}} - \Xi_{r_{j-1}, r_{j+1}}\right\rvert}
&\leq {\left\lvertR^{\phi}_{r_{j-1}, r_j} x_{r_j, r_{j+1}}\right\rvert} + {\left\lvertR^{\phi'}_{r_{j-1}, r_j} \mathbf{x}^2_{r_j, r_{j+1}}\right\rvert} + {\left\lvert\phi''_{r_{j-1}, r_j} \mathbf{x}^3_{r_j, r_{j+1}}\right\rvert} \\
&\leq 3 \, \omega(r_{j-1}, r_{j+1})^{\theta}
\leq 3 \left( \frac{2}{k-1} \right)^{\theta} \omega(s, t)^{\theta}.\end{aligned}$$ Appropriately extracting points from the partition until $[s, t]$ remains gives us the bound $$\begin{aligned}
\label{controlledOldBound2}
{\left\lvert\int_{\pi} \phi_r {\circ \mathrm{d} \mathbf{x}_r} - \left( \phi_s x_{s,t} + \phi'_s \mathbf{x}^2_{s,t} + \phi''_s \mathbf{x}^3_{s,t} \right)\right\rvert}
< C \, \zeta(\theta) \, \omega(s, t)^{\theta},\end{aligned}$$ where $$\begin{aligned}
\int_{\pi} \phi_r {\circ \mathrm{d} \mathbf{x}_r}
:= \sum_i \phi_{r_i} x_{r_i, r_{i+1}} + \phi'_{r_i} \mathbf{x}^2_{r_i, r_{i+1}} + \phi''_{r_i} \mathbf{x}^3_{r_i, r_{i+1}},\end{aligned}$$ and is proved as in Theorem 3.3 of [@cl2018].
If we define $$\begin{aligned}
\label{remainder2}
R^z_{s,t} := \int_s^t \phi_r {\circ \mathrm{d} x_r} - \left( \phi_s x_{s,t} + \phi'_s \mathbf{x}^2_{s,t} \right),\end{aligned}$$ we get $$\begin{aligned}
z_{s,t} = z'_s x_{s,t} + z''_s \mathbf{x}^2_{s,t} + R^z_{s,t},\end{aligned}$$ and from , ${\left\lvertR^z_{s,t}\right\rvert}^q$ is bounded above by $$\begin{aligned}
C_{p,q} \left[ {\left\lVert\phi''\right\rVert}^q_{\mathcal{V}^p} \left( {\left\lVert\mathbf{x}^3\right\rVert}^{\frac{p}{3}}_{\frac{p}{3}-var; [s,t]} \right)^{\frac{3q}{p}} + {\left\lVertx\right\rVert}^q_{p-var;[0, T]} {\left\lVertR^{\phi}\right\rVert}^q_{q-var; [s,t]} + {\left\lVertR^{\phi'}\right\rVert}^q_{\frac{p}{2}-var;[s,t]} {\left\lVert\mathbf{x}^2\right\rVert}_{\frac{p}{2}-var;[s,t]}^q \right].\end{aligned}$$ Since $q \geq \frac{p}{3}$, the right side of the above expression is a control and is thus super-additive. Furthermore, ${\left\lVertR^z\right\rVert}_{q-var; [0, T]}$ is bounded above by $$\begin{aligned}
C_{p, q} \left( {\left\lVert\phi''\right\rVert}_{\mathcal{V}^p} {\left\lVert\mathbf{x}^3\right\rVert}_{\frac{p}{3}-var; [0, T]} + {\left\lVertx\right\rVert}_{p-var;[0, T]} {\left\lVertR^{\phi}\right\rVert}_{q-var; [0, T]} + {\left\lVertR^{\phi'}\right\rVert}_{\frac{p}{2}-var;[0, T]} {\left\lVert\mathbf{x}^2\right\rVert}_{\frac{p}{2}-var;[0, T]} \right).\end{aligned}$$ Continuing, we define $$\begin{aligned}
R^{z'}_{s,t} := \phi''_s \mathbf{x}^2_{s, t} + R_{s,t}^{\phi},\end{aligned}$$ which gives $$\begin{aligned}
z'_{s,t} = z''_s x_{s,t} + R^{z'}_{s,t},\end{aligned}$$ as well as $$\begin{aligned}
{\left\lVertR^{z'}\right\rVert}_{\frac{p}{2}-var; [0, T]}
\leq {\left\lVert\phi''\right\rVert}_{\mathcal{V}^p; [0, T]} {\left\lVert\mathbf{x}^2\right\rVert}_{\frac{p}{2}-var; [0, T]} + {\left\lVertR^{\phi}\right\rVert}_{\frac{p}{2}-var; [0, T]}.\end{aligned}$$
For the next proposition, given maps $A \in\mathcal{L}(\mathbb{R}^{d}; \mathcal{L}(\mathcal{U}; \mathcal{V}))$ and $B \in\mathcal{L}(\mathbb{R}^{d}; \mathcal{U})$, we will identify them as tensors (either $\mathcal{L} (\mathcal{U}; \mathcal{V})$-valued or $\mathcal{U}$-valued) $$\begin{aligned}
& A = \sum_{j=1}^{d} a_{j} \, \mathrm{d} e_{j}, \quad a_{j} \in\mathcal{L}(\mathcal{U}; \mathcal{V}),\\
& B = \sum_{j=1}^{d} b_{j} \, \mathrm{d} e_{j}, \quad b_{j} \in\mathcal{U},\end{aligned}$$ and adopt the following notation $$\begin{aligned}
AB := a_{i} (b_{j}) \, \mathrm{d} e_{i} \otimes\mathrm{d} e_{j},\\
BA := a_{j} (b_{i}) \, \mathrm{d} e_{i} \otimes\mathrm{d} e_{j},\\
\mathrm{Sym} (AB) := \frac{1}{2} (AB + BA).\end{aligned}$$
\[controlledLeibniz2\] (Leibniz rule) For $3 \leq p < 4$, let $$\begin{aligned}
& \phi\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L}(\mathcal{U};
\mathcal{V}) \right) ,\\
& \phi^{\prime}\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L}(\mathbb{R}^{d}; \mathcal{L}(\mathcal{U}; \mathcal{V})) \right) ,
\quad\mathrm{and}\\
& \phi^{\prime\prime}\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L}(\mathbb{R}^{d} \otimes\mathbb{R}^{d}; \mathcal{L}(\mathcal{U}; \mathcal{V}))
\right) .\end{aligned}$$ Assume that $\left( \phi, \phi^{\prime}, \phi^{\prime\prime}\right) $ is controlled by $\mathbf{x} \in\mathcal{C}^{p-var} \left( [0, T]; G^{3}\left(
\mathbb{R}^{d} \right) \right) $, with remainder terms $R^{\phi}$ and $R^{\phi^{\prime}}$ of bounded $q$-variation and $\frac{p}{2}$-variation respectively, where $\frac{1}{p} + \frac{1}{q} > 1$ and $q \geq\frac{p}{3}$.
(i) Let $$\begin{aligned}
& \psi\in\mathcal{C}^{p-var}\left( [0,T]; \mathcal{U} \right) ,\\
& \psi^{\prime}\in\mathcal{C}^{p-var}\left( [0,T]; \mathcal{L}(\mathbb{R}^{d}; \mathcal{U} )\right) , \quad\mathrm{and}\\
& \psi^{\prime\prime}\in\mathcal{C}^{p-var}\left( [0,T]; \mathcal{L}(\mathbb{R}^{d} \otimes\mathbb{R}^{d}; \mathcal{U} )\right) .\end{aligned}$$ If $\left( \psi,\psi^{\prime}, \psi^{\prime\prime}\right) $ is controlled by $\mathbf{x}$, then the path $\phi\psi\in\mathcal{C}^{p-var}([0,T];
\mathcal{V})$ given by the composition of $\phi$ and $\psi$ is also controlled by $\mathbf{x}$, with derivative process $$\begin{aligned}
(\phi\psi)^{\prime\prime}\psi+\phi\psi^{\prime}$$ and second derivative process $$\begin{aligned}
(\phi\psi)^{\prime\prime}= \phi^{\prime\prime}\psi+ 2 \, \mathrm{Sym}
(\phi^{\prime}\psi^{\prime}) + \phi\psi^{\prime\prime}.\end{aligned}$$ In addition, we have the bound $$\begin{aligned}
\label{leibnizBound2a}
\left\| \phi\psi\right\| _{p,q-cvar} \leq4 \left\|
\phi\right\| _{p,q-cvar} \left\| \psi\right\| _{p,q-cvar} \left( 1 +
\left\| x\right\| _{p-var; [0, T]} + \left\| \mathbf{x}^{2}\right\|
_{\frac{p}{2}-var; [0,T]} \right) .\end{aligned}$$
(ii) Suppose that $\psi\in\mathcal{C}^{q-var}([0,T]; \mathcal{U})$. Then $\phi\psi\in\mathcal{C}^{p-var}([0,T]; \mathcal{V})$ is also controlled by $\mathbf{x}$, with derivative process $$\begin{aligned}
(\phi\psi)^{\prime}= \phi^{\prime}\psi\end{aligned}$$ and second derivative process $$\begin{aligned}
(\phi\psi)^{\prime\prime}= \phi^{\prime\prime}\psi.\end{aligned}$$ Moreover, we have the bound $$\begin{aligned}
\label{leibnizBound2b}\left\| \phi\psi\right\| _{p,q-cvar} \leq\left\|
\phi\right\| _{p,q-cvar} \left\| \psi\right\| _{\mathcal{V}^{q}; [0, T]}.\end{aligned}$$
\(i) It is trivial to see that ${\left\lVert\phi'\psi + \phi\psi'\right\rVert}_{p-var;[0, T]}$ and ${\left\lVert\phi'' \psi + 2 \, \mathrm{Sym} (\phi' \psi') + \phi \psi''\right\rVert}_{p-var; [0, T]}$ satisfy . First we denote $$\begin{aligned}
\tilde{R}^{\phi}_{s,t} := \phi''_s \mathbf{x}^2_{s,t} + R^{\phi}_{s,t}, \\
\tilde{R}^{\psi}_{s,t} := \psi''_s \mathbf{x}^2_{s,t} + R^{\psi}_{s,t},\end{aligned}$$ and since ${\left\lVert\cdot\right\rVert}_{\frac{p}{2}-var;[s, t]} \leq {\left\lVert\cdot\right\rVert}_{q-var;[s, t]}$, we have the bounds $$\begin{aligned}
{\left\lVert\tilde{R}^{\phi}\right\rVert}_{\frac{p}{2}-var;[s, t]}
\leq {\left\lVert\phi''\right\rVert}_{\infty} {\left\lVert\mathbf{x}^2\right\rVert}_{\frac{p}{2}-var; [s, t]} + {\left\lVertR^{\phi}\right\rVert}_{q-var; [s,t]}, \\
{\left\lVert\tilde{R}^{\psi}\right\rVert}_{\frac{p}{2}-var; [s, t]}
\leq {\left\lVert\psi''\right\rVert}_{\infty} {\left\lVert\mathbf{x}^2\right\rVert}_{\frac{p}{2}-var; [s, t]} + {\left\lVertR^{\psi}\right\rVert}_{q-var; [s, t]},\end{aligned}$$ for all $s, t$ in $[0, T]$. Continuing, we compute $$\begin{aligned}
\label{phipsiInc}
\begin{split}
(\phi \psi)_{s,t}
&= \phi_{s,t} \psi_{s,t} +\phi_{s,t} \psi_s + \phi_s \psi_{s,t} \\
&= \left(\phi'_s x_{s,t} + \tilde{R}^{\phi}_{s,t} \right) \left(\psi'_s x_{s,t} + \tilde{R}^{\psi}_{s,t} \right) + \left(\phi'_s x_{s,t} + \phi''_s \mathbf{x}_{s,t}^2 + R^{\phi}_{s,t} \right) \psi_s + \phi_s \left(\psi'_s x_{s,t} + \psi''_s \mathbf{x}_{s,t}^2 + R^{\psi}_{s,t} \right) \\
&= \phi'_s x_{s,t} \left( \psi'_s x_{s,t} \right) + \left( \phi'_s \psi_s + \phi_s \psi'_s \right) x_{s,t} + \left(\phi''_s \psi_s + \phi_s \psi''_s \right) \mathbf{x}^2_{s,t} \\
&\qquad + \phi'_s x_{s,t} \tilde{R}^{\psi}_{s,t} + \tilde{R}^{\phi}_{s,t} \psi'_s x_{s,t} + \tilde{R}^{\phi}_{s,t} \tilde{R}^{\psi}_{s,t} + R^{\phi}_{s,t} \psi_s + \phi_s R^{\psi}_{s,t}.
\end{split}\end{aligned}$$ Denoting $$\begin{aligned}
&\phi'_s = \sum_{i=1}^d \phi'_i (s) \, \mathrm{d} e_i, \quad \phi'_i(s) \in \mathcal{L}(\mathcal{U}; \mathcal{V}) \:\: \forall i \in \{1, \ldots, d \} \, \mathrm{and} \, s \in [0, T], \\
&\psi'_s = \sum_{i=1}^d \psi'_i (s) \, \mathrm{d} e_i, \quad \psi'_i(s) \in \mathcal{U} \:\: \forall i \in \{1, \ldots, d \} \, \mathrm{and} \, s \in [0, T],\end{aligned}$$ we have $$\begin{aligned}
\phi'_s x_{s,t} \left( \psi'_s x_{s,t} \right)
&= \sum_{i,j=1}^d \phi'_i (s) ( \psi'_j (s)) \, x^{(i)}_{s,t} x^{(j)}_{s,t} \\
&= \left( \sum_{i, j =1}^d \left( \phi'_i (s) (\psi'_j(s)) + \phi'_j(s) (\psi'_i(s)) \right) \mathrm{d} e_i \otimes \mathrm{d} e_j \right) \frac{1}{2} \left( x_{s,t} \otimes x_{s,t} \right) \\
&= 2 \, \mathrm{Sym} (\phi'_s \psi'_s) \, \mathbf{x}^2_{s,t}\end{aligned}$$ Thus, continuing from , we have $$\begin{aligned}
(\phi \psi)_{s,t}
= \left( \phi'_s \psi_s + \phi_s \psi'_s \right) x_{s,t} + \left(\phi''_s \psi_s + 2 \, \mathrm{Sym} (\phi'_s \psi'_s) + \phi_s \psi''_s \right) \mathbf{x}^2_{s,t} + R^{\phi\psi}_{s,t},\end{aligned}$$ where $$\begin{aligned}
R^{\phi\psi}_{s,t}
:= \phi'_s x_{s,t} \tilde{R}^{\psi}_{s,t} + \tilde{R}^{\phi}_{s,t} \psi'_s x_{s,t} + \tilde{R}^{\phi}_{s,t} \tilde{R}^{\psi}_{s,t} + R^{\phi}_{s,t} \psi_s + \phi_s R^{\psi}_{s,t}.\end{aligned}$$ We can use the fact that $\frac{4q}{p} > \frac{3q}{p} \geq 1$ to show that ${\left\lVertR^{\phi\psi}\right\rVert}_{q-var; [0, T]}$ is bounded above by $$\begin{aligned}
{\left\lVertx\right\rVert}_{p} \left( {\left\lVert\phi'\right\rVert}_{\infty} {\left\lVert\tilde{R}^{\psi}\right\rVert}_{\frac{p}{2}} + {\left\lVert\psi'\right\rVert}_{\infty} {\left\lVert\tilde{R}^{\phi}\right\rVert}_{\frac{p}{2}} \right)
+ {\left\lVert\tilde{R}^{\phi}\right\rVert}_{\frac{p}{2}} {\left\lVert\tilde{R}^{\psi}\right\rVert}_{\frac{p}{2}} + {\left\lVert\psi\right\rVert}_{\infty} {\left\lVertR^{\phi}\right\rVert}_{q} + {\left\lVert\phi\right\rVert}_{\infty} {\left\lVertR^{\psi}\right\rVert}_{q},\end{aligned}$$ where here we use ${\left\lVert\cdot\right\rVert}_p$ as short-hand for ${\left\lVert\cdot\right\rVert}_{p-var; [0, T]}$.
Moving on, we need to show that $$\begin{aligned}
(\phi \psi)'_{s,t} = (\phi \psi)''_s x_{s,t} + R^{(\phi\psi)'}_{s,t}.\end{aligned}$$ We have $$\begin{aligned}
\left( \phi \psi \right)'_{s,t}
&= \phi'_t \psi_t + \phi_t \psi'_t - \phi'_s \psi_s - \phi_s \psi'_s \\
&= \phi'_{s,t} \psi_s + \phi_s \psi'_{s,t} + \phi'_s \psi_{s,t} + \phi_{s,t} \psi'_s + \phi'_{s,t} \psi_{s,t} + \phi_{s,t} \psi_{s,t} \\
&= \left( \phi''_s x_{s,t} + R^{\phi'}_{s,t} \right) \psi_s + \phi_s \left( \psi''_s x_{s,t} + R^{\psi'}_{s,t} \right) + \phi'_s \left( \psi'_s x_{s,t} + \tilde{R}^{\psi}_{s,t} \right) \\
&\qquad \qquad+ \left( \phi'_s x_{s,t} + \tilde{R}^{\phi}_{s,t} \right) \psi'_s + \phi'_{s,t} \psi_{s,t} + \phi_{s,t} \psi'_{s,t} \\
&= \left( \phi''_s \psi_s + \phi_s \psi''_s \right) x_{s,t} + \phi'_s (\psi'_s x_{s,t}) + \phi'_s x_{s,t} (\psi'_s) \\
&\qquad \qquad+ \underset{\displaystyle =: R_{s,t}^{(\phi\psi)'}}{\underbrace{ \phi'_s \tilde{R}^{\psi}_{s,t} + \tilde{R}^{\phi}_{s,t} \psi'_s + R^{\phi'}_{s,t} \psi_s + \phi_s R^{\psi'}_{s,t} + \phi'_{s,t} \psi_{s,t} + \phi_{s,t} \psi'_{s,t}}}.\end{aligned}$$ We have $$\begin{aligned}
\phi'_s (\psi'_s x_{s,t}) + \phi'_s x_{s,t} (\psi'_s)
&= \sum_{i,j=1}^d \phi'_j(s) (\psi'_i(s)) x^{(i)}_{s,t} \, \mathrm{d} e_j + \sum_{i,j=1}^d \phi'_i(s) (\psi'_j(s)) x^{(i)}_{s,t} \, \mathrm{d} e_j \\
&= 2 \, \mathrm{Sym} (\phi'_s \psi'_s) x_{s,t},\end{aligned}$$ and again ${\left\lVertR^{(\phi\psi)'}\right\rVert}_{\frac{p}{2}-var; [0, T]}$ is bounded above by $$\begin{aligned}
{\left\lVert\phi'\right\rVert}_{\infty} {\left\lVert\tilde{R}^{\psi}\right\rVert}_{\frac{p}{2}} + {\left\lVert\psi'\right\rVert}_{\infty} {\left\lVert\tilde{R}^{\phi}\right\rVert}_{\frac{p}{2}} + {\left\lVert\psi\right\rVert}_{\infty} {\left\lVertR^{\phi'}\right\rVert}_{\frac{p}{2}}
+ {\left\lVert\phi\right\rVert}_{\infty} {\left\lVertR^{\psi'}\right\rVert}_{\frac{p}{2}} + {\left\lVert\phi'\right\rVert}_{p} {\left\lVert\psi\right\rVert}_{p} + {\left\lVert\phi\right\rVert}_{p} {\left\lVert\psi'\right\rVert}_{p}.\end{aligned}$$ (ii) Note that ${\left\lVert\phi' \psi\right\rVert}_{p-var; [0, T]}$ and ${\left\lVert\phi'' \psi\right\rVert}_{p-var; [0, T]}$ satisfy . Moreover, we have $$\begin{aligned}
\left( \phi \psi \right)_{s,t}
&= \phi_{s,t} \psi_s + \phi_t \psi_{s,t} \\
&= \left( \phi'_s x_{s,t} + \phi''_s \mathbf{x}^2_{s,t} + R^{\phi}_{s, t} \right) \psi_s + \phi_t \psi_{s,t} \\
&= \phi'_s \psi_s x_{s,t} + \phi''_s \psi_s \mathbf{x}^2_{s,t} + \underset{\displaystyle =: R^{\phi\psi}_{s,t}}{\underbrace{R^{\phi}_{s,t} \psi_s + \phi_t \psi_{s,t}}},\end{aligned}$$ and thus $$\begin{aligned}
{\left\lVert R^{\phi\psi}\right\rVert}_{q-var; [0, T]}
\leq {\left\lVert\psi\right\rVert}_{\infty} {\left\lVertR^{\phi}\right\rVert}_{q-var; [0, T]} + {\left\lVert\phi\right\rVert}_{\infty} {\left\lVert\psi\right\rVert}_{q-var; [0, T]}.\end{aligned}$$ Continuing, we have $$\begin{aligned}
(\phi \psi)'_{s,t}
&= \phi'_{s,t} \psi_s + \phi'_t \psi_{s,t} \\
&= \left(\phi''_s x_{s,t} + R^{\phi'}_{s,t} \right) \psi_s + \phi'_t \psi_{s,t} \\
&= \phi''_s \psi_s x_{s,t} + \underset{\displaystyle =: R^{(\phi\psi)'}_{s,t}}{\underbrace{R^{\phi'}_{s,t} \psi_s + \phi'_t \psi_{s,t}}},\end{aligned}$$ which implies that $$\begin{aligned}
{\left\lVertR^{(\phi \psi)'}\right\rVert}_{\frac{p}{2}-var; [0, T]}
\leq {\left\lVert\psi\right\rVert}_{\infty} {\left\lVertR^{\phi'}\right\rVert}_{\frac{p}{2}-var; [0, T]} + {\left\lVert\phi'\right\rVert}_{\infty} {\left\lVert\psi\right\rVert}_{q-var; [0, T]}.\end{aligned}$$
\[controlledSmoothMap2\] Let $\mathbf{x} \in\mathcal{C}^{p-var} \left(
[0, T]; G^{3}\left( \mathbb{R}^{d} \right) \right) $ where $3 \leq p < 4$. We assume that $$\begin{aligned}
& y \in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{U} \right) ,\\
& y^{\prime}\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} \left(
\mathbb{R}^{d}, \mathcal{U} \right) \right) ,\\
& y^{\prime\prime}\in\mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} \left(
\mathbb{R}^{d} \otimes\mathbb{R}^{d}; \mathcal{U} \right) \right) ,\end{aligned}$$ and $\left( y, y^{\prime}, y^{\prime\prime}\right) $ is controlled by $\mathbf{x}$ with remainder terms $R^{y}$ and $R^{y^{\prime}}$ of bounded $q$-variation and $\frac{p}{2}$-variation respectively, where $\frac{1}{p} +
\frac{1}{q} > 1$ and $q \geq\frac{p}{3}$. Let $\phi\in\mathcal{C}_{b}^{3}
\left( \mathcal{U}, \mathcal{V} \right) $ and define $$\begin{aligned}
\left( z_{s}, z_{s}^{\prime}, z_{s}^{\prime\prime}\right) := \left(
\phi\left( y_{s}\right) ,\nabla\phi\left( y_{s}\right) y_{s}^{\prime
},\nabla\phi\left( y_{s}\right) y_{s}^{\prime\prime}+\nabla^{2}\phi\left(
y_{s}\right) \left( y_{s}^{\prime},y_{s}^{\prime}\right) \right)\end{aligned}$$ for all $s \in[0, T]$. Then $\left( z, z^{\prime}, z^{\prime\prime}\right) $ is controlled by $\mathbf{x}$, and we have the following bounds $$\begin{aligned}
\left\| z\right\| _{p-var; \left[ 0,T \right] } & \leq\left\|
\phi\right\| _{\mathcal{C}_{b}^{3}} \left\| y\right\| _{p-var; \left[
0,T\right] },\\
\left\| z^{\prime}\right\| _{p-var; \left[ 0,T\right] } & \leq\left\|
\phi\right\| _{\mathcal{C}_{b}^{3}} \left\| y\right\| _{\mathcal{V}^{p}; [0,
T]} \left\| y^{\prime}\right\| _{\mathcal{V}^{p}; [0, T]},\\
\left\| z^{\prime\prime}\right\| _{p-var; \left[ 0,T\right] } &
\leq\left\| \phi\right\| _{\mathcal{C}_{b}^{3}} \left\| y\right\|
_{\mathcal{V}^{p}; [0, T]} \left( \left\| y^{\prime\prime}\right\|
_{\mathcal{V}^{p}; [0, T]} + \left\| y^{\prime}\right\| ^{2}_{\mathcal{V}^{p}; [0, T]} \right) ,\end{aligned}$$ and $$\begin{aligned}
\label{est Rz}\left\| R^{z}\right\| _{q-var; \left[ 0,T\right] }, \left\|
R^{z^{\prime}}\right\| _{\frac{p}{2}-var; \left[ 0,T\right] } \leq\left\|
\phi\right\| _{\mathcal{C}_{b}^{3}} \left( 1 + \left\| y\right\| _{p,q-cvar}
\right) ^{3} \left( 1 + \left\| x\right\| _{p-var; \left[ 0,T\right] } +
\left\| \mathbf{x}^{2}\right\| _{\frac{p}{2}-var; \left[ 0,T\right] }
\right) ^{2}.\end{aligned}$$
From Taylor’s theorem we have $$\begin{aligned}
\label{z2}
z_{s,t}=\nabla \phi \left( y_{s}\right) y_{s,t}+\frac{1}{2}\nabla ^{2}\phi
\left( y_{s}\right) \left( y_{s,t}, y_{s,t} \right) + R_{s,t}^{Taylor}\end{aligned}$$ for all $s < t$ in $\left[ 0,T\right]$, where ${\left\lvertR_{s,t}^{Taylor}\right\rvert} \leq {\left\lVert\phi\right\rVert}_{\mathcal{C}_b^3} {\left\lverty_{s,t}\right\rvert}^3$. From this it follows that $$\begin{aligned}
\label{Tremainder2}
{\left\lVertR^{Taylor}\right\rVert}_{q-var; \left[ 0,T\right] }
\leq {\left\lVert\phi\right\rVert}_{\mathcal{C}_b^3} {\left\lVerty\right\rVert}_{p-var; \left[ 0, T \right] }^3.\end{aligned}$$ As in the previous proposition, we define $$\begin{aligned}
\tilde{R}^{y}_{s, t} := y''_s \mathbf{x}^2_{s,t} + R^y_{s,t},\end{aligned}$$ and note that $y_{s,t} = y'_s x_{s,t} + \tilde{R}^y_{s,t}$ and $$\begin{aligned}
\label{rTildeEst}
{\left\lVert\tilde{R}^y\right\rVert}_{\frac{p}{2}-var; [s, t]} \leq {\left\lVerty''\right\rVert}_{\infty} {\left\lVert\mathbf{x}^2\right\rVert}_{\frac{p}{2}-var; [s, t]} + {\left\lVertR^y\right\rVert}_{\frac{p}{2}-var; [s, t]}\end{aligned}$$ for all $s, t$ in $[0, T]$. We next use the fact that $(y, y' , y'')$ is controlled by $\mathbf{x}$ in equation , which yields $$\begin{aligned}
z_{s,t}
&= \nabla \phi \left( y_{s}\right) \left( y'_s x_{s,t} + y''_s \mathbf{x}_{s,t}^2 + R_{s,t}^y \right) +\frac{1}{2} \nabla^2 \phi \left( y_{s}\right) \left( y'_s x_{s,t} + \tilde{R}^y_{s,t}, y'_s x_{s,t} + \tilde{R}^y_{s,t} \right) + R_{s,t}^{Taylor} \\
&= \underset{\displaystyle = z'_s}{\underbrace{\nabla \phi \left( y_{s}\right) y'_s}} x_{s,t} + \underset{\displaystyle = z''_s}{\underbrace{\left( \nabla \phi \left( y_{s}\right) y''_s + \nabla^2 \phi \left( y_{s}\right) \left( y_{s}^{\prime }, y_{s}^{\prime
}\right) \right) }}\mathbf{x}_{s,t}^{2}+\underset{\displaystyle =: R_{s,t}^z}{\underbrace{\nabla \phi \left( y_{s}\right) R_{s,t}^{y}+E_{s,t}+R_{s,t}^{Taylor}}},\end{aligned}$$ where $$\begin{aligned}
E_{s,t}
&:= \nabla^2 \phi \left( y_s\right) \left( y'_{s} x_{s,t}, \tilde{R}^y_{s,t} \right) + \frac{1}{2} \nabla^2 \phi \left( y_{s}\right) \left( \tilde{R}_{s,t}^y, \tilde{R}_{s,t}^y \right),\end{aligned}$$ and we have used the fact that $\frac{1}{2} \nabla ^{2}\phi \left( y_{s}\right) \left( y'_{s} x_{s,t}, y'_{s} x_{s,t} \right) = \nabla^2 \phi \left( y_{s}\right) \left( y'_s, y'_s \right) \mathbf{x}_{s,t}^2$ (recall that $\mathbf{x}$ is assumed to be weakly-geometric). The stated estimates on the $p^{th}$-variation of $(z, z', z'')$ are then easily derived. We get $$\begin{aligned}
{\left\lVertE\right\rVert}_{q-var; [0, T]}
\leq {\left\lVert\phi\right\rVert}_{\mathcal{C}^3_b} \left( {\left\lVerty'\right\rVert}_{\infty} {\left\lVertx\right\rVert}_{p-var; [0, T]} {\left\lVert\tilde{R}^y\right\rVert}_{\frac{p}{2}-var; [0, T]} + {\left\lVert\tilde{R}^y\right\rVert}^2_{\frac{p}{2}-var; [0, T]} \right).\end{aligned}$$ After using and adding the $q^{th}$-variation bounds of $\nabla \phi(y) R^y$ and $R^{Taylor}$, we get bound for $R^z$.
Proceeding, we can apply Lemma 3.5 from [@cl2018] to $\nabla \phi (y)$ to obtain $$\begin{aligned}
R^{\nabla \phi(y)} \leq {\left\lVert\phi\right\rVert}_{\mathcal{C}^3_b} \left( {\left\lVerty\right\rVert}^2_{p-var; [0, T]} + {\left\lVert\tilde{R}^y\right\rVert}_{\frac{p}{2}-var; [0, T]} \right).\end{aligned}$$ Furthermore, if we apply Lemma 3.6 from [@cl2018] with $\phi$ replaced with $\nabla \phi(y)$ and $\psi$ replaced by $y'$, we obtain $$\begin{aligned}
z'_{s,t} = \left( \nabla \phi(y) \, y' \right)_{s,t}
&=: (\phi \psi)_{s,t} \\
&= (\phi'_s \psi_s + \phi_s \psi'_s) \, x_{s,t} + R^{\phi\psi}_{s,t} \\
&= \underset{\displaystyle = z''_s}{\underbrace{\left( \nabla^2 \phi (y_s) \left(y'_s, y'_s \right) + \nabla \phi(y_s) \, y''_s \right)}} \, x_{s,t} + \underset{\displaystyle =: R^{z'}_{s,t}}{\underbrace{R^{\phi\psi}_{s,t}}}\end{aligned}$$ and $$\begin{aligned}
{\left\lVertR^{z'}\right\rVert}_{\frac{p}{2}-var; [0, T]}
&\leq \left( {\left\lVert\nabla \phi (y)\right\rVert}_{\mathcal{V}^p; [0, T]} + {\left\lVertR^{\nabla \phi(y)}\right\rVert}_{\frac{p}{2}-var; [0, T]} \right) \left( {\left\lVerty'\right\rVert}_{\mathcal{V}^p; [0, T]} + {\left\lVertR^{y'}\right\rVert}_{\frac{p}{2}-var; [0, T]} \right) \\
&\leq {\left\lVert\phi\right\rVert}_{\mathcal{C}^3_b} \left( \left( 1 + {\left\lVerty\right\rVert}_{p-var; [0, T]} \right)^2 + {\left\lVert\tilde{R}^y\right\rVert}_{\frac{p}{2}-var; [0, T]} \right) \left( {\left\lVerty'\right\rVert}_{\mathcal{V}^p; [0, T]} + {\left\lVertR^{y'}\right\rVert}_{\frac{p}{2}-var; [0, T]} \right).\end{aligned}$$
The following theorem extends Theorem 3.1 in [@cl2018].
\[controlledRDE\] Consider the system of RDEs $$\begin{aligned}
& \mathrm{d} y_{t} = V(y_{t}) \circ\mathrm{d} \mathbf{x}_{t}, \quad y_{0} = a
\in\mathbb{R}^{e}, \\
& \mathrm{d} J^{\mathbf{x}}_{t} = \nabla V (y_{t}) \left( \circ\mathrm{d} \mathbf{x}_{t} \right) J^{\mathbf{x}}_{t}, \quad J^{\mathbf{x}}_{0} = \mathcal{I}_{e},\end{aligned}$$ where $V=(V_{1},\ldots,V_{d})$ is a collection of $\mathbb{R}^{e}$-valued vector fields. If $\mathbf{x} = \left( 1, x,\mathbf{x}^{2}, \mathbf{x}^{3} \right) \in\mathcal{C}^{p-var} \left( [0, T]; G^{3} \left( \mathbb{R}^{d}\right) \right) $, $3 \leq p < 4$, and $V$ is in $\mathcal{C}^{4}_{b}$, then both $(y, V(y), V^2(y) )$ and\
$\left( J^{\mathbf{x}}, \left( J^{\mathbf{x}} \right) ^{\prime}, \left( J^{\mathbf{x}} \right)^{\prime\prime}\right) $ are controlled by $\mathbf{x}$. In addition, $$\begin{aligned}
\label{boundedVF2}
\left\| y\right\| _{p,q-cvar} \leq C_{p} \left( 1 + \left\Vert V\right\Vert _{C_{b}^{3}} \right) ^{10} \left( 1 + \left\Vert \mathbf{x}\right\Vert_{p-var;[0,T]}\right)^{8},\end{aligned}$$ and $$\begin{aligned}
\label{linearVF2}
\left\| J^{\mathbf{x}}\right\| _{p,q-cvar} \leq C_{1} \left( 1 + \exp\left( C_{2} N^{\mathbf{x}}_{1; [0, T]} \right) \right) ^{10} \left( 1 + \left\Vert \mathbf{x}\right\Vert _{p-var;[0,T]}\right)^{8},\end{aligned}$$ where $C_{1}$, $C_{2}$ depend on $p$ and $\left\| V\right\| _{\mathcal{C}^{4}_{b}}$.
Using Corollary 10.15 in [@fv2010b], for $\gamma > p$ and $s,t \in [0, T]$, we have $$\begin{aligned}
{\left\lverty_{s,t} - V(y_s) x_{s,t} - V^2(y_s) \mathbf{x}^2_{s,t} - V^3(y_s) \mathbf{x}^3_{s, t}\right\rvert}
\leq C_p \left( {\left\lVertV\right\rVert}_{\mathcal{C}^3_b} {\left\lVert\mathbf{x}\right\rVert}_{p-var;[s, t]} \right)^{\gamma},\end{aligned}$$ where $V^2(y_s)$ and $V^3(y_s)$ denote the following tensors $$\begin{aligned}
&V^2(y_s) := \nabla V(y_s) (V(y_s)) \in \mathbb{R}^e \otimes \mathbb{R}^d \otimes \mathbb{R}^d \quad \mathrm{and} \\
&V^3(y_s) := \nabla^2 V(y_s) (V(y_s), V(y_s)) + \nabla V(y_s) \left[ \nabla V(y_s) (V(y_s)) \right] \in \mathbb{R}^e \otimes \mathbb{R}^d \otimes \mathbb{R}^d \otimes \mathbb{R}^d\end{aligned}$$ respectively.
This implies that $$\begin{aligned}
\label{REst2}
\begin{split}
{\left\lvertR^y_{s,t}\right\rvert}^q
&\leq C_q \left( {\left\lvert V^3(y_s) \mathbf{x}^3_{s,t}\right\rvert}^q + \left( {\left\lVertV\right\rVert}_{\mathcal{C}^3_b} {\left\lVert\mathbf{x}\right\rVert}_{p-var; [s,t]} \right)^{\gamma q} \right) \\
&\leq C_q \left( {\left\lVertV\right\rVert}_{\mathcal{C}^3_b}^{3q} {\left\lVert\mathbf{x}\right\rVert}_{p-var;[s,t]}^{3q} + {\left\lVertV\right\rVert}_{\mathcal{C}^3_b}^{\gamma q} {\left\lVert\mathbf{x}\right\rVert}_{p-var;[s,t]}^{\gamma q} \right),
\end{split}\end{aligned}$$ and thus $$\begin{aligned}
{\left\lVertR^y\right\rVert}_{q-var; [0, T]} \leq C_q \left( {\left\lVertV\right\rVert}^3_{\mathcal{C}^3_b} {\left\lVert\mathbf{x}\right\rVert}_{p-var; [0, T]}^3 \vee {\left\lVertV\right\rVert}_{\mathcal{C}^3_b}^{\gamma} {\left\lVert\mathbf{x}\right\rVert}_{p-var; [0, T]}^{\gamma} \right),\end{aligned}$$ from the super-additivity of the right side of . Observe that $$\begin{aligned}
&{\left\lVertV(y)\right\rVert}_{p-var; [0, T]} \leq {\left\lVertV\right\rVert}_{\mathcal{C}^3_b} {\left\lVerty\right\rVert}_{p-var; [0, T]}, \\
&{\left\lVertV^2(y)\right\rVert}_{p-var; [0, T]} \leq {\left\lVertV\right\rVert}_{\mathcal{C}^3_b}^2 \left(1 + {\left\lVerty\right\rVert}_{p-var; [0, T]} \right)^2,\end{aligned}$$ and $$\begin{aligned}
{\left\lVerty\right\rVert}_{p-var;[0, T]}
\leq C_p \left( {\left\lVertV\right\rVert}_{\mathcal{C}^3_b} {\left\lVert\mathbf{x}\right\rVert}_{p-var; [0, T]} \vee {\left\lVertV\right\rVert}_{\mathcal{C}^3_b}^p {\left\lVert\mathbf{x}\right\rVert}^p_{p-var; [0, T]} \right).\end{aligned}$$ Applying Proposition 2.21 from [@cl2018], we also have $$\begin{aligned}
V(y)_{s, t} = V^2(y_s) x_{s,t} + R^{V(y)}_{s, t}, \quad \forall \, s < t \in [0, T],\end{aligned}$$ where $$\begin{aligned}
{\left\lVertR^{V(y)}\right\rVert}_{\frac{p}{2}-var;[0,T]}
\leq {\left\lVertV\right\rVert}_{\mathcal{C}_{b}^3} \left( {\left\lVerty\right\rVert}_{p-var;[0,T]}^2 + {\left\lVertR^y\right\rVert}_{q-var;[0,T]}\right),\end{aligned}$$ and thus is satisfied.
Note that $q \leq \frac{p}{p-1}$ implies that $q \leq \frac{p}{2}$ for all $p \geq 3$. Collecting all the estimates above and choosing $\gamma$ to be in $(p, 4)$ gives us the bound in .
The proof for the Jacobian is the same as that in Theorem 3.1 of [@cl2018]. From Proposition 5 in [@fr2013], we can obtain the bound $$\begin{aligned}
\label{Jsup}
{\left\lVertJ^{\mathbf{x}}\right\rVert}_{\infty} \leq 1 + \exp \left( C_{p, {\left\lVertV\right\rVert}_{\mathcal{C}^3_b}} \left( N^{\mathbf{x}}_{1; [0, T]} + 1 \right) \right) =: 1 + M,\end{aligned}$$ and construct $\mathcal{C}^4_b$ vector fields $U_i(y_t)$ which equal the linear vector field $z \mapsto \nabla V_i (y_t) z$ on the set $\mathcal{W} = \left\{ z \in \mathbb{R}^{e^2} \; \big\vert \; {\left\lvertz\right\rvert} < M + 2 \right\}$, and which satisfy $$\begin{aligned}
{\left\lVertU_i\right\rVert}_{\mathcal{C}^4_b}
\leq {\left\lVertV\right\rVert}_{\mathcal{C}^4_b} (M + 3), \quad i = 1, \ldots, d.\end{aligned}$$ Then the system of RDEs can be rewritten as a bounded RDE $$\begin{aligned}
\mathrm{d} \tilde{y}_t = \tilde{V} (\tilde{y}_t) {\circ \mathrm{d} \mathbf{x}_t}, \quad \tilde{y}_0 = (a, \mathcal{I}_e),\end{aligned}$$ where $\tilde{y} = \left( y, J^{\mathbf{x}} \right) \in \mathbb{R}^{e + e^2}$ and ${\left\lVert\tilde{V}\right\rVert}_{\mathcal{C}^4_b} \leq {\left\lVertV\right\rVert}_{\mathcal{C}^4_b} (M + 3)$.
Finally, an application of to the above equation yields .
Upper bounds on the high-order Malliavin derivatives {#HOMD}
----------------------------------------------------
We now use the results from the proceeding section to obtain upper bounds on the directional derivative. We first recall the following results from [@cl2018] for the formula of $\mathcal{D}_{h_{1},\ldots, h_{n}}^{n}y_{t}$.
\[mdFormula\] Let $p \geq1$ and $q \geq 1$ be such that $\frac{1}{p} + \frac{1}{q} > 1$, and let $n \in \mathbb{N}$ such that $n \geq 2$. Assume $\mathbf{x} \in\mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^{d}\right) \right) $ and suppose $y$ is the path-level solution to the RDE $$\begin{aligned}
\mathrm{d}y_{t}=V\left( y_{t}\right) \circ\mathrm{d}\mathbf{x}_{t}, \quad y_{0}\in\mathbb{R}^{e}\text{ given},\end{aligned}$$ where $V \in\mathcal{C}^{\lfloor p \rfloor+ n}_{b} \left( \mathbb{R}^{e}; \mathbb{R}^{e} \otimes\mathbb{R}^{d} \right) $. Suppose that $g_{1}, \ldots, g_{n} \in \mathcal{C}^{q-var}([0, T]; \mathbb{R}^d)$. Then $\mathcal{D}_{g_{1}, \ldots, g_{n}}^{n} y_{t}$ satisfies the RDE $$\begin{aligned}
\label{df}
\begin{split}
&\mathrm{d} \mathcal{D}_{g_{1}, \ldots, g_{n}}^{n} y_{t}
= \sum_{i=1}^{n} \nabla^{i} V\left( y_{t} \right) A_{i}^{n}\left( t\right) \circ\mathrm{d} \mathbf{x}_{t} + \sum_{i=0}^{n-1}\sum_{j=1}^{n} \nabla^{i} V\left( y_{t} \right) B_{i,j}^{n} \left( t \right) \, \mathrm{d} g_{j} (t), \\
&\mathcal{D}^n_{g_1, \ldots, g_n} y_0 = 0,
\end{split}\end{aligned}$$ where $A_{i}^{n}$ and $B_{i,j}^{n}$ are respectively defined as $$\begin{aligned}
\label{A}A_{i}^{n}\left( t\right) :=\sum_{\pi=\left\{ \pi_{1},\ldots ,\pi_{i}\right\} \in\mathcal{P}\left( \left\{ g_{1},\ldots, g_{n}\right\} \right) }\mathcal{D}_{\pi_{1}}^{\left\vert \pi_{1}\right\vert }y_{t} \tilde{\otimes}\cdots\tilde{\otimes}\mathcal{D}_{\pi_{i}}^{\left\vert \pi_{i}\right\vert } y_{t}, \, t \in \left[ 0, T\right] ,\end{aligned}$$ and $$\begin{aligned}
\label{Bb}
B_{i,j}^{n}\left( t\right) :=\sum_{\pi=\left\{ \pi_{1},\ldots, \pi_{i}\right\} \in\mathcal{P}\left( \left\{ g_{1}, \ldots, g_{j-1}, g_{j+1}, \ldots, g_{n} \right\} \right) }\mathcal{D}_{\pi_{1}}^{\left\vert \pi_{1}\right\vert }y_{t}\tilde{\otimes}\cdots\tilde{\otimes}\mathcal{D}_{\pi_{i}}^{\left\vert \pi_{i}\right\vert } y_{t}.\end{aligned}$$
\[cf\] Under the conditions of the preceding theorem, $\mathcal{D}_{g_{1}, \ldots, g_{n}}^{n} y_{t}$ equals $$\begin{aligned}
\label{exp2}
\sum_{i=2}^{n}\int_{0}^{t} J_{t}^{\mathbf{x}} \left(J_{s}^{\mathbf{x}} \right) ^{-1} \nabla^{i}V\left( y_{s}\right) A_{i}^{n} \left( s\right) \circ\mathrm{d} \mathbf{x}_{s} + \sum_{i=1}^{n-1} \sum_{j=1}^{n} \int_{0}^{t} J_{t}^{\mathbf{x}} \left( J_{s}^{\mathbf{x}} \right)^{-1} \nabla^{i} V\left( y_{s} \right) B_{i,j}^{n}\left(s \right) \, \mathrm{d} g_{j}(s)\end{aligned}$$ for all $n\geq2$.
We now arrive at the main result of this section, which extends Proposition 3.5 in [@cl2018].
\[dir der est\] Let $p \in[2, 4)$ and $q$ be such that $\frac{1}{p} + \frac{1}{q} > 1$. Let $\mathbf{x} \in\mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^{d}\right) \right) $, and $y$ be the solution to the RDE $$\begin{aligned}
\mathrm{d}y_{t}=V\left( y_{t}\right) \circ\mathrm{d}\mathbf{x}_{t}, \quad y_{0}\in\mathbb{R}^{e}\text{ given},\end{aligned}$$ where $V \in\mathcal{C}^{\lfloor p \rfloor+ n}_{b} (\mathbb{R}^{e}; \mathbb{R}^{e} \otimes\mathbb{R}^{d})$. Then there exists a polynomial $P_{d(n)}: \mathbb{R}_{+} \times\mathbb{R}_{+} \rightarrow\mathbb{R}_{+}$ of finite degree $d(n)$ for which $$\begin{aligned}
\label{ubound}
\left\| \mathcal{D}_{g_{1}, \ldots, g_{n}}^{n} y_{\cdot} \right\|_{\mathcal{V}^{p}; \left[ 0,T\right] } \leq P_{d(n)} \left( \left\| \mathbf{x}\right\| _{p-var; [0, T]}, \exp\left( C \, N^{\mathbf{x}}_{1; [0, T]} \right) \right) \prod\limits_{i=1}^{n} \left\| g_i \right\| _{q-var; [0, T]},\end{aligned}$$ for any $g_{1}, \ldots, g_{n} \in \mathcal{C}^{q-var}([0, T]; \mathbb{R}^d)$. Here $N_{1}^{\mathbf{x}}$ is defined as in , and the constant $C$ as well as the coefficients of $P_{d(n)}$ depend only on $\left\| V\right\|_{\mathcal{C}^{\lfloor p \rfloor+ n}_{b}}$, $p$ and $q$ ($= \frac{p}{2}$ when $2 \leq p < 3 $).
We shall omit the proof as it proceeds in virtually the same manner as that of Proposition 3.5 in [@cl2018], for which the reader is invited to consult. The only difference is that for the case $p \geq 3$, one will have to use Theorems 5.2 and 5.5, as well as Propositions 5.3 and 5.4 in lieu of Theorem 2.20 and 3.5, and Propositions 2.22 and 2.21 respectively from [@cl2018].
Augmenting the higher-order iterated integrals
----------------------------------------------
For this section, we will use $\pi(n) := \left\{ t^{n}_{i} \right\} $ to denote the $n^{th}$ dyadic partition of $[0, T]$, i.e. $t^{n}_{i} = \frac{iT}{2^{n}}$ for $i = 0, \ldots, 2^{n}$, and $\Delta^{n}_{i}$ to denote the interval $\left[ t^{n}_{i}, t^{n}_{i+1}\right] $.
In addition, $\rho^{\prime}$ will denote the Hölder conjugate of $\rho$, i.e. $\frac{1}{\rho} + \frac{1}{\rho^{\prime}} = 1$.
The following proposition giving bounds for the compensated second-order terms was proven in [@cl2018].
\[2ndlevel\] Let $X$ be a continuous, centered Gaussian process in $\mathbb{R}^{d}$ with i.i.d. components, and for $p \in[2, 4)$, let $\mathbf{X} \in\mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^{d}\right) \right) $ denote the geometric rough path constructed from the limit of the piecewise-linear approximations of $X$.
Let $\rho$ and $q$ be such that $\rho\in\left[ 1, 2 \right) $ and $\frac{1}{p} + \frac{1}{q} > 1$. We assume that the covariance function of $X$ satisfies
(a) $\left\| R\right\| _{\rho-var; [0, T]^{2}} < \infty$,
(b) $\left\| R(t, \cdot) - R (s, \cdot)\right\| _{q-var; [0, T]} \leq C
\left| t - s\right| ^{\frac{1}{\rho}}$, for all $s, t \in[0, T]$.
Now let $\psi: \Omega\times[0, T] \rightarrow\mathbb{R}^{d} \otimes \mathbb{R}^{d}$ be a stochastic process satisfying $\displaystyle \psi_{t} = \sum_{a, b = 1}^{d} \psi_{t}^{(a, b)} \mathrm{d} e_{a} \otimes\mathrm{d} e_{b}
\in\mathbb{D}^{4,2} (\mathbb{R}^{d} \otimes\mathbb{R}^{d})$ for all $t \in [0, T]$. Furthermore, assume there exists $C < \infty$ such that for all $s, t \in [0, T]$ and $a, b = 1, \ldots, d$, we have $$\begin{aligned}
\label{propCond0}
\left| \mathbb{E} \left[ \psi^{(a, b)}_{s} \psi^{(a, b)}_{t}\right] \right| \leq C,\end{aligned}$$ and for $k = 2, 4$, we have $$\begin{aligned}
\label{propCond}
\left| \mathbb{E} \left[ \mathcal{D}^{k}_{h_{1}, \ldots, h_{k}} \left( \psi^{(a, b)}_{s} \psi^{(a, b)}_{t} \right) \right] \right|
\leq C \prod_{i=1}^{k} \left\| \Phi(h_{i})\right\|_{q-var; [0, T]},\end{aligned}$$ for all $h_{1}, \ldots, h_{k} \in\mathcal{H}_{1}^{d}$. Then $$\begin{aligned}
\lim_{n \rightarrow \infty} \left\| \sum_{i=0}^{2^{n} - 1} \psi_{t^{n}_{i}}
\left( \mathbf{X}^{2}_{t^{n}_{i}, t^{n}_{i+1}} - \frac{1}{2} \sigma^{2}
\left( t^{n}_{i}, t^{n}_{i+1} \right) \mathcal{I}_{d} \right) \right\|_{L^{2}(\Omega)} = 0.\end{aligned}$$
We will now proceed to give a similar estimate for the third-order terms. We first begin with the following lemma.
\[productExp\] For any $h_{1}, \ldots, h_{6} \in\mathcal{H}_{1}^{d}$, we have $$\begin{aligned}
\mathbb{E} \left[ \psi_{t^{n}_{i}} \psi_{t^{n}_{j}} \prod_{k=1}^{6}
I_{1}(h_{k})\right] = \mathbb{E} \left[ \mathcal{D}_{h_{1}, h_{2}, h_{3},
h_{4}, h_{5}, h_{6}}^{6} \psi_{t^{n}_{i}} \psi_{t^{n}_{j}}\right] +
\sum_{\sigma\in\mathcal{S}_{6}} C_{\sigma, 1} A_{\sigma, 1} + C_{\sigma, 2}
A_{\sigma, 2} + C_{\sigma, 3} A_{\sigma, 3},\end{aligned}$$ where $$\begin{aligned}
& A_{\sigma, 1} := \mathbb{E} \left[ \mathcal{D}_{h_{\sigma(1)},
h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)}}^{4} \psi_{t^{n}_{i}} \psi
_{t^{n}_{j}}\right] \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle
_{\mathcal{H}_{1}^{d}},\\
& A_{\sigma, 2} := \mathbb{E} \left[ \mathcal{D}_{h_{\sigma(1)},
h_{\sigma(2)}}^{2} \psi_{t^{n}_{i}} \psi_{t^{n}_{j}}\right] \left\langle
h_{\sigma(3)}, h_{\sigma(4)} \right\rangle _{\mathcal{H}_{1}^{d}} \left\langle
h_{\sigma(5)}, h_{\sigma(6)} \right\rangle _{\mathcal{H}_{1}^{d}},\\
& A_{\sigma, 3} := \mathbb{E} \left[ \psi_{t^{n}_{i}} \psi_{t^{n}_{j}
}\right] \left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle
_{\mathcal{H}_{1}^{d}} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle
_{\mathcal{H}_{1}^{d}} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle
_{\mathcal{H}_{1}^{d}},\end{aligned}$$ $\mathcal{S}_{6}$ denotes the symmetric group of permutations on $\{1, \ldots, 6\}$, and the $C_{\sigma, k}$’s are constants that depend on the permutation $\sigma$.
From the product formula we have $$\begin{aligned}
\label{product6}
\begin{split}
\prod_{k=1}^6 I_1(h_k)
&= I_6 (h_1 \otimes h_2 \otimes h_3 \otimes h_4 \otimes h_5 \otimes h_6) \\
&\qquad + \sum_{\sigma \in \mathcal{S}_6} C_{\sigma, 1} \, I_4 \left( h_{\sigma(1)} \otimes h_{\sigma(2)} \otimes h_{\sigma(3)} \otimes h_{\sigma(4)} \right) \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\qquad \qquad+ C_{\sigma, 2} \, I_2 \left( h_{\sigma(1)} \otimes h_{\sigma(2)} \right) \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\qquad\qquad + C_{\sigma, 3} \, \left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}.
\end{split}\end{aligned}$$ Applying integration-by-parts finishes the proof.
\[3rdlevel\] Let $\mathbf{X} \in\mathcal{C}^{0, p-var} \left( [0, T]; G^{3} \left( \mathbb{R}^{d}\right) \right) $, $3 \leq p < 4$, be a geometric Gaussian rough path which satisfies Condition \[newCond1\], and assume that its covariance function satisfies, for all $s, t \in[0, T]$, $$\begin{aligned}
\left\| R(t, \cdot) - R (s, \cdot) \right\| _{q-var; [0, T]} \leq C \left| t - s\right| ^{\frac{1}{\rho}},\end{aligned}$$ for some finite constant $C$ and $\rho\in[1, 2)$.
Let $\psi: \Omega\times[0, T] \rightarrow\mathbb{R}^{d} \otimes\mathbb{R}^{d} \otimes\mathbb{R}^{d}$ be a stochastic process satisfying $\displaystyle \psi_{t} = \sum_{a,b,c = 1}^{d} \psi_{t}^{(a,b,c)} \mathrm{d} e_{a} \otimes \mathrm{d} e_{b} \otimes \mathrm{d} e_{c} \in \mathbb{D}^{6,2} (\mathbb{R}^{d} \otimes \mathbb{R}^{d} \otimes \mathbb{R}^{d})$ for all $t \in [0, T]$. Furthermore, assume there exists $C < \infty$ such that we have $$\begin{aligned}
\left| \mathbb{E} \left[ \psi^{(a, b)}_{s} \psi^{(a, b)}_{t}\right] \right| \leq C,\end{aligned}$$ and for $k = 2, 4, 6$, we have $$\begin{aligned}
\label{propCond2}
\left| \mathbb{E} \left[ \mathcal{D}^{k}_{h_{1}, \ldots, h_{k}} \left( \psi^{(a, b, c)}_{s} \psi^{(a, b, c)}_{t} \right) \right] \right|
\leq C \prod_{i=1}^{k} \left\| \Phi(h_{i})\right\|_{q-var; [0, T]},\end{aligned}$$ for all $h_{1}, \ldots, h_{k} \in\mathcal{H}_{1}^{d}$, $s, t \in[0, T]$ and $a, b, c = 1, \ldots, d$. Then $$\begin{aligned}
\lim_{n \rightarrow\infty} \left\| \sum_{i=0}^{2^{n} - 1} \psi_{t^{n}_{i}}
\left( \mathbf{X}^{3}_{t^{n}_{i}, t^{n}_{i+1}} \right) \right\|_{L^{2}(\Omega)} = 0.\end{aligned}$$
First note that $$\begin{aligned}
\label{mainQuant2}
\begin{split}
\left\| \sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right) \right\|_{L^2(\Omega)}
\leq \left\| \sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \left(\mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^S \right) \right\|_{L^2(\Omega)}
+ \left\| \sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^{NS} \right) \right\|_{L^2(\Omega)},
\end{split}\end{aligned}$$ where $\left( \mathbf{X}^3 \right)^S$ denotes the symmetric part of $\mathbf{X}^3$ and $$\begin{aligned}
\left( \mathbf{X}^3_{s,t} \right)^{NS} = \mathbf{X}^3_{s,t} - \left( \mathbf{X}^3_{s,t} \right)^S\end{aligned}$$ denotes the non-symmetric part. The two parts will be tackled separately, and since $$\begin{aligned}
{\left\lVert\sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^S \right)\right\rVert}_{L^2(\Omega)}
\leq \sum_{a, b, c = 1}^d {\left\lVert \sum_{i=0}^{2^n - 1} \psi_{t^n_i}^{(a, b, c)} \left( \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^S \right)^{(a, b, c)} \right\rVert}_{L^2(\Omega)},\end{aligned}$$ and similarly for the non-symmetric part, we will study the convergence of each fixed $(a, b, c)^{th}$ tensor component individually and henceforth suppress the notation for the component in the superscript of $\psi$.
\(a) To begin, we will prove that $$\begin{aligned}
\lim_{n \rightarrow \infty} {\left\lVert\sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left(
\mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^S\right\rVert}_{L^2(\Omega)} = 0.\end{aligned}$$ Since $$\begin{aligned}
\left( \mathbf{X}^3_{s,t} \right)^S = \frac{1}{6} X_{s, t} \otimes X_{s,t} \otimes X_{s,t},\end{aligned}$$ this is equivalent to showing that $$\begin{aligned}
{\mathbb{E} \left[ \left( \sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{S} \right)^{(a, b, c)} \right)^2\right]}
= \frac{1}{36} \sum_{i, j=0}^{2^n - 1} {\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} X^{(a)}_{\Delta^n_i} X^{(b)}_{\Delta^n_i} X^{(c)}_{\Delta^n_i} X^{(a)}_{\Delta^n_j} X^{(b)}_{\Delta^n_j} X^{(c)}_{\Delta^n_j}\right]}\end{aligned}$$ converges to zero as $n \rightarrow \infty$. First define $$\begin{aligned}
&h_1 := \mathds{1}_{\Delta^n_i}^{(a)}, \quad h_2 := \mathds{1}_{\Delta^n_i}^{(b)}, \quad h_3 := \mathds{1}_{\Delta^n_i}^{(c)}, \\
&h_4 := \mathds{1}_{\Delta^n_j}^{(a)}, \quad h_5 := \mathds{1}_{\Delta^n_j}^{(b)}, \quad h_6 := \mathds{1}_{\Delta^n_j}^{(c)}.\end{aligned}$$ Note that for $k = 1, \ldots, 6$, we have $$\begin{aligned}
{\left\lVert\Phi(h_k)\right\rVert}_{q-var; [0, T]}
&= {\left\lVertR\left( t^n_{i+1}, \cdot\right) - R\left( t^n_i, \cdot\right)\right\rVert}_{q-var; [0, T]} \quad \mathrm{or} \quad {\left\lVertR\left( t^n_{j+1}, \cdot\right) - R\left( t^n_j, \cdot\right)\right\rVert}_{q-var; [0, T]} \\
&\leq C \, 2^{-\frac{n}{\rho}}\end{aligned}$$ and $$\begin{aligned}
{\left\lVerth_k\right\rVert}_{\mathcal{H}_1^d}
&= \sqrt{\sigma^2 \left( t_i^n, t_{i+1}^n \right)} \quad \mathrm{or} \quad \sqrt{\sigma^2 \left( t_j^n, t_{j+1}^n \right)} \\
&\leq \sqrt{{\left\lVertR\left( t^n_{i+1}, \cdot\right) - R\left( t^n_i, \cdot\right)\right\rVert}_{q-var; [0, T]}} \quad \mathrm{or} \quad \sqrt{{\left\lVertR\left( t^n_{j+1}, \cdot\right) - R\left( t^n_j, \cdot\right)\right\rVert}_{q-var; [0, T]}} \\
&\leq C \, 2^{-\frac{n}{2\rho}}.\end{aligned}$$ Recall from Lemma \[productExp\] that $$\begin{aligned}
&{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} \prod_{k=1}^6 I_1(h_k)\right]}
=: {\mathbb{E} \left[ \mathcal{D}_{h_1, h_2, h_3, h_4, h_5, h_6}^6 \psi_{t^n_i} \psi_{t^n_j}\right]} + \sum_{\sigma \in \mathcal{S}_6} C_{\sigma, 1} A_{\sigma, 1} + C_{\sigma, 2} A_{\sigma, 2} + C_{\sigma, 3} A_{\sigma, 3}.\end{aligned}$$ For the first term on the right side of the above expression, we have $$\begin{aligned}
\sum_{i, j = 0}^{2^n - 1} {\mathbb{E} \left[ \mathcal{D}_{h_1, h_2, h_3, h_4, h_5, h_6}^6 \psi_{t^n_i} \psi_{t^n_j}\right]}
&\leq C \sum_{i, j = 0}^{2^n - 1} \prod_{k=1}^6 {\left\lVert\Phi(h_k)\right\rVert}_{q-var; [0, T]} \\
&\leq C\, 2^{-2n \left( \frac{3}{\rho} - 1 \right)},\end{aligned}$$ which vanishes as $n \rightarrow \infty$ since $\rho < 2$.
For the $A_{\sigma, 1}$ terms we have $$\begin{aligned}
&\sum_{i, j= 0}^{2^n-1} {\mathbb{E} \left[ \mathcal{D}^4_{h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)}} \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&{\qquad \qquad \qquad \qquad}\leq \sum_{i, j = 0}^{2^n-1} \prod_{k=1}^4 {\left\lVert\Phi(h_{\sigma(k)})\right\rVert}_{q-var; [0, T]} {\left\lVerth_{\sigma(5)}\right\rVert}_{\mathcal{H}_1^d} {\left\lVerth_{\sigma(6)}\right\rVert}_{\mathcal{H}_1^d} \\
&{\qquad \qquad \qquad \qquad}\leq C \, 2^{-2n \left( \frac{5}{2\rho} -1 \right)} \rightarrow 0,\end{aligned}$$ and similarly for the $A_{\sigma, 2}$ terms we have $$\begin{aligned}
&\sum_{i, j= 0}^{2^n-1} {\mathbb{E} \left[ \mathcal{D}^2_{h_{\sigma(1)}, h_{\sigma(2)}} \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\qquad \qquad \leq \sum_{i, j = 0}^{2^n-1} {\left\lVert\Phi(h_{\sigma(1)})\right\rVert}_{q-var; [0, T]} {\left\lVert\Phi(h_{\sigma(2)})\right\rVert}_{q-var; [0, T]} \prod_{k=3}^6 {\left\lVerth_{\sigma(k)}\right\rVert}_{\mathcal{H}_1^d} \\
&\qquad \qquad \leq C \, 2^{-2n \left( \frac{2}{\rho} -1 \right)} \rightarrow 0.\end{aligned}$$ Finally for the $A_{\sigma, 3}$ terms we have two cases: either $$\begin{aligned}
\left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
t^n_i & t^n_{i+1} \\
t^n_j & t^n_{j+1}
\end{pmatrix}^3,\end{aligned}$$ or $$\begin{aligned}
\left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
t^n_i & t^n_{i+1} \\
t^n_j & t^n_{j+1}
\end{pmatrix}
\sigma^2 \left( t^n_i, t^n_{i+1} \right) \sigma^2 \left( t^n_j, t^n_{j+1} \right).\end{aligned}$$ In either case, since $$\begin{aligned}
{\left\lvertR \begin{pmatrix}
t^n_i & t^n_{i+1} \\
t^n_j & t^n_{j+1}
\end{pmatrix}\right\rvert},
\sigma^2 \left( t^n_i, t^n_{i+1} \right), \sigma^2 \left( t^n_j, t^n_{j+1} \right)
\leq \frac{C}{2^{\frac{n}{\rho}}},\end{aligned}$$ we have $$\begin{aligned}
\sum_{i, j = 0}^{2^n-1}
{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} &\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\leq C \left( \sum_{i,j = 0}^{2^n-1}
R \begin{pmatrix}
t^n_i & t^n_{i+1} \\
t^n_j & t^n_{j+1}
\end{pmatrix}^{\rho} \right)^{\frac{1}{\rho}}
\left( \sum_{i, j=0}^{2^n-1} 2^{\frac{-2n \rho'}{\rho}} \right)^{\frac{1}{\rho'}} \\
&\leq C \, {\left\lVertR\right\rVert}_{\rho-var;[0, T]^2} 2^{-2n\left( \frac{1}{\rho} - \frac{1}{\rho'}\right)} \\
&\leq C \, {\left\lVertR\right\rVert}_{\rho-var;[0, T]^2} 2^{-2n \left( \frac{2}{\rho} - 1\right)} \rightarrow 0.\end{aligned}$$ (b) We will now move on to show that $$\begin{aligned}
\lim_{n \rightarrow \infty} {\left\lVert\sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left(
\mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^{NS}\right\rVert}_{L^2(\Omega)} = 0.\end{aligned}$$ Let $X^{\pi(k)}$ denote the piece-wise linear approximation of $X$ over $\pi(k)$, and let\
$\mathbf{X}^{\pi(k)} = \left(1, \, \mathbf{X}^1 (\pi(k)), \, \mathbf{X}^2 (\pi(k)), \, \mathbf{X}^3 (\pi(k)) \right) = S_3 \left( X^{\pi(k)} \right)$ denote its canonical lift to a geometric rough path.
Next, define $$\begin{aligned}
\left( \mathbf{X}^3_{s,t} \right)^{NS} (\Delta_l) := \left( \mathbf{X}^3_{s,t} \right)^{NS} \left( \pi (l+1) \right) - \left( \mathbf{X}^3_{s,t} \right)^{NS} \left( \pi (l) \right).\end{aligned}$$ Since $\left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \pi(n) \right) = 0$, we have $$\begin{aligned}
\left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} = \lim_{m\rightarrow\infty}\sum_{k=1}^m
\left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \Delta_{n+k} \right) \text{ for every }n\in \mathbb{N} \; \mathrm{and} \; i = 0, 1, \ldots, 2^n - 1,\end{aligned}$$ where the limit is taken in $L^2 (\Omega)$.
We want to show that $$\begin{aligned}
{\left\lVert\sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} (\pi (n+m)) \right)^{(a, b, c)}\right\rVert}_{L^2(\Omega)} \rightarrow 0\end{aligned}$$ uniformly for all $m$ as $n \rightarrow \infty$. To begin, let $$\begin{aligned}
\label{sDefn}
s_u^{k, i} := t_i^n + \frac{u}{2^{n+k}} = t_{u + i 2^k}^{n+k},\end{aligned}$$ and we will denote the intervals $$\begin{aligned}
\label{LRDefn}
\begin{split}
&\Delta_{u^L}^i := \left[ s^{k+1, i}_{2u}, s^{k+1, i}_{2u+1} \right], \quad \Delta_{u^R}^i := \left[ s^{k+1, i}_{2u+1}, s^{k+1, i}_{2u+2} \right], \\
&\Delta_u^i := \Delta_{u^L}^i \cup \Delta_{u^R}^i = \left[ s^{k, i}_u, s^{k, i}_{u+1} \right] \subset \left[ t^n_i, t^n_{i+1} \right], \quad \forall u = 0, \ldots 2^k - 1.
\end{split}\end{aligned}$$ Note that we suppress the dependence on $k$ and $n$ in the notation for the variables on the left. The following computation on $G^3(\mathbb{R}^d)$ can be verified easily; for $f, g \in \mathbb{R}^d$ we have $$\begin{aligned}
&\exp(f) \otimes \exp(g)
= \left( 1, \, f + g, \, \frac{(f + g)^{\otimes 2}}{2} + \frac{1}{2} [f, g], \, \frac{(f + g)^{\otimes 3}}{6} + N(f, g) \right),\end{aligned}$$ where $$\begin{aligned}
N(f, g) := \frac{1}{4} \big( (f+g) \otimes [f, g] + [f, g] \otimes (f + g) \big) + \frac{1}{12} \big( [f, [f, g]] + [g, [g, f]] \big).\end{aligned}$$ Using the above expression with $f = X_{\Delta_{u^L}^i}$ and $g = X_{\Delta_{u^R}^i}$, for $k = 1, \ldots, m$ we obtain the following identity on $T^3 \left( \mathbb{R}^d \right)$: $$\begin{aligned}
&\bigotimes_{u=0}^{2^k-1} \exp \left( X_{\Delta_{u^L}^i} \right) \otimes \exp \left( X_{\Delta^i_{u^R}} \right) - \bigotimes_{u=0}^{2^k-1} \exp \left( X_{\Delta_u^i} \right) \\
&\quad \qquad = \bigotimes_{u=0}^{2^k-1} \left( 1, X_{\Delta_u^i}, \frac{1}{2} X_{\Delta_u^i}^{\otimes 2}, \frac{1}{6} X_{\Delta_u^i}^{\otimes 3} \right)
+ \left(0, 0, \frac{1}{2} \left[ X_{\Delta^i_{u^L}}, X_{\Delta_{u^R}^i} \right], 0 \right)
+ \left(0, 0, 0, N \left( X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right) \right) \\
&{\qquad \qquad \qquad \qquad}- \bigotimes_{u = 0}^{2^k-1} \left( 1, X_{\Delta_u^i}, \frac{1}{2} X_{\Delta_u^i}^{\otimes 2}, \frac{1}{6} X_{\Delta_u^i}^{\otimes 3} \right) \\
&\quad \qquad = \sum_{u=0}^{2^k-1} \left( 0, \, 0, \, \frac{1}{2} \left[ X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right], \, M\left( X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right) + N\left( X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right) \right)\end{aligned}$$ where $$\begin{aligned}
M\left( X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right)
&:= \sum_{r=0}^{u-1} X_{\Delta^i_r} \otimes \frac{1}{2} \left[ X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right] + \frac{1}{2} \left[ X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right] \otimes \sum_{r=u+1}^{2^k-1} X_{\Delta^i_r} \\
&= X_{t^n_i, s^{k, i}_u} \otimes \frac{1}{2} \left[ X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right] + \frac{1}{2} \left[ X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right] \otimes X_{s^{k,i}_{u+1}, t^n_{i+1}}.\end{aligned}$$ This means that $$\begin{aligned}
\mathbf{X}^3_{t^n_i, t^n_{i+1}} (\pi(n+k+1)) - \mathbf{X}^3_{t^n_i, t^n_{i+1}} (\pi(n+k))
&= \sum_{u=0}^{2^k-1} M_u + N_u,\end{aligned}$$ where we use $M_u$ and $N_u$ as short-hand for $M\left( X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right)$ and $N\left( X_{\Delta^i_{u^L}}, X_{\Delta^i_{u^R}} \right)$ respectively. This in turn gives us $$\begin{aligned}
\left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \Delta_{n+k} \right) = \sum_{u=0}^{2^k-1} M_u + N_u,\end{aligned}$$ since $$\begin{aligned}
\mathbf{X}^3_{t^n_i, t^n_{i+1}} (\pi(n+k+1))& - \mathbf{X}^3_{t^n_i, t^n_{i+1}} (\pi(n+k)) \\
&= \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^S (\pi(n+k+1)) - \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)^S (\pi(n+k)) + \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \Delta_{n+k} \right) \\
&= \exp \left( X_{t^n_i, t^n_{i+1}} \right) - \exp \left( X_{t^n_i, t^n_{i+1}} \right) + \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \Delta_{n+k} \right) \\
&= \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \Delta_{n+k} \right).\end{aligned}$$ Thus, we obtain $$\begin{aligned}
\label{3rdQuant}
\begin{split}
&{\mathbb{E} \left[ \left( \sum_{i=0}^{2^n - 1} \psi_{t^n_i} \left( \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} (\pi (n+m)) \right)^{(a, b, c)} \right)^2\right]} \\
&\quad \qquad \qquad \qquad= {\mathbb{E} \left[ \left( \sum_{i=0}^{2^n - 1} \psi_{t^n_i} \sum_{k=1}^m \left( \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} (\Delta_{n+k}) \right)^{(a, b, c)} \right)^2\right]} \\
&\quad \qquad \qquad \qquad= \sum_{i, j=0}^{2^n - 1} {\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} \sum_{k=1}^m \left( \left( \mathbf{X}^3_{t_i^n, t_{i+1}^n} \right)^{NS} \left( \Delta_{n+k} \right) \right)^{(a, b, c)} \sum_{l=1}^m \left( \left( \mathbf{X}^3_{t_j^n, t_{j+1}^n} \right)^{NS} \left( \Delta_{n+l} \right) \right)^{(a, b, c)}\right]} \\
&\quad \qquad \qquad \qquad= \sum_{i,j=0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} {\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} \left( M_u + N_u\right)^{(a, b, c)} \left( M_v + N_v \right)^{(a, b, c)}\right]}.
\end{split}\end{aligned}$$ In what follows, it does not matter to the analysis which particular subinterval of $\Delta_u^i$, $\Delta_v^j$, $\Delta^n_i$ or $\Delta^n_j$ is present in the terms. Hence we will use the notation $$\begin{aligned}
&\Delta_{u^*} = \Delta_{u^L}^i, \Delta_{u^R}^i \: \mathrm{or} \: \Delta_u^i, \quad \Delta_{v^*} = \Delta_{v^L}^j, \Delta_{v^R}^j \: \mathrm{or} \: \Delta_v^j, \\
&\Delta_{i^*} = \left[t^n_i, s^{k,i}_u \right] \: \mathrm{or} \: \left[ s^{k,i}_{u+1}, t^n_{i+1} \right], \quad \Delta_{j^*} = \left[t^n_j, s^{l,j}_v \right] \: \mathrm{or} \: \left[ s^{l,j}_{v+1}, t^n_{j+1} \right],\end{aligned}$$ and $$\begin{aligned}
R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}
:= \left\langle \mathds{1}_{\Delta_u^*}, \mathds{1}_{\Delta_v^*} \right\rangle_{\mathcal{H}_1}
= R \begin{pmatrix}
a_1 & a_2 \\
b_1 & b_2
\end{pmatrix},\end{aligned}$$ where $[a_1, a_2] = \Delta_{u^L}^i, \Delta_{u^R}^i \: \mathrm{or} \: \Delta_u^i$, and $[b_1, b_2] = \Delta_{v^L}^j, \Delta_{v^R}^j \: \mathrm{or} \: \Delta_v^j$.
$R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix}$, $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{j^*}
\end{pmatrix}$, $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$, $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix}$, $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix}$ and $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}$ can be defined in the same manner, and we have the bounds $$\begin{aligned}
\label{qrBounds}
\begin{split}
{\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert},
{\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}\right\rvert},
{\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix}\right\rvert},
{\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{j^*}
\end{pmatrix}\right\rvert}
&\leq {\left\lVertR \left( \Delta_{u^*}, \cdot \right)\right\rVert}_{q-var; [0, T]} \\
&= {\left\lVert\Phi \left( \mathds{1}_{\Delta_{u^*}} \right)\right\rVert}_{q-var; [0, T]}
\leq \frac{C}{2^{\frac{n+k}{\rho}}}, \\
{\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert},
{\left\lvertR \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert},
{\left\lvertR \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix}\right\rvert},
{\left\lvertR \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix}\right\rvert}
&\leq {\left\lVertR \left( \Delta_{v^*}, \cdot \right)\right\rVert}_{q-var; [0, T]} \\
&= {\left\lVert\Phi \left( \mathds{1}_{\Delta_{v^*}} \right)\right\rVert}_{q-var; [0, T]}
\leq \frac{C}{2^{\frac{n+l}{\rho}}}.
\end{split}\end{aligned}$$ Using the notation $$\begin{aligned}
R_{\Delta_u^i \times \Delta_v^j} : = {\left\lvertR \begin{pmatrix}
s^{k, i}_{2u} & s^{k, i}_{2u+1} \\
s^{l, j}_{2v} & s^{l, j}_{2v+1}
\end{pmatrix}\right\rvert} +
{\left\lvertR \begin{pmatrix}
s^{k, i}_{2u+1} & s^{k, i}_{2u+2} \\
s^{l, j}_{2v} & s^{l, j}_{2v+1}
\end{pmatrix}\right\rvert} +
{\left\lvertR \begin{pmatrix}
s^{k, i}_{2u} & s^{k, i}_{2u+1} \\
s^{l, j}_{2v+1} & s^{l, j}_{2v+2}
\end{pmatrix}\right\rvert} +
{\left\lvertR \begin{pmatrix}
s^{k, i}_{2u+1} & s^{k, i}_{2u+2} \\
s^{l, j}_{2v+1} & s^{l, j}_{2v+2}
\end{pmatrix}\right\rvert},\end{aligned}$$ note that $$\begin{aligned}
\sum_{i, j=0}^{2^n-1} \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} {\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{\rho}
\leq \sum_{i, j=0}^{2^n-1} \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} R^{\rho}_{\Delta^i_u \times \Delta_v^j}
\leq 4^{\rho} {\left\lVertR\right\rVert}_{\rho-var; [0, T]^2}^{\rho}\end{aligned}$$ for all $k,l \in \mathbb{N}$.
For $k = 1, \ldots 6$, let $y_k$ denote $a, b$ or $c$. Returning to , we see that the last line can be expanded to include terms of the type $M_u^{(a, b, c)} M_v^{(a, b, c)}$: $$\begin{aligned}
\label{typeMM}
{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} X^{(y_1)}_{\Delta_{u^*}} X^{(y_2)}_{\Delta_{u^*}} X^{(y_3)}_{\Delta_{i^*}} X^{(y_4)}_{\Delta_{v^*}} X^{(y_5)}_{\Delta_{v^*}} X^{(y_6)}_{\Delta_{j^*}} \right]},\end{aligned}$$ terms coming from $N_u^{(a, b, c)} N_v^{(a, b, c)}$: $$\begin{aligned}
\label{typeNN}
{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} X^{(y_1)}_{\Delta_{u^*}} X^{(y_2)}_{\Delta_{u^*}} X^{(y_3)}_{\Delta_{u^*}} X^{(y_4)}_{\Delta_{v^*}} X^{(y_5)}_{\Delta_{v^*}} X^{(y_6)}_{\Delta_{v^*}} \right]},\end{aligned}$$ and terms arising from $M_u^{(a, b, c)} N_v^{(a, b, c)}$: $$\begin{aligned}
\label{typeMN}
{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} X^{(y_1)}_{\Delta_{u^*}} X^{(y_2)}_{\Delta_{u^*}} X^{(y_3)}_{\Delta_{i^*}} X^{(y_4)}_{\Delta_{v^*}} X^{(y_5)}_{\Delta_{v^*}} X^{(y_6)}_{\Delta_{v^*}} \right]}.\end{aligned}$$ To account for the remaining $N_u^{(a, b, c)} M_v^{(a, b, c)}$ terms, we simply swap $u$ and $v$ in the third case. Note also that with our short-hand notation, as an example, $X_{\Delta_{u^*}}^{(y_1)}$ may not be equal to $X_{\Delta_{u^*}}^{(y_2)}$ even if $y_1 = y_2$ since $\Delta_{u^*}$ may be one of several intervals.
Since $M_u$ is anti-symmetric with respect to $X_{\Delta_{u^L}}$ and $X_{\Delta_{u^R}}$, we can assume that $y_1 \neq y_2$ in and , and $y_4 \neq y_5$ in . In each of the three cases, we will use $I_1(h_k)$ to denote $X^{(y_k)}$ for $k = 1, \ldots, 6$; for example in , $h_1 := \mathds{1}_{\Delta_u^*}^{(y_1)}$ and $I_1 (h_1) = X^{(y_1)}_{\Delta_{u^*}}$. Now applying Lemma \[productExp\], we have $$\begin{aligned}
&{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j} \prod_{k=1}^6 I_1(h_k)\right]}
=: {\mathbb{E} \left[ \mathcal{D}_{h_1, h_2, h_3, h_4, h_5, h_6}^6 \psi_{t^n_i} \psi_{t^n_j}\right]} + \sum_{\sigma \in \mathcal{S}_6} C_{\sigma, 1} A_{\sigma, 1} + C_{\sigma, 2} A_{\sigma, 2} + C_{\sigma, 3} A_{\sigma, 3},\end{aligned}$$ where we recall that $$\begin{aligned}
&A_{\sigma, 1} := {\mathbb{E} \left[ \mathcal{D}_{h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)}}^4 \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}, \\
&A_{\sigma, 2} := {\mathbb{E} \left[ \mathcal{D}_{h_{\sigma(1)}, h_{\sigma(2)}}^2 \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}, \\
&A_{\sigma, 3} := {\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}.\end{aligned}$$ We will show that for each of these terms, the sum over all the sub-intervals converges to zero as $n \rightarrow \infty$.
For the first term, from we have $$\begin{aligned}
{\mathbb{E} \left[ \mathcal{D}_{h_1, h_2, h_3, h_4, h_5, h_6}^6 \psi_{t^n_i} \psi_{t^n_j}\right]}
\leq C \prod_{i=1}^6 {\left\lVert\Phi(h_i)\right\rVert}_{q-var; [0, T]}.\end{aligned}$$ Looking at each of the three types of terms , and , we see that at least two of the $h_i$’s are $\mathds{1}_{\Delta_{u^*}}$, and another two of the $h_i$’s are $\mathds{1}_{\Delta_{v^*}}$. Thus we obtain $$\begin{aligned}
\sum_{i, j = 0}^{2^n - 1} \sum_{k,l = 1}^m \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} {\mathbb{E} \left[ \mathcal{D}_{h_1, h_2, h_3, h_4, h_5, h_6}^6 \psi_{t^n_i} \psi_{t^n_j}\right]}
& \leq C \sum_{i, j = 0}^{2^n - 1} \sum_{k,l = 1}^m \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} \frac{1}{2^{(n+k)\frac{2}{\rho}} } \frac{1}{2^{(n+l)\frac{2}{\rho}} } \\
&\leq C \sum_{i, j = 0}^{2^n - 1} \frac{1}{2^{2n \left( \frac{2}{\rho}\right)}} \sum_{k,l=1}^{\infty} \frac{1}{2^{k \left( \frac{2}{\rho} - 1 \right)}} \frac{1}{2^{l \left( \frac{2}{\rho} - 1 \right)}} \\
& \leq \frac{C}{2^{2n \left( \frac{2}{\rho} - 1\right)}} \rightarrow 0\end{aligned}$$ since $\rho < 2$.
For the $A_{\sigma, 1}$ terms, we have two cases:
(i) $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} = R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$:
In all three types of terms , and , at least one of $\left\{ h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)} \right\}$ equals $\mathds{1}_{\Delta_{u^*}}$, and another one in the set equals $\mathds{1}_{\Delta_{v^*}}$. Applying the bounds in , we get $$\begin{aligned}
\sum_{i, j = 0}^{2^n - 1} \sum_{k,l = 1}^m \sum_{u = 0}^{2^k - 1} \sum_{v=0}^{2^l-1} &{\mathbb{E} \left[ \mathcal{D}_{h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)}}^4 \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}, \\
&\leq C \sum_{i, j = 0}^{2^n - 1} \sum_{k,l = 1}^m \sum_{u = 0}^{2^k - 1} \sum_{v=0}^{2^l-1} \prod_{r=1}^4 {\left\lVert\Phi(h_{\sigma(r)})\right\rVert}_{q-var; [0, T]} {\left\lvert\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert} \\
&\leq C \sum_{i, j = 0}^{2^n - 1} \sum_{k,l = 1}^m \sum_{u = 0}^{2^k - 1} \sum_{v=0}^{2^l-1}
2^{\frac{-(n+k)}{\rho}} 2^{\frac{-(n+l)}{\rho}}
{\left\lvert R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert} \\
&\leq C \sum_{k,l = 1}^m \left( \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}^{\rho} \right)^{\frac{1}{\rho}}
\left( \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} 2^{-(n+k)\left(\frac{\rho'}{\rho}\right)} 2^{-(n+l)\left(\frac{\rho'}{\rho}\right)} \right)^{\frac{1}{\rho'}} \\
&\leq C \, 2^{-2n \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} {\left\lVertR\right\rVert}_{\rho-var;[0, T]^2} \sum_{k,l=1}^m 2^{-k \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} 2^{-l \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} \\
&\leq C \, 2^{-2n \left( \frac{2}{\rho} - 1 \right)} {\left\lVertR\right\rVert}_{\rho-var;[0, T]^2} \sum_{k,l=1}^{\infty} 2^{-k \left( \frac{2}{\rho} - 1 \right)} 2^{-l \left( \frac{2}{\rho} - 1 \right)} \rightarrow 0\end{aligned}$$
(ii) $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \neq R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$:
We will go through each of the three types of terms , and to count the number of quantities with increments $\Delta_{u^*}$ or $\Delta_{v^*}$, which yield the factors $2^{\frac{-(n+k)}{\rho}}$ and $2^{\frac{-(n+l)}{\rho}}$ respectively.
(a) $M_u^{(a, b, c)} M_v^{(a, b, c)}$ terms:
We have five possibilities: $$\begin{aligned}
\label{56choices}
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{j^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{i^*} \\
\Delta_{j^*}
\end{pmatrix};\end{aligned}$$ we need not consider the cases $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$ or $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}$ since $y_1 \neq y_2$ and $y_4 \neq y_5$ in .
If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to either of the first two quantities on the right of , then one of $\left\{ h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)} \right\}$ must be equal to $\mathds{1}_{\Delta_{u^*}}$ and another two in the set must be equal to $\mathds{1}_{\Delta_{v^*}}$. If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to the third or the fourth quantity in , we have the same count with $u$ and $v$ switched.
If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} = R \begin{pmatrix}
\Delta_{i^*} \\
\Delta_{j^*}
\end{pmatrix}$, then without loss of generality, $$\begin{aligned}
h_{\sigma(1)} \: \mathrm{and} \: h_{\sigma(2)} = \mathds{1}_{\Delta_{u^*}}, \quad
h_{\sigma(3)} \: \mathrm{and} \: h_{\sigma(4)} = \mathds{1}_{\Delta_{v^*}}.\end{aligned}$$
(b) $N_u^{(a, b, c)} N_v^{(a, b, c)}$ terms:
If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$, then one of $\left\{ h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)} \right\}$ must equal $\mathds{1}_{\Delta_{u^*}}$ and another two in the set must equal $\mathds{1}_{\Delta_{v^*}}$. By switching $u$ and $v$, we can resolve the only other case $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}$ similarly.
(c) $M_u^{(a, b, c)} N_v^{(a, b, c)}$ terms:
There are only three possibilities $$\begin{aligned}
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix},\end{aligned}$$ and we need not consider the case $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$ since $y_1 \neq y_2$ in . If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix}$, then one of $\left\{ h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)} \right\}$ must be equal to $\mathds{1}_{\Delta_{u^*}}$ and another two in the set must be equal to $\mathds{1}_{\Delta_{v^*}}$. If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to the second or third quantity, the same count applies with $u$ and $v$ switched.
Thus in each case, applying the bounds in yields $$\begin{aligned}
\label{A1Bound}
\prod_{r=1}^4 {\left\lVert\Phi(h_{\sigma(r)})\right\rVert}_{q-var; [0, T]} {\left\lvert\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}
\leq C \, 2^{\frac{-2(n+k)}{\rho}} 2^{\frac{-2(n+l)}{\rho}},\end{aligned}$$ which gives us $$\begin{aligned}
&\sum_{i, j= 0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2l-1} {\mathbb{E} \left[ \mathcal{D}^4_{h_{\sigma(1)}, h_{\sigma(2)}, h_{\sigma(3)}, h_{\sigma(4)}} \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&{\qquad \qquad \qquad \qquad}\leq C \sum_{i, j= 0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2l-1} 2^{-(n+k) \frac{2}{\rho}} 2^{-(n+l) \frac{2}{\rho}} \\
&{\qquad \qquad \qquad \qquad}\leq C \, 2^{-2n \left( \frac{2}{\rho} -1 \right)} \sum_{k,l=1}^{\infty} 2^{-k \left(\frac{2}{\rho} - 1 \right)} 2^{-l \left(\frac{2}{\rho} - 1 \right)} \rightarrow 0.\end{aligned}$$
For the $A_{\sigma, 2}$ terms, when we consider $\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}$ and $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$, we have three cases: either both, one, or none of them are equal to $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$.
(i) $\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \: \mathrm{and} \: \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} = R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$:
(Note that this does not imply that they are equal to one another since $\Delta_{u^*}$ and $\Delta_{v^*}$ can be one of several intervals.)
Observe that $$\begin{aligned}
{\left\lvert\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}
\leq {\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{\frac{\rho}{2}} {\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{1 - \frac{\rho}{2}}
\leq C \, R^{\frac{\rho}{2}}_{\Delta_u^i \times \Delta_v^j} \, 2^{\frac{-(n+k)}{\rho} \left( 1 - \frac{\rho}{2} \right) },\end{aligned}$$ and $$\begin{aligned}
{\left\lvert\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}
\leq {\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{\frac{\rho}{2}} {\left\lvertR \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{1 - \frac{\rho}{2}}
\leq C \, R^{\frac{\rho}{2}}_{\Delta_u^i \times \Delta_v^j} \, 2^{\frac{-(n+l)}{\rho} \left( 1 - \frac{\rho}{2} \right) }.\end{aligned}$$ Thus we obtain $$\begin{aligned}
\sum_{i, j = 0}^{2^n - 1} \sum_{k,l=1}^m \sum_{u = 0}^{2^k - 1} \sum_{v=0}^{2^l-1} &{\mathbb{E} \left[ \mathcal{D}_{h_{\sigma(1)}, h_{\sigma(2)}}^2 \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\leq C \sum_{k,l=1}^m 2^{\frac{-(n+k)}{\rho} \left( 1 - \frac{\rho}{2} \right) } 2^{\frac{-(n+l)}{\rho} \left( 1 - \frac{\rho}{2} \right) } \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} R^{\rho}_{\Delta_u^i \times \Delta_v^j} \\
&\leq C \, 2^{-2n \left( \frac{1}{\rho} - \frac{1}{2} \right)} \sum_{k,l=1}^{\infty} 2^{-k \left( \frac{1}{\rho} - \frac{1}{2} \right)} 2^{-l \left( \frac{1}{\rho} - \frac{1}{2}\right)} {\left\lVertR\right\rVert}^{\rho}_{\rho-var;[0,T]^2} \rightarrow 0,\end{aligned}$$ since $\frac{1}{\rho} - \frac{1}{2} > 0$.
(ii) WLOG, assume $\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} = R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}, \quad
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \neq R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$:
As before, we will use the bounds in to show that $$\begin{aligned}
\label{A2Bound}
{\left\lVert\Phi(h_{\sigma(1)})\right\rVert}_{q-var; [0, T]} {\left\lVert\Phi(h_{\sigma(2)})\right\rVert}_{q-var; [0, T]} {\left\lvert\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}
\leq C \, 2^{\frac{-(n+k)}{\rho}} 2^{\frac{-(n+l)}{\rho}}.\end{aligned}$$
(a) $M_u^{(a, b, c)} M_v^{(a, b, c)}$ terms:
Again we have five possibilities, $$\begin{aligned}
\label{56choices2}
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{j^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{i^*} \\
\Delta_{j^*}
\end{pmatrix},\end{aligned}$$ and we need not consider the cases $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$ or $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}$ since $y_1 \neq y_2$ and $y_4 \neq y_5$ in .
If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to either of the first two quantities on the right of , then either $h_{\sigma(1)} $ or $h_{\sigma(2)} $ is equal to $\mathds{1}_{\Delta_{v^*}}$. If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to the third or fourth quantity, then either $h_{\sigma(1)} $ or $h_{\sigma(2)} $ is equal to $\mathds{1}_{\Delta_{u^*}}$.
If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} = R \begin{pmatrix}
\Delta_{i^*} \\
\Delta_{j^*}
\end{pmatrix}$, then we must have $h_{\sigma(1)} = \mathds{1}_{\Delta_{u^*}}$ and $h_{\sigma(2)} = \mathds{1}_{\Delta_{v^*}}$, or vice versa.
(b) $N_u^{(a, b, c)} N_v^{(a, b, c)}$ terms:
If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$ (resp. $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}$), then both $h_{\sigma(1)} $ and $h_{\sigma(2)} $ must be equal to $\mathds{1}_{\Delta_{v^*}}$ (resp. $\mathds{1}_{\Delta_{u^*}}$).
(c) $M_u^{(a, b, c)} N_v^{(a, b, c)}$ terms:
There are only three possibilities, $$\begin{aligned}
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix},
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix},\end{aligned}$$ and we need not consider the case $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}$ since $y_1 \neq y_2$ in . If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix}$, then both $h_{\sigma(1)}$ and $h_{\sigma(2)}$ are equal to $\mathds{1}_{\Delta_{v^*}}$. If $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$ is equal to $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}$ or $R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix}$, then either $h_{\sigma(1)}$ or $h_{\sigma(2)}$ is equal to $\mathds{1}_{\Delta_{u^*}}$.
Thus we obtain $$\begin{aligned}
\sum_{i, j= 0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} &{\mathbb{E} \left[ \mathcal{D}^2_{h_{\sigma(1)}, h_{\sigma(2)}} \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\leq \sum_{k,l = 1}^m \left( \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} {\left\lvert R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{\rho} \right)^{\frac{1}{\rho}}
\left( \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} 2^{-(n+k)\left(\frac{\rho'}{\rho}\right)} 2^{-(n+l)\left(\frac{\rho'}{\rho}\right)} \right)^{\frac{1}{\rho'}} \\
&\leq {\left\lVertR\right\rVert}_{\rho-var; [0,T]^2} \sum_{k,l=1}^m 2^{-2n \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} 2^{-k \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} 2^{-l \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} \\
&\leq C {\left\lVertR\right\rVert}_{\rho-var; [0,T]^2} 2^{-2n \left( \frac{2}{\rho} - 1 \right)} \rightarrow 0.\end{aligned}$$
(iii) $\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \: \mathrm{and} \: \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
\neq R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$:
We will show that $$\begin{aligned}
\label{A22Bound}
{\left\lVert\Phi \left( h_{\sigma(1)} \right)\right\rVert}_{q-var; [0, T]} {\left\lVert\Phi \left( h_{\sigma(2)} \right)\right\rVert}_{q-var; [0, T]} {\left\lvert \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}\right\rvert} {\left\lvert \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}\end{aligned}$$ is bounded above by $C \, 2^{\frac{-2(n+k)}{\rho}} 2^{\frac{-2(n+l)}{\rho}}$, which gives us $$\begin{aligned}
\sum_{i, j= 0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} &{\mathbb{E} \left[ \mathcal{D}^2_{h_{\sigma(1)}, h_{\sigma(2)}} \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\leq C \sum_{i, j= 0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} 2^{-(n+k) \frac{2}{\rho}} 2^{-(n+l) \frac{2}{\rho}} \\
&\leq C \, 2^{-2n \left( \frac{2}{\rho} -1 \right)} \sum_{k,l=1}^m 2^{-k \left(\frac{2}{\rho} - 1 \right)} 2^{-l \left(\frac{2}{\rho} - 1 \right)} \rightarrow 0.\end{aligned}$$
(a) $M_u^{(a, b, c)} M_v^{(a, b, c)}$ terms:
Note that in this scenario, neither $h_{\sigma(1)}$ nor $h_{\sigma(2)}$ can be equal to $\mathds{1}_{\Delta_{i^*}}$ or $\mathds{1}_{\Delta_{j^*}}$, so we essentially have two cases.
If $h_{\sigma(1)} \: \mathrm{and} \: h_{\sigma(2)} = \mathds{1}_{\Delta_{u^*}}$, we must have $$\begin{aligned}
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix} \quad \mathrm{and} \quad \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix},\end{aligned}$$ or vice versa. (The case $h_{\sigma(1)} \: \mathrm{and} \: h_{\sigma(2)} = \mathds{1}_{\Delta_{v^*}}$ can be resolved similarly by swapping $u$ and $v$.)
If instead we have $h_{\sigma(1)} = \mathds{1}_{\Delta_{u^*}}$ and $h_{\sigma(2)} = \mathds{1}_{\Delta_{v^*}}$, or vice versa, then without loss of generality, it must be the case that $$\begin{aligned}
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{j^*}
\end{pmatrix},
\quad \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix}.\end{aligned}$$
(b) $N_u^{(a, b, c)} N_v^{(a, b, c)}$ terms:
Without loss of generality, we have $$\begin{aligned}
h_{\sigma(1)} = \mathds{1}_{\Delta_{u^*}}, \:\:
h_{\sigma(2)} = \mathds{1}_{\Delta_{v^*}}, \:\:
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}, \:\:
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}.\end{aligned}$$
(c) $M_u^{(a, b, c)} N_v^{(a, b, c)}$ terms:
Without loss of generality, either $$\begin{aligned}
h_{\sigma(1)}, \, h_{\sigma(2)} = \mathds{1}_{\Delta_{u^*}}, \quad
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix} \:\: \mathrm{and} \:\: \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix},\end{aligned}$$ or $$\begin{aligned}
&h_{\sigma(1)} = \mathds{1}_{\Delta_{u^*}}, \: h_{\sigma(2)} = \mathds{1}_{\Delta_{v^*}}, \\
&\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix} \:\: \mathrm{and} \:\: \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}.\end{aligned}$$
For the $A_{\sigma, 3}$ terms, when we consider the three inner products $\left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d}$, $\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}$ and $\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}$, we have two cases: either one of them is equal to $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$, or two or more of them are. Observe that it is not possible for none of them to equal $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$.
(i) If two or more of the inner products are equal to $R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$, then we can use the same computation in the first case for the $A_{\sigma, 2}$ terms to show that $$\begin{aligned}
{\left\lvert \left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d}\right\rvert} {\left\lvert \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \right\rvert} {\left\lvert \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}
\leq C \, R^{\rho}_{\Delta_u^i \times \Delta_v^j} \, 2^{\frac{-(n+k)}{\rho} \left( 1 - \frac{\rho}{2} \right) } 2^{\frac{-(n+l)}{\rho} \left( 1 - \frac{\rho}{2} \right) },\end{aligned}$$ and this gives us $$\begin{aligned}
\sum_{i, j = 0}^{2^n - 1} \sum_{k,l=1}^m \sum_{u = 0}^{2^k - 1} \sum_{v=0}^{2^l-1} &{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\leq C \, 2^{-2n \left( \frac{1}{\rho} - \frac{1}{2} \right)} \sum_{k,l=1}^{\infty} 2^{-k \left( \frac{1}{\rho} - \frac{1}{2} \right)} 2^{-l \left( \frac{1}{\rho} - \frac{1}{2}\right)} {\left\lVertR\right\rVert}^{\rho}_{\rho-var;[0,T]^2} \rightarrow 0.\end{aligned}$$
(ii) Assume that $\left\langle h_{\sigma(1)}, h_{\sigma(2)} \right\rangle_{\mathcal{H}_1^d} = R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}, \: \mathrm{and} \:
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}, \, \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
\neq R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}$.
Then without loss of generality, we have:
(a) $M_u^{(a, b, c)} M_v^{(a, b, c)}$ terms: $$\begin{aligned}
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{j^*}
\end{pmatrix}, \quad
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}^1_d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{i^*}
\end{pmatrix} \: \mathrm{or} \:
R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{j^*}
\end{pmatrix}.\end{aligned}$$
(b) $N_u^{(a, b, c)} N_v^{(a, b, c)}$ terms: $$\begin{aligned}
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{u^*}
\end{pmatrix}, \quad
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}.\end{aligned}$$
(c) $M_u^{(a, b, c)} N_v^{(a, b, c)}$ terms: $$\begin{aligned}
\left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{i^*}
\end{pmatrix}, \quad
\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}
= R \begin{pmatrix}
\Delta_{v^*} \\
\Delta_{v^*}
\end{pmatrix}.\end{aligned}$$
In each case, applying the bounds in gives us $$\begin{aligned}
{\left\lvert \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d}\right\rvert} {\left\lvert\left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d}\right\rvert}
\leq C \, 2^{\frac{-(n+k)}{\rho}} 2^{\frac{-(n+l)}{\rho}},\end{aligned}$$ which in turn yields $$\begin{aligned}
\sum_{i, j= 0}^{2^n-1} \sum_{k,l=1}^m \sum_{u=0}^{2^k-1} \sum_{v=0}^{2^l-1} &{\mathbb{E} \left[ \psi_{t^n_i} \psi_{t^n_j}\right]} \left\langle h_{\sigma(1)}, h_{\sigma(2)}\right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(3)}, h_{\sigma(4)} \right\rangle_{\mathcal{H}_1^d} \left\langle h_{\sigma(5)}, h_{\sigma(6)} \right\rangle_{\mathcal{H}_1^d} \\
&\leq \sum_{k,l = 1}^m \left( \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} {\left\lvert R \begin{pmatrix}
\Delta_{u^*} \\
\Delta_{v^*}
\end{pmatrix}\right\rvert}^{\rho} \right)^{\frac{1}{\rho}}
\left( \sum_{i, j = 0}^{2^n - 1} \sum_{u = 0}^{2^k - 1} \sum_{v = 0}^{2^l-1} 2^{-(n+k)\left(\frac{\rho'}{\rho}\right)} 2^{-(n+l)\left(\frac{\rho'}{\rho}\right)} \right)^{\frac{1}{\rho'}} \\
&\leq {\left\lVertR\right\rVert}_{\rho-var; [0,T]^2} \sum_{k,l=1}^m 2^{-2n \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} 2^{-k \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} 2^{-l \left( \frac{1}{\rho} - \frac{1}{\rho'} \right)} \\
&\leq C {\left\lVertR\right\rVert}_{\rho-var; [0,T]^2} 2^{-2n \left( \frac{2}{\rho} - 1 \right)} \rightarrow 0.\end{aligned}$$
\[skorodhodLimitEnhanced\] For $2 \leq p < 4$, let $Y \in \mathcal{C} ^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^d; \mathbb{R}^m) \right)$ denote the path-level solution to $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ where $\mathbf{X} \in \mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left(\mathbb{R}^d \right)\right)$ satisfies Condition \[newCond1\] and its covariance function satisfies $$\begin{aligned}
\left\| R(t, \cdot) - R (s, \cdot) \right\|_{q-var; [0, T]}
\leq C \left| t - s\right|^{\frac{1}{\rho}}, \quad \forall s, t \in [0, T].\end{aligned}$$ Then if $V \in \mathcal{C}^{\lfloor p \rfloor + 4}_b \left( \mathbb{R}^{md}; \mathbb{R}^{md} \otimes \mathbb{R}^d \right)$, we have $$\begin{aligned}
\label{enhancedSkoEq}
\lim_{\left\| \pi(n)\right\| \rightarrow 0} {\left\lVert\sum_i V(Y_{t^n_i}) \left( \mathbf{X}^2_{t^n_i, t^n_{i+1}} - \frac{1}{2} \sigma^2 \left( t^n_i, t^n_{i+1} \right) \mathcal{I}_d \right)\right\rVert}_{L^2(\Omega)} = 0.\end{aligned}$$ Furthermore, if $3 \leq p < 4$ and $V \in \mathcal{C}^9_b \left( \mathbb{R}^{md}; \mathbb{R}^{md} \otimes \mathbb{R}^d \right)$, we have $$\begin{aligned}
\label{enhancedSkoEq2}
\lim_{\left\| \pi(n)\right\| \rightarrow 0} {\left\lVert\sum_i \nabla V(Y_{t^n_i}) \left( V(Y_{t^n_i}) \right) \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right)\right\rVert}_{L^2(\Omega)} = 0.\end{aligned}$$
We have to show that bounds and in Proposition \[2ndlevel\] are satisfied with $$\begin{aligned}
\psi_t = \left[ V(Y_t) \right]_j \in \mathbb{R}^d \otimes \mathbb{R}^d, \quad j = 1, \ldots, m,\end{aligned}$$ to show . Similarly, proving that bound in Proposition \[3rdlevel\] is satisfied with $$\begin{aligned}
\psi_t = \left[ \nabla V(Y_t) \left( V(Y_t) \right) \right]_j \in \mathbb{R}^d \otimes \mathbb{R}^d \otimes \mathbb{R}^d, \quad j = 1, \ldots, m,\end{aligned}$$ will yield .
is trivially true since $V \in \mathcal{C}^1_b$. To show that the bounds hold for the higher Malliavin derivatives, recall Proposition \[dir der est\], which states that almost surely we have $$\begin{aligned}
{\left\lVert\mathcal{D}_{h_1, \ldots, h_n}^n Y_{\cdot}\right\rVert}_{\infty}
\leq P_{d(n)} \left( {\left\lVert\mathbf{X}\right\rVert}_{p-var; [0, T]}, \exp \left( C \, N^{\mathbf{X}}_{1; [0, T]} \right)\right) \prod\limits_{i=1}^{n} \left\| \Phi(h_{i})\right\|_{q-var; [0, T]}.\end{aligned}$$ As both ${\left\lVert\mathbf{X}\right\rVert}_{p-var; [0, T]}$ and $\exp \left( C \, N^{\mathbf{X}}_{1; [0, T]} \right)$ belong to $\bigcap_{r>0}L^{r}\left( \Omega \right)$, we have $$\begin{aligned}
\label{bound}
{\left\lVert\mathcal{D}_{h_1, \ldots, h_n}^n Y_t\right\rVert}_{L^{r}\left( \Omega \right) }
\leq C_{n, q} \prod \limits_{i=1}^{n} {\left\lVert\Phi(h_i)\right\rVert}_{q-var; \left[ 0,T\right] }\end{aligned}$$ for any $r > 0$. Now we simply use the product and chain rule of Malliavin differentiation in conjunction with the fact that $V$ has bounded derivatives up to the appropriate order.
Correction formula {#corr formula}
==================
We are now ready to prove the main result of the paper. As before, $\pi(n) := \left\{ t^{n}_{i} \right\} , t^{n}_{i} := \frac{iT}{2^{n}}$, denotes the sequence of dyadic partitions on $[0, T]$.
Main theorem
------------
\[mainThm3rdLevel\] For $3 \leq p < 4$, let $Y \in \mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^d; \mathbb{R}^m) \right)$ denote the path-level solution to $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ where $V \in \mathcal{C}^9_b \left( \mathbb{R}^{md}; \mathbb{R}^{md} \otimes \mathbb{R}^d \right)$, and $\mathbf{X} \in \mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left(\mathbb{R}^d \right)\right)$ is a Volterra process which satisfies Condition \[newCond1\], and whose kernel satisfies Condition \[amnCond\] with $\alpha < \frac{1}{p}$. Furthermore, we assume the covariance function satisfies $$\begin{aligned}
\left\| R(t, \cdot) - R (s, \cdot) \right\|_{q-var; [0, T]} \leq C \left| t - s\right|^{\frac{1}{\rho}},\end{aligned}$$ for all $s, t \in [0, T]$, and ${\left\lVertR(\cdot)\right\rVert}_{q-var; [0, T]} < \infty$. Then almost surely, we have $$\begin{aligned}
\label{cflr}
\int_0^T Y_t \circ \mathrm{d} \mathbf{X}_t
&= \int_0^T Y_t \, \mathrm{d} X_t + \sum_{j=1}^m \left( \frac{1}{2} \int_0^T \mathrm{tr} \left[ V(Y_s) \right]_j \, \mathrm{d} R(s) + U^{(j)}_T \right) e_j,\end{aligned}$$ where for $j = 1, \ldots, m$, $U^{(j)}_T$ is the limit in $L^2(\Omega)$ of $$\begin{aligned}
\label{2ndTermApprox}
\sum_{i} \int_0^{t^n_i} \mathrm{tr} \left[J^{\mathbf{X}}_{t^n_i \leftarrow s} V(Y_s) - V \left( Y_{t^n_i} \right) \right]_j R(\Delta^n_i, \mathrm{d}s) \end{aligned}$$ along the dyadic partitions $\{ t^n_i\}$ of $[0, T]$.
Using bounds , together with the integrability of $\mathbf{X}$, we can apply dominated convergence theorem to in Theorem \[controlledThm2\] to show that $\int_0^T Y_t {\circ \mathrm{d} \textbf{X}_t} $ is the $L^2(\Omega)$ limit of $$\begin{aligned}
\lim_{n \rightarrow \infty} \sum_i
\, Y_{t^n_i} \left( X_{t^n_i, t^n_{i+1}} \right) + V(Y_{t^n_i}) \left( \mathbf{X}^2_{t^n_i, t^n_{i+1}} \right) + \nabla V \left( Y_{t^n_i} \right) \left( V \left( Y_{t^n_i} \right) \right) \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right).\end{aligned}$$ Now applying Proposition \[skorohodLimit3rdLevel\] in conjunction with Corollary \[skorodhodLimitEnhanced\] gives us $$\begin{aligned}
\int_0^T Y_t {\, \mathrm{d} X_t}
&= \lim_{n \rightarrow \infty} \sum_i \left[ Y_{t^n_i} \left( X_{t^n_i, t^n_{i+1}} \right) - \sum_{j=1}^m \left( \int_0^{t^n_i} \mathrm{tr} \left[J^{\mathbf{X}}_{t^n_i \leftarrow s} V(Y_s)\right]_j R\left( \Delta^n_i, \mathrm{d} s \right) \right) e_j + A_i \right], \end{aligned}$$ where the limit is also in $L^2(\Omega)$ and $$\begin{aligned}
A_i := V(Y_{t^n_i}) \left( \left( \mathbf{X}^2_{t^n_i, t^n_{i+1}} \right) - \frac{1}{2} \sigma^2 \left( t^n_i, t^n_{i+1} \right) \mathcal{I}_d \right) + \nabla V \left( Y_{t^n_i} \right) \left( V \left( Y_{t^n_i} \right) \right) \left( \mathbf{X}^3_{t^n_i, t^n_{i+1}} \right).\end{aligned}$$ Following the procedure in Theorem 6.1 of [@cl2018], subtracting the two integrals and re-balancing the terms gives us $$\begin{aligned}
\label{diff}
\begin{split}
\int_0^T Y_t {\circ \mathrm{d} \textbf{X}_t} \, - &\int_0^T Y_t {\, \mathrm{d} X_t} \\
&= \sum_{j=1}^m \left( \lim_{n \rightarrow \infty} \sum_i \int_0^{t^n_i} \mathrm{tr} \left[ J^{\mathbf{X}}_{t^n_i \leftarrow s} V(Y_s) \right]_j R(\Delta^n_i, {\, \mathrm{d} s} ) + \frac{1}{2} \sigma^2 \left( t^n_i, t^n_{i+1} \right) \mathrm{tr} \left[ V(Y_{t^n_i}) \right]_j \right) e_j, \\
&= \sum_{j=1}^m \left( \lim_{n \rightarrow \infty} \sum_i \int_0^{t^n_i} \mathrm{tr} \left[ J^{\mathbf{X}}_{t^n_i \leftarrow s} V(Y_s) - V\left(Y_{t^n_i} \right) \right]_j R(\Delta^n_i, {\, \mathrm{d} s} ) \right. \\
&\left. {\qquad \qquad \qquad \qquad}+ \frac{1}{2} \lim_{n \rightarrow \infty} \sum_i \mathrm{tr} \left[ V(Y_{t^n_i}) \right]_j \big( R\left( t^n_{i+1}, t^n_{i+1} \right) - R \left(t^n_i, t^n_i \right) \big) \right) e_j.
\end{split}\end{aligned}$$ The second term in the last line of the expression above is dominated by $$\begin{aligned}
C {\left\lVertV(Y_{\cdot})\right\rVert}_{p-var; [0, T]} {\left\lVertR(\cdot)\right\rVert}_{q-var; [0, T]}\end{aligned}$$ by Young’s inequality, and thus converges in $L^2 (\Omega)$ to $$\begin{aligned}
\frac{1}{2} \int_0^T \mathrm{tr} \left[ V(Y_s) \right]_j \, \mathrm{d} R(s).\end{aligned}$$ This in turn guarantees the convergence of the first term in $L^2(\Omega)$ to the random variable $U_T^{(j)}$. Now extracting an almost sure subsequence allows us to equate both sides of almost surely, and the proof is thus complete.
In the more regular case $2 \leq p < 3$, we can be more precise in identifying the second term .
\[volterra2ndTerm\] For $2 \leq p < 3$, let $Y \in \mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^d; \mathbb{R}^m) \right)$ denote the path-level solution to $\mathbf{X}$ $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ where $\mathbf{X} \in \mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left(\mathbb{R}^d \right)\right)$ is a Volterra process satisfying Condition \[newCond1\] with some $\rho \in \left[ 1, \frac{3}{2} \right)$, and whose kernel satisfies Condition \[amnCond\] with $\alpha < \frac{1}{2p}$.
Furthermore, assume that $V \in \mathcal{C}^6_b \left( \mathbb{R}^{md}; \mathbb{R}^{md} \otimes \mathbb{R}^d \right)$, and the covariance function satisfies $$\begin{aligned}
\left\| R(t, \cdot) - R (s, \cdot) \right\|_{\rho-var; [0, T]}
\leq C \left| t - s\right|^{\frac{1}{\rho}},\end{aligned}$$ for all $s, t \in [0, T]$. Then almost surely we have $$\begin{aligned}
\int_0^T Y_t \circ \mathrm{d} \mathbf{X}_t
&= \int_0^T Y_t \, \mathrm{d} X_t + Z_T,\end{aligned}$$ where the correction term is given by $$\begin{aligned}
\label{Zdef}
\begin{split}
Z^{(j)}_T
&= \frac{1}{2} \int_0^T \mathrm{tr} \left[ V(Y_s) \right]_j \, \mathrm{d} R(s) + \int_{[0, T]^2} h_j(s, t) \, \mathrm{d} R(s,t) \\
&= \frac{1}{2} \int_0^T \mathrm{tr} \left[ V(Y_s) \right]_j \, \mathrm{d} R(s) + \int_0^T \mathcal{K}^* \otimes \mathcal{K}^* h_j (r, r) {\, \mathrm{d} r}, \quad j = 1, \ldots, m.
\end{split}\end{aligned}$$ with $$\begin{aligned}
h_j (s, t) := \mathds{1}_{[0, t)} (s) \, \mathrm{tr} \left[J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) - V(Y_t) \right]_j.\end{aligned}$$
Under the conditions of the theorem, we can invoke Theorem 6.1 from [@cl2018] to obtain the first line of . To obtain the second line, we will use Proposition 4.3 from [@lim2018], which states that if $\phi:[0,T]^{2} \rightarrow \mathbb{R}$ is a $\lambda$-Hölder bi-continuous function (one that satisfies Definition \[biHolderDef\] without necessarily satisfying ) with $\lambda > 2 \alpha$, then $$\begin{aligned}
\label{volEqDR}
\int_{[0, T]^{2}}\phi(s, t)\,\mathrm{d} R(s, t) = \int_{0}^{T}\mathcal{K}^* \otimes\mathcal{K}^* \phi(r,r)\,\mathrm{d}r.\end{aligned}$$ Thus, the proof is complete once we show that $h_j (s, t)$ is $\frac{1}{p}$-Hölder bi-continuous for all $j = 1, \ldots, m$ since $\frac{1}{p} > 2\alpha$. Using the fact that $$\begin{aligned}
\tilde{h}_j (s, t) := \mathrm{tr} \left[J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) - V(Y_t) \right]_j\end{aligned}$$ is $\frac{1}{p}$-Hölder bi-continuous, we have, assuming $v_2 > v_1$ without loss of generality, $$\begin{aligned}
{\left\lverth_j (u, v_2) - h_j (u, v_1)\right\rvert}
&= {\left\lvert\tilde{h}_j (u, v_2 \vee u) - \tilde{h}_j (u, v_1 \vee u)\right\rvert}, \quad u, v_1, v_2 \in [0, T], \\
&\leq C_1 {\left\lvertv_2 \vee u - v_1 \vee u\right\rvert}^{\frac{1}{p}} \\
&\leq C_2 {\left\lvertv_2 - v_1\right\rvert}^{\frac{1}{p}},\end{aligned}$$ and similarly, $$\begin{aligned}
{\left\lverth_j (u_2, v) - h_j (u_1, v)\right\rvert} \leq C {\left\lvertu_2 - u_1\right\rvert}^{\frac{1}{p}}, \quad v, u_1, u_2 \in [0, T].\end{aligned}$$
In the case $3 \leq p < 4$, due to the lack of complementary regularity, we cannot apply Theorem \[2Dintegral\] although $\mathds{1}_{[0, t)} (s) \mathrm{tr} \left[J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) - V \left( Y_t \right) \right]_j$ is continuous almost surely on $[0, T]^2$. Furthermore, although the integrand is strongly $\frac{1}{p}$-Hölder bi-continuous away from the diagonal, one can check that in general, fails at the diagonal, which means that we cannot employ from Proposition 4.3 in [@lim2018] (it can also be verified that there would be insufficient Hölder regularity in the weaker sense). Hence, we can only show convergence in $L^2(\Omega)$ rather than almost surely. The question of whether the second part of the correction term can be identified as a proper 2D Young-Stieltjes integral requires further investigation.
An interesting special case of Theorem is when the vector fields defining the RDE commute. In this situation the $U_{T}$ terms in the correction formula disappear.
\[thmSimple\] Under the conditions of Theorem \[mainThm3rdLevel\], if in addition the vector fields commute, i.e. $[V_{i},V_{j}]=0$ for all $i,j=1,\ldots,d$, then $$\begin{aligned}
\int_{0}^{T}Y_{t}\circ\mathrm{d}\mathbf{X}_{t}=\int_{0}^{T}Y_{t} \, \mathrm{d}X_{t}+\frac{1}{2}\sum_{j=1}^{m}\left( \int_{0}^{T}\mathrm{tr} \left[ V(Y_{s})\right]_{j}\,\mathrm{d}R(s)\right) e_{j},\end{aligned}$$
For any vector field $W \in \mathcal{C}^1 \left(\mathbb{R}^{md}; \mathbb{R}^{md} \right)$, we have $$\begin{aligned}
\left( J^{\mathbf{X}}_t\right)^{-1} W(Y_t) = W(y_0) + \sum_{i=1}^d \int_0^t \left( J^{\mathbf{X}}_s \right)^{-1} [V_i, W] {\circ \mathrm{d} \mathbf{X}^{(i)}_s},\end{aligned}$$ which can be computed using the RDEs satisfied by $Y$ and $\left( J^{\mathbf{X}} \right)^{-1}$, cf. Chapter 20 (Section 4.2) in [@fv2010b]. Hence, if the $V_i$’s commute, then each $V_i$ is invariant under the flow of $Y$, and we have $$\begin{aligned}
J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) = V(Y_t), \quad 0 \leq s < t \leq T.\end{aligned}$$
Applications of the correction formula
--------------------------------------
We present applications of the main theorem to two important special cases. The first is to fractional Brownian motion in the regime $H > \frac{1}{4}$. Given the interest in this in recent years, especially in volatility models in mathematical finance, this result may also find practical uses among the wider areas of applied probability. The second application is to use the commuting case discussed in Corollary \[thmSimple\] to obtain Itô formulas for Gaussian processes. This links our correction formula to the prolific stream of recent work mentioned in the introduction.
\[Correction formula fBM, $H > \frac{1}{4}$\] For $1 \leq p < 4$, let $Y \in \mathcal{C}^{p-var} \left( [0, T]; \mathcal{L} (\mathbb{R}^d; \mathbb{R}^m) \right)$ denote the path-level solution to $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = y_0,\end{aligned}$$ where we assume that $V \in \mathcal{C}^k_b \left( \mathbb{R}^{md}; \mathbb{R}^{md} \otimes \mathbb{R}^d \right)$, with $$\begin{aligned}
\label{kDefn}
k = \begin{cases}
\, 2, \quad 1 \leq p < 2, \\
\, 6, \quad 2 \leq p < 3, \\
\, 9, \quad 3 \leq p < 4,
\end{cases}\end{aligned}$$ and $\mathbf{X} \in \mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left(\mathbb{R}^d \right)\right)$ is the geometric rough path constructed from the limit of the piecewise-linear approximations of standard fractional Brownian motion with Hurst parameter $H > \frac{1}{4}$. Then almost surely, we have $$\begin{aligned}
\int_0^T Y_t \circ \mathrm{d} \mathbf{X}_t
&= \int_0^T Y_t \, \mathrm{d} X_t + Z_T,\end{aligned}$$ where the correction term $Z_T = \left( Z^{(1)}_T, \ldots, Z^{(m)}_T \right)$ is given by $$\begin{aligned}
\label{cTerm}
\begin{split}
Z^{(j)}_T
&= H \int_0^T \mathrm{tr} \left[ V(Y_s) \right]_j \, s^{2H - 1} {\, \mathrm{d} s} + \int_{[0, T]^2} h_j(s,t) \, \mathrm{d} R(s, t), \quad j = 1, \ldots, m, \\
&= H \int_0^T \mathrm{tr} \left[ V(Y_s) \right]_j \, s^{2H - 1} {\, \mathrm{d} s} + \int_0^T \mathcal{K}^*\otimes \mathcal{K}^* h_j (r, r) {\, \mathrm{d} r}, \quad \left( \mathrm{when} \, \frac{1}{3} < H \leq \frac{1}{2} \right),
\end{split}\end{aligned}$$ with $$\begin{aligned}
h_j (s,t) := \mathds{1}_{[0, t)} (s) \mathrm{tr} \left[J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) - V(Y_t) \right]_j, \quad j = 1, \ldots, m.\end{aligned}$$
For simplicity, we use the same notation for the second term of $Z_{T}^{(j)}$ for all $H > \frac{1}{4}$, with the understanding that it denotes the $L^{2}(\Omega)$ limit of when $\frac{1}{4} < H \leq \frac{1}{3}$.
The proof rests entirely on Proposition \[fBMCond\], which tells us that fractional Brownian motion fulfills all the requirements needed to apply Theorem 6.1 of [@cl2018] when $H > \frac{1}{3}$, and Theorem \[mainThm3rdLevel\] when $\frac{1}{4} < H \leq \frac{1}{3}$.
We now show that we can recover Itô’s formulas.
\[Itô formulas for Gaussian processes\] For $1 \leq p < 4$, let $\mathbf{X} \in \mathcal{C}^{0, p-var} \left( [0, T]; G^{\lfloor p \rfloor} \left( \mathbb{R}^d \right)\right)$ satisfy Condition \[newCond1\]. Depending on $p$, we further impose the following conditions:
(i) $1 \leq p < 2$: $\sigma^2(s, t) \leq C {\left\lvertt -s\right\rvert}^{\theta}$ for some $\theta > 1$ and ${\left\lVertR(\cdot)\right\rVert}_{q-var; [0, T]} < \infty$.
(ii) $2 \leq p < 3$: The covariance function satisfies $$\begin{aligned}
\label{cond2}
\left\| R(t, \cdot) - R (s, \cdot) \right\|_{q-var; [0, T]}
\leq C \left| t - s\right|^{\frac{1}{\rho}},\end{aligned}$$ for all $s, t \in [0, T]$.
(iii) $3 \leq p < 4$: $\mathbf{X}$ is a Volterra process whose kernel satisfies Condition \[amnCond\] with $\alpha < \frac{1}{p}$. Furthermore, its covariance function satisfies and ${\left\lVertR(\cdot)\right\rVert}_{q-var; [0, T]} < \infty$.
Then almost surely, for $f \in \mathcal{C}^{k+2}_b \left( \mathbb{R}^d; \mathbb{R} \right)$, $k$ defined as in , we have $$\begin{aligned}
f(X_T) - f(0) = \int_0^T \left\langle \nabla f(X_t), \circ \mathrm{d} \mathbf{X}_t \right\rangle
&= \int_0^T \left\langle \nabla f(X_t), \, \mathrm{d} X_t \right\rangle + \frac{1}{2} \int_0^T \Delta f (X_t) \, \mathrm{d} R(t).\end{aligned}$$
Let $Y_t = \left( Y^{(1)}_t, \ldots, Y^{(2d)}_t \right)$ denote the augmented process $$\begin{aligned}
\left( {\frac{\partial f}{\partial e_1}} (Y_t), \ldots, {\frac{\partial f}{\partial e_d}} (Y_t), X^{(1)}_t, \ldots X^{(d)}_t \right),\end{aligned}$$ in which case $Y$ satisfies the RDE $$\begin{aligned}
\mathrm{d} Y_t = V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}, \quad Y_0 = \left( y_0, 0 \right),\end{aligned}$$ where $V(Y) \in \mathbb{R}^{2d} \otimes \mathbb{R}^d$ is represented by the $2d$-by-$d$ matrix $$\begin{aligned}
V(Y_t) = \begin{bmatrix}
\begin{array} {c}
\nabla^2 f (Y_t) \\
\hdashline
\mathcal{I}_d
\end{array}
\end{bmatrix}.\end{aligned}$$ Now one can check that $[V_i, V_j] = 0$ for all $i, j = 1, \ldots, d$, and apply Corollary \[thmSimple\].
Alternatively, since $\nabla V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}$ is the upper-triangular $2d$-by-$2d$ matrix $$\begin{aligned}
\nabla V(Y_t) {\circ \mathrm{d} \mathbf{X}_t}
= \begin{bmatrix}
\begin{array}{c:c}
0 & \nabla^3 f(X_t) {\circ \mathrm{d} \mathbf{X}_t} \\
\hdashline
\phantom{123456} 0 \phantom{123456} & 0
\end{array}
\end{bmatrix},\end{aligned}$$ where $$\begin{aligned}
\left( \nabla^3 f (X_t) {\circ \mathrm{d} \mathbf{X}_t} \right)_{ij} = \sum_{k=1}^d \frac{\partial^3 f}{\partial e_k \partial e_i \partial e_j} (X_t) {\circ \mathrm{d} \mathbf{X}^{(k)}_t}, \quad i, j = 1, \ldots, d,\end{aligned}$$ one can directly compute the solution to the RDE satisfied by the Jacobian as $$\begin{aligned}
J^{\mathbf{X}}_t = \begin{bmatrix}
\begin{array}{c:c}
\mathcal{I}_d & \nabla^2 f(X_t) - \nabla^2 f(0) \\
\hdashline
\phantom{1234567} 0 \phantom{1234567} & \mathcal{I}_d
\end{array}
\end{bmatrix}, \quad t \in [0, T].\end{aligned}$$ Now since the inverse is given by $$\begin{aligned}
\left( J^{\mathbf{X}}_t \right)^{-1} = \begin{bmatrix}
\begin{array}{c:c}
\mathcal{I}_d & \nabla^2 f(0) - \nabla^2 f (X_t) \\
\hdashline
\phantom{1234567} 0 \phantom{1234567} & \mathcal{I}_d
\end{array}
\end{bmatrix}, \quad t \in [0, T],\end{aligned}$$ we also obtain, for $0 \leq s < t \leq T$, $$\begin{aligned}
J^{\mathbf{X}}_{t \leftarrow s} V(Y_s) = V(Y_t).\end{aligned}$$
[99]{} <span style="font-variant:small-caps;">Alòs, E., Mazet, O.</span> and <span style="font-variant:small-caps;">Nualart, D.</span> (2000). Stochastic calculus with respect to fractional Brownian motion with Hurst parameter lesser than $\frac{1}{2}$. *Stochastic Processes and Their Applications* **86**(1), 121–139.
<span style="font-variant:small-caps;">Alòs, E., Mazet, O.</span> and <span style="font-variant:small-caps;">Nualart, D.</span> (2001). Stochastic calculus with respect to Gaussian processes. *The Annals of Probability* **29**(2), 766-–801.
<span style="font-variant:small-caps;">Bayer, C.</span>, <span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Gatheral, J.</span> (2016). Pricing under rough volatility. *Quantitative Finance* **16**(6), 887–904.
<span style="font-variant:small-caps;">Carmona, P., Coutin, L.</span> and <span style="font-variant:small-caps;">Montseny, G.</span> (2003). Stochastic integration with respect to fractional Brownian motion. *Annales de l’Institut Henri Poincaré – Probabilitiés et Statistiques* **39**(1), 27–-68.
<span style="font-variant:small-caps;">Cass, T.</span> and <span style="font-variant:small-caps;">Friz, P.</span> (2011). Malliavin calculus and rough paths. *Bulletin Des Sciences Mathematiques* **135**(6–7), 542–556.
<span style="font-variant:small-caps;">Cass, T.</span>, <span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Victoir, N.</span> (2009). Non-degeneracy of Wiener functionals arising from rough differential equations. *Transactions of the American Mathematical Society* **361**(6), 3359–3359.
<span style="font-variant:small-caps;">Cass, T.</span>, <span style="font-variant:small-caps;">Hairer, M.</span>, <span style="font-variant:small-caps;">Litterer, C.</span> and <span style="font-variant:small-caps;">Tindel, S.</span> (2015). Smoothness of the density for solutions to Gaussian rough differential equations. *The Annals of Probability* **43**(1), 188–239.
<span style="font-variant:small-caps;">T. Cass</span> and <span style="font-variant:small-caps;">N. Lim</span> (2018). A Stratonovich-Skorohod integral formula for Gaussian rough paths. *To appear in The Annals of Probability*.
<span style="font-variant:small-caps;">Cass, T., Litterer, C.</span> and <span style="font-variant:small-caps;">Lyons, T.</span> (2013). Integrability and tail estimates for Gaussian rough differential equations. *The Annals of Probability* **41**(4), 3026-–3050.
<span style="font-variant:small-caps;">Coutin, L.</span> and <span style="font-variant:small-caps;">Qian, Z.</span> (2002). Stochastic Analysis, Rough Path Analysis and Fractional Brownian Motions. *Probability Theory and Related Fields* **122**, 108–140.
<span style="font-variant:small-caps;">Decreusefond, L.</span> (2005). Stochastic Integration with respect to Volterra processes. *Annales de l’Institut Henri Poincaré – Probabilitiés et Statistiques* **41**(2), 123–-149.
<span style="font-variant:small-caps;">Decreusefond, L.</span> and <span style="font-variant:small-caps;">Üstünel, A. S.</span> (1999). Stochastic Analysis of the Fractional Brownian Motion. *Potential Analysis* **10**(2), 177–214.
<span style="font-variant:small-caps;">Elliott, R. J.</span> and <span style="font-variant:small-caps;">Van Der Hoek, J.</span> (2003). A general fractional white noise theory and applications to finance. *Mathematical Finance* **13**(2), 301–330.
<span style="font-variant:small-caps;">Friz, P.</span>, <span style="font-variant:small-caps;">Gess, B.</span>, <span style="font-variant:small-caps;">Gulisashvili, A.</span> and <span style="font-variant:small-caps;">Riedel, S.</span> (2016). The Jain-Monrad Criterion for rough paths and applications. *The Annals of Probability* **44**(1), 684–738.
<span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Hairer, M.</span> (2014). *A Course on Rough Paths*, Springer.
<span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Riedel, S.</span> (2013). Integrability of (Non-)Linear Rough Differential Equations and Integrals. *Stochastic Analysis and Applications* **31**(2), 336–358.
<span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Victoir, N.</span> (2010). Differential equations driven by Gaussian signals. *Annales de l’Institut Henri Poincaré – Probabilitiés et Statistiques* **46**(2), 369-–413. <span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Victoir, N.</span> (2010). *Multidimensional Stochastic Processes as Rough Paths: Theory and Applications*, 1st ed., Cambridge Studies in Advanced Mathematics **120**, Cambridge University Press.
<span style="font-variant:small-caps;">Friz, P.</span> and <span style="font-variant:small-caps;">Victoir, N.</span> (2011). A note on higher dimensional p-variation. *Electronic Journal of Probability* **16**, 1880–1899.
<span style="font-variant:small-caps;">Gubinelli, M.</span> (2004). Controlling Rough Paths. *Journal of Functional Analysis* **216**(1), 86–140.
<span style="font-variant:small-caps;">Hu, Y.</span>, <span style="font-variant:small-caps;">Jolis, M.</span> and <span style="font-variant:small-caps;">Tindel, S.</span> (2013). On Stratonovich and Skorohod Stochastic Calculus for Gaussian Processes. *The Annals of Probability* **41**(3A), 1656–1693.
<span style="font-variant:small-caps;">Hairer, M.</span> and <span style="font-variant:small-caps;">Pillai, N.S.</span> (2013). Regularity of laws and ergodicity of hypoelliptic SDES driven by rough paths. *The Annals of Probability* **41**(4), 2544-–2598.
<span style="font-variant:small-caps;">Inahama, Y.</span> (2014). Malliavin differentiability of solutions of rough differential equations. *Journal of Functional Analysis* **267**(5), 1566–1584.
<span style="font-variant:small-caps;">Lim, N.</span> (2016). A Stratonovich-Skorohod integral formula for Gaussian rough paths. *Ph.D. thesis*, Imperial College London.
<span style="font-variant:small-caps;">Lim, N.</span> (2018). Young-Stieltjes integrals with respect to Volterra covariance functions. *Preprint*.
<span style="font-variant:small-caps;">Lyons, T.</span> (1998). Differential equations driven by rough signals. *Rev. Mat. Iberoamericana* **14**(2), 215–-310.
<span style="font-variant:small-caps;">Lyons, T., Caruana, M.</span> and <span style="font-variant:small-caps;">Lévy, T.</span> (2007). *Differential Equations Driven by Rough Paths*, Springer.
<span style="font-variant:small-caps;">Lyons, T.</span> and <span style="font-variant:small-caps;">Qian, Z.</span> (2003). *System Control and Rough Paths*, Oxford Mathematical Monographs, Oxford University Press.
<span style="font-variant:small-caps;">Nourdin, I.</span>, <span style="font-variant:small-caps;">Nualart, D.</span> and <span style="font-variant:small-caps;">Tudor, C. A.</span> (2010). Central and non-central limit theorems for weighted power variations of fractional Brownian motion. *Annales de l’Institut Henri Poincaré – Probabilitiés et Statistiques* **46**(4), 1055–1079.
<span style="font-variant:small-caps;">Nualart, D.</span> (2006). *The Malliavin Calculus and Related Topics*, 2nd ed., Probability and Its Applications, Springer.
<span style="font-variant:small-caps;">Nualart, D.</span> and <span style="font-variant:small-caps;">Ortiz-Latorre, S.</span> (2011). Multidimensional Wick-Itô formula for Gaussian processes. *Stochastic Analysis, Stochastic Systems and Applications to Finance*, World Scientific.
<span style="font-variant:small-caps;">Privault, N.</span> (1998). Skorohod stochastic integration with respect to non-adapted processes on Wiener space. *Stochastics and Stochastic Reports* **65**, 13–39.
<span style="font-variant:small-caps;">Towghi, N.</span> (2002). Multidimensional extension of L.C. Young’s inequality. *Journal Of Inequalities In Pure And Applied Mathematics* **4**(1), 1–-7.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Deeply bound KNN, KNNN and KNNNN states are discussed. The effective force exerted by the K meson on the nucleons is calculated with static nucleons. Next the binding energies are obtained by solving the Schrödinger equation or by variational calculations.
The dominant attraction comes from the S-wave $\Lambda(1405)$ and an additional contribution is due to $\Sigma(1385)$. The latter state is formed at the nuclear peripheries and absorbs a sizable piece of the binding energy. It also generates new branches of quasi-bound states. The lowest binding energies based on a phenomenological KN input fall into the 40-80 MeV range for KNN, 90-150 MeV for KNNN and 120-220 MeV for K$\alpha$ systems. The uncertainties are due to unknown KN interactions in the distant subthreshold energy region.
address:
- 'Andrzej So[ł]{}tan Institute for Nuclear Studies 00-681 Warsaw, Hoza 69, Poland [^1]\'
- 'Helsinki Institute of Physics, P.O. Box 64, FIN-00014, Finland [^2]\'
author:
- 'S. Wycech'
- 'A.M. Green'
title: 'Variational calculations for K-few-nucleon systems'
---
Introduction
=============
In this paper a quantitative understanding of Kaon-few-nucleon quasi-bound states is attempted. In recent years, the existence of such states has been vividly discussed. It was initiated by the KEK finding of peaks in the nucleon spectra of K$^-$ absorption in $^4$He, [@KEK04; @KEK05]. Additional evidence was given by the FINUDA measurement of the invariant mass distribution of the $\Lambda p $ produced in K$^-$absorption by light nuclei [@FIN05]. The existence of such bound states have been expected as the kaon-nucleon and the kaon-nucleus interactions have been known to be strongly attractive, [@WYC86]. This is now firmly confirmed on the basis of kaonic atom data [@FRI07]. However, the KEK and FINUDA experiments indicate unexpectedly strong bindings of the order of 100, 150 MeV in the lightest KNN, KNNN systems. These experiments require further confirmation. Also, the interpretation of the observed peaks has been disputed in Refs. [@VAL06A], [@VAL06B] while the initial interpretation is defended in Ref. [@AKA07].
Calculations indicate that such states are expected, albeit these might be very broad and difficult to detect. The first calculations performed by Akaishi and Yamazaki in Ref. [@AKA02] were followed by several subsequent publications. These calculations exploited essentially the S - wave resonant attraction related to the $\Lambda(1405)$ state. With an optical model type of approach it was shown that the K-meson optical potential at the center of small nuclei may be as strong as 500 MeV generating very strong binding of the meson and a strong contraction of the few-nucleon systems. However, to reproduce the KEK data, these calculations involved some relaxation of the NN repulsion at short distances which would allow the existence of strongly bound and very dense systems. These calculations raise the important question on how to implement a realistic short range NN repulsion in the kaonic systems.
Another open question is related to the strength and range of KN interactions. Any mathematical description of few body systems requires knowledge of NN and KN off-shell scattering amplitudes. Those related to NN interactions are controlled fairly well in terms of modern NN potentials. For a bound K-meson the amplitudes needed involve the subthreshold energy region $$\label{1}
f_{KN}= f_{KN}( - E_B - E_{\rm{recoil}}),$$ where $E_B$ is the KN separation energy and $E_{\rm{recoil}} $ the recoil energy of the KN pair relative to the rest of the system. If the separation energy is as large as 100 MeV, meson momenta become $\approx 250 $ MeV/c and $ E_{\rm{recoil}}$ may be as large as $40$ MeV. The energies of interest for $ (- E_B -
E_{\rm{recoil}})$ are then located well below the $\Lambda(1405)$ state. The amplitudes there are strongly attractive and so when used in a standard optical potential approach may support very strong bindings. One problem that arises at this stage is of a technical character. As these amplitudes are energy dependent, it is hard to account for that in the optical model approach. There exists another, more serious, problem which is common to all approaches. As the energies involved are far away from the physical region tested in KN scattering, the uncertainties in the KN scattering amplitudes are sizable. For instance, if the $\Lambda(1405)$ is a KN bound state, then the amplitude far below the resonance is given not only by the position of the singularity but to a greater extent by the Born term, which indicates a strong dependence on the uncertain interaction range $r_o$. An old multi-channel potential model of Ref. [@KRZ75] indicates that the available scattering data do not allow one to fix the precise value of $r_o$. This unfortunate situation is still actual. The $r_o$ is expected to be close to the inverse vector meson mass. However, even though a change of 20$\%$ in $r_o$ would not affect the scattering region, it results in a 30$\%$ change of $f_{KN}$ in the deep subthreshold region. The corresponding uncertainty in the binding energy then amounts to $\approx 30$ MeV. As indicated by few body calculations of Ref. [@WEI06] this problem strongly affects the outcome. The uncertainties of $f_{KN}$ require further coherent experimental and theoretical studies of KN and K-few-N interactions. It becomes one of the most important purposes of the K meson physics.
On the other hand, there is one consequence of Eq.(\[1\]) which is model independent. If the binding and recoil are so large the $f_{KN}$ amplitudes involve the energies below the thresholds of meson-hyperon decay channels. As a consequence the dominant decay modes are blocked and the lifetimes of nuclear K meson systems are determined only by multi-nucleon captures. This leads to the expectation that such states may live long enough to be detectable, [@WYC86].
There exists several calculations of KNN binding energies. These states are named K$^{-}$pp although in reality they correspond to isospin $I_{NN}=1 $ and total isospin $I_{tot}=1/2$. The first prediction by Akaishi and Yamazaki lead to $( E_B, \Gamma) =
(48,60) $ MeV [@AKA02]. With a similar, molecular type method, Dote and Weise [@DOT07] obtain $ E_B, < 50 $ MeV and indicate a strong dependence of this result on the short range NN repulsion. On the other hand, the recent three body calculations based on Faddeev or AGS methods yield larger bindings. Thus Schevchenko *et al.* [@SHE07] obtain $( E_B, \Gamma) =
(55-70, 95-110) $ MeV while Ikeda and Sato [@IKE07] calculate $( E_B, \Gamma) = (\sim 80, \sim 73) $ MeV. Later in the text we show that the discrepancy of these two groups of results is due to: different $\Lambda(1405)$ properties, explicit description of the multiple scattering in decay channels and possibly to an incompatible treatment of the NN repulsion.
There are two new elements introduced in this paper. First, the P-wave interactions due to $\Sigma(1385)$ have been indicated as a possible source of the strong binding. Here, these are introduced explicitly. Second, the stress is put on the strong KN spacial correlations induced by the S and P wave resonances.
Leaving aside the interpretation of the peaks attributed to bound KNN and KNNN systems the essential theoretical questions are:
$(a)$ What is the binding mechanism?
$(b)$ Are the technical questions under control?
$(c)$ Can the widths be narrow?
This paper attempts an answer to these questions and the following results are obtained :
1\) To account properly for the KN force range, short range KN correlations and the NN repulsion, a two step calculation is performed. First a wave function involving strongly correlated K-N subsystems is found in a fixed nucleon approximation. This step also allows one to find potentials due to the K meson which tend to contract the nucleons. Next, these correlated wave functions and contracting potentials are used as the input in variational calculations for the K-few nucleon binding. In the KNN case the binding energy and width are found by solving the Schrödinger equation.
2\) While the dominant mechanism of attraction is related to the $\Lambda(1405)$ state, it is found that another resonant state, the $\Sigma(1385)$, contributes significantly to the structure of the bound states but much less to the binding in KNN and K-few-N systems. In addition the $\Sigma(1385)$ generates new branches of nuclear states that could not be generated by the $\Lambda(1405)$ alone.
3\) The binding energy is determined to a large extent by the attraction and the repulsive core in NN interactions. With the Argonne NN potential [@ARG95] one obtains the lowest state of KNN bound by about 40-80 MeV and a KNNN state bound by about 90-150 MeV. Moderate dependence on the KN interactions is found, provided these are constrained by the shape of $\Lambda(1405)$ and the value of the KN scattering length. However, the position of $\Lambda(1405)$ itself is not well known and this becomes the source of a large uncertainty. The effect of $ \Sigma(1385)$ on the binding energy is limited, In the states bound via $\Lambda(1405)$ it adds some 5-10 MeV contribution to the KNN binding and 10-20 MeV to KNNN binding. In this sense the suggestions of Ref. [@WYC07] are not fully supported. However, the effect of $ \Sigma(1385)$ on the space structure of deeply bound kaonic states is strong. The $ \Sigma(1385)$ is formed in peripheral regions and it absorbs a large fraction of the total K meson binding. In consequence the radii of these systems are fairly large and the nucleon densities are comparable to those met in the $^4$He nuclei.
4\) The problem of uncertainties related to the large recoil momenta entering Eq.([\[1\]]{}) is only partly removed. Large kaon momenta are hidden inside the resonant structures. In principle these may be kept under control with the help of other experiments. In practice it is not the case. The other sector of large momenta, due to the strong binding, is partly screened by the short range NN repulsion. The main consequence is a strong dependence of the meson binding energies on the position of the $\Lambda(1405)$ resonance. In principle the shape of $\Lambda(1405)$ is tested by the invariant mass distribution in the decay $\Sigma\pi$ channel. In practice it is not so as the relevant energy region is located close to the $\Sigma\pi$ threshold. In this region the theoretical and experimental uncertainties are large.
5\) These states are very broad if the binding energies are less than 100 MeV. For stronger bindings, which are possible under the current values of the KN parameters the main mesonic decay modes may be closed. The widths for non-mesonic modes are hard to calculate and extrapolations from the emulsion data are not very reliable. New experiments are needed.
A simple physical picture emerges from this approach. The mesons are strongly correlated to slowly moving nucleons. The correlations are of the $\Lambda(1405)$ type at large densities, and of the $ \Sigma(1385)$ type in the peripheries. Each K,N pair has a good chance to stay also in the $\Sigma ,\pi$ form. The structure is rather loose as sizable fractions of the binding energies are hidden in the short ranged correlations.
The KNN bound state
====================
This section presents an introduction to the method used in this work. Several steps describe the increasing degree of precision and also the increasing level of technical complications:
$\bullet$ At first the KNN levels are found within the fixed nucleon approximation with a simple $S$ wave KN interaction.
$\bullet$ The nucleon degrees of freedom and NN interactions are introduced and a related Schrödinger equation is solved.
$\bullet$ The method is extended to multiple channel situations.
$\bullet$ Both $S$ and $P$ wave KN interactions are allowed.
Consider scattering of a light meson on two identical, heavy nucleons. To begin with, the nucleons are fixed at coordinates $\textbf{x}_i (i=1,2) $ and the wave function is assumed to be in the form
$$\label{a2} \Psi(\textbf{x},\textbf{ x}_1 ,\textbf{x}_2) =
\chi_K(\textbf{x};\textbf{ x}_1, \textbf{x}_2)~
\chi_{NN}(\textbf{x}_1,\textbf{x}_2),$$
where $\textbf{x}$ is the meson coordinate. The notation is simplified and some possible indices are suppressed. The meson wave function $\chi_K $ is given by the solution of the multiple scattering equation $$\label{a3}
\chi_K(\textbf{x},\textbf{x}_1,\textbf{x}_2)= \chi_K(\textbf{x})^o - \Sigma_i ~ \int \textbf{dy }\frac
{\exp[ip\mid \textbf{x-y}\mid ]}
{4\pi \mid \textbf{x-y}\mid }~ U_{KN}(\textbf{y},\textbf{x}_i)~
\chi_K(\textbf{y},\textbf{x}_1,\textbf{x}_2)$$ obtained with fixed positions of the nucleons. An equation of similar structure with a zero range meson-nucleon pseudo-potential $U$ was used by Brueckner [@BRU] to calculate the scattering length of a meson on two nucleons. For a high energy scattering it was extensively discussed by Foldy and Walecka, who used finite range separable interactions $U$ [@FOL69]. With such interactions equation (\[a3\]) allows for semi-analytic solutions in the NN, and also in few nucleon cases. Here, the method is extended to the bound state problem. One looks for solutions of Eq. (\[a3\]) with no incident wave $
\chi_K(\textbf{x})^o$. The momentum $p$ becomes a complex eigenvalue $p(x_i)$ which determines the energy and width of the system for given nucleon positions $x_i$.
Equation (\[a3\]) is written in terms of the Klein-Gordon or Schrödinger propagator. The difference arises when the relation of energy and momentum is established. Reasons of simplicity, which will become clear later, favor the non-relativistic relation in the KN center of mass system. Thus, the interaction is presented as $U_{KN} =2\mu_{KN} V_{KN}$ where $\mu_{KN}$ is the reduced mass. Corrections for relativity may be introduced at a later stage. The potential $V_{KN}$ for an $S$ wave interaction is chosen in a separable form $$\label{a4}
V_{KN}( \bf{x-x_i}, \bf{x'-x_i}) =
\lambda ~\upsilon(\bf{x-x_i})~\upsilon(\bf{x'-x_i}),$$ where $\upsilon$ is a form-factor and $\lambda$ is a strength parameter. The eigenvalue equation is now reduced to $$\label{a5}
\chi_K(\textbf{x},\textbf{x}_1,\textbf{x}_2) + \Sigma_i ~ \lambda ~\int \bf{dy}~ \frac {\exp[ip(\textbf{x}_1,\textbf{x}_2)\mid \bf{x-y}\mid ]}
{4\pi \mid \bf{x-y}\mid } ~\upsilon(\bf{y-x_i})~ \int \bf{dy'}~ \upsilon(\bf{y'-x_i})~\chi_K( \bf{y'},\textbf{x}_1,\textbf{x}_2) =0.$$ Equation (\[a5\]) becomes a matrix equation for wave amplitudes $\psi_i$ defined at each scatterer $i$ by $$\label{a6}
\psi_i = \lambda~ \int \bf{dx}~ \upsilon( \bf{x-x}_i) ~\chi_K( \bf{x},\textbf{x}_1,\textbf{x}_2).$$ To find the equations for $\psi_i $ one introduces the off-shell KN scattering matrices $f$ and matrix elements of the propagator $$\label{a7}
G_{i,j}(\bf{x_i,x_j}) = \int\bf{dy }\bf{dx } ~ \upsilon( \bf{x-x}_i)~
\frac {\exp(ik\mid \bf{x-y}\mid )}{4\pi \mid \bf{x-y}\mid
} ~\upsilon(\bf{y-x}_j).$$ The diagonal value, $G_{i,i}\equiv G$, determines the meson nucleon scattering matrix $t$ by the well known (see e.g. Ref. [@FOL69]) relation $$\label{a7t}
t(E) = (1+\lambda ~ G )^{-1}~ \lambda$$ and this yields the full off-shell scattering amplitude $ f $ $$\label{a7f}
f(k,E,k') =\upsilon(k)~t(E)~\upsilon(k').$$ Here, $k,k'$ are the initial and final momenta while the form-factor $v(k)$ is given by the Fourier transform of $\upsilon(r)$. The Yamaguchi form $\upsilon(k)=
1/(1+k^2/\kappa^2)$ with a free parameter $\kappa$ will be used in this paper. At zero momenta and at the threshold this choice normalizes $f$ (and $t$) to the scattering length. Unfortunately for historical reasons the standard convention in the K-N system is to define the scattering length by $$\label{a7g}
a+ib = - f(k=0,E=0,k'=0)\equiv F(0,0,0)$$ and the capital $F$ will be used in several places to comply with the standard KN parameters.
In order to cast Eq. (\[a5\]) into a standard multiple scattering equation for $\psi_i$ one carries out the following three steps: 1) Integrates Eq. (\[a5\]) over the i-th form-factor $\upsilon(\bf{x-x_i})$. 2) Selects the i-th term from the R.H. side. 3) Multiplies Eq. (\[a5\]) by $ (1+\lambda
G)^{-1}$. In this way the kernel of the multiple scattering equation can be expressed in terms of scattering amplitudes $t_i$ at each nucleon $i$ and propagators describing the passage from the nucleon $i$ to the other nucleon $j$. One now arrives at a set of linear equations $$\label{a8}
\psi_i +
\Sigma_{j\neq i} ~ t_j~G_{i,j}~ \psi_j =0,$$ which may be solved numerically. For the Yamaguchi form-factors, propagators $G_{i,j}$ allow analytic expressions $$\label{a9}
G_{1,2}(r,k) =
\frac{1}{ r}~ \upsilon(k)^2 ~[ \exp(ikr)-\exp(-\kappa r)- r
\frac{\kappa^2 +k^2}{2\kappa} \exp(-\kappa r)] ~\equiv~G (r,k),$$ where $ \textbf{r }= \textbf{x}_2- \textbf{x}_1 $. For the sake of illustration, the KNN case is presented in some detail. The condition for a bound state with two amplitudes $\psi_i$ leads to a pair of equations $$\label{a10}
\psi_1 + t~G ~\psi_2 = 0, ~~~~ \psi_2+t~G~\psi_1 = 0.$$ When the determinant $$\label{a11a} D= 1-( t~G)^2$$ is put to zero, the binding “momenta” $p(r)$ may be obtained numerically. Two different solutions corresponding to $ 1+ tG = 0 $ or $ 1-tG = 0$ may exist. The first solution is symmetric $\psi_2= \psi_1$ and describes the meson in the $S$ wave state with respect to the NN center of mass. The second solution is antisymmetric $\psi_2= -
\psi_1$ and describes a $P$ wave solution. With the rank one separable interaction this latter solution does not exist in the full range of $r$. However, it arises with the more complicated rank two interactions discussed later.
Eigenvalues corresponding to unstable quasi-bound states are obtained in the second quadrant of complex $ p(r)\equiv p = p_R +i p_I$ plane. In this quadrant the kernel $$\label{a12}
tG =f(p) ~[ \exp(-p_Ir) \exp(ip_Rr)-\exp(-\kappa r)~( 1+ r \frac{\kappa^2 - p^2}{2\kappa})]/r$$ is exponentially damped at large distances as required by the asymptotic form of the bound state wave function $\chi_K$. At short distances $ G $ is regularized by the KN form-factor.
If the scattering amplitude is dominated by a quasi-bound state, such as $\Lambda(1405)$, the related pole dominates and in some energy region $ f\simeq \gamma^2/(E-E^*)$, where $\gamma$ is a coupling constant and $E^*=E_r-i\Gamma_r/2$ is a complex $\Lambda(1405)$ binding energy. The full KNN binding energy, $V_K$, is given by the equation $ 1+ tG= 0 $ which becomes $$\label{a13} V_K(r) \simeq E^* - \gamma^2 G(r,p).$$ Since Re $ G(r,p)$ close to the resonance is positive this solution offers binding stronger than the $\approx$ 28 MeV binding in the $\Lambda(1405)$. Asymptotically, for $r\rightarrow\infty$ one obtains $ V_K \rightarrow E^* $, that is a kaon bound to a nucleon to form a $\Lambda(1405)$. This type of asymptotic behavior occurs in all of the K-few-nucleon systems of practical interest. Hence, the separation energy is understood here as the separation of the K-N-N system into the N-$\Lambda(1405)$ system.
The limits $ r \rightarrow0$ in Eqs. (\[a12\],\[a13\]) are regular. However, a joint limit of zero range KN interactions, $\kappa \rightarrow \infty$, and $ r\rightarrow0$ is singular and the KNN system collapses. Therefore, some care is necessary when this limit is taken. Here, we stay within a phenomenological approach and the standard expectation that the range of KN interactions is determined by vector meson exchange. In equation (\[a9\]) for $G$ the range of interactions enters twice, first as a cutoff at small distances and second in terms of the form-factor $v(k)^2$. We find in a numerical way that these two effects cancel and $fG$ is very stable within the range $ 3< \kappa < 6 ~fm $. As the KN interaction range is very short, but finite, the uncertainties related to the actual value of $\kappa$ are additionally eliminated by the short range repulsion in the NN systems. This yields an important stability in the few-body calculations described here.
With nucleons fixed at a distance $r$ the eigen-value condition determines $ p(r)$ which in turn generates the potential $V_K(r)$ contracting the NN system to a smaller radius. The form and strength of this potential depends on the form of the kinetic energy. In the KN C.M. system the $ E_{KN}$ energy is given by $ \sqrt{M^2 +p^2 } + \sqrt{m^2 + p^2},$ where $m,M$ are masses of the meson and nucleon. The same form is kept in the large nucleon mass $M$ limit. In the few nucleon systems the non-relativistic form of the nucleon energy is used and the problem of large nucleon mass disappears.
The meson propagator in Eq. (\[a3\]) is chosen to make the multiple scattering equation (\[a8\]) equivalent to a differential equation $$\label{a14}
[ - \Delta + \sum_i ~ 2\mu_{KN}~ V_{ KN_i }~]~\chi = p(r)^2$$ and the contracting potential becomes $V_{K}= p(r)^2/ 2\mu_{KN}$ . The advantage of this choice is discussed in the next section.
Schrödinger equation
---------------------
The solution of the full KNN bound state problem is given by equation $$\label{s15}
( - \frac{\Delta_x}{2m} - \frac{\Delta_1}{2M}- \frac{\Delta_2}{2 M} + V_{KN1}+V_{KN2}+ V_{NN} ) \Psi = E \Psi.$$ The wave function is assumed in the form $\Psi= \chi_K(x,x_i) \chi_{NN}(r)$ as given in Eq. (\[a2\]). Multiplying Eq. (\[s15\]) on the left by $ \chi_K$ and integrating over the meson coordinate $x$ one obtains the Schrödinger equation for the NN wave function $$\label{s16}
\chi_{NN} (r) = \int \textbf{dx } \chi_K( x,x_i) \Psi (x,x_i)$$ in the form $$\label{s17}
[ E - V_K(r) + \Delta_1 / 2 M + \Delta_2 / 2 M - V_{NN}]~\chi_{NN} + \Delta E_{kin} \chi_{NN} =
0.$$ where the last term $\Delta E_{kin}$ is a correction to the kinetic energies. This correction is small due to the choice of the meson kinetic energies. In the Schrödinger equation (\[s15\]) it is given by the meson mass $m$. On the other hand, to determine the $\Lambda(1405)$ properties and to solve the scattering equation (\[a3\]) the reduced mass $\mu_{KN}$ is used. Due to this, the correction term $ \Delta E_{kin} $ is of the order of $1/M$. In addition, the meson wave function satisfies the relation $$\label{s19}
\Delta_x \chi_K = \sum_i \Delta_i\chi_K,$$ which may be obtained by partial integration over coordinate $y$ in Eq. (\[a5\]). In this way $$\label{s18} \Delta E_{kin} \chi_{NN} =-\frac{1}{M} \Sigma_i~ \int
\bf{dx }~\chi_K~\overrightarrow{\partial}_i\chi_K
~\overrightarrow{\partial}_i \chi_{NN},$$ which is very small due to angular averaging and sign changes in the derivatives. In more detail this correction reduces to $$\label{s18a} \Delta E_{kin} \chi_{NN} =-\frac{2}{M} \int
d\bf{\xi}~ \frac{\overrightarrow{\xi } \overrightarrow{r}}{\xi
r}~~\frac{G(\bf{\xi}-\bf{r})\partial_{\xi} G(\xi)} { \int
d\bf{\eta}[ G(\bf{\eta}-\bf{r})+G(\bf{\eta})]^2 } ~ ~\partial_r
\chi_{NN}(r),$$ and is suppressed by the angular average over $\xi$ and at large $r$ by the small overlap of $G(\bf{\xi}-\bf{r})$ and $G(\xi)$. The $\Delta E_{kin}$ makes a contribution $ \approx 0.2 $ MeV to the binding energy. Such twice damped, small terms of similar type, arise also in more involved versions of this calculation. The $\Delta
E_{kin}$ is of the same order but is given by very lengthy formulas. Since it is very small in comparison to the dominant uncertainties in $V_K$ it is dropped, leading to a significant simplification of the variational approach.
As the next step, equation (\[s17\]) is solved with an S-wave interaction based on the more realistic NN potential of Argonne [@ARG95]. This solution is also compared to another, variational solution with the intention of checking the variational method used in heavier systems. The actual interaction used, in the notation of Ref. [@ARG95], has the form $$v(NN)=v^{EM}(NN)+v^{\pi}(NN)+v^{R}(NN),
\label{Wir1}$$ where the electromagnetic part $v^{EM}$ only includes the dominant term proportional to $F_C(r)$ in Eq. (4) of Ref. [@ARG95], the OPE term $v^{\pi}$ is given by Eq. (18) of Ref. [@ARG95] and the phenomenological short range term $v^{R}$ from Eq. (20) with the parameters in Table II - again all from Ref. [@ARG95]. This gives directly the S-wave T=1, S=0 interaction $v(\rm{S-wave, \
T=1, \ S=0})$. However, in the T=0, S=1 deuteron channel, the effect of the tensor interaction $\upsilon_t(T=0, \ S=1)$ on the central component $
\upsilon_c(\rm{T=0, \ S=1})$ is incorporated by the closure approximation to give $$V(\rm{S-wave, \ Deuteron})= \upsilon_c(\rm{T=0, \ S=1})-
\frac{8\upsilon_t(\rm{T=0, \ S=1})^2}{\rm{Den}}, \label{Wir2}$$ where the energy denominator was adjusted to Den=338 MeV to ensure the correct binding energy of the deuteron.
The precision of variational estimates for E ( used in the next sections) may be checked against numerical solutions of the Schrödinger equation. It is about 0.3 MeV, compared with the overall binding of $\sim 50$ MeV. The width of the state is calculated as $$\Gamma/2 = < \chi_{NN}\mid Im~ V_K \mid \chi_{NN} > .
\label{gamma}$$
Interactions in the decay channels
----------------------------------
The decay channel $ \Sigma\pi$ coupled to the basic KN channel is now introduced explicitly. The wave function at each scattering center has two components one in the KN the other in the $
\Sigma\pi$ channel. The scattering amplitudes are two dimensional vectors $ \psi_i \rightarrow [\psi_i^K,~ \psi_i^\pi
$\] at each nucleon. Multiple scattering equations given in the previous section are now changed accordingly. One has $$\label{t1}
\psi_1^K + t^{K,K}~G^{K,K}~ \psi_2^K + t^{K,\pi}~G^{\pi,\pi}~ \psi_2^\pi = 0$$ $$\label{t2}
\psi_1^\pi +G^{\pi,\pi}~t^{\pi,\pi}~ \psi_2^\pi + t^{\pi,K}~G^{K,K}~ \psi_2^K = 0$$ and an analogous pair with $1\leftrightarrow 2$. The notation has been changed to describe channel indices and the $2\times 2
$ scattering matrix $\hat{f}$. The determinant related to these equations gives the complex eigenvalue $p(x_i)$ in the KN channel. The eigen-equation is now more complicated. Introducing a new notation in channel indices $ U^{a,b} =G^{a,a}t^{a,b}$ the determinant becomes $$\label{t3} D=
[(1+ U^{K,K})(1+ U^{\pi,\pi})~
- U^{\pi,K}U^{K,\pi}][(1-U^{K,K})(1-U^{\pi,\pi})~
-U^{K,\pi}U^{\pi,K}].$$ The $D=0$ condition is more transparent close to the singularity in the case of a scattering amplitude given by $$\label{t4}
f^{a,b} \approx \frac{\gamma_a \gamma_b} { E - E_o+ i \Gamma/2}.$$ Consistency requires the width to be $ \Gamma/2=
p_\pi(\gamma_{\pi})^2$, where $p_\pi $ is the momentum in the decay channel. The singular term (\[t4\]) permits one to find a solution of Eq. (\[t3\]) in a fairly simple form. It is presented below in the limit of zero range KN ( and $ \Sigma\pi$ ) force. The binding energy $$\label{t5} \emph{Re } E = E_o - (\gamma_K)^2\frac{cos(p_Rr)}{r}
\exp(-p_Ir) - (\gamma_{\pi})^2 \frac{cos(p_\pi r)}{r}$$ becomes larger than the binding of the resonance but the collisions in the decay channel indicates oscillations. This oscillatory behavior is also seen in the width of the system $$\label{t6} \emph{Im}~E = - (\gamma_{\pi})^2 ~ p_\pi~[
1+\frac{sin(p_{\pi}r)}{p_{\pi}r}] - (\gamma_K)^2\frac{sin(p_Rr)}{r}\exp(-p_Ir).$$ The effect of KN scattering represented by the second term enlarges the width as $p_R$ is negative. The contribution from multiple scattering in the decay channel is sizable in general but it oscillates and may under some conditions reduce the total width. That is an effect of interference in the decay channel. Scattering in the decay channel turns out to be constructive in the KNN case but it is not necessarily so in some heavier systems.
S and P wave interactions
--------------------------
With the KN interactions allowed in both $S$ and $P$ waves the scattering equation (\[a8\]) is a $4\otimes 4$ matrix equation relating four amplitudes $ \psi_i$. The amplitudes for $S$ waves are now denoted by $ \psi_1^s ,\psi_2^s$. For $P$ wave interactions the corresponding amplitudes are vectors. As there is only one vector in the $NN$ system, the relative separation, the $P$ amplitudes are chosen to be $ \textbf{r}\psi_1^p $ and $
\textbf{r}\psi_2^p $. The scattering is now described by three types of propagators $ G^{a,b}$ related to consecutive collisions in the $(S,S)$ $(S,P)$ and $(P,P)$ waves. The scattering equations are $$\label{sp1}
\psi_1^s ~+ ~f^s ~G^{ss}~\psi_2^s - ~f^s~G^{sp}~r^2 ~\psi_2^p =0$$
$$\label{sp2}
\psi_2^s ~+ ~f^s ~G^{ss}~\psi_1^s + ~f^s~G^{sp}~r^2 ~\psi_1^p =0$$
$$\label{sp3}
\psi_1^p ~+ ~f^p ~G^{pp}~\psi_2^p + ~f^p~G^{sp} ~\psi_2^s =0$$
$$\label{sp4} \psi_2^p ~+ ~f^p ~G^{pp}~\psi_1^p - ~f^p~G^{sp}
~\psi_1^s =0,$$
where the propagation in between two $P$ wave interactions is described by $ G^{pp} = G^{pp}_O + r^2 G^{pp}_T$. Indices numbering the nucleons have been suppressed. The propagator $G^{ss}$ is given in Eq. (\[a9\]) and explicit formulas for $
G^{sp},G^{pp}_O , G^{pp}_T $ may be found in the appendix. All these functions are regular in the $ r\rightarrow 0$ limit. The determinant $D$ of this system factorizes into two terms $$\label{sp5}
D = D_S~D_P,$$ where $$\label{sp6}
D_S = ( 1 ~+~ G^{ss}~ f ^s)( 1~-~ G^{pp}~ f ^p )~ - ~ G^{sp}~r^2 ~f^s~ f ^p,$$ $$\label{sp7}
D_P = ( 1~ -~ G^{ss}~ f ^s)( 1~ + ~ G^{pp}~f ^p )~-~G^{sp}~r^2~ f^s~ f^p.$$ Let us consider the solution of $D_S=0$ close to the $\Lambda(1405)$ resonance. It is given by an equation analogous to (\[a13\]) $$\label{sp8} E = E^* - b^2~ G^{ss}(r,p(E)~ [ 1 ~-~ \frac{r^2(G^{sp})^2
f^p}{1 - G^{pp} f^p}~].$$ The second term in parentheses describes the effect of $P$ wave interactions. At energies below $\Sigma(1385)$ the amplitude $f^p$ is negative and generates an additional attraction.
Isospin symmetry simplifies the algebraic structure of the scattering equations which are (see next sections) expressed by appropriate isospin combinations of the isospin KN scattering amplitudes $f$. Equations (\[sp1\]-\[sp4\]) allow for a simple symmetry of the total KNN wave function. Thus, under the eigenvalue condition (\[sp6\]) the coordinate wave function becomes symmetric with respect to the exchange of nucleon coordinates. This condition allows solutions in terms of two amplitudes $\psi^s = \psi^s_1 =\psi^s_2$ and $ \psi^p =\psi^p_1 =
- \psi^p_2 $. Wave functions for the KNN system have the form $$\label{I6}
\Psi(\bf{r},\bf{x}) = \chi_{NN}(r)
[ G(\bf{x}-\bf{r}/2) + G(\bf{x}+\bf{r}/2)] \psi_s +
\chi_{NN}(r) \overrightarrow{r} \overrightarrow {\partial}_x [
G(\bf{x}-\bf{r}/2) - G(\bf{x}+\bf{r}/2)] \psi_p,$$ where $\bf{x}$ is the meson coordinate in the NN center of mass system, $ \chi_{NN}(r)$ is the NN wave function. To make this formula more transparent the zero range force limit is taken. The two terms in Eq.(\[I6\]) follow the KN interactions in $S$ and $P$ waves. The weight of the $P$ wave contribution is given by $$\label{I7} \frac{\psi_p }{ \psi_s} = \frac{ f^p~G^{sp}}{1 -
f^p~G^{pp}},$$ which becomes dominant close to the zero of the denominator in this equation. At large NN separations it happens almost at the singularity in $ f_p$. In this region the lowest energy, symmetric, solution of $D_S=0 $ is given essentially by the situation $1- G^{pp} f ^p \approx 0
$. Such a solution exists for $r\geq 1.6 fm$ in the proper quadrant of the complex momentum. This implies that, at large separations, it is energetically profitable for the KNN system to exist in the N $\Sigma(1385)$ configuration, with the nucleon and $\Sigma(1385)$ weakly repelling each other. At shorter distances the condition $ 1 + G^{ss} f ^s \approx 0 $ determines the attraction generated by $\Lambda(1405)$. Despite repulsive effects of the $P$ wave interaction such a solution yields the strongest binding, since a large piece of the binding energy is hidden within the structure of $\Sigma(1385)$.
The KNN system is built on short range KN and NN correlations and the $ \Psi(\bf{r},\bf{x})$ contains a large number of partial waves coupled to zero total angular momentum. In Jacobi coordinates $(r,s)$ the $L_r \otimes L_s $ decomposition of the first term of $ \Psi(\bf{r},\bf{x})$ is mainly $ S \otimes S $. Both terms involve even values of $L_r $ and the spin-isospin structure of the NN pair is either $ I_{NN}=0 , S_{NN}=1$ or $ I_{NN}=1 , S_{NN}=0$. Both types of states may be formed.
Other solutions are determined by $D_P=0$. This condition allows amplitudes of different symmetry $\psi^s = \psi^s_1 =- \psi^s_2$ and $ \psi^p =\psi^p_1 = \psi^p_2 $. In comparison with Eq. \[I6\], the wave functions for the KNN system now become $$\label{I8}
\Psi(\bf{r},\bf{x}) = \chi_{NN}(r)
[ G(\bf{x}-\bf{r}/2) - G(\bf{x}+\bf{r}/2)] \psi_s +
\chi_{NN}(r) \overrightarrow{r} \overrightarrow {\partial}_x [
G(\bf{x}-\bf{r}/2) + G(\bf{x}+\bf{r}/2)] \psi_p$$ are now antisymmetric in the nucleon coordinates and contain odd angular momenta $L_r, L_s $ in the $L_r \otimes L_s$ decomposition. The spin-isospin structure of the NN pair is either $ I_{NN}=0 , S_{NN}=0$ or $ I_{NN}=1 , S_{NN}=1$. The NN interactions in the $ I_{NN}=0 , S_{NN}=0$ states are repulsive and do not support any KNN bound states. On the other hand, $
I_{NN}=1 , S_{NN}=1$ states may be formed.
The results of the two previous subsections may be unified. The notation used in Eq.(\[t3\]) is now extended to include the partial wave index in channel KN : $U^{p,s}= G^{p,p} t^s ,~
U^{s,p}= G^{s,s} t^p ,~ U^{p,p}= G^{p,p} t^p, U^{s,s}\equiv
U^{K,K}$. The last equivalence indicates that the $P$-wave multiple scattering is included only in the basic KN channel. For the determinant of scattering equations one obtains $ D= D_S~D_P
$ where now $$\label{I9} D_S = [(1+ U^{K,K})(1+ U^{\pi,\pi})~ -
U^{\pi,K}U^{K,\pi}][1+U^{p,p}]- [(1-U^{\pi,\pi})U^{p,s}U^{s,p}]$$ $$\label{I10} D_P =
[(1- U^{K,K})(1- U^{\pi,\pi})~
- U^{\pi,K}U^{K,\pi}][1+U^{p,p}]- [(1+U^{\pi,\pi})U^{p,s}U^{s,p}].$$ The solutions of the corresponding eigen-value equations retain the symmetries indicated in the previous section.
Figure 1 indicates typical contracting potentials obtained with an A.Martin solution (see next section and Ref. [@MAR81]). Asymptotically these reproduce the separation energies of the K meson hidden either in the $S$-wave $\Lambda(1405)$ or in the $P$-wave $\Sigma(1385)$. The real binding is generated by the difference $ V_K(r)- V_K(\infty)$ which jointly with $V_{NN}$ determines the nucleon wave function $\chi_{NN}$.
![Contracting potential $V_K$(r) in the NN, $I_{NN}=1,
I_{KNN}=1/2$ state, for the symmetric meson wave solutions. The $V_S$ line shows *Re* $V_K(r)$ for S wave interactions. The $ V_{SP}$ line shows *Re* $V_K(r)$ for S + P wave interactions. The upper curve $-\Gamma$ shows 2 *Im* $V_K(r)$. These results are based on the A. Martin amplitudes [@MAR81].](knn.epsi){width="40.00000%"}
\[fig:VK\]
KN interactions
===============
The coupled multichannel $KN , \Sigma\pi,\Lambda\pi $ system is the easiest to describe in terms of the $\hat{K}$ matrix related to the scattering matrix $\hat{T}$ by the algebraic Heitler equation $$\label{K1} \hat{T} = \hat{K} - \hat{K} i\hat{ Q} \hat{T},$$ where $Q$ is a diagonal matrix of channel momenta in the C.M. system. Early parameterizations involved constant $K$-matrix elements chosen to fit the scattering data. Later these were improved by an effective range expansion. As the data were (and still are) poor such fits were supplemented by additional consistency conditions formulated in terms of dispersion relations [@MAR81; @MSA69]. Such solutions can be tested above the KN threshold and to some extent in the $\Sigma\pi $ channel. For the dominant isospin $0$ interactions there are two types of solutions. These are given in Table \[KN\] in terms of the inverse $\hat{M}= \hat{K}^{-1}$ matrix which in turn determines the scattering matrix $$\label{K2} \hat{T}^{-1} = \hat{M} + i\hat{ Q}.$$
Extrapolations into the complex energy plane display a similar $\Lambda(1405)$ pole position. However, the physics in both solutions indicates different interplay of the main KN with the hyperon pion channels. The position of the singularity is given essentially by the attractive and, in both cases, large $K_{KN,KN}$ element. This allows one to interpret $\Lambda (1405)$ as a $KN$ quasi-bound state. In principle there exists an alternative possibility - the $\Lambda$ as a quark state. If this is the case it may be introduced into the $K$ matrix as an external pole in $\hat{K}\sim 1/(E-E^*)$. However, the scattering data exclude such a term or limit it to a very small contribution [@MAR81]. Amplitudes below the KN threshold may be tested indirectly, either in the elastic $\Sigma\pi $ channel or in the $KN
\rightarrow\Sigma\pi $ transitions on bound nucleons [@STA87]. These reactions support the bound state interpretation but are not very restrictive on the position of the singularity. In particular, the analysis of Dalitz and Deloff [@DAL91] shows that several models offer comparable descriptions of the $\Sigma\pi $ data in the resonance region. The $M$ matrix model given in the DD column of Table \[KN\] is only slightly favored by the authors of ref. [@DAL91]. [^3]. The KWW column in Table \[KN\] comes from a quasi-relativistic separable potential model. It belongs to a second type of solution and was based on the B. Martin – Sakitt solution. In this work we use off-shell extrapolations of both types of solutions.
The separable off-shell extension
---------------------------------
The three-channel or two-channel separable model is used here to extend the phenomenological S-wave KN interactions off the energy shell. This method is standard in momentum space but here we have already used the coordinate representation. In a single KN channel case the potential equivalent to those of Eq.(\[a3\]) is described by $$\label{S1} V( k , k' ) = \lambda
v(k)v(k').$$ The Yamaguchi form-factors $v(k) = \kappa^2 /(\kappa^2 + k^2)$ are convenient to perform explicit analytic calculations in both representations. The related Fourier transforms are of Yukawa form $ v(r)= \kappa^2\exp(-\kappa r)/ ( 4 \pi r) $ and are normalized to delta functions in the limit of zero range forces. The off-shell scattering amplitude becomes $$\label{S3}
f_{KN} =\frac{ v(k)v(k')}{ \lambda^{-1} + G(E)}.$$ With the non-relativistic form of the kinetic energy $E_{kin} = q^2/(2\mu_{KN}) $ one obtains $$\label{S4}
G(E) = \int d \bf{\tau} \frac{ v(t)^2}{2\pi^2(\tau^2-q^2)} =
\frac{ \kappa}{2( 1 - i q/ \kappa )^2} .$$ At the KN threshold energy $ E_t=M_K+M_N $, the standard convention requires one to define the scattering length as $
a_o+ib_o = -~f_{KN}(E_t)$. Below the threshold $1/\lambda + G$ is forced to have a zero corresponding to the $\Lambda(1405)$ state. The complex momentum at this point is $$\label{S5}
p_B = [\frac {\lambda_S \kappa^3}{2}]^{1/2} - \kappa.$$ The next step in order to improve the absorptive part is to guarantee that it vanishes below the $\Sigma \pi$ threshold. That is achieved by scaling the absorption strength Im $\lambda $ by a phase space factor $f_{\varrho} = q_{\Sigma\pi}(E) / q_{\Sigma\pi }( E_t) $ where $q_{\Sigma\pi}$ is the momentum in the decay channel. The values $ \lambda_S = - 0.602\exp(i~0.12~f_{\varrho})~fm $ and $ \kappa = 4.5 fm^{-1}$ give a good reproduction of the PDG recommended E = 1405 and $\Gamma=50 $ MeV values [@DAL91; @PDG].
This one channel amplitude may serve as a guide, but to describe finer details and for a better comparison with the scattering data one needs multi-channel separable models. Below, two types of multi-channel reaction matrices are extrapolated off the energy-shell.
$\circ$ One solution has been given by Krzyzanowski $et \ al.$ [@KRZ75] in terms of a quasi-relativistic multi-channel separable potential. The $G(E)$ used there differs from the solution (\[S4\]) by an invariant momentum phase space and the use of quasi- relativistic intermediate meson energies $E =
q^2/(2M_{N}) + \sqrt{m^2+q^2} $. This solution was motivated by the early B. Martin and Sakitt [@MSA69] K- matrix (BM in Table \[KN\]) and as may be seen in this table it offers similar on-shell parameters.
$\circ$ Another solution is based on the commonly used A. Martin’s K-Matrix [@MAR81]. However, in the decay channels this matrix is not well reproduced by simple rank one separable potentials. Instead, we use the extrapolation $ K^{off}(k,E,k')
= v(k)~K^{on}~ v(k')$, with the Yamaguchi form-factors and $\kappa =$ 4.5 fm$^{-1}$ as obtained above in the one channel case.
P-wave KN interactions
-----------------------
To account for the P-wave interactions dominated by $\Sigma(1385)$ the scattering amplitude of equation (\[a7f\]) is generalized to $$\label{P1}
f_{KN} = f_S + f_P = f_S + 2 {\bf k} {\bf k}' f_P^{l+}.$$ The last term is a consequence of the $j= l+1/2$ total spin of $\Sigma(1385)$. It involves an $l+1$ factor instead of $2l+1$ typical for spin zero situations. The omitted piece contains spin flip amplitudes and is expected to be small in the few body context. The $f_P^{l+}$ term is described here by a separable single-channel K - matrix which in the KN channel is given by $$\label{P2} K( \textbf{k}, \textbf{k'}) = \gamma_K^2 \frac{
\textbf{k} \textbf{k'}v_P(k)v_P(k')}{ E- E_o+ i \Gamma_\pi/2},$$ where the form-factor is $$\label{P3} v_P(k) = \frac{\kappa^4 }{(\kappa^2 + k^2)^2},$$ and $E_o$ is a phenomenological parameter which determines the position of $\Sigma(1385)$. The width $\Gamma_{\pi}$ is strongly energy dependent $$\label{P4} \Gamma_\pi = \gamma_{\Lambda\pi}^2 q_{\Lambda\pi}^3 + \gamma_{\Sigma\pi}^2 q_{\Sigma\pi}^3 .$$ In these equations the $q$’s are the channel momenta and $
\gamma$ are couplings of the resonance to the $\Sigma\pi$ and $
\Lambda\pi$ channels. The latter are determined by the experimental decay width of about 36($\pm 2$) MeV and the $\Sigma\pi$ branching ratio of 0.13($ \pm$ 0.01 ) [@PDG]. For the coupling to the KN channel the SU(3) value $ \gamma_{KN}^2/
\gamma_{\Sigma\pi}^2 = 2/3$ is taken. This value is consistent with the experimental result of Brown, which yields $ 0.57( \pm
0.18)$ [@BRO79].
The coupling to the KN channel generates the off-shell scattering amplitude $$\label{P5} f_P(\textbf{k}, E, \textbf{k}') = \frac{2\textbf{k}
\textbf{k}' v_P(k)v_P(k')
\gamma_{KN}^2 }{ E- E_o+ \gamma_{KN}^2~\int
\frac{d\bf{\tau} \tau^2 ~v_P(\tau)^2 } {2\pi^2(\tau^2-k_{KN}^2)}+ i~ \Gamma_\pi/2 }
\equiv \textbf{k} \textbf{k}'v_P(k)f_P(E)v_P(k').$$ Let us notice that below the $KN$ threshold the integral in Eq.(\[P5\]) deforms significantly the shape of the resonance profile. The range parameter $\kappa = 3.8 $ fm$^{-1}$and $ E_o=
1505.2$ MeV are used to reproduces the profile tested experimentally by Cameron [*et al.*]{} [@CAM78] in the $\Lambda \pi$ channel. For further applications the coordinate representation is needed, which is given by equation $$\label{P6} f( \textbf{x}, E, \textbf{x}') = \lambda_P
\overleftarrow{\partial}v_P(x)f_P(E)
v_P(x')\overrightarrow{\partial'}.$$ in terms of the Fourier transforms of form-factors (\[P3\]).
Few nucleon systems
====================
The procedure presented in the KNN section is now extended to systems consisting of several nucleons. Practical calculations are done for three and four nucleons. At first the multiple scattering equations similar to Eqs. (\[a3\] - \[a5\]) in the previous section are solved in fixed nucleon systems. The bound K-meson wave function $\chi_K$ is a solution of $$\label{f1}
\chi_K(\bf{x},x_1....x_n ) =
- \Sigma_i\Sigma_{\beta} \int \bf{dy }\frac {\exp(ik\mid \bf{x-y}\mid )}{4\pi \mid \bf{x-y}\mid
}~ v( \bf{x-x}_i)_{\alpha}~\lambda^{\alpha\beta}~ v(\bf{y-x}_i)_{\beta} ~\chi_K( \bf{y}, x_1....x_n
),$$ where indices $ \alpha, \beta $ denote channels and partial waves of the meson-baryon pair. An index related to the symmetry of $\chi$ is suppressed. By analogy with Eqs. (6) and (7), equation (\[f1\]) may be reduced to a matrix equation for the wave amplitudes defined at each scatterer as $$\label{f2}
\psi_i^{\alpha} = \Sigma_{\beta}\int \textbf{dy }~\lambda^{\alpha\beta}~ v(\bf{y-x}_i)_{\beta} ~\chi_K( \bf{y}, x_i ).$$ The kernel of the scattering equation can be now expressed in terms of scattering amplitudes $f_i^{\alpha, \beta}$ at each nucleon $i$ and propagators describing the passage from nucleon $i$ to another nucleon $j$. The latter are given by $$\label{f3}
G_{i,j}^{\alpha,\beta} =
\int \textbf{dy }\textbf{dx } v( \textbf{x-x}_i)_{\alpha }
\frac {\exp(ik\mid \textbf{x-y}\mid )}{4\pi \mid \textbf{x-y}\mid
} v(\textbf{y-x}_j)_{ \beta}.$$ The procedure explained in Sect. II leads to a set of linear equations $$\label{f4}
\psi_i^{\alpha} + \Sigma_{\beta\gamma}~
\Sigma_j ~ f_j^{\alpha,\beta}~ G_{i,j}^{\beta\gamma} ~ \psi_j^{\gamma} =0 ,$$ which are solved numerically. This matrix equation is simplified as the $G$’s are diagonal in channel indices and the $f$ are diagonal in the partial wave index. Still, the corresponding determinants are complicated algebraic expressions involving functions $ G$ and $f$. Numerical solutions become a difficult problem. It has been solved in the following, approximate way. The determinant consists of many terms which are arranged according to the number of collisions. With up to four collisions in the main channel we retain the structure found in the KNN situation and the determinant $D$ of this system is presented as $$\label{f5}
D = 1 + \Sigma_{pairs} (D_S~D_P-1) + O_{higher~ orders},$$ where $D_S$ and $D_P$ are defined in Eqs. (\[I9\]) and (\[I10\]). The main term is composed of collisions in the KNN subsystems which allows one to keep track of the wave function symmetry. The terms of higher order in $f$ are dropped.
The solution of the full K-few-N bound state problem is given by the equation $$\label{f6}
[ E + \frac{\Delta_x}{2 m}+ \sum_i \frac{\Delta_i}{2M}- \sum_i
V_{KN_i}-
\sum_{i,j} V_{N_iN_j} ] \Psi(x,x_1,..,x_n) = 0.$$ Again we assume the wave function to be given by eqs.(\[a2\]) and (\[f1\]) i.e. $\chi_K(x,x_1..x_n) \chi_{N}(x_1,..,x_n))$. Projecting Eq.(\[f6\]) on $\psi_K$ one obtains the Schrödinger equation for the few nucleon wave function $$\label{f7}
\chi_{N} (x_i) = \int \textbf{dx } \chi_K( x,x_i) \Psi (x,x_i)$$ in the form $$\label{f8}
[ E - E^{c}(x_i) + \sum_i \frac{\Delta_i}{2M} - \sum V_{NN}]\chi_N + \Delta E_{kin} \chi_N =0 ,$$ where $\Delta E_{kin}$ is a correction to the nucleon kinetic energies. As in Eq. (\[s18\]) it is given by $$\label{f9} \Delta E_{kin} \chi_N =
-\frac{1}{M} \Sigma_i \int \bf{dx }\partial_i\psi_K\partial_i \Psi$$ and as before turns out to be very small due to angular averaging and sign changes in both the derivatives. As discussed in the KNN situation this correction has been dropped. In deriving equation (\[f8\]) a special form (\[f1\]) of the meson wave function is used. As in Eq. (\[s19\]) it satisfies the relation $$\label{f10}
\Delta_x \chi_K =\sum_i\Delta_i\chi_K.$$
In the next step, equation (\[f8\]) is solved by a variational method with the NN potential of Argonne [@ARG95]. The trial wave function is of the form $$\label{f11}
\chi_N =\prod_{i,j, i\neq j}[ 1-\exp(-\lambda_c^2
(\bf{x}_i-\bf{x}_j)^2)]\exp(-\lambda_l
\mid\bf{x}_i-\bf{x}_j \mid) /\mid\bf{x}_i-\bf{x}_j \mid,$$ where $\lambda_c,\lambda_l$ are variational parameters. This form is chosen to give the correct asymptotic limit for large $ \mid
\bf{x}_i-\bf{x}_j \mid$ and also gives a vanishing wave function for small $ \mid \bf{x}_i-\bf{x}_j \mid$ as expected for a strong repulsion in the NN potential.
Results
========
KNN
---
In this section the calculations of KNN levels are presented. The sensitivity to KN input parameters is studied and the states of different symmetry are discussed.
Contracting potentials $V_K(r)$ were calculated with several solutions for the phenomenological S-wave $KN$ reaction matrices presented in Table \[KN\]. The solutions of second type may be well fitted with a rank one separable potential. Here, the calculations are done with the quasi-relativistic model of [@KRZ75]. This model is based on the $K$-matrix of B. Martin [@MSA69]. Numerically it is fairly close to the separable potential of Ref. [@ALB76]. For the the first type of solutions in Table \[KN\] - due to A. Martin [@MAR81] no satisfactory rank one separable approximation is found. This difficulty is related to the large effective range parameters involved in this $K$ matrix. In order to retain the physics involved, a simple off-shell extension is adopted $ K_{i,j}
\rightarrow v(k_i)K_{i,j}v(k_j)$. The Yamaguchi form-factors have been used and the inverse range parameter $\kappa$ was varied over the range of $3-6$ fm$^{-1}$. The actual value of $\kappa$ affects the multiple scattering via propagator $G_{ss}(k,r)$ of Eq. (\[a9\]). Larger values reduce the form-factor $v(k)$ but enhance the significance of the small $r$ region in $G_{ss}(k,r)$. On average these two effects balance very well and one finds a very weak ($\sim 1$ MeV) dependence of the total binding energy on the actual value of $\kappa$. The results given in Fig.1 and in the tables which follow are obtained with $
\kappa=4.5$ fm$^{-1}$.
The energies of the most strongly bound KNN, $I_{tot}=1/2$, $I_{NN}=1$, quasi-bound states are given in Tables \[KNNS1\] and \[KNNlow\]. The first table describes several steps of the approximation while the second table indicates the dependence of binding on the KN input parameters.
The first line in Table \[KNNS1\] is determined essentially by the effects of $\Lambda(1405)$ excitations described in the elastic channel only. The second line describes additional effects due to multiple scattering in the $\Sigma\pi$ channel. The other two lines include the P wave interactions. The energies of $I_{tot}=1/2$, $I_{NN}=1$ states given by the S wave interactions and described by multiple scattering in the single, KN, channel span the region of 30-50 MeV. These results are consistent with the findings of Akaishi and Yamazaki [@AKA02] and Dote and Weise [@DOT07] obtained with different methods. The differences within this range are due to a different KN and/or NN input. As seen in the second and fourth rows, significant changes arise with the explicit inclusion of the multiple scattering in the $\Sigma,\pi $ channels. The binding rises by 10 to 20 MeV and the effect of collision broadening is large.
solution
--------------------- ------- ---------- ----------- ------- ---------- -----------
$E_B$ $\Gamma$ $R_{rms}$ $E_B$ $\Gamma$ $R_{rms}$
$KN; ~~S$ 27 36 3.1 35.5 37 2.4
$KN,\Sigma\pi;~S $ 37 42 2.5 43.1 47 2.1
$KN; ~~S,P$ 49 36 3.7 49.7 36 3.3
$KN,\Sigma\pi;~S,P$ 52 37 2.9 56.5 39 2.3
$KN; ~~S$ 47 47 2.3
: Binding energies and widths \[in MeV\] of the KNN, $I_{tot}=1/2$, $I_{NN}=1$ space-symmetric states. The results on the left are based on [@MAR81] parameters, the results on the right follow KWW [@KRZ75] parameters discussed in Table \[KN\]. The first column specifies the channels explicitly involved in the multiple scattering and meson-nucleon partial waves. $R_{rms}$ is the radius mean squared of the N-N separation \[in fm \]. The last line is obtained with the simplest separable potential discussed in the text and the $I_{KN}=1$ amplitudes from AM.
\[KNNS1\]
Table \[KNNlow\] shows binding energies obtained with the “canonical” $\Lambda$ pole position $E = 1405 $ MeV [@DAL91; @PDG] which is lower than the position obtained in other parameterizations. The result given in the second line of this table is comparable to the results obtained, with a similar input, by Schevchenko [*et al.*]{}[@SHE07]. The latter work employs a superior Faddeev technique, but a more detailed comparison of results is not easy since that calculation uses a rank one separable potential to describe the NN interactions.
solution KWW$^*$
--------------------- ------- ---------- -----------
$E_B$ $\Gamma$ $R_{rms}$
$KN; ~~S$ 50 51 2.05
$KN,\Sigma\pi;~S $ 71 85 1.81
$KN; ~~S,P$ 65 43 2.09
$KN,\Sigma\pi;~S,P$ 78 60 1.88
: Binding energies and widths \[in MeV\] of the KNN, $I_{tot}=1/2$, $I_{NN}=1$ space-symmetric states. These results are based on KWW [@KRZ75] parameters modified to set the pole of $\Lambda(1405)$ at 1405 MeV and the width at 48 MeV and given in Table \[KN\] (KWW$^*$). The first column specifies the channels explicitly involved in the multiple scattering and the meson-nucleon partial waves. $R_{rms}$ is the radius mean squared of the N-N separation \[in fm\].
\[KNNlow\]
The inclusion of resonant P wave interactions increases the binding by some 10 MeV. There is some room for different values as the experimental $K N \Sigma(1385)$ coupling is not certain. However, the main effect of $\Sigma(1385)$ is a change of structure in the KNN systems. A sizable portion of the binding energy is contained in the structure of this resonance. On the other hand the system is dissolved as the inclusion of P waves enlarges the NN separation and the formation of $\Sigma(1385)$ is essentially a peripheral effect. The KN correlations for $ r>1.6 $ fm are mostly of the $\Sigma(1385)$ type. The other effect of P wave interactions is a formation of additional KNN states. These are given in Tables \[KNNS0\],\[KNNP\] and discussed below.
The energies of KNN quasi-bound states with $I_{NN}=0$ given in Table \[KNNS0\] are determined essentially by the $\Sigma(1385)$ excitations. Let us notice that the result is unstable against the KN input. The state is still more likely to exist with a lower value of the $\Lambda(1405)$ energy. In any case it is a very loose structure that might be a quasi-bound or a virtual state.
$E_B$ $\Gamma$ $R_{rms}$ $E_B$ $\Gamma$ $R_{rms}$
---------- ---------- ----------- ------- ---------- -----------
no b.s. - - 47.1 36 $\sim $7
: Binding energy and width \[in MeV\] of the KNN, $I_{tot}=1/2$, $I_{NN}=0$ space-symmetric states. Results on the left are based on A.Martin parameters and do not support a bound state (no b.s.). The results on the right follow the KWW parameters. This result involves $S+P$ wave interactions and external interactions in two meson-nucleon channels. $R_{rms}$ is the radius mean squared of the N-N separation \[in fm\].
\[KNNS0\]
$I_{NNK}$ $I_{NN}$ $E_B[MeV]$ $\Gamma[MeV]$ $R_{rms}[fm]$
--------- ----------- ---------- ------------ --------------- ---------------
$K^-nn$ 3/2 1 48.5 36 4.9
: The binding energy and width \[in MeV\] of the $KNN$ $I_{tot}=3/2$, $I_{NN}=1$ space-asymmetric states. In the $NN$ subsystem $^{2S+1}L_J= ^3P_2$. The last column gives the radius mean squared of the N-N separation, \[in fm\].
\[KNNP\]
The energy of an asymmetric quasi-bound state is given in Table \[KNNP\]. It is determined essentially by the $\Sigma(1385)$ excitations. The table in Appendix B indicates that the $K^-nn$ state has the largest possible $\Sigma$ component which offers the strongest $\Sigma(1385)-N$ attractive potential. It reaches a maximum depth of about $-10$ MeV at $~ 1 ~fm $ distance, but it is not strong enough to overcome the NN $P$-wave barrier and generate a quasi-bound state. To obtain real binding, assistance from the $I_{KN}=1 $ $ S$-wave state and the $NN$ $P$-wave attraction is necessary. Thus, the $NN$ interactions repulsive at large distances in the $ I_{NN}=0$, $
S_{NN}=0,L_{NN}=1$ waves do not support bound states. However, such states may be generated by $ I_{NN}=1$ , $S_{NN}=1,L_{NN}=1$ interactions. Here, the analysis becomes more subtle as the $NN$ interaction is strongly spin dependent. The energy given in Table \[KNNP\] corresponds to $J=2$ ($^3P_2$) wave in the NN subsystem where the interaction is the most attractive. This KNN system is large and loosely bound, by about $1.5$ MeV in the $\Sigma(1385)-N$ configuration. In this calculation the S wave parameters come from AM [@MAR81] and the calculated energy is uncertain, as the involved $K^-n$ parameters are poorly known. The experimental detection would be very difficult, nevertheless, a more precise analysis of the spin and isospin structure of such states is of interest in the context of $K^-D$ atoms and will be performed elsewhere.
KNNN and KNNNN systems
----------------------
The discussion of these systems is limited to the states of the simplest symmetry. The fixed nucleon model generates a contracting potential which in KNNN systems may be, to a good approximation, presented in the form $$\label{R1} E^c(R_x, R_y, R_z) = - V_{NNN}[1- C\exp(-\lambda_{s}(
R_x+R_y+ R_z))][\exp(-\lambda_{l}R_x) + \exp(-\lambda_{l}+ R_y)+
\exp(-\lambda_{l}R_z)] - V(\infty) ,$$ where $R_x, \ R_y, \ R_z $ are the inter-nucleon distances. The short range behavior at the triple coincidence may be obtained analytically and $C= 0.42$, other parameters being numerical.
With the KN parameters of Refs.[@KRZ75],[@MAR81] and $I_{tot}=0$ the parameters are obtained in the range $V_{NNN}\sim150-200$ MeV, $\lambda_{s}\simeq4.5$ fm , $\lambda_{l}\sim 1.8-1.9$ fm. For $I_{tot}=1$ one has $V_{NNN}\sim 50-60$. $V(\infty)$ is the binding of KN into $\Lambda(1405)$ in the S wave case or $\Sigma(1385)$ in the S+P wave case. For $I_{tot}=0$ the corresponding binding energies are given on the left side of Table \[KNNN0c\]. These numbers may be compared to the simplest version of this model - the $S$ wave interactions described by the single KN channel - which produce 91 MeV binding.
The modified version of the KWW model with parameters from Table \[KN\] fixed to set the $\Lambda(1405)$ energy to 1405 MeV yields much stronger contracting forces, $V_{NNN}\sim250-350$ MeV and $\lambda_{l}\sim
2.1-2.3$ fm. The states indicated on the right side in Table \[KNNN0c\] are bound very deeply. The basic NNN systems obtained with our variational wave function are over-bound by about 2 MeV and this value has already been subtracted from the numerical KNNN energies.
$E_B$ $\Gamma$ $E_B$ $\Gamma$
------- ------- ---------- ------- ----------
$S$ 103 29 142 25
$S+P$ 119 23 153 21
: Binding energies and widths \[in MeV\] of the $I_{tot}=
0$, KNNN, space-symmetric states obtained with the two channel KN , $\Sigma\pi$ channel multiple scattering formulation. The results on the left are based on the modified KWW [@KRZ75] parameters with $\Lambda(1405)$ energy set to 1405 MeV. The first column specifies meson-nucleon partial waves involved. The widths do not describe non-mesic capture modes.
\[KNNN0c\]
$E_B$ $\Gamma$
------- ------- ----------
$S+P$ 63 38
: Binding energies and widths \[in MeV\] of the $I_{tot}=
1$, KNNN, space-symmetric states obtained with the two channel KN , $\Sigma\pi$ channel multiple scattering formulation. These states are formed as a result of the $P$ wave interactions with some assistance of the $S$ wave attraction.
\[KNNN1c\]
There may exist a number of states in the KNNNN systems. In Table \[KNNNN\] one finds only the states with the simplest symmetry, which involve wave functions symmetric under exchange of the nucleon coordinates. In the absence of the K meson the basic $\alpha$ particle structure is used and only the $S$ wave NN interactions are included. With the tensor interactions described by Eq.(\[Wir2\]) this $\alpha$ system is over-bound by about 10 MeV, and this value has been subtracted from the calculated KNNNN levels.
$E_B$ $\Gamma$ $E_B$ $\Gamma$
------- ------- ---------- ------- ----------
$S$ 121 25 170 10
$S+P$ 136 20 172 10
: Binding energies and widths \[in MeV\] of the KNNNN , space-symmetric, $S_{tot}=0$, $I_{tot}=1/2$ states. See captions to Table \[KNNN0c\].
\[KNNNN\]
Level widths
------------
Level widths are calculated as twice the expectation value of $ Im
~V_K$. The KN resonant lifetimes are strongly energy dependent, being very short at the resonance, becoming longer below the resonant energies and staying infinite below the thresholds of the decay channels. This trend is reflected by $ Im
~V_K$ in Fig. \[fig:VK\]. The energy dependence contained in the amplitude $f_{KN}( - E_B- E_{\rm{recoil}})$ of Eq. (\[1\]) is traded into the space dependence $f_{KN}(V_K)$. These two types of averaging give fairly close results provided the final binding energy is located well above the threshold of the decay channels. Let us indicate some consequences of this relation.
The states generated by the P wave interactions given in Tables \[KNNS0\],\[KNNP\], and \[KNNN1c\] correspond to a fairly loosely bound $\Sigma(1385)$ and the widths of quasi-bound states are essentially equal to the width of the $\Sigma(1385)$. This comes as a result of the peripheral binding and weak effects of the collision broadening in the $P$ wave resonances. In these states, the $V_K$ underestimates slightly the average $ - E_B-
E_{\rm{recoil}}$ and the real widths might be smaller. For the binding energies in the range 60-90 MeV the $V_K$ is too small at large distances and too large at small distances with a reasonably good average. Let us notice that the level widths generated by the $S+P$ interactions are smaller than those generated by the $S$ waves alone. This is due to three factors: the width of $\Lambda(1405)$ is larger than the width of $\Sigma(1385)$, the collision broadening in $P$ waves is small and the systems due to the $S+P$ interactions are less compressed. Let us also notice very strong sensitivity to the input KN amplitudes. The few examples of $ Im ~F_{100}$ given in Table \[KN\] and the differences of the widths in the $ I_{tot}=1/2$, $ I_{NN}= 0$, KNN state visualize this point.
In cases of very large binding, in the range of 120-200 MeV, one ($\Sigma\pi$) or two ($\Sigma\pi,\Lambda\pi$) decay channels are blocked and the widths calculated here are over estimated. Such a situation is likely to happen in the K-$\alpha$ state. To account for that effect, the calculation of the contracting potential was repeated in an optical potential manner. So, the momenta in the decay channels $\Sigma\pi,\Lambda\pi$ were related to the binding energy $E_B$ and allowed no outgoing waves. Such a calculation results in a stronger binding. In the KWW$^*$ model one obtains binding of 220 instead of the 170 given in Table \[KNNNN\]. The real decay width is now given by the multi-nucleon capture mode.
The multi-nucleon captures are initiated by the non-mesic KNN $\rightarrow$ Y N mode and the branching ratio for this process is known from emulsion studies to be about $20\%$ in light nuclei [@ROO79]. The emulsion data are obtained with stopped K mesons and pertain to the nuclear surfaces. An extrapolation in terms of a characteristic nuclear densities $\rho$ and two body phase space $L$ $$\label{R2} \Gamma_{multi} \simeq L ~ \varrho^2 ~ \gamma$$ for this decay was attempted in Ref.[@WYC86]. A constant $\gamma$ may be fixed to the emulsion branching ratio and a 20 MeV level width in the nuclear matter at 90 MeV binding was obtained. In the strongly bound, few-body systems the kinematics of the decay is different since the residual nucleons also take sizable recoil energies. Roughly, for a three body decay $ L \sim Q ^2$ where $Q$ is the decay energy. Again, an extrapolation from the emulsion data in terms of the available phase space and the involved nuclear density yields non-mesic capture widths in K-$\alpha$ in the 10-30 MeV range. These estimates are somewhat larger than the 12 MeV obtained for the KNNN system in Ref.[@AKA02]. Unfortunately such extrapolations are uncertain as the energy dependence in $\gamma$ might be large and the $Q$ value is not known. Help from new experiments is necessary to settle these questions.
Conclusions
===========
In this paper a new method to calculate the deeply bound KNN, KNNN and KNNNN states has been presented. The calculation consists of two steps. First a wave function involving strongly correlated K-N subsystems is found in a fixed nucleon approximation. This step also allows one to find potentials due to the K meson which tend to contract the inter-nucleon distances. Next, these correlated wave functions and contracting potentials are used as input in the Schrödinger or variational calculations for the K-few-nucleon binding.
The lowest binding energies based on a phenomenological KN input fall into the 40-80 MeV range for KNN, 90-150 MeV for KNNN and 120-220 MeV for K$\alpha$ systems. The uncertainties are due to unknown KN interactions in the distant subthreshold energy region.
We obtain at least partial answers to the basic questions presented in the introduction.
$(a)$ The binding mechanism: the dominant mechanism of the attraction is related to the $\Lambda(1405)$ state. This fact has been known for a long time. In addition, it is found here, that the $\Sigma(1385)$ contributes significantly to the structure of the K-few-N bound states but much less to the actual binding energies. The bound states are built from the strongly correlated KN subsystems. At central densities these correlations resemble the $\Lambda(1405)$ and at peripheries the correlations are made by the quasi-free $\Sigma(1385)$. Sizable fractions of the binding energies are contained in the KN correlations. One consequence is that even with the strong bindings the nucleon densities are not dramatically enhanced as in Ref. [@AKA02] but can become a factor 2-4 larger than the standard nuclear matter density $\rho_o$.
The presence of $\Sigma(1385)$ resonances in the few nucleon systems generates new states. These are predominantly P-wave states or states built upon the P-wave N-N interactions and are usually broad and loosely bound developing long tails built from the $\Sigma(1385)$ correlation.
$(b)$ The control of technical questions: the choice of correlated wave functions removes the difficulties related to the uncertain K-N interaction range and allows one to use realistic N-N interactions. The recoil energy of the KN subsystems with respect to the residual nucleons is described only in an average sense. This seems to be the weakest part of this method.
$(c)$ The widths are related to the lifetimes of the $\Lambda(1405)$ and $\Sigma(1385)$ enhanced by the collision broadenings. Under the phenomenological KN interactions ( the $\Lambda $ pole located at $\sim 1412 - 17i $ MeV ) the K-few-N states are $\sim~40 $ MeV wide.
On the other hand, the models with the $\Lambda $ pole located at $ \sim 1405 - 25i $ MeV generate K-few-N states which are more deeply bound. These may be either very broad or quite narrow. With the binding energies in the 60-80 MeV range ( the KNN case) very broad - up to 90 MeV - states are obtained. However, with the bindings of 120 MeV (KNNN) or 170 MeV (KNNNN) the single nucleon decay modes are effectively blocked. The widths are strongly reduced and the main decay modes are due to multi-nucleon K captures. These widths are hard to predict, a simple model suggested here generate widths of about 20 MeV.
A simple physical picture emerges from this approach. The mesons are strongly correlated to slowly moving nucleons. The correlations are of the $\Lambda(1405)$ type at large densities, and of the $ \Sigma(1385)$ type in the peripheries. Each K,N pair has a good chance to stay also in the $\Sigma ,\pi$ form. The structure is rather loose as sizable fractions of the binding energies are hidden in the short ranged correlations.
[0]{}
T. Suzuki $ et \ al$., [*Phys.Lett.*]{} [**B597**]{}(2004)263 : M. Iwasaki, [*Proc.EXA 05, Austr. Acad. Sc. Press, 2005* ]{}p. 93: M. Agnello for FINUDA Collab. [*Phys.Rev.Lett.*]{} [**94**]{}(2005)212303 S. Wycech, [*Nucl.Phys.*]{} [**A 450**]{}, 399c, (1986) E. Friedmann and A.Gal, [*Physics Reports* ]{} [**452**]{} (2007)89 E. Oset and H.Toki [*Phys. Rev.* ]{} [**C74**]{}(2006)015207 V.K. Magas, E. Oset, A. Ramos and H. Toki [*Phys. Rev.* ]{} [ **C74**]{}(2006)025206, Y. Akaishi and T. Yamazaki, [*Nucl.Phys.* ]{} [**A792**]{}(2007)229
Y. Akaishi and T. Yamazaki, [*Phys. Rev.* ]{} [**C65**]{}(2002)044005 W. Weise, [*Proc.EXA 05, Austr. Acad. Sc. Press, 2005* ]{}p. 35
W. Krzyzanowski, J. Wrzecionko and S. Wycech [*et al*]{}., [*Acta.Phys.Pol.*]{} [**B6**]{}(1975)259 A. Dote and W. Weise, [*EPJ A* ]{}, nucl-th/0701050 [*Prog.Theor. Phys. Suppl.* ]{} [**168**]{}, 593(2007)
N. Shevchenko, A. Gal and J. Mares, [*Phys. Rev.* ]{} [**C76**]{}, 044004(2007)
Y. Ikeda and T. Sato, [*Phys. Rev.* ]{} [**C76**]{}, 035203(2007)
R.B. Wiringa [*et al*]{}., [*Phys. Rev.* ]{} [**C51**]{}(1995)38 S. Wycech and A.M. Green, [*Proc.EXA 05, Austr. Acad. Sc. Press, 2005* ]{}p. 101, [*IJMP*]{}[**A22**]{}, 629(2007)
K.A. Brueckner, [*Phys. Rev.*]{} [**89**]{}, 834(1953)
W.L. Foldy and J.D. Walecka, [*Ann.Phys.*]{} [**54**]{}(1969)447 A.D. Martin [*Nucl.Phys.* ]{} [**B94**]{}(1975)413 B.R. Martin and M. Sakitt, [*Phys.Rev* ]{} [**183**]{}(1969)307, B.R. Martin in [*Proceedings of Herceg Novi Summer School 1972.* ]{} B.R. Martin, [*Nucl.Phys.* ]{} [**B184**]{}(1981)33 M. Alberg, E. Henley and L. Wilets, [*Ann. of Phys.* ]{} [ **96**]{}(1976)43.
R.H. Dalitz and A. Deloff, [*J.Phys.G* ]{} [**17**]{}(1991)289
O. Brown, [*Nucl.Phys* ]{} [**129**]{}(1979)1 W. Cameron [*et al*]{}. , [*Nucl.Phys.* ]{} [**143**]{}(1978)189
Particle Data Group, [*Phys.Rev .* ]{}[**D 50**]{}, 1734(1990)
E. Friedmann, A. Gal, J. Mares and A. Cieply, [*Phys.Rev .* ]{} [**60**]{}, 024314(1990)
J. Mares, E. Friedmann and A. Gal, [*Nucl. Phys.* ]{} [**A 777**]{}, 84(2006)
R. Staronski and S. Wycech, [*Journ. Phys.*]{} [**G13**]{}, (2004)1361 :
T.Yamazaki and Y. Akaishi, [*Phys. Rev.* ]{} [**C76**]{}, 045201(2007)
R. Roosen [*et al.*]{}, [*Nuovo. Cim.* ]{} [**49A**]{}, 217(1979)
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the KBN grant 1P0 3B 04229 and the EU Contract No. MRTN-CT-2006-035482, FLAVIAnet”. Part of this work was carried out while one of the authors (SW) was at the Helsinki Institute of Physics.
Propagators
===========
Several formulas for kernels of multiple scattering equations are collected in this appendix.
For Yamaguchi form-factors, the propagators $G_{i,j}$ yield analytic expressions. Thus for two consecutive S-wave interactions one has $$\label{22}
G(r,k)^{ss} =
\frac{1}{4\pi r} v(k)^2 \Big[ \exp(ikr)-\exp(\kappa r)- r
\frac{\kappa^2 +k^2}{2\kappa} \exp(-\kappa r) \Big],$$ where$ \mathbf{r }= \mathbf{x}_i- \mathbf{x}_j $, and the indices $i,j$ referring to the nucleon sites are suppressed. For an initial S wave scattering followed by a P wave scattering $G$ becomes a vector $$\label{22a}
\textbf{G}(r,k)^{sp} = \textbf{r}G^{sp}(r,k),$$ $$\label{22b} G^{sp} = \frac{1}{4\pi r} v(k)^2 \Big[ \exp(ikr)
(ikr-1) -\exp(\kappa r)[1 +\kappa r + \frac{r^2(\kappa^2 +
k^2)}{2} + \frac{r^3(\kappa^2 + k^2)^2}{8\kappa} ]\Big].$$ For two consecutive P wave interactions the propagator is a tensor of the form $$\label{22c}
\textbf{G}(r,k)^{pp}|_{nm} = v(k)^2[ \delta_{nm}~G^{pp}_{O} + r_{n}
r_{m}~G^{pp}_{T} ].$$ These functions may be expressed in terms of basic integrals $$\label{22d}
g_n(r) = \frac{4\pi}{(2\pi)^3} \int d{\bf p}
\exp(ikr) \{\frac{\kappa^2 }{\kappa^2+p^2}\}^{n}
\frac{1}{p^2- k^2} ,$$ which give by recurrence $$\label{23}
g_{1}(r) = \frac{ \exp(ikr)-\exp(\kappa r)}{ r}, \ \
g_{2}(r)= g_{1}(r) - \frac{\kappa^2 +k^2}{2\kappa} \exp(-\kappa r)$$ $$\label{23b}
g_{3}(r)= g_{2}(r) -
\frac{(\kappa^2 +k^2)^2}{8\kappa^3}(1+\kappa r) \exp(-\kappa r), \
\ g_{4}(r)= g_{3}(r) -
\frac{(\kappa^2 +k^2)^3}{16\kappa^5}
(1+ \kappa r + \frac {(\kappa r)^2}{3} )\exp(-\kappa r)$$ and finally $$\label{24}
G^{pp}_{O} = \frac{g_{4}(r)'}{r}, \ \ G^{pp}_T = \frac{ g_{4}(r)''}{r^2} -\frac{ g_{4}(r)'}{r^3}.$$
Isospin symmetry
================
It is assumed here that the isospin is conserved in the quasi-bound states of K mesons. In the lowest S-wave states of the KNN systems the isospin wave functions may be built upon iso-singlet or iso-triplet NN states. From the experimental point of view, the most interesting one seems to be $$\label{IT1}
\Psi^{1/2}_1 = \{\{NN\}^{1}K\}^{1/2} =
\sqrt{3}/2\{\{NK\}^{0}N\}^{1/2}+1/2\{\{NK\}^{1}N\}^{1/2},$$ where in $\Psi^{1/2}_1$ the upper index denotes isospin $
I_{nucl}$ and the lower index denotes the spin of the two nucleons. On the right side the isospin content in the KN subsystem is given. This state is a mixture of $K^-pp $ and $K^0np$ and is frequently named $K^-pp $ since it can be experimentally accessed via this entrance channel. The NN spin in this state is $S=0$ and the effective KN interaction amplitude obtained from Eq.(\[IT1\]) becomes $$\label{IT2}
f_{KN} = 3/4 f^0_{KN} +1/4 f^1_{KN}.$$
Another KNN state of interest is built upon the NN iso-singlet $$\label{IT3}
\Psi^{1/2}_0 = \{\{NN\}^{0}K\}^{1/2} =
-1/2 \{\{NK\}^{0}N\}^{1/2}+\sqrt{3}/2\{\{NK\}^{1}N\}^{1/2} .$$ This state is a mixture of $K^-np $ and $K^0nn $ which might be reached by the $K^-np $ entrance channel. Now the NN spin is $S=1$ and the effective KN interaction amplitude obtained from Eq. \[IT3\] becomes $$\label{IT4}
f_{KN} = 1/4 f^0_{KN} +3/4 f^1_{KN}.$$ The S - wave KN interaction in the $\Psi^{1/2}_0 $ state is much less attractive than in the $\Psi^{1/2}_1 $ state since the $\Lambda(1405)$ contribution is reduced. However, this is offset by the strong short range attraction in the NN system due to the tensor force. An additional attractive force is due to a larger contribution from the $\Sigma(1385)$ resonance.
Finally one may have total isospin 3/2 states of the type $K^-nn $ or $K^0pp$ $$\label{IT5}
\Psi^{3/2}_1 = \{\{NN\}^{1}K\}^{3/2} = \{\{NK\}^{1}N\}^{3/2}.$$ Those states, involve weakly attractive and uncertain, S - wave KN I = 1 amplitudes. A deeper state can in principle be built upon the stronger P - wave interactions. Its existence and the chances for detection present a situation that is more difficult than the other cases.
For the three nucleon problem we retain the dominant structure of the triton and helium isospin symmetry. The KNNN wave function is assumed to be of the form $$\label{IT7}
\Psi^{T} = \frac{1}{\sqrt{2}}\{
\{\{\{NN\}^{0,1}N\}^{1/2}+\{\{NN\}^{1,0}N\}^{1/2}\}K\}^{T},$$ where the pair of indices denote spin and isospin of the $NN$ pair.
Re-coupling to the KN system leads in the total T=0 state to the relation
$$\label{IT8}
\Psi^{0} = \sqrt{1/2}\{\{NN\}^{1} \{NK\}^{1} +
\{NN\}^{0} \{NK\}^{0}\}^0$$
and in this case the KN interaction amplitude is $$\label{IT9}
f^s = 1/2 f^0_{KN} +1/2 f^1_{KN} .$$ Likewise for the total isospin 1 system $$\label{IT10}
f^s = 1/6 f^0_{KN} +5/6 f^1_{KN}$$ These amplitudes are collected into the table.
\[IX\]
System $I_{tot}$ $I_{nucl}$ $f_{KN}$
-------- --------------- --------------- ---------------------------------------
KNN $\frac{3}{2}$ 1 $f_{1}$
KNN $\frac{1}{2}$ 1 $\frac{3}{4}f_{0}+ \frac{1}{4}f_{1}$
KNN $\frac{1}{2}$ 0 $\frac{1}{4}f_{0}+ \frac{3}{4}f_{1}$
KNNN $ 0 $ $\frac{1}{2}$ $\frac{1}{2}f_{0}+ \frac{1}{2}f_{1}$
KNNN $1 $ $\frac{1}{2}$ $\frac{1}{6}f_{0}+ \frac{5}{6}f_{1}$
KNNNN $\frac{1}{2}$ $0$ $\frac{1}{3}f_{0}+ \frac{2}{3}f_{1}$
: Isospin composition of Kaon nucleon scattering amplitudes. $I_{tot}$ = total isospin, $I_{nucl}$ = isospin of nucleons, $f_{i}$ = KN amplitudes of isospin, $i$.
Three nucleons, S-wave interactions
===================================
The energy eigenvalue is obtained by the simultaneous solution of three equations $$\label{3a}
\psi^s_1+ G^{ss}_{1,2}f^s \psi_2 + G^{ss}_{1,3}f ^s \psi^s_3 =0$$ $$\label{3b}
\psi^s_2 + G^{ss}_{2,3}f ^s \psi^s_3 + G^{ss}_{2,1}f ^s \psi^s_1 =0$$ $$\label{3c}
\psi^s_3+ G^{ss}_{3,1}f ^s\psi^s_1 + G^{ss}_{3,2}f ^s \psi^s_2 =0,$$ which require the eigenvalue condition $$\label{d3s}
D_{3s}\equiv 1 - (f ^s)^2[ G^{ss}_{1,2} G^{ss}_{1,2}+ G^{ss}_{1,3}G^{ss}_{1,3}+G^{ss}_{3,2}G^{ss}_{3,2}] +2(f ^s)^3
G^{ss}_{1,2}G^{ss}_{2,3}G^{ss}_{3,1} =0.$$ This equation is to be solved numerically. A helpful guide to find the symmetry of two physically meaningful solutions is the situation of two equal $NN$ separations $ r_{12} = r_{13}$. Dropping the upper indices one obtains the factorized form $$\label{3d}
D_{3s}= (1 - f~ G_{1,2})( 1+ f~G_{1,2} - 2~f ^2~ G_{1,3}^2).$$ The first factor corresponds to an antisymmetric solution with the meson sticking to two nucleons only. The second factor generates a solution symmetric with the interchange of nucleons 1 and 2. These solutions are a direct continuation of the two solutions obtained in the $KNN$ systems.
[^1]: [email protected]
[^2]: [email protected]
[^3]: We thank Andrzej Deloff for supplying amplitudes of the DD model
|
{
"pile_set_name": "ArXiv"
}
|
=6truept 0truept 0truept
[**[Chromospheric activity of ROSAT discovered weak-lined T Tauri stars]{}**]{}
[*[ D. Montes$^{1,2}$, L.W. Ramsey$^{1}$ ]{}*]{}
$^1$ [The Pennsylvania State University, Department of Astronomy and Astrophysics, 525 Davey Laboratory, University Park, PA 16802, USA]{}\
$^2$ [Departamento de Astrofísica, Facultad de Físicas, Universidad Complutense de Madrid, E-28040 Madrid, Spain]{}
To be published in ASP Conf. Ser., Solar and Stellar Activity: Similarities and Differences (meeting dedicated to Brendan Byrne, Armagh 2-4th September 1998) C.J. Butler and J.G. Doyle, eds
=0.6truecm
------------------------------------------------------------------------
[**Abstract**]{}
We have started a high resolution optical observation program dedicated to the study of chromospheric activity in weak-lined T Tauri stars (WTTS) recently discovered by the ROSAT All-Sky Survey (RASS). It is our purpose to quantify the phenomenology of the chromospheric activity of each star determining stellar surface fluxes in the more important chromospheric activity indicators (Ca [ii]{} H & K, H$\beta$, H$\alpha$, Ca [ii]{} IRT) as well as obtain the Li [i]{} abundance, a better determination of the stellar parameters, spectral type, and possible binarity. With this information we can study in detail the flux-flux and rotation-activity relations for this kind of objects and compare it with the corresponding relations in the well studied RS CVn systems.
A large number of WTTS have been discovered by the RASS in and around different star formation clouds. Whether these stars are really WTTS, or post-TTS, or even young main sequence stars is a matter of ongoing debate. However, we have centered our study only on objects for which very recent studies, of Li [i]{} abundance (greater than Pleiads of the same spectral type) or radio properties, clearly confirmed their pre-main sequence (PMS) nature.
In this contribution we present preliminary results of our January 1998 high resolution echelle spectroscopic observations at the 2.1m telescope of the McDonald Observatory. We have analysed, using the spectral subtraction technique, the H$\alpha$ and Ca [ii]{} IRT lines of six WTTS (RXJ0312.8-0414NW, SE; RXJ0333.1+1036; RXJ0348.5+0832; RXJ0512.0+1020; RXJ0444.9+2717) located in and around the Taurus-Auriga molecular clouds. A broad and variable double-picked H$\alpha$ emission is observed in RXJ0444.9+2717. Emission above the continuum in H$\alpha$ and Ca [ii]{} IRT lines is detected in RXJ0333.1+1036 and a filling-in of these lines is present in the rest of the stars. Our spectral type and Li [i]{} EW deterninations confirm the PMS nature of these objects.
------------------------------------------------------------------------
------------------------------------------------------------------------
[**Introduction**]{}
Weak-lined T Tauri stars (WTTS) are low mass pre-main sequence stars (PMS) with H$\alpha$ equivalent widths $\leq$ 10 [Å]{} in which no signs of accretion are observed. The emission spectrum of these stars is not affected by the complications of star-disk interaction which often masks the underlying absorption lines as well as extincts the stellar light in classical T Tauri stars (CTTS). The WTTS are thus ideal targets to study the behavior of surface activity in the PMS stage of the stellar evolution. While there are a large number of studies in UV, X-ray and radio wavelengths, little research has been directed towards the study of the chromospheric activity using optical observations. Those which have been done are based on low resolution spectroscopic observations. Only some recent higher resolution studies centered in bona-fide WTTS in Taurus are available (see Feigelson et al. 1994; Welty 1995; Welty & Ramsey 1995, 1998; Poncet et al. 1998; Montes & Miranda 1999). In order to improve the knowledge of the WTTS chromospheres high resolution optical observations are needed. The WTTS discovered very recently by the ROSAT All-Sky Survey (RASS) are good targets to accomplish these objectives. A large number of them have been found far away from the star formation clouds (Neuhäuser et al. 1995; Alcalá et al. 1995, 1996; Wichmann et al. 1996; Magazzù et al. 1997; Krautter et al. 1997). Whether these stars are really WTTS, or post TTS, or even young main sequence stars is a matter of ongoing debate (Feigelson 1996, Briceño et al. 1997, Favata et al. 1997). However, we will study only those in the Taurus Auriga Molecular Cloud for which very recent studies clearly confirmed their PMS nature.
In this contribution we present preliminary results of our high resolution echelle spectroscopic observations of RX J0312.8-0414NW, SE; RX J0333.1+1036; RX J0348.5+0832; RX J0512.0+1020; and RX J0444.9+2717.
------------------------------------------------------------------------
------------------------------------------------------------------------
[**Observations**]{}
The spectroscopic observations were obtained during a 10 night run 12-21 January 1998 using the 2.1m telescope at McDonald Observatory and the Sandiford Cassegrain Echelle Spectrograph (McCarthy et al. 1993). This instrument is a prism crossed-dispersed echelle mounted at the Cassegrain focus and it is used with a 1200X400 Reticon CCD. The spectrograph setup was chosen to cover the H$\alpha$ (6563 Å) and Ca [ii]{} IRT (8498, 8542, 8662 Å) lines. The wavelength coverage is about 6400-8800Å$\ $ and the reciprocal dispersion ranges from 0.06 to 0.08 Å/pixel. The spectral resolution, determined by the FWHM of the arc comparison lines, ranges from 0.13 to 0.20 Å$\ $ (resolving power R=$\lambda$/$\Delta\lambda$ of 50000 to 31000) in the H$\alpha$ line region. In one of the nights we changed the spectrograph setup to include the He[i]{} D$_{3}$ (5876 Å) and Na [i]{} D$_{1}$ and D$_{2}$ (5896, 5890 Å) with a wavelength coverage of 5600-7000 Å. The spectra have been extracted using the standard reduction procedures in the IRAF package (bias subtraction, flat-field division, and optimal extraction of the spectra). The wavelength calibration was obtained by taking spectra of a Th-Ar lamp. Finally, the spectra have been normalized by a low-order polynomial fit to the observed continuum. The chromospheric contribution in these features is determined using the spectral subtraction technique (Huenemoerder & Ramsey 1987; Montes et al. 1995; 1997). The synthesized spectrum was constructed using the program STARMOD developed at Penn State (Barden 1985).
------------------------------------------------------------------------
------------------------------------------------------------------------
[**Results**]{}
We have analysed the H$\alpha$ and Ca [ii]{} IRT lines of six WTTS (RX J0312.8-0414NW, SE; RX J0333.1+1036; RX J0348.5+0832; RX J0512.0+1020; RX J0444.9+2717) located in and around the Taurus-Auriga molecular clouds. These targets were selected from two sources:
\(1) From the ROSAT detected late-type stars south of the Taurus (Neuhäuser et al. 1995; Magazzù et al. 1997 (hereafter M97)) we selected the stars that the spectroscopic studies of M97 and Neuhäuser et al. (1997, hereafter N97) clearly identified as WTTS from their greater Li abundance than Pleiades of the same spectral type. Some of them have been classified by these authors as single- and double-lined spectroscopic binaries (SB1, and SB2) and others are visual binaries (Sterzik et al. 1997, hereafter S97).
\(2) From the list of Wichmann et al. (1996) (hereafter W96) of new WTTS stars in Taurus, we selected the stars in which radio emission was detected by Carkner et al. (1997, hereafter C97) supporting their identification as genuine WTTS rather than ZANS.
Representative spectra of these stars are plotted in Fig. 1 (H$\alpha$), Fig. 2 (Li [i]{} 6708 Å), and Fig. 3 (Ca [ii]{} IRT). A K1V reference star is also plotted for comparison. The observed and subtracted spectra for the case of RX J0444.9+2717 are plotted in Fig. 4.
------------------------------------------------------------------------
------------------------------------------------------------------------
[****]{}
RX J0312.8-0414 is a visual binary with components NW and SE separated by 14” (M97, S97). The NW component is a G0V with [*v*]{}sin[*i*]{} = 33 km s$^{-1}$ and is a SB2. The SE component is a G8V with [*v*]{}sin[*i*]{} = 11 km s$^{-1}$. Both components exhibit H$\alpha$ absorption with a EW(H$\alpha$) of 3.5 and of of 2.5 Å respectively (M97; N97).
Our spectra exhibit a strong Li [i]{} 6708 Å line confirming the PMS nature of these objects. However, the level of chromospheric activity is very low, only a small filling-in of H$\alpha$ and Ca [ii]{} IRT lines is detected, in agreement with the earlier spectral type of both stars.
[****]{}
This star is classified as a confirmed PMS star by M97 and N97 on the basis of its Li [i]{} abundance, however, C97 detect no radio emission. M97 give a spectral type of K3 and observed the H$\alpha$ line in emission with a EW of -0.8 Å. N97 measured a [*v*]{}sin[*i*]{} of 20 km s$^{-1}$
Emission above the continuum in H$\alpha$ and Ca [ii]{} IRT lines is detected in our five spectra from January 14 to January 20 1998 with small variations from night to night.
[****]{}
This PMS is a rapidly-rotating star ([*v*]{}sin[*i*]{} = 127 km s$^{-1}$, N97) of spectral type G7 and with a small emission in the H$\alpha$ line (EW = -0.1 Å, M97).
In our five spectra (from 01/12/98 to 01/18/98) we observe the H$\alpha$ line always in absorption, but filled in. The Ca [ii]{} IRT lines are also filled in by chromospheric emission.
[****]{}
M97 give a spectral type K2 for this star and observed a small emission in the H$\alpha$ line (EW = -0.1 Å). N97 measured a rotational velocity of 57 km s$^{-1}$.
In our three spectra (from 01/14/98 to 01/17/98) we observe a variable filling-in of the H$\alpha$ and Ca [ii]{} IRT lines. The H$\alpha$ line shows emission in the blue wing in one of the spectra.
------------------------------------------------------------------------
------------------------------------------------------------------------
[****]{}
This is a K1 star with H$\alpha$ emission above the continuum (EW = -2.1 Å) and classified as a PMS star by W96 on the basis of its Li [i]{} abundance. The detection of radio emission by C97 confirm its PMS nature. Kohler & Leinert (1998) found a IR companion with a separation of 1.754” and a brignness ratio at K of 0.102.
We have eight spectra of this star available (from 01/12/98 to 01/20/98). The observed spectra are well matched using a K1V reference star with a rotational broadening of 65 km s$^{-1}$. Some of the more intense photospheric lines exhibit a flat-bottomed core (i.e. the core is noticeable filled in with respect to the reference profile) as is observed in other rapidly-rotating and spotted stars. A broad and variable double-picked H$\alpha$ emission above the continuum is observed (see Fig. 4). The H$\alpha$ EW in the observed spectra changes from -1.2 Å to -2.6 Å. The Ca [ii]{} IRT lines exhibit a strong filling-in.
------------------------------------------------------------------------
Alcalá, J. M., et al. 1995, A&AS, 114, 109
Alcalá, J. M., et al. 1996, A&AS, 119, 7
Briceño, C., Hartmann, L. W., Stauffer, J. R., et al. 1997, AJ, 113, 740
Barden, S. C. 1985, ApJ, 295, 162
Carkner, L., Mamajek, E., Feigelson, E. D., et al. 1997, ApJ, 490, 735
Favata, F., Micela, G., Sciortino, S. 1997, A&A, 326, 647
Feigelson, E. D., et al. 1994, , 432, 373
Feigelson, E. D. 1996, ApJ, 468, 306
Huenemoerder, D. P., & Ramsey, L. W. 1987, ApJ, 319, 392
Kohler, R. & Leinert, C. 1998, , 331, 977
Krautter, J., 1997, A&AS, 123, 329
Magazzù, A., Martín, E. L., Sterzik, M. F., et al. 1997, A&AS, 124, 449
McCarthy, J. K., Sandiford, B. A., Boyd, D., & Booth, J. 1993, PASP, 105, 881
Montes, D., Fernández-Figueroa, M. J., De Castro, E., & Cornide, M. 1995, A&A, 294, 165
Montes, D., Fernández-Figueroa, M. J., De Castro, E., & Sanz-Forcada J. 1997, A&AS, 125, 263
Montes, D., Miranda, L. F. 1999, , in preparation
Neuhäuser, R., Sterzik, M. F., Torres, G., & Martín, E. L. 1995, A&A, 299, L13
Neuhäuser, R., Torres, G., Sterzik, M. F., & Randich, S. 1997, A&A, 325, 647
Poncet, A., Montes, D., Fernández-Figueroa, M. J., & Miranda, L. F. 1998, in ASP Conf. Ser. 155, Cool Stars, Stellar Systems, and the Sun, 10th Cambridge Workshop, eds. R. A. Donahue & J. A. Bookbinder (San Francisco: ASP), CD-1772
Sterzik, M. F., Durisen, R. H., Brandner, W., et al. 1997, AJ, 114, 1673
Welty, A. D., 1995, , 110, 776
Welty, A. D., & Ramsey, L. W. 1995, AJ, 110, 336
Welty, A. D., & Ramsey, L. W. 1998, AJ, in press
Wichmann, R., Krautter, J., Schmitt, J. H. M. M., et al. 1996, A&A, 312, 439
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'J. M. Pittard, and S. M. Dougherty'
title: Particle acceleration in the colliding winds binary WR140
---
Introduction {#sec:intro}
============
WR140 (HD193793) is the archetype of long-period CWB systems. It consists of a WC7 star and an O4-5 star in a highly eccentric orbit ($e \approx 0.88$), and exhibits dramatic variations in its emission from near-IR to radio wavelengths [@Williams:1990; @White:1995], and also at X-ray energies [@Zhekov:2000; @Pollock:2002; @Pollock:2005] during its 7.9-year orbit. The variability appears to be linked to the WCR, which experiences significant changes as the stellar separation varies between 2 and 30 AU. The orbit modulation of the synchrotron flux has yet to be understood, but is likely due to a number of mechanisms. The changing free-free opacity along the line of sight through the extended stellar winds is certain to play a role, as will the strong inverse Compton cooling of the NT electrons. In addition, the intrinisic synchrotron luminosity and the spectral index of the NT electron energy distribution may alter. Recently developed models of CWBs that are based on hydrodynamical simulations of the stellar winds and the WCR have provided a more accurate representation of the thermal and NT emission [@Dougherty:2003; @Pittard:2006], and are a first step towards distentangling these mechanims and ultimately understanding the acceleration processes in detail.
Due to the non-unique solutions which can arise from these models, it is essential that observations across a broad energy spectrum are available to provide as many contraints as possible. Fortunately, recent observations of WR140 with the VLBA enable a full orbit definition, including inclination, along with robust distance and luminosity estimates [@Dougherty:2005]. Together, these parameters provide essential constraints that are currently unavailable for any other wide CWB. However, thermal X-ray and NT X-ray and $\gamma$-ray observations are also critical to establish some of the model parameters.
![Model fits to radio data of WR140 at orbital phase 0.837. In model A $\eta=0.22$, while model B has $\eta=0.02$. The latter is preferred (see text for details).[]{data-label="fig:radioAB"}](wr140_spec_phi0.837_ffabs.eps){width="\columnwidth"}
Modelling the Non-thermal Emission {#sec:models}
==================================
We have applied our newly developed radiative transfer models of CWBs to WR140 in order to investigate the emission and absorption processes which act to govern the radio variations. Full details can be found in @Pittard:2006b. Rather than attempting to model the radio lightcurve in one step, our first aim was to obtain good fits to the radio data at phase 0.837. This phase was chosen on the basis that the observed radio emission is close to maximum, that an X-ray spectrum exists, and that orbital-induced curvature of the WCR is negligible. The thermal X-ray flux is used to obtain a family of solutions with varying WR and O-star mass-loss rates as a function of wind momentum ratio, $\eta$. With these constraints, fits to the radio emission (see Fig. \[fig:radioAB\]) lead us to favour models with a relatively small value of $\eta$ (e.g. 0.02), on the basis that the required line-of-sight angle is more consistent with the orbital solution of @Dougherty:2005. Unfortunately, the VLBA images do not have the sensitivity to directly constrain the value of $\eta$ due to the rapid decline in the surface brightness of the WCR with off-axis distance. In these models, free-free absorption is responsible for the turnover between 1.6 and 5 GHz - the Razin effect can be ruled out because it places an unacceptably large fraction of energy into NT electrons. The post-shock B-field at the apex of the WCR is estimated to be about 1 G, while somewhat less than 1% of the total available energy is transferred to NT electrons.
A key finding is that the slope of the energy spectrum of the NT electrons, $p$, is flatter than the canonical value anticipated for diffusive shock acceleration (i.e. $p<2$). Several mechanisms can produce such distributions. The most likely are re-acceleration at multiple shocks within the WCR [e.g. @Schneider:1993], and second-order Fermi acceleration resulting from magnetic scattering off turbulent cells [e.g. @Scott:1975; @Dolginov:1990]. Both of these processes could occur together if the WCR is highly structured, as occurs naturally when clumps within the winds impact the WCR and introduce vorticity [@Pittard:2007]. Magnetic reconnection, perhaps through resistive MHD, is another possibility. Reconnection probably occurs throughout the volume of the turbulent WCR, and not just at a hypothetical contact discontinuity, and may provide additional energy for generating and maintaining the magnetic fluctuations which drive stochastic acceleration.
A remaining problem is that the B-field and $p$ are somewhat ill constrained by the radio data alone. However, their degeneracy is broken at $\gamma$-ray energies, and thus tighter constraints can be made if future observations with GLAST are utilized. While we generally predict lower $\gamma$-ray fluxes than previous works, detectability should not be an issue. However, we conclude that WR140 is unlikely to be detected as a TeV source with VERITAS-4, though it may be brighter at phases closer to periastron. Since the high stellar photon fluxes prevent the acceleration of electrons beyond Lorentz factors $\gamma \gtsimm
10^{5}-10^{6}$, TeV emission from CWB systems will provide unambiguous evidence of pion-decay emission from accelerated ions.
Future Directions {#sec:future}
=================
While these models have provided key insights into the particle acceleration process(es) occuring in WR140, much work remains. In forthcoming work we will apply our model to other orbital phases, and will investigate the effects of particle acceleration [*within*]{} the WCR.
Dolginov, A. Z., & Silant’ev, N. A. 1990, A&A, 236, 519 Dougherty, S. M., Beasley, A. J., Claussen, M. J., Zauderer, B. A., & Bolingbroke, N. J. 2005, ApJ, 623, 447 Dougherty, S. M., Pittard, J. M., Kasian, L., Coker, R. F., Williams, P. M., & Lloyd, H. M. 2003, A&A, 409, 217 Pittard, J. M. 2007, submitted to ApJL Pittard, J. M., & Dougherty, S. M. 2006, MNRAS, 372, 801 Pittard, J. M., Dougherty, S. M., Coker, R. F., O’Connor, E., & Bolingbroke, N. J. 2006, A&A, 446, 1001 Pollock, A. M. T., Corcoran, M. F., & Stevens, I. R 2002, in Moffat A. F. J., St-Louis N., eds, ASP. Conf. Ser. 260, 537 Pollock, A. M. T., Corcoran, M. F., Stevens, I. R, & Williams, P. M. 2005, ApJ, 629, 482 Schneider, P. 1993, A&A, 278, 315 Scott, J. S., & Chevalier, R. A. 1975, ApJL, 197, L5 White, R. L., & Becker, R. H. 1995, ApJ, 451, 352 Williams, P. M., van der Hucht, K. A., Pollock, A. M. T., Florkowski, D. R., van der Woerd, H., & Wamsteker, W. M. 1990, MNRAS, 243, 662 Zhekov, S. A., & Skinner, S. L. 2000, ApJ, 538, 808
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We provide a classification of compact Euclidean submanifolds $M^n\subset{{\mathbb{R}}}^{n+2}$ with nonnegative sectional curvature, for $n\ge 3$. The classification is in terms of the induced metric (including the diffeomorphism classification of the manifold), and we study the structure of the immersions as well. In particular, we provide the first known example of a nonorientable quotient $({\mathbb{S}}^{n-1}\times{\mathbb{S}}^1)/{{{\mathbb{Z}}}_2}\subset{{\mathbb{R}}}^{n+2}$ with nonnegative curvature. For the 3-dimensional case, we show that either the universal cover is isometric to ${\mathbb{S}}^2\times{{\mathbb{R}}}$, or $M^3$ is diffeomorphic to a lens space, and the complement of the (nonempty) set of flat points is isometric to a twisted cylinder $(N^2\times{{\mathbb{R}}})/{{\mathbb{Z}}}$. As a consequence we conclude that, if the set of flat points is not too big, there exists a unique flat totally geodesic surface in $ M^3$ whose complement is the union of one or two twisted cylinders over disks.'
address:
- 'IMPA: Est. Dona Castorina 110, 22460-320, Rio de Janeiro, Brazil'
- 'University of Pennsylvania: Philadelphia, PA 19104, USA'
author:
- 'Luis A. Florit'
- Wolfgang Ziller
title: |
Nonnegatively curved Euclidean submanifolds\
in codimension two
---
[^1]
It is well known that a compact immersed positively curved hypersurface in Euclidean space is diffeomorphic to a sphere (and is in fact the boundary of a convex body), the diffeomorphism simply given by the Gauss map. A deeper result, that such a submanifold still has to be homotopy equivalent to a sphere in codimension 2, is due to A. Weinstein [@We] and D. Moore [@m2]. However, the situation where the sectional curvature of the submanifold is only nonnegative is more delicate. Although it is still not known whether there exists an isometric immersion of ${\mathbb{R\mkern1mu P}}^2$ into ${{\mathbb{R}}}^4$ with nonnegative curvature, the higher dimensional problem was studied in [@bm1; @bm2]. Using recent results about the Ricci flow, we strengthen these by showing the following.
\[a\] Let $f:M^n\to{{\mathbb{R}}}^{n+2}$, $n\geq 3$, be an [isometric immersion ]{}of a compact Riemannian manifold with nonnegative sectional curvature. Then one of the following holds:
- $M^n$ is diffeomorphic to ${\mathbb{S}}^n$;
- $M^n$ is isometric to a product metric on ${\mathbb{S}}^k\times{\mathbb{S}}^{n-k}$ for some $2\le k\le n-2$, and $f$ is the product embedding of two convex Euclidean hypersurfaces;
- $M^n$ is isometric to $({\mathbb{S}}^{n-1}\times{{\mathbb{R}}})/\Gamma$ with a product metric on ${\mathbb{S}}^{n-1}\times{{\mathbb{R}}}$ and $\Gamma\simeq{{\mathbb{Z}}}$ acting isometrically. As a manifold, $M^n$ is diffeomorphic to ${\mathbb{S}}^{n-1}\times{\mathbb{S}}^1$ if orientable, or diffeomorphic to the nonorientable quotient $({\mathbb{S}}^{n-1}\times{\mathbb{S}}^1)/\Delta{{\mathbb{Z}}}_2$ otherwise, where $\Delta{{\mathbb{Z}}}_2$ denotes some diagonal ${{\mathbb{Z}}}_2$ action;
- $M^n$ is diffeomorphic to a 3 dimensional lens space ${\mathbb{S}}^3/{{\mathbb{Z}}}_k$.
Not only is this statement stronger than previously known results, but thanks to the use of the Ricci flow its proof becomes substantially simpler as well. More importantly, in the process we will also provide new strong restrictions on the structure of these submanifolds. In fact, the particular nature of the submanifolds in case $(c)$, and even if a nonorientable immersion exists, was unsettled in the literature. Furthermore, if case $(d)$ is possible at all was never discussed. The main purpose of our paper is to address both issues.
Regarding case $(a)$, there are plenty of immersions of ${\mathbb{S}}^n$ with nonnegative curvature in codimension two. For example, take any compact convex hypersurface and then a composition with a flat (not necessarily complete) hypersurface. On the other hand, the submanifolds in case $(b)$ are isometrically rigid.
It was the study of the structure of the immersions that led us to the following non-orientable example in case $(c)$.
[*Example 1\[ex1\]: The nonorientable quotient $({\mathbb{S}}^{n-1}\times{\mathbb{S}}^1)/\Delta{{\mathbb{Z}}}_2$ embedded in ${{\mathbb{R}}}^{n+2}$.*]{}
Consider a flat strip isometrically immersed in ${{\mathbb{R}}}^3$ $$\beta:{{\mathbb{R}}}\times(-\epsilon,\epsilon)\to{{\mathbb{R}}}^3$$ such that $ \beta(x_0+1,-x_1) = \beta(x_0,x_1).$ The image is a flat Moebius band immersed or embedded in ${{\mathbb{R}}}^3$. Large families of analytic Moebius bands of this type, together with some classification results, have been given in [@ck; @sc; @wu].
The product immersion of $\beta$ with the identity gives a flat hypersurface $$h=\beta\times\operatorname{Id}_{{{\mathbb{R}}}^{n-1}}:{{\mathbb{R}}}\times(-\epsilon,\epsilon)\times{{\mathbb{R}}}^{n-1}
\to{{\mathbb{R}}}^3\times{{\mathbb{R}}}^{n-1}={{\mathbb{R}}}^{n+2}.$$ Using a convex hypersurface $g\colon N^{n-1}\cong{\mathbb{S}}^{n-1}\to (-\epsilon,\epsilon)\times{{\mathbb{R}}}^{n-1}$ invariant under the reflection in the first coordinate of ${{\mathbb{R}}}^{n-1}$, we define the cylinder $\operatorname{Id}_{{\mathbb{R}}}\times g\colon{{\mathbb{R}}}\times
N^{n-1}\to{{\mathbb{R}}}\times(-\epsilon,\epsilon)\times{{\mathbb{R}}}^{n-1}$. The composition $$f=h\circ(\operatorname{Id}_{{\mathbb{R}}}\times g)\colon {{\mathbb{R}}}\times N^{n-1}\to {{\mathbb{R}}}^{n+2}$$ then satisfies $f\circ\tau=f$, where $\tau$ is the map $\tau(x_0,x_1,x_2,\dots,x_n)=(x_0+1,-x_1,x_2,\dots,x_n)$. Hence the image of $f$ is isometric to the [*twisted cylinder*]{} $({{\mathbb{R}}}\times N^{n-1})/{{\mathbb{Z}}}$, where ${{\mathbb{Z}}}$ is generated by the orientation reversing isometry of $N^{n-1}$ induced by $\tau$. Thus $h\circ f$ descends to the desired immersion $$f':({\mathbb{S}}^1\times N^{n-1})/\Delta{{\mathbb{Z}}}_2\simeq ({{\mathbb{R}}}\times N^{n-1})/{{\mathbb{Z}}}\to{{\mathbb{R}}}^{n+2}.$$ Observe that we can also choose $\beta$ as a cylinder and obtain immersions of $(N^{n-1}\times{{\mathbb{R}}})/{{\mathbb{Z}}}\simeq {\mathbb{S}}^{n-1}\times{\mathbb{S}}^1$ which are not products of immersions as in part $(b)$. Actually, they are not even locally product of immersions, as is the case in Example 2 below.
We say that an isometric immersion $f$ is a [*composition (of $j$)*]{} when $f=h\circ j$, where $j:M^n\to N_0^{n+1}$ is an isometric immersion into a (not necessarily complete) flat Euclidean hypersurface $h\colon N_0^{n+1}\to {{\mathbb{R}}}^{n+2}$. We will see that all submanifolds in [Theorem \[a\]]{} $(c)$ are almost everywhere compositions of a cylinder over a strictly convex Euclidean hypersurface when the Ricci curvature is 2-positive. For complete simply connected nowhere flat nonnegatively curved manifolds which split off a line, this structure of a composition was shown in [@bdt] under certain regularity assumptions. It is not clear though what additional restrictions hold, or whether regularity is necessary, if the immersion $f$ descends to a compact quotient, as required in .
The following example, where the induced intrinsic metric is well known in the theory of graph manifolds, illustrates the structure necessary for case $(d)$ to occur.
[*Example 2: The switched ${\mathbb{S}}^3$ in ${{\mathbb{R}}}^5$.*]{}
Consider a closed strictly convex hemisphere inside a closed halfspace, $${\mathbb{S}}^2_+\subset {{\mathbb{R}}}^3_+={{\mathbb{R}}}^2\times {{\mathbb{R}}}_+,$$ such that the boundary is a closed geodesic in ${\mathbb{S}}^2_+$ along which the Gauss curvature vanishes to infinite order, and with image contained in ${{\mathbb{R}}}^2\times\{0\}$. Its product with another ${\mathbb{S}}^1\subset{{\mathbb{R}}}^2$ gives the nonnegatively curved three manifold $$N^3_+={\mathbb{S}}^2_+\times{\mathbb{S}}^1\subset{{\mathbb{R}}}^5_+={{\mathbb{R}}}^2\times{{\mathbb{R}}}_{\geq 0}\times {{\mathbb{R}}}^2,$$ whose boundary is the totally geodesic flat torus $$T^2={\mathbb{S}}^1\times{\mathbb{S}}^1\subset {{\mathbb{R}}}^2\times\{0\}\times{{\mathbb{R}}}^2={{\mathbb{R}}}^4\subset{{\mathbb{R}}}^5.$$ Now, reversing the role of the ${{\mathbb{R}}}^2$ factors we similarly construct another manifold $N^3_-={\mathbb{S}}^1\times{\mathbb{S}}^2_-\subset {{\mathbb{R}}}^5_-={{\mathbb{R}}}^2\times {{\mathbb{R}}}_{\leq 0}\times
{{\mathbb{R}}}^2$ with the same boundary $T^2\subset{{\mathbb{R}}}^4$. Thus $$M^3=N^3_+\cup N^3_-\subset{{\mathbb{R}}}^5$$ gives an embedded Euclidean submanifold with nonnegative sectional curvature:
{width="30.00000%"}
A switched ${\mathbb{S}}^3$ isometrically embedded in ${{\mathbb{R}}}^5$
As a smooth manifold $M^3$ is diffeomorphic to the 3-sphere since this is the usual description of $\ {\mathbb{S}}^3$ as the union of 2 solid tori. The set of flat points of $M^3$ contains the totally geodesic flat torus $T^2\subset M^3$ which disconnects $M^3$. Moreover, each connected component of $M^3\setminus T^2$ is obtained as a composition of the cylinder ${{\mathbb{R}}}\times {\mathbb{S}}^2_+\subset{{\mathbb{R}}}^4$ with a local [isometric immersion ]{}from ${{\mathbb{R}}}^4$ into ${{\mathbb{R}}}^5$. By replacing the torus $T^2$ with a small flat cylinder $C_\epsilon={\mathbb{S}}^1\times[0,\epsilon]\times{\mathbb{S}}^1\subset {{\mathbb{R}}}^5$, we obtain $${\mathbb{S}}^3\cong M^3_\epsilon=N^3_+\cup C_{\epsilon}\cup N^3_-\subset{{\mathbb{R}}}^5.$$
The resulting embeddings $M^3_{\epsilon}\subset{{\mathbb{R}}}^5$ for ${\epsilon}\geq 0$ have thus the following properties:
- $M^3_\epsilon$ is compact with nonnegative sectional curvature;
- The set of flat points contains a totally geodesic flat torus that disconnects $M^3_\epsilon$;
- $M^3_\epsilon$ is locally an isometric product of two manifolds for ${\epsilon}>0$ (almost everywhere for ${\epsilon}=0$);
- The embedding $M^3_\epsilon\subset{{\mathbb{R}}}^5$ is locally, yet not globally, a composition of a cylindrical hypersurface for ${\epsilon}>0$ (almost everywhere for ${\epsilon}=0$);
- $M^3_\epsilon$ has no points with positive Ricci curvature;
- $M^3_\epsilon$ has 2-positive Ricci curvature outside the set of flat points.
The submanifold $M^3_\epsilon$ can be changed further so that the boundary of the set of flat points is not regular, and arbitrarily complicated sets of flat points can be introduced in ${\mathbb{S}}^2_\pm$ as well. Observe that, by taking the two disks ${\mathbb{S}}^2_\pm$ with strictly convex boundaries as plane curves, all these properties can be achieved even with 1-regular full embeddings, i.e., with the dimension of the space spanned by the second fundamental form having constant dimension 2 everywhere.
The following result recovers some of this structure for any immersion of a lens space with nonnegative sectional curvature.
\[3a\] Let $f:M^3={\mathbb{S}}^3/{{\mathbb{Z}}}_k \to {{\mathbb{R}}}^5$ with $k>1$ be an isometric immersion with nonnegative sectional curvature. Then the set $M_0$ of flat points of $M^3$ is nonempty, and $M^3\setminus M_0$ is isometric to a twisted cylinder $(N^2\times{{\mathbb{R}}})/{{\mathbb{Z}}}$, where $N^2$ is a (not necessarily connected) surface with positive Gaussian curvature. Moreover, $f$ is a composition almost everywhere on $M^3\setminus M_0$.
The natural intrinsic problem raised by this result is to try to understand how twisted cylinders can be glued together with flat regions in order to build a compact manifold. One way is to glue the cylinders through compact totally geodesic hypersurfaces, as we did for $M^3_{\epsilon}$ in Example 2 above. In [@fz] we study this intrinsic structure under the additional assumption that the set of nonflat points $M\setminus M_0$ is dense and locally finite (without assuming that the curvature is necessarily nonnegative). The main result in [@fz] is that under this assumption there exist pairwise disjoint flat totally geodesic compact surfaces in $M^3$ whose complement is a union of twisted cylinders. We also give examples of metrics on $T^3$ which show that the assumption on the nonflat points is necessary. In the case of nonnegative curvature, Theorem B above, together with Theorem C in [@fz], implies that:
\[main2\] Let $f:M^3={\mathbb{S}}^3/{{\mathbb{Z}}}_k \to {{\mathbb{R}}}^5$, $k>1$, be an isometric immersion with nonnegative sectional curvature, such that $M^3\setminus M_0$ is dense and its number of connected components is locally finite. Then, either $M_0$ contains a unique flat totally geodesic torus $T^2$ that disconnects $M^3$ and $M^3=V_1\sqcup T^2 \sqcup V_2$, or $M^3$ is the lens space $L(4p,2p+1)$ and $M_0$ contains a unique flat totally geodesic Klein bottle $K^2$ satisfying $M^3=V_1\sqcup K^2$. In both cases, $V_i=(D_i^2\times{{\mathbb{R}}})/{{\mathbb{Z}}}$ is a twisted cylinder, where $D_i^2$ is a disc with nonnegative Gaussian curvature and boundary a closed geodesic.
Yet, despite the rigid structure described above, we still do not know if a nonnegatively curved lens space in ${{\mathbb{R}}}^5$ actually exists.
For the proof of [Theorem \[a\]]{} we will need a fundamental property of our submanifolds that was proved in [@Ch] and [@m2]. In order to state it, recall that we have the [*type numbers*]{} $\tau_k$ of the immersion: if $h_v={\langle}v,\cdot\,{\rangle}:M^n\to{{\mathbb{R}}}$ is a height function, and $\mu_k(v)$ the number of critical points of $h_v$ with index $k$, then $\tau_k$ is the average of $\mu_k(v)$, integrated over the unit sphere in ${{\mathbb{R}}}^{n+2}$. If $M^n$ is compact with nonnegative sectional curvature immersed in ${{\mathbb{R}}}^{n+2}$, then the type numbers satisfy $$\tau_0+\tau_n\ge \tau_1 + \dots + \tau_{n-1}.$$ Furthermore, if the inequality is strict, we will see in [Section \[RG\]]{} that $M^n$ is diffeomorphic to a sphere. We are thus mainly interested in the case of equality, and call such immersions [*wide*]{}. In the situation of [Theorem \[3a\]]{}, we will show that $\tau_k=\frac{1}{8\pi^2}\int_{N^2}K(x)\kappa_g(x)dx$ for $0\leq k\leq3$, where $\kappa_g(x)$ denotes the total curvature of the image of the ${{\mathbb{R}}}$ factor through $x\in N^2$, and $K$ the Gaussian curvature of $N^2$.
For codimension two Euclidean submanifolds with nonnegative curvature, A. Weinstein [@We] showed that at every point there exists a basis $\{\xi,\eta\}$ of the normal space such that its shape operators $A$ and $B$ are nonnegative, and hence the curvature operator is nonnegative. The crucial new property that we will prove is that for a wide immersion we have in addition that $$\text{either} \ \ker A\cap\ker B \neq 0, \ \ \text{or} \
\ \ker A\oplus \ker B = T_p M\ \, \text{with} \ A,B\neq 0,$$
and call the (not necessarily complete) immersions which satisfy only this property [*locally wide*]{}. These immersions already have strong rigidity properties. We examine their behavior on the open subsets $U_k = \{p\in M^n: {\ensuremath{\operatorname{rank}}}\, A(p) = n-k,\,{\ensuremath{\operatorname{rank}}}\, B(p)
= k \ \ \text{and} \ \ker A(p)\cap \ker B(p) = 0\}$, which fill out the complement of the points of positive relative nullity. We will show in particular that:
- If $2\leq k \leq [n/2]$, then $f|_{U_k}$ is locally a product immersion of two strictly convex Euclidean hypersurfaces;
- If $\pi:\tilde U_1\to U_1$ is the universal cover, then $(f\circ\pi)$ is a composition of a convex Euclidean hypersurface with constant index of relative nullity one (see [Theorem \[b1\]]{});
- If $M^n$ is complete and has no points of positive Ricci curvature, but has 2-positive Ricci curvature, then $\tilde M^n={\mathbb{S}}^{n-1}\times{{\mathbb{R}}}$, where $g:{\mathbb{S}}^{n-1}\to{{\mathbb{R}}}^n$ is a compact strictly convex hypersurface.
Preliminaries
=============
Let $f:M^n\to{{\mathbb{R}}}^{n+2}$, $n\geq 3$, be an [isometric immersion ]{}of a Riemannian manifold with nonnegative sectional curvature. A. Weinstein observed that, in this case, the curvature operator $\hat R$ is nonnegative. Indeed, if $\alpha$ is the second fundamental form, the Gauss equations imply that ${\langle}\alpha(X,X),\alpha(Y,Y){\rangle}\ge 0$ at every $p\in M$. Hence, in the connected subset $\{\alpha(X,X):X\in T_pM\}\subset
T^\perp_p M$ intersected with the unit circle all points have distance at most $\pi/2$. This implies that this set lies in the first quadrant with respect to some orthonormal basis $\{\xi,\eta\}$, while it lies in its interior if it has positive sectional curvature. We thus have the shape operators $$\label{AB}
A=A_{\xi} {\rm \ \ \ and \ \ \ } B=A_{\eta}, \ \ \
{\rm \ \ \ with\ \ \ } A\ge 0 {\rm \ \ \ and \ \ \ } B\ge 0.$$ since ${\langle}A_\tau(X),Y{\rangle}={\langle}\alpha(X,Y),\tau{\rangle}$.
By the Gauss equations we have $\hat{R}=\Lambda^2A+\Lambda^2B$, and hence $\hat{R}\ge 0$ (and $\hat{R}>0$ if the sectional curvature is positive). Furthermore, the cone of shape operators which are positive semidefinite contains the first quadrant, while the cone of shape operators which are negative semidefinite contains the third.
We will see that the interesting immersions are quite rigid as a consequence of the following purely algebraic key result which will be applied to the shape operators in .
\[l:eq\] Let $A$ and $B$ be nonnegative semidefinite self–adjoint operators on an Euclidean space $V$. Then, $|\det\left(A+tB\right)|\geq|\det\left(A-tB\right)|,\ \ \forall\, t>0.$ Moreover, equality holds for all $t>0$ if and only if either $\ker A \cap \ker B \neq 0$, or $\ker A \oplus \ker B=V$.
First, observe that the lemma follows easily if one of the operators is nonsingular. Indeed, if, say, $B$ is invertible, $\det\left(A+tB\right) = \det(B)\phi(t)$, where $\phi$ is the characteristic polynomial of the positive semidefinite operator $B^{-1/2}AB^{-1/2}$. Since all the coefficients of $\phi$ are nonnegative the inequality holds, with equality if and only if $A=0$. Therefore, we can assume that both $A$ and $B$ are singular.
Let $K=\ker B$, and decompose $A$ as its block endomorphisms $(C,D,D^t,E)$, according to the orthogonal decomposition $K\oplus K^\perp$, $$A(u,v) = (Cu+Dv,D^tu+Ev), \ \ \ \forall u\in K, v\in K^\perp.$$ Since $A\geq 0$, we have $C\geq0$ and $E\geq 0$.
For $u\in\ker C$ we get $0\leq{\langle}A(su,v), (su,v){\rangle}= 2s{\langle}D^tu,v{\rangle}+{\langle}Ev,v{\rangle}$, for all $s\in{{\mathbb{R}}}$. Hence, $D^tu=0$ and $u \in K \cap \ker A$. Therefore either $C>0$ or the kernels intersect and equality holds. So let us then assume further that $C>0$ and $K\cap \ker A = 0$.
If we call $\hat B$ the symmetric operator $B$ restricted and projected to $K^\perp$, then $\hat B>0$ and, for all $u\in K, v\in K^\perp$, $$(A+tB)(u,v) = (Cu+Dv,D^tu+(E+t\hat B)v).$$ Thus, $$\det(A+tB)=\det(C)\det\left(E+t\hat B-D^tC^{-1}D\right)
= \det(C)\det(\hat A+t\hat B),$$ where $\hat A$ is the symmetric operator on $K^\perp$ given by $\hat A=E-D^tC^{-1}D$. Now, observe that $\hat A$ is nonnegative since $${\langle}(E-D^tC^{-1}D)v,v{\rangle}= {\langle}Ev,v{\rangle}- {\langle}C^{-1}Dv,Dv{\rangle}= {\langle}A(-C^{-1}Dv,v),(-C^{-1}Dv,v){\rangle}\geq 0.$$ This gives the inequality in the lemma as in the nonsingular case. Moreover, this computation also shows that $v\in\ker \hat A$ if and only if $(-C^{-1}Dv,v)\in\ker A$. Thus, $\det(A+tB)$ is again, up to a positive constant, the characteristic polynomial of the positive semidefinite operator $\hat B^{-1/2}\hat A\hat B^{-1/2}$, which is an odd or even function if and only if $\hat A=0$. In this case, $\dim \ker A = \dim K^\perp$, as claimed.
The converse of the equality case is straightforward.
Since $\hat{R}\ge 0$ we can use the following rigidity result, where $hol$ denotes the Lie algebra of the holonomy group of $M$ (see Theorem 1.13 in [@Wi]):
\[rigidity\] If $M$ is a compact and simply connected Riemannian manifold with $\hat{R}\ge 0$, then one of the following holds:
- $hol={{\mathfrak{so}}}(n)$ and $M$ is diffeomorphic to ${\mathbb{S}}^n$, with the Ricci flow converging to a metric with constant positive curvature;
- $hol={{\mathfrak{u}}}(n)$ and $M$ is diffeomorphic to ${\mathbb{C\mkern1mu P}}^n$, with the Ricci flow converging to the Fubini Study metric;
- $M$ is isometric to an irreducible symmetric space $G/K$ with $hol={{\mathfrak{k}}}$;
- Arbitrary isometric products of the above cases.
Notice in particular that if $hol={{\mathfrak{so}}}(n)$ in the above, then $M$ is diffeomorphic to a sphere since, if the Lie algebra ${{\mathfrak{k}}}$ of $K$ for a symmetric space $G/K$ satisfies ${{\mathfrak{k}}}={{\mathfrak{so}}}(n)$, then $G/K$ is in fact isometric to a round $n$-sphere.
We will also need the next result due to Bishop ([@Bi]) about the holonomy group of a general compact Euclidean submanifold in codimension two.
\[bishop\] If $M^n$ is a compact submanifold immersed in ${{\mathbb{R}}}^{n+2}$, then either $n=4$ and $hol=\mathfrak{u}(2)$, or $hol={{\mathfrak{so}}}(k)\times{{\mathfrak{so}}}(n-k)$ for some $0\leq k \leq [n/2]$.
Here $k=0$ corresponds to $hol={{\mathfrak{so}}}(n)$ and $k=1$ to $hol={{\mathfrak{so}}}(n-1)$, i.e., the manifold splits a flat factor locally. In particular, there can be at most a one dimensional flat factor.
Rigidity {#RG}
========
In this section we provide the proof of [Theorem \[a\]]{}. In the process, we reprove, simplify and strengthen known results about compact Euclidean submanifolds in codimension 2 with nonnegative sectional curvature by using [Proposition \[rigidity\]]{}. Along the way, we will see how the proof implies further rigidity of the immersion.
Let $f$ and $M^n$ be as in [Theorem \[a\]]{}. Since $M^n$ is compact with nonnegative sectional curvature, a well-known consequence of the Cheeger–Gromoll splitting theorem, see [@CG] Theorem C, implies that its universal cover $\tilde{M}^n$ splits isometrically as $$\label{soul}
\tilde{M}^n={{\mathbb{R}}}^\ell\times N^{n-\ell},$$ where $N^{n-\ell}$ is compact and simply connected. Since the Lie algebra of the holonomy group of $M$ and $\tilde M$ coincide, [Proposition \[bishop\]]{} implies that $\ell\le 1$.
Now observe that in our context $n=4$ and $hol=u(2)$ is not possible in [Proposition \[bishop\]]{} since then the universal cover would be compact, and by [Proposition \[rigidity\]]{} diffeomorphic to ${\mathbb{C\mkern1mu P}}^2$. But ${\mathbb{C\mkern1mu P}}^2$ does not admit an immersion in ${{\mathbb{R}}}^6$ as follows, e.g., from $p_1({\mathbb{C\mkern1mu P}}^2)=3$, the product formula for Pontrjagin classes, and $p_1(E)=e(E)^2$ for the normal 2 plane bundle $E$ of the immersion.
Therefore, the existence of an [isometric immersion ]{}implies that $$\label{k}
hol={{\mathfrak{so}}}(k)\oplus{{\mathfrak{so}}}(n-k),$$ for some $0\leq k \leq [n/2]$. We can now combine , and [Proposition \[rigidity\]]{}. First, it follows that $\ell\le 1$ and $\ell=1$ if and only if $k=1$.
If $\ell=k=0$, then $\tilde{M}^n$ is compact and hence diffeomorphic to ${\mathbb{S}}^n$, while if $\ell=0$ and $k\ge 2$, then $\tilde{M}^n$ is diffeomorphic to ${\mathbb{S}}^k\times{\mathbb{S}}^{n-k}$. In the latter case, by [@am] the immersion must be a product immersion of convex hypersurfaces and, in particular, $M^n$ itself is simply connected and diffeomorphic to ${\mathbb{S}}^k\times{\mathbb{S}}^{n-k}$.
If $\ell=k=1$, and hence $hol={{\mathfrak{so}}}(n-1)$, implies that $\tilde{M}^n$ splits isometrically $\tilde{M}^n=N^{n-1}\times {{\mathbb{R}}}$, with $N^{n-1}$ compact and simply connected. But then $hol(N^{n-1})={{\mathfrak{so}}}(n-1)$, and again by [Proposition \[rigidity\]]{} we get that $N^{n-1}$ is diffeomorphic to ${\mathbb{S}}^{n-1}$. Thus $M^n=({\mathbb{S}}^{n-1}\times{{\mathbb{R}}})/\Gamma$ isometrically, with $\Gamma$ a discrete group acting by isometries. We will see in \[z\] below that $\Gamma={{\mathbb{Z}}}$ and that $M^n$ is diffeomorphic to either ${\mathbb{S}}^{n-1}\times{\mathbb{S}}^1$ if $M^n$ is orientable, or to the nonorientable quotient $({\mathbb{S}}^{n-1}\times{\mathbb{S}}^1)/\Delta{{\mathbb{Z}}}_2$ otherwise.
To conclude the proof of [Theorem \[a\]]{}, it remains to study the possibilities for the fundamental group when $\ell=k=0$ and $\tilde{M}^n={\mathbb{S}}^n$, or when $\ell=k=1$ and $\tilde{M}^n={\mathbb{S}}^{n-1}\times{{\mathbb{R}}}$ for some metrics of nonnegative curvature on the spheres. In order to do this, we first recall the relationship, due to Kuiper, between height functions and shape operators of compact Euclidean submanifolds.
Following [@m2], for each $v\in{\mathbb{S}}^{n+1}$ consider the height function $h_v={\langle}v,\cdot\,{\rangle}:M^n\to{{\mathbb{R}}}$ and define $$\mu_k(v)=\text{number of critical points of $h_v$ with index $k$},$$ and $$\tau_k=\frac{1}{vol({\mathbb{S}}^{n+1})}\int_{{\mathbb{S}}^{n+1}}\mu_k(v)dv.$$ Since on a set of full measure the height functions are nondegenerate, we can apply the usual Morse inequalities for Morse functions. After integrating them we get $$\tau_k\ge b_k$$ and $$\sum_{k=0}^{\ell}(-1)^{\ell-k}\,\tau_k\ge
\sum_{k=0}^{\ell}(-1)^{\ell-k}\, b_k,\ \ \ \forall\ \ell = 1,\dots n,$$ where $b_k=\dim H_k(M,F)$ for some field $F$.
Clearly, $p$ is a critical point of $h_v$ if and only if $v=\beta$ for some $(p,\beta)\in T_1^\perp M$, where $T_1^\perp M$ is the normal bundle of the immersion. Moreover, $p$ is a nondegenerate critical point of $h_\beta$ if and only if $A_\beta$ is nonsingular, and the index of the critical point is equal to $ind(A_\beta)$, the number of negative eigenvalues of $A_\beta$. We now use the Gauss map $G\colon T^\perp_1M\to {\mathbb{S}}^{n+1}$ with $G(p,\beta)=\beta$, which for compact submanifolds is surjective. Since $G_{*(p,\beta)}={\ensuremath{\operatorname{diag}}}(A_\beta,\operatorname{Id})$, we have that $p$ is a nondegenerate critical point of $h_\beta$ if and only if $(p,\beta)$ is a regular point of $G$. Thus if $C\subset
{\mathbb{S}}^{n+1}$ is the set of regular values of $G$ (an open and dense set of full measure), then $h_v$ with $v\in C$ is a Morse function and hence $\mu_k$ is constant on every connected component of $C$. Furthermore, for $v\in C$, the set $G^{-1}(v)$ contains $\mu_k(v)$ many points of index $k$. Thus by change of variables, $$\label{form}
\tau_k=\int_C\mu_k(v)dv=\frac{1}{vol({\mathbb{S}}^{n+1})}
\int_{ind(A_\beta)=k }{\left\vert\det A_\beta\right\vert}dvol_{T_1^\perp M}.$$ Since $\mu_k(v)=\mu_{n-k}(-v)$, we also have $\tau_{n-k}=\tau_k$ for all $k$.
We proceed by applying [Lemma \[l:eq\]]{} to the shape operators $A,B$ in in order to estimate ${\left\vert\det A_\beta\right\vert}$ for $(p,\beta)\in T_1^\perp M$, by first integrating over a circle in the normal space $T_p^\perp M$, and then over $p\in M$. For a fixed normal space $T_p^\perp M$, [Lemma \[l:eq\]]{} implies that $$\label{detcomp}
|\det\left(\cos\theta A+\sin\theta B\right)|\geq|\det\left(\cos\theta
A-\sin\theta B\right)| \ \ \text{ for all }\ 0 \le \theta\le
\frac{\pi}{2}.$$ Furthermore, if a regular value $\beta=\cos(\theta)\xi+\sin(\theta)\eta$ lies in the first quadrant, then , and if it lies in the third quadrant then $ind(A_\beta)=n$, whereas all saddle points (and possibly some local maxima and minima as well) lie in the second and fourth quadrants. Thus , together with $\tau_{n-k}=\tau_k$, implies that $$\label{chen}
\tau_0+\tau_n\ge \tau_1 + \dots + \tau_{n-1}.$$
As we will see, in the case of strict inequality $M$ is diffeomorphic to a sphere. We are thus interested from now on in the equality case, where one expects a certain amount of rigidity. This motivates our next key definition.
[*Definition.*]{} Given a compact Riemannian manifold $M^n$ with nonnegative sectional curvature, an isometric immersion $f:M^n\to{{\mathbb{R}}}^{n+2}$ is said to be [*[wide]{}*]{} if $\tau_0+\tau_n= \tau_1 + \dots + \tau_{n-1}$.
We can now use [Lemma \[l:eq\]]{} to express the rigidity in terms of the shape operators $A$ and $B$ in .
\[wide\] If $f \colon M^n\to {{\mathbb{R}}}^{n+2}$ is wide, then at each point $p\in M^n$ we have that either $\ker A\oplus \ker B = T_p M$ with $A\neq0$ and $B\neq0$, or $\ker A\cap\ker B \neq 0$.
Equality in holds if and only if we have equality in at every point and for every angle $\theta$. Thus [Lemma \[l:eq\]]{} implies that either $\ker A\oplus \ker B = T_p M$ or $\ker A\cap\ker B \neq 0$. Furthermore, if $A=0$ at some point, then $B$ is nonsingular, and hence in a neighborhood of the point as well. But then this neighborhood will not have any saddle points, and we would have a strict inequality in .
Notice that, along the (open) set where the first case holds, the orthonormal basis $\{\xi,\eta\}$ in , and hence $A$ and $B$, are unique and thus smooth. On the other hand, no such information can be obtained in the second case.
Altogether, we have the following improvement of a result in [@m2], where we obtain additional information on the structure of the immersion.
\[main\] If $M^n$ is compact, nonnegatively curved and immersed in Euclidean space in codimension $2$, then we have for any coefficient field $F$ over which $M^n$ is orientable that $$b_1+b_2+\dots +b_{n-1}\le 2,$$ with equality if and only if $$\tau_k=b_k \ \ \forall\ 2\leq k\leq n-2, \ \ \tau_1-\tau_0= b_1-1
\text{ and }\ \tau_0+\tau_n= \tau_1 + \dots + \tau_{n-1}.$$ In particular, equality implies that the immersion is [wide]{}.
By the Morse inequalities $\tau_k\ge b_k$ for $k=2,\dots n-2$ and $\tau_1-\tau_0\ge b_1-b_0$, i.e. $b_1\le \tau_1-\tau_0+1$. If $M^n$ is compact and orientable mod $F$, then $b_k=b_{n-k}$ and, in general, $\tau_k=\tau_{n-k}$. Hence, $$b_1+(b_2+\dots +b_{n-2}) +b_{n-1}\le \tau_1-\tau_0+1 +(\tau_2+\dots
+\tau_{n-2}) + \tau_{n-1}-\tau_n+1,$$ which together with yields $$b_1+b_2+\dots+b_{n-1}\le\tau_1\dots+\tau_{n-1}-(\tau_0+\tau_n)+2\le 2,$$ with equality as claimed.
So far we have shown that there are three situations: either $\tilde M^n
\cong {\mathbb{S}}^n$ diffeomorphically, or $\tilde M^n$ splits isometrically as either $\tilde M^n\equiv{\mathbb{S}}^{n-1}\times{{\mathbb{R}}}$ or $M^n\equiv{\mathbb{S}}^k\times{\mathbb{S}}^{n-k}$, for certain metrics of nonnegative curvature on the spheres. As a corollary of [Proposition \[main\]]{}, we conclude that in the first case $M$ itself is a sphere when $n\geq 4$:
\[sn\] If $n\ge 4$ and $\pi_1(M^n)$ is finite, then $M^n$ is simply connected. In particular, either $M^n$ is diffeomorphic to ${\mathbb{S}}^n$, or isometric to a Riemannian product , $2\leq k\leq
n-2$, for certain metrics of nonnegative sectional curvature on the spheres. In the latter case, the immersion is a product of two convex hypersurfaces, and thus it is wide.
Assume $M^n$ is not simply connected, and let $g\in\pi_1(M^n)$ be an element of prime order $q$, so ${\langle}g{\rangle}\cong{{\mathbb{Z}}}_q\subseteq \pi_1(M)$. Then, the lens space ${\mathbb{S}}^n/{{\mathbb{Z}}}_q$ is a finite cover of $M^n$ and with the covering metric it has an [isometric immersion ]{}into ${{\mathbb{R}}}^{n+2}$. But $b_k({\mathbb{S}}^n/{{\mathbb{Z}}}_q,{{\mathbb{Z}}}_q)=1$ for all $k$ which contradicts [Proposition \[main\]]{} if $n\geq 4$. Notice that if $q>2$ the lens space is orientable, and if $q=2$ it is orientable mod $2$.
In the case where $M^n$ is isometric to a Riemannian product , [@am] implies that the immersion is a product of two convex hypersurfaces. One easily sees that in this case the immersion is wide.
For the three dimensional case we have the following.
\[lens\] If $n=3$ and $\pi_1(M^3)\neq 0$ is finite, then $M^3$ is diffeomorphic to a lens space ${\mathbb{S}}^3/{{\mathbb{Z}}}_k$, and the immersion is wide with $\tau_0=\tau_1=\tau_2=\tau_3$. Furthermore, the height functions satisfy $\mu_0=\mu_1$ as well as $\mu_{2}=\mu_3$.
$M^3$ is again covered by a lens space with $b_1=b_2=1$ and we have equality in [Proposition \[main\]]{}, which implies that $\tau_0=\tau_1=\tau_2=\tau_3$ (they do not have to be $1$, though). The Morse inequality tells us that $\mu_1(v)-\mu_0(v)\geq b_1-b_0=0$ for a.e. $v\in{\mathbb{S}}^{n+1}\subset{{\mathbb{R}}}^{n+2}$, and since $\tau_1=\tau_0$, it follows that $\mu_1(v)=\mu_0(v)$. From $\mu_k(v) = \mu_{n-k}(-v)$ we get $\mu_3=\mu_2$ for a.e. $v$. Therefore for almost all $v$ the Morse function $h_v$ has the same number of critical points of index 0 and 1 (first for the lens space, and then for $M^3$ as well). But if this is the case, the cancellation result in [@mm] says that there exists another (abstract) Morse function on $M^3$ with only one critical point of index 0 and 1. In particular, this means that $M^3$ is a CW complex whose 1 skeleton is a circle, which implies by transversality that the map $\pi_1(S^1)\to \pi_1(M^3)$ induced by the inclusion is onto and hence $\pi_1(M^3)$ is finite cyclic.
For infinite fundamental group, we have:
\[z\] If $\pi_1(M^n)$ is infinite, then $\pi_1(M^n)\cong {{\mathbb{Z}}}$ is also cyclic. Moreover, $\tilde M^n = {\mathbb{S}}^{n-1}\times{{\mathbb{R}}}$ splits isometrically for some metric of nonnegative curvature in ${\mathbb{S}}^{n-1}$, and $M^n$ is diffeomorphic to ${\mathbb{S}}^{n-1}\times{\mathbb{S}}^1$ if $M^n$ is orientable, or to the nonorientable quotient $({\mathbb{S}}^{n-1}\times{\mathbb{S}}^1)/\Delta{{\mathbb{Z}}}_2$ otherwise. In addition, $\tau_0=\tau_1=\tau_{n-1}=\tau_n$, and $\tau_2=\cdots=\tau_{n-2}=0$, and hence the immersion is wide. Furthermore, the height functions satisfy $\mu_0=\mu_1$ as well as $\mu_{n-1}=\mu_n$, and $\mu_2= \cdots =\mu_{n-2}=0$.
We already saw that under this assumption, $M^n$ is diffeomorphic to $({\mathbb{S}}^{n-1}\times{{\mathbb{R}}})/\Gamma$ for some group $\Gamma$ acting properly discontinuously. Since the quotient is compact, there exists a subgroup $\Gamma'\simeq{{\mathbb{Z}}}$ which acts via translations on the ${{\mathbb{R}}}$ factor. Choose a covering $f\colon M^*\to M$ such that the image of the fundamental group under $f_*$ is equal to $\Gamma'$. Thus $b_1(M^*,{{\mathbb{Z}}}_2)=b_{n-1}(M^*,{{\mathbb{Z}}}_2)=1$, and we can apply [Proposition \[main\]]{} to the induced immersion of $M^*$. As in the proof of \[lens\], it follows that for the height functions on $M^*$ we have $\mu_1(v)=\mu_0(v)$ and $\mu_{n-1}(v)=\mu_n(v)$ for a.e. $v$. The same thus holds for the height functions of the immersion of $M^n$, and again as in the proof of \[lens\], we see that $\pi_1(M)$ is cyclic, hence $\Gamma$ is isomorphic to ${{\mathbb{Z}}}$. Finally, since the functions $\mu_i$ are nonnegative, and $\tau_2= \cdots =\tau_{n-2}=0$, it follows that $\mu_2= \cdots
=\mu_{n-2}=0$ as well.
Now, projection onto the first factor gives rise to a fiber bundle $M^n\to {\mathbb{S}}^1$ with fiber ${\mathbb{S}}^{n-1}$ which is hence isomorphic to ${\mathbb{S}}^{n-1}\times[0,1]/(p,0)\sim (\sigma(p),1)$ for some diffeomorphism $\sigma$. If $\sigma$ is orientation preserving, $\sigma$ is homotopic to the identity and hence the bundle is trivial, in which case $M^n$ is diffeomorphic to ${\mathbb{S}}^{n-1}\times{\mathbb{S}}^1$. If on the other hand $\sigma$ is orientation reversing, $M^n$ is nonorientable and the orientable double cover is diffeomorphic to ${\mathbb{S}}^{n-1}\times{\mathbb{S}}^1$ and hence $M^n$ is diffeomorphic to $({\mathbb{S}}^{n-1}\times{\mathbb{S}}^1)/\Delta{{\mathbb{Z}}}_2$. Notice that, in both cases, the product structure is not necessarily isometric.
The presence of points with positive curvatures imposes further restrictions.
\[posric\] If there exists a point with positive Ricci curvature, then either $M^n$ is diffeomorphic to ${\mathbb{S}}^n$, or $M^n$ is isometric to a product ${\mathbb{S}}^k\times{\mathbb{S}}^{n-k}$ and $f$ is a product of two convex hypersurfaces, for some $2\leq k\leq n-2$. In particular, if there is a point with positive sectional curvature, then $M^n$ is diffeomorphic to ${\mathbb{S}}^n$.
The fundamental group must be finite since otherwise \[z\] implies that $\tilde M^n$ splits off a real line. If $n=3$, then [Proposition \[wide\]]{} implies that at a point with $Ric>0$, one of $A,B$, say $A$, has rank 2, and $B$ has rank one. But then $\hat{R}=\Lambda^2A+\Lambda^2B=\Lambda^2A$ and hence all 2-planes containing $\ker A$ have curvature 0, thus contradicting $Ric>0$. In the remaining case \[sn\] proves our claim.
In particular, Corollaries \[sn\], \[lens\] and \[z\] imply the following.
\[mustbewide\] If $M^n$ is not diffeomorphic to a sphere, then the immersion is wide.
0.1cm
Wide and locally wide immersions
================================
In this section we obtain further information about the way the manifold $M^n$ is immersed, that is, about the [isometric immersion ]{}$f$. As we saw in \[mustbewide\], unless $M^n$ is diffeomorphic to the sphere, for which the space of nonnegatively curved immersions is quite rich, the [isometric immersion ]{}$f$ must be [wide]{}. Hence the purpose of this section is to understand these immersions. We have chosen this approach since there are interesting wide immersions of spheres, as shown in Example 2.
We assume from now on that $M^n$ is a Riemannian manifold with nonnegative sectional curvature, $n\geq 3$, and $f:M^n\to{{\mathbb{R}}}^{n+2}$ is an isometric immersion. To provide a deeper understanding of the local phenomena, we do not require $M^n$ to be complete or compact unless otherwise stated. Instead, we only assume that $f$ satisfies the local consequence in [Proposition \[wide\]]{} of being wide. In other words, again following the notations in , we say that $f$ is [*locally wide*]{} if, at every point $p\in M$, either $$\label{localpepe}
\ker A\cap\ker B \neq 0,
\ \ \ \text{or}
\ \ \ \ker A\oplus \ker B = T_p M\ \ \text{with}\ A,B\neq 0.$$ From now on we will also assume in the second case that ${\ensuremath{\operatorname{rank}}}B\le {\ensuremath{\operatorname{rank}}}A$.
Recall that the [*index of nullity*]{} $\mu$ and the [*nullity*]{} $\Gamma$ of $M$ are defined as $$\Gamma(p)=\{u\in T_pM
\colon R(a,b)u=0 \ \ \forall\ a,b\in T_pM\}, \ \text{ and }\
\mu(p)=\dim \Gamma(p).$$ Furthermore, we have the [*index of relative nullity*]{} $\nu(p)$ of $f$ at $p$ defined as the dimension of the [*relative nullity*]{} $\Delta(p)$ of $f$ at $p$, $$\Delta(p)= \{u\in T_pM\colon \alpha(u,v)=0 \ \, \forall \, v\in T_pM\}=
\ker A(p)\cap\ker B(p).$$ Notice that $\mu$ is an intrinsic invariant, while $\nu$ is extrinsic. By the Gauss equation, $\Delta\subseteq\Gamma$ and hence $\nu\le\mu$. We will also use the well know fact (see e.g. [@chk; @ma]) that the nullity distribution, as well as the relative nullity distribution, is integrable on any open set where it has constant dimension, and its leaves are totally geodesic in $M^n$. In the case of the relative nullity, the images of the leaves under $f$ are open subsets of affine subspaces of the Euclidean space as well. In addition, if $M^n$ is complete, the leaves of both distributions are complete on the open set of minimal nullity or minimal relative nullity.
Choosing an orthonormal normal frame $\{\xi,\eta\}$ as in , by $M^n$ can be written as the disjoint union $$M^n = K\cup U_1\cup \cdots\cup U_{[n/2]},$$ where $K$ is the subset of positive index of relative nullity, $$K = \{p\in M^n: \nu(p) > 0\},$$ and, for $1\leq k \leq [n/2]$, $U_k$ is given by $$U_k = \{p\in M^n: {\ensuremath{\operatorname{rank}}}\, A(p) = n-k,\,{\ensuremath{\operatorname{rank}}}\, B(p) = k \ \ \text{and} \ \ker A(p)\cap \ker B(p) = 0\}.$$ From the continuity of the eigenvalues and $\hat{R}=\Lambda^2A+\Lambda^2B$ it follows that:
- $K$ is closed, and the sets $U_k$ are open with $\partial U_k \subset K$;
- $U_2\cup\cdots\cup U_{[n/2]}$ is the set of all points with positive Ricci curvature, $Ric>0$;
- On $U_1$ we have $\hat{R}=\Lambda^2A$ and hence $\Gamma=\ker A $ and $\mu=1$, $\nu=0$.
\[KU\] [The above discussion and \[posric\] imply that, if $M^n$ is compact and $f$ is wide, then $f$ is isometric to a product of convex hypersurfaces unless $U_k=\emptyset$ for all $k\geq 2$. Therefore, for the immersions of types $(c)$ and $(d)$ in [Theorem \[a\]]{} we have that $M^n=K\cup U_1$, and, in particular, $\mu\geq 1$ everywhere.]{}
In what follows, on $M^n\setminus K$ where we have seen that the frame $\{\xi,\eta\}$ is unique and smooth, $w$ will denote the normal connection form of $f$ given by $$w(X):={\langle}\nabla^\perp_X\xi,\eta{\rangle}$$ for $X\in TM^n$, and thus $\nabla^\perp_X\xi=w(X)\eta$ and $\nabla^\perp_X\eta=-w(X)\xi$. Hence the Codazzi equations $(\nabla_X
A)(Y,\beta)=(\nabla_Y A)(X,\beta)$ become $$\nabla_X AY-A\nabla_XY-w(X)BY = \nabla_Y AX-A\nabla_YX-w(Y)BX,$$ $$\nabla_X BY-B\nabla_XY+w(X)AY = \nabla_Y BX-B\nabla_YX + w(Y)AX,$$ or, equivalently, $$\nabla_X AY-\nabla_Y AX=A([X,Y])+B(w(X)Y-w(Y)X)$$ $$\nabla_X BY-\nabla_Y BX=B([X,Y])-A(w(X)Y-w(Y)X),$$ while the Ricci equation is $${\langle}R^\perp(X,Y)\xi,\eta{\rangle}= dw(X,Y) = {\langle}[A,B]X,Y{\rangle}.$$
For the sake of completeness, let us first analyze the local behaviour of $f$ on $U_k$ for $k\geq 2$. We will see that it is already a product of strictly convex hypersurfaces, hence showing that case $(b)$ in [Theorem \[a\]]{} has its roots in a local phenomenon.
\[prod\] Assume that $f$ is locally wide and $2\leq k \leq [n/2]$. Then the immersion $f|_{U_k}$ is locally a product immersion of two strictly convex Euclidean hypersurfaces whose respective normal vectors are $\xi$ and $\eta$.
Since their dimensions are constant, the distributions $\ker A$ and $\ker
B$ are smooth, have rank bigger or equal than two, and satisfy $\ker A\oplus \ker B = TU_k$.
For $X,Y\in \ker A$, the Codazzi equations imply that $A[X,Y] =
B(w(X)Y-w(Y)X)$. Since ${\ensuremath{\operatorname{Im}}}A \cap {\ensuremath{\operatorname{Im}}}B=0$, both sides have to be $0$ and hence $\ker A$ is integrable. Furthermore, $\ker A \subset \ker w$ since $\ker A \cap \ker B = 0$ and $\dim\ker A\geq 2$. Indeed, If $AX=0$, choose a linearly independent $Y\in \ker A \cap\ker\omega$, which implies $w(X)BY=w(Y)BX$ and hence $\omega(X)=0$. Analogously, $\ker B$ is integrable and $\ker B \subset \ker w$. Therefore, $w|_{U_k}=0$ which implies that $f|_{U_k}$ has flat normal bundle. From the Ricci equation it now follows that $A$ and $B$ commute, and hence $\ker A \perp \ker B$.
For $X,Y\in \ker A= {\ensuremath{\operatorname{Im}}}B $ and $U\in \ker B={\ensuremath{\operatorname{Im}}}A $ the Codazzi equations imply that $A[X,U] = \nabla_X AU$. Since $U$ is arbitrary, $\nabla_XU\in {\ensuremath{\operatorname{Im}}}A$ as well. Hence ${\langle}\nabla_X Y,U{\rangle}=-{\langle}Y,\nabla_X U{\rangle}=0$ which implies that the distribution $\ker A$ is totally geodesic, and similarly so is $\ker B$. We conclude that both $\ker A$ and $\ker B$ are mutually orthogonal transversal totally geodesic distributions, and hence both are parallel, and $M^n$ is locally a product. Furthermore, $\alpha_f(X,U)={\langle}AX,U{\rangle}\xi+{\langle}BX,U{\rangle}\eta=0$ for all $X\in \ker A$ and $U\in\ker B$. The proposition then follows from the Main Lemma in [@m1]. Observe that each factor is strictly convex since $A$ and $B$ are positive definite on the corresponding factors.
Before continuing we point out the following consequence.
\[ric\] If $M^n$ is complete, $Ric_M>0$, and $f$ is locally wide, then $f$ is a global product of two strictly convex embedded hypersurfaces. In particular, $f$ is rigid.
Consider $\pi:\tilde M^n\to M^n$ the universal cover of $M^n$ with the covering metric, and $\tilde f = f \circ \pi$. By hypothesis, $K$ and $U_1$ are empty. Since $\partial U_k\subset K$ for all $k$, there is $k_0\geq 2$ such that $U_{k_0}=\tilde M^n$. But then [Proposition \[prod\]]{} and the deRham decomposition theorem imply that $\tilde M^n$ is globally a product. From [@am] it follows that $\tilde f$ is a product immersion of two strictly convex Euclidean hypersurfaces. By [@sa1], a complete strictly convex Euclidean hypersurface is the boundary of a strictly convex body and hence embedded. Thus $\tilde f$ is injective and so $\tilde f = f$ and $\tilde M^n=M^n$.
The description of $f$ on the set $U_1$ is considerably more delicate. This case is of main interest to us since, as pointed out in [Remark \[KU\]]{}, for the immersion in case $(c)$ and $(d)$ of [Theorem \[a\]]{}, we have $M=K\cup U_1$. Furthermore, $U_1$ is nonempty since otherwise the open subset of minimal relative nullity in $K$ would be foliated by complete straight lines in Euclidean space, contradicting compactness.
We need the following definition from [@df1].
\[comp\] [Given an [isometric immersion ]{}$g\colon\,M^n\to{{\mathbb{R}}}^{n+1}$, we say that another [isometric immersion ]{}$f\colon\,M^n\to{{\mathbb{R}}}^{n+2}$ is a [*composition (of $g$)*]{} when there is an isometric embedding $g'\colon M^n\hookrightarrow
N_0^{n+1}$ into a flat manifold $N_0^{n+1}$, an [isometric immersion ]{}$j\colon\,N_0^{n+1}\to{{\mathbb{R}}}^{n+1}$ (that is, a local isometry) satisfying $g=j\circ g'$, and an [isometric immersion ]{}$h\colon\,N_0^{n+1}\to{{\mathbb{R}}}^{n+2}$ such that $f=h\circ g'$.]{}
Observe that, for any open subset $U\subset M^n$ where $g_{|U}$ in the above is an embedding, we can assume that $N_0^{n+1}$ is an open subset on ${{\mathbb{R}}}^{n+1}$ and $j$ is the inclusion.
Compare the following local description of $f$ on $U_1$ with the structure of Example 1 in the introduction.
\[b0\] If $\pi:\tilde U_1\to U_1$ is the universal cover of $U_1$, then $f\circ\pi$ is a composition of a convex Euclidean hypersurface with constant index of relative nullity one.
Since ${\ensuremath{\operatorname{rank}}}B=1$, and hence $\hat{R}=\Lambda^2A$, the shape operator $A$ alone satisfies the Gauss equation along $U_1\subset M^n$. We claim that $A$ also satisfies the Codazzi equation for hypersurfaces, that is, the skew symmetric tensor $S(X,Y):=\nabla_XAY - \nabla_YAX - A[X,Y]$ vanishes.
For any $X,Y\in TU_1$, the Codazzi equation for $A$ tells us that $S(X,Y)
= B(w(X)Y-w(Y)X)$. For $X,Y\in\ker B$, the Codazzi equation for $B$ says that $B[X,Y]=-A(w(X)Y-w(Y)X)$. Since $\dim \ker B\ge 2$, it follows as in the proof of [Proposition \[prod\]]{} that $\ker B \subset \ker w$. Thus either $\ker B = \ker w$ or $w=0$. In the first case, since $S$ is skew symmetric, we can assume that $X$ and $Y$ are linearly independent, and hence can assume that $X\in\ker\omega$. But then $S(X,Y)=0$. Altogether, $S$ vanishes.
By the Fundamental Theorem of Submanifolds, locally on $U_1$ (or globally on $U_1$ if it is simply connected), there exists a Euclidean hypersurface $g$ whose second fundamental form is $A$. Since ${\ensuremath{\operatorname{rk}}}A = n-1 \geq 2$, $\nu_g=\mu|_{U_1}\equiv 1$. Now, since $\ker B \subset \ker w$, by Proposition 8 in [@df1] we have that $f$ is a composition of $g$ (see also Proposition 9 in [@df]).
Since $\nu_g=1$, the nullity geodesics in $U_1$ are (locally) mapped by $g$ into straight lines of ${{\mathbb{R}}}^{n+1}$. If $M=U_1$ is complete and simply connected, we will see in the proof of [Theorem \[b1\]]{} that these straight lines are parallel and $g$ is globally a cylinder.
We will also discuss properties of the metric and the immersion on the set of index of nullity $1$, i.e., on $V:=\mu^{-1}(1)$, and the open set $U_1'$ where $B$ does not vanish (and hence $f$ is 1-regular, i.e., $\dim{\mbox{span}}{\ensuremath{\operatorname{Im}}}(\alpha) = 2$), $$U_1'= \{p\in V: {\ensuremath{\operatorname{rk}}}B(p) = 1\}\supseteq U_1.$$ Recall that by the Gauss equations on the set $U_1$ we have $\mu=1$ and, by definition, $\nu=0$ as well. Thus, $$U_1\subseteq U_1'\subseteq V\subseteq K\cup U_1 \subseteq M^n.$$ On the complement of $U_1$ in $V$ we have $\nu=\mu=1$. In any case, $V$ is the set of minimal index of nullity of $M^n$, so its leaves of nullity are complete if $M^n$ is complete.
The global version of [Proposition \[b0\]]{} is the following. This in particular applies to any immersion in cases $(c)$ and $(d)$ of [Theorem \[a\]]{}.
\[b1\] Assume that $M^n$ is compact with nonnegative sectional curvature, and $f$ is locally wide. Furthermore, assume that $M^n$ has no points with positive Ricci curvature, i.e., $M=U_1\cup K$. Let $\pi\colon\tilde V \to V$ the universal cover of $V$, $\tilde f=f\circ\pi$ the lift of $f|_V$, and set $\tilde{U}_1'=\pi^{-1}(U_1')\subset \tilde V$. Then we have:
- $\tilde V$ splits globally and isometrically as a product $\tilde V = N^{n-1} \times {{\mathbb{R}}}$, where $g\colon N^{n-1}\to {{\mathbb{R}}}^n$ is a strictly locally convex hypersurface. In particular, $\tilde V$ is itself an Euclidean hypersurface via the cylinder over $g$, $g\times\operatorname{Id}_{{\mathbb{R}}}:\tilde V\to{{\mathbb{R}}}^{n+1}$;
- The restriction $\tilde f|_{\tilde{U}_1'}$ is a composition of the cylinder over $g|_{\tilde{U}_1'}$;
- Along each connected component $W$ of the interior of $\tilde V\setminus \tilde U_1'$, $\tilde f|_W$ is a composition of the cylinder over $g|_W$ with a linear inclusion ${{\mathbb{R}}}^{n+1}\subset{{\mathbb{R}}}^{n+2}$.
Here we continue with the notations in the proof of [Proposition \[b0\]]{}. Define on $V$ the [*splitting tensor*]{} of the (totally geodesic and complete) nullity distribution, $C:\Gamma^\perp \to \Gamma^\perp$ given by $CX=-\nabla_XT$, where $T$ is a unit vector field tangent to $\Gamma$. Since $\Gamma$ is totally geodesic, the distribution $\Gamma^\perp$ is totally geodesic if and only if $C$ vanishes.
Observe that $C$ satisfies the Riccati type differential equation $C'=C^2$ when restricted to a geodesic with $\gamma' = T\in\Gamma$. Indeed, $C'X = \nabla_TCX - C \nabla_TX = -\nabla_T\nabla_XT +
\nabla_{\nabla_TX}T = -\nabla_X\nabla_TT - \nabla_{[T,X]}T -R(T,X)T+
\nabla_{\nabla_TX}T = \nabla_{\nabla_XT}T=C^2X$ since $\nabla_TT=0$ and $R(T,X)T=0$. Since $\gamma$ is complete, this ODE holds true over the entire real line, and then $C(\gamma(t))=C_0(I-tC_0)^{-1}$, where $C_0=C(\gamma(0))$. Therefore, along each nullity geodesic, all real eigenvalues of $C$ vanish.
We claim that $C\equiv 0$. Observe first that every nullity line in $V$ has to intersect $U_1$ since in $V\setminus U_1$ we have $\nu=1$ and hence a complete nullity line is also a relative nullity line. But then its image under $f$ is a straight line in ${{\mathbb{R}}}^{n+2}$, contradicting compactness of $M^n$. As we saw in the proof of [Proposition \[b0\]]{}, on $U_1$ the shape operator $A$ satisfies the Codazzi equation of a hypersurface and hence $\nabla_TAX=\nabla_XAT+A[T,X]=A[T,X]$ since on $U_1$ we also have $AT=0$. Thus if $X\in TU_1$ and $X\in\Gamma^\perp$ we have $$\label{a'}
A'X = \nabla_{T} AX - A\nabla_{T}X =A[T,X]-A\nabla_{T}X
= -A\nabla_XT = ACX.$$ Hence, from the symmetry of $A'$, we have $A C=(A C)^t=C^tA $. Furthermore, since $A>0$ on $\Gamma^\perp$, we have the inner product ${\langle}X,Y{\rangle}_1={\langle}AX,Y{\rangle}$, positive definite on $\Gamma^\perp$. But then $C$ is self adjoint since ${\langle}X,CY{\rangle}_1={\langle}AX,CY{\rangle}={\langle}C^tAX,Y{\rangle}={\langle}ACX,Y{\rangle}={\langle}CX,Y{\rangle}_1$. Thus all eigenvalues of $C$ are real, and hence $C\equiv 0$ on $V$ which proves the claim.
In particular, $V$ is locally a Riemannian product of a line and a manifold with positive sectional curvature. Moreover, by , $A$ is parallel along $\gamma$ in $U_1$, which implies in particular that $\xi$ and $\eta$ are parallel along $U_1$ as well.
Now, in $V\setminus U_1$, $T$ spans the relative nullity of $f$. For any parallel normal vector field $\sigma$ along $\gamma$ we obtain as in that $A_\sigma'=A_\sigma C = 0$. In particular, by taking the parallel transport of $\xi$ we see that, along the whole $V$, there is a unique smooth unit normal vector field $\xi$ such that $A=A_\xi\geq 0$ has rank equal to $n-1$, and for a local unit normal vector field $\eta\perp\xi$, $B=A_\eta$ has rank at most 1. In addition, $A$ is parallel along the complete lines of nullity in $V$, and $w(T)=0$ along $V\setminus U_1$.
We now show that $\tilde V$ is a cylinder over a strictly convex Euclidean hypersurface, thus proving part $(i)$. First, the Ricci equation along $V\setminus U_1$ tells us that $$0=R^\perp(X,T)=T(w(X))-X(w(T))-w(\nabla_TX-\nabla_XT)
=T(w(X))-w(\nabla_TX),$$ since $\Gamma=\Delta$ in $V\setminus U_1$. Therefore, $\ker w$ is parallel along $\gamma$. Moreover, since $B'=BC=0$, $\ker B$ is also parallel along $\gamma$. But in the proof of [Proposition \[b0\]]{} we showed that, on $U_1$, $\ker B\subset\ker w$. We conclude that $\ker B\subset\ker w$ on the whole $V$, and, again as in the proof of [Proposition \[b0\]]{}, $A$ satisfies Gauss and Codazzi equations for $V$. So, the tensor $\tilde A=\pi^*A$ satisfies Gauss and Codazzi equations for $\tilde V$, and thus there exists an isometric immersion $g':\tilde V\to{{\mathbb{R}}}^{n+1}$ with gauss map $\tilde\xi$ whose second fundamental form is $\tilde A$. By construction, $\Gamma$ is now the relative nullity of $g'$, and hence the complete geodesics of nullity are mapped into complete straight affine lines of ${{\mathbb{R}}}^{n+1}$. On the other hand, since $$\tilde \nabla_X g'_*T = {\langle}\tilde AT,X{\rangle}\tilde\xi = 0,
\ \ \ \forall X\in TV,$$ for the standard connection $\tilde \nabla$ of ${{\mathbb{R}}}^{n+1}$, the complete leaves of $\Gamma$ are mapped by $g'$ into locally parallel lines of ${{\mathbb{R}}}^{n+1}$. Therefore, $\tilde V$ is globally a product $\tilde V =
N^{n-1}\times {{\mathbb{R}}}$ and $g'$ is a cylinder, i.e., $g'=g\circ\operatorname{Id}_{{\mathbb{R}}}$, where $g:N^{n-1}\to {{\mathbb{R}}}^n$ is a strictly convex hypersurface, as claimed.
Finally, since $\ker B \subset \ker w$, part $(ii)$ follows from Proposition 8 in [@df1] as in [Proposition \[b0\]]{}, while part $(iii)$ is immediate from $B=0$ and $w=0$ (i.e. $\eta$ is locally constant) on $V\setminus U_1$.
\[1reg\] [The map $h$ in [Definition \[comp\]]{} may fail to be an immersion along the image of $j$ on the boundary points of $\tilde U_1'$.]{}
The above easily implies the following corollary. Note that this applies in particular to the immersions in part $(c)$ of [Theorem \[a\]]{} if the metric on ${\mathbb{S}}^{n-1}$ has positive sectional curvature.
\[2pos\] Under the assumptions of [Theorem \[b1\]]{}, suppose further that $M^n$ has 2-positive Ricci curvature. Then, for the universal cover $\pi:\tilde M^n\to M^n$ we have $\tilde M^n={\mathbb{S}}^{n-1}\times{{\mathbb{R}}}$, where $g:{\mathbb{S}}^{n-1}\to{{\mathbb{R}}}^n$ is a compact strictly convex embedded hypersurface. Moreover, if $f$ is 1-regular, then $f\circ\pi$ is globally a composition of its cylinder: $f\circ\pi=h\circ (g\times \operatorname{Id}_{{\mathbb{R}}})$.
The hypothesis on the Ricci curvature is equivalent to ${\ensuremath{\operatorname{rank}}}A=n-1$ everywhere, and hence $V=M^n$. The 1-regularity of $f$ implies that $U_1'=V=M^n$. The result then follows from [Theorem \[b1\]]{}.
For $n=3$, if $M^3$ is compact but not diffeomorphic to a sphere, we have the decomposition $$M^3=K\cup U_1=F\cup V,$$ where $F$ is the set of flat points of $M^3$. Indeed, in dimension 3, $\mu(p)>1$ implies that $p\in F$.
Compare our next result with Examples 1 and 2 in the introduction.
\[3\] Let $f:M^3 \to {{\mathbb{R}}}^5$ be an [isometric immersion ]{}of a compact manifold with nonnegative sectional curvature. If $M^3$ has no complete geodesic of flat points, then $M^3$ is either diffeomorphic to ${\mathbb{S}}^3$, or its universal cover $\tilde M^3$ is isometric to ${\mathbb{S}}^2\times{{\mathbb{R}}}$ for some metric of positive Gaussian curvature on the sphere ${\mathbb{S}}^2$.
By \[mustbewide\], $f$ is wide, hence locally wide. As we saw in the proof of [Theorem \[b1\]]{}, $V$ is then foliated by complete geodesics of nullity, and therefore its boundary as well. But since $M^3=F\cup V$, the boundary of $V$ is made of flat points. Therefore, $M^3$ has no flat points, and $\mu=1$ everywhere. By part $(i)$ of [Theorem \[b1\]]{}, $\tilde M^3=\tilde V=N^2\times{{\mathbb{R}}}$ splits globally and isometrically. That $N^2$ is a sphere is a consequence of [Theorem \[a\]]{}.
Part $(i)$ of [Theorem \[b1\]]{} also immediately implies the next corollary, which has [Theorem \[3a\]]{} as a consequence.
\[lens1\] Let $f:M^3={\mathbb{S}}^3/{{\mathbb{Z}}}_k \to {{\mathbb{R}}}^5$ with $k>1$ be an isometric immersion with nonnegative sectional curvature. Then $M^3$ has flat points, and the complement $V$ of its flat points is isometric to a twisted cylinder $(N^2\times{{\mathbb{R}}})/{{\mathbb{Z}}}$, where $N^2 \subset {{\mathbb{R}}}^3$ is a surface with positive Gaussian curvature.
We can now use the above and \[lens\] to compute the type numbers $\tau_k$ for case $(d)$ in [Theorem \[a\]]{}.
\[taus\] In the situation of \[lens1\] and hence [Theorem \[a\]]{} part $(d)$, the type numbers are given by $\tau_k=\frac{1}{8\pi^2}\int_{N^2}K(x)\kappa_g(x)dx$ for all $0\leq k \leq 3$. Here, $K$ denotes the Gaussian curvature of $N^2$ and $\kappa_g(x)$ the total curvature of the leaf of $\Gamma$ through $x\in N^2$, considered as a curve in ${{\mathbb{R}}}^5$.
Notice first that the computation of $\tau_k$ in only takes into account the normal bundle over the set with vanishing relative nullity, which in our case is $U_1\subset V$. Writing $\beta=\cos(\theta)\xi+\sin(\theta)\eta$, a direct computation using ${\ensuremath{\operatorname{rank}}}A\leq 2$ and ${\ensuremath{\operatorname{rank}}}B\leq 1$ shows that $\det A_\beta=\cos(\theta)^2\sin(\theta)\det(\hat A){\langle}BZ,Z{\rangle}$ on $V$, where $\hat A$ is the restriction of $A$ to $\Gamma^\perp=TN^2$ and $Z=\gamma_x'\in\Gamma=\ker A$ for $\gamma_x(t)=[(x,t)]$. Along $V$ the surface $N^2$ in \[lens1\] has $\det(\hat A)=K\geq 0$. Furthermore, $|{\langle}BZ,Z{\rangle}|= \|\tilde\gamma_x''\|$ is the geodesic curvature of $\tilde\gamma_x:=f\circ\gamma_x$ in ${{\mathbb{R}}}^5$ since $AZ=0$ and hence $\tilde\gamma_x''=\alpha (\gamma',\gamma')={\langle}BZ,Z{\rangle}\eta$. Thus implies that $$\tau_k = \frac{1}{4}\sum_{i=0}^3\tau_i=
\frac{3}{32\pi^2}\int_{T_1^\perp\!V}{\left\vert\det A_\beta\right\vert}=
\frac{1}{8\pi^2}\int_VK(x)\|\tilde\gamma_x''(t)\|=
\frac{1}{8\pi^2}\int_{N^2}K(x)\kappa_g(x)dx.$$ Here, $\kappa_g(x)=\int_0^a\|\tilde\gamma_x''(t)\|dt$ is the total geodesic curvature of $\tilde\gamma_x\subset{{\mathbb{R}}}^5$ until the return time $a>0$, i.e., the cyclic group ${{\mathbb{Z}}}$ in \[lens1\] is spanned by $(x,t)\mapsto(j(x),t+a)\in\operatorname{Iso}(N\times{{\mathbb{R}}})$.
In particular, for the switched 3-sphere $M^3_{\epsilon}$ in Example 2 we get $\tau_k=1$, and hence the immersion is tight.
[9999l]{} S. Alexander and R. Maltz, *Isometric immersions of Riemannian products in Euclidean space*. J. Diff. Geom. 11 (1976), 47–57.
Y.Y. Baldin and F. Mercuri, [*Isometric immersions in codimension two with nonnegative curvature*]{}. Math. Z. [**173**]{} (1980), no. 2, 111–117.
Y.Y. Baldin and F. Mercuri, [*Codimension two nonorientable submanifolds with nonnegative curvature*]{}. Proc.Amer.Math.Soc. [**103**]{} (1988), 918–920.
J. L. Barbosa, M. Dajczer, and R. Tojeiro, [*Isometric immersions of Riemannian products revisited*]{}. Comm. Math. Helv. [**69**]{} (1994), 281–290.
R.L. Bishop, [*The holonomy algebra of immersed manifolds of codimension two*]{}. J. Differ. Geom [**6**]{} (1971), 119–128.
C. Böhm and B. Wilking, [*Manifolds with positive curvature operators are space forms*]{}. Ann. of Math. [**167**]{} (2008), 1079–1097.
C.S. Chen, [*On tight isometric immersions of codimension two*]{}. Amer. J. Math. [**94**]{} (1972), 974–990.
J. Cheeger and D. Gromoll, [*The splitting theorem for manifolds of nonnegative Ricci curvature*]{}. J. Differential Geometry [**6**]{} (1972), 119–128.
S.S. Chern and N. Kuiper, [*Some theorems on the isometric imbedding of compact Riemann manifolds in euclidean space*]{}. Ann. of Math. (2) [**56**]{} (1952), 422–430.
C. Chicone and N.J. Kalton, [*Flat embeddings of the Möbius strip in ${{\mathbb{R}}}^3$*]{}. Comm. Appl. Nonlinear Anal. [**9**]{} (2002), no. 2, 31–50.
M. Dajczer and L. Florit, [*Compositions of isometric immersions in higher codimension*]{}. Manuscripta Math. [**105**]{} (2001), 507–517.
M. Dajczer and L. Florit, [*Genuine deformations of submanifolds*]{}. Comm. Anal. Geometry [**12**]{} (2004), 1105–1129.
L. Florit and W. Ziller, [*Manifolds with conullity at most two as graph manifolds*]{}. Preprint.
R. Maltz, [*The nullity spaces of curvature-like tensors*]{}. J. Diff. Geom. [**7**]{} (1972), 519–523.
J.D. Moore, [*Isometric immersions of Riemannian products*]{}. J. Diff. Geom. [**5**]{} (1971), 159–168.
J.D. Moore, [*Codimension two submanifolds of positive curvature*]{}. Proc. Amer. Math. Soc. [**70**]{} (1978), 72–74.
M. Morse, [*The existence of polar non-degenerate functions on differentiable manifolds*]{}. Ann. of Math. (2) [**71**]{} (1960), 352–383.
R. Sacksteder, [*On hypersurfaces with no negative sectional curvatures*]{}. Amer. J. Math. [**82**]{} (1960), 609–630.
G. Schwarz, [*A pretender to the title “canonical Moebius strip"*]{}. Pacific J. Math. [**143**]{} (1990), no. 1, 195–200.
A. Weinstein, [*Positively curved n-manifolds in ${{\mathbb{R}}}^{n+2}$*]{}. J. Diff. Geom. [**192**]{} (1970), 1–4.
B. Wilking, [*Nonnegatively and Positively Curved Manifolds*]{}. in: Metric and Comparison Geometry, Surv. Differ. Geom. 11, ed. K.Grove and J.Cheeger, International Press, (2007), 25–62.
W. Wunderlich, [*Über ein abwickelbares Möbiusband*]{}. Monatsh. Math. [**66**]{} (1962) 276–289.
[^1]: The first author was supported by CNPq-Brazil, and the second author by a grant from the National Science Foundation, by IMPA, as well as CAPES.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this article, we treat $G_2$-geometry as a special case of multisymplectic geometry and make a number of remarks regarding Hamiltonian multivector fields and Hamiltonian differential forms on manifolds with an integrable $G_2$-structure; in particular, we discuss existence and make a number of identifications of the spaces of Hamiltonian structures associated to the two multisymplectic structures associated to an integrable $G_2$-structure. Along the way, we prove some results in multisymplectic geometry that are generalizations of results from symplectic geometry.'
address:
- 'Department of Mathematics, University of Rochester, Rochester, NY, $14627$'
- 'Department of Mathematics, University of Rochester, Rochester, NY, $14627$'
- 'Department of Mathematics, University of California - Riverside, Riverside, CA, $92521$'
author:
- 'Hyunjoo Cho, Sema Salur, and A. J. Todd'
bibliography:
- 'ham.bib'
title: 'Remarks on Hamiltonian Structures in $G_2$-Geometry'
---
\#1[[(\[\#1\])]{}]{} \[section\] \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Corollary]{} \[thm\][Definition]{} Ø Ł
Introduction
============
Let $M$ be a $7$-dimensional manifold admitting a smooth differential $3$-form $\vp$ such that, for all $p\in M$, the pair $(T_pM,\vp)$ is isomorphic as an oriented vector space to the pair $(\R^7,\vp_0)$ where $$\vp_0=\d x^{123}+\d x^{145}+\d x^{167}+\d x^{246}-\d x^{257}-\d x^{347}-\d x^{356}
\label{G23form}$$ with $\d x^{ijk}=\d x^i\w \d x^j\w \d x^k$. In [@Br1], it is shown that the Lie group $G_2$ can be defined as the set of all elements of $GL(7,\R)$ that preserve $\vp_0$, so for a manifold admitting such a $3$-form, there is a reduction in the structure group of the tangent bundle to the exceptional Lie group $G_2$; hence, the pair $(M,\vp)$ is called a *manifold with $G_2$-structure*. Using the theory of $G$-structures and the inclusion of $G_2$ in $SO(7)$, all manifolds with $G_2$-structure are necessarily orientable and spin, any orientable $7$-manifold with spin structure admits a $G_2$-structure, and associated to a given $G_2$-structure $\vp$ are a metric $g_{\vp}$ called the *$G_2$-metric*, satisfying $$(X\lrcorner\vp)\w(Y\lrcorner\vp)\w\vp=6g_{\vp}(X,Y)\Vol_{\vp}
\label{G2metric}$$ for any vector fields $X$ and $Y$ on $M$, a $4$-form $$\sv_0=\d x^{4567}+\d x^{2367}+\d x^{2345}+\d x^{1357}-\d x^{1346}-\d x^{1256}-\d x^{1247}
\label{G24form}$$ where $\star$ is the Hodge star operator associated to $g_{\vp}$ and finally a $2$-fold vector cross product $\t$. A natural geometric requirement is that the $3$-form $\vp$ be covariant constant with respect to the Levi-Civita connection of the $G_2$-metric $g_{\vp}$; if this is so, we say that the $G_2$-structure is *integrable* and call the pair $(M,\vp)$ a *$G_2$-manifold*. It is a nontrivial fact that the integrability of the $G_2$-structure is equivalent to the holonomy of $g_{\vp}$ being a subgroup of $G_2$ as well as $\vp$ being simultaneously closed and coclosed, that is, $\d\vp=0$ and $\d^*\vp=0$ respectively, where $\d^*$ is the adjoint to the exterior derivative $\d$ defined in terms of the Hodge star $\star$ of $g_{\vp}$; the condition that $\vp$ be coclosed then implies that the $4$-form $\sv$ is closed. See [@Jo1], [@Jo2], [@Kari1], [@Kari2], [@Sa] for more information on these constructions and conditions.
Let $M$ be an $n$-dimensional smooth manifold, and let $\om$ be a closed $(k+1)$-form on $M$ satisfying the nondegeneracy condition $$X\lrcorner\om=0\text{ iff }X=0
\label{nd}$$ for $X$ a vector field on $M$. The pair $(M,\om)$ is called a *multisymplectic* $n$-manifold of degree $k+1$; such manifolds can be viewed geometrically as a natural generalization of symplectic manifolds which, in this language, are even-dimensional multisymplectic manifolds of degree $2$. Note that there are the $k$ associated linear maps $$\begin{array}{cccc}
\widehat{\om_j}:&\Om^j(TM)&\to&\Om^{k+1-j}(T^*M)\\
&Q&\mapsto& Q\lrcorner\om \\
\end{array}
\label{almaps}$$ for all $j=1,\ldots,k$ and that the nondegeneracy of $\om$ forces $\widehat{\om_1}$ to be injective and $\widehat{\om_k}$ to be surjective. Also, for multisymplectic forms of degree $\geq 3$, there is a stronger notion of nondegeneracy than that used here, cf. [@MaSw2], [@MaSw3], [@MaSw1]. Exact multisymplectic manifolds, that is, multisymplectic manifolds where the multisymplectic $(k+1)$-form is exact, arise naturally in physics as multiphase spaces which are bundles of higher-degree differential forms equipped with an exact multisymplectic form that are generalizations of the standard phase space given by the cotangent bundle equipped with the canonical symplectic form, e. g. [@CIdL1], [@CIdL2], [@FPR2], [@FPR1], [@FoRo]. It is important for the purpose of this article to note that Equations \[G2metric\], \[QsvQvpsveq\] imply the nondegeneracy of $\vp$ and $\sv$ in the sense of Equation \[nd\] which means that a $G_2$-manifold is simultaneously a multisymplectic manifold of degree $3$ and of degree $4$ with the multisymplectic $3$-form $\vp$ and multisymplectic $4$-form $\sv$ respectively.
A vector field $X$ and a real-valued function $H$ on a symplectic manifold $(M,\om)$ satisfying $$X\lrcorner\om=\d H$$ are referred to as a *Hamiltonian* vector field and a *Hamiltonian* function respectively. Similarly, a $(k-l)$-multivector field $Q$ and a differential $l$-form $\al$ on a multisymplectic manifold $(M,\om)$ of degree $k+1$ satisfying $$Q\lrcorner\om=\d\al$$ are referred to as a *Hamiltonian* $(k-l)$-multivector field and a *Hamiltonian* $l$-form respectively. The spaces of Hamiltonian vector fields and Hamiltonian functions, along with associated algebraic structures on these spaces, are important and fundamental concepts in symplectic geometry from both the mathematical and the physical perspective and arise from the Hamiltonian formulation of classical mechanics where a Hamiltonian function represents the total energy of a given mechanical system, e. g. [@Arno], [@daSi], [@McSa]; similarly, the spaces of Hamiltonian multivector fields and Hamiltonian differential forms, along with associated algebraic structures on these spaces, arise from the covariant Hamiltonian formulation of classical mechanics, and indeed much of the interest in the subject of multisymplectic geometry has come from various areas of physics, e. g. [@AtWi], [@BHR], [@BR], [@CIdL2], [@CIdL1], [@CCI], [@FPR2], [@FPR1], [@GYZ], [@PR]. The work providing the foundations for study of these spaces of Hamiltonian multivector fields and Hamiltonian differential forms on general multisymplectic manifolds has been completed by a number of authors, cf. [@CIdL1], [@FPR1], [@PR].
The purpose of this paper then is to treat $G_2$-geometry as a special case of multisymplectic geometry, consider Hamiltonian multivector fields and Hamiltonian differential forms on manifolds with an integrable $G_2$-structure, called *Rochesterian*/*coRochesterian* multivector fields and *Rochesterian*/*coRochesterian* differential forms to distinguish them from the general multisymplectic setting, and prove some results in multisymplectic geometry that are generalizations of results from symplectic geometry.
This paper comes from a research plan to view $G_2$-geometry using the perspectives of symplectic and contact geometry, cf. [@ACS], [@ACS2] for more on contact structures on manifolds with a $G_2$-structure. Treating symplectic geometry and $G_2$-geometry as analogues however is certainly not new. For example, let $V$ be a real, finite-dimensional vector space with inner product $\langle\cdot,\cdot\rangle$ and define a $k$-fold vector cross product to be an alternating, multilinear map $$P:\underbrace{V\t\cdots\t V}_{k\text{ times}}\to V$$ satisfying $$\langle P(v_1,\ldots,v_k), v_i\rangle=0\text{ for all }i=1,\ldots,k$$ $$\langle P(v_1,\ldots,v_k),P(v_1,\ldots,v_k)\rangle=\Vol(v_1,\ldots,v_k)$$ where $\Vol$ is the volume form on $V$ with respect to $\langle\cdot,\cdot\rangle$. Using this inner product, there is then also the associated $(k+1)$-exterior form $\om$ on $V$ $$\om(v_1,\ldots,v_{k+1})=\langle P(v_1,\ldots,v_k),v_{k+1}\rangle
\label{efatavcp}$$ Such objects can be attached to smooth Riemannian manifolds in the standard way by attaching them to the tangent spaces at each point and requiring that they vary smoothly across the manifold; note that a $k$-fold cross product on a Riemannian manifold yields a differential $(k+1)$-form defined point-wise by Equation \[efatavcp\]. $k$-fold vector cross products on linear spaces are studied by Brown and Gray [@BrGr] and on manifolds by Gray [@Gray]. They show that a $k$-fold vector cross product on an $n$-dimensional space only exists for certain pairs $(k,n)$: $(n-1,n)$, $(1,2m)$, $(2,7)$ and $(3,8)$. The geometry of an $(n-1)$-fold vector cross product on an $n$-dimensional smooth manifold is simply the Riemannian geometry of the original manifold since the associated differential form has degree $n$ and must therefore be a multiple of the volume form associated to the Riemannian metric; note that the standard two-fold vector cross product on $\R^3$ falls into this category. $1$-fold vector cross products are better known as almost complex structures, and the associated differential $2$-form, when closed with respect to the exterior derivative, is then a symplectic form.
The other two cases arise naturally from the octonions $\O$, an $8$-dimensional, real, non-associative, non-commutative, normed algebra with unit vector $1$ where the $1$-dimensional subspace spanned by the vector $1$ is denoted $\Re\O$ and the $7$-dimensional orthogonal complement is denoted $\Im\O$. Using the algebra multiplication on $\O$, one can define a two-fold vector cross product on $\Im\O$ [@HaLa Definition $B.1$] and a three-fold vector cross product on $\O$ [@HaLa Definition $B.1$]. Results of Gray [@Gray] show in particular that the existence of a two-fold vector cross product on a $7$-dimensional manifold is equivalent to a reduction in the structure group of the tangent bundle to the exceptional Lie group $G_2$. Work of Fernández and Gray [@FeGr] classifies $G_2$-structures into $16$ classes such as closed $G_2$-structures, coclosed $G_2$-structures and integrable $G_2$-structures and, as they mention, should be compared to the various classes of Kähler structures; examples of manifolds with $G_2$-structures satisfying various of these conditions have been extensively studied, e. g., [@Br1], [@Br2], [@BrSa], [@BrXu], [@CMS], [@ClIv], [@Fe1], [@Fe2], [@FeIg], [@Jo1], [@Jo2], [@Sa]. Finally, links between Calabi-Yau geometry and $G_2$-geometry in the context of mirror symmetry have been actively explored, e. g. [@AkSa], [@AtWi], [@GYZ], [@LeLe], [@Leun].
The body of this article consists of two main sections. Section \[sectionbackground\] is background material for the paper and is itself divided into two subsections. Section \[multivectorfields\] gives introductory material on multivector fields and associated operations including the Schouten-Nijenhuis bracket and a generalization of [@Kari1 Lemma $A.0.8$] to general $q$-multivector fields
For $Q\in\Om^q(TM)$ and $\al\in\Om^l(T^*M)$
1. $\star(Q\lrcorner\al)=(-1)^{q(l-q)}Q^{\flat}\w\star\al$
2. $\star(Q\lrcorner\star\al)=(-1)^{q(n-l-q)+l(n-l)}Q^{\flat}\w\al$
3. $Q\lrcorner\al=(-1)^{(l-q)(n-l)}\star(Q^{\flat}\w\star\al)$
4. $Q\lrcorner\star\al=(-1)^{lq}\star(Q^{\flat}\w\al)$
Section \[HSMG\] is a discussion of Hamiltonian multivector fields and Hamiltonian differential forms on multisymplectic manifolds largely paralleling the exposition of [@CIdL1] together with the multisymplectic version of a well-known result from symplectic geometry
Let $(M,\om)$ be a multisymplectic manifold of order $(k+1)$, and let $\al_i\in\Om^{k-q_i}_{H}(T^*M)$ with associated Hamiltonian multivector fields $Q_i\in\Om^{q_i}_H(TM)$, $i=1,2$ such that $q_1+q_2=k+1$. Then $\{\al_1,\al_2\}=0$ if and only if $\L_{Q_2}\al_1=0$ (if and only if $\L_{Q_1}\al_2=0$).
Finally, Section \[sectionmain\] is the main section of our paper wherein we consider Rochesterian multivector fields and Rochesterian differential forms. We first prove the helpful lemma
$\widehat{\sv}_2:\Om^2(TM)\to\Om^2(T^*M)$ is an isomorphism.
Our main results then consist of the following theorem, corollary and proposition.
Let $(M,\vp)$ be a $G_2$-manifold. $$\begin{array}{rcccl}
\Om^1_R(TM) = & \widetilde{\Om}^1_R(TM) & \hookrightarrow & \widetilde{\Om}^2_{cR}(TM) & =\Om^2_{cR}(TM) \\
& \updownarrow_{\cong} & & \updownarrow_{\cong} & \\
& \widetilde{\Om}^1_R(T^*M) & \subset & \widetilde{\Om}^1_{cR}(T^*M) & =\Om^1(T^*M)/Z^1(M) \\
\end{array}$$ $$\begin{array}{cccl}
\widetilde{\Om}^2_R(TM) & \cong & \widetilde{\Om}^0_R(T^*M) & \\
\updownarrow_{\cong}& & \shortparallel & = C^{\infty}(M)/\{f\text{ locally constant}\}\\
\widetilde{\Om}^3_{cR}(TM) & \cong & \widetilde{\Om}^0_{cR}(T^*M) & \\
\end{array}$$ $$\widetilde{\Om}^1_{cR}(TM)=\Om^1_{cR}(TM)\cong\widetilde{\Om}^2_{cR}(T^*M)$$ \[thmmain\]
There are no nonzero Rochesterian vector fields, and hence no nonzero Rochesterian $1$-forms, on a closed manifold $M$ with closed $G_2$-structure $\vp$. \[RocVecNonexistence\]
Let $(M,\vp)$ be a closed $G_2$-manifold. Then the only $2$-multivector fields $Q$ that are Rochesterian and coRochesterian are those whose contraction with $\vp$ is zero. \[RcR2mvfdisjoint\]
Background Material {#sectionbackground}
===================
Multivector Fields & Operations {#multivectorfields}
-------------------------------
References for the following material include [@MaSw1], [@Marl]; other references are given throughout. Let $V$ be an $n$-dimensional vector space over $\R$ with dual space $V^*$. Using the exterior product $\w$, there are the vector spaces of $l^{th}$-exterior powers $\La^l(V)$ and $\La^l(V^*)$ of $V$ and $V^*$ respectively for $l\geq 0$ whose elements are called *$l$-multivectors* and *exterior $l$-forms* respectively. Explicitly, we can write $$\La^l(V)=\sp\left\{v_1\w\cdots\w v_l:v_i\in V\right\}$$ $$\La^l(V^*)=\sp\left\{v^1\w\cdots\w v^l:v^i\in V^*\right\}$$ where $\w$ satisfies $v\w w=-w\w v$; this property implies that $\La^l(V)=\La^l(V^*)=\{0\}$ for all $l>n$. The collections of these spaces then forms the associative graded algebras $$\La(V)=\bigop_{l\geq 0}\La^l(V)$$ and $$\La(V^*)=\bigop_{l\geq 0}\La^l(V^*).$$
Let $M$ be an $n$-dimensional manifold, so for each $p\in M$, we have the vector spaces given by the tangent space $T_pM$ and the cotangent space $T_p^*M$ which by the above construction yield the $l^{th}$-exterior powers $\La^l(T_pM)$ and $\La^l(T_p^*M)$ respectively. The collection of these spaces gives the associated vector bundles $\La^l(TM)$ and $\La^l(T^*M)$. Smooth sections of these bundles, denoted by $\Om^l(TM)$ and $\Om^l(T^*M)$, are called *$l$-multivector fields* and *differential $l$-forms* respectively where $\Om^0(TM)=\Om^0(T^*M)=C^{\infty}(M)$ are smooth real-valued functions on $M$, $\Om^1(TM)$ is the space of vector fields on $M$, $\Om^l(TM)\cong\Om^l(T^*M)\cong\{0\}$ for all $l>n$; an $l$-multivector field that can be written as $X_1\w\cdots\w X_l$ for $X_1,\ldots,X_l\in\Om^1(TM)$ is called *decomposable*. As above, we can take the collection of multivector fields and differential forms for all $l\geq 0$ to get the associative graded algebras $$\Om(TM)=\bigop_{l\geq0}\Om^l(TM)$$ and $$\Om(T^*M)=\bigop_{l\geq0}\Om^l(T^*M)$$ of multivector fields and differential forms respectively.
For a differential $\tilde{l}$-form $\phi$ and a decomposable $l$-multivector $X_1\w\cdots\w X_l$ define the contraction of $\phi$ by $X_1\w\cdots\w X_l$ $$(X_1\w\cdots\w X_l)\lrcorner\phi=X_l\lrcorner\cdots\lrcorner X_1\lrcorner\phi$$ where $\lrcorner$ denotes the standard contraction of a differential form by a vector field; if $f\in\Om^0(TM)$, then contraction reduces to multiplication $$f\lrcorner\phi=f\phi,$$ and for a $q$-multivector $Q$ with $q>\tilde{l}$, $$Q\lrcorner\phi=0.$$
We next consider the Schouten-Nijenhuis bracket on the collection of multivector fields, a natural extension of the standard Lie bracket on vector fields. The construction given here follows that of [@Va]; see [@FPR2 Appendix A] for an equivalent, up to signs, perspective on this bracket and related operations. For $X\in\Om^1(TM)$ and $Q\in\Om^q(TM)$ define an extension of the standard Lie derivative by $$\left(\L_XQ\right)(p)=\left.\frac{d}{dt}\right|_{t=0}\left(\exp(-tX)_*Q_{\exp(tX)(p)}\right)$$ where $p\in M$ and $\{\exp(tX)\}$ denotes the one-parameter subgroup of diffeomorphisms generated by $X$. For a decomposable $l$-multivector field $X_1\w\cdots\w X_{l}$ then define $$[X_1\w\cdots\w X_{l},Q]=\sum_{i=1}^{l}(-1)^{i+1}X_1\w\cdots\w \hat{X}_i\w\cdots\w X_{l}\w[X_i,Q]$$ where $[X_i,Q]=\L_{X_i}Q$. Extend this operation to a general multivector field by linearity, and note that for $Q_i\in\Om^{q_i}(TM)$, $i=1,2$, we have $[Q_1,Q_2]\in\Om^{q_1+q_2-1}(TM)$. This bracket, called the *Schouten-Nijenhuis bracket*, satisfies the following properties $$[Q_1,Q_2]=(-1)^{q_1q_2}[Q_2,Q_1]$$ $$[Q_1,Q_2\w Q_3]=[Q_1,Q_2]\w Q_3 + (-1)^{q_1q_2+q_2}Q_2\w[Q_1,Q_3]$$ $$\begin{split}
(-1)^{q_1(q_3-1)}&[Q_1,[Q_2,Q_3]]+(-1)^{q_2(q_1-1)}[Q_2,[Q_3,Q_1]] \\
&+(-1)^{q_3(q_2-1)}[Q_3,[Q_1,Q_2]]=0 \\
\end{split}$$ for $Q_i\in\Om^{q_i}(TM)$, $i=1,2,3$. The Schouten-Nijenhuis bracket gives the algebra of multivector fields $\Om(TM)$ the structure of a *Gerstenhaber algebra*, cf. [@Marl], [@VYL], [@Xu].
To describe the relationship between the Schouten-Nijenhuis bracket and the contraction operation, we use an extension of the standard Lie derivative of a differential form along a vector field to that along a multivector field. Our sign conventions follow those given by [@Mich]. For $Q\in\Om^q(TM)$ and $\al\in\Om^l(T^*M)$ define $$\L_Q\al=Q\lrcorner\d\al-(-1)^q\d(Q\lrcorner\al)
\label{LQal}$$ This operator satisfies the following properties $$\d\L_Q\al=(-1)^{q+1}\L_Q\d\al$$ $$\L_{Q_1\w Q_2}\al=Q_2\lrcorner\L_{Q_1}\al+(-1)^{q_1}\L_{Q_2}Q_1\lrcorner\al$$ $$[Q_1,Q_2]\lrcorner\al=(-1)^{q_1q_2+q_2}\L_{Q_1}Q_2\lrcorner\al-Q_2\lrcorner\L_{Q_1}\al
\label{SNbipeLd}$$ for all $Q\in\Om^q(TM)$, $Q_i\in\Om^{q_i}(TM)$, $i=1,2$, $\al\in\Om^l(T^*M)$.
Finally, there are a number of relations that are straightforward generalizations to multivector fields of those relations given in [@Kari1 Lemma $A.0.8$] for vector fields that will be useful in Section \[sectionmain\]. Let $M$ be an $n$-dimensional Riemannian manifold with Riemannian metric $g$ where $g$ is used to also denote the extension of Riemannian metric to the spaces of $j$-multivectors and differential $j$-forms. Let $\flat:\Om^j(TM)\to\Om^j(T^*M)$ be the isomorphism such that $$g(Q,Q)=Q^{\flat}(Q)=g(Q^{\flat},Q^{\flat})\text{ for all }Q\in\Om^j{TM)},
\label{flat}$$ $\sharp:\Om^j(T^*M)\to\Om^j(TM)$ the corresponding isomorphism such that $$g(\al^{\sharp},\al^{\sharp})=\al(\al^{\sharp})=g(\al,\al)\text{ for all }\al\in\Om^j(T^*M)
\label{sharp}$$ and $\star$ the Hodge star isomorphism associated to the Riemannian metric $g$ where, for $\al,\be\in\Om^l(T^*M)$, $\star$ satisfies $$\be\w\star\al=g(\be,\al)\Vol_M=g(\al,\be)\Vol_M=\al\w\star\be$$
For $Q\in\Om^q(TM)$ and $\al\in\Om^l(T^*M)$
1. $\star(Q\lrcorner\al)=(-1)^{q(l-q)}Q^{\flat}\w\star\al$
2. $\star(Q\lrcorner\star\al)=(-1)^{q(n-l-q)+l(n-l)}Q^{\flat}\w\al$
3. $Q\lrcorner\al=(-1)^{(l-q)(n-l)}\star(Q^{\flat}\w\star\al)$
4. $Q\lrcorner\star\al=(-1)^{lq}\star(Q^{\flat}\w\al)$
\[lemmaA08gen\]
For $\be\in\Om^{l-q}(T^*M)$, $$\begin{split}
\be&\w\star(Q\lrcorner\al)=g(\be,Q\lrcorner\al)\Vol_M \\
&=(Q\lrcorner\al)(\be^{\sharp})\Vol_M=\al(Q\w\be^{\sharp})\Vol_M \\
&=g(\al,Q^{\flat}\w\be)\Vol_M=(Q^{\flat}\w\be)\w\star\al \\
&=(-1)^{q(l-q)}\be\w Q^{\flat}\w\star\al \\
\end{split}$$ from which we get the relation $$\star(Q\lrcorner\al)=(-1)^{q(l-q)}Q^{\flat}\w\star\al
\label{Qflat1}$$ Recall that $\star\star\al=(-1)^{l(n-l)}\al$ which, together with Equation \[Qflat1\], implies that $$\star(Q\lrcorner\star\al)=(-1)^{q(n-l-q)+l(n-l)}Q^{\flat}\w\al
\label{Qflat2}$$ Taking $\star$ of Equations \[Qflat1\], \[Qflat2\] gives $$Q\lrcorner\al=(-1)^{(l-q)(n-l)}\star(Q^{\flat}\w\star\al)
\label{Qflat3}$$ $$Q\lrcorner\star\al=(-1)^{lq}\star(Q^{\flat}\w\al)
\label{Qflat4}$$
Hamiltonian Structures in Multisymplectic Geometry {#HSMG}
--------------------------------------------------
Our presentation parallels that of [@CIdL1]; [@FPR1] also covers this material using coordinates on multiphase spaces. Let $(M,\om)$ be an $n$-dimensional multisymplectic manifold of degree $k+1$. Given a $(k-l)$-multivector field $Q\in\Om^{k-l}(TM)$ with $0\leq l\leq k-1$, we say that $Q$ is *locally Hamiltonian* if the $(l+1)$-form $Q\lrcorner\om$ is closed; we say that $Q$ is *Hamiltonian* if there is an $l$-form $\al$ such that $$Q\lrcorner\om=\d\al
\label{ham}$$ An $l$-form $\al$ is *Hamiltonian* if there is a $(k-l)$-multivector field $Q$ satisfying Equation \[ham\]. We denote the collection of locally Hamiltonian $(k-l)$-multivector fields by $\Om^{k-l}_{lH}(TM)$, the collection of Hamiltonian $(k-l)$-multivector fields by $\Om^{k-l}_{H}(TM)$ and the collection of Hamiltonian $l$-forms by $\Om^l_H(T^*M)$. A number of properties are immediately evident from these definitions. $\d^2=0$ gives $$\Om^{k-l}_{H}(TM)\subseteq\Om^{k-l}_{lH}(TM)$$ for each $l$, and if $H^{l+1}_{dR}(M)=\{0\}$, that is, if the $(l+1)$-de Rham cohomology space of $M$ is trivial, then $\Om^{k-l}_H(TM)=\Om^{k-l}_{lH}(TM)$. Linearity of the exterior derivative $\d$ and the interior product $\lrcorner$ gives that $\Om^{k-l}_{lH}(TM)$, $\Om^{k-l}_{H}(TM)$ and $\Om^l_H(T^*M)$ are vector spaces over $\R$, but in contrast to $\Om^{k-l}(TM)$ and $\Om^{l}(T^*M)$, we cannot view any of $\Om^{k-l}_{lH}(TM)$, $\Om^{k-l}_{H}(TM)$, $\Om^l_H(T^*M)$, in general, as modules over the commutative ring $C^{\infty}(M)$ of real-valued functions on $M$; moreover, while we can consider the associated graded vector spaces $\Om_{lH}(TM)$, $\Om_{H}(TM)$, $\Om_H(T^*M)$, these spaces are not in general closed under the wedge product $\w$, so they cannot be considered as subalgebras of multivector fields/differential forms.
Notice that, given $Q\in\Om^{k-l}_H(TM)$, the associated Hamiltonian $l$-form $\al_Q\in\Om^l_H(T^*M)$ is only defined up to the addition of a closed $l$-form. Let $Z^l(M)$ denote the space of closed $l$-forms which are a subspace of $\Om^l_H(T^*M)$ and consider the quotient space $\widetilde{\Om}^l_H(T^*M)$ given by $$\widetilde{\Om}^l_H(T^*M)=\Om^l_H(T^*M)/Z^l(M)$$ Thus, given $Q\in\Om^{k-l}_H(TM)$, there exists a unique $\tilde{\al}\in\widetilde{\Om}^l_H(T^*M)$ such that $Q\lrcorner\om=\d\tilde{\al}$. Conversely, given $\al\in\Om^l_H(T^*M)$, the associated Hamiltonian $(k-l)$-multivector field $Q_{\al}\in\Om^{k-l}(TM)$ is only defined up to a $(k-l)$-multivector field whose contraction with $\om$ is zero. Such $(k-1)$-multivector fields form a subspace of $\Om^{k-l}_H(TM)$, so we can consider the quotient space $\widetilde{\Om}^{k-l}_H(TM)$. These considerations yield the isomorphism of vector spaces $$\widetilde{\Om}^l_H(T^*M)\cong\widetilde{\Om}^{k-l}_H(TM)\text{ for all }0\leq l\leq k-1$$ In the specific case of $l=k-1$, injectivity of $\widehat{\om}_1$ implies that the Hamiltonian $1$-multivector field, i. e., Hamiltonian vector field, associated to a given Hamiltonian $(k-1)$-form is uniquely determined by that form, that is, $$\widetilde{\Om}^{k-1}_H(T^*M)\cong\widetilde{\Om}^{1}_H(TM)\cong\Om^1_H(TM)$$ and in the case of $l=0$, surjectivity of $\widehat{\om}_{k}$ implies that $$\Om^0_H(T^*M)=C^{\infty}(M)$$ and hence that $$\widetilde{\Om}^0_{H}(T^*M)=C^{\infty}(M)/\{f:M\to\R|f\text{ is locally constant}\}.$$
The injectivity of $\widehat{\om}_1$ implies that there exists an injective map $$\Om^1_{lH}(TM)\hookrightarrow Z^k(M)$$ and the surjectivity of $\widehat{\om}_k$ implies that $$\widetilde{\Om}^k_{lH}(TM)\cong Z^1(M)$$ where $\widetilde{\Om}^{k}_{lH}(TM)$ denotes the quotient space of $\Om^{k}_{lH}(TM)$ by the subspace of all elements of $\Om^{k}(TM)$ whose contraction with $\om$ is zero.
Next, let $Q_i\in\Om^{q_i}_{lH}(TM)$ for $i=1,2$. Then $$\begin{split}
[Q_1,Q_2]\lrcorner\om&=(-1)^{q_1q_2+q_2}\L_{Q_1}(Q_2\lrcorner\om)-Q_2\lrcorner(\underbrace{\L_{Q_1}\om}_{=0}) \\
&=(-1)^{q_1q_2+q_2}\L_{Q_1}(Q_2\lrcorner\om) \\
&=(-1)^{q_1q_2+q_2}(Q_1\lrcorner\underbrace{\d(Q_2\lrcorner\om)}_{=0}-(-1)^{q_1}\d(Q_1\lrcorner Q_2\lrcorner\om)) \\
&=(-1)^{q_1q_2+q_2+q_1+1}\d(Q_2\w Q_1\lrcorner\om) \\
&=(-1)^{q_1+q_2+1}\d(Q_1\w Q_2\lrcorner\om) \\
\end{split}
\label{snh}$$ shows $[Q_1,Q_2]\in\Om^{q_1+q_2-1}_{H}(TM)$. If the contraction of $\om$ by either $Q_1$ or $Q_2$ is zero, then $[Q_1,Q_2]\lrcorner\om=0$ showing that $[\cdot,\cdot]$ is well-defined on the quotient spaces $\widetilde{\Om}^{k-l}_H(TM)$. Also, with $q_1=q_2=1$, this shows $(\Om^1_{H}(TM),[\cdot,\cdot])$ and $(\Om^1_{lH}(TM),[\cdot,\cdot])$ are Lie subalgebras of the Lie algebra of vector fields $(\Om^1(TM),[\cdot,\cdot])$.
Let $\al_i\in\Om^{k-q_i}_H(T^*M)$ with associated Hamiltonian $q_i$-multivector fields $Q_i$ for $i=1,2$. Define a bracket operation on $\Om_H(T^*M)$ by $$\{\al_1,\al_2\}=(-1)^{q_1+q_2+1}(Q_1\w Q_2)\lrcorner\om
\label{hfb}$$ with $\{\al_1,\al_2\}=0$ if $q_1+q_2>k+1$. This definition is independent of the choices of associated Hamiltonian multivector field and satisfies $$\begin{split}
\{\al_2,\al_1\}&=(-1)^{q_2+q_1+1}(Q_2\w Q_1)\lrcorner\om \\
&=(-1)^{q_1q_2}(-1)^{q_1+q_2+1}(Q_1\w Q_2)\lrcorner\om=(-1)^{q_1q_2}\{\al_1,\al_2\}\\
\end{split}$$ Further, $\{\al_1,\al_2\}$ is a Hamiltonian $(k+1-q_1-q_2)$-form because $$\d\{\al_1,\al_2\}=(-1)^{q_1+q_2+1}\d(Q_1\w Q_2\lrcorner\om)=[Q_1,Q_2]\lrcorner\om$$ This bracket is well-defined on the quotient spaces $\widetilde{\Om}^{l}_H(T*M)$ because if, for example, we assume that $\d\al_1=0$, then by Equations \[ham\], \[hfb\], $$\begin{split}
\{\al_1,\al_2\}&=(-1)^{q_1+q_2+1}(Q_1\w Q_2\lrcorner\om)\\
&=(-1)^{q_1+q_2+1}(Q_2\lrcorner Q_1\lrcorner\om)\\
&=(-1)^{q_1+q_2+1}(Q_2\lrcorner\d\al_1)=0
\end{split}$$ A similar argument can also be applied in the case that the second entry is closed to get the claim; then define $$\{\tilde{\al_1},\tilde{\al_2}\}=\widetilde{\{\al_1,\al_2\}}$$
Note that the above means that we need to take the Lie degree of $\al\in\Om^{k-q}_H(T^*M)$ to be $q$, that is, the degree of $\al\in\Om^{k-q}_H(T^*M)$ as an element of the graded Lie algebra of Hamiltonian forms is defined as $k$ minus the tensor degree of $\al$; the Lie degree of $\{\al_1,\al_2\}$ is then $k-(k+1-q_1-q_2)=q_1+q_2-1$. We last consider the Jacobi identity for this bracket, so let $\al_i\in\Om^{k-q_i}_{H}(T^*M)$ with associated Hamiltonian $q_i$-multivector fields $Q_i=Q_{\al_i}\in\Om^{q_i}_H(TM)$ for $i=1,2,3$. $$\begin{split}
&(-1)^{q_1(q_3-1)}\{\al_1,\{\al_2,\al_3\}\} \\
=&(-1)^{q_1(q_3-1)}(-1)^{q_1+q_2+q_3}(Q_1\w [Q_2,Q_3])\lrcorner\om \\
=&(-1)^{q_1(q_3-1)}(-1)^{q_1+q_2+q_3}(-1)^{q_1(q_2+q_3-1)}(-1)^{q_2q_3}([Q_3,Q_2]\w Q_1)\lrcorner\om \\
=&(-1)^{q_1+q_2+q_3+q_1q_2+q_2q_3}Q_1\lrcorner [Q_3,Q_2]\lrcorner\om \\
%=&(-1)^{q_1+q_2+q_3+q_1q_2+q_2q_3}Q_1\lrcorner\d\{\al_3,\al_2\} \\
=&(-1)^{q_1+q_2+q_3+q_1q_2+q_2q_3}(-1)^{q_3+q_2+1}Q_1\lrcorner\d((Q_3\w Q_2)\lrcorner\om) \\
=&(-1)^{q_1+q_1q_2+q_2q_3+1}Q_1\lrcorner\d(Q_2\lrcorner Q_3\lrcorner\om) \\
=&(-1)^{q_1+q_1q_2+q_2q_3+1}Q_1\lrcorner\d(Q_2\lrcorner\d\al_3) \\
\end{split}$$ Similarly, $$\begin{split}
&(-1)^{q_2(q_1-1)}\{\al_2,\{\al_3,\al_1\}\}=(-1)^{q_2+q_2q_3+1}Q_2\lrcorner\d(Q_1\lrcorner\d\al_3) \\
\end{split}$$ Finally, with the use of Equations \[LQal\], \[SNbipeLd\] $$\begin{split}
&(-1)^{q_3(q_2-1)}\{\al_3,\{\al_1,\al_2\}\} \\
=&(-1)^{q_3(q_2-1)}(-1)^{q_3+q_1+q_2}(Q_3\w [Q_1,Q_2])\lrcorner\om \\
=&(-1)^{q_3q_2+q_1+q_2}[Q_1,Q_2]\lrcorner Q_3\lrcorner\om \\
=&(-1)^{q_3q_2+q_1+q_2}[Q_1,Q_2]\lrcorner\d\al_3 \\
=&(-1)^{q_3q_2+q_1+q_2}((-1)^{q_1q_2+q_2}\L_{Q_1}(Q_2\lrcorner\d\al_3)-Q_2\lrcorner\L_{Q_1}\d\al_3) \\
=&(-1)^{q_3q_2+q_1+q_1q_2}\L_{Q_1}(Q_2\lrcorner\d\al_3)-(-1)^{q_3q_2+q_1+q_2}Q_2\lrcorner\L_{Q_1}\d\al_3) \\
=&(-1)^{q_3q_2+q_1+q_1q_2}(Q_1\lrcorner\d(Q_2\lrcorner\d\al_3)-(-1)^{q_1}\d(Q_1\lrcorner\d(Q_2\lrcorner\d\al_3))) \\ &-(-1)^{q_3q_2+q_1+q_2}Q_2\lrcorner(Q_1\lrcorner\d(\d\al_3)-(-1)^{q_1}\d(Q_1\lrcorner\d\al_3)) \\
=&(-1)^{q_3q_2+q_1+q_1q_2}Q_1\lrcorner\d(Q_2\lrcorner\d\al_3)-(-1)^{q_3q_2+q_1q_2}\d(Q_1\lrcorner\d(Q_2\lrcorner\d\al_3)) \\
&+(-1)^{q_3q_2+q_2}Q_2\lrcorner\d(Q_1\lrcorner\d\al_3)) \\
\end{split}$$ These equations combine to yield the relation $$\begin{split}
(-1)&^{q_1(q_3-1)}\{\al_1,\{\al_2,\al_3\}\} +(-1)^{q_2(q_1-1)}\{\al_2,\{\al_3,\al_1\}\} \\
&+(-1)^{q_3(q_2-1)}\{\al_3,\{\al_1,\al_2\}\}=(-1)^{q_3q_2+q_1q_2+1}\d(Q_1\lrcorner\d(Q_2\lrcorner\d\al_3))
\end{split}$$ which is zero in the quotient space $\widetilde{\Om}_H(T^*M)$. This bracket and grading thus give the quotient space $\widetilde{\Om}_H(T^*M)$ the structure of a graded Lie algebra which can be identified with the graded Lie algebra of the quotient space $\widetilde{\Om}_H(TM)$ together with the Schouten-Nijenhuis bracket.
This bracket is a generalization of the *semibracket* defined in [@BHR] for the specific case of $q_1=q_2=1$, and the proof of the previous result for that case, of which our proof is a straightforward generalization, can be found in [@BHR Proposition 3.7].
Let $(M,\om)$ be a multisymplectic manifold of order $(k+1)$, and let $\al_i\in\Om^{k-q_i}_{H}(T^*M)$ with associated Hamiltonian multivector fields $Q_i\in\Om^{q_i}_H(TM)$, $i=1,2$ such that $q_1+q_2=k+1$. Then $\{\al_1,\al_2\}=0$ if and only if $\L_{Q_2}\al_1=0$ (if and only if $\L_{Q_1}\al_2=0$). \[PropBracketZero\]
Note that $q_1+q_2=k+1$ implies $q_2=k+1-q_1>k-q_1$; in particular, this means that $Q_2\lrcorner\al=0$ which, together with Equation \[LQal\] gives $$\begin{split}
\{\al_1,\al_2\}&=(-1)^{q_1+q_2+1}Q_1\w Q_2\lrcorner\om=(-1)^{q_1+q_2+1}Q_2\lrcorner Q_1\lrcorner\om \\
&=(-1)^{q_1+q_2+1}Q_2\lrcorner\d\al_1=(-1)^{q_1+q_2+1}\L_{Q_2}\al_1 \\
\end{split}$$
We remark that in the case $q_1+q_2<k+1$, then the above calculation yields $$\{al_1,\al_2\}=(-1)^{q_1+q_2+1}\L_{Q_2}\al_1+(-1)^{q_1+q}\d(Q_2\lrcorner\d\al_1)$$ showing only that $\widetilde{\{\al_1,\al_2\}}\in\widetilde{\Om}^{k+1-q_1-q_2}_H(T^*M)$ is zero if and only if $\widetilde{\L_{Q_2}\al_1}\in\widetilde{\Om}^{k+1-q_1-q_2}_H(T^*M)$ is zero.
(Co)Rochesterian Multivector Fields and (Co)Rochesterian Differential Forms on $G_2$-Manifolds {#sectionmain}
==============================================================================================
Let $(M,\vp)$ be a manifold with closed $G_2$-structure.
1. An $l$-multivector field $Q$ is called a *$G_2$ $l$-multivector field*, $l=1,2$, if the $(3-l)$-form $Q\lrcorner\vp$ is closed; $\Om^l_{G_2}(TM)$ will denote the set of $G_2$ $l$-multivector fields on $M$.
2. An $l$-multivector field $Q$ is called a *Rochesterian $l$-multivector field*, $l=1,2$, if there is a $(2-l)$-form $\al$ satisfying $Q\lrcorner\vp=\d\al$; $\Om^l_{R}(TM)$ will denote the set of Rochesterian $l$-multivector fields on $M$.
3. A differential $l$-form $\al$ is called a *Rochesterian $l$-form*, $l=0,1$, if there is a $(2-l)$-multivector field $Q$ satisfying $\d\al=Q\lrcorner\vp$; $\Om^l_{R}(T^*M)$ will denote the set of Rochesterian $l$-forms on $M$.
Corresponding to the multisymplectic $7$-manifold $(M,\vp)$ of degree $3$ there are the spaces $$\Om_{G_2}(TM)=\Om^1_{G_2}(TM)\op\Om^2_{G_2}(TM)$$ $$\Om_{R}(TM)=\Om^1_{R}(TM)\op\Om^2_{R}(TM)$$ $$\Om_{R}(T^*M)=\Om^0_{R}(T^*M)\op\Om^1_{R}(T^*M)$$ together with the quotient spaces as in Section \[HSMG\]; moreover, there are the linear maps as in Equation \[almaps\] $$\widehat{\vp}_1:\Om^1(TM)\to\Om^2(T^*M)$$ $$\widehat{\vp}_2:\Om^2(TM)\to\Om^1(T^*M)$$ with $\widehat{\vp}_1$ injective and $\widehat{\vp}_2$ surjective.
Let $(M,\vp)$ be a manifold with coclosed $G_2$-structure.
1. An $l$-multivector field $Q$ is called a *co$G_2$ $l$-multivector field*, $l=1,2,3$, if the $(4-l)$-form $Q\lrcorner\sv$ is closed; $\Om^l_{cG_2}(TM)$ will denote the set of co$G_2$ $l$-multivector fields on $M$.
2. An $l$-multivector field $Q$ is called a *coRochesterian $l$-multivector field*, $l=1,2,3$, if there is a $(3-l)$-form $\al$ satisfying $Q\lrcorner\sv=\d\al$; $\Om^l_{cR}(TM)$ will denote the set of coRochesterian $l$-multivector fields on $M$.
3. A differential $l$-form $\al$ is called a *coRochesterian $l$-form*, $l=0,1,2$, if there is a $(3-l)$-vector field $Q$ satisfying $\d\al=Q\lrcorner\vp$; $\Om^l_{R}(T^*M)$ will denote the set of Rochesterian $l$-forms on $M$.
Corresponding to the multisymplectic $7$-manifold $(M,\sv)$ of degree $4$ there are the spaces $$\Om_{cG_2}(TM)=\Om^1_{cG_2}(TM)\op\Om^2_{cG_2}(TM)\op\Om^3_{G_2}(TM)$$ $$\Om_{cR}(TM)=\Om^1_{cR}(TM)\op\Om^2_{cR}(TM)\op\Om^3_{cR}(TM)$$ $$\Om_{cR}(T^*M)=\Om^0_{cR}(T^*M)\op\Om^1_{cR}(T^*M)\op\Om^2_{cR}(T^*M)$$ together with the quotient spaces as in Section \[HSMG\]; moreover, there are the linear maps as in Equation \[almaps\] $$\widehat{\sv}_1:\Om^1(TM)\to\Om^3(T^*M)$$ $$\widehat{\sv}_2:\Om^2(TM)\to\Om^2(T^*M)$$ $$\widehat{\sv}_3:\Om^3(TM)\to\Om^1(T^*M)$$ with $\widehat{\vp}_1$ injective and $\widehat{\vp}_3$ surjective.
$\widehat{\sv}_2:\Om^2(TM)\to\Om^2(T^*M)$ is an isomorphism. \[sv2iso\]
Recall that the action of $G_2$ on $\R^7$ induces an action of $G_2$ on spaces of differential forms on a manifold with $G_2$-structure, so we can decompose each space of $k$-forms into irreducible $G_2$-representations. References for this material include [@FeGr], [@Jo1] and [@Sa]. In particular, as in [@Kari1], it is the case that $$\Om^2(T^*M)=\Om^2_7(T^*M)\op\Om^2_{14}(T^*M)
\label{2formdecomp}$$ where $\Om^2_l$ denote the $l$-dimensional irreducible $G_2$-representation in $\Om^2$ and $$\Om^2_7=\{\be\in\Om^2:\star(\vp\w\be)=2\be\}
\label{O27}$$ $$\Om^2_{14}=\{\be\in\Om^2:\star(\vp\w\be)=-\be\}
\label{O214}$$ These descriptions then yield the following relations for $\be\in\Om^2(T^*M)$ $$\star(\vp\w\be)=2\pi_7(\be)-\pi_{14}(\be)
\label{svpbe}$$ $$\pi_7(\be)=\frac{\be+\star(\vp\w\be)}{3}
\label{svpbe7}$$ $$\pi_{14}(\be)=\frac{2\be-\star(\vp\w\be)}{3}
\label{svpbe14}$$ with $$\pi_7:\Om^2(T^*M)\to\Om^2_7(T^*M)$$ $$\pi_{14}:\Om^2(T^*M)\to\Om^2_{14}(T^*M)$$ the natural projection maps.
If $\widehat{\sv}_2(Q)=Q\lrcorner\sv=0$, then $\star(\vp\w Q^{\flat})=0$ by Equation \[Qflat4\]; this, together with Equations \[svpbe\], \[svpbe7\], \[svpbe14\], implies that $$2\pi_7(Q^{\flat})=\pi_{14}(Q^{\flat})=\frac{2}{3}Q^{\flat}$$ showing that $Q^{\flat}\in\Om^2_7\cap\Om^2_{14}=\{0\}$. Thus $Q=0$, and hence $\widehat{\sv}_2$ is injective. Now, let $\be\in\Om^2(T^*M)$, so we can write $\be=\be_7+\be_{14}$ by Equation \[2formdecomp\] where $\be_7=\pi_7(\be)$ and $\be_{14}=\pi_{14}(\be)$. Using Equations \[O27\], \[O214\], \[Qflat4\] and the map $\sharp$ of Equation \[sharp\] $$\be_7=\frac{1}{2}\star(\vp\w\be_7)=\frac{1}{2}\star(\be_7\w\vp)=\frac{1}{2}\be_7^{\sharp}\lrcorner\sv$$ $$\be_{14}=-\star(\vp\w\be_{14})=-\star(\be_{14}\w\vp)=-\be_{14}^{\sharp}\lrcorner\sv$$ from which it follows that $$\be=\be_7+\be_{14}=(\frac{1}{2}\be_7^{\sharp}-\be_{14}^{\sharp})\lrcorner\sv$$ Thus $\widehat{\sv}_2$ is surjective and hence bijective.
$$\widetilde{\Om}^2_{cR}(TM)=\Om^2_{cR}(TM)$$
$$\Om^1_{cR}(T^*M)=\Om^1(T^*M)$$
$$\Om^2_{cG_2}(TM)\cong Z^2(M)$$
\[corOm2cR\]
$$\Om^1_R(T^*M)\subset\Om^1_{cR}(T^*M).
\label{R1cR1}$$
The reverse inclusion does not hold in general. For example, consider $(\R^7,\vp_0)$ where $\vp_0$ is the $3$-form defined in Equation \[G23form\], let $Q$ be the $2$-multivector field given by $Q=\frac{\pd}{\pd x^6}\w\frac{\pd}{\pd x^6}$ and let $\al$ be the $1$-form defined by $\al=x^4\d x^5+x^2\d x^3$. Then $\al$ is coRochesterian since $$Q\lrcorner\sv_0=\d x^{45}+\d x^{23}=\d\al$$ where $\sv_0$ is the $4$-form defined in Equation \[G24form\]; however, $\al$ is not Rochesterian since for a general vector field $X=\sum_{i=1}^7X_i\frac{\pd}{\pd x^i}$, $$\begin{split}
X\lrcorner\vp_0&=X_1(\d x^{23}+\d x^{45}+\d x^{67})+X_2(-\d x^{13}+\d x^{46}-\d x^{57}) \\
&+X_3(\d x^{12}-\d x^{47}-\d x^{56})+X_4(-\d x^{15}-\d x^{26}+\d x^{37}) \\
&+X_5(\d x^{14}+\d x^{27}+\d x^{36})+X_6(-\d x^{17}+\d x^{24}-\d x^{35}) \\
&+X_7(\d x^{16}-\d x^{25}-\d x^{34})
\end{split}$$ which can never be equal to $\d x^{45}+\d x^{23}$. This example also emphasizes the fact that the space of Rochesterian multivector fields is not in general closed under the wedge product since on $(\R^7,\vp_0)$, straightforward calculations show that every coordinate vector field is a Rochesterian vector field. From [@CST], if $X$ is a $3$-dimensional manifold, then $(T^*X\t\R,\vp=\Re\Om+\om\w\d t)$ is a $7$-manifold with closed $G_2$-structure where $\Om$ is a certain complex $3$-form and $\om$ is the tautological $2$-form on $T^*X$. Then the vector field $\frac{\pd}{\pd t}$ is Rochesterian with an associated Rochesterian $1$-form given by the tautological $1$-form $\al$ on $T^*X$; however, because $\widehat{\vp}_1$ is not surjective, the existence of nonzero Rochesterian vector fields, and hence Rochesterian $1$-forms, on manifolds with a closed $G_2$-structure is not guaranteed as is seen as a consequence of the following theorem (see [@ACS Theorem 2.4] for the original form of the statement and proof of this theorem).
Let $M$ be a closed manifold, and let $\vp$ be a closed $G_2$-structure on $M$. Then $X\lrcorner\vp$ is exact if and only if $X$ is the zero vector field.
If $X$ is the zero vector field, then $X\lrcorner\vp=0$ at every point; in this case, we can take $X\lrcorner\vp=\d f$ where $f:M\to \R$ is the constant function $f(p)=0$ for all $p\in M$. Conversely, assume that $X$ is an arbitrary vector field such that $X\lrcorner\vp$ is exact. Then there exists some $1$-form $\al$ such that $X\lrcorner\vp=\d\al$. Using the $G_2$-metric defined by $\vp$, we have that $$\begin{split}
6\langle X, X\rangle \Vol_{\vp}&=(X\lrcorner\vp)\w(X\lrcorner\vp)\w\vp \\
&=\d\al\w\d\al\w\vp=\d(\al\w\d\al\w\vp)\\
\end{split}$$ From here, we find that, since $\pd M=\emptyset$, an application of Stokes’ Theorem yields $$\begin{split}
0&\leq 6||X||^2_{L^2}\Vol(M)=\int_M6\langle X, X\rangle \Vol_{\vp} \\
&=\int_M\d(\al\w\d\al\w\vp)=\int_{\pd M}\al\w\d\al\w\vp=0
\end{split}$$ Since $\Vol(M)\neq 0$, we must have that $||X||_{L^2}=0$ proving that $X=0$ as desired.
Corollary \[RocVecNonexistence\] now follows immediately.
As pointed out by an anonymous referee, there is another instance of nonexistence, this time, for $G_2$ vector fields. Specifically, in the case of a compact torsion-free $G_2$-structure, i. e., a $G_2$-manifold, $G_2$ vector fields, being by definition Killing vector fields, will be parallel since $G_2$-manifolds are Ricci flat. In the case of a nontrivial $G_2$ vector field, there would necessarily be a reduction in the holonomy to a proper subgroup of $G_2$. Thus there are no nontrivial $G_2$ vector fields on a compact $G_2$-manifold with full $G_2$-holonomy.
The bracket constructed on Hamiltonian differential forms in Section \[HSMG\] reduces to the usual Poisson bracket on functions in the case of a multisymplectic manifold of degree $2$, that is, a symplectic manifold, and it is a fundamental fact that diffeomorphisms which preserve the symplectic structure can be identified with the diffeomorphisms which preserve this Poisson bracket; however, no such characterization is available for diffeomorphisms which preserve the multisymplectic structure of a general multisymplectic manifold as can be seen using Corollary \[RocVecNonexistence\]. Let $(M,\vp)$ be a closed manifold with closed $G_2$-structure, then the bracket operation operates either on a pair of Rochesterian $1$-forms, on a Rochesterian $1$-form and Rochesterian function or on a pair of Rochesterian functions; since the bracket on Rochesterian functions is necessarily zero and there are no nonzero Rochesterian $1$-forms, any smooth map $\Phi:M\to M$ then trivially preserves this bracket.
By the considerations in Section \[HSMG\], there are the following identifications $$\widetilde{\Om}^1_R(TM)=\Om^1_R(TM)\cong\widetilde{\Om}^1_R(T^*M)
\label{R11}$$ $$\widetilde{\Om}^2_R(TM)\cong\widetilde{\Om}^0_R(T^*M)\cong C^{\infty}(M)/\{f:M\to\R|f\text{ is locally constant}\}
\label{R20}$$ $$\widetilde{\Om}^1_{cR}(TM)=\Om^1_{cR}(TM)\cong\widetilde{\Om}^2_{cR}(T^*M)
\label{cR12}$$ $$\widetilde{\Om}^2_{cR}(TM)\cong\widetilde{\Om}^1_{cR}(T^*M)
\label{cR21}$$ $$\widetilde{\Om}^3_{cR}(TM)\cong\widetilde{\Om}^0_{cR}(T^*M)\cong C^{\infty}(M)/\{f:M\to\R|f\text{ is locally constant}\}
\label{cR30}$$ Equations \[R20\], \[cR30\] immediately give the correspondence $$\widetilde{\Om}^2_R(TM)\cong\widetilde{\Om}^3_{cR}(TM)\cong C^{\infty}(M)/\{f:M\to\R|f\text{ is locally constant}\}$$ Equations \[R11\], \[cR21\] and \[R1cR1\] give an injective map $$\Om^1_R(TM)\hookrightarrow \Om^2_{cR}(TM)$$ Explicitly, if $X\in\Om^1_R(TM)$, then by definition there exists a unique $\al\in\widetilde{\Om}^1_R(T^*M)$ such that $$X\lrcorner\vp=\d\al$$ and, since $\widehat{\sv}_2$ is an isomorphism, there exists a unique $U\in\Om^2(TM)$ such that $$U\lrcorner\sv=\d\al.$$ This completes the identifications of Theorem \[thmmain\]. We remark that there is also the description of $\widetilde{\Om}^1_R(T^*M)$ as $$\widetilde{\Om}^1_R(T^*M)=\{\al\in\Om^1(T^*M):\d\al\in\Om^2_7(T^*M)\}/Z^1(M)$$
For our final consideration, we will need the following lemma which is a slight generalization of [@Kari1 Lemma $2.4.6$] to general $2$-multivector fields, and the proof given here a direct adaptation of that proof.
For $Q\in\Om^2(TM)$, $$(Q\lrcorner\sv)\w(Q\lrcorner\vp)\w\sv=2|Q\lrcorner\vp|_{\vp}^2\Vol_{\vp}
\label{QsvQvpsveq}$$ where $|\cdot|_{\vp}$ is the norm induced by the $G_2$-metric $g_{\vp}$. \[QsvQvpsv\]
By Equations \[Qflat4\], \[svpbe\] $$Q\lrcorner\sv=\star(Q^{\flat}\w\vp)=2\be_7-\be_{14}$$ where $\be_i=\pi_i(Q^{\flat})$ and $\pi_i$ is the projection $\pi_i:\Om^2(T^*M)\to\Om^2_7(T^*M)$. Writing $Q=\be_7+\be_{14}$ and using Equations \[Qflat3\], \[O27\], \[O214\], $$\begin{split}
(Q\lrcorner&\vp)\w\sv=\star(Q^{\flat}\w\sv)\w\sv=\star((\be_7+\be_{14})\w\sv)\w\sv \\
&=\star(\be_7\w\sv)\w\sv+\star(\be_{14}\w\sv)\w\sv=\star\vp\w\star(\sv\w\be_7) \\
&=3\star\be_7 \\
\end{split}$$ which implies $$|(Q\lrcorner\vp)\w\sv|^2_{\vp}=9|\be_7|^2$$ By [@Kari1 Equation $(2.8)$], $$|(Q\lrcorner\vp)\w\sv|_{\vp}^2=3|Q\lrcorner\vp|_{\vp}^2$$ which together yield that $$|\be_7|_{\vp}^2=\frac{|Q\lrcorner\vp|^2}{3}$$ Putting the pieces together yields $$\begin{split}
(Q\lrcorner\sv)&\w(Q\lrcorner\vp)\w\sv=(2\be_7-\be_{14})\w(3\star\be_7) \\
&=6\be_7\w\star\be_7+3\be_{14}\w\star\be_7=6|\be_7|_{\vp}^2\Vol_{\vp}+3g_{\vp}(\be_{14},\be_7)\Vol_{\vp} \\
&=6|\be_7|^2_{\vp}\Vol_{\vp}=2|Q\lrcorner\vp|^2_{\vp}\Vol_{\vp} \\
\end{split}$$
Let $Q\in\Om^2_R(TM)\cap\Om^2_{cR}(TM)$. Then there exist $f\in\Om^0_R(T^*M)$ and $\al\in\Om^1_{cR}(T^*M)$ such that $$Q\lrcorner\vp=\d f$$ and $$Q\lrcorner\sv=\d\al$$ Together with Lemma \[QsvQvpsv\], this implies that $$2|Q\lrcorner\vp|_{\vp}^2\Vol_{\vp}=(Q\lrcorner\sv)\w(Q\lrcorner\vp)\w\sv=\d\al\w\d f\w\sv=\d(\al\w\d f\w\sv)$$ Using Stokes’ Theorem, we find that $$\begin{split}
0\leq&2||Q\lrcorner\vp||_{L^2}\Vol(M)=\int_M2|Q\lrcorner\vp|_{\vp}^2\Vol_{\vp} \\
&=\int_M\d(\al\w\d f\w\sv)=\int_{\pd M}\al\w\d f\w\sv=0
\end{split}$$ Given that $\Vol(M)>0$, this implies that $Q\lrcorner\vp=0$.
This is not necessarily the case for a general $G_2$-manifold. Again, consider $(\R^7,\vp_0)$ with $Q=\frac{\pd}{\pd x^6}\w\frac{\pd}{\pd x^7}$. Then $Q\in\Om^2_R(T\R^7)\cap\Om^2_{cR}(T\R^7)$ because $$Q\lrcorner\vp_0=\d x^1$$ and $$Q\lrcorner\sv_0=\d(x^4\d x^5+x^2\d x^3)$$ which also implies that $Q$ is nonzero in the quotient.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the problem of describing excursion sets of a real-valued function $f$, i.e. the set of inputs where $f$ is above a fixed threshold. Such regions are hard to visualize if the input space dimension, $d$, is higher than 2. For a given projection matrix from the input space to a lower dimensional (usually $1,2$) subspace, we introduce profile sup (inf) functions that associate to each point in the projection’s image the sup (inf) of the function constrained over the pre-image of this point by the considered projection. Plots of profile extrema functions convey a simple, although intrinsically partial, visualization of the set. We consider expensive to evaluate functions where only a very limited number of evaluations, $n$, is available, e.g. $n<100d$, and we surrogate $f$ with a posterior quantity of a Gaussian process (GP) model. We first compute profile extrema functions for the posterior mean given $n$ evaluations of $f$. We quantify the uncertainty on such estimates by studying the distribution of GP profile extrema with posterior quasi-realizations obtained from an approximating process. We control such approximation with a bound inherited from the Borell-TIS inequality. The technique is applied to analytical functions ($d=2,3$) and to a $5$-dimensional coastal flooding test case for a site located on the Atlantic French coast. Here $f$ is a numerical model returning the area of flooded surface in the coastal region given some offshore conditions. Profile extrema functions allowed us to better understand which offshore conditions impact large flooding events.'
author:
- |
Dario Azzimonti\
Istituto Dalle Molle di studi sull’Intelligenza Artificiale (IDSIA)\
Galleria 2, Via Cantonale 2c, 6928 Manno, Switzerland,\
and\
David Ginsbourger\
UQOD, Idiap Research Institute\
Centre du Parc, Rue Marconi 19, 1920 Martigny, Switzerland; and\
IMSV, Department of Mathematics and Statistics, University of Bern\
Alpeneggstrasse 22, 3012 Bern, Switzerland,\
and\
Jérémy Rohmer\
BRGM\
3 Av. C. Guillemin, BP 36009, 45060 Orléans Cedex 2, France,\
and\
Déborah Idier\
BRGM\
3 Av. C. Guillemin, BP 36009, 45060 Orléans Cedex 2, France.
bibliography:
- 'biblio.bib'
title: '**Profile extrema for visualizing and quantifying uncertainties on excursion regions. Application to coastal flooding**'
---
\#1
0
[0]{}
1
[0]{}
[**Profile extrema for visualizing and quantifying uncertainties on excursion regions. Application to coastal flooding**]{}
[*Keywords:*]{} excursion set; set estimation; Gaussian process; profile extrema function; Graphical Methods.
Introduction {#sec:intro}
============
Problem statement
-----------------
In many domains of application (structural reliability, @dubourg2013metamodel, nuclear safety, @Chevalier.etal2014, environmental problems, @BolinLindgren2014, natural risk, @RohmerIdier2012 [@bayarri.etal2009using]) the region in the parameter space leading to response values above or below a given threshold is of crucial interest. In particular here, we assume that a phenomenon can be modelled by a function $f: D \subset {\mathbb{R}}^d \rightarrow {\mathbb{R}}$, where $D$ is a compact set and we focus on excursion sets of the form $$\varGamma = \{ {\mathbf{x}}\in D : f({\mathbf{x}}) \geq {\tau}\},$$ where ${\tau}\in {\mathbb{R}}$ is a prescribed threshold depending on the application at hand. This work presents a method to gain visual insights on $\varGamma$ when the input space dimension $d \geq 3$.
Visualization of point clouds and manifolds in high dimensions is a very active field of research. Principal components analysis and its kernel version [@Scholkopf1997] play a fundamental role in dimensionality reduction for high dimensional data. Parallel coordinates [@Inselberg1985; @Inselberg2009] are widely used for exploratory data analysis. In order to visualize more complex structure, techniques relying on the topological concept of Morse-Smale complex [@Edelsbrunner2008; @Gerber.etal2010] provide powerful, but hard to interpret, tools. Projection based techniques, in particular tours [see, e.g., @Asimov1985; @Cook_etal1995; @Cook2008], are also a prominent method to visualize data in high-dimensional Euclidean spaces. All these techniques, however, are based on a finite set of data points while, in our case, we provide a representation that exploits the structure of $\varGamma$.
The contribution of this paper is two-fold: first we show how to use profile extrema functions for visualizing excursion sets of known deterministic functions; second, we implement these tools for expensive computer experiments with Gaussian process (GP) emulation.
For a given projection matrix $\Psi \in {\mathbb{R}}^{d\times p}$, a profile sup (inf respectively) is a function ${P^{\sup}}_\Psi f \ $ $({P^{\inf}}_\Psi f)$ that associates to each point $\eta$ in the image of $\Psi^T$, the $\sup$ ($\inf$ respectively) of $f$ over all points with projection equal to $\eta$, i.e. ${P^{\sup}}_\Psi f (\eta) = \sup_{\Psi^T {\mathbf{x}}= \eta} f({\mathbf{x}})$. For example, if $\Psi^T = [1, 0, \ldots, 0] \in {\mathbb{R}}^{1 \times d}$ is the canonical projection over the first coordinate, then ${P^{\sup}}_\Psi f(\eta)$ is the $\sup$ of $f$ over all points in $D$ with first coordinate equal to $\eta$. In this case we can plot the values of ${P^{\sup}}_\Psi f$ and ${P^{\inf}}_\Psi f$ over their $1$-dimensional ($p$ dimensional, in general) domain and study such functions. In particular, all $\tilde{\eta}$ such that ${P^{\sup}}_\Psi f(\tilde{\eta}) <{\tau}$, i.e. all inputs ${\mathbf{x}}\in D$ such that $\Psi^T{\mathbf{x}}= \tilde{\eta}$, identify regions of non-excursion. Analogously ${P^{\inf}}_\Psi f(\eta)>{\tau}$ selects regions of excursion. We compute profile extrema for known, fast-to-evaluate functions $f$ with numerical optimization, as described in Section \[sec:profiles\].
In this paper, however, we focus on the case where the function $f$ is very expensive to evaluate and thus the number of evaluations, $n$, is limited, e.g. $n < 100d$. Following the classical computer experiments literature [see, e.g., @Sacks.etal1989; @Santner2013design], we define a design of experiments ${X}_n = ({\mathbf{x}}_1, ,\ldots, {\mathbf{x}}_n ) \in D^n$ and denote with ${\mathbf{y}}_n= (f({\mathbf{x}}_1), \ldots, f({\mathbf{x}}_n))^T$ the corresponding vector of evaluations. We assume that $f$ is a realization of a GP $(Z_{\mathbf{x}})_{{\mathbf{x}}\in D} \sim GP({\mu},{\mathfrak{K}})$ with prior mean function ${\mu}: D \subset {\mathbb{R}}^d \rightarrow {\mathbb{R}}$ and covariance kernel ${\mathfrak{K}}: D \times D \rightarrow {\mathbb{R}}$. The set $\varGamma$ can be estimated from the posterior GP distribution [see, e.g. @Bingham.etal2014; @BolinLindgren2014; @Azzimonti2016], in particular, here we consider the plug-in estimate $$\hat{\varGamma}_n = \{{\mathbf{x}}\in D : \gamma_n({\mathbf{x}}) \geq {\tau}\},$$ where $\gamma_n$ is a posterior quantity related to $Z$. Examples of interesting quantities $\gamma_n$ are the posterior mean, the $1-\alpha$ posterior quantiles or a posterior realization of $Z$. If $d=1,2$, a visualization of $\hat{\varGamma}_n$ only requires evaluations of $\gamma_n$ on a grid; however, when $d\geq 3$, this approach is problematic. We propose a technique to compute and plot profile extrema functions for $\gamma_n$ and we detail how to use profile extrema to identify regions of interest.
A GP model not only provides an estimate for $\varGamma$, but it also allows for a quantification of the associated uncertainty. In particular, here we study the distribution of the stochastic object ${P^{\sup}}_\Psi Z$ and we provide point-wise confidence statements on profile extrema functions. The quantities ${P^{\sup}}_\Psi Z,\ {P^{\inf}}_\Psi Z$ are strongly related to extreme values of Gaussian processes indexed by compact sets which have been widely studied in probability theory [see, e.g., @Adler.Taylor2007; @Azais.Wschebor2009] and have been the subject of recent interest in the computer experiments literature, see, e.g. @Chevalier2013 [Chapter 6] and @Ginsbourger_etal2014. Here we obtain point-wise confidence intervals for profile extrema functions by using posterior quasi-realizations of an approximating process obtained as in @Azzimonti.etal2016. Moreover we introduce a probabilistic bound for the quantiles of the posterior profile function based on the sample quantiles of the approximating process. This procedure enabled the identification of regions of interests for the coastal flooding test case presented in Section \[sec:motivating\].
Outline of the paper
--------------------
In Section \[sec:profiles\] we introduce profile extrema functions for a generic function $\gamma$ and we apply them to a synthetic test case. In this section, we also discuss implementation details. Section \[sec:UQ\] describes the GP model and the quantification of uncertainty on the profiles with approximate posterior GP simulations. Moreover we provide a probabilistic bound (Theorem \[theo:bound\] and Corollary \[cor:bounds\], proofs in Appendix \[sec:proofs\]) for this approximation. In Section \[sec:motivating\] we apply the procedure on a coastal flooding problem where $f$ is evaluated with an expensive-to-evaluate computer experiment. Further details are in Appendix \[sec:full5dRes\]. Finally, in Section \[sec:discussion\], we discuss advantages and drawbacks of the current method and propose possible extensions.
Profile extrema functions {#sec:profiles}
=========================
Definitions {#subsec:defProf}
-----------
In this section, we introduce the concept of profile extrema for a continuously differentiable function $\gamma: D \rightarrow {\mathbb{R}}$, $D \subset {\mathbb{R}}^ d$ compact. While $\gamma$ can be any function, in this paper we mainly consider $\gamma$ as the posterior mean, a quantile or a posterior realization of a GP. The first example of profile extrema are coordinate profiles defined below, i.e. $2 d$ univariate functions describing the extremal behavior of $\gamma$ for each coordinate.
For each $i \in \{1, \ldots, d \}$, let us denote with ${\mathbf{e}}_i$ the $i$-th canonical coordinate vector, then the *i-th coordinate profile extrema* functions are $$\begin{aligned}
{P^{\sup}}_{{\mathbf{e}}_i} \gamma = {P^{\sup}}_i \gamma &: \ \eta \in E_{{\mathbf{e}}_i} \longrightarrow \ {P^{\sup}}_i \gamma(\eta) := \sup_{x_i = \eta}\gamma({\mathbf{x}}) \\
{P^{\inf}}_{{\mathbf{e}}_i} \gamma = {P^{\inf}}_i \gamma &: \ \eta \in E_{{\mathbf{e}}_i} \longrightarrow \ {P^{\inf}}_i \gamma(\eta) := \inf_{x_i = \eta} \gamma({\mathbf{x}}).
\end{aligned}$$ where $E_{{\mathbf{e}}_i} = \{ \eta \in {\mathbb{R}}: {\mathbf{e}}_i^T {\mathbf{x}}= \eta, \ {\mathbf{x}}\in D \}$ and ${\mathbf{x}}= (x_1, \ldots, x_i, \ldots, x_d)$.
By studying these functions we can split the input space $D$ in simple regions providing a marginal visualization of $\varGamma$. Consider the first coordinate profile sup function ${P^{\sup}}_1 \gamma$. If there exists $\eta^*$ such that ${P^{\sup}}_1 \gamma(\eta^*) < {\tau}$, the function $\gamma$ is below ${\tau}$ for all ${\mathbf{x}}\in D_{i,\eta^ *}= \{ {\mathbf{x}}\in D : x_i = \eta^* \}$. In a symmetric way, if $\eta^*$ is such that ${P^{\inf}}_1 \gamma(\eta^*) \geq {\tau}$, the function $\gamma$ is above ${\tau}$ over the whole $D_{i,\eta^*}$. Since ${P^{\sup}}_1$ and ${P^{\inf}}_1$ are univariate functions we can evaluate them over a discretization of $E_{{\mathbf{e}}_1}$ and with a simple plot we can locate those regions. Profile functions could also lead to more undetermined situations: for example, for $\tilde{\eta}$ such that ${P^{\sup}}_1 \gamma(\tilde{\eta}) \geq {\tau}$, there exists at least one point $\bar{{\mathbf{x}}}^{\tilde{\eta}} = (\tilde{\eta}, \bar{x}_2, \ldots, \bar{x}_d) \in D_{1,\tilde{\eta}}$ such that $\gamma(\bar{{\mathbf{x}}}^{\tilde{\eta}}) \geq {\tau}$ and we cannot flag the region $D_{1,\tilde{\eta}}$ neither as excursion nor as non-excursion. Coordinate profile extrema are the results of constrained optimization problems over particular projections on the coordinate axes. Profile extrema functions further generalize this concept.
Consider a full column rank matrix $\Psi \in {\mathbb{R}}^{d \times p}$ with $p < d$. Let $E_\Psi := \{\Psi^T {\mathbf{x}}: {\mathbf{x}}\in D \}$ and $D_{\Psi,\eta} := \{ {\mathbf{x}}\in D : \Psi^T {\mathbf{x}}= \eta \}$, then the functions $$\begin{aligned}
{P^{\sup}}_\Psi \gamma &: \eta \in E_\Psi \longrightarrow \ {P^{\sup}}_\Psi \gamma(\eta) := \sup_{{\mathbf{x}}\in D_{\Psi, \eta}}\gamma({\mathbf{x}}) \\
{P^{\inf}}_\Psi \gamma &: \eta \in E_\Psi \longrightarrow \ {P^{\inf}}_\Psi \gamma(\eta) := \inf_{{\mathbf{x}}\in D_{\Psi, \eta}}\gamma({\mathbf{x}}),
\end{aligned}$$ are called *Profile sup* and *Profile inf* functions of $\gamma$ respectively. The set $E_\Psi$ is the image of $D$ under $\Psi^T$ and $D_{\Psi,\eta}$ is the pre-image of $\eta$ under $\Psi^T$ restricted to $D$. \[def:profExtrema\]
Since we are interested in visualization here we only consider $p \in \{1,2\}$ and $\Psi$ as orthogonal projections because they have a more direct interpretation.
Analytical example {#subsec:analExample}
------------------
Let us review the concepts just introduced on an analytical test case. Consider $$\gamma(\mathbf{x}) = \sin\left( a \mathbf{v_1}^T \mathbf{x} +b\right) + \cos\left(c \mathbf{v_2}^T \mathbf{x} +d \right) \qquad \mathbf{x} \in [0,1]^2, \ \mathbf{v_1}, \mathbf{v_2} \in \mathbb{R}^2
\label{eq:analExample}$$ where $a,b,c,d \in {\mathbb{R}}$, $\mathbf{v_1} = [\cos(\theta),\sin(\theta)]^T$, $\mathbf{v_2} = [\cos(\theta+\pi/2),\sin(\theta+\pi/2)]^T$ and $\theta= \pi/6$.
Figure \[fig:analyticalFun\] shows contour lines of $\gamma$ with parameters chosen as $a=1,b=0,c=10,d=0$. We fix a threshold ${\tau}=0$ and we compute the coordinate profile extrema, plotted in Figure \[fig:coordAnalytic\]. In this case, ${P^{\sup}}_1 \gamma(\eta) <{\tau}$ for $\eta \in [0,0.13]$ and ${P^{\sup}}_2 \gamma(\eta) <{\tau}$ for $\eta \in [0,0.25]\cup [0.70,0.80]$. The second coordinate seems more informative on the excursion: ${P^{\sup}}_1 \gamma$ flags as non-excursion a region with volume $0.13$ while ${P^{\sup}}_2 \gamma$ flags a region with volume $0.35$. The true set has volume $0.127$, i.e. the region of non excursion has volume $0.873$. This example shows that canonical directions might not be the most appropriate ones to visualize this function. Figure \[fig:obliqueAnalytic\] shows profile extrema computed along the (oblique) generating directions $\mathbf{v_1},\mathbf{v_2}$, i.e. profile extrema in Definition \[def:profExtrema\] with $\Psi= \mathbf{v_1}$ or $\mathbf{v_2}$. By using oblique directions we can say that for ${\mathbf{x}}$ such that $\mathbf{v_1}^T{\mathbf{x}}\in [0,0.52] \cup [1.22,1.37]$ (area excluded $0.33$) and for ${\mathbf{x}}$ such that $\mathbf{v_2}^T {\mathbf{x}}\in [-0.5,-0.1] \cup [0.11,0.54] \cup [0.71,0.87]$ (area excluded $0.68$) there is no excursion. As shown in Figure \[fig:anObliqueFun\] profile extrema along this direction allow us to exclude a much larger portion (total area excluded $0.79$ versus $0.44$ with coordinate profiles) of the input space. In Appendix \[sec:bivEx3d\] we combine coordinate, oblique and bivariate profiles extrema for a finer visualization of the excursion set on a $3$-dimensional analytical function inspired by .
![Profile extrema functions for $\gamma$ (black solid) and their approximation (blue dashed) from $k=15$ points (circles). Threshold ${\tau}=0$, horizontal solid line.[]{data-label="fig:coordAnalytic"}](./FunctionCoordProfiles.pdf){width="\linewidth"}
![Profile extrema functions for $\gamma$ (black solid) and their approximation (blue dashed) from $k=15$ points (circles). Threshold ${\tau}=0$, horizontal solid line.[]{data-label="fig:coordAnalytic"}](./coordinateProfiles.pdf){width="\linewidth"}
![Oblique profile extrema functions for $\gamma$ (black solid lines) and their approximation (blue dashed lines) from $k=15$ points.[]{data-label="fig:obliqueAnalytic"}](./FunctionObliqueProf.pdf){width="\linewidth"}
![Oblique profile extrema functions for $\gamma$ (black solid lines) and their approximation (blue dashed lines) from $k=15$ points.[]{data-label="fig:obliqueAnalytic"}](./ObliqueProfiles.pdf){width="\linewidth"}
Implementation {#subsec:implementation}
--------------
The value of the profile sup ${P^{\sup}}_{\Psi} \gamma(\eta)$, for a fixed matrix $\Psi \in {\mathbb{R}}^{d\times p}$, is the solution of a constrained optimization problem. In general, $\gamma$ is not convex, therefore we are not guaranteed to achieve a global optimum. We implement a local method to obtain profile extrema functions for $\Psi ={\mathbf{e}}_i \in {\mathbb{R}}^d, i \in \{1,\ldots,d\}$, and for a matrix $\Psi \in {\mathbb{R}}^{d\times p}$ of rank $p$.
In the first case we obtain the value for ${P^{\sup}}_i \gamma$, for a coordinate $i$ with the following procedure. For each $\eta$, define a new function $\tilde{\gamma}_{i,\eta}: {\mathbb{R}}^{d-1} \rightarrow {\mathbb{R}}$ such that $\tilde{\gamma}_i({\mathbf{z}}) := \gamma(z_1, \ldots,z_{i-1},\eta,z_{i}, \ldots, z_{d-1})$ for any ${\mathbf{z}}= (z_1, \ldots, z_{d-1}) \in D_{-i}$, where $D_{-i}$ is the $d-1$ dimensional restriction of $D$ without the $i$th coordinate. We can estimate the value ${P^{\sup}}_i \gamma(\eta)$ with a L-BFGS-B algorithm. If the gradient of $\gamma$ is available it is used in the optimization, otherwise numerical derivatives are used. This method depends on the starting point chosen for the optimization. In general we are interested in the evaluation of the functions ${P^{\sup}}_i \gamma(\eta)$ for all $\eta$ and for all $i \in \{1, \ldots, d\}$. By exploiting the smoothness of $\gamma$ we re-use the previously obtained points of optimum as starting point for the subsequent optimizations.
In the second case, we assume that the matrix $\Psi$ has full column rank and that $D$ is a hyper-cube. The value ${P^{\sup}}_{\Psi} \gamma(\eta)$ is the solution of the constrained optimization problem $$\text{maximize } \gamma({\mathbf{x}}) \qquad \text{ subject to } \Psi^T{\mathbf{x}}= \eta, \ {\mathbf{x}}\in D.
\label{eq:optFormulation}$$ We transform in an unconstrained optimization problem first by computing an arbitrary $\xi \in {\mathbb{R}}^d$ such that $\Psi^T\xi = \eta$ and then by observing that if ${\mathbf{x}}$ satisfies the equality constraint then ${\mathbf{x}}= \xi + \operatorname{Null}(\Psi^T){\mathbf{z}}$, where $\operatorname{Null}(\Psi^T) \in {\mathbb{R}}^{d \times (d-p)}$ is the matrix representing the null space of $\Psi^T$ and ${\mathbf{z}}\in {\mathbb{R}}^{d-p}$. We can then maximize $\gamma$ over the $d-p$ dimensional space using a (hard) barrier function for the hyper-cube inequality constraints. Both $\xi$ and the starting point(s) for the interior point method are computed by solving auxiliary linear programs.
In both cases, a multi-start optimization procedure provides a more reliable estimate of the global optimum at an increased computational cost. The two methods are implemented in the R programming language [@Rcore] in the package `profExtrema`, available on CRAN. The first method uses the L-BFGS-B [@Byrd.etal1995] implementation in the function `optim`, base package; the second method solves the auxiliary linear programs with `lpcdd` from the package `rcdd` (a R interface for `cddlib`, @Fukuda_cddlib) and optimizes the barrier function with the BFGS implementation in `optim`.
Approximation of profile extrema functions {#subsec:approxProf}
------------------------------------------
The cost of evaluating the profile extrema functions at each $\eta$ increases as $d - p$ grows larger. Plots of profile extrema functions with $\Psi\in {\mathbb{R}}^{d\times p}$ with $p=1$ or $2$ require evaluations over $E_\Psi \subset {\mathbb{R}}^p$. If the input dimension $d$ is high, each evaluation is a high dimensional optimization and visualizing the function at sufficient resolution becomes expensive. We mitigate this issue by approximating ${P^{\sup}}_\Psi \gamma (\eta)$ at any $\eta$ with an interpolation of profile function evaluations at few space filling points $\eta_1, \ldots, \eta_k \in E_\Psi$.
Coordinate profile extrema functions can be seen as a special type of optimal value functions. First order characterizations of such objects have been studied in optimization and perturbation theory, see, e.g., @Danskin1967 [@BonnansShapiro1998]. In particular, consider the $i$-th coordinate profile extrema function ${P^{\sup}}_i \gamma$. We can write $\gamma: D_{\mid -i} \times D_{\mid i} \rightarrow {\mathbb{R}}$ and ${P^{\sup}}_i \gamma(\eta) = \sup_{y\in D_{\mid -i}} \gamma(y,\eta)$. If $\gamma(y,\cdot)$ is differentiable with $\nabla_\eta \gamma(y,\cdot)$ continuous for $y \in D_{\mid -i}$ and there exists a unique $y^* = \arg \max \gamma(y,\eta)$, then $\nabla_\eta {P^{\sup}}_i \gamma(\eta) = \nabla_\eta \gamma(\cdot,\cdot)_{\arrowvert_{(y^*,\eta)}}$. We can then evaluate ${P^{\sup}}_{i} \gamma(\eta)$ and $\nabla_\eta\eta {P^{\sup}}_{i}\gamma(\eta)$ at few space filling points $\eta_1, \ldots, \eta_k$, for a small $k>0$, and interpolate the value ${P^{\sup}}_{i} \gamma(\eta)$ at any $\eta \in E_\Psi$ using a first order approximation. For arbitrary $\Psi$, we approximate ${P^{\sup}}_\Psi \gamma (\eta)$ at any $\eta$ with cubic splines or kriging interpolators as they do not require $\nabla_\eta {P^{\sup}}_\Psi$.
Below we present an example for the case $p=1$ with a cubic spline approximation. The blue dashed lines in Figures \[fig:coordAnalytic\], \[fig:obliqueAnalytic\] show the approximation of coordinate and oblique profile extrema functions, the exact functions are also plotted in black solid lines. We select $k=15$ points $\eta_1, \ldots, \eta_k \in E_{{\mathbf{e}}_i}$ with Latin hyper-cube sampling, the approximate profiles at any point are computed with a cubic spline regression from $k$ exact calculations. Exact coordinate profiles ${P^{\sup}}_i,{P^{\inf}}_i$, for $i=1,2$ over a grid of equispaced $100$ points on $[0,1]$ require $0.25$ seconds and oblique profiles ${P^{\sup}}_{\mathbf{v_i}},{P^{\inf}}_{\mathbf{v_i}}$, $i=1,2$ require $0.64$ seconds on a laptop with 2.6 GHz Intel Core i5-7300U CPU. On the same laptop, cubic spline approximations plotted on the same grid require $0.04$ seconds for coordinate profiles ($0.13$ oblique). On this example, the maximum absolute median errors achieved with the approximation of ${P^{\inf}}\gamma$, ${P^{\sup}}\gamma$ are $0.07 \%$ and $0.54 \%$ for coordinate profiles and $0.19 \%$ and $5.08 \%$ for oblique profiles. The number of approximation points, $k$, is an important parameter for the approximation, the choice of $k$ boils down to a trade-off between computational time and precision. For 1-dimensional profile extrema ($p=1$) the default choice implemented is $k=10\sqrt{d}$. When the profile functions are smooth this heuristics often leads to good results, however it is generally better to compare different values for $k$. Cubic splines are implemented here only for $p=1$, for $p=2$ we use kriging.
Uncertainty quantification for profile extrema {#sec:UQ}
==============================================
Profile extrema for expensive-to-evaluate functions {#subsec:introGP}
---------------------------------------------------
In the previous section we introduced profile extrema functions to visualize excursion sets of arbitrary functions. This method, however, requires performing potentially very high numbers of function evaluations which can be prohibitive in many computer experiments, as, for example, in the motivating application of Section \[sec:motivating\]. We therefore rely on a Gaussian process emulator based on few evaluations of the expensive function. We consider a design of experiments (DoE) with $n$ points ${X}_n \in D^n$ and the corresponding function evaluations $\mathbf{y}_n$ as described in Section \[sec:intro\]. We select a prior GP $(Z_{{\mathbf{x}}})_{{\mathbf{x}}\in D} \sim GP({\mu},{\mathfrak{K}})$ with mean ${\mu}({\mathbf{x}}) = \sum_{j=1}^m c_j h_j({\mathbf{x}})$, where $\mathbf{c} = (c_1, \ldots, c_m)$ is a vector of unknown coefficients and $h_j$ are known basis functions. The covariance kernel is typically selected from a stationary parametric family, e.g. Matérn family, and its hyper-parameters are estimated with maximum likelihood. Here we use the R package `DiceKriging` [@Roustant.etal2012] for fitting and prediction. Given the data and hyper-parameters the posterior mean and covariance kernel can be computed with the following universal kriging (UK) equations. $$\begin{aligned}
\label{eq:postMean}
{\mu}_n({\mathbf{x}}) &= \mathbf{h}({\mathbf{x}})^T\widehat{\mathbf{c}} + {\mathfrak{K}}({\mathbf{x}}, {X}_n){\mathfrak{K}}({X}_n,{X}_n)^{-1} \left({\mathbf{y}}_n - {H}\hat{\mathbf{c}} \right) \\ \label{eq:postCov}
{\mathfrak{K}}_n({\mathbf{x}},{\mathbf{x}}^\prime) &= {\mathfrak{K}}({\mathbf{x}},{\mathbf{x}}^\prime) - {\mathfrak{K}}({\mathbf{x}}, {X}_n){\mathfrak{K}}({X}_n, {X}_n)^{-1}{\mathfrak{K}}({\mathbf{x}}^\prime, {X}_n)^T + \\ \nonumber
&\boldsymbol{\lambda}({\mathbf{x}},{X}_n)({H}^T{\mathfrak{K}}({X}_n, {X}_n)^{-1}{H})^{-1}\boldsymbol{\lambda}({\mathbf{x}}^\prime,{X}_n)^T, \ \text{for } {\mathbf{x}},{\mathbf{x}}^\prime \in D \qquad \text{with} \\ \nonumber
\boldsymbol{\lambda}({\mathbf{x}},{X}_n) &= (\mathbf{h}({\mathbf{x}})^T - {\mathfrak{K}}({\mathbf{x}}, {X}_n){\mathfrak{K}}({X}_n, {X}_n)^{-1}{H}) \qquad \text{ and} \\ \nonumber
\widehat{\mathbf{c}} &= ({H}^T{\mathfrak{K}}({X}_n, {X}_n)^{-1}{H})^{-1}{H}^T{\mathfrak{K}}({X}_n, {X}_n)^{-1}{\mathbf{y}}_n\end{aligned}$$ where $\mathbf{h}({\mathbf{x}}) = (h_1({\mathbf{x}}), \ldots, h_m({\mathbf{x}}))^T$, ${H}= [h_j({\mathbf{x}}_i)]_{i=1,\ldots,n, j=1,\ldots, m}$, ${\mathfrak{K}}({\mathbf{x}},{X}_n) = ({\mathfrak{K}}({\mathbf{x}},{\mathbf{x}}_1), \ldots, {\mathfrak{K}}({\mathbf{x}},{\mathbf{x}}_n) )$, for ${\mathbf{x}}\in D$ and ${\mathfrak{K}}({X}_n,{X}_n) = ({\mathfrak{K}}({\mathbf{x}}_i,{\mathbf{x}}_j))_{i,j=1, \ldots, n} \in {\mathbb{R}}^{n \times n}$. If $m=1$ and $h_1 \equiv 1$ we have ordinary kriging (OK) equations.
Once we have fitted the GP, profile extrema functions can be directly computed with the posterior mean, leading to a visualization of a plug-in estimate $\hat{\varGamma} = \{ {\mathbf{x}}\in D : {\mu}_n(x) \geq {\tau}\}$ for $\varGamma$. The GP assumption allows us to also quantify the uncertainty on profile extrema functions, at fixed hyper-parameters. In particular, in the next section, we develop a Monte Carlo method that exploits approximate posterior realizations to obtain point-wise confidence intervals for ${P^{\sup}}_{{\mathbf{e}}_i} f$ and ${P^{\inf}}_{{\mathbf{e}}_i} f$, for $i=1, \ldots, d$.
Approximation for posterior field realizations {#subsec:approxReals}
----------------------------------------------
Consider the GP $(Z_{\mathbf{x}})_{{\mathbf{x}}\in D}$, we denote with $Z_{\mathbf{x}}(\omega)$ a realization of the process at ${\mathbf{x}}$, for $\omega \in \Omega$, the underlying probability space. We are interested in posterior realizations of $Z_{\mathbf{x}}$ given $Z_{{X}_n} = \mathbf{y}_n$. A brute force approach would involve simulating ${s}$ times the field over a space filling design on $D$ and then evaluating the profile extrema function by taking the discrete extrema. This procedure however is very costly as it requires ${s}$ exact simulations of the field over many points and its results strongly depend on the chosen discretization. Following @Azzimonti.etal2016 we use an approximating process in order to reduce the computational cost. We consider the sequence of points $G= ({\mathbf{g}}_1, \ldots, {\mathbf{g}}_\ell) \in D^\ell$ and the *approximating process of $Z$ based on $G$* $$\widetilde{Z}_{\mathbf{x}}= a({\mathbf{x}}) + \mathbf{b}^T({\mathbf{x}}) Z_G, \qquad {\mathbf{x}}\in D,
\label{eq:tildeZ}$$ where $a: D \rightarrow {\mathbb{R}}$ is a trend function and $\mathbf{b}: D \rightarrow {\mathbb{R}}^\ell$ is a continuous vector-valued function of deterministic weights, and $Z_G = (Z_{{\mathbf{g}}_1}, \ldots, Z_{{\mathbf{g}}_\ell})$ is the $\ell$-dimensional random vector given by the values of the original process $Z$ at $G$. Here we select $a({\mathbf{x}}) = {\mu}({\mathbf{x}})$, $\mathbf{b}({\mathbf{x}}) = {\mathfrak{K}}(G,G)^{-1}{\mathfrak{K}}({\mathbf{x}},G)^T$ where ${\mathfrak{K}}({\mathbf{x}},G) = ({\mathfrak{K}}({\mathbf{x}},{\mathbf{g}}_1), \ldots, {\mathfrak{K}}({\mathbf{x}},{\mathbf{g}}_l)) \in {\mathbb{R}}^{1\times \ell}$ and ${\mathfrak{K}}(G,G) = [{\mathfrak{K}}({\mathbf{g}}_i,{\mathbf{g}}_j)]_{i,j =1, \ldots, \ell} \in {\mathbb{R}}^{\ell \times \ell}$. In what follows, the points $G$ are called *pilot points*, borrowing the term from geostatistics [see,e.g. @Scheidt2006 Chapter 4.2], and here they are selected with Algorihtm B from @Azzimonti.etal2016. The number of pilot points $\ell$ can be empirically chosen by stopping when the optimum of Algorithm B’s objective function stabilizes around a value. In moderate dimensions, usually $\ell$ between $50$ and $150$ leads to distributions close to the true one. The bound introduced in section \[subsec:bound\] and, in particular, the related tightness indicator, equation , can also be used to choose the number of pilot points. For a fixed $\omega \in \Omega$, $\widetilde{Z}_{\mathbf{x}}(\omega)$ is a function of ${\mathbf{x}}$ that has an analytical expression and only requires posterior simulations of the original field $Z$ at $G$ and linear operations with pre-calculated kriging weights. Moreover the gradient of $\widetilde{Z}_{\mathbf{x}}(\omega)$ with respect to ${\mathbf{x}}$ is known, see Appendix \[sec:gradTildeZ\], therefore we compute profile extrema for each realization, ${P^{\sup}}_\Psi \widetilde{Z}_{\mathbf{x}}(\omega)$ and ${P^{\inf}}_\Psi \widetilde{Z}_{\mathbf{x}}(\omega)$, with the algorithms introduced in Section \[subsec:implementation\].
Analytical example under uncertainty {#subsec:analExUQ}
------------------------------------
Let us consider the analytical example illustrated in Section \[subsec:analExample\]. We evaluate the function in Equation only at $n$ points and we approximate it on a dense grid $100\times 100$ with a GP. We consider two DoE with $n=20$ and $n=90$, both obtained with (randomized) maximin Latin hypercube sampling (R package `DiceDesign`, @Dupuy_dDesign_2015) and we choose a prior GP with constant mean and Matèrn covariance kernel with $\nu=5/2$. We estimate the covariance hyper-parameters with maximum likelihood and the constant mean value with the ordinary kriging formulae. The model with $n=20$ evaluations provides a rough approximation for $f$, while the model with $n=90$ is an accurate reconstruction. The $Q^2$ criterion[^1] computed over a dense grid of $N_{test}=10000$ test data is equal to $0.86$ for the posterior mean with $n=20$ and $0.99$ for $n=90$. For both models, we consider $\ell=80$ pilot points and we compute the point-wise confidence intervals for the profile extrema on the mean with Monte Carlo simulations. Figure \[fig:ex2dmean\_20\] shows the posterior GP mean obtained from $n=20$ evaluations of the function in Equation , the regions delimited by the oblique profile extrema on the mean along the directions $\mathbf{v_1},\mathbf{v_2}$ and its $90\%$ point-wise confidence intervals for the threshold ${\tau}= 0$. Figure \[fig:ex2dprof\_20\] shows the profile extrema for the mean, the profile extrema for ${s}=150$ approximate realizations of the posterior process and the empirical point-wise $90\%$ confidence intervals (dark shaded tube, red). Figure \[fig:ex2dprof\_90\] shows profile extrema for the mean on the model with $n=90$ observations. In this case the confidence intervals (dark shaded tube, red) are much tighter indicating a smaller uncertainty.
![[Profile extrema functions ($n=20$, ${\tau}=0$): mean (black, solid), breakpoints (vertical green; mean: solid, quantiles: dashed), quantiles ($90 \%$, black dashed, CI dark red tube), upper-lower bound (light blue tube).]{}[]{data-label="fig:ex2dprof_20"}](./profMean_UQ_20.pdf){width="\linewidth"}
![[Profile extrema functions ($n=20$, ${\tau}=0$): mean (black, solid), breakpoints (vertical green; mean: solid, quantiles: dashed), quantiles ($90 \%$, black dashed, CI dark red tube), upper-lower bound (light blue tube).]{}[]{data-label="fig:ex2dprof_20"}](./prof_UQ_bound_approx_20.pdf){width="\linewidth"}
![Profile extrema functions ($n=90$, ${\tau}=0$): mean (black, solid), breakpoints (vertical green; mean: solid, quantiles: dashed), quantiles ($90 \%$, black dashed, CI dark red tube), upper-lower bound (light blue tube).[]{data-label="fig:ex2dprof_90"}](./profMean_UQ_90.pdf){width="\linewidth"}
![Profile extrema functions ($n=90$, ${\tau}=0$): mean (black, solid), breakpoints (vertical green; mean: solid, quantiles: dashed), quantiles ($90 \%$, black dashed, CI dark red tube), upper-lower bound (light blue tube).[]{data-label="fig:ex2dprof_90"}](./prof_UQ_bound_approx_90.pdf){width="\linewidth"}
Bounds for the approximation {#subsec:bound}
----------------------------
The approximating process $\widetilde{Z}$ does not provide proper probabilistic statements for ${P^{\sup}}_\Psi Z, {P^{\inf}}_\Psi Z$ as it is based on the $\ell$-dimensional random vector $Z_G$. However since the distribution of the error process $Z-\widetilde{Z}$ can be expressed in closed form, we can control the approximating error with the following probabilistic bound. Proofs are in Appendix \[sec:proofs\].
Consider a Gaussian process $(Z_{\mathbf{x}})_{{\mathbf{x}}\in D} \sim GP({\mu},{\mathfrak{K}})$, the approximating process $\widetilde{Z}$ of $Z$ based on the points $G$ (Equation ) and $T \subset D$, then for any $u>{\mu}^{\widetilde{\Delta}}_T$ $$Pr\left( \left\lvert \sup_{{\mathbf{x}}\in T} Z_{\mathbf{x}}- \sup_{{\mathbf{x}}\in T} \widetilde{Z}_{\mathbf{x}}\right\rvert > u \right) \leq 2 \exp \left( -\frac{(u- {\mu}^{\widetilde{\Delta}}_T )^2 }{2 (\sigma^{\widetilde{\Delta}}_T)^2} \right),
\label{eq:supBound}$$ where $$\begin{aligned}
\label{eq:mmvvDelta}
{\mu}^{\widetilde{\Delta}}_T &=\sup_{{\mathbf{x}}\in T} \lvert {\mu}^{\widetilde{\Delta}}({\mathbf{x}}) \rvert \text{ and } (\sigma^{\widetilde{\Delta}}_T)^2 = \sup_{{\mathbf{x}}\in T} {\mathfrak{K}}^{\widetilde{\Delta}}({\mathbf{x}},{\mathbf{x}}) \text{ with } \\ \nonumber
{\mu}^{\widetilde{\Delta}}({\mathbf{x}}) &= {\mathbb{E}}[Z_{\mathbf{x}}- \widetilde{Z}_{\mathbf{x}}] = {\mu}({\mathbf{x}}) - a({\mathbf{x}}) - \mathbf{b}^T({\mathbf{x}}){\mu}(G) \qquad {\mathbf{x}}\in T \\ \nonumber
{\mathfrak{K}}^{\widetilde{\Delta}}({\mathbf{x}},{\mathbf{y}}) &=
{\mathfrak{K}}({\mathbf{x}},{\mathbf{y}}) - {\mathfrak{K}}({\mathbf{y}},G)\mathbf{b}({\mathbf{x}}) - {\mathfrak{K}}({\mathbf{x}},G)\mathbf{b}({\mathbf{y}}) + \mathbf{b}^T({\mathbf{x}}){\mathfrak{K}}(G,G)\mathbf{b}({\mathbf{y}}), \qquad {\mathbf{x}},{\mathbf{y}}\in T
\end{aligned}$$ \[theo:bound\]
If the approximating process $\widetilde{Z}$ is unbiased then ${\mu}^{\widetilde{\Delta}}({\mathbf{x}}) \equiv 0$ and the inequality in is valid for any $u>0$.
\[cor:bounds\] Consider the profile sup random functions defined as $${P^{\sup}}_{\Psi} \widetilde{Z} (\eta) := \sup_{ \{{\mathbf{x}}\in D : \Psi^T{\mathbf{x}}= \eta \}} \widetilde{Z}_{\mathbf{x}}\ \text{ and } \ {P^{\sup}}_{\Psi} Z (\eta) := \sup_{ \{{\mathbf{x}}\in D : \Psi^T{\mathbf{x}}= \eta \}} Z_{\mathbf{x}}, \qquad \eta \in E_\Psi,$$ where $\widetilde{Z}$ is an unbiased approximate process in the form of Equation . For any $\eta^* \in E_\Psi$ and $\alpha \in (0,1)$ $$\begin{gathered}
\nonumber
Pr\big({P^{\sup}}_{\Psi} Z (\eta^*) \leq u^+_{\alpha} \big) \geq 1- \alpha, \qquad \text{ with } \ u^+_{\alpha} = \tilde{u}^+_\beta+\sqrt{2 \left(\sigma^{\widetilde{\Delta}}_T\right)^2 \log\left(\frac{2}{\alpha -\beta}\right)} \text{ and } \\
Pr\big({P^{\sup}}_{\Psi} Z (\eta^*) \geq u^-_{\alpha} \big) \geq 1- \alpha, \qquad \text{ with } \ u^-_{\alpha} = \tilde{u}^-_\beta -\sqrt{2 \left(\sigma^{\widetilde{\Delta}}_T\right)^2 \log\left(\frac{2}{\alpha -2\beta}\right)}
\label{eq:profBoundCor}
\end{gathered}$$ with $(\sigma^{\widetilde{\Delta}}_T)^2$ as in , $T= \{ {\mathbf{x}}\in D: \Psi^T{\mathbf{x}}= \eta \}$, $\alpha > 2\beta$ and $\tilde{u}^+_\beta,\tilde{u}^-_\beta$ are the $1-\beta, \beta$ quantiles for ${P^{\sup}}_{\Psi} \widetilde{Z} (\eta^*)$ respectively. Equations imply $Pr\big({P^{\sup}}_{\Psi} Z (\eta^*) \in [u^-_\alpha, u^+_\alpha] \big) \geq 1-2\alpha$.
In practice the quantiles $\tilde{u}^\pm_\beta$ in equation are estimated with sample quantiles $\widehat{\tilde{u}^\pm_\beta}$ from the realizations of $\widetilde{Z}$ and $u^\pm_\alpha$ are estimated by plugging-in $\widehat{\tilde{u}^\pm_\beta}$ in equation .
Figures \[fig:ex2dprof\_20\] and \[fig:ex2dprof\_90\] show the conservative bound in equation on ${P^{\sup}}_i Z$ (skyblue, lightly shaded tube) and ${P^{\inf}}_i Z$ (seagreen, lightly shaded), $i=1,2$, on the example presented in Section \[subsec:analExUQ\]. The uncertainty is much smaller with $n=90$, however the bound is still very conservative. An indicator for the bound tightness is the quantity $(\sigma^{\widetilde{\Delta}}_T)^2(\eta,\ell)$, for each $\eta \in E_\Psi$. Here we explicit the dependency of the approximation on $\ell$. In particular, we study the integral of this map, i.e. $$I(\sigma^{\widetilde{\Delta}}_T)^2(\ell) := \int_{E_\Psi} (\sigma^{\widetilde{\Delta}}_T)^2 (\eta,\ell) d\eta
\label{eq:integratedVarDiff}$$ Appendix \[sec:deltaTex2d\] shows a comparison of $I(\sigma^{\widetilde{\Delta}}_T)^2$ from the example in figure \[fig:ex2dprof\_90\], as a function of $\ell$. Figure \[fig:ex2dDeltaTcomparison\], in particular, shows exponential decrease in $I(\sigma^{\widetilde{\Delta}}_T)^2$ as $\ell$ grows to $80$.
Discussion on method’s parameters
---------------------------------
Profile extrema functions for an expensive-to-evaluate function require the user to set several parameters summarized in Table \[tab:paramSummary\], Appendix \[sec:deltaTex2d\]. The DoE selected to train the GP model has the biggest influence on the profile extrema uncertainty. Figures \[fig:ex2dprof\_20\] and \[fig:ex2dprof\_90\] show that a GP model trained on more function evaluations, reduces the uncertainty on profiles. The bounds on profile extrema, equation , however, are also controlled by two parameters chosen by the user given a fixed DoE: the number of pilot points, $\ell$, and of GP realizations, ${s}$. The number of posterior GP realizations, ${s}$, controls the accuracy of the empirical quantiles $\tilde{u}^\pm_\beta$. More importantly, the number of pilot points $\ell$ controls directly the tightness of the bounds in equation . This parameter plays a more prominent role in uncertainty quantification as more pilot points lead to a smaller $\sigma^{\widetilde{\Delta}}_T$ and a tighter bound, see, e.g., Figure \[fig:ex2dDeltaTcomparison\], Appendix \[sec:deltaTex2d\].
Motivating application: coastal flooding {#sec:motivating}
========================================
Motivation and study case
-------------------------
Coastal flooding models experienced recent progresses opening new research and applications perspectives. However, their computational cost ($>$ hours) hinders their use when a large number of simulations is required for estimating the excursion region corresponding to the critical forcing conditions leading to inundation or when forecast is needed [@RohmerIdier2012; @idier13].
We focus here on coastal flooding induced by overflow and we consider the Boucholeurs area (French Atlantic coast, see Figure \[fig:LOC\]). This area is located close to La Rochelle and was flooded during the 2010 Xynthia storm event. This event was characterized by a high storm surge ($>1.5 {\text{ m}}$ at La Rochelle tide gauge) in phase with a high spring tide [@bertin14]. Here, we focus on these primary drivers (tide and storm surge) and on how they affect the resulting flooded surface ($Y$, in square meters).
![\[fig:LOC\]Study site location (left) and computational domain limits (right, in white) and location of the forcing conditions ${\mathbf{x}}$ (right, in blue).](./Fig_LOC.pdf){width="0.85\linewidth"}
### The forcing conditions
The offshore forcing conditions correspond to the tide and storm surge temporal evolution (see Figure \[fig:X\]a) are denoted with ${\mathbf{x}}= (T,S,t_0,t_+,t_-)$. They are parametrized as follows:
- the tide is simplified by a sinusoidal signal parametrised by its high tide level $T \in [0.95 {\text{ m}},3.70 {\text{ m}}]$, (see Figure \[fig:X\]a);
- the surge signal is assumed to be described by a triangular model (see Figure \[fig:X\]a) using four parameters: the peak amplitude $S \in [0.65 {\text{ m}},2.50 {\text{ m}}]$, the phase difference $t_0\in [-6,6]$ hours, between the surge peak and the high tide, the time duration of the raising part $t_- \in [-12.0,-0.5]$ hours, and the falling part $t_+ \in [0.5,12.0]$ hours.
 Schematic representation of the tide and surge temporal signals and the different parameters describing them. (b) Maps of inland water height for given values of the ${\mathbf{x}}$ parameters, and deduced $Y$ value of flood surface (in square meter).](./Fig_X.pdf){width="0.85\linewidth"}
### The numerical model
The numerical modelling of the coastal flood relies on the MARS model [@lazure08]. This finite-difference model solves the shallow-water equations and was originally designed to compute regional coastal hydrodynamics, e.g., tide and storm induced water level and currents. The MARS model has here been adapted to account for the specificities of local coastal flooding processes: hydraulic processes around connections like culverts and weirs, coastal defence breaching. This model has been implemented on the study site (white box in Figure \[fig:LOC\]) with a spatial resolution of $25 {\text{ m}}$ and a total number of mesh cells of $>$39,000. The land cover effect on the flood is taken into account by using a spatially varying friction coefficient. The different hydraulic connections (e.g., the hydraulic culverts below the roads, dike, railway,…) are taken into account in the modeling. The forcing conditions (time series deduced from the parameters ${\mathbf{x}}$, Figure \[fig:X\]a) are uniform over the open boundaries of the domain in blue on Figure \[fig:LOC\]. A single model run takes about 30-60 minutes of computation using a single CPU. For more details on the study site, the model set-up and validation, see @Rohmer.etal2018. It should be noted that when wave overtopping is dominant in the flooding processes, other types of models should be used [@leroy15], with computation times 2 orders of magnitude larger. An overflow case allows setting up statistical developments which will useful also for the more expensive models. Figure \[fig:X\]b provides examples of the inundation depth ($H$) computed in each cell for given forcing conditions ${\mathbf{x}}$, as well as the resulting flood surface value ($Y$). We consider the threshold values ${\tau}^{(Y)}_1=1.2\times 10^6 {\text{ m}}^2, \ldots, {\tau}^{(Y)}_4=6.5\times 10^6 {\text{ m}}^2$ introduced in Figure \[fig:X\]b.
### Gaussian process model {#subsec:model}
We consider a rescaled input space $D= [0,1]^5$ and the function $\check{f}: D \subset {\mathbb{R}}^5 \rightarrow {\mathbb{R}}_+$ with $Y = \check{f}({\mathbf{x}})$ for each ${\mathbf{x}}= (x_1, \ldots, x_5) = ({T}, {S}, {t_o}, {t_+},{t_-}) \in D$. In the remainder we keep the notation $({T}, {S}, {t_o}, {t_+},{t_-})$ for the rescaled input. We are interested in estimating $$\varGamma = \{ {\mathbf{x}}\in D : \check{f}({\mathbf{x}}) \geq {\tau}\}, \ \text{ where } \ {\tau}= {\tau}^{(Y)}_k,\ \text{ for } \ k=1, \ldots, 4.
\label{eq:Gamma5d}$$
We consider the square root transformed output data $f({\mathbf{x}}) = \sqrt{\check{f}({\mathbf{x}})} = \sqrt{Y}$ for ${\mathbf{x}}\in D$ and the square root thresholds ${\tau}_k = \sqrt{{\tau}^{(Y)}_k}$. This transformation was chosen, after fitting the model on different scales for $Y$, because it provided the best cross-validation metrics.
We fix an initial DoE ${X}_n \in D^n$, with $n=200$ points obtained by evaluating the function $f$ on the first $500$ points of the $5$-d Sobol’ sequence and by selecting the first $n=200$ points leading to a flood of any magnitude. The evaluations $f({X}_n)$ at ${X}_n$ are denoted with $\mathbf{y}_n \in {\mathbb{R}}^n$. We consider a GP model with a tensor product prior Matérn covariance kernel $\nu=3/2$ and prior mean of the form $$\mu({\mathbf{x}}) = c_0+c_1 x_1 + c_2 x_2 + c_3 x_3^2 + c_4 x_4 + c_5 x_5 = c_0 + \sum_{j=1}^5c_j h_j({\mathbf{x}}).
\label{eq:priorMean}$$ The covariance kernel hyper-parameters are estimated with maximum likelihood from $({X}_n, \mathbf{y}_n)$ and the posterior mean and covariance kernel are obtained with Equations , . The GP and the basis functions $h_j$ were selected using expert-based information achieving a $Q^2 = 0.958$. A comparison of different model fits is shown in Appendix \[sec:full5dRes\].
We estimate $\varGamma$ from the posterior mean with $$\hat{\varGamma}_{n,{\tau}_k} = \{ {\mathbf{x}}\in D : {\mu}_n({\mathbf{x}}) \geq {\tau}_k\}, \ \text{ for }k=1, \ldots, 4.$$
Procedure overview
------------------
![Flow chart for the procedure used in the test case.[]{data-label="fig:flowChart"}](./flowchart.pdf){width="0.85\linewidth"}
In the following sections profile extrema are used to explore visually estimates for $\varGamma$ and to quantify their uncertainty. The proposed procedure is summarized in Figure \[fig:flowChart\]:
step 1: design of experiment and emulation.
: Select a DoE ${X}_n$ ($n=200$) and run the MARS model (numerical simulator) to compute the flooded area $Y_n$; Fit a GP (\[subsec:model\]) on ${X}_n,Y_n$, evaluate the emulation quality with leave-one-out-cross validation.
step 2: univariate profile extrema.
: Compute coordinate profile extrema (i.e. along the canonical directions) on the GP mean and give a first visual indication on the excursion set. Uncertainty quantification on profile extrema (Section \[sec:UQ\]) guides expert knowledge in identifying possible regions of excursion. A comparison of several profile extrema plots with a different number of pilot points $\ell$ increases the understanding of profile functions uncertainty without additional numerical simulator runs. The conclusions drawn from 1d profile extrema can be used to refine the DoE in step 1 and ultimately decrease the uncertainty on the excursion set. This first analysis, for example, shows that some offshore conditions do not influence the excursion. Oblique profile extrema could be used in this phase, see, e.g., section \[subsec:analExample\], if more informative directions are known in advance.
step 3: bivariate profile extrema.
: Explore combinations of input variables that lead to excursion with bivariate profile extrema functions. Orthogonal projections can potentially be used as shown in Section \[subsec:analExample\] on the analytical example. Similarly as for step 3, refinements of the DoE in the regions of interest can be performed.
Results on coastal flooding test case {#subsec:results}
-------------------------------------
### Univariate profile extrema
We start the analysis of $\hat{\varGamma}_{n,{\tau}_k}$, for the thresholds ${\tau}_1, \ldots, {\tau}_4$, with coordinate profile extrema on the posterior GP mean. The uncertainty is visualized with posterior quantiles of the profile extrema (Section \[subsec:approxReals\]) and with the upper and lower bound of equation .
Figure \[fig:ex5dUQfun\_approx\] shows the coordinate profile extrema for the posterior mean of the process based on the design with $n=200$ points described in Section \[subsec:model\], with the universal kriging prior mean defined in Equation and the lowest and highest threshold ${\tau}_1,{\tau}_4$.
Let us consider ${\tau}_4=2549.5 = \sqrt{6.5e6}$, the highest threshold in dark red. Coordinate profile extrema on the posterior mean tell us that if ${T}<0.57$, there is no excursion independently of the other coordinates. The $90\%$ point-wise confidence interval is based on ${s}=600$ approximate posterior realizations generated with $\ell=300$ pilot points. The point-wise confidence intervals identify a possible ($90\%$) region of non-excursion $\{ {\mathbf{x}}\in D : {T}\in [0, \lambda] \text{ with }\lambda \in [0.52,0.6] \}$. If we consider the variable ${S}$, a possible region of non-excursion (above $\tau_4$) is $\{ {\mathbf{x}}\in D : {S}\in [0,\lambda], \ \lambda \in [0.28,0.43] \}$. This region does not exist for low threshold values ($<\tau_2$), thus indicating that small values of surge peak only play a role in moderate flooding events, i.e. $<\tau_2$. Similar assessments are available for the other coordinates and the other thresholds, see Table \[tab:5dSummary\] in Appendix \[sec:full5dRes\] for a summary. Note that the variables ${t_+},{t_-}$ do not bring information on the excursion regions as ${P^{\inf}}_{{\mathbf{e}}_{4,5}}\gamma_n(\eta)$ and ${P^{\sup}}_{{\mathbf{e}}_{4,5}}\gamma_n(\eta)$ are consistently below and above the thresholds ${\tau}_1, {\tau}_4$ respectively. The bounds on the approximating process from equation , plotted as wide light blue tube, show that there is still uncertainty on this assessment. In fact, by accounting also for the approximation uncertainty, the possible region of non-excursion becomes $\{ {\mathbf{x}}\in D: {T}\in [0,0.19 ] \}$. The tightness of the bound is evaluated by looking the integrated variance of the difference, i.e. equation . We chose $\ell=300$ as the resulting integrated variance is small enough and at the same time the method is not computationally too expensive. For example the average integrated variance over all dimensions is $41.6 \%$ and $18.3 \%$ smaller than with $\ell=37$ for ${P^{\inf}}$ and ${P^{\sup}}$ respectively. On the other hand, the computational time for $(\sigma_T^{\widetilde{\Delta}})$ grew from $302$ seconds ($\ell=37$) to $1595$ seconds for $\ell=300$. More details in Appendix \[sec:full5dRes\].
![Coordinate profiles for $5$-dimensional test case. Excursion thresholds (red, light ${\tau}_1$, dark ${\tau}_4$), breakpoints ${P^{\sup}}_i/{P^{\inf}}_i$ on mean (vertical green lines) and on $95\% - 5\%$ quantiles (vertical green lines, dashed), CI dark red tube, bound range light blue tube.[]{data-label="fig:ex5dUQfun_approx"}](./5dBRGM_UQ_quantReg_noM_2T.pdf){width="0.85\linewidth"}
### Bivariate profile extrema
We now focus on the combination of variables ${T}, {S}$ and ${T}, {t_o}$, whose more prominent role on the excursion was outlined by the coordinate profile extrema, and we consider ${\tau}=\tau_1$. We explore which values of these combinations lead to excursion with bivariate profile extrema functions. For example, the profile extrema for ${T}, {S}$ is obtained with $$\Psi = \begin{bmatrix}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0
\end{bmatrix}^T.$$ We compute the empirical quantile maps and the bound with $\ell=300$ and ${s}=420$ approximate posterior realizations, see Section \[subsec:approxReals\]. Figures \[fig:ex5dbivProf\] (a,b,c) show the contour lines of ${P^{\sup}}_\Psi$ and ${P^{\inf}}_\Psi$ for the posterior GP mean based on the DoE described in Section \[subsec:model\]. The background colors indicate different heuristic measures of uncertainty: (a) the weighted inter-quantile range (i.e. the empirical inter-quantile range for profile extrema maps if the threshold is between the upper and lower quantile or zero, otherwise); (b) the upper-lower bound range (i.e. the difference between the upper and lower bound in if ${\tau}$ is between them, zero otherwise); (c) the standard deviation of the difference $(\sigma^{\widetilde{\Delta}}_T) (\eta)$, $\eta \in E_\Psi$.
(a)
![Bivariate (${T},{S}$) profile extrema mean (contour lines). Threshold ${\tau}_1$ dashed red line, ${X}_n$ locations (crosses, green above ${\tau}_1$). Background: (a) weighted inter-quantile range ($95\% - 5\%$), (b) bound range (eq. ), (c) approximation error standard deviation.[]{data-label="fig:ex5dbivProf"}](./5dBRGM_bivUQiqr_TS.pdf){width="0.85\linewidth"}
\
(b)
![Bivariate (${T},{S}$) profile extrema mean (contour lines). Threshold ${\tau}_1$ dashed red line, ${X}_n$ locations (crosses, green above ${\tau}_1$). Background: (a) weighted inter-quantile range ($95\% - 5\%$), (b) bound range (eq. ), (c) approximation error standard deviation.[]{data-label="fig:ex5dbivProf"}](./5dBRGM_bivUQbound_TS.pdf){width="0.85\linewidth"}
\
(c)
![Bivariate (${T},{S}$) profile extrema mean (contour lines). Threshold ${\tau}_1$ dashed red line, ${X}_n$ locations (crosses, green above ${\tau}_1$). Background: (a) weighted inter-quantile range ($95\% - 5\%$), (b) bound range (eq. ), (c) approximation error standard deviation.[]{data-label="fig:ex5dbivProf"}](./5dBRGM_bivUQsigma_TS.pdf){width="0.85\linewidth"}
Let us start by analyzing ${P^{\sup}}_\Psi,{P^{\inf}}_\Psi$ on the posterior mean, shown in the contour lines in Figures \[fig:ex5dbivProf\]. The map on the left suggests that the region below the critical dashed red contour can be excluded, i.e. the region is outside the excursion. In a symmetric way, ${P^{\inf}}_\Psi$ (Figure \[fig:ex5dbivProf\], right) suggests that high values for ${T}$ ($>0.8$) along with values for ${S}$ in the range $[0.05,0.7]$ should lead to an excursion (i.e. flood). The bivariate profile extrema for ${T},{t_o}$ in Figure \[fig:ex5dbivProfTP\], instead, show (left) a non-flood region on the left of the critical contour and suggest (right) excursion in the region ${T}>0.95$, ${t_o}\in[0.42,0.68]$. These conclusions are consistent with our physical knowledge, except for the excursion domain on ${T},{S}$ in Figure \[fig:ex5dbivProf\] (right). Indeed, an excursion domain bounded by maximum surge does not have a physical explanation because $Y$ should increase with ${S}$, with no maximum bound of ${S}$.
The above analysis is based only on ${P^{\sup}}$ and ${P^{\inf}}$ for the posterior GP mean and, since the DoE is small and non-adaptive, there could be high uncertainty in some parts of the input domain. The indicators plotted as background colors allow us to quantify this uncertainty. Both the weighted inter-quantile range and the bound range for ${P^{\inf}}_\Psi$ in Figure \[fig:ex5dbivProf\] (right) are high the region of large ${T}$ and ${S}$. Moreover, as shown in Figure \[fig:ex5dbivProf\](c), the values of $(\sigma^{\widetilde{\Delta}}_T)$ are not high in that region indicating that the uncertainty due to the approximating process is not very high. Those insights collectively suggest that more function evaluations should be added in the ${P^{\inf}}_\Psi$ uncertain region.
### Summary of results
Coordinate profile extrema functions on the coastal flooding test case enabled: (i) to highlight the major role of the high tide level, ${T}$, whatever the considered thresholds, i.e., for small to large flooding events; (ii) to highlight the key role of the surge peak, ${S}$, only for moderate thresholds i.e., moderate flooding events; (iii) to highlight the moderate role of the phase difference ${t_o}$ alone; (iv) to exclude a strong influence of ${t_-}$ and ${t_+}$ alone for the excursion of the response whatever the considered thresholds. Moreover by studying bivariate profile extrema we could: (i) exclude regions where ${T}$ and ${S}$ are simultaneously small (e.g. ${T}<0.4$ and ${S}<0.6$); (ii) highlight the role of phase difference ${t_o}$ and tide ${T}$ combined with a possible excursion in the region ${T}> 0.95$, ${S}\in [0.42,0.68]$. Furthermore, the comparison of approximation error indicators (Figure \[fig:ex5dbivProf\](c)) with uncertainty measures on profile extrema (Figure \[fig:ex5dbivProf\](b)) enabled us to nuance the results and to track the main uncertainty source. In the coastal flooding test case, the uncertainty unlikely stems from the approximating process, but rather from the lack of function evaluations.
Discussion {#sec:discussion}
==========
In this work we introduced a visualization technique, based on profile extrema functions, for excursion regions of expensive-to-evaluate functions. The main idea is to study the constrained optima of the functions on lower dimensional subspaces resulting from projections on lines or planes. By plotting profile extrema functions we can select regions of interest. If the function is expensive to evaluate, a GP model is used to emulate the function and we showed how profile extrema can be computed on the GP model. In this case the conclusions strongly depend on the chosen GP model and, as we show in Appendix \[sec:full5dRes\], the profile extrema uncertainty is affected by the modeling choices.
In the coastal flooding test case, as sketched in Figure \[fig:flowChart\], we selected directions of interest from coordinate profiles. For example, the profile sup along the direction of coordinate ${T}$ is below the threshold of interest for some values. This indicates that ${T}$ is a direction of interest for the excursion phenomena. This procedure was repeated for each coordinate. In test cases where canonical directions are not meaningful we could explore the excursion by looking at oblique directions as we did in the analytical example in Section 3.3.
The bivariate profile extrema maps in Figure \[fig:ex5dbivProf\] also suggest a principled way to develop adaptive design of experiments that sequentially reduce uncertainties. For example, we could select the next evaluations as the minimizers of the integrated (or maximum of) $(\sigma^{\widetilde{\Delta}}_T)$. Such criteria should be analytically tractable and could lead to adaptive designs similar to classic IMSE (MSE) strategies, @Sacks.etal1989. Alternative criteria could be obtained by minimizing bound-range uncertainties (Figure \[fig:ex5dbivProf\](b)), as they provide direct information on the excursion, however their tractability is still unclear.
Profile extrema functions require a continuously differentiable function on a compact domain. If the domain is not compact, then the optimization in the definition of profile extrema might not be well posed. In case of an open domain with a probability distribution on the inputs, profile extrema functions could be extended by using quantiles in place of the maximum [@RoyNotz2014]. Profile extrema functions could be extended to non-linear subsets, however it is technically not straightforward and it might result in visualizations that are much harder to interpret. The overall approach developed here is a one-step procedure, and it could become part of an exploratory work flow. As shown in Appendix \[sec:bivEx3d\], coordinate profile extrema, oblique and bivariate profiles can be combined to convey more information on the excursion set in simpler terms. A possible future extension could involve a treed procedure where the input space is restricted with constrained coordinate profile functions. Oblique coordinate profiles, i.e. profiles along non-canonical directions, require the user to choose which directions to explore. This choice could be driven by expert knowledge, such as in the motivating test case presented here. However, when such knowledge is not available, we could envisage a procedure following similar steps to projection pursuit [@Cook_etal1995] where we obtain the most informative direction.
Proofs {#sec:proofs}
======
Consider the GP regression set-up as described in Section \[subsec:model\]. For ease of notation let us denote the posterior process as $(Z_{\mathbf{x}})_{{\mathbf{x}}\in T}\sim GP({\mu},{\mathfrak{K}})$, where we drop $n$, the number of observations as it is fixed in this section. Recall that the proposed approximate field $(\widetilde{Z}_{\mathbf{x}})_{{\mathbf{x}}\in T}$ is defined as follows $$\widetilde{Z}_{\mathbf{x}}= a({\mathbf{x}}) + \mathbf{b}^T({\mathbf{x}})Z_G \qquad {\mathbf{x}}\in T,$$ where $a: T \rightarrow {\mathbb{R}}$ is a continuous trend function, $\mathbf{b}: T \rightarrow {\mathbb{R}}^\ell$ is a continuous vector-valued function of deterministic weights, $G = ({\mathbf{g}}_1, \ldots, {\mathbf{g}}_\ell ) \in T^\ell$ is a fixed sequence of points in $T$ and $Z_G = (Z_{{\mathbf{g}}_1}, \ldots, Z_{{\mathbf{g}}_\ell})^T$ is a $\ell$-dimensional random vector. Let us consider the difference process $(\widetilde{\Delta}_{\mathbf{x}})_{{\mathbf{x}}\in T}$, defined as $\widetilde{\Delta}_{\mathbf{x}}:= Z_{\mathbf{x}}- \widetilde{Z}_{\mathbf{x}}$, for each ${\mathbf{x}}\in T$. The mean function and covariance kernel of $\widetilde{\Delta}$ are $$\begin{aligned}
{\mu}^{\widetilde{\Delta}}({\mathbf{x}}) &= {\mathbb{E}}[Z_{\mathbf{x}}- \widetilde{Z}_{\mathbf{x}}] = {\mu}({\mathbf{x}}) - a({\mathbf{x}}) -\mathbf{b}({\mathbf{x}})^T {\mu}(G) \qquad {\mathbf{x}}\in T \\
{\mathfrak{K}}^{\widetilde{\Delta}}({\mathbf{x}},{\mathbf{y}}) &= {\mathfrak{K}}({\mathbf{x}},{\mathbf{y}}) - {\mathfrak{K}}({\mathbf{y}},G)\mathbf{b}({\mathbf{x}}) - {\mathfrak{K}}({\mathbf{x}},G)\mathbf{b}({\mathbf{y}}) + \mathbf{b}^T({\mathbf{x}}){\mathfrak{K}}(G,G) \mathbf{b}({\mathbf{y}})
\qquad {\mathbf{x}},{\mathbf{y}}\in T\end{aligned}$$
First of all notice that $$Pr\left(\left\lvert \sup_{{\mathbf{x}}\in T} Z_{\mathbf{x}}- \sup_{{\mathbf{x}}\in T} \widetilde{Z}_{\mathbf{x}}\right\rvert >u \right) \leq Pr\left(\sup_{{\mathbf{x}}\in T} \lvert \widetilde{\Delta}_{\mathbf{x}}\rvert >u \right)
\label{eq:boundIneq1}$$
Let us now consider the centred process $(\widetilde{\Delta}^C_{\mathbf{x}})_{{\mathbf{x}}\in T} := (\widetilde{\Delta}_{\mathbf{x}}- {\mu}^{\widetilde{\Delta}}({\mathbf{x}}))_{{\mathbf{x}}\in T}$. We have that $$Pr\left(\sup_{{\mathbf{x}}\in T} \lvert \widetilde{\Delta}^C_{\mathbf{x}}\rvert > u\right) \leq 2Pr\left(\sup_{{\mathbf{x}}\in T} \widetilde{\Delta}^C_{\mathbf{x}}> u\right) \leq 2 e^{-u^2 /2(\sigma^{\widetilde{\Delta}}_T)^2}, \qquad u>0
\label{eq:boundIneq2}$$ where $(\sigma^{\widetilde{\Delta}}_T)^2 = \sup_{{\mathbf{x}}\in T} {\mathbb{E}}[(\widetilde{\Delta}^C_{\mathbf{x}})^2] = \sup_{{\mathbf{x}}\in T} {\mathfrak{K}}^{\widetilde{\Delta}}({\mathbf{x}},{\mathbf{x}})$. The first inequality follows from the symmetric distribution of the centered field $\widetilde{\Delta}^C_{\mathbf{x}}$ and the second is the Borell-TIS inequality, see, e.g., @Adler.Taylor2007, Chapter 2 for more detail.
Since we have $$\sup_{{\mathbf{x}}\in T} \lvert \widetilde{\Delta}_{\mathbf{x}}\rvert \leq \sup_{{\mathbf{x}}\in T} \lvert \widetilde{\Delta}^C_{\mathbf{x}}\rvert +\sup_{{\mathbf{x}}\in T} \lvert {\mu}^{\widetilde{\Delta}}({\mathbf{x}}) \rvert
\label{eq:boundIneq3}$$ then following Equations , , , if $u> {\mu}^{\widetilde{\Delta}}({\mathbf{x}})$ $$\begin{aligned}
Pr\left(\left\lvert \sup_{{\mathbf{x}}\in T} Z_{\mathbf{x}}- \sup_{{\mathbf{x}}\in T} \widetilde{Z}_{\mathbf{x}}\right\rvert >u \right)
&\leq Pr\left( \sup_{{\mathbf{x}}\in T} \lvert \widetilde{\Delta}^C_{\mathbf{x}}\rvert +\sup_{{\mathbf{x}}\in T} \lvert {\mu}^{\widetilde{\Delta}}({\mathbf{x}}) \rvert >u \right) \\
&\leq Pr\left( \sup_{{\mathbf{x}}\in T} \lvert \widetilde{\Delta}^C_{\mathbf{x}}\rvert > u - \sup_{{\mathbf{x}}\in T} \lvert {\mu}^{\widetilde{\Delta}}({\mathbf{x}}) \rvert \right) \\
&\leq 2 \exp\left(-\frac{\left(u- \sup_{{\mathbf{x}}\in T} \lvert {\mu}^{\widetilde{\Delta}}({\mathbf{x}}) \rvert\right)^2}{2(\sigma^{\widetilde{\Delta}}_T)^2} \right).\end{aligned}$$ If $\widetilde{Z}$ is an unbiased approximation for $Z$, then ${\mu}^{\widetilde{\Delta}}({\mathbf{x}}) =0$ and the inequality is valid for any $u>0$.
Let us denote with $A = {P^{\sup}}_{\Psi} Z(\eta^*)$ and $B = {P^{\sup}}_{\Psi} \widetilde{Z}(\eta^*)$. By using Theorem \[theo:bound\] we have that for $u>0$ $$Pr(\lvert A- B \rvert \geq u) \leq 2 e^{- \frac{(u - {\mu}_T^{\widetilde{\Delta}})^2}{2 (\sigma_T^{\widetilde{\Delta}})^2 }}$$ Moreover $$\begin{aligned}
Pr(A \leq u^+_\alpha ) &= 1- Pr(A \geq u^+_\alpha) = 1 - Pr(A-B+B \geq u^+_\alpha) \\
&= 1- Pr(A-B+B \geq u^+_\alpha \mid B \geq \tilde{u}^+_\beta) Pr (B \geq \tilde{u}^+_\beta) \\
&- Pr(A-B+B \geq u^+_\alpha, B \leq \tilde{u}^+_\beta) \\
&\geq 1-\beta - Pr(A-B+\tilde{u}^+_\beta \geq u^+_\alpha, B \leq \tilde{u}^+_\beta) \\
&\geq 1-\beta - Pr(\lvert A-B \rvert \geq u^+_\alpha -\tilde{u}^+_\beta, B \leq \tilde{u}^+_\beta) \\
&\geq 1- \beta - 2e^{- \tfrac{(u^+_\alpha - \tilde{u}^+_\beta - {\mu}_T^{\widetilde{\Delta}})^2}{2 (\sigma_T^{\widetilde{\Delta}})^2 }}
\end{aligned}$$ By solving for $u^+_\alpha$ in $\beta + 2e^{- \tfrac{(u^+_\alpha - \tilde{u}^+_\beta - {\mu}_T^{\widetilde{\Delta}})^2}{2 (\sigma_T^{\widetilde{\Delta}})^2 }} = \alpha$ under the constraint $u^+_\alpha > \tilde{u}^+_\beta$, we obtain . For the other side notice that $$\begin{aligned}
Pr(A \geq u^-_\alpha ) &\geq Pr(B\geq \tilde{u}^-_\beta, \lvert A - B \rvert \leq \tilde{u}^-_\beta - u^-_\alpha) \quad = 1- Pr(B > \tilde{u}^-_\beta, \lvert A-B \rvert \geq \tilde{u}^-_\beta - u^-_\alpha) \\
&- Pr(B \leq \tilde{u}^-_\beta, \lvert A-B \rvert \geq \tilde{u}^-_\beta - u^-_\alpha) - Pr(B \leq \tilde{u}^-_\beta, \lvert A-B \rvert \leq \tilde{u}^-_\beta - u^-_\alpha) \\
&\geq 1 - Pr(\lvert A-B \rvert \geq \tilde{u}^-_\beta - u^-_\alpha) - \beta - \beta \quad \geq 1- 2\beta - 2e^{- \tfrac{(u^-_\alpha - \tilde{u}^-_\beta)^2}{2 (\sigma_T^{\widetilde{\Delta}})^2 }}
\end{aligned}$$ By solving for $u^-_\alpha$ in $2\beta + 2e^{- \tfrac{(u^-_\alpha - \tilde{u}^-_\beta )^2}{2 (\sigma_T^{\widetilde{\Delta}})^2 }} = \alpha$ under the constraint $u^-_\alpha < \tilde{u}^-_\beta$, we obtain .
Gradient of $\widetilde{Z}_{\mathbf{x}}$ with respect to ${\mathbf{x}}$ {#sec:gradTildeZ}
=======================================================================
We are interested in the approximating process for the posterior distribution of $Z$ conditioned on $({X}_n,\mathbf{y}_n)$. If $a,\mathbf{b}$ are chosen as the posterior mean and the kriging weights respectively, then we can write $\widetilde{Z}$ as $$\begin{aligned}
\widetilde{Z}_{\mathbf{x}}&= \Lambda({\mathbf{x}})^T \begin{bmatrix}
\mathbf{y}_n \\ Z_G
\end{bmatrix} \qquad \text{with} \\
\Lambda({\mathbf{x}}) &= K_{n+\ell}^{-1} \left( {\mathfrak{K}}_{n+\ell}({\mathbf{x}}) + {H}_{n+\ell} K_{H} \left( \mathbf{h}({\mathbf{x}}) - {H}_{n+\ell}^T K_{n+\ell}^ {-1}{\mathfrak{K}}_{n+\ell}({\mathbf{x}}) \right) \right) \\
\text{where } \quad K_{H} &= \left( {H}_{n+\ell}^T K_{n+\ell}^ {-1} {H}_{n+\ell} \right)^{-1}, \quad K_{n+\ell} = {\mathfrak{K}}(A_{n+\ell},A_{n+\ell}), \qquad {\mathfrak{K}}_{n+\ell}({\mathbf{x}}) = {\mathfrak{K}}(x,A_{n +\ell}), \\
{H}_{n+\ell} &= [h_j(A_{n+\ell})]_{j=1, \ldots, m} \in {\mathbb{R}}^{(n+\ell) \times m} \quad
\text{and } \quad A_{n+\ell} = [{X}_n^T, G^T]^T \in {\mathbb{R}}^{(n+\ell)\times d}\end{aligned}$$
Then $\nabla_x \widetilde{Z}_{\mathbf{x}}= \nabla_x \Lambda({\mathbf{x}})^T\begin{bmatrix}
\mathbf{y}_n \\ Z_G .
\end{bmatrix}$ and it suffices to compute the gradient of $\Lambda$ and $$\begin{gathered}
K_{n+\ell} \nabla_x\Lambda({\mathbf{x}}) = \left( \nabla_x{\mathfrak{K}}({\mathbf{x}},A_{n+\ell}) + {H}_{n+\ell} K_{H} \left( \nabla_x\mathbf{h}({\mathbf{x}}) - {H}_{n+\ell}^T K_{n+\ell}^ {-1}\nabla_x{\mathfrak{K}}({\mathbf{x}},A_{n+\ell}) \right) \right) \\
\text{with } \nabla_x{\mathfrak{K}}({\mathbf{x}},A_{n+\ell}) = [\nabla_x{\mathfrak{K}}({\mathbf{x}}, a_1), \ldots, \nabla_x{\mathfrak{K}}({\mathbf{x}}, a_{n+\ell})]^T \in {\mathbb{R}}^{(n+\ell) \times d} \\ \text{ and } \nabla_x\mathbf{h}({\mathbf{x}}) = [\nabla_x h_1({\mathbf{x}}), \ldots, h_m({\mathbf{x}})] \in {\mathbb{R}}^{m\times d}.\end{gathered}$$
Full results on flooding test case {#sec:full5dRes}
==================================
In this section, we report more details on the GP model used in the flooding test case in Section \[sec:motivating\] and we show the profile extrema functions for all thresholds ${\tau}_1,\ldots, {\tau}_4$.
In table \[tab:5dmodels\] we compare different GP models according to two metrics: $Q^2$ on leave-one-out predictions and log likelihood ($logLik$). According to those metrics the prior mean function in equation \[eq:priorMean\] (universal kriging, UK) results in better fits than a constant prior mean (ordinary kriging, OK). On the other hand the difference between using a smoothness parameter $\nu=3/2$ or $\nu=5/2$ is very small. Here we chose the parameter $\nu=3/2$ as it leads to standardized model residuals that have a distribution closer to the normal one.
\[-1.8ex\] Matern 3/2, OK Matern 3/2, UK Matern 5/2, OK Matern 5/2, UK
------------ ---------------- ------------------------ ---------------- -----------------
$Q^2$ $0.95$ $\mathbf{0.96}$ $0.94$ $\mathbf{0.96}$
$logLik$ $$-$1,406.02$ $$-$\textbf{1,363.73}$ $$-$1,410.71$ $$-$1,365.32$
: Comparison of different GP models for flooding test case. OK: constant mean; UK: mean function as in equation \[eq:priorMean\]. Best values in bold.[]{data-label="tab:5dmodels"}
If the model fit is worse, then we obtain coordinate profiles with larger confidence bands and thus more uncertainty. We checked this assumption by computing coordinate profile extrema on each of the models in Table \[tab:5dmodels\]. Figures \[fig:ex5dcoord\_mat32OK\] and \[fig:ex5dcoord\_mat52OK\] show the first two coordinate profile plots for Matérn $3/2$, OK and Matérn $5/2$, OK. A quick glance already shows that the confidence bands are larger than in Figure \[fig:ex5dUQfun\_approx\]. We further compared the integrated inter-quantile range for each model and we ranked them in increasing order. The average ranks over all coordinates for profile sup and profile inf are shown in Table \[tab:5diqrModels\]. Note how the chosen model (Matérn $3/2$ model with UK prior mean) has an average rank $1.8$ for ${P^{\sup}}f$ and $1.4$ for ${P^{\inf}}f$.
\[-1.8ex\] Matern 3/2, OK Matern 3/2, UK Matern 5/2, OK Matern 5/2, UK
--------------- ---------------- ---------------- ---------------- ----------------
${P^{\sup}}f$ $2.4$ $\mathbf{1.8}$ $3$ $2.8$
${P^{\inf}}f$ $3.8$ $\textbf{1.4}$ $2.8$ $2$
: Average rank of integrated inter-quantile range for coordinate profile extrema.[]{data-label="tab:5diqrModels"}
![Coordinate profiles on GP model with Matérn $\nu=5/2$ and constant mean.[]{data-label="fig:ex5dcoord_mat52OK"}](./5dBRGM_mat32_ok_cp.pdf){width="\linewidth"}
![Coordinate profiles on GP model with Matérn $\nu=5/2$ and constant mean.[]{data-label="fig:ex5dcoord_mat52OK"}](./5dBRGM_mat52_ok_cp.pdf){width="\linewidth"}
Let us now focus on the chosen model: Matérn $\nu=3/2$ with prior mean function as described in equation \[eq:priorMean\]. Table \[tab:5dSummary\] summarizes the intervals selected with the coordinate profile extrema functions. Figure \[fig:ex5dUQfun\_full\] shows the coordinate profile extrema functions, the $5\%, 95\%$ quantiles, the boundary of the subsets selected (vertical lines) by posterior median and quantiles. The values of the bound with confidence level $\alpha=0.1$ are shown with the sky blue (${P^{\sup}}_i Z$) and sea green (${P^{\inf}}_i Z$) shaded regions. The bound on the profile inf function is almost overlapping with the $90\%$ confidence intervals in red. The bound on the profile sup function instead provides an higher quantile and identifies as region of possible non excursion ${T}\in [0,0.19]$ above ${\tau}_4$. For the other coordinates the bound does not provide information on regions of non excursion.
\[-1.8ex\] ${T}\in [0,\lambda_{{T},k}]$ ${S}\in [0,\lambda_{{S},k}] $ ${t_o}\in [0,\lambda^a_{{t_o},k}] \cup [\lambda^b_{{t_o},k},1]$ ${t_+}$ ${t_-}$
------------ ----------------------------------- ----------------------------------- ----------------------------------------------------------------- --------- ---------
${\tau}_1$ $\lambda_{{T},1} \in [0.17,0.26]$ $-$ $\lambda^a_{{t_o},1}=0, \lambda^b_{{t_o},1}=1$ $-$ $-$
${\tau}_2$ $\lambda_{{T},2} \in [0.24,0.32]$ $\lambda_{{S},2} \in [0,0.06]$ $\lambda^a_{{t_o},2} \in [0,0.11], \lambda^b_{{t_o},2}=1$ $-$ $-$
${\tau}_3$ $\lambda_{{T},3} \in [0.33,0.40]$ $\lambda_{{S},2} \in [0,0.14]$ $\lambda^a_{{t_o},3} \in [0,0.12], \lambda^b_{{t_o},3}=1$ $-$ $-$
${\tau}_4$ $\lambda_{{T},4} \in [0.52,0.60]$ $\lambda_{{S},2} \in [0.28,0.43]$ $\lambda^a_{{t_o},4} \in [0,0.16],$ $\lambda^b_{{t_o},4} =1$ $-$ $-$
: \[tab:5dSummary\] Regions excluded with profile extrema functions on $5$-d test case. Interval defined in the top line, for each threshold the interval for the boundary computed from the $90\%$ approximate confidence intervals for the profiles is reported.
![Coordinate profiles for $5$-dimensional test case. Thresholds (red, light to dark for ${\tau}_1, \ldots, {\tau}_4$), regions selected by profile extrema on mean (vertical green lines) and regions selected by profile extrema on $95\% - 5\%$ quantiles (vertical green lines, dashed). Bound on the confidence region ($\alpha=0.05$) for ${P^{\sup}}_i Z$ (sky blue, shaded) and for ${P^{\inf}}_i Z$ (sea green, shaded).[]{data-label="fig:ex5dUQfun_full"}](./5dBRGM_UQ_quantReg_noM.pdf){width="0.8\linewidth"}
Section \[subsec:bound\] introduced the integrated variance of the difference, equation , as an indicator for the bound tightness. Figure \[fig:integVariance5d\] shows the value $I(\sigma_T^{\widetilde{\Delta}})^2$ averaged over all dimensions for ${P^{\sup}}$ and ${P^{\inf}}$ versus the number of pilot points $\ell$. Note the “diminishing returns” as $\ell$ increases. In particular, here we chose $\ell=300$ for computational reasons.
![Average integrated variance of the difference, equation , for profile inf (dashed, triangles) and profile sup (solid, circles) versus $\ell$ $(37,75,150,300)$. Computational time for $(\sigma_T^{\tilde{\Delta}})^2$ in legend. Profile coordinates, 5-d test case.[]{data-label="fig:integVariance5d"}](./5dBRGM_comp_nSimPts.pdf){width="0.5\linewidth"}
![Cut with ${T}=0,0.5,1$ of ${P^{\sup}}_{\Psi_1},{P^{\inf}}_{\Psi_1}$ on the profile mean, see also Figure \[fig:ex5dbivProf\].[]{data-label="fig:ex5dbivPE_TScuts"}](./cut_TS_T0.pdf){width="1.1\linewidth"}
![Cut with ${T}=0,0.5,1$ of ${P^{\sup}}_{\Psi_1},{P^{\inf}}_{\Psi_1}$ on the profile mean, see also Figure \[fig:ex5dbivProf\].[]{data-label="fig:ex5dbivPE_TScuts"}](./cut_TS_T0_5.pdf){width="1.1\linewidth"}
![Cut with ${T}=0,0.5,1$ of ${P^{\sup}}_{\Psi_1},{P^{\inf}}_{\Psi_1}$ on the profile mean, see also Figure \[fig:ex5dbivProf\].[]{data-label="fig:ex5dbivPE_TScuts"}](./cut_TS_T1.pdf){width="1.1\linewidth"}
![Bivariate profile extrema for $5$-dimensional test case with ${T},{t_o}$. Excursion threshold ${\tau}_1$ in red, bivariate profile mean values in contour lines, ${X}_n$ locations as crosses (green if above ${\tau}_1$), background color denotes the upper-lower bound range (equation ).[]{data-label="fig:ex5dbivProfTP"}](./5dBRGM_bivUQbound_TP.pdf){width="0.85\linewidth"}
![Cut with ${T}=0,0.5,1$ of ${P^{\sup}}_{\Psi_2},{P^{\inf}}_{\Psi_2}$ on the profile mean, see also Figure \[fig:ex5dbivProfTP\].[]{data-label="fig:ex5dbivPE_TPcuts"}](./cut_TP_T0.pdf){width="1.1\linewidth"}
![Cut with ${T}=0,0.5,1$ of ${P^{\sup}}_{\Psi_2},{P^{\inf}}_{\Psi_2}$ on the profile mean, see also Figure \[fig:ex5dbivProfTP\].[]{data-label="fig:ex5dbivPE_TPcuts"}](./cut_TP_T0_5.pdf){width="1.1\linewidth"}
![Cut with ${T}=0,0.5,1$ of ${P^{\sup}}_{\Psi_2},{P^{\inf}}_{\Psi_2}$ on the profile mean, see also Figure \[fig:ex5dbivProfTP\].[]{data-label="fig:ex5dbivPE_TPcuts"}](./cut_TP_T1.pdf){width="1.1\linewidth"}
Figure \[fig:ex5dbivPE\_TScuts\] shows the cuts at ${T}=0,0.5,1$ of the bivariate profile functions on the mean shown in Figure \[fig:ex5dbivProf\](b) with their respective upper and lower bounds. Such plots can be used to zoom in particular regions of the bivariate profiles. They do not plot more information than bivariate profile maps, however they are simpler to read. For example, Figure \[fig:ex5dbivPE\_TScuts\] shows that $\{{\mathbf{x}}\in D: {T}=0 \text{ and } {S}< 0.79 \}$ is a non-excursion region with high probability.
Figure \[fig:ex5dbivProfTP\] shows the bivariate profile extrema for the variables ${T}, {t_o}$. In particular the left-most map allows us to exclude all input regions on the left of the red dotted curve. In a symmetric way, the right-most plot tells us that values for ${T},{t_o}$ in the region on the right-hand side of the red dotted line lead to an excursion. Also in this case the uncertainty quantification adds more insights on the result. The profile inf shows much uncertainty around the threshold, therefore while region of interest could be selected the uncertainty must be accounted for. The profile sup also tells us that close to the threshold the uncertainty is still high, however the top and bottom left triangles are not in the excursion with very little uncertainty. A very conservative assessment should select only this input region as non-excursion. Figure \[fig:ex5dbivPE\_TPcuts\] shows the cuts at ${T}=0,0.5,1$ of these bivariate profile functions. As in the other case such plots confirm information shown in bivariate maps. For example, we can clearly see that for ${T}=0$ and ${t_o}\in [0,0.37] \cup [0.63,1]$ there is no excursion with high probability.
An indicator of tightness for the bound {#sec:deltaTex2d}
=======================================
\[-1.8ex\] Symbol Meaning Section
------------------- ---------------------------------------------------------------------------------- ------------------------
$n$ Number of expensive computer experiment simulations \[sec:intro\]
$d$ Dimension of input space \[sec:intro\]
$p$ Dimension of profile extrema projection, usually $1$ or $2$ \[sec:profiles\]
$k$ Number of points used for deterministic approximation of ${P^{\sup}},{P^{\inf}}$ \[subsec:approxProf\]
$\ell$ Number of pilot points for GP approximation in \[subsec:approxReals\]
$s$ Number of posterior approximate GP realizations \[subsec:approxReals\]
: \[tab:paramSummary\] Summary of profile extrema parameters.
![Values of $I(\sigma^{\widetilde{\Delta}}_T)^2(\ell)$, Eq. , for $\Psi=\mathbf{v}_1$ (solid, circles) and $\Psi=\mathbf{v}_2$ (dashed, triangles) as functions of $\ell$ . Profile sup (left) and Profile inf (right).[]{data-label="fig:ex2dDeltaTcomparison"}](./comp_nSimPts_max.pdf){width="\linewidth"}
![Values of $I(\sigma^{\widetilde{\Delta}}_T)^2(\ell)$, Eq. , for $\Psi=\mathbf{v}_1$ (solid, circles) and $\Psi=\mathbf{v}_2$ (dashed, triangles) as functions of $\ell$ . Profile sup (left) and Profile inf (right).[]{data-label="fig:ex2dDeltaTcomparison"}](./comp_nSimPts_min.pdf){width="\linewidth"}
The bounds for the uncertainty on ${P^{\sup}}_\Psi Z$ are computed with equation . For a given $\ell$, the tightness of this approximation is controlled by $(\sigma^{\widetilde{\Delta}}_T)^2(\eta,\ell)$, i.e. the sup of the difference variance. This is a non-negative function of $\eta$, as the difference depends on where we evaluate ${P^{\sup}}_\Psi Z (\eta)$. The integrated variance $I(\sigma^{\widetilde{\Delta}}_T)^2(\ell)$, Equation , is a summary of this quantity. Since $(\sigma^{\widetilde{\Delta}}_T)^2(\eta,\ell)$ is non-negative for each $\eta \in E_\Psi$, a smaller integral implies less variability and a tighter bound. We can control $I(\sigma^{\widetilde{\Delta}}_T)^2(\ell)$ with the number of pilot points $\ell$ chosen for the approximate realizations. Figure \[fig:ex2dDeltaTcomparison\] shows the estimated integral $I(\sigma^{\widetilde{\Delta}}_T)^2(\ell)=\int_{E_{\mathbf{v}_i}}(\sigma^{\widetilde{\Delta}}_T)^2(\eta,\ell)d\eta$, $i=1,2$ for the synthetic example introduced in Section \[subsec:analExUQ\] with $n=90$ as a function of $\ell$. More pilot points lead to a smaller integral for each coordinate and for both profile extrema, however the rate of decrease is test case dependent.
Uni and bivariate profiles, a combined approach {#sec:bivEx3d}
===============================================
In this section we consider the following function $$\gamma(\mathbf{x}) = \sin\left( a \mathbf{v_1}^T \mathbf{x} +b\right) + \cos\left(c \mathbf{v_2}^T \mathbf{x} +d \right) +\sin \left( e \mathbf{v_3}^T \mathbf{x} + f\right) -1.5 \qquad \mathbf{x} \in [0,1]^3, \ \mathbf{v_1}, \mathbf{v_2}, \mathbf{v_3} \in \mathbb{R}^3
\label{eq:analExample3d}$$ where $a,b,c,d,e,f \in {\mathbb{R}}$ and $$\mathbf{v_1} = \begin{bmatrix}
\sin(\theta)\cos(\phi) \\
\sin(\theta)\sin(\phi) \\
\cos(\theta)
\end{bmatrix}, \ \mathbf{v_2} = \begin{bmatrix}
\cos(\theta)\cos(\phi) \\
\cos(\theta)\sin(\phi) \\
-\sin(\theta)
\end{bmatrix}, \ \mathbf{v_3} =\begin{bmatrix}
-\sin(\phi) \\
\cos(\phi) \\
0
\end{bmatrix}$$ with $\theta= \pi/4$, $\phi=\pi/4$. We fix $[a,b,c,d,e,f]=[1,0,10,0,1,0]$ and we study the excursion set above ${\tau}=0$. Figure \[fig:ex3dFunction\] shows three cuts corresponding to the planes $z=0,0.49,1$.
![Cuts of the function in Equation .[]{data-label="fig:ex3dFunction"}](./3dEx_Fun_cut_z_00.pdf){width="1.1\linewidth"}
![Cuts of the function in Equation .[]{data-label="fig:ex3dFunction"}](./3dEx_Fun_cut_z_49.pdf){width="1.1\linewidth"}
![Cuts of the function in Equation .[]{data-label="fig:ex3dFunction"}](./3dEx_Fun_cut_z_100.pdf){width="1.1\linewidth"}
Since the excursion set sits along directions which are oblique with respect to the canonical coordinates, the coordinate profile extrema do not provide information on this set. Figure \[fig:ex3dCoord\] shows that ${P^{\sup}}_i f, i=1,2,3$ are always above the threshold and ${P^{\inf}}_i f, i=1,2,3$ are always below the threshold, thus we cannot exclude any region. In such situations we propose two approaches: compute oblique profiles according to more informative directions or compute bivariate coordinate profiles. Figure \[fig:ex3dOblique\] show the oblique profile extrema along the generating directions $\mathbf{v_1}, \mathbf{v_2}, \mathbf{v_3}$. This plot excludes part of the input space by considering the planes delimited by the interceptions between the profile sup and the threshold, e.g. all points between the planes $\mathbf{v_1}{\mathbf{x}}=0$ and $\mathbf{v_1}{\mathbf{x}}=0.31$.
![Profile extrema for the function in Equation along the directions $\mathbf{v_1},\mathbf{v_2},\mathbf{v_3}$.[]{data-label="fig:ex3dOblique"}](./3dEx_coordinateProfiles.pdf){width="\linewidth"}
![Profile extrema for the function in Equation along the directions $\mathbf{v_1},\mathbf{v_2},\mathbf{v_3}$.[]{data-label="fig:ex3dOblique"}](./3dEx_obliqueProfiles.pdf){width="\linewidth"}
The second option involves computing the profile extrema with projections on $2$ dimensional subspaces. Here we consider all combinations of bivariate projections on canonical axes. Figures \[fig:ex3dBivCoord\_sup\], \[fig:ex3dBivCoord\_inf\] show the maps obtained for the ${P^{\sup}}_\Psi f, {P^{\inf}}_\Psi f$ for the three matrices build by combining the canonical directions.
![Bivariate profile inf for the function in Equation along projections on combinations of three canonical axes.[]{data-label="fig:ex3dBivCoord_inf"}](./3dEx_bivCanonical_sup.pdf){width="\linewidth"}
![Bivariate profile inf for the function in Equation along projections on combinations of three canonical axes.[]{data-label="fig:ex3dBivCoord_inf"}](./3dEx_bivCanonical_inf.pdf){width="\linewidth"}
Figure \[fig:ex3dBivCoord\_sup\] in particular highlights darker shaded (red) regions that are not part of the excursion set. Consider ${P^{\sup}}_{\Psi_1}\gamma$, for all $x_1,x_2$ in the shaded region the profile tells us that there is no excursion. For example if $x_1 \in [0.115,1]$ and $x_2 = 0.2$, the segment highlighted in the top left plot of Figure \[fig:ex3dBivCoord\_sup\], there is no excursion. In this case, ${P^{\inf}}_\Psi f$, Figure \[fig:ex3dBivCoord\_inf\], does not allow us to select any excursion region.
[^1]: $Q^2 = 1 - \sum_{j=1}^{N_{test}} ({\mu}_{n}(x_j) - y^{test}_j)^2 / (\sum_{j=1}^{N_{test}} (y^{test}_j - \overline{y^{test}} )^2)$ with $\overline{y^{test} }= 1/{N_{test}}\sum_{j=1}^{N_{test}} y^{test}_j$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A cryogenic achromatic half-wave plate (HWP) for submillimetre astronomical polarimetry has been designed, manufactured, tested, and deployed in the Balloon-borne Large-Aperture Submillimeter Telescope for Polarimetry (BLASTPol). The design is based on the five-slab Pancharatnam recipe and it works in the wavelength range 200–600$\mu$m, making it the broadest-band HWP built to date at (sub)millimetre wavelengths. The frequency behaviour of the HWP has been fully characterised at room and cryogenic temperatures with incoherent radiation from a polarising Fourier transform spectrometer. We develop a novel empirical model, complementary to the physical and analytical ones available in the literature, that allows us to recover the HWP Mueller matrix and phase shift as a function of frequency and extrapolated to 4K. We show that most of the HWP non-idealities can be modelled by quantifying one wavelength-dependent parameter, the position of the HWP equivalent axes, which is then readily implemented in a map-making algorithm. We derive this parameter for a range of spectral signatures of input astronomical sources relevant to BLASTPol, and provide a benchmark example of how our method can yield improved accuracy on measurements of the polarisation angle on the sky at submillimetre wavelengths.'
author:
- |
Lorenzo Moncelsi$^{1,2}$[^1], Peter A. R. Ade$^{2}$, Francesco E. Angilè$^{3}$, Steven J. Benton$^{4}$, Mark J. Devlin$^{3}$, Laura M. Fissel$^{5}$, Natalie N. Gandilo$^{5}$, Joshua O. Gundersen$^{6}$, Tristan G. Matthews$^{7}$, C. Barth Netterfield$^{4,5}$, Giles Novak$^{7}$, David Nutter$^{2}$, Enzo Pascale$^{2}$, Frédérick Poidevin$^{8}$, Giorgio Savini$^{8}$, Douglas Scott$^{9}$, Juan Diego Soler$^{5}$, Locke D. Spencer$^{2,10}$, Matthew D. P. Truch$^{3}$, Gregory S. Tucker$^{11}$, Jin Zhang$^{2}$\
$^{1}$ California Institute of Technology, 1200 E. California Blvd., Pasadena, CA 91125, USA\
$^{2}$ Department of Physics and Astronomy, Cardiff University, 5 The Parade, Cardiff, CF24 3AA, UK\
$^{3}$ Department of Physics and Astronomy, University of Pennsylvania, 209 South 33rd Street, Philadelphia, PA 19104, USA\
$^{4}$ Department of Physics, University of Toronto, 60 St. George Street, Toronto, ON M5S 1A7, Canada\
$^{5}$ Department of Astronomy and Astrophysics, University of Toronto, 50 St. George Street, Toronto, ON M5S 3H4, Canada\
$^{6}$ Department of Physics, University of Miami, 1320 Campo Sano Drive, Coral Gables, FL 33146, USA\
$^{7}$ Department of Physics and Astronomy, Northwestern University, Evanston, IL 60208, USA\
$^{8}$ Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, UK\
$^{9}$ Department of Physics and Astronomy, University of British Columbia, 6224 Agricultural Road, Vancouver, BC V6T 1Z1, Canada\
$^{10}$Department of Physics and Astronomy, University of Lethbridge, 4401 University Drive, Lethbridge, Alberta, Canada, T1K 3M4\
$^{11}$Department of Physics, Brown University, 182 Hope Street, Providence, RI 02912, USA
bibliography:
- 'refs.bib'
date: 'Accepted . Received ; in original form '
title: 'Empirical modelling of the BLASTPol achromatic half-wave plate for precision submillimetre polarimetry'
---
\[firstpage\]
instrumentation: polarimeters — techniques: polarimetric — balloons — magnetic fields — polarisation — submillimetre.
Introduction
============
Galactic magnetic fields are believed to play a crucial role in the evolution of star-forming molecular clouds, perhaps controlling the rate at which stars are born and even determining their mass [@Crutcher2004b; @McKee2007]. However, magnetic fields are very difficult to probe on the spatial scales relevant to the star-forming processes, especially within obscured molecular clouds [e.g., @Crutcher2004a; @Whittet2008], hence their influence on star formation has not yet been clearly established observationally.
Zeeman splitting of molecular lines, which allows a direct measurement of the strength of the line-of-sight component of the local magnetic field, has been carried out successfully for a number of molecular cloud cores [e.g., @Crutcher1999a], though the technique is difficult and often limited to very bright regions [@Crutcher2012]. A promising alternative method is to observe clouds with a far-infrared/submillimetre (FIR/submm) polarimeter [e.g., @Hildebrand1984; @Hildebrand2000; @WardThompson2000]. By tracing the linearly polarised thermal emission from aspherical dust grains aligned with respect to the local magnetic fields, we can estimate the direction of the plane-of-the-sky component of the field within the cloud [@Davis1951; @Dolginov1976; @Lazarian2007], and its strength via the Chandrasekhar & Fermi [CF; -@Chandrasekhar1953] technique, provided that ancillary measurements of the turbulent motion velocity are available. The observed morphology in submm polarisation maps can also be used, in synergy with magnetohydrodynamic simulations, to study the imprint of turbulence and magnetization on the formation of structure in the cloud [@Houde2009; @Soler2013].
Ground-based observations with the SCUBA polarimeter [@Murray1997; @Greaves2003] and the Submillimeter Polarimeter for Antarctic Remote Observations [SPARO; @Novak2003] show that the submm emission from prestellar cores and giant molecular clouds (GMCs) is indeed polarised to a few percent [@WardThompson2000; @Li2006]. [*Planck*]{} [@Planck2011] will provide coarse resolution ($\sim$5) submm polarimetry maps of the entire Galaxy. The Atacama Large Millimeter/submillimeter Array [ALMA; @Wootten2009] will provide sub-arcsecond millimetre (mm) and submm polarimetry, capable of resolving fields within cores and circumstellar disks, but will not be sensitive to cloud-scale fields.
The Balloon-borne Large-Aperture Submillimeter Telescope for Polarimetry [BLASTPol; @Marsden2008; @Fissel2010; @Pascale2012], with its arcminute resolution, is the first submm polarimeter to map the large-scale magnetic fields within molecular clouds with unique combined sensitivity and mapping speed, and sufficient angular resolution to observe into the dense cores. BLASTPol will be able to trace magnetic structures in the cold interstellar medium from scales of 0.05pc out to 5pc, thus providing a much needed bridge between the large-area but coarse-resolution polarimetry provided by [*Planck*]{} and the high-resolution but limited field-of-view maps of ALMA.
BLASTPol successfully completed two science flights over Antarctica during the austral summers of 2010 and 2012, mapping the polarised dust emission at 250, 350, and 500$\mu$m over a wide range of column densities corresponding to $A_V$$\gtrsim$4mag, yielding hundreds to thousands of independent polarisation pseudo-vectors per cloud, for a dozen between GMCs and dark clouds. The first scientific results from the 2010 campaign are soon to be released [@Matthews2013 and Poidevin et al. in preparation], while the 2012 data are still under analysis (an overview of the 2012 observations can be found in Angilè et al. in preparation).
The BLASTPol linear polarisation modulation scheme comprises a stepped cryogenic achromatic half-wave plate (HWP) and photolithographed polarising grids placed in front of the detector arrays, acting as analysers. The grids are patterned to alternate the polarisation angle sampled by 90$^{\circ}$ from bolometer-to-bolometer along the scan direction. BLASTPol scans so that a source on the sky passes along a row of detectors, and thus the time required to measure one Stokes parameter (either $Q$ or $U$) is just equal to the separation between bolometers divided by the scan speed. During operations, we carry out spatial scans at four HWP rotation angles spanning 90$^{\circ}$ (22.5$^{\circ}$ steps), allowing us to measure the other Stokes parameter through polarisation rotation.
The use of a continuously rotating or stepped HWP as a polarisation modulator is a widespread technique at (sub)mm wavelengths [e.g., @Renbarger2004; @Hanany2005; @Pisano2006; @Savini2006; @Savini2009; @Johnson2007; @Li2008a; @Matsumura2009]. A thorough account of the HWP non-idealities and its inherent polarisation systematics, especially for very achromatic designs, has become necessary as the accuracy and sensitivity of (sub)mm instruments have soared in recent years.
The literature offers numerous efforts to address, through simulations, the impact of the inevitable instrumental systematic errors due to the polarisation modulation strategy in the unbiased recovery of the Stokes parameters $Q,U$ on the sky, especially for cosmic microwave background (CMB) polarisation experiments [e.g., @Odea2007; @Odea2011; @Brown2009]. In addition, physical and analytical models have been developed to retrieve the frequency-dependent modulation function of achromatic HWPs and estimate the corrections due to non-flat source spectral indices [@Savini2006; @Savini2009; @Matsumura2009].
Nevertheless, little work has been published on incorporating the [*measured*]{} HWP non-idealities in a data-analysis pipeline and ultimately in a map-making algorithm. @Bryan2010 derive an analytic model that parametrises the frequency-dependent non-idealities of a monochromatic HWP and present a map-making algorithm that accounts for these. @Bao2012 carefully simulate the impact of the spectral dependence of the polarisation modulation induced by an achromatic HWP on measurements of the CMB polarisation in the presence of astrophysical foregrounds, such as Galactic dust. However, both these works assume the nominal design values for the build parameters of the HWP plus anti-reflection coating (ARC) assembly.
While this assumption is a reasonable one when no spectral measurements of the HWP as-built are available, several studies clearly show that the complex multi-slab crystal HWP and its typically multi-layer ARC are practically impossible to manufacture [*exactly*]{} to the desired specifications. In particular, @Savini2006 [@Savini2009] and @Pisano2006 caution against the finite precision to which the multiple crystal substrates composing an achromatic HWP can be aligned relative to each other in the Pancharatnam [-@Pancharatnam1955] scheme. In addition, @Zhang2009 show how some of the design parameters in the ARC can slightly change during the bonding of layers, achieved via a hot-pressing technique [@Ade2006]. We will briefly cover these points and discuss the repercussions on the HWP performance.
This work describes a novel empirical method that allows the reconstruction of the Mueller matrix[^2] of a generic HWP as a function of frequency through spectral transmission measurements of the HWP rotated by different angles with respect to the input polarised light. Not only does this method give complete and quantitative information on the [*measured*]{} spectral performance of the HWP, but it also provides a direct avenue to accounting for the non-idealities of the HWP [*as-built*]{} in a map-making algorithm. This empirical approach is applied to the BLASTPol HWP and will help improve the accuracy on astronomical measurements of polarisation angles at submm wavelengths.
The layout of this paper is as follows. In Section \[sec:intro\_HWP\], we give an overview of the manufacturing process for BLASTPol’s five-slab sapphire HWP. Section \[sec:HWP\_spectra\] describes the spectral measurements, while Section \[sec:empirical\_model\] presents the empirical model as well as the main results of the paper. Finally, in Section \[sec:map\_maker\], we describe the algorithm for the naive-binning map-making technique implemented by BLASTPol, which naturally accounts for the [*measured*]{} HWP non-idealities. Section \[sec:concl\_HWP\] contains our conclusions.
The BLASTP Half-Wave Plate {#sec:intro_HWP}
==========================
Wave plates (or retarders) are optical elements used to change the polarisation state of an incident wave, by inducing a predetermined phase difference between two perpendicular polarisation components. A (monochromatic) wave plate can be simply obtained with a single slab of uniaxial birefringent crystal of specific thickness, which depends upon the wavelength and the index of refraction of the crystal. A birefringent crystal is characterised by four parameters, $n_{\rm e}$, $n_{\rm o}$, $\alpha_{\rm e}$, $\alpha_{\rm o}$, the real part of the indices of refraction and the absorption coefficient (in cm$^{-1}$) for the extraordinary and ordinary axes of the crystal. At a specific wavelength $\lambda_0$, the phase shift induced by a slab is determined uniquely by its thickness $d$, and reads: $$\label{eq:phase_HWP}
\Delta\varphi\left(\lambda_0\right) = \frac{2\,\pi\,d}{\lambda_0}
\left(n_{\rm e} - n_{\rm o}\right)~.$$Given the operating wavelength $\lambda_0$, the required phase shift for the wave plate is achieved by tuning the thickness $d$.
While monochromatic wave plates have been (and are still being) used in (sub)mm astronomical polarimeters [see e.g., @Renbarger2004; @Li2008a; @BryanSPIE; @Bryan2010; @Dowell2010], the inherent dependence of the phase shift on wavelength, expressed in Equation \[eq:phase\_HWP\], constitutes an intrinsic limit in designing a polarisation modulator that operates in a broad spectral range (i.e., is achromatic).
Achromatic half-wave plate design {#sec:HWP_design}
---------------------------------
Achromaticity is necessary for wave plates that are designed for use with multi-band bolometric receivers, such as BLASTPol, PILOT [@Bernard2007], or SCUBA-2 [@Bastien2005; @Savini2009]. To achieve a broadband performance, multiple-slab solutions have been conceived in the past [@Pancharatnam1955; @Title1981] to compensate and keep the phase shift approximately constant across the bandwidth, by stacking an odd number (usually 3 or 5) of birefringent substrates of the same material, which are rotated with respect to each other about their optical[^3] axes by a frequency-dependent set of angles.
Achromatic wave plates have been designed and built for astronomical polarimeters at (sub)mm wavelengths by many authors in the last decade [@Hanany2005; @Pisano2006; @Savini2006; @Savini2009; @Matsumura2009], following the Poincaré sphere (PS) method first introduced by Pancharatnam [-@Pancharatnam1955]. Because the four parameters that characterise a crystal, $n_{\rm e}$, $n_{\rm o}$, $\alpha_{\rm
e}$, $\alpha_{\rm o}$, all depend upon wavelength (in particular, the different frequency-dependence of the ordinary and extraordinary refraction indices enters Equation \[eq:phase\_HWP\] in a non-trivial way, as we will illustrate in detail for sapphire), the design of an achromatic HWP becomes progressively more difficult as the bandwidth increases.
Using the PS method, we have designed a HWP for the BLASTPol instrument, which requires an extended frequency range to cover three adjacent 30% wide spectral bands at 250, 350, and 500$\mu$m. A Pancharatnam [-@Pancharatnam1955] five-slab design is chosen with axis orientations of $\Phi_0=0^{\circ}$, $\Phi_1=26^{\circ}$, $\Phi_2=90.3^{\circ}$, $\Phi_1=26^{\circ}$, and $\Phi_0=0^{\circ}$; these angles are optimised using the physical and analytical model developed by @Savini2006 for an achromatic HWP, which in turn is based on the work of @Kennaugh1960. In Fig. \[fig:HWP\_exploded\_plates\] we show an exploded view of the HWP assembly; to our knowledge, with its $\sim$100% bandwidth, this is the broadest-band HWP manufactured to date at (sub)mm wavelengths for which measurements are published[^4].
![Exploded view of the BLASTPol HWP. We also show the two-layer anti-reflection coating described in Section \[sec:HWP\_manu\]. [Figure modified from @Savini2006].[]{data-label="fig:HWP_exploded_plates"}](HWP_exploded_plates_Phi.pdf){width="1.\linewidth"}
HWP manufacture {#sec:HWP_manu}
---------------
In addition to the broad spectral range of operation, the BLASTPol HWP is required to function at cryogenic temperatures [4K; see @Fissel2010] for two main reasons: (1) reduce the thermal emission from a warm element placed in the optical path, which would constitute a significant background load on the bolometers; and (2) reduce the losses in transmission due to absorption from the stack of five crystal substrates, which drops dramatically with temperature. The absorption in a crystal at FIR wavelengths is the result of the interactive coupling between the incident radiation and phonons – the thermally induced vibrations of the constituent atoms of the substrate crystal lattice. Because the phonon population is much reduced at low temperatures, cooling the crystal effectively reduces the absorption.
The two obvious candidates (uniaxial birefringent) crystals are sapphire and quartz, because of their favorable optical properties in the FIR/submm. Sapphire is chosen over quartz due to its larger difference between ordinary and extraordinary refraction indices [$\Delta n_{\rm e-o} \approx 0.32$ for sapphire, and $\approx 0.032$ for quartz; @Loewenstein1973], which implies a smaller thickness for the substrates (see Equation \[eq:phase\_HWP\]). Since quartz and sapphire have a comparable level of absorption at cryogenic temperatures in the wavelength range of 200–600$\mu$m [@Loewenstein1973], thinner substrates are desirable to minimise absorption losses. Nonetheless, the thin sapphire substrates chosen for the BLASTPol HWP do indeed show appreciable absorption, especially at the shortest wavelengths (250$\mu$m band; see Section \[sec:HWP\_spectra\]).
The five slabs of the Pancharatnam [-@Pancharatnam1955] design all have the same thickness. To cover the broad wavelength range of 200–600$\mu$m, a substrate thickness is chosen to produce a HWP at the central wavelength of the central band, 350$\mu$m. By using the values of the refractive indices for cold sapphire published by @Loewenstein1973 and @Cook1985 [$\Delta n^{\rm 350\,\mu m}_{\rm e-o} \approx
0.32$], and imposing the required phase shift of 180$^{\circ}$ between the two orthogonal polarisations travelling through the plate, Equation \[eq:phase\_HWP\] yields for the thickness of a single substrate a value $\sim$0.547mm. The nearest available thickness on the market is 0.5mm, and sapphire can not be easily ground to the desired thickness due to its brittleness. A deviation of $\sim$0.047mm from the desired thickness would naively translate into a departure of up to $\sim$15$^{\circ}$ from the ideal phase shift of 180$^{\circ}$ at 350$\mu$m. However, the case of a multi-slab Pancharatnam HWP is more complex, since the phase shift becomes an “effective” one and requires proper modelling as a function of frequency, which we have included in Section \[sec:phase\_shift\].
The orientation of the optic axis on each sapphire substrate is determined with a polarising Fourier transform spectrometer (pFTS hereafter), which is briefly described in Section \[sec:cold\_spectra\]. Each substrate is rotated between two aligned polarisers at the pFTS output until a maximum signal is achieved. The use of two polarisers avoids any complication from a partially polarised detecting system and any cross polarisation incurred from the pFTS output mirrors. The HWP is assembled by marking the side of each substrate with its reference optic axis and rotating each element according to the Pancharatnam design described in the previous section. The stack of five carefully-oriented sapphire substrates, interspersed with 6$\mu$m layers of polyethylene, are bonded together with a hot-pressing technique used in standard FIR/submm filter production [@Ade2006]. The polyethylene has negligible effects on the final optical performance of the HWP, because when heated it seeps into the roughened surfaces of the adjacent substrates.
In order to improve the robustness of the bond, the individual substrates are sandblasted with aluminium oxide (Al$_2$O$_3$) prior to fusion; this procedure dramatically improves the grip of the polyethylene between adjacent crystal surfaces. Careful cleansing and degreasing of all the crystal surfaces is required after sandblasting; in particular, we found trichloroethylene to be most effective in removing the traces of oily substances due to the sandblasting process.
A two-layer broadband anti-reflection coating (ARC), necessary to maximise the in-band transmission of the HWP, is also hot-pressed to the front and back surfaces of the assembled plate, again using 6$\mu$m layers of polyethylene. The layer adjacent to the sapphire is an artificial dielectric metamaterial (ADM) composed of metal-mesh patterned onto polypropylene sheets [@Zhang2009], while the outer layer is a thin film of porous polytetrafluoroethylene (PTFE). The thickness of the final stack (coated HWP) is $2.80\pm 0.01$mm. The diameter of the ARC is set to $88.0\pm 0.1$mm, slightly smaller than that of the HWP ($100.0\pm 0.1$mm) to avoid any contact between the coating and the HWP mount [see @Fissel2010]; the ARC is bonded concentrically to the HWP and thus its diameter defines the optically-active area of the HWP.
Because of the thermal expansion mismatch between the sapphire and the polypropylene, the HWP assembly has been cryogenically cycled numerous times prior to the flight to test the robustness of the bond at liquid helium temperatures. The HWP has been successfully installed in the BLASTPol cryogenic receiver and has been flown twice from a balloon platform, without delamination of the ARC or damage to the assembly.
However, for cryogenic crystal HWPs much larger than the BLASTPol one, the application of a metal-mesh ADM as an ARC has proven extremely challenging. Therefore, extending previous work by @Pisano2008, we have recently designed and realised a prototype polypropylene-embedded metal-mesh broadband achromatic HWP for mm wavelengths [@Zhang2011]; this will allow next generation experiments with large-aperture detector arrays to be equipped with large-format ($\gtrsim$20cm in diameter) HWPs for broadband polarisation modulation.
Spectral characterisation {#sec:HWP_spectra}
=========================
The first step to retrieving the frequency-dependent Mueller matrix and phase shift of the BLASTPol HWP is to measure its transmission as a function of frequency and incoming polarisation state. Because of the strong dependence of the sapphire absorption coefficient on temperature, we can not limit ourselves to determining the room-temperature response of the HWP, which is designed to operate at cryogenic temperatures. Therefore, we measure its spectral response in a vacuum cavity, cooled to temperatures as low as currently possible with our experimental apparatus ($\sim$120K)[^5].
Experimental setup {#sec:cold_spectra}
------------------
We fully characterise the spectral performance of the BLASTPol HWP by using a pFTS of the Martin-Puplett [-@MartinPuplett1970] type. A schematic drawing of the experimental setup is shown in Fig. \[fig:FTS\_scheme\_HWP\_measurements\]; in the following, we describe each element in sequential order from the source to the detector system.
![Schematic drawing of the spectral measurements setup. The $f/3.5$ horizontally polarised output of a pFTS illuminates the HWP with a $\sim$8$^{\circ}$ incidence angle and is focused directly onto the horn aperture of the bolometric detector. The analyser alternatively parallel or perpendicular to the pFTS output polariser creates the necessary polarisation selection for the “co-pol” and “cross-pol” sets of measurements, as depicted in the table above. The lines in the circles indicate the selected polarisation, so that for the photolithographed polarisers the wire grid orientation is perpendicular to the lines.[]{data-label="fig:FTS_scheme_HWP_measurements"}](cold_finger_schematics.pdf){width="1.\linewidth"}
The source is an incoherent mercury arc lamp with an aperture of 10mm, whose emission is well approximated by a blackbody spectrum at $T_{\rm eff}\approx
2000$K; a low-pass filter blocks radiation from the source at wavelengths shorter than $\sim$3.4$\mu$m. The interferometer is equipped with a P1[^6] beam divider, a P2 input polariser (at the source), and a P10 output polariser. The pFTS has a (horizontally) polarised output focused beam with $f/3.5$ or, in other words, a converging beam with angles $\theta\la8^{\circ}$.
This beam spread is conveniently close to the $\sim$5.7$^{\circ}$ incidence angle that the HWP is illuminated by in the $f/5$ BLASTPol optics box [see @Fissel2010; @Pascale2012], therefore it is ideal optically to place the HWP between the pFTS output polariser and the detector system the beam focuses onto, without the need for additional optical elements; this also ensures an even illumination of the entire HWP optically-active area.
We position the HWP in a liquid nitrogen-cooled removable module retrofitted in the vacuum cavity at the output port of the pFTS; a photograph and a brief description of the module, which we refer to as “cold finger”, are given in Fig. \[fig:cold\_finger\]. The manually-driven rotating mount can rotate the HWP about its optical axis to obtain the polarisation modulation, with a resolution on the rotation angle of 0.06$^{\circ}$.
![Photograph of the “cold finger” module, which fits in the vacuum cavity at the output port of the pFTS. The central cylinder is hollow and has to be regularly replenished with liquid nitrogen to maintain the temperature of the HWP at $\sim$120K. Aluminium insulation and a thick copper strap improve the thermal performance of the module. Two thermometers monitor the temperature at the bottom of the cylinder (base plate) at the edge of the copper HWP holder. The rotator is manually driven via a gear train and a vacuum-seal shaft leading to a manual knob outside the module. The resolution of the analogue encoder on the rotation angle is 0.06$^{\circ}$. The presence of a thermometer on the rotating element prevents rotations greater than $\sim$180$^{\circ}$.[]{data-label="fig:cold_finger"}](cold_finger_front.pdf){width="1.\linewidth"}
The HWP is placed centrally between the pFTS output polariser and a P10 analyser, installed at the exit port of the vacuum cavity; the efficiency of these polarisers is separately determined to exceed 99.8% over the range of frequencies of interest, with a cross-polarisation of less than 0.1%. The polarisers are initially aligned with respect to each other, with the grid wires vertical (thus selecting horizontal polarisation) with respect to the optical bench.
A small cryostat, connected with no air gaps to the exit port of the vacuum cavity, houses a feedhorn-coupled composite bolometer cooled to 1.5K by pumping on the liquid helium bath. The spectral coverage of the data is thus defined by the cut-off frequency of the light collector waveguide (5cm$^{-1}$) and by a low-pass filter (60cm$^{-1}$) installed in the cryostat to minimise photon noise.
Finally, the rapid-scan system records interferograms with a 8$\mu$m sampling interval over a 10cm optical path difference, at a scan speed of 2cms$^{-1}$; this results in a Nyquist frequency of 625cm$^{-1}$ and a spectral resolution of 0.05cm$^{-1}$.
Measurement strategy and results {#sec:measurements_and_results}
--------------------------------
After the roughly two hours needed for the cold finger module to thermalise, its base plate reaches temperatures close to 77K, while the HWP holder thermalises at about 120K, despite the thermal insulation and high thermal conductivity link to the base plate. Other cryogenic tests conducted by bonding a thermometer at the center of a single slab of sapphire ensure that the temperature measured at the edge of an aluminium or copper holder closely matches that of the sapphire substrate at its center. While maintaining a constant level of liquid nitrogen in the cold finger, we can characterise the spectral response of the cold HWP, by rotating it inside the vacuum cavity.
Following the convention depicted in Fig. \[fig:FTS\_scheme\_HWP\_measurements\], measurements with aligned polarisers are referred to as “co-pol” transmission, $T_{\rm cp}$. The HWP has a complementary response when the analyser is rotated by 90$^{\circ}$ about the optical axis of the system (i.e., horizontal wires, selecting vertical polarisation); data taken with this configuration are necessary to completely characterise the HWP, and are referred to as “cross-pol” transmission[^7], $T_{\rm
xp}$.
The very first dataset, which we refer to as the background spectrum, must be obtained in co-pol configuration by scanning the spectrometer in the absence of the HWP. This dataset defines the pFTS reference spectral envelope, and it is the set against which all the following spectra are divided in order to account for the spectral features of the source, pFTS optics, and detector system. Subsequently, the HWP is inserted in between the polarisers in co-pol configuration, and spectra are acquired at many different HWP rotation angles (resulting in a data cube). To enhance the spectral signal-to-noise ratio, each dataset at a given angle is obtained by computing the Fourier transform of an (apodised and phase-corrected) average of 60 interferograms with the mirror scanned in both the forward and backward directions. As anticipated, the resulting spectra are divided by the background dataset, which in turn is the average of three spectra, to obtain the transmission of the coated HWP alone as a function of frequency.
Over two days of measurements, we acquire data cubes for co-pol and cross-pol transmissions, shown respectively in Fig. \[fig:BLAST\_pol\_spareHWP\_ARC\_copol\_spectra\] and Fig. \[fig:BLAST\_pol\_spareHWP\_ARC\_xpol\_spectra\], where we only display spectra taken at rotation angles near the HWP maxima and minima, for visual clarity. The full datasets, including spectra taken at intermediate angles (roughly every 10$^{\circ}$ between 0$^{\circ}$ and $\sim$180$^{\circ}$), are shown in Figs. \[fig:spareHWP\_copol\_spectra\_surface\] and \[fig:spareHWP\_xpol\_spectra\_surface\] as 3-D surfaces.
Because of the controlled environment in the vacuum cavity, our measurements are not susceptible to changes in the external environment; however, we repeat background scans at the very end of our measurement session to monitor drifts in the bolometer responsivity and other potential systematic effects. Prior to inserting the HWP in the cavity, we have also characterised the instrumental cross-pol of this setup by rotating the analyser by 90$^{\circ}$ in cross-pol configuration and acquiring three spectra. By averaging these cross-pol spectra and dividing by the co-pol background, we measure a cross-pol level of 0.2% or less across the entire spectral range of interest (5–60cm$^{-1}$); we include the resulting cross-pol spectrum in Figs. \[fig:BLAST\_pol\_spareHWP\_ARC\_copol\_spectra\] and \[fig:BLAST\_pol\_spareHWP\_ARC\_xpol\_spectra\] (dark pink line).
![Measured co-pol transmission spectra of the coated BLASTPol HWP cooled to $\sim$120K. Each line is obtained at a different HWP rotation angle by computing the Fourier transform of an (apodised and phase-corrected) average of 60 interferograms. For visual clarity, we only show here spectra at rotation angles near the HWP maxima and minima, omitting data taken at intermediate angles (shown in Fig. \[fig:spareHWP\_copol\_spectra\_surface\]). The solid vertical black lines show the approximate extent of the three BLASTPol bands.[]{data-label="fig:BLAST_pol_spareHWP_ARC_copol_spectra"}](BLAST_pol_spareHWP_ARC_cold_copol_20101021_log_vertical_legend_mod.pdf){width="1.\linewidth"}
![Measured spectra of the HWP cooled to $\sim$120K equivalent to those shown in Fig. \[fig:BLAST\_pol\_spareHWP\_ARC\_copol\_spectra\] but for cross-pol transmission.[]{data-label="fig:BLAST_pol_spareHWP_ARC_xpol_spectra"}](BLAST_pol_spareHWP_ARC_cold_xpol_20101022_log_vertical_legend_mod.pdf){width="1.\linewidth"}
![Data cube represented by a surface obtained by stacking a set of spectral co-pol transmissions of the HWP at different angles. Each measured spectra is a slice of the surface perpendicular to the angle axis.[]{data-label="fig:spareHWP_copol_spectra_surface"}](spareHWP_copol_spectra_surface.pdf){width="1.\linewidth"}
![Equivalent data cube to that shown in Fig. \[fig:spareHWP\_copol\_spectra\_surface\] but for cross-pol transmission. Note how the two surfaces are complementarily in counterphase to each other. Each measured spectra is a slice of the surface perpendicular to the angle axis.[]{data-label="fig:spareHWP_xpol_spectra_surface"}](spareHWP_xpol_spectra_surface.pdf){width="1.\linewidth"}
An ideal HWP modulates the polarisation at 4$\theta$, therefore in a complete revolution there are four maxima (and minima), two for each of the birefringent axes. The (arbitrary) zero angle in Figs. \[fig:BLAST\_pol\_spareHWP\_ARC\_copol\_spectra\], \[fig:BLAST\_pol\_spareHWP\_ARC\_xpol\_spectra\], \[fig:spareHWP\_copol\_spectra\_surface\], and \[fig:spareHWP\_xpol\_spectra\_surface\] does not (need to) coincide with a HWP maximum (minimum), which is the HWP angle at which we measure maximum (minimum) total power on the detector; this of course includes signal outside of the HWP bands (in the range 5–60cm$^{-1}$). As we show in Section \[sec:beta\_ea\], the position of the equivalent axes of the sapphire plate stack (and hence the position of the HWP maxima/minima) depends upon the wavelength. Therefore the HWP maxima (and minima) we assign while taking spectra are just rough approximations. Although we increase the angle sampling rate in the vicinities of a maximum or minimum (with steps of 3$^{\circ}$ rather than 10$^{\circ}$), in order to fully characterise the HWP it is not necessary to take spectra [*exactly*]{} at its maxima or minima.
We verify that the experimental setup is symmetric with respect to the HWP rotation and that there are no artifacts arising from misalignments in the optical setup by measuring no appreciable change in pairs of datasets taken at angles that are exactly 180$^{\circ}$ apart, due to polarisation symmetry.
In the surfaces depicted in Figs. \[fig:spareHWP\_copol\_spectra\_surface\] and \[fig:spareHWP\_xpol\_spectra\_surface\], slices of the data cube along the wavenumber axis constitute the measured spectra at different HWP angles, while slices along the angle axis represent the modulation function of the HWP at a given frequency or, more precisely, within a narrow band of frequencies defined by a combination of spectral resolution and the spectrometer’s instrument response function.
The features visible in all spectra are spectral fringes due to standing waves generated inside the stack of dielectric substrates (even with a quasi-perfect impedance matching coating on the outer surfaces); the presence of several interspersed layers of polyethylene enhances the amplitude of the fringes by introducing small amounts of absorption at every internal reflection.
Uncertainties on the measured spectra {#sec:uncertainties _spectra}
-------------------------------------
Because we average 60 interferograms to obtain the final spectrum at each HWP position, the statistical uncertainty associated with the average on a single dataset is found to be negligible, as expected. Rather, we decide to average together all the available background interferograms that are collected over one day of measurements, and take their statistical dispersion as our estimate of the uncertainty associated with all the spectra collected on that day. Because the thermodynamic conditions in the cavity under vacuum are not susceptible to changes in the external environment, this procedure allows us to account for drifts in the bolometer responsivity and other potential systematic effects. We report in Fig. \[fig:plot\_noise\_backs\_HWP\] the mean background spectra and the associated error (shown as 10$\sigma$ error bars for visual clarity) for the co-pol and cross-pol measurement sessions. These errors (1$\sigma$) are used in Section \[sec:empirical\_model\] to estimate the uncertainties on the HWP Mueller matrix coefficients.
![Noise estimation for the spectra shown in Figs. \[fig:BLAST\_pol\_spareHWP\_ARC\_copol\_spectra\] and \[fig:BLAST\_pol\_spareHWP\_ARC\_xpol\_spectra\]. We plot the mean background spectra (in arbitrary units) for the co-pol (black solid line) and cross-pol (yellow solid line, shifted by 1 in the positive $y$ direction for visual clarity) as a function of wavenumber. The (10$\sigma$) error bars (in red) are quantified as the statistical error on the mean. Also shown for reference is the relative spectral response of the three BLASTPol channels, in arbitrary units. Henceforth, we adopt a colour code in the plots whereby curves referring to the three BLASTPol bands, 250, 350, and 500$\mu$m are drawn in blue, green, and red, respectively.[]{data-label="fig:plot_noise_backs_HWP"}](plot_noise_backs_HWP.pdf){width="1.\linewidth"}
Comparison with design parameters {#sec:physical_model}
---------------------------------
While the major goal of this work is to provide an avenue for including the measured non-idealities of the BLASTPol HWP [*as-built*]{} in a map-making algorithm, it is useful at this stage to compare the nominal values of the build parameters we assumed to design the HWP with the actual values that can be estimated via the physical and analytical model developed by @Savini2006 [@Savini2009 which we refer to for a complete account of the formalism and implementation].
In this work, the physical model is fit to the spectral data described in Section \[sec:measurements\_and\_results\] by allowing the HWP build parameters to vary in a physical way around the nominal values. We report in Fig. \[fig:BLAST\_pol\_spareHWP\_ARC\_copol\_spectra\_Giorgio\] a comparison between the measured co-pol transmission spectra near the HWP maxima/minima and the corresponding physical model, whose best-fit parameters are listed in Table \[tab:physical\_parameters\_HWP\], along with the nominal ones.
![Comparison between the measured co-pol transmission spectra near the HWP maxima/minima and the corresponding physical model of @Savini2006, which is fit to the data by allowing the build parameters of the HWP (refraction index and thickness of all the materials, orientation of the birefringent substrates) to vary in a physical way around the nominal values.[]{data-label="fig:BLAST_pol_spareHWP_ARC_copol_spectra_Giorgio"}](BLAST_pol_spareHWP_ARC_cold_copol_20101021_Giorgio.pdf){width="1.\linewidth"}
parameter units nominal best-fit uncertainty
----------------- -------- --------- ---------- -------------
$n^0_o$ 3.052 3.065 0.002
$n^1_o$ cm 0.00207 0.00008
$n^0_e$ 3.372 3.444 0.001
$n^1_e$ cm 0.00127 0.00008
$d$ $\mu$m 500.0 490.4 0.5
$t_{\rm PE}$ $\mu$m 6.0 3.5 0.5
$\Phi_1$ deg 26 22 1
$\Phi_2$ deg 90.3 87 1
$\Phi_3$ deg 26 23 1
$\Phi_4$ deg 0 -1 1
$n_{\rm ARC}^1$ 1.375 1.26 0.01
$t_{\rm ARC}^1$ $\mu$m 54 72 2
$n_{\rm ARC}^2$ 1.62 0.01
$t_{\rm ARC}^2$ $\mu$m 37 3
$n_{\rm ARC}^3$ 2.28 0.01
$t_{\rm ARC}^3$ $\mu$m 36 2
: Best-fit build parameters of the BLASTPol HWP estimated using the physical model of @Savini2006. $n^0_{e(o)}$ and $n^1_{e(o)}$ are, respectively, the constant and linear terms of the refraction index for the sapphire (extra)ordinary axis, modelled with a linear dependence on wavenumber; $d$ is the thickness of the sapphire substrates; $t_{\rm PE}$ is the thickness of the polyethylene layers; $\Phi_i$ ($i=1$, ..., 4) are the four angles at which the sapphire substrates are oriented (assuming $\Phi_0=0^{\circ}$); $n_{\rm ARC}^i$ and $t_{\rm ARC}^i$ ($i=1$, 2, 3) are, respectively, the refraction indeces and the thicknesses of the i-th layer of ARC (in order of penetration into the HWP), which is modelled as series of equivalent dielectrics.
\[tab:physical\_parameters\_HWP\]
Among all the parameters presented in Table \[tab:physical\_parameters\_HWP\], let us focus on the angles at which the sapphire substrates are oriented, $\Phi_i$ ($i=1$, ..., 4; see Section \[sec:HWP\_manu\]), since they have a greater impact on what follows. The best-fit orientation angles are up to 3–4$^\circ$ smaller than the design goal. While this is not surprising given the practical challenge of keeping a stack of five plates (interspersed with thin slippery layers of polyethylene) aligned to within $\sim$3mm (linear length of a 3.5$^\circ$ arc for a 100mm diameter), such a deviation from the desired values will affect the HWP performance, and in particular its phase shift as a function of frequency.
In the following sections we present all the HWP performance parameters, for which the physical model presented in this section is verified to be in general agreement with the empirical model we develop in Section \[sec:empirical\_model\]. However, we will show in Section \[sec:phase\_shift\] that the empirical model is inadequate to retrieve the HWP phase shift and we will have to resort to the physical model again to compare the design and best-fit phase shift versus frequency. This will give us a chance to expand more on which performance parameters are more directly affected by plate alignment errors.
Modulation function and efficiency {#sec:modulation_function}
----------------------------------
We can reduce the dependence on frequency of our data cubes by integrating over the spectral bands of BLASTPol, as follows: $$\label{eq:modulation_curve}
\overline{T}^{\rm ch}_{\rm cp} \left(\theta\right)=
\frac{\int_0^{\infty} \Sigma^{\rm ch}\left(\nu\right)\,T_{\rm
cp}\left(\theta,\nu\right)\,d\nu}{\int_0^{\infty} \Sigma^{\rm
ch}\left(\nu\right)\,d\nu}~.$$ Here the superscript “ch” refers to one among 250, 350, and 500$\mu$m, $\Sigma^{\rm ch}\left(\nu\right)$ is the measured spectral response of each of the BLASTPol bands [see @Pascale2008], and $T_{\rm
cp}\left(\theta,\nu\right)$ are points on the co-pol surface depicted in Fig. \[fig:spareHWP\_copol\_spectra\_surface\]. A similar expression can be written for the cross-pol band-integrated transmission. By performing this integration at every angle for which spectral data have been obtained, the interpolation of these data points will result in the modulation functions of the HWP at $\sim$120K for each of the BLASTPol bands; these curves are shown in Figs. \[fig:modulation\_curves\_copol\_spareHWP\] (co-pol) and \[fig:modulation\_curves\_xpol\_spareHWP\] (cross-pol).
![Band-integrated co-pol modulation functions of the BLASTPol HWP at $\sim$120K. The curves show the HWP polarisation modulation functions for a fully polarised source (with a flat spectrum) parallel to the analyser in the three spectral bands. Note how the positions of the maxima (and minima) depend on the wavelength, even when considering a flat-spectrum polarised input source; the dotted vertical lines show the band-integrated positions of the HWP extrema (shown in Fig. \[fig:HWP\_offset\_vs\_freq\]), which result from the fitting routine described in the next sections.[]{data-label="fig:modulation_curves_copol_spareHWP"}](modulation_curves_copol_spareHWP_last.pdf){width="1.\linewidth"}
![Band-integrated modulation functions equivalent to those shown in Fig. \[fig:modulation\_curves\_copol\_spareHWP\] but for cross-pol transmission.[]{data-label="fig:modulation_curves_xpol_spareHWP"}](modulation_curves_xpol_spareHWP_last.pdf){width="1.\linewidth"}
The modulation curves presented here are valid for input sources that have a flat spectrum in the BLASTPol bands. Equation \[eq:modulation\_curve\] can be generalised to include the known (or assumed) spectral signature of a given astronomical or calibration source [see also e.g., @Novak1989a Equation 2]. More generally, all the band-averaged quantities that we have defined here and will be defined in the following are potentially affected by the spectral shape of the input source. However, we will see how the HWP transmission and modulation efficiency are very weakly dependent on the spectral index of the input source, whereas the position of the equivalent axes of the sapphire plate stack is more significantly affected [see also the analysis carried out by @Savini2009], especially at 250 and 500$\mu$m.
Figs. \[fig:modulation\_curves\_copol\_spareHWP\] and \[fig:modulation\_curves\_xpol\_spareHWP\] clearly show that there is a significant dependence of the position of the HWP maxima and minima upon frequency, even when considering a flat-spectrum polarised input source. These effects are particularly important for a “HWP step and integrate” experiment such as BLASTPol, and a polarisation calibration must be performed by using information from the characterisation of the HWP. We begin to tackle this problem in the next section, where we outline a relatively simple solution to account for most of the HWP non-idealities in the data-analysis pipeline, and in particular in the map-making algorithm (see Section \[sec:map\_maker\]).
The spectral transmission datasets of the HWP cooled to $\sim$120K, when compared to those taken with the HWP at room temperature (not reported here for brevity), show a definite abatement of the in-band losses due to absorption from sapphire, as expected. However the effect is still appreciable, especially above $\sim$25cm$^{-1}$. As we will show in the following, we have evidence that the residual absorption nearly vanishes when the sapphire is further cooled to 4K, as it is when the HWP is installed in the BLASTPol cryostat. While it is not currently feasible for us to measure the spectral response of the HWP cooled to 4K, the unique quality and completeness of our dataset allow us to fully characterise the performance of the BLASTPol HWP.
![Ordinary (solid) and extraordinary (dashed) sapphire absorption coefficient as a function of wavenumber, at cryogenic temperatures. The two analytical relations covering the whole frequency range are derived by Savini (2010, pers. comm.) from a set of spectral measurements of a sample at 80K, and, strictly speaking, only apply at frequencies $\lesssim$1THz (dotted vertical line). For reference, we plot measurements from @Loewenstein1973 [purple] at 1.5K and @Cook1985 [ochre] at 60K, displaced by 0.002cm$^{-1}$ in $y$ for visual clarity; the lines connecting these data points follow the convention shown in the legend.[]{data-label="fig:sapphire_absorption_coefficient_cold"}](sapphire_absorption_coefficient_cold.pdf){width="1.\linewidth"}
We extrapolate our “cold” dataset to 4K, using the analytical relations shown in Fig. \[fig:sapphire\_absorption\_coefficient\_cold\][^8]. The HWP modulation efficiency is defined as ($T^{0^{\circ}}_{\rm
cp}-T^{0^{\circ}}_{\rm xp}$)/($T^{0^{\circ}}_{\rm
cp}+T^{0^{\circ}}_{\rm xp}$), where the “co-pol” and “cross-pol” transmissions, $T^{0^{\circ}}_{\rm cp}$ and $T^{0^{\circ}}_{\rm xp}$, are the spectral responses of the HWP near one of the transmission maxima (called 0$^{\circ}$ here), between parallel and perpendicular polarisers, respectively. The inferred co-pol/cross-pol transmissions and modulation efficiency of the BLASTPol HWP (with its axis at 0$^{\circ}$) at 4K are shown in Fig. \[fig:modeff\]. For a flat-spectrum input source, the band-integrated transmission of the HWP at its maxima is $\sim$0.87, $\sim$0.91, and $\sim$0.95 at 250, 350, and 500$\mu$m, respectively; whereas the band-integrated cross-pol is $\la$0.5%, $\la$0.2%, and $\la$0.5%, respectively; finally, the band-integrated modulation efficiency is $\sim$98.8% $\sim$99.5%, and $\sim$99.0%, respectively.
Empirical modelling {#sec:empirical_model}
===================
The final goal of this section is to provide a set of usable parameters that completely describe the performance of the HWP as measured in the laboratory. This set of parameters consists of the 16 coefficients of the Mueller matrix of a generic HWP, and the actual phase shift. For an ideal HWP, the Mueller matrix at $\theta=0^{\circ}$ reads [@Collett1993] $$\label{eq:ideal_HWP}
{\mathbfss M_{\rm HWP}}=
\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & -1 & 0 \\
0 & 0 & 0 & -1 \\
\end{pmatrix}~,$$ and the phase shift is $\Delta\varphi=180^{\circ}$. For a real HWP, these parameters always depart from the ideal case to some extent, and certainly depend upon frequency. In the following we describe an empirical model that we develop specifically for the characterisation of the BLASTPol HWP, though we note that it can be applied to any HWP to recover its frequency-dependent descriptive parameters. Such an empirical model is complementary to the physical and analytical one developed by @Savini2006 [@Savini2009], which produces an analogous output by modelling the non-idealities of the components of the HWP assembly and their optical parameters.
Mueller matrix characterisation {#sec:mueller_HWP}
-------------------------------
By recalling the Stokes formalism, we can formalise the experimental apparatus described in Section \[sec:cold\_spectra\] as a series of matrix products, as follows: $$\begin{aligned}
S^{\rm\,cp}_{\rm out}&=&\vec{\mathbfit D}^{\rm T}\,\cdot\,{\mathbfss M^{\rm h}_{\rm p}}\,\cdot\,{\mathbfss R}\left(-\theta\right)\,\cdot\,{\mathbfss M_{\rm HWP}}\,\cdot\,{\mathbfss R}\left(\theta\right)\,\cdot\vec{\mathbfit S}^{\rm\,h}_{\rm in}~;\label{eq:mueller_HWP_1a}\\
S^{\rm\,xp}_{\rm out}&=&\vec{\mathbfit D}^{\rm
T}\,\cdot\,{\mathbfss M^{\rm v}_{\rm p}}\,\cdot\,{\mathbfss
R}\left(-\theta\right)\,\cdot\,{\mathbfss M_{\rm
HWP}}\,\cdot\,{\mathbfss
R}\left(\theta\right)\,\cdot\vec{\mathbfit S}^{\rm\,h}_{\rm
in}~.\label{eq:mueller_HWP_1b}\end{aligned}$$ Here $\vec{\mathbfit D}$ is the Stokes vector for a bolometric (polarisation insensitive) intensity detector, ${\mathbfss M^{\rm
h}_{\rm p}}$ is the Mueller matrix of an ideal horizontal polariser, ${\mathbfss M^{\rm v}_{\rm p}}$ is that of an ideal vertical polariser, ${\mathbfss R}\left(\theta\right)$ is the generic Mueller rotation matrix, and $\vec{\mathbfit S}_{\rm in}$ is the horizontally polarised input beam from the pFTS. By expanding all the matrices in Equation \[eq:mueller\_HWP\_1a\], $$\begin{aligned}
\label{eq:mueller_HWP_2}
S^{\rm\,cp}_{\rm out}=\begin{pmatrix}
1 & 0 & 0 & 0
\end{pmatrix}\cdot
\begin{pmatrix}
1 & 1 & 0 & 0 \\
1 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
\end{pmatrix}\cdot~~~~~~~~~~~~~~~~\\
\nonumber\cdot
\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & \cos(2\theta) & \sin(2\theta) & 0 \\
0 & -\sin(2\theta) & \cos(2\theta) & 0 \\
0 & 0 & 0 & 1 \\
\end{pmatrix}
\cdot
\begin{pmatrix}
a_{00} & a_{01} & a_{02} & a_{03} \\
a_{10} & a_{11} & a_{12} & a_{13} \\
a_{20} & a_{21} & a_{22} & a_{23} \\
a_{30} & a_{31} & a_{32} & a_{33} \\
\end{pmatrix}\cdot\\
\nonumber\cdot
\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & \cos(2\theta) & -\sin(2\theta) & 0 \\
0 & \sin(2\theta) & \cos(2\theta) & 0 \\
0 & 0 & 0 & 1 \\
\end{pmatrix}
\cdot
\begin{pmatrix}
1 \\
1 \\
0 \\
0 \\
\end{pmatrix}~,~~~~~~~~~~~~~~~~~~~~~~~~\end{aligned}$$ we can compute the products and rearrange as $$\begin{aligned}
S^{\rm\,cp}_{\rm out}=\frac{1}{2} \Big[2 a_{00}+a_{11}+a_{22}+2
(a_{01}+a_{10})
\cos2\theta+\label{eq:mueller_HWP_4a}~~~~~\\
\nonumber+(a_{11}-a_{22}) \cos4\theta-2 (a_{02}+a_{20})
\sin2\theta -(a_{12}+a_{21}) \sin4\theta\Big]\\
=A+B \sin2\theta +C \cos2\theta + D \sin4\theta + E
\cos4\theta~,~~~~~\label{eq:mueller_HWP_4b}\end{aligned}$$ with $$\begin{aligned}
\label{eq:mueller_HWP_5}
\nonumber A &\equiv& a_{00}+\frac{a_{11}}{2}+\frac{a_{22}}{2}~,\\
B&\equiv&-\left(a_{02} + a_{20}\right),~~~~C\equiv a_{01}+a_{10}~,\\
\nonumber D&\equiv&-\frac{1}{2}\left(a_{12} +
a_{21}\right),~~~~E\equiv\frac{1}{2}\left(a_{11}-a_{22}\right)~.\end{aligned}$$ Similarly, noting that $${\mathbfss M^{\rm v}_{\rm p}}=\begin{pmatrix}
1 & -1 & 0 & 0 \\
-1 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
\end{pmatrix}~,$$ we rearrange Equation \[eq:mueller\_HWP\_1b\] as $$\begin{aligned}
S^{\rm\,xp}_{\rm out}=A'+B' \sin2\theta +C' \cos2\theta + D'
\sin4\theta + E' \cos4\theta~,~~~~~\label{eq:mueller_HWP_8b}\end{aligned}$$ with $$\begin{aligned}
\label{eq:mueller_HWP_9}
\nonumber A' &\equiv& a_{00}-\frac{a_{11}}{2}-\frac{a_{22}}{2}~,\\
B'&\equiv& a_{20} - a_{02},~~~~C'\equiv a_{01}-a_{10}~,\\
\nonumber D'&\equiv&\frac{1}{2}\left(a_{12} +
a_{21}\right),~~~~E'\equiv\frac{1}{2}\left(a_{22}-a_{11}\right).\end{aligned}$$
Finally, by performing linear combinations of the quantities defined in Equations \[eq:mueller\_HWP\_5\] and \[eq:mueller\_HWP\_9\], one can write the individual elements that compose the Mueller matrix of a generic HWP as $$\begin{aligned}
\label{eq:mueller_HWP_10}
\nonumber a_{00} &=& \frac{1}{2}\,\left(A+A'\right),~~~a_{01}=\frac{1}{2}\,\left(C+C'\right)~,\\
a_{10} &=& \frac{1}{2}\,\left(C-C'\right),~~~a_{11}=\frac{1}{2}\,\left(A-A'+E-E'\right)~,\\
\nonumber a_{02} &=& -\frac{1}{2}\,\left(B+B'\right),~~~a_{20}=\frac{1}{2}\,\left(B'-B\right)~,\\
\nonumber a_{22} &=&
\frac{1}{2}\,\left(A-A'-E+E'\right),~~~a_{12}=a_{21}=\frac{1}{2}\,\left(D'-D\right)~,\end{aligned}$$ where in the last equality we currently assume the symmetry of two coefficients, $a_{12}=a_{21}$. This degeneracy may be broken by imposing the conservation of energy, i.e. by requiring the output Stokes vector resulting from a generic polarised input travelling through the recovered HWP Mueller matrix to satisfy $I^2 = Q^2 +
U^2$. Alternatively, the degeneracy can be broken by taking spectra at an intermediate configuration between co- and cross-pol; this additional constraint will be included in a future work (Spencer et al., in preparation). Also, because our experimental setup is sensitive to linear but not circular polarisation, this method only allows us to constrain the 9 elements of the Mueller matrix associated with $\left[I,Q,U\right]$. The remaining 7 coefficients associated with $V$ can only be measured with the use of a quarter-wave plate, which induces a phase shift of 90$^{\circ}$ between the two orthogonal polarisations travelling through the plate; this measurement is beyond the scope of this paper and not pertinent to the needs of BLASTPol.
We want to estimate the 9 coefficients derived in Equation \[eq:mueller\_HWP\_10\] from the co-pol and cross-pol data cubes described in Section \[sec:measurements\_and\_results\]. Equations \[eq:mueller\_HWP\_4b\] and \[eq:mueller\_HWP\_8b\] encode a simple dependence of $S^{\rm\,cp}_{\rm out}$ and $S^{\rm\,xp}_{\rm out}$ upon $\theta$, the HWP rotation angle. Therefore, for a given frequency, a fitting routine can be applied to the measured transmission curves as a function of $\theta$, to determine the parameter sets $[A,B,C,D,E]$ and $[A',B',C',D',E']$ for the co-pol and cross-pol configurations, respectively. By repeating the fit for every frequency, we have an estimate of the 9 coefficients as a function of wavelength. However, this procedure does not allow us to associate an uncertainty to our estimates.
A better approach to this problem is to use a Monte Carlo simulation. We repeat the above fitting procedure 1000 times; every time we add to every individual transmission curve a realisation of white noise, scaled to the 1$\sigma$ spectral uncertainty as estimated in Fig. \[fig:plot\_noise\_backs\_HWP\], and compute the fit using this newly generated transmission curve. In addition, for every frequency we introduce a random jitter on the rotation angle that has a 1$\sigma$ amplitude of 1$^{\circ}$. The dispersion in the fitted parameters due to these two types of uncertainties, which are inherent to the measurement process, provides a realistic estimate of the uncertainty associated with each of the 9 coefficients. In particular, at each frequency, we produce 9 histograms of the 1000 fitted values. We use the mode of each distribution as our best estimate for the corresponding coefficient at that frequency, and the 68% confidence interval as the associated 1$\sigma$ error.
In Fig. \[fig:spare\_HWP\_MM\_coeff\_no\_offset\] we show a graphical representation of the 9-element Mueller matrix of the BLASTPol HWP at a given angle ($\theta=0^{\circ}$), as a function of wavenumber. In Fig. \[fig:histo\_spare\_HWP\_MM\_histos\_20.01cm-1\] we show the resulting histograms for the 9 coefficients at 20cm$^{-1}$, central frequency of the 500$\mu$m BLASTPol band; histograms at 28.57cm$^{-1}$ (350$\mu$m) and 40.02cm$^{-1}$ (250$\mu$m) look very similar and thus are not presented here for brevity.
![Graphical representation of the Mueller matrix of the BLASTPol HWP at a given angle ($\theta=0^{\circ}$), as a function of wavenumber. The (10$\sigma$) error bars (in red) are quantified via a Monte Carlo, which accounts for random errors in the spectra of amplitude as given in Fig. \[fig:plot\_noise\_backs\_HWP\], and random errors of amplitude 1$^{\circ}$ in the rotation angle.[]{data-label="fig:spare_HWP_MM_coeff_no_offset"}](spare_HWP_MM_coeff_no_offset_crop.pdf){width="1.\linewidth"}
![Histograms at 20cm$^{-1}$ (central frequency of the 500$\mu$m BLASTPol band) resulting from the Monte Carlo fit of the HWP parameters. For every histogram, the dashed red line indicates the mode of the distribution, which we adopt as our best estimate for the corresponding coefficient at that frequency, while the two dotted red lines indicate the 68% confidence interval, which we use as the uncertainty on the retrieved coefficient.[]{data-label="fig:histo_spare_HWP_MM_histos_20.01cm-1"}](histo_spare_HWP_MM_histos_20cm-1_crop.pdf){width="1.\linewidth"}
Position of the HWP equivalent axes {#sec:beta_ea}
-----------------------------------
The behaviour of the coefficients as a function of wavenumber shown in Fig. \[fig:spare\_HWP\_MM\_coeff\_no\_offset\] confirms that the position of the HWP equivalent axes, $\beta_{\rm ea}$ hereafter, has an inherent frequency dependence, which we must investigate. $\beta_{\rm ea}$ can be readily retrieved at each frequency by locating the rotation angle that corresponds to the first minimum in the fitted transmission curve. Hence, $\beta_{\rm ea}$ is measured with respect to an arbitrary constant offset that is inherent to the specific experimental setup; we set this offset to be zero at 25cm$^{-1}$. Operationally, this means that the HWP zero angle in the instrument reference frame ($\beta_0$; see Equation \[eq:polangle\]) must be calibrated using the 350$\mu$m band. A plot of $\beta_{\rm ea}$ as a function of wavenumber is given in Fig. \[fig:HWP\_offset\_vs\_freq\].
As anticipated, it is of crucial importance to derive the band-averaged value of $\beta_{\rm ea}$ for input sources with different spectral signature, as follows: $$\label{eq:beta_ea}
\overline{\beta}_{\rm ea}^{\rm ch}= \frac{\int_0^{\infty}
\Sigma^{\rm ch}\left(\nu\right)\,\beta_{\rm
ea}\left(\nu\right)\,\varsigma\left(\nu\right)\,d\nu}{\int_0^{\infty}
\Sigma^{\rm
ch}\left(\nu\right)\,\varsigma\left(\nu\right)\,d\nu}~,$$ where we adopt the same notation as in Equation \[eq:modulation\_curve\] and the known (or assumed) spectrum of an astronomical or calibration source is modelled as $\varsigma\left(\nu\right) \propto \nu^{\alpha}$. We compute Equation \[eq:beta\_ea\] for a range of spectral indices of interest: $\alpha=0$ for a flat spectrum; $\alpha=2$ for the Rayleigh-Jeans tail of a blackbody; $\alpha=4$ for interstellar dust, modelled as a modified blackbody with emissivity $\beta=2$ [@Hildebrand1983]; and finally $\alpha=-2$ as a replacement for the mid-infrared exponential on the Wien side of a blackbody to account for the variability of dust temperatures within a galaxy [@Blain1999c; @Blain2003]. The results of this analysis are shown in Fig. \[fig:HWP\_offset\_vs\_freq\] and in Table \[tab:HWP\_offset\_vs\_freq\].
![Position of the HWP equivalent axis, $\beta_{\rm ea}$, as a function of wavenumber (solid black line). Note that this quantity is defined with respect to an arbitrary constant offset that is inherent to the specific experimental setup; we set this offset to be zero at 25cm$^{-1}$. The band-averaged values for input sources with different spectral index ($\alpha$; see legend) are drawn as thick horizontal lines. Also shown for reference is the relative spectral response of the three BLASTPol channels, in arbitrary units.[]{data-label="fig:HWP_offset_vs_freq"}](HWP_offset_vs_freq.pdf){width="1.\linewidth"}
$\overline{\beta}_{\rm ea}$ \[deg\]
---------- ----------- ------------------------------------- -----------
$\alpha$ 250$\mu$m 350$\mu$m 500$\mu$m
$-$2 4.9 0.30 2.7
0 5.7 0.35 2.3
$+$2 6.6 0.39 1.9
$+$4 7.5 0.44 1.6
: Band-averaged position of the HWP equivalent axis for sources with different spectral index. The input source is assumed to have a spectrum $\varsigma\propto
\nu^{\alpha}$.[]{data-label="tab:HWP_offset_vs_freq"}
As expected, the impact of different input spectral signatures is minimal at 350$\mu$m, where the HWP has been designed to function optimally (see Section \[sec:HWP\_manu\]); whereas the spectral dependence is more pronounced at 250 and 500$\mu$m, and, if neglected, it may lead to an arbitrary rotation of the retrieved polarisation angle on the sky of magnitude $2\,\overline{\beta}_{\rm ea} =10$–15$^{\circ}$ (3–5$^{\circ}$) at 250 (500)$\mu$m (see Equation \[eq:polangle\]).
We have thus confirmed that the dependence of the HWP equivalent axes upon wavelength is inherent to the achromatic design. We now postulate that most of the non-idealities we see in the measured HWP Mueller matrix (Fig. \[fig:spare\_HWP\_MM\_coeff\_no\_offset\]) are primarily due to the wavelength dependence of $\beta_{\rm
ea}$, along with the residual absorption from sapphire at $\sim$120K. One can imagine that the HWP performance would approach the ideal case once this effect is corrected for.
![Graphical representation of the Mueller matrix of the BLASTPol HWP, equivalent to that shown in Fig. \[fig:spare\_HWP\_MM\_coeff\_no\_offset\], but including in the fit the frequency-dependent position of the HWP equivalent axes (see Fig. \[fig:HWP\_offset\_vs\_freq\]).[]{data-label="fig:spare_HWP_MM_coeff"}](spare_HWP_MM_coeff_crop.pdf){width="1.\linewidth"}
Therefore, we include $\beta_{\rm ea}\left(\nu\right)$ in our Monte Carlo as a frequency-dependent offset in the array of rotation angles (so that $\theta \rightarrow \theta - \beta_{\rm
ea}$), and repeat our simulations. The results, presented in Fig. \[fig:spare\_HWP\_MM\_coeff\], can now be qualitatively compared to the Mueller matrix of an ideal HWP (Equation \[eq:ideal\_HWP\]). The improvement is noticeable, especially in the off-diagonal elements, and the resemblance to an ideal HWP is remarkable across the entire spectral range of interest; this procedure effectively acts to diagonalise the HWP Mueller matrix. However, the transmission losses due to absorption from the sapphire at $\sim$120K still affect the diagonal elements of the matrix, as expected.
As a final improvement, we extrapolate the $\beta_{\rm
ea}$-corrected HWP Mueller matrix to 4K by including in our Monte Carlo a correction for the residual sapphire absorption (using the data presented in Fig. \[fig:sapphire\_absorption\_coefficient\_cold\]). The results are shown in Fig. \[fig:spare\_HWP\_MM\_coeff\_cold\]. Although there still seems to be residual transmission losses due to sapphire absorption at 250 and 350$\mu$m, the retrieved HWP Mueller matrix is nearly that of an ideal HWP. The band-averaged values of the matrix coefficients for a flat-spectrum input source are reported in Table \[tab:spare\_HWP\_MM\_coeff\_cold\], along with their propagated uncertainty; the off-diagonal elements are always consistent with zero within 2$\sigma$ and the modulus of the three diagonal coefficients is always $>$0.8. The combination of these coefficients with the band-averaged values of $\beta_{\rm ea}$ given in Table \[tab:HWP\_offset\_vs\_freq\] gives a complete account of the HWP non-idealities to the best of our ability.
![Graphical representation of the Mueller matrix of the BLASTPol HWP, equivalent to that shown in Fig. \[fig:spare\_HWP\_MM\_coeff\], but including a correction for the temperature dependence of the sapphire absorption coefficient (see Fig. \[fig:sapphire\_absorption\_coefficient\_cold\]).[]{data-label="fig:spare_HWP_MM_coeff_cold"}](spare_HWP_MM_coeff_cold_crop.pdf){width="1.\linewidth"}
We repeat the calculation of the band-averaged coefficients for the other spectral indices discussed in Fig. \[fig:HWP\_offset\_vs\_freq\]; we find values that are always within 1–2% of those reported in Table \[tab:spare\_HWP\_MM\_coeff\_cold\], and thus we do not explicitly report them here. Because the three diagonal elements of the HWP Mueller matrix effectively determine the HWP co-pol/cross-pol transmission and modulation efficiency, this analysis confirms that these quantities are very weakly dependent on the spectral index of the input source; these findings are in very good agreement with those of @Savini2009. We will see in Section \[sec:map\_maker\] how $\overline{a}_{00}$, $\overline{a}_{11}$, and $\overline{a}_{22}$ can be incorporated in the map-making algorithm in terms of optical efficiency, $\eta$, and polarisation efficiency, $\varepsilon$, of each detector.
Band 250$\mu$m 350$\mu$m 500$\mu$m
--------------------- ---------------------- ---------------------- ----------------------
$\overline{a}_{00}$ 0.905 $\pm$ 0.006 1.001 $\pm$ 0.006 1.008 $\pm$ 0.007
$\overline{a}_{01}$ 0.012 $\pm$ 0.010 0.017 $\pm$ 0.010 0.014 $\pm$ 0.011
$\overline{a}_{02}$ $-$0.002 $\pm$ 0.008 0.006 $\pm$ 0.009 0.001 $\pm$ 0.009
$\overline{a}_{10}$ $-$0.016 $\pm$ 0.010 $-$0.021 $\pm$ 0.010 $-$0.020 $\pm$ 0.011
$\overline{a}_{11}$ 0.806 $\pm$ 0.011 0.928 $\pm$ 0.010 0.935 $\pm$ 0.012
$\overline{a}_{12}$ $-$0.007 $\pm$ 0.011 $-$0.009 $\pm$ 0.014 $-$0.011 $\pm$ 0.014
$\overline{a}_{20}$ $-$0.008 $\pm$ 0.008 $-$0.022 $\pm$ 0.010 $-$0.021 $\pm$ 0.010
$\overline{a}_{21}$ $-$0.007 $\pm$ 0.011 $-$0.009 $\pm$ 0.014 $-$0.011 $\pm$ 0.014
$\overline{a}_{22}$ $-$0.808 $\pm$ 0.008 $-$0.960 $\pm$ 0.009 $-$0.979 $\pm$ 0.010
: Band-averaged Mueller matrix coefficients. These values are relative to Fig. \[fig:spare\_HWP\_MM\_coeff\_cold\]. The input source is assumed to have a flat spectrum in frequency.[]{data-label="tab:spare_HWP_MM_coeff_cold"}
Effective HWP phase shift {#sec:phase_shift}
-------------------------
Finally, we discuss a potential limitation to any linear polarisation modulator, i.e. the leakage between axes. In a HWP, the phase shift between the two axes should be as close to 180$^{\circ}$ as possible to avoid transforming linear into elliptical polarisation, hence losing modulation efficiency. The phase can not be directly measured in a pFTS, but it can be indirectly inferred from the HWP Mueller matrix.
In order to recover the wavelength-dependent phase shift of the HWP, we recall the Mueller matrix of a non-ideal impedance-matched single birefringent slab [@Savini2009 at $\theta=0^{\circ}$]: $$\label{eq:mueller_HWP_11}
{\mathbfss M_{\rm
slab}}\left(\theta=0^{\circ},\Delta\varphi\right)=\frac{1}{2}\times~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$$ $$\nonumber\times\begin{pmatrix}
\tau_{\|}^2+\tau_{\bot}^2 & \tau_{\|}^2-\tau_{\bot}^2 & 0 & 0 \\
\tau_{\|}^2-\tau_{\bot}^2 & \tau_{\|}^2+\tau_{\bot}^2 & 0 & 0 \\
0 & 0 & 2\,\tau_{\|}\,\tau_{\bot}\cos\Delta\varphi & 2\,\tau_{\|}\,\tau_{\bot}\sin\Delta\varphi \\
0 & 0 & -2\,\tau_{\|}\,\tau_{\bot}\sin\Delta\varphi & 2\,\tau_{\|}\,\tau_{\bot}\cos\Delta\varphi \\
\end{pmatrix}~,~~~~~~~~~~~~~~~~~$$ where $\tau_{\|}^2$ and $\tau_{\bot}^2$ are the measured transmissions of orthogonal polarisations aligned with the birefringent axes. By comparing the matrix in Equation \[eq:mueller\_HWP\_11\] with that of a generic HWP, we can solve for the HWP phase shift as $$\label{eq:mueller_HWP_12}
\cos\Delta\varphi =\frac{a_{22}}{2}
\left({\frac{a_{00}+a_{01}}{2}}\right)^{-\frac{1}{2}}
\left({\frac{a_{00}-a_{01}}{2}}\right)^{-\frac{1}{2}}~.$$
Equation \[eq:mueller\_HWP\_12\] allows us to recover the phase shift from our knowledge of $a_{00}$, $a_{01}$ and $a_{22}$. Fig. \[fig:phase\_spareHWP\] shows the estimated phase shift of the BLASTPol HWP as a function of wavenumber, after the introduction in our Monte Carlo routine of the wavelength-dependent position of the HWP equivalent axes depicted in Fig. \[fig:HWP\_offset\_vs\_freq\].
![Phase shift of the BLASTPol HWP as a function of wavenumber. The purple line shows the design goal and the orange line shows the [*as-built*]{} phase shift, estimated using the physical model of @Savini2006 [see Section \[sec:physical\_model\]], which accounts for the depolarisation effects due to standing waves between sapphire substrates by estimating the circular polarisation ($V$) portion of the HWP Mueller matrix. On the other hand, our empirical model (in black, with 3$\sigma$ error bars in yellow) underestimates the phase shift because (in the absence of the $V$ terms) we have assumed a pure cosine modulation, which is not adequate in a multi-slab Pancharatnam stack. Also shown for reference are the ideal phase shift for a HWP (180$^{\circ}$) and the relative spectral response of the three BLASTPol channels, in arbitrary units.[]{data-label="fig:phase_spareHWP"}](phase_spareHWP_Giorgio.pdf){width="1.\linewidth"}
Clearly, the $\beta_{\rm ea}$-corrected phase shift (in black) appreciably departs from 180$^{\circ}$. However, recall that Equation \[eq:mueller\_HWP\_12\] strictly applies only to a single birefringent plate. In a multi-slab Pancharatnam stack, $\Delta\varphi$ becomes an “effective” phase shift as we are no longer in the presence of a pure cosine modulation, thus slightly skewing the HWP modulation function and resulting in an artificially higher leakage between axes when the cosine function is inverted. Our empirical method, which does not estimate the circular polarisation ($V$) portion of the HWP Mueller matrix, is prone to underestimating the phase shift by mistaking the depolarisation effects due to standing waves between sapphire substrates (enhanced by the presence of the interspersed polyethylene layers) for a phase deficit. In fact, were the phase shift really $\lesssim$160$^{\circ}$, we would get a substantially higher cross-pol on the FTS measurements than the $\la$0.5% we measure in Figs. \[fig:BLAST\_pol\_spareHWP\_ARC\_xpol\_spectra\] and \[fig:modeff\]a.
While none of the other quantities estimated in the previous sections are affected by this deficiency in our model, we redress for the inadequacy of our empirical model in retrieving the phase shift by resorting to the physical model of @Savini2006 [discussed in Section \[sec:physical\_model\]]. The purple line in Fig. \[fig:phase\_spareHWP\] shows the design goal for the BLASTPol HWP, which is computed by assuming the nominal values for the HWP build parameters (see Table \[tab:physical\_parameters\_HWP\]), while the orange line shows the as-built phase shift, which is obtained by fitting the spectral data and allowing the build parameters to vary in a physical way around the nominal values.
The physical model is able to reconstruct the inevitable substrate alignment errors that occur during the HWP assembly and provides a much improved estimate of the actual phase shift, which is within $\sim$5$^{\circ}$ of the ideal value for the central BLASTPol band (350$\mu$m, where the HWP is optimized). Although the side bands show worse performance (within $\sim$15$^{\circ}$ of 180$^{\circ}$), we have indications that the modulation efficiency of the HWP at 4K is only mildly affected by this departure from ideality. From Fig. \[fig:modeff\]b we see that the extrapolated HWP modulation efficiency is always above 95% across the whole spectral range of interest, with band-integrated values exceeding 98%. Moreover, phase shift deviations of similar amplitude are measured in most (sub)mm-wave achromatic half-wave plates manufactured to date [e.g., @Savini2009; @Zhang2011].
In addition to the non-ideal substrate alignment, the physical model yields a best-fit thickness for the individual sapphire substrates that is 9.6$\mu$m smaller than the nominal value of 500$\mu$m available on the market[^9]. While it is beyond the scope of this paper to quantify exactly the relative contributions of plate misalignment and reduced thickness to the overall non-ideal behaviour of the phase shift, we can not rule out a contribution from both effects. Furthermore, and more importantly, neither of these two effects has nearly the same impact on the other performance parameters presented in the previous sections (the 9 Mueller matrix coefficients associated with linear polarisation) because of their weak dependence on the phase shift.
Incidentally, we verify that our methodology does not violate conservation of energy by ensuring that the output Stokes vector resulting from a generic polarised input travelling through the recovered HWP Mueller matrix satisfies $I^2 \geq Q^2 + U^2$ in every instance described above.
Map-making algorithm {#sec:map_maker}
====================
Map-making is the operation that generates an astronomical map, which contains in every pixel an estimate of the sky emission, and is obtained by combining data from all detectors available at a given wavelength channel, their noise properties and the pointing information. The raw BLAST(Pol) data consist of bolometer time-ordered streams, which are cleaned and pre-processed before being fed into the map-maker; the details are extensively described elsewhere [@Rex2007; @Truch2007; @Wiebe2008; @Pascale2008], and we refer to these works for a complete account of the low-level data reduction.
In the following, we focus on the mathematical formalism of the map-making technique, and its algorithmic implementation in the specific case of BLASTPol.
For a non-ideal polarisation experiment, by adopting the Stokes formalism and assuming that no circular ($V$) polarisation is present, we can model the data as $$\label{eq:timeline}
d^i_t = \frac{\eta^i}{2}A^i_{tp}\left[ I_p + \varepsilon^i \left(Q_p\,\cos2\gamma^i_t + U_p\,\sin 2\gamma^i_t \right)\right] + n^i_t~.$$ Here: $i$, $t$ and $p$ label detector index, time, and map pixel respectively; $d^i_t$ are the time-ordered data for a given channel, related to the sky maps $\left[I_p,Q_p,U_p\right]$ by the pointing operator $A^i_{tp}$; $\eta^i$ is the optical efficiency of each detector; $\varepsilon^i$ is the polarisation efficiency of each detector with its polarising grid (analyser); and $n^i_t$ represents a generic time-dependent noise term. Throughout this discussion it is assumed that the term within square brackets is the convolution of the sky emission with the telescope point-spread function (PSF). $\gamma^i_t$ is the time-ordered vector of the observed polarisation angle, defined as the angle between the polarisation reference vector at the sky pixel $p$ (in the chosen celestial frame) and the polarimeter transmission axis. $\gamma^i_t$ is given by $$\label{eq:polangle}
\gamma^i_t = \alpha_t^i + 2\left[\beta_t-\beta_0-\overline{\beta}_{\rm ea}\right] +\delta^i_{\rm grid}~,$$ where $\alpha_t^i$ is the angle between the reference vector at pixel $p$ and a vector pointing from $p$ to the zenith along a great circle, $\beta_t$ is the HWP orientation angle in the instrument frame, $\beta_0$ is the HWP zero angle in the instrument frame, $\overline{\beta}_{\rm ea}$ is the band-averaged position of the equivalent axes of the HWP (dependent on the known or assumed spectral signature of the input source; see Section \[sec:beta\_ea\]), and $\delta^i_{\rm grid}=[0,\pi/2]$ accounts for the transmission axis of the polarising grids being parallel/perpendicular to the zenith angle.
The notation outlined above can be connected to the Mueller formalism developed in Section \[sec:mueller\_HWP\] to determine under which circumstances Equation \[eq:timeline\] is valid in the presence of a real (i.e., non-ideal) HWP. Because we have included in Equation \[eq:polangle\] the band-averaged position of the equivalent axes of the HWP, $\overline{\beta}_{\rm ea}$, the Mueller matrix of the BLASTPol HWP can be considered almost that of an ideal HWP, as discussed in Section \[sec:beta\_ea\]. Nonetheless, we have shown that the band-averaged values of the three diagonal matrix coefficients are not identically unity (but always $>$0.8 in modulus), probably as a result of residual absorption from sapphire, especially in the 250 and 350$\mu$m bands (although we have corrected for it to the best of our knowledge). In the light of these considerations, we now want to compare Equation \[eq:timeline\] to Equation \[eq:mueller\_HWP\_4b\], which both represent the signal measured by a polarisation insensitive intensity detector when illuminated by a polarised input that propagates through a rotating HWP and an analyser. A term-by-term comparison shows that these two expressions are equivalent when the coefficients $B$ and $C$ (defined in Equation \[eq:mueller\_HWP\_5\]) are zero, i.e. when the HWP modulates the polarisation purely at four times the rotation angle, with no leakage in the second harmonic (twice the rotation angle) and thus no leakage of $I$ into $Q$ and $U$. These two coefficients are linear combinations of the HWP Mueller matrix elements $\overline{a}_{01}, \overline{a}_{10}, \overline{a}_{02},
\overline{a}_{20}$, which we have shown in Table \[tab:spare\_HWP\_MM\_coeff\_cold\] to be all compatible with zero within 2$\sigma$. In addition, their amplitude is at most $\sim$2% of that of the diagonal matrix elements, so to first order the coefficients $B$ and $C$ can be neglected, and the two expressions can be considered equivalent. Nonetheless, these generally moderate levels of $I\rightarrow Q,U$ leakage can be readily accounted for by incorporating in the map-making algorithm a correction for the “instrumental polarisation” [IP; see @Matthews2013 and Angilè et al. in preparation].
In addition, after some simple algebra, it can be shown that $\eta={a}_{00}+\frac{{a}_{11}}{2}+\frac{{a}_{22}}{2}$ and $\eta\,\varepsilon=\frac{{a}_{11}}{2}-\frac{{a}_{22}}{2}$. As anticipated in Section \[sec:beta\_ea\], the knowledge of the band-averaged values of the three diagonal matrix elements, $\overline{a}_{00}, \overline{a}_{11}, \overline{a}_{22}$ (which we have shown to depend weakly on the spectral index of the input source), can be readily incorporated in the map-making algorithm in terms of optical efficiency, $\eta$, and polarisation efficiency, $\varepsilon$, of the HWP; these can be factored into the overall optical and polarisation efficiency of each detector, which are the product of several factors (e.g., bolometer absorption efficiency, HWP efficiency, absorption in the optical chain, etc.). From the values listed in Table \[tab:spare\_HWP\_MM\_coeff\_cold\], in our case we find $[\eta_{\rm hwp},\varepsilon_{\rm hwp}]=[0.904,0.893],
[0.985,0.958]$, and $[0.986,0.971]$ at 250, 350, and 500$\mu$m, respectively.
Finally, the comparison of Equations \[eq:timeline\] and Equation \[eq:mueller\_HWP\_4b\] also yields the relation $\eta\,\varepsilon\,\chi=-{a}_{12}=-{a}_{21}$, where we have introduced a new parameter, $\chi$, which quantifies the amplitude of the mixing of $Q$ and $U$. From Table \[tab:spare\_HWP\_MM\_coeff\_cold\], we see that $\overline{a}_{12}=\overline{a}_{21}$ are always compatible with zero within 1$\sigma$, and their amplitude is at most $\sim$1% of that of the diagonal matrix elements. Nonetheless we quantify the amplitude of the $Q \leftrightarrow U$ mixing to be $\chi_{\rm
hwp}=0.009,0.010$, and 0.011 at 250, 350, and 500$\mu$m, respectively. While this correction is not currently included in our algorithm, we indicate that it can be implemented in a relatively straightforward way by modifying Equation \[eq:timeline\] with a double change of variable, i.e. $Q\rightarrow Q+\chi U$ and $U\rightarrow U+\chi Q$. If $\chi$ is estimated to the required accuracy, the unmixed $Q$ and $U$ can be retrieved unbiasedly. This correction may be very relevant to CMB polarisation experiments, where any $Q \leftrightarrow U$ leakage leads to a spurious mixing of the $EE$ and $BB$ modes.
We remind the reader that the above factors have been computed directly from the band-averaged coefficients of the inferred HWP Mueller matrix extrapolated to 4K, and offer a direct way to include the modelled HWP non-idealities in a map-making algorithm. On the other hand, the band-averaged HWP maximum transmission, polarisation efficiency and cross-pol quoted at the end of Section \[sec:HWP\_spectra\] are estimated directly from the spectra extrapolated to 4K, and are mostly informative from an experimental point of view rather than for data analysis purposes.
Consider now one map pixel $p$ that is observed in one band by $k$ detectors ($i=1,...,k$); let us define the generalised pointing matrix [A]{}$_{tp}$, which includes the trigonometric functions along with the efficiencies, $$\label{eq:generalized_pointing_matrix}
{\mathbfss A}_{tp} \equiv \frac{1}{2}
\begin{pmatrix}
\eta^1\,A^1_{tp} & \eta^1\,\varepsilon^1\,A^1_{tp}\,\cos2\gamma^1_t & \eta^1\,\varepsilon^1\,A^1_{tp}\,\sin2\gamma^1_t \\
\vdots & \vdots & \vdots \\
\eta^i\,A^i_{tp} & \eta^i\,\varepsilon^i\,A^i_{tp}\,\cos2\gamma^i_t & \eta^i\,\varepsilon^i\,A^i_{tp}\,\sin2\gamma^i_t \\
\vdots & \vdots & \vdots \\
\eta^k\,A^k_{tp} & \eta^k\,\varepsilon^k\,A^k_{tp}\,\cos2\gamma^k_t & \eta^k\,\varepsilon^k\,A^k_{tp}\,\sin2\gamma^k_t \\
\end{pmatrix}~,$$ and the map triplet ${\mathbfss S}_{p}$, along with the combined detector (${\mathbfit D}_t$) and noise (${\mathbfit n}_t$) timelines, $$\label{eq:map_triplet}
{\mathbfss S}_{p} \equiv
\begin{pmatrix}
I_p\\
Q_p\\
U_p\\
\end{pmatrix}~,~~~~
{\mathbfit D}_t \equiv
\begin{pmatrix}
d^1_t\\
\vdots\\
d^i_t\\
\vdots\\
d^k_t\\
\end{pmatrix}~,~~~~
{\mathbfit n}_t \equiv
\begin{pmatrix}
n^1_t\\
\vdots\\
n^i_t\\
\vdots\\
n^k_t\\
\end{pmatrix}~.$$ Equation \[eq:timeline\] can then be rewritten in compact form as $$\label{eq:timeline_matrix}
{\mathbfit D}_t = {\mathbfss A}_{tp}\,{\mathbfss S}_{p} + {\mathbfit n}_t~.$$
Under the assumption that the noise is Gaussian and stationary, the likelihood of [S]{}$_{p}$ given the data can be maximised, thus yielding the well-known generalised least squares (GLS) estimator for [S]{}$_{p}$: $$\label{eq:GLS}
{\tilde{\mathbfss S}}_{p} = \left({\mathbfss A}_{tp}^{\rm T}\,{\mathbfss N}^{-1}\,{\mathbfss A}_{tp}\right)^{-1}\,{\mathbfss A}_{tp}^{\rm T}\,{\mathbfss
N}^{-1}\,{\mathbfit D}_t~,$$ where ${\mathbfss N} \equiv \langle{\mathbfit n}_t\,{\mathbfit
n}_{t'}\rangle$ is the noise covariance matrix of the data in the time domain, with $t,t'$ running over the detector time samples (typically $N_{\rm s}\sim10^6$–10$^7$).
Computation of the solution to Equation \[eq:GLS\] is far from trivial in most astronomical applications, due to [N]{} being a very large matrix, of size $k N_{\rm s} \times k N_{\rm
s}$. Understandably, it is computationally challenging to invert this matrix, especially when there are correlations among detectors, and a number of “optimal” map-making techniques have been developed in the literature to tackle this problem [e.g., @Natoli2001; @Natoli2009; @Masi2006; @Johnson2007; @Wu2007; @Patanchon2008; @Cantalupo2010].
Naive binning
-------------
In the simple case that the noise is uncorrelated between different detectors, the matrix [N]{} reduces to block diagonal: $$\label{eq:noise_covariance_diag}
\langle n^i_t\,n^j_{t'}\rangle = \langle n^j_t\,n^i_{t'}\rangle = 0~~~~~(i\neq j)~.$$ In addition, let us assume that there is no correlation between noise of different samples acquired by the same detector, or, in other words, that the noise in each detector is white. In this case, each “block” of the noise covariance matrix collapses into one value, which is the timeline variance for each detector. Hence, [N]{} becomes a $k \times k$ diagonal matrix where the diagonal elements are the sample variances of the detectors, $\sigma^2_i$, and weights can thus be defined as the inverse of those variances, $w^i \equiv 1/\sigma^2_i$.
Therefore, under the assumption that the noise is white and uncorrelated among detectors, Equation reduces to a simple, weighted binning [“naive” binning; see also @Benoit2004; @Pascale2011] of the map: $$\label{eq:naive_binning}
{\mathbfss S}_{p} =
\begin{pmatrix}
I_p\\
Q_p\\
U_p\\
\end{pmatrix}
= \frac{\sum\limits_{i=1}^{k} \sum\limits_{t=1}^{N_{\rm s}} w^i
\frac{({\mathbfss A}^i_{tp})^{\rm T}\,d^i_t}{({\mathbfss
A}^i_{tp})^{\rm T}\,{\mathbfss A}^i_{tp}}}{\sum\limits_{i=1}^{k}
w^i}~.$$
In the light of these considerations, let us go back to Equation and model the generic time-dependent noise term as $n^i_t = u_t + \xi^i \rho_t$ , where $u_t$ represents a time-dependent noise term, completely uncorrelated among different detectors, while $\rho_t$ describes the correlated noise (varying over timescales larger than the ratio of the size of the detector array to the scan speed), coupled to each detector via the $\xi^i$ parameter, peculiar to each bolometer.
Let us define the following quantity for every pixel $p$ in the map: $$\label{eq:linear_system1}
{\mathbfss S}_{p}^{\rm e} =
\begin{pmatrix}
I_p^{\rm e}\\
Q_p^{\rm e}\\
U_p^{\rm e}\\
\end{pmatrix}
\equiv
\begin{pmatrix}
\sum\limits_{i=1}^{k} \sum\limits_{t=1}^{N_{\rm s}}d^i_t \\
\sum\limits_{i=1}^{k} \sum\limits_{t=1}^{N_{\rm s}}d^i_t\,\cos2\gamma^i_t\\
\sum\limits_{i=1}^{k} \sum\limits_{t=1}^{N_{\rm s}}d^i_t\,\sin2\gamma^i_t \\
\end{pmatrix}~,$$ where $N_{\rm s}$ is now the number of samples in each detector timeline that fall within pixel $p$, and the superscript “e” stands for “estimated”. The above quantities can be computed directly from the detector timelines. Recalling Equation , we can outline the following linear system of $3$ equations with $3$ unknowns: $$\begin{aligned}
\label{eq:linear_system2}
\nonumber\begin{pmatrix}
I_p^{\rm e}\\
Q_p^{\rm e}\\
U_p^{\rm e}\\
\end{pmatrix}
=
\begin{pmatrix}
\sum\limits_{i,t} 1 & \sum\limits_{i,t} \cos2\gamma^i_t & \sum\limits_{i,t} \sin2\gamma^i_t \\
\sum\limits_{i,t} \cos2\gamma^i_t & \sum\limits_{i,t} \cos^2 2\gamma^i_t & \sum\limits_{i,t} \cos2\gamma^i_t\,\sin2\gamma^i_t \\
\sum\limits_{i,t} \sin2\gamma^i_t & \sum\limits_{i,t} \cos2\gamma^i_t\,\sin2\gamma^i_t & \sum\limits_{i,t} \sin^2 2\gamma^i_t \\
\end{pmatrix}\\
\cdot \frac{1}{2}
\begin{pmatrix}
I_p\\
Q_p\\
U_p\\
\end{pmatrix}
+
\begin{pmatrix}
\sum\limits_{i,t} (u_t + \xi^i \rho_t)\\
\sum\limits_{i,t} (u_t + \xi^i \rho_t)\,\cos2\gamma^i_t\\
\sum\limits_{i,t} (u_t + \xi^i \rho_t)\,\sin2\gamma^i_t\\
\end{pmatrix}~,~~~~~~~~~~~~~~~~~~~\end{aligned}$$ where we have temporarily assumed $\eta^i=\varepsilon^i=w^i=1$ and combined the two sums in one, with the indices $i$ and $t$ running, respectively, over the bolometers and the samples in each detector timeline.
If we now define the quantities, $$\begin{aligned}
\label{eq:linear_system3}
\nonumber N_{\rm hit}\equiv \sum_{i,t} \frac{1}{2},~~~~~~c
\equiv \sum_{i,t} \frac{1}{2} \cos2\gamma^i_t,\\
\nonumber c_2 \equiv\sum_{i,t} \frac{1}{2} \cos^2
2\gamma^i_t,~~~~~~
s \equiv \sum_{i,t} \frac{1}{2} \sin2\gamma^i_t,\\
\nonumber s_2\equiv\sum_{i,t} \frac{1}{2} \sin^2 2\gamma^i_t =
N_{\rm hit}-c_2,~~~~~~U\equiv\sum_{i,t} u_t,\\
\nonumber m\equiv\sum_{i,t} \frac{1}{2}
\cos2\gamma^i_t\,\sin2\gamma^i_t
\nonumber,~~~~C^u_2\equiv\sum_{i,t} u_t\,\cos2\gamma^i_t\\
\nonumber S^u_2\equiv\sum_{i,t} u_t\,\sin2\gamma^i_t,~~~~~~
P\equiv\sum_{i,t} \xi^i \rho_t,\\
C^{\rho}_2\equiv\sum_{i,t} \xi^i
\rho_t\,\cos2\gamma^i_t,~~~~~~S^{\rho}_2\equiv\sum_{i,t} \xi^i
\rho_t\,\sin2\gamma^i_t,\end{aligned}$$ then the system in Equation can be rewritten in compact form as $$\label{eq:linear_system4}
\begin{pmatrix}
I_p^{\rm e}\\
Q_p^{\rm e}\\
U_p^{\rm e}\\
\end{pmatrix}
=
\begin{pmatrix}
N_{\rm hit} & c & s \\
c & c_2 & m \\
s & m & N_{\rm hit}-c_2 \\
\end{pmatrix}
\cdot
\begin{pmatrix}
I_p\\
Q_p\\
U_p\\
\end{pmatrix}
+
\begin{pmatrix}
U+P\\
C^u_2 + C^{\rho}_2\\
S^u_2 + S^{\rho}_2\\
\end{pmatrix}~.$$
In order to retrieve an estimate of ${\mathbfss S}_{p}$ from the quantities computed in Equation , the above system has to be solved for every pixel $p$ in the map. One can already see the computational advantage of inverting a $3\times3$ matrix $N_{\rm pix}\times N_{\rm pix}$ times, with respect the inversion of a generic $k N_{\rm s} \times k N_{\rm
s}$ matrix [for detectors having uncorrelated $1/f$ noise as well as a common-mode $1/f$ noise; @Patanchon2008], or $k$ matrices of size $N_{\rm s} \times N_{\rm s}$ [for detectors having only uncorrelated $1/f$ noise; @Cantalupo2010]. The main difficulties are, of course, in estimating the noise terms $U,P,C^u_2,C^{\rho}_2,S^u_2,S^{\rho}_2$. However, recalling Equation and the fact that adjacent detectors have orthogonal polarising grids ($\delta^i_{\rm
grid}=[0,\pi/2]$), we note that, in the sum over $i$, adjacent detectors have equal and opposite contributions to $C^{\rho}_2$ and $S^{\rho}_2$, under the following assumptions: (i) the timescale over which the correlated noise is approximately constant is larger than the time elapsed while scanning the same patch of sky with two adjacent detectors, and (ii) $\xi^i$ is not too dissimilar between adjacent bolometers.
This means that the terms $C^{\rho}_2$ and $S^{\rho}_2$ can be neglected, under the above assumptions, while estimating the $[Q,U]$ maps. In particular, as a first step, we can solve for $I$ only by high-pass filtering the timelines, in order to suppress the correlated noise term in $I$, $P$. Subsequently, $I$ can be assumed known, and the $[Q,U]$ maps can be computed without filtering the timelines, so that polarised signal at large angular scales is not suppressed.
The other assumption required for the naive binning is that the noise is white, at least on the timescales relevant to BLASTPol’s scan strategy. An analysis of the bolometer timelines from the 2010 campaign shows that the knee of the $1/f$ noise in the difference between two adjacent detectors is typically located at frequencies $\lesssim$0.1Hz; assuming a typical scan speed of 0.1$^{\circ}$s$^{-1}$, this corresponds to angular scales of $\gtrsim$1$^{\circ}$ on the sky. The regions mapped by BLASTPol hardly exceed 1$^{\circ}$ in size [see Table 1.1 in @Moncelsi2011b and Angilè et al. in preparation], hence here we stipulate that the noise in the difference between pairs of adjacent detectors is white.
Therefore, under the assumptions above, we can solve the linear system outlined in Equation ; by defining the following quantities, $$\begin{aligned}
\label{eq:linear_system5}
\nonumber \Delta & \equiv & c^2\left(c_2-N_{\rm h}\right)-N_{\rm h}\left(c_2^2+m^2-c_2\,N_{\rm h}\right)+2\,c\,s\,m-c_2\,s^2~, \\
\nonumber A &\equiv&-\left(c_2^2+m^2-c_2\,N_{\rm h}\right),~~~~B\equiv c\left(c_2-N_{\rm h}\right)+s\,m~,\\
C &\equiv& c\,m-s\,c_2 ,~~~~D\equiv -\left[\left(c_2-N_{\rm h}\right)N_{\rm h}+s^2\right]~,\\
\nonumber E &\equiv& c\,s-m\,N_{\rm h},~~~~F\equiv c_2\,N_{\rm
h}-c^2~,\end{aligned}$$ the solution to the system can be written in compact form: $$\label{eq:linear_system6}
{\mathbfss S}_{p} =
\begin{pmatrix}
I_p\\
Q_p\\
U_p\\
\end{pmatrix}
=
\begin{pmatrix}
\frac{A\,I^{\rm e}_p+B\,Q^{\rm e}_p+C\,U^{\rm e}_p}{\Delta}\\
\frac{B\,I^{\rm e}_p+D\,Q^{\rm e}_p+E\,U^{\rm e}_p}{\Delta}\\
\frac{C\,I^{\rm e}_p+E\,Q^{\rm e}_p+F\,U^{\rm e}_p}{\Delta}\\
\end{pmatrix}~,$$ where we have renamed $N_{\rm hit} \rightarrow N_{\rm h}$ for brevity.
Weights and uncertainties
-------------------------
The solution for ${\mathbfss S}_{p}$ given in Equation is a simple, unweighted binning of the data into the map pixels. In reality, as anticipated in Equation , we want to perform a weighted binning, where the weight of each detector is given by the inverse of its timeline variance, which can be easily measured as the bolometer’s white noise floor level. In our formalism, the weighted binning is simply achieved by defining $[I^{\rm
e}_p,Q^{\rm e}_p,U^{\rm e}_p]$ in Equation , as well as each of the quantities $N_{\rm h}$, $c$, $s$, $c_2$, $s_2$, and $m$ introduced in Equation , to include $w^i$ in the sums. Similarly, the measured values of the optical efficiencies $\eta^i$ and polarisation efficiencies $\varepsilon^i$ can readily be inserted in Equation to account for the non-idealities of the optical system.
The introduction of the weights allows us to derive the expression for the statistical error on ${\mathbfss S}_{p}$, in the continued assumption of uncorrelated noise, following the usual error propagation formula [e.g., @NumericalRecipes here we omit the sum over $t$ for simplicity]: $$\label{eq:map_maker_uncert1}
\sigma^2_p = \sum_{i} \frac{1}{w^i}\left(\frac{\partial {\mathbfss
S}_{p}}{\partial d^i}\right)^2~.$$ After some tedious algebra, the expression for the statistical error is $$\begin{aligned}
\label{eq:map_maker_uncert2}
\nonumber\sigma^2_{p} =
\begin{pmatrix}
{\rm Var}^I_p\\
{\rm Var}^Q_p\\
{\rm Var}^U_p\\
\end{pmatrix}
=~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\
\nonumber
\begin{pmatrix}
\frac{2}{\Delta^2}\left(A^2\,N_{\rm h}+B^2\,c_2+C^2\,s_2+2\,A\,B\,c+2\,A\,C\,s+2\,B\,C\,m\right)\\
\frac{2}{\Delta^2}\left(B^2\,N_{\rm h}+D^2\,c_2+E^2\,s_2+2\,B\,D\,c+2\,B\,E\,s+2\,D\,E\,m\right)\\
\frac{2}{\Delta^2}\left(C^2\,N_{\rm h}+E^2\,c_2+F^2\,s_2+2\,C\,E\,c+2\,C\,F\,s+2\,E\,F\,m\right)\\
\end{pmatrix},\end{aligned}$$ where $s_2\equiv N_{\rm h}-c_2$, as noted in Equation . To first order, these expressions can be used to quantify the uncertainty of $[I,Q,U]$ in each map pixel $p$. A more comprehensive account of the correlations in the noise, as well as a thorough validation of the assumptions made here, is beyond the scope of this paper and will be treated in a future work.
Preliminary map {#sec:map}
---------------
As a proof of concept of the naive binning technique for the BLASTPol polarised map-maker, which includes all the corrections due to HWP non-idealities discussed in this work, we present a *preliminary* map at 500$\mu$m of one of the giant molecular clouds observed by BLASTPol, the Carina Nebula. In this specific case, we calculate $\overline{\beta}^{\rm 500\,\mu
m}_{\rm ea} = 1.7$ based on a dust emissivity spectral index of 1.37 [@Salatino2012] for a modified blackbody spectral energy distribution.
The map in Fig. \[fig:CarinaNeb\_500um\_pol\_map\] is presented as contour levels of the intensity map $I$, on which we superimpose pseudo-vectors indicating the inferred magnetic field direction on the sky [assumed to be perpendicular to the measured polarisation direction; see @Lazarian2007]. The sky polarisation angle is given by $\phi=\frac{1}{2}\arctan\frac{U}{Q}$. Because the absolute flux calibration has not been finalised yet, we choose not to report here the intensity values corresponding to each contour level. This map should not be considered of any scientific value as it is not calibrated in flux. Nonetheless, we find a remarkable agreement between the BLASTPol polarisation directions at 500$\mu$m and those produced by the Submillimeter Polarimeter for Antarctic Remote Observations [SPARO; @Novak2003] at 450$\mu$m, which are shown in Fig. 1 of @Li2006 and whose pseudo-vectors are reported in Fig. \[fig:CarinaNeb\_500um\_pol\_map\] for comparison. Here the original BLASTPol $\left[I,Q,U\right]$ maps have been smoothed with a kernel of 4 (FWHM) for a more direct comparison with the SPARO data.
![Preliminary BLASTPol map at 500$\mu$m of the Carina Nebula, a giant molecular cloud approximately centred at coordinates $[10^{\rm h} 42^{\rm m} 35^{\rm s}, -59^{\circ}
42\arcmin 15\arcsec]$. The intensity contour levels are shown in the background. The pseudo-vectors indicate the inferred magnetic field direction: the BLASTPol measurements (red) are in excellent agreement with those from the SPARO polarimeter at 450$\mu$m [black; @Li2006]. This map should not be considered of any scientific value as no absolute calibration has been applied to the intensity contours; the map is only shown as a proof of concept for the map-maker, which includes all the corrections due to HWP non-idealities discussed in this work. Here we adopt $\overline{\beta}^{\rm 500\,\mu m}_{\rm ea} = 1.7$, calculated assuming a spectral index for the dust emissivity of 1.37 [@Salatino2012].[]{data-label="fig:CarinaNeb_500um_pol_map"}](CarinaNeb_500um_P36_respoff_pol_map_clockwise.pdf){width="0.86\linewidth"}
Concluding Remarks {#sec:concl_HWP}
==================
The goal of the first part of this work was to identify and measure the parameters that fully characterise the spectral performance of the linear polarisation modulator integrated in the BLASTPol instrument, a cryogenic achromatic HWP.
We have described in detail the design and manufacturing process of a five-slab sapphire HWP, which is, to our knowledge, the most achromatic built to date at (sub)mm wavelengths. In the same context, we have provided a useful collection of spectral data from the literature for the sapphire absorption coefficient at cryogenic temperatures.
Using a polarising FTS, we have fully characterised the spectral response of the anti-reflection coated BLASTPol HWP at room temperature and at 120K; we have acquired data cubes by measuring spectra while rotating the HWP to produce the polarisation modulation.
The cold dataset contains measurements in both co-pol and cross-pol configurations; we have used these two data cubes to estimate 9 out of 16 elements of the HWP Mueller matrix as a function of frequency. We have developed an ad-hoc Monte Carlo algorithm that returns for every frequency the best estimate of each matrix element and the associated error, which is a combination of the uncertainty on the measured spectra and a random jitter on the rotation angle.
We have measured how the position of the equivalent axes of the HWP, $\beta_{\rm ea}$, changes as a function of frequency, an effect that is inherent to any achromatic design. Once this dependence is accounted for in the Monte Carlo, and a correction is implemented for the residual absorption from sapphire, the Mueller matrix of the HWP approaches that of an ideal HWP, at all wavelengths of interest. In particular, the (band-averaged) off-diagonal elements are always consistent with zero within 2$\sigma$ and the modulus of the three diagonal coefficients is always $>$0.8. Therefore, we have introduced in the BLASTPol map-making algorithm the band-integrated values of $\beta_{\rm
ea}$ as an additional parameter in the evaluation of the polarisation angle. To first order, this approach allows us to account for most of the non-idealities in the HWP.
We have investigated the impact of input sources with different spectral signatures on $\beta_{\rm ea}$ and on the HWP Mueller matrix coefficients. We find that the HWP transmission and modulation efficiency are very weakly dependent on the spectral index of the input source, whereas the position of the equivalent axes of the sapphire plate stack is more significantly affected. This latter dependence, if neglected, may lead to an arbitrary rotation of the retrieved polarisation angle on the sky of magnitude $2\,\overline{\beta}_{\rm ea} =10$–15$^{\circ}$ (3–5$^{\circ}$) at 250 (500)$\mu$m. The 350$\mu$m band, however, is minimally perturbed by this effect.
In principle, the measured Mueller matrix can be used to generate a synthetic time-ordered template of the polarisation modulation produced by the HWP as if it were continuously rotated at a mechanical frequency $f=\omega\,t$. Continuous rotation of the HWP allows the rejection of all the noise components modulated at harmonics different than 4$f$ (synchronous demodulation) and is typically employed by experiments optimised to measure the polarisation of the CMB [e.g., @Johnson2007; @Reichborn2010]. In such experiments, the HWP modulation curve leaves a definite synchronous imprint on the time-ordered bolometer data streams, hence it is of utter importance to characterise the template and remove it from the raw data. However, a time-ordered HWP template would be of no use to a step-and-integrate experiment such as BLASTPol, whose timelines are not dominated by the HWP synchronous signal.
We have measured the phase shift of the HWP across the wavelength range of interest to be within 5$^{\circ}$ of the ideal 180$^{\circ}$ for the central BLASTPol band, and within 15$^{\circ}$ for the side bands. This is due to a combination of alignment errors of the sapphire substrates, which are hard to avoid in the manufacture of a five-slab stack, and their lower than ideal thickness. However, the modulation efficiency of the HWP is only mildly affected by this departure from ideality, being above 98% in all three BLASTPol bands. Moreover, departures of similar amplitude are not uncommon for HWPs at (sub)mm wavelengths.
The goal of the second part of this work was to include the measured non-idealities of the HWP as-built in a map-making algorithm. We have focused on the implementation of a naive binning technique for the case of BLASTPol, under the assumption of white and uncorrelated noise. As a proof of concept, we have presented a preliminary polarisation map for one of the scientific targets observed by BLASTPol during its first Antarctic flight, completed in January 2011. The inferred direction for the local magnetic field in the Carina Nebula star-forming region is in excellent agreement with the results obtained by @Li2006 with the SPARO instrument. The empirical approach presented in this paper will help improve the accuracy on astronomical measurements of the polarisation angle on the sky at submm wavelengths.
Acknowledgments {#acknowledgments .unnumbered}
===============
The BLASTPol collaboration acknowledges the support of NASA through grant numbers NAG5-12785, NAG5-13301 and NNGO-6GI11G, the Canadian Space Agency (CSA), the Science and Technology Facilities Council (STFC), Canada’s Natural Sciences and Engineering Research Council (NSERC), the Canada Foundation for Innovation, the Ontario Innovation Trust, the Puerto Rico Space Grant Consortium, the Fondo Istitucional para la Investigacion of the University of Puerto Rico, the National Science Foundation Office of Polar Programs, and the Canadian Institute for Advanced Research.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: We adopt the Stokes [-@Stokes1852] formalism to represent the time-averaged polarisation state of electromagnetic radiation; for a review of polarisation basics we refer the reader to @Collett1993.
[^3]: We distinguish between “optic” axis of a crystal, i.e. the direction in which a ray of transmitted light experiences no birefringence, and “optical” axis, i.e. the imaginary line along which there is some degree of rotational symmetry in the optical system.
[^4]: The HWP designed for the mm-wave E and B EXperiment [EBEX; @Matsumura2009; @Reichborn2010] is nominally slightly more achromatic, with a $\sim$110% bandwidth. However, a comprehensive spectral characterisation of the final (as-flown) AR-coated HWP assembly has yet to be published.
[^5]: Room-temperature spectra were acquired and can be made available to the reader, but are neither reported here nor used in the subsequent analysis, because plagued by significant in-band transmission loss due to the absorption from sapphire.
[^6]: The notation P[*n*]{} refers to a wire grid polariser that has a grid period of [*n*]{}\[$\mu$m\], with [*n*]{}/2 copper strips and [*n*]{}/2 gaps, photolithographed on a 1.5$\mu$m mylar substrate.
[^7]: We note that this definition of cross-pol may differ from other conventions adopted in the literature (e.g., that of @Masi2006, who operate without a HWP).
[^8]: The analytical relations apply, strictly speaking, at 80K and for $k\lesssim33$cm$^{-1}$, thus we corroborate them at higher frequencies with the data points, which apply at $\lesssim$60K. It is evident that the sapphire absorption coefficient has a very weak dependence on temperature below 80K, and in particular data points collected at 1.5K are in good enough agreement (within 2% on the resulting absorption for $d=2.5$mm) with those collected at higher temperatures (up to 80K). Therefore we can safely claim that for our application the two analytical relations shown in Figure \[fig:sapphire\_absorption\_coefficient\_cold\] are a good representation of the sapphire absorption at 4K.
[^9]: We verify that the sum of the thicknesses from Table \[tab:physical\_parameters\_HWP\] is compatible with the overall measured thickness of the HWP stack.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Presently available high-energy proton beams in circular accelerators carry enough momentum to accelerate high-intensity electron and positron beams to the TeV energy scale over several hundred meters of the plasma with a density of about $10^{15}~\mathrm{cm}^{-3}$. However, the plasma wavelength at this density is 100-1000 times shorter than the typical longitudinal size of the high-energy proton beam. Therefore the self-modulation instability (SMI) of a long ($\sim$10 cm) proton beam in the plasma should be used to create the train of micro-bunches which would then drive the plasma wake resonantly. Changing the plasma density profile offers a simple way to control the development of the SMI and the acceleration of particles during this process. We present simulations of the possible use of a plasma density gradient as a way to control the acceleration of the electron beam during the development of the SMI of a 400 GeV proton beam in a 10 m long plasma. This work is done in the context of the AWAKE project — the proof-of-principle experiment on proton driven plasma wakefield acceleration at CERN.'
address:
- 'CERN, CH-1211 Geneva 23, Switzerland'
- 'Budker Institute of Nuclear Physics SB RAS, 630090 Novosibirsk, Russia'
- 'Novosibirsk State University, 630090 Novosibirsk, Russia'
author:
- 'A. Petrenko'
- 'K. Lotov'
- 'A. Sosedkin'
title: 'Numerical Studies of Electron Acceleration Behind Self-Modulating Proton Beam in Plasma with a Density Gradient'
---
AWAKE experiment ,Self-Modulation Instability ,CERN ,Proton-Driven Plasma Wakefield Acceleration
Introduction
============
Proton beams in the modern high-energy accelerators carry a large amount of energy. For example, the design parameters of the Large Hadron Collider correspond to the total proton beam energy of 360 MJ stored in the 2800 bunches at the top energy of 7 TeV. This is comparable to the kinetic energy of a typical fully loaded airliner (80 t) at the take-off speed of 300 km/h (280 MJ). Even the single bunch ($3 \cdot 10^{11}$ protons at 400 GeV) in the 40 year old Super Proton Synchrotron (SPS) at CERN carries an order of magnitude more energy than the single bunch in the proposed International Linear Collider ($2 \cdot 10^{10}$ electrons or positrons at 250 GeV). This makes it possible to accelerate a substantial amount of particles to a TeV-scale energy in a single accelerating stage using a beam-driven scheme powered by a relativistic proton beam. Several such techniques were suggested earlier as a way of transferring the proton beam energy to other particles (electrons/positrons, muons, and pions) [@proton_klystron; @coherent_acceleration] which cannot be accelerated in a circular machine either because of the prohibitively high energy loss due to the synchrotron radiation (electrons/positrons) or because of the short life-time of the unstable particles. The proton-driven plasma wakefield acceleration (PDPWFA) [@PDPWFA_nature] is a recently proposed method promising a GeV/m rate of acceleration to a TeV-scale energy in a single plasma stage. Plasma wakefields with GV/m amplitude correspond to the plasma wavelength around 1 mm, however the typical proton beam bunches in a storage ring have the longitudinal size in the order of 10 cm. In the original PDPWFA proposal [@PDPWFA_nature] the proton beam was assumed to be compressed by a factor of 1000 down to the longitudinal size of 0.1 mm. Such an extreme bunch compression is very challenging (although technically feasible), but the similarly high plasma wakefields can be excited resonantly with a sequence of sub-millimeter long microbunches produced from the long proton beam as a result of the beam self-modulation instability (SMI) [@EPAC98-806; @SMI_PRL2010] in plasma. The self-modulated beam can drive the plasma wakefields for long distances [@density_step] even without the very strong external focusing needed in the case of the compressed proton bunch.
The AWAKE experiment [@AWAKE_PoP2014; @Edda_EAAC2015; @AWAKE_history] at CERN will be the first proof-of-principle demonstration of this technique using the self-modulation of the 400 GeV proton bunch from the SPS accelerator. A 10 m long plasma section based on the continuous flow of rubidium vapor at 200 $^{\circ}$C will be used. The selected plasma section design is already capable of creating plasma density profiles with a constant gradient along the whole 10 m long section [@Gennady_report; @AWAKE_history]. This gradient naturally appears if the continuous flow of rubidium vapor through the orifices at the beam entrance and exit is unbalanced.
In general, plasma density gradients are well known to be useful for the control of electron injection and acceleration in the laser- and beam-driven plasma wakefield acceleration [@LWFA_gradient; @Katsouleas_PRA_1986]. In this article we study the effect of a plasma density gradient on the acceleration of electrons in the plasma wake created by the self-modulating proton beam. The details of electron beam injection and acceleration in the uniform plasma were already given in [@on_axis_PoP2014]. More complicated longitudinal plasma density profiles were suggested earlier for the SMI-based PDPWFA [@Schroeder_PoP2012]. However we focus on the case of a constant plasma density gradient which is relevant for the AWAKE experiment. The process is studied numerically with the particle-in-cell version of 2d3v quasi-static code LCODE [@LCODE; @PRST-AB6-061301; @IPAC13-1238] assuming the cylindrical symmetry.
The AWAKE experiment configuration
==================================
The AWAKE experiment will be conducted at CERN in the deep underground CNGS facility [@Edda_EAAC2015]. An LHC-type proton bunch of 400 GeV/$c$ momentum but higher intensity ($3\cdot10^{11}$ protons/bunch) is extracted from the SPS and sent towards a 10 m long rubidium vapor cell. A high power (2 TW) laser pulse, co-propagating and co-axial with the proton beam, is used to ionize the (initially neutral) rubidium gas in the plasma cell and also to generate a seed for the proton bunch self-modulation. A several millimeter long bunch with $\sim10^{9}$ electrons at 10-20 MeV produced by the photo-injector serves as a witness beam and is accelerated in the wake of the proton bunches. Several diagnostics are installed downstream the plasma cell to measure the proton bunch self-modulation and the accelerated electron bunch properties. The geometry and the baseline parameters of the experiment are given in Fig. \[fig:layout\].
Effect of density gradient on the plasma wakefields
===================================================
The plasma wavelength is inversely proportional to the square root of the plasma density $n$: $$\lambda_p = \sqrt{\frac{\pi}{n r_e}},$$ where $r_e$ is the classical electron radius. Small change in the plasma density $\delta n$ results in the corresponding change in the plasma wavelength $$\lambda_p + \delta \lambda_p = \sqrt{\frac{\pi}{(n + \delta n) r_e}} \approx \lambda_p \left( 1 - \frac{\delta n}{2n} \right).$$ The phase shift of the wakefield corresponding to this plasma density change is $$\label{eq:Delta_psi}
\Delta \psi = N\delta \psi = 2\pi N \frac {\delta \lambda_p} {\lambda_p} = -\pi N \frac{\delta n} {n},$$ where $N$ is the number of plasma oscillations behind the ionizing laser pulse. Here for simplicity we consider free plasma oscillations driven by the seed pulse. Another cause of phase shifts is the evolution of the proton beam [@PoP22-103110]. During the SMI growth, the plasma wakefield phase and amplitude change even in the uniform plasma. In our case the typical phase shift due to the SMI is one plasma wavelength for 100 micro-bunches [@on_axis_PoP2014]. This shift accumulates over first 4 meters of the plasma. Only half of the wakefield period is decelerating for protons. Therefore, the number of proton beam micro-bunches driving the plasma wake resonantly in a non-uniform plasma can be estimated from $\Delta \psi \sim \pi$ and (\[eq:Delta\_psi\]) as $$N \sim n/\delta n.$$ Since the baseline AWAKE proton bunch has $\sigma_{zb}=12$ cm, the number of proton micro-bunches at the baseline density is around 100. We can expect that the typical plasma density gradients which can still support a significant wakefield amplitude are several percent over the entire length of the plasma section.
The simulated amplitude of the longitudinal electric field excited by the self-modulating proton beam in plasmas with different density profiles is shown in Fig. \[fig:Ez\_max\]. As expected, self-modulation of the proton beam with the baseline AWAKE parameters (Fig. \[fig:layout\]) produces wakefields with significant amplitudes for plasma density gradients up to several percent over 10 m (the 10% gradient reduces the maximum field by a factor of three). However, the process of electron acceleration is more sensitive to the density profile [@PoP20-013102], especially to the sign of the density gradient which defines the phase velocity of the plasma wakefield. Fig. \[fig:plus\_minus\_2\_percent\] shows the final accelerated electron energy as a function of the longitudinal position along the beam.
Density gradients offer a convenient way to control the phase of the plasma wakefield during the development of the self-modulation instability. In order to have a significant energy gain (> 1 GeV) in the uniform plasma, one has to inject electrons around 130 periods of plasma oscillations behind the laser pulse [@on_axis_PoP2014], while in the case of a positive density gradient it is possible to shift the optimum injection delay closer to the laser pulse. This effect can be seen in Fig. \[fig:plus\_minus\_2\_percent\].
The longitudinal motion of an individual electron as well as the wakefield phase and amplitude for positive and negative plasma density gradients are shown in Fig. \[fig:plus\_minus\_10\_percent\]. As one can see, the amplitude of the electric field is approximately the same in both cases, however, the electron gains energy only in the case of the positive density gradient.
The negative density gradient results in a gradual increase of the plasma wavelength and corresponding continuous drift of the wakefield phase towards the tail of the bunch (i. e., the phase velocity of the wakefield is slower than the speed of light). This limits the achievable energy, because, as soon as the injected electron gains some energy, it outruns the plasma wake and enters the decelerating (but still focusing) phase where it is decelerated. Such cycle of electron acceleration and deceleration can be repeated several times as one can see in the left side of Fig. \[fig:plus\_minus\_10\_percent\].
The positive density gradient makes the phase velocity of the plasma wake equal or slightly faster than the speed of light over a long part of the plasma section at some specific delay $\xi$ behind the laser pulse. Therefore it becomes possible for electrons to stay in phase with the wakefield and gain energy continuously until the end of the plasma section (right part of Fig. \[fig:plus\_minus\_10\_percent\]). The steeper the plasma density gradient, the closer the optimal electron injection delay is to the laser pulse.
In general, the acceleration of electrons is easier to achieve by injecting particles closer to the laser pulse. The less micro-bunches are used for acceleration, the more stable and reproducible this process becomes with respect to various perturbations like small scale plasma non-uniformities or beam parameter variations from shot to shot. With a proton beam longitudinally compressed by a factor of two, the 20% plasma density gradient (corresponding to the optimal injection at $\xi_e \approx -3$ cm) provides electron acceleration to more than 1 GeV (Fig. \[fig:compressed\]). In this case only 25 proton micro-bunches are required to accelerate electrons above 1 GeV. This is a factor of four reduction with respect to 100 micro-bunches necessary in the baseline case of uniform plasma and proton beam with $\sigma_{zb} = 12$ cm. Correspondingly, the requirement on small scale density uniformity can be four times relaxed.
Conclusions
===========
Using numerical simulation we have shown that the plasma density gradients offer a simple and efficient way to control the development of the proton beam self-modulation instability and acceleration of electrons during this process.
For the baseline parameters of the AWAKE experiment it is possible to accelerate electrons to a high energy (>1 GeV) in a plasma section with a density gradient of up to 10% along 10 m. Using small positive gradients (below 10% over 10 m) should improve the overall stability and reproducibility of electron acceleration experiment because it will rely on significantly lower number of micro-bunches to drive the plasma wakefield resonantly (optimum electron injection position shifts closer to the laser pulse). With longitudinally compressed proton bunches it is possible to operate at higher density gradients (20% with 2x compression) –— this can further improve the control over the acceleration process.
Even a small negative plasma density gradient ($-2$% over 10 m for example) dramatically reduces the achievable electron energy because in this case the phase velocity of the plasma wake is always slower than the speed of light.
Acknowledgements
================
The authors thank AWAKE collaboration for fruitful discussions. The contribution of Novosibirsk team to this work is supported by The Russian Science Foundation, grant No. 14-12-00043. Some of the simulations are performed at the Siberian Supercomputer Center SB RAS.
[00]{}
E. A. Perevedentsev, A. N. Skrinsky, On Possibility of Using Intense Beams of Large Proton Accelerators for Excitation of Linear Accelerating Structures. Proceedings of VI Nat. Accelerator Meeting, Dubna, 1978. Preprint INP 79-80, Novosibirsk, 1979.
V. E. Balakin, A. V. Novokhatsky, The Method of Accelerating Electrons with Maximum High Gradient by a Proton Beam. Preprint INP 79-86, Novosibirsk, 1979.
A. Caldwell, K. Lotov, A. Pukhov, and F. Simon, Proton-Driven Plasma-Wakefield Acceleration. Nature Phys. 5, 363 (2009).
K. V. Lotov, Instability of long driving beams in plasma wakefield accelerators. Proc. 6th European Particle Accelerator Conference (Stockholm, 1998), p.806-808.
N. Kumar, A. Pukhov and K. Lotov, Self-Modulation Instability of a Long Proton Bunch in Plasmas. Phys. Rev. Lett. 104 255003 (2010).
A. Caldwell and K. V. Lotov, Plasma wakefield acceleration with a modulated proton bunch. Physics of Plasmas, Phys. Plasmas 18(10), 103101 (2011).
R. Assmann et al. (AWAKE Collaboration), Proton-driven plasma wakefield acceleration: a path to the future of high-energy particle physics. Plasma Phys. Control. Fusion 56, 084013 (2014).
E. Gschwendtner et al. (AWAKE Collaboration). “AWAKE, the Advanced Proton-Driven Plasma Wakefield Acceleration Experiment at CERN” (this issue).
A. Caldwell, et al. “Path to AWAKE: Evolution of the concept” (this issue).
G. Plyushchev, Rubidium flow simulations in the AWAKE plasma section. Talk presented at the 11th AWAKE Physics Board Meeting, CERN, 2015. Presentation and the report available online from the indico webpage `https://indico.cern.ch/event/403300/`
C. G. R. Geddes, K. Nakamura, G. R. Plateau, Cs. Toth, E. Cormier-Michel, E. Esarey, C. B. Schroeder, J. R. Cary, and W. P. Leemans, Plasma-Density-Gradient Injection of Low Absolute-Momentum-Spread Electron Bunches. Phys. Rev. Lett. 100, 215004 (2008).
T. Katsouleas, Physical mechanism in the plasma wake-field accelerator. Phys. Rev. A., v.33 (1986), p.2056–2064.
K. V. Lotov, A. P. Sosedkin, A. V. Petrenko, L. D. Amorim, J. Vieira, R. A. Fonseca, L. O. Silva, E. Gschwendtner and P. Muggli, Electron trapping and acceleration by the plasma wakefield of a self-modulating proton beam. Physics of Plasmas 21, 123116 (2014).
C. B. Schroeder, C. Benedetti, E. Esarey, F. J. Gruner, and W. P. Leemans, Particle beam self-modulation instability in tapered and inhomogeneous plasma. Physics of Plasmas, 19, 010703 (2012).
`www.inp.nsk.su/~lotov/lcode/` K. V. Lotov, Phys. Rev. ST Accel. Beams [**6**]{}, 061301 (2003). K. V. Lotov, A. Sosedkin, E.Mesyats, Simulation of Self-modulating Particle Beams in Plasma Wakefield Accelerators. Proceedings of IPAC’2013 (Shanghai, China), p. 1238.
K. V. Lotov, Physics of beam self-modulation in plasma wakefield accelerators. Phys. Plasmas 22, 103110 (2015).
K. V. Lotov, A. Pukhov, and A. Caldwell, Effect of plasma inhomogeneity on plasma wakefield acceleration driven by long bunches. Phys. Plasmas 20(1), 013102 (2013).
|
{
"pile_set_name": "ArXiv"
}
|
---
address:
- 'Institut für Theoretische Physik, Universität Würzburg, Am Hubland, 97074 Würzburg, Germany'
- 'Theoretische Physik, ETH Zürich, CH-8093 Zürich, Switzerland '
-
-
-
-
author:
- Ansgar Dorneich
- Matthias Troyer
title: |
Accessing the dynamics of large many-particle systems\
using Stochastic Series Expansion
---
The SSE technique {#sse_tech}
=================
Since their first formulation in the early eighties[@suzu.76; @hirs.82] Quantum Monte Carlo (QMC) methods have become one of the most powerful numerical simulation techniques and tools in many-body physics. The first QMC algorithms were based on a discretization in imaginary time (“Trotter decomposition”) and used purely local update steps to sample the system’s statistically relevant states. These methods require a delicate extrapolation to zero discretization in order to reduce systematic errors. Furthermore, the purely local updates often proove incapable to traverse the accessible states in an efficient way: autocorrelation times grow rapidly with increasing system size.
A more recent class of QMC algorithms, the so-called “loop algorithms”[@ever.93; @hge_more; @wiese; @Hubbard; @nk1; @nk2; @nk3; @LoopReview], use non-local cluster or loop update schemes, thus reducing autocorrelation times by several orders of magnitude in some cases. Unfortunately, it is often highly non-trivial to construct a loop algorithm for a new Hamiltonian, and some important interactions cannot be incorporated into the loop scheme. These interactions have to be added as [*a posteriori*]{} acceptance probabilities after the construction of the loop, which can seriously decrease overall efficiency of the simulation. Some loop algorithms also suffer from “freezing”[@ever.93; @kohn.97] when the probability is high that a certain type of cluster occupies almost the whole system.
These insufficiencies can be overcome using the “sto- chastic series expansion” (SSE) approach together with a loop-type updating scheme (see Ref. and earlier works referenced therein).
- SSE is (almost) as efficient as loop algorithms on large systems.
- It is a numerically exact method without any discretization error.
- It is as easy to construct and general in applicability as world-line methods.
Following Sandvik[@sand.99; @sand.92; @sand.97] we briefly outline the basic ideas of SSE now. The central quantity to be sampled in a QMC simulation is the partition function $$Z = \mbox{Tr}(\mbox{e}^{-\beta{\hat{H} }}),$$ where ${\hat{H} }$ is the system’s Hamiltonian and $\beta=1/T$ the inverse temperature. Standard QMC techniques[@vdL.92] split up the exponential into a product of many “imaginary time slices” $\mbox{e}^{-\Delta\tau{\hat{H} }}$ and truncate the Taylor expansion of this expression after a certain order in $\Delta\tau$, thereby introducing a discretization error of order $\Delta\tau^n$. In SSE, however, one chooses a convenient Hilbert base $\{{ |\!\; {\alpha} \!\;\rangle}\}$ (for example the $S^z$ eigenbase $\{{ |\!\; {\alpha} \!\;\rangle}\}=\{{ |\!\; {S^z_1,S^z_2,...,S^z_N} \!\;\rangle}\}$) and expands $Z$ into the power series $$\label{power_series}
Z = \sum_\alpha \sum_{n=0}^{\infty} \frac{(-\beta)^n}{n!} { {\left\langle\!\: {\alpha} \!\;\right|} }
{\hat{H} }^n { |\!\; {\alpha} \!\;\rangle}.$$ The statistically relevant exponents of this power series are centered around $$\label{mean_n}
\langle n \rangle \propto N_s \beta,$$ where $N_s$ is the number of sites (or orbitals) in the system. (This follows from Eq. (\[E\_mea\]) and from $\langle E \rangle \propto N_s$.) We can thus truncate the infinite sum over $n$ at a finite cut-off length $L\propto N_s \beta$ without introducing any systematic error for practical computations. The best value for $L$ can be determined and adjusted during an initial thermalization phase of the QMC simulation: beginning with a relatively small value of $L$ one can start the QMC update process, stop it whenever the cut-off $L$ is exceeded and continue with $L$ increased by $10...20\%$.
Now let ${\hat{H} }$ be composed of a certain number of elementary interactions involving one or two sites (such as on-site potentials, nearest neighbor hopping etc.). In order to obtain a uniform notation we combine those interactions affecting only one site to new “bond” interactions. (One can, for example, take two chemical potential terms $\mu\cdot \hat{n}\mbox{\small (site1)}$ and $\mu\cdot \hat{n}\mbox{\small (site2)}$ and form the bond term $\frac{1}{C}\mu(\hat{n}\mbox{\small (site1)}
+\hat{n}\mbox{\small (site2)})$ with the constant $C$ assuring that the sum over all new bond terms equals the sum over all initial on-site terms.) We can thus assume in the following that ${\hat{H} }$ is a finite sum of “bond” terms ${\hat{H} }_b$ and that the operator strings ${\hat{H} }^n$ in (\[power\_series\]) can be split into terms of the form $$\prod_{i=1}^n {\hat{H} }_{b_i}^{(a_i)},$$ where $b_i$ labels the bond on which the elementary interaction term operates and $a_i$ the operator type (e.g. density–density interaction or hopping). By introducing “empty” unit operators ${\hat{H} }^{(0)}=\mbox{id}$ one can artificially grow all operator strings to length $L$ and obtain[@sand.97] $$Z = \sum_\alpha \sum_{\{S_L\}} \frac{\beta^n (L-n)!}{L!}
{ {\left\langle\!\: {\alpha} \!\;\right|} } \prod_{i=0}^L (-{\hat{H} }^{(a_i)}_{b_i}) { |\!\; {\alpha} \!\;\rangle}.$$ Here $\{S_L\}$ denotes the set of all concatenations of $L$ bond operators ${\hat{H} }_{b}^{(a)}$ and $n$ is the number of non-unit operators in $S_L$.
If we want to sample the $(\alpha,S_L)$ according to their relative weights with a Monte Carlo procedure we have to make sure that the matrix element of each bond operator is zero or negative since in order to fulfill detailed balance we choose the acceptance probability $p$ of a bond interaction to be proportional to its negative matrix element. This requires however that all matrix elements be non-positive. Does a simple redefinition of the zero of energy help? For the diagonal operators we can indeed add the same negative constant $C$ to each of them without changing the system’s properties, and thus make all matrix elements negative or zero. Unfortunately, for the non-diagonal terms an equally simple remedy does not exist. If one can show, however, that such a non-diagonal operator must appear pairwise for the matrix element to be non-zero, its matrix element can be multiplied by $-1$ without changing the physics of the system. (This corresponds to a gauge transformation on all lattice sites with odd parity.) On non-frustrated lattices this trick is widely applicable, which considerably increases the set of Hamiltonians suitable for SSE. If there are valid world-line configurations carrying an odd number of non-diagonal vertices with positive matrix elemet – which is typical for Hamiltonians and lattices with frustrations – only the conventional approach of dealing with the sign problem helps[@hirs.82; @taka.86; @hata.94]: one simulates a new system with the acceptance probabilty $p'=|\:\!p\:\!|$ and obtains the estimate of a physical quantity $Q$ in the form $$\langle Q \rangle = \frac{\langle Q\, \mbox{sign}\,p\rangle}
{\langle\mbox{sign}\,p\rangle}.$$ Unfortunately, $\langle \mbox{sign}\, p \rangle$ tends to zero exponentially with increasing system size $N_s$ and inverse temperature $\beta$, so that the computation time needed to achieve a certain accuracy exponentially increases with $N_s \beta$ and the practically accessible range of system sizes and temperatures is rather limited.
Loop updates {#loop_upd}
============
Having outlined the basic idea of SSE we review the non-local updating updating scheme proposed by Sandvik.[@sand.99] In the following figures we illustrate the proceeding by means of a simple physical model: a system of two types of hard-core bosons on a 6-site chain with periodic boundary conditions and Hamiltonian $$\begin{aligned}
H&=&-t\sum_{\alpha=1,2}\sum_{i}{\cal P}
\left[a_{\alpha,i}^{\dag}a_{\alpha,i+1}+H.c.\right]
{\cal P} \nonumber \\ && +
\sum_{\alpha=1,2}\mu_{\alpha}\sum_{i}n_{\alpha,i}
\\ &&+
\sum_{\alpha=1,2}\eta_{\alpha}\sum_{i}
{\cal P}\left[a_{\alpha,i}^{\dag}a_{\alpha,i+1}^{\dag} + H.c.\right]
{\cal P}\nonumber\end{aligned}$$
with $${\cal P} = \sum_{i}(1-n_{1,i}n_{2,i})\,.$$ The creation operator $a_{\alpha,i}^{\dag}$ creates a hardcore boson of type $\alpha=1$ or $2$ on site $i$. The first term ($t$) is a nearest neighbor hopping term, the second term ($\mu_{\alpha}$) the chemical potential and the third term ($\eta_{\alpha}$) pair creation and annihilation. The projection operator ${\cal P}$ implements the hard core constraints between the two types of bosons. In the world-line representation – in which the horizontal axis represents the spatial dimension and the vertical axis the propagation level $l\!=\!1 ... L$ – we symbolize type-1 bosons by single solid lines, type-2 bosons by double lines and empty sites by dotted lines (see Fig.\[fig1\]).
Sandvik separates the set of all bond operators into three classes: empty operators ${\hat{H} }^{(0)}$, diagonal operators ${\hat{H} }^{(d)}$ and non-diagonal operators ${\hat{H} }^{(nd)}$. The QMC process starts with an arbitrarily chosen initial state ${ |\!\; {\alpha} \!\;\rangle}$ and an empty operator string: in Fig. \[fig1\], for example, three sites are occupied with type-1 bosons, two sites are empty and on site 2 is occupied by a type-2 particle. Now two different update steps are performed in alternating order: a diagonal update exchanging empty and diagonal bond operators and an operator loop update transforming and exchanging diagonal and non-diagonal operators.
In the diagonal update step the operator string positions $l=1...L$ are traversed in ascending order. If the current bond operator is a non-diagonal one it is left unchanged; if it is an empty or diagonal operator it is replaced by a diagonal or empty one with a certain probability satisfying detailed balance (i.e. an operator with lower energy is more likely to be maintained or inserted than an operator with higher energy) (Fig.\[fig2\]).
Following Sandvik[@sand.99] we use the notation $${ |\!\; {\alpha(l)} \!\;\rangle} = \prod_{i=1}^l {\hat{H} }_{b_i}^{(a_i)} { |\!\; {\alpha} \!\;\rangle}$$
for the state obtained by acting on ${ |\!\; {\!\alpha\!} \!\;\rangle}$ with the first $l$ bond operators and ${ |\!\; {\!\alpha_b(l)\!} \!\;\rangle}$ for the restriction of ${ |\!\; {\!\alpha(l)\!} \!\;\rangle}$ to the bond $b$. Let $M$ be the total number of interacting bonds on the lattice. Then the detailed balance conditions for the diagonal update read
$$\begin{aligned}
P({\hat{H} }^{(0)}{{ \scriptstyle }(l)}\rightarrow {\hat{H} }^{(d)}_b{{ \scriptstyle }(l)}) &&\,= \\
\min\Big(1 && ,\frac{M\beta{ {\left\langle\!\: {\alpha_b(l)} \!\;\right|} }{\hat{H} }^{(d)}_b
{ |\!\; {\alpha_b(l)} \!\;\rangle}}{L-n}\Big)\\
P({\hat{H} }^{(d)}_b{{ \scriptstyle }(l)}\rightarrow {\hat{H} }^{(0)}{{ \scriptstyle }(l)}) &&\,= \\
\min\Big(1 && ,\frac{L-n+1}{M\beta{ {\left\langle\!\: {\alpha_b(l)} \!\;\right|} }{\hat{H} }^{(d)}_b
{ |\!\; {\alpha_b(l)} \!\;\rangle}}\Big)\,.\end{aligned}$$
$$$$
Non-diagonal bond operators cannot simply be inserted into the world line configuration as diagonal operators can: their insertion and modification requires local changes of the world-line occupations. We discussed earlier in this paper that concatenated local changes along a closed path (or loop) through the network of world-lines and interaction vertices are much more efficient than independent purely local changes. Sandvik proposed the following method to construct such a loop: a certain world-line and a propagation level $l$ on it is chosen arbitrarily; at the chosen point one disturbs the world-line by a local change – for example the creation or annihilation of a particle. Then one chooses a direction (up or down in propagation direction) and starts moving the disturbation in this direction (Fig. \[fig3\]). The aim is to move this disturbation (we will call it “loop head” in the following) through the network of world-lines and interaction vertices until the initial discontinuity is reached again and healed up.
Whenever the loop head reaches an interaction vertex we must decide how to go on; in the situation shown in Fig. \[fig3\] the path “bounce” is always possible since it results in an unchanged vertex. The path “straight” results in a diagonal vertex, and the path is possible if the matrix element of that vertex is nonzero. The path “turn” is only allowed if the Hamiltonian contains nearest neighbor hopping terms for particle type 2, while path “jump” is forbidden unless the Hamiltonian also allows for pair creation of particle type 2. The choice among the allowed paths must again satisfy detailed balance.
In our model – in which both pair creation and hopping are allowed – we might end up with the series “turn”, “jump”, “turn”, “turn” of path choices, after which the starting point is regained and the world-line discontinuity healed up (Fig. \[fig4\]). The overall result of this loop is that we have replaced 4 diagonal interactions by 4 non-diagonal interactions (marked “n.d.”) in Fig. \[fig4\].
Sandvik’s method implicitly assumes that running with a world-line change into an interaction vertex always requires choosing an outgoing leg and a change on it and continuing the loop. But what if the encircled vertex in Fig. \[fig3\] with three empty legs and one leg occupied by particle 2 is also a valid vertex? Then we have to add a fifth possibility to the list of allowed path choices: the “stop here”. If this last alternative is chosen the loop has reached a dead end. In this case our SSE code terminates the loop here, goes back to the starting point and moves in the opposite direction until either another dead end is encounterd or the starting point is reached again and the initial discontinuity is healed up.
From Fig. \[fig3\] one can see that when choosing a path through the current vertex there is always the possibility to undo the current change on the incoming leg of the vertex and to “bounce” backwards on the same leg. This path choice is normally not very helpful since it means one step backwards in the construction of the current loop. Fortunately, all “bounce” paths can be suppressed without violating detailed balance if on each bond all nonzero matrix elements are equal, or can be made equal after a suitable energy shift of the diagonal vertices. As an additional benefit, without the “bounce” path the algorithm becomes equivalent to the loop algorithms. For each vertex a path can be chosen according to detailed balance, after which the loop construction becomes deterministic. All the Heisenberg models studied in section \[sec\_scal\] are examples for this class of “optimizable” physical systems.
A further improvement of the update scheme is possible in the limit of high temperatures, i.e. $\beta\rightarrow 0$. Equation (\[mean\_n\]) tells us that the average number of (non-empty) vertices is rather small in this situation, and a large part of all world-lines is not connected to any vertex at all. The loop update will not be very efficient here since it essentially needs a sufficient number of vertices interconnecting the world-lines. For this reason our SSE code additionally performs a so-called “free world-line update” on each world line carrying no vertex at all. In this update the occupation of the entire world-line is changed to a randomly selected new occupation.
We have stressed several times that all the local path choices satisfy detailed balance. What remains to be shown is that the updating mechanism is ergodic in the grand canonical ensemble, i.e. that all bond operator strings $S_L$ and all states ${ |\!\; {\alpha} \!\;\rangle}$ can be reached. In order to demonstrate that we remind the reader that loops crossing the boundary between first and last propagation level $l$ modify the initial state ${ |\!\; {\alpha} \!\;\rangle}$ for the next update cycle. Therefore, the loops sample not only $S_L$ but also ${ |\!\; {\alpha} \!\;\rangle}$, and starting from a completely empty system [*any*]{} allowed configuration can be generated by a series of loops traversing one entire world-line each.
Numerical tests of the loop-update mechanism described above show that for large system sizes and if there are elementary interactions with very different energy scales, the loop construction sometimes gets stuck and the loop head does not find its way back to the starting point even after millions of steps. In order to avoid this trapping loops that exceed a critical length are aborted and the original state of the vertices is restored. This causes no systematic errors for measurements done between loop updates as detailed balance is not violated. The measurements of Green’s functions $G(r)$, however, which are performed “on the run” during loop construction (see Sec. \[sec\_GF\]), are biased if large loops are thrown away. Since the large loops are more likely to reach regions of the systems far away from the starting point than short loops, the values of $G(r)$ for large distances $r$ are systematically under-estimated if a considerable amount of large loops is aborted. Hence the total number of aborted loops has to be checked before one can trust in the recorded Green’s functions.
Measurements {#sect_meas}
============
Efficient estimators for many static observables within the SSE mechanism have been derived by Sandvik.[@sand.97a]
- All observables ${\hat{H} }^{(a)}$ appearing as elementary interactions in the system’s Hamiltonian can be measured very easily by counting the corresponding interaction vertices in the bond operator string $S_L$: if $S_L$ contains on average $\langle N(a) \rangle$ such vertices one obtains $$\langle {\hat{H} }^{(a)} \rangle = -\frac{1}{\beta}\langle N(a) \rangle.$$
- Summing over all elementary terms ${\hat{H} }^{(a)}$ gives an estimator for the internal energy $E$: $$\label{E_mea}
E = -\frac{1}{\beta} \langle n \rangle,$$ where $n$ is the number of non-empty interaction vertices in $S_L$. (This equation can be derived very easily from $\langle E \rangle = \frac{\partial}{\partial\beta}\mbox{ln} Z$.)
- For the heat capacity $C_V$ we additionally have to measure the fluctuations of $n$: $$C_V = \langle n^2 \rangle - \langle n \rangle ^2 - \langle n \rangle.$$
- Equal time correlations of two diagonal operators ${\hat{D} }_1$ and ${\hat{D} }_2$ can be measured via $$\langle {\hat{D} }_1 {\hat{D} }_2 \rangle = \left\langle \frac{1}{n+1}
\sum_{l=0}^{n} d_2[l]\,d_1[l] \right\rangle,$$ where $d_i[l]=\langle\alpha(l)|{\hat{D} }_i|\alpha(l)\rangle$.
Are there equally efficient estimators for time-dependent observables? In SSE the propagation index $l$ describes the evolution of an initial state when a series of elementary terms of the Hamiltonian is acting on it; thus $l$ plays a role analogous to imaginary time in a standard path integral. More detailed calculations[@sand.92] show that an imaginary time separation $\tau$ corresponds to a binomial distribution of propagation distances $\Delta l$; the time-dependent correlation $\langle {\hat{D} }_2(\tau){\hat{D} }_1(0)\rangle$, for example, is related to the correlator $$C_{12}(\Delta l) = \frac{1}{n+1} \sum_{l=0}^{n} d_2[l+\Delta l]\,d_1[l]$$ via $$\label{time_dep_corr}
\langle {\hat{D} }_2(\tau){\hat{D} }_1(0) \rangle \!=\! \left\langle
\sum_{\Delta l=0}^n\!\! {n \choose \Delta l}\!\!\left(\frac{\tau}{\beta}\right)^{\!\!\Delta l}\!\!\!
\left(1\!-\!\frac{\tau}{\beta}\right)^{\!\!n-\Delta l}\!\!\!\!\!
C_{12}(\Delta l)\! \right\rangle.$$ Instead of working in a representation with varying $n$ a fixed string size $L$ can be chosen, as the identity vertices are uniformly distributed and do not influence the mapping from index to imaginary time.
The corresponding generalized susceptibilities can be calculated straight forward by integrating $\langle {\hat{D} }_2(\tau){\hat{D} }_1(0)\rangle$ over $\tau$, $$\chi_{12}=\int_0^\beta \langle {\hat{D} }_2(\tau){\hat{D} }_1(0) \rangle\, d\tau.$$ which gives[@sand.92] $$\begin{aligned}
\chi_{12} =&& \Bigg\langle {\beta\over n(n+1)}
\left (\sum\limits_{l=0}^{n-1} d_2[l] \right )
\left (\sum\limits_{l=0}^{n-1} d_1[l] \right )\nonumber\\
&&+{\beta\over (n+1)^2} \sum\limits_{l=0}^{n} d_2[l]\,d_1[l] \Bigg\rangle .\end{aligned}$$
Scaling Behavior {#sec_scal}
================
One decisive criterion for the performance of a QMC simulation technique is the behavior of computation time $C$ as a function of system size $N_s$ or inverse temperature $\beta$; To facilitate a hardware-independent measurement of $C$ and a comparison to other QMC techniques we define $C$ as the number of elementary update operations needed to transform a given state ${ |\!\; {\!\alpha^{(n)}\!} \!\;\rangle}$ into an new state ${ |\!\; {\!\alpha^{(n+1)}\!} \!\;\rangle}$ in such a way that the mean autocorrelation time $\tau$ is equal to 1. In SSE the number of elementary update operations is the number of diagonal vertices tested for replacement plus the number of vertices traversed during the loop update. In the following we compare SSE to the loop-algorithm, which is known to show an excellent scaling behavior for many benchmark problems. As test models we choose isotropic antiferromagnetic Heisenberg models in one, two and three dimensions with up to 4096 sites and $\beta$ up to $64$ in a vanishing or finite external magnetic field.
Following Ref. we describe the scaling behavior of the two algorithms by means of the dynamical exponent $z$ defined from $$\label{eq_dyn_exp}
\tau\!\cdot\!C\propto \beta\!\cdot\!l^D l^z\,.$$ Here, $\tau\!\cdot\!C$ is the computational effort (i.e. the number of elementary update steps) needed to achieve a mean autocorrelation time of $\tau\!=\!1$ for the measurements of the studied quantity; $D$ is the spatial dimension of the simulated system and $l=\sqrt[D]{N_s}$ its length in each dimension. From Table \[tab\_scal2Dh0\] we see that both simulation techniques show an approximately equal performance and an very good scaling behavior: since the ratio $C\tau/(\beta N_s)$ is approximately constant we obtain $z\approx 0$ in both cases.
----------------------------- --------------------------------------------------- ------------------- --------------------------------------------------- -------------------
$\beta\cdot N_s$ $\displaystyle \frac{C\cdot\tau}{\beta\cdot N_s}$ $\displaystyle \frac{C\cdot\tau}{\beta\cdot N_s}$
\[1.7ex\]
\[-2.2ex\] $4\cdot\,\; 4^2$ 1.00 $0.040(46\pm 16)$ 1.00 $0.040(20\pm 10)$
$8\cdot\,\; 8^2$ 0.61 $0.044(83\pm 15)$ 0.90 $0.044(92\pm 8)$
$16\cdot 16^2$ 0.40 $0.044(72\pm 12)$ 0.56 $0.044(68\pm 6)$
$32\cdot 32^2$ 0.40 $0.044(19\pm 11)$ 0.53 $0.044(24\pm 6)$
$64\cdot 64^2$ 0.36 $0.044(01\pm 23)$ 0.42 $0.044(07\pm 14)$
----------------------------- --------------------------------------------------- ------------------- --------------------------------------------------- -------------------
: 2D antiferromagnetic Heisenberg model at vanishing magnetic field $h=0$: calculation of the uniform magnetic susceptibility $\langle\chi\rangle = \frac{\partial \langle M\rangle}{\partial h}\big|_{h=0}$ from QMC simulations with 1000000 (120000 in the case $\beta=L=64$) update-measurement cycles (left: SSE, right: loop-algorithm). $C\cdot\tau$ is the number of elementary update operations per cycle needed to achieve a mean autocorrelation time $\tau=1$ for the measurements of $\chi$.[]{data-label="tab_scal2Dh0"}
Next we enlarge the square lattice into the third spatial dimension and examine a bilayer quantum Heisenberg antiferromagnet at the quantum critical point separating the spin gap phase from the magnetucally ordered one.[@bilayer] Our aim is to measure scaling behavior and dynamical exponents exactly at this quantum critical point. This point is of particular interest since the immediate neighborhood of a phase transition often leads to the so-called “critical slowing down” of QMC simulations, i.e. exploding autocorrelation times and thus a dramatic decrease of efficiency of the QMC update process.
The results in Table \[tab\_scalBih0\] show that the scaling behavior for both algorithms is still almost linear in $\beta N_s$. The scaling for SSE looks slightly superior to the loop algorithm. This difference can most probably be attributed to the fact that improved estimators were used in the loop algorithm simulation, leading to slightly smaller errors but larger autocorrelation times. There is no sign of critical slowing down in either algorithm.
[rclcl]{} & &\
$\beta\cdot N_s$ &
------------------------------------------------------------------------
$\displaystyle \frac{C\cdot\tau}{\beta\cdot N_s}$ & & $\displaystyle \frac{C\cdot\tau}{\beta\cdot N_s}$ &\
\
$4\cdot\,\; 2\cdot 4^2$ & 1.00 & $0.0115(6\pm 7)$ & 1.00 & $0.0114(6\pm 5)$\
$8\cdot\,\; 2\cdot 8^2$ & 0.96& $0.0068(2\pm 6)$ & 1.03 & $0.0069(2\pm 2)$\
$16\cdot 2\cdot 16^2$ & 0.68 & $0.0036(8\pm 4)$ & 1.20 & $0.0036(6\pm 2)$\
$32\cdot 2\cdot 32^2$ & 0.56 & $0.0018(5\pm 3)$ & 1.20 & $0.0018(3\pm 1)$\
As we have mentioned in the introduction one of the major advantages of SSE is that external potentials (and magnetic fields in spin models) can be included without a loss of performance. To verify this assertion we now examine the antiferromagnetic Heisenberg model on a chain and a square lattice in a finite magnetic field $h\ne0$. For the loop-algorithm we expect to find a rapidly increasing autocorrelation time and decreasing performance if the product of magnetic field $h$ and inverse temperature is much larger than 1. This is due to the fact that the external field is incorporated into the loop-algorithm via a-posteriori acceptance probabilities for each constructed loop. for $\beta h \ll 1$ these probabilities are still large, whereas at $\beta h \approx 1$ they begin to decrease considerably.
-------------- ---- ---- ---- ----------------------------------------------------------- ----------------- --------- ------------------
$\displaystyle \frac{C\!\cdot\!\tau}{C_0\!\cdot\!\tau_0}$
\[1.7ex\]
\[-2.2ex\] 0 02 0 32 1.00 $0.008(2\pm 5)$ 1.00 $0.0083(7\pm 7)$
0 04 0 64 1.02 $0.016(4\pm 6)$ 0.97 $0.017(6\pm 2)$
0 1 1 6 1.22 $0.057(8\pm 8)$ 1.84 $0.057(7\pm 5)$
0 2 3 2 1.98 $0.24(4\pm 2) $ 7.55 $0.24(3\pm 2)$
0 4 6 4 1.37 $0.893(8\pm 9)$ 167.40 $0.89(4\pm 3)$
1 0 16 0 0.86 $2.069(0\pm 8)$ 2338.74 $2.(24\pm 12)$
-------------- ---- ---- ---- ----------------------------------------------------------- ----------------- --------- ------------------
: *1D chain antiferromagnetic Heisenberg model in a magnetic field $h$: calculation of magnetization $M$ from QMC simulations with 1000000 update-measurement cycles (left: SSE, right: loop-algorithm) for a system with $\beta=N_s=16$. $C\cdot\tau$ is the number of elementary update operations per cycle needed to achieve a mean autocorrelation time $\tau=1$ for the measurements of $M$.*[]{data-label="tab_scal1Dh"}
Indeed, the numerical results in Table \[tab\_scal1Dh\] demonstrate that at $\beta h\approx 10$ the loop algorithm cannot be used any more because the autocorrelation times get too long. For SSE, on the contrary, we do not expect any negative effect by introducing a magnetic field whose strength is of the order of the other elementary interactions, $h\approx J$, since no a-posteriori acceptance decision is neccessary. We rather presume that performance is slightly worse for $h/J\ll 1$ because there are elementary interaction vertices with very different energy scales. Both predictions are verified by the data in Table \[tab\_scal1Dh\]. For weak fields $h \ll J$ it might be preferable to construct loops in zero field, and to introduce the field via an a-posteriori Metropolis decision on whether to accept loops which change the magnetization, as it is done in the loop algorithm.
For sake of completeness we also show the corresponding data for the 2-dimensional Heisenberg model in Table \[tab\_scal2Dh\]. The results demonstrate that the different behavior of SSE and loop-algorithm described in the one dimensional case is even more severe in two dimensions: the loop-algorithm loses its efficiency already at $\beta h\approx 1.5$.
-------------- ---- ---- ---- ----------------------------------------------------------- ---- --------------- --------- -----------------
$\displaystyle \frac{C\!\cdot\!\tau}{C_0\!\cdot\!\tau_0}$
\[1.7ex\]
\[-2.2ex\] 0 02 0 32 1.00 0 $22(4 \pm 8)$ 1.00 $0.231(3\pm 3)$
0 04 0 64 0.83 0 $47(9 \pm 8)$ 1.38 $0.477(7\pm 7)$
0 1 1 6 0.94 1 $41(0 \pm 9)$ 5.61 $1.42(2\pm 2)$
0 2 3 2 0.44 3 $47(0 \pm 7)$ 34.25 $3.48(6\pm 7)$
0 4 6 4 0.18 7 $73(7 \pm 4)$ 1691.66 $7.7(8\pm 7)$
1 0 16 0 0.12 22 $10(6 \pm 4)$ $---$ $-----$
-------------- ---- ---- ---- ----------------------------------------------------------- ---- --------------- --------- -----------------
: Square lattice antiferromagnetic Heisenberg model in a magnetic field $h$: calculation of magnetization $M$ from QMC simulations with 1000000 update-measurement cycles for a system with inverse temperature $\beta J=16$ and $N_s=16^2$ lattice sites. $C\cdot\tau$ is the number of elementary update operations per cycle needed to achieve a mean autocorrelation time $\tau=1$ for the measurements of $M$.[]{data-label="tab_scal2Dh"}
In some cases other performance measurements are more interesting. One could ask how the computation time till a certain accuracy in a certain measured variable is reached scales with $\beta$ and $N_s$. This is studied in Fig. \[fig\_scal\_beta\_vol\]. For the square lattice Heisenberg antiferromagnet we trace the computation time to reach an accuracy of 4 digits in energy as a function of $\beta$ (Fig. \[fig\_scal\_beta\_vol\] top) and $N_s$ (Fig. \[fig\_scal\_beta\_vol\] bottom). The exponents $\kappa{\scriptstyle (\beta)}$ in $C\propto\beta^{\kappa{\scriptstyle (\beta)}}$ and $\kappa{\scriptstyle (N_s)}$ in $C\propto\beta^{\kappa{\scriptstyle (N_s)}}$ derived from Fig. \[fig\_scal\_beta\_vol\] are $$\begin{aligned}
\kappa{\scriptstyle (\beta)} &=& 0.34 \pm 0.05, \\
\kappa{\scriptstyle (N_s)} &=& 0.48 \pm 0.05. \end{aligned}$$ Both quantities are smaller than 1, and Eq. (\[eq\_dyn\_exp\]) would return a negative dynamical exponents $z$. This is due to self-averaging: in a large system local fluctuations of a physical observable around its mean value on different subregions of the lattice can compensate and average out each other, thereby lowering the observable’s measured variance. The computational effort needed to get thermodynamical averages to a certain relative error scales sublinearly with system size and inverse temperature, so that systems of several thousand sites or at temperatures of not more than $0.001 J$ can be simulated within minutes or a few hours on a standard PC or workstation.
Green’s functions {#sec_GF}
=================
The observables listed listed in section \[sect\_meas\] serve to access important static thermodynamic properties of the studied system. However, properties such as photo emission $\langle a^\dagger{{ \scriptstyle }({\mathbf{k} },\omega)}\, a{{ \scriptstyle }(0,0)} \rangle$ or spin flip $\langle S^-{{ \scriptstyle }({\mathbf{k} },\omega)}\, S^+{{ \scriptstyle }(0,0)} \rangle$ are often even more interesting as they provide insights into the system’s dynamics. Within the framework of SSE measuring these Green’s functions $G({\mathbf{k} },\omega)$ requires the insertion of local changes on certain world-lines (such as removing a particle at propagation level $l_1$ on world-line $w_1$ and re-inserting it at propagation level $l_2$ on world-line $w_2$). Performing these insertions is a highly non-trivial task since on the one hand detailed balance must be assured, on the other hand the whole process has to sample all distances $r=w_2-w_1$ and all propagation differences $\Delta l=l_2-l_1$ efficiently. Both requirements are already fulfilled by the loop update steps. Since this update inserts and moves local changes on the network of world-lines and connecting interaction vertices it can be used to record the corresponding Green’s functions $G(r,\Delta l)$ “on the fly” while constructing the loop update.
As an example we reconsider the hard-core boson model from section \[loop\_upd\] and in particular the operator loop shown in Figs. \[fig3\] and \[fig4\] which starts with the removal of a type-2 particle on propagation level 6 of world-line 2; our cut-off power in the series expansion was $L=9$, and previous diagonal updates have produced $n=7$ “non-identity” interaction vertices.
Taking level 6, the starting point of the loop, as zero point for the propagation direction we are now able to measure quantities of type $\langle a^\dagger_1{{ \scriptstyle }(r,\Delta l)}\, a^{}_2{{ \scriptstyle }(0,0)} \rangle$ and $\langle a^\dagger_2{{ \scriptstyle }(r,\Delta l)}\, a^{}_2{{ \scriptstyle }(0,0)} \rangle$ during the construction of this loop. Fig. \[fig6\] shows that for $\Delta l=0$ exactly 2 measurements of $\langle a^\dagger_2{{ \scriptstyle }(r,\Delta l=0)} \,
a^{}_2{{ \scriptstyle }(0,0)} \rangle$ and one of $\langle a^\dagger_1{{ \scriptstyle }(r,\Delta l=0)} \, a^{}_2{{ \scriptstyle }(0,0)} \rangle$ can be performed during the loop: one at the start (or end) of the loop at distance $r=0$, two on adjacent world-lines ($r=1$) while moving down (right) and up (left). The recorded value at each measurement is the product of the matrix elements of the creation/annihilation operators inserted at the open ends of the loop under construction. We denote the state at propagation level 6 in our example before inserting the two creation/annihilation operators as $| \alpha(6)\rangle$ and the state after insertion of the operators as $| \tilde{\alpha}(6)\rangle$. Then the $\Delta l\!=\! 0$ matrix element $\langle a^\dagger_2{{ \scriptstyle }(r\!=\!1,\Delta l\!=\!0)} \,a^{}_2{{ \scriptstyle }(0,0)}\rangle$ – measured when the loop head moves down on world line 3 – is $$\begin{aligned}
\langle a^\dagger_2{{ \scriptstyle }(r=1,\Delta l=0)} \,a^{}_2{{ \scriptstyle }(0,0)} \rangle
= \langle\;\!\tilde{\alpha}(6) |\;\! a^\dagger_2{{ \scriptstyle }(r=1,\Delta l=0)}
\,a^{}_2{{ \scriptstyle }(0,0)}\;\! | \;\!\alpha(6) \rangle\,. \end{aligned}$$ Stepping down by one more propagation level on world line 3 we can record the $\Delta l\!\ne\! 0$ matrix element $$\begin{aligned}
\langle a^\dagger_2&&{{ \scriptstyle }(r=1,\Delta l=-1)}\,a^{}_2{{ \scriptstyle }(0,0)}\rangle\\
&&= \langle\;\!\tilde{\alpha}(5) |\;\! a^\dagger_2{{ \scriptstyle }(r=1,\Delta l=-1)}\;\!
|\;\! \alpha(5) \rangle\, \langle\;\! \tilde{\alpha}(6) | \,a^{}_2{{ \scriptstyle }(0,0)}
|\;\! \alpha(6) \rangle\,. \end{aligned}$$ Leaving our hardcore boson example behind and returning to the general case we conclude this paragraph with the remark that for the creation or annihilation of a fermion the recorded matrix elements are always equal to 1, while they can adopt other values for spin flips or the creation/annihilation of bosons.
Having measured and recorded the quantities $G(r,\Delta l)$ (or a correlation function $C(r,\Delta l)=\langle {\hat{D} }_2(r,\tau){\hat{D} }_1(0,0) \rangle$) we still have to perform a couple of non-trivial transformation steps till we obtain the desired quantities $G(k,\omega)$ and $C(k,\omega)$ which describe the dynamical response of the system to external perturbations. First we have to relate propagation levels $\Delta l$ to imaginary times $\tau$, then a Fourier transform brings us from $r$-space to $k$-space; finally we need an inverse Laplace transform to step from imaginary time $\tau$ to exitation energy $\omega$.
Efficiently Accessing the system’s dynamics
===========================================
In this section we will discuss efficient implementation strategies for recording $G(r,\Delta l)$ and for the adjacent transformation steps mentioned above.
The transformation from propagation levels $\Delta l$ to imaginary time $\tau$ requires the same weight factors as discussed earlier for diagonal correlation functions: $$\begin{aligned}
\label{GF_tf}
G(r,\tau) &=& \sum_{\Delta l=0}^n
{L \choose \Delta l}\!\left(\frac{\tau}{\beta}\right)^{\Delta l}\!\!
\left(1\!-\!\frac{\tau}{\beta}\right)^{L-\Delta l'}\!\!G(r,\Delta l) \\
&\equiv& \sum_{\Delta l=0}^L w(\tau,\Delta l)\; G(r,\Delta l)\,. \nonumber\end{aligned}$$ where $$w(\tau,\Delta l) =
{L \choose \Delta l}\!\left(\frac{\tau}{\beta}\right)^{\Delta l}\!\!
\left(1\!-\!\frac{\tau}{\beta}\right)^{L-\Delta l}.$$
Working in a fixed string size representetion with fixed $L$ instead of varying $n$ is more convenient because the binomial weight prefactors are fixed during the entire simulation and can easily be calculated once at the beginning of the simulation.
There are several possible ways to implement the recording of $G(r,\Delta l)$ measurements and the adjacent transformation to $G(r,\tau)$. The easiest and at first glance fastest way simply writes all recorded $G(r,\Delta l)$ data into a two-dimensional array with dimensions $N_s$ and $L\propto N_s\beta$. The transformation to $G(r,\tau)$ can then be performed once at the end of the simulation. However, this method has two problems. A separate measurement has to be recorded each time the loop head steps up or down by one level on a world-line and whenever it traverses an interaction vertex. Recording all these measurements drastically slows down the loop update process. Second, for large systems ($N_s \approx 5000$) and low temperatures ($\beta\approx 40$) the two-dimensional array needed to store $G(r,\Delta l)$ contains about 1 billion elements and needs more memory than available on many computer systems.
In order to overcome these problems one can replace the “brute force” recording of data on [*all*]{} traversed $(r,\Delta l)$ points by a Monte Carlo sampling: in each loop-update a distance $\Delta l$ is chosen randomly, according to the probabilities in Eq. (\[GF\_tf\]), for each of the times $\tau$ of interest. Measurements are then performed only at these $\Delta l$ and transformed directly into $\tau$.
In our code we have adopted a third strategy: we perform [*all*]{} possible $G(r,\Delta l)$ measurements (thereby exploiting the fact that $G(r,\Delta l)$ is constant on the entire world-line fragment between tho adjacent vertices) and directly transform these into $G(r,\tau)$ at the end of each loop update step. The transformation after each QMC update step is necessary to keep memory requirements low.
Simply applying Eq. (\[GF\_tf\]) with its computationally expensive operations (divisions,powers,binomial coefficients,large sums) would now cost by far too much computation time. Instead we remember that $G(r,\Delta l)$ is composed of a relatively small number of intervals $I=]\Delta l_1{{ \scriptstyle }(I)},\Delta l_2{{ \scriptstyle }(I)}]$ with constant function value (Fig. \[fig7\]b)). Therefore we can compute the contribution of an entire $\Delta l$-interval to $G(r,\tau)$ in one step: $$G(r,\tau) = \sum_{I} G(r,I) \big(W(\tau,\Delta l_2{{ \scriptstyle }(I)}) -
W(\tau,\Delta l_1{{ \scriptstyle }(I)}) \big),$$ where $W$ is the “integrated weight function” $$W(\tau,\Delta l) = \sum_{m=0}^{\Delta l} w(\tau,m).$$ The $\Delta l$-range in which $W(\tau,\Delta l)$ considerably differs from 0 and 1 is determined by mean value and standard deviation of the binomial distribution $w(\tau,\Delta l)$ $$\begin{aligned}
\langle \Delta l \rangle &=& L \frac{\tau}{\beta}\\
\sigma_{\!\Delta l} &=& \sqrt{L \frac{\tau}{\beta}\left(1-\frac{\tau}{\beta}
\right)}.\end{aligned}$$ Below $\langle \Delta l \rangle - 5\,\sigma_{\!\Delta l}$ the integrated weight is zero, above $\langle \Delta l \rangle + 5\,\sigma_{\!\Delta l}$ it is 1 (up to an error of less than $10^{-7}$). The remaining interval rarely contains more than fifty or hundred $\Delta l$-points (see Fig. \[fig7\]d); these values can easily be stored after having been computed once for each $\tau$. Thus $W(\tau,\Delta l)$ can be calculated very rapidly with nothing but a couple of “cheap” elementary operations.
For very large systems and very low temperatures the “relevant” $\Delta l$-ranges might become so large that it is unfavorable to store all needed $W(\tau,\Delta l)$ values – for example because accessing the large array $W[\tau_i,\Delta l]$ would caust too many cache misses. In this case one can store the coefficients of some interpolation functions for $W(\tau,\Delta l)$ instead of the function values themselves. Practical tests have shown that dividing the relevant interval $[\langle \Delta l \rangle - 5\sigma_{\Delta l},
\langle \Delta l \rangle + 5\sigma_{\Delta l}]$ into six sub-intervals with boundaries $\langle \Delta l \rangle - 5\,\sigma_{\!\Delta l}$, $\langle \Delta l \rangle - 2.8\,\sigma_{\!\Delta l}$, $\langle \Delta l \rangle - 1.3\,\sigma_{\!\Delta l}$, $\langle \Delta l \rangle$, $\langle \Delta l \rangle +1.3\,\sigma_{\!\Delta l}$, $\langle \Delta l \rangle +2.8\,\sigma_{\!\Delta l}$ and $\langle \Delta l \rangle +5\,\sigma_{\!\Delta l}$ and interpolating $W$ in each sub-interval by a fifth-order polynomial is a good compromise between evaluation speed (about 15 elementary operations), storage requirements (36 floating point numbers for each $\tau$) and interpolation accuracy (better than $2..3\times 10^{-7}$).
The next transformation step, Fourier transform from $G(r)$ to $G(k)$, is a well known standard method that does not impose any fundamental problems. However, standard Fast Fourier Transform (FFT) algorithms perform best if [*all*]{} $G(k)$ values are to be calculated, whereas in practice one rarely needs all $k$-values and is interested only in one $k$-point or in some special points of the Brillouin zone, e.g. the point $k=(\pi,\pi)$ and its immediate neighborhood. Then one can save a lot of computation time by not recurring to FFT but using optimized algorithms designed particularly for these cases. If we are interested in only one or a few $k$-points we can use a simple Fourier transform to get $\{ G(k,\tau) \}$ from $\{G(r,\tau)\}$ in $\cal{O}$$(N_s\cdot n_k)$ operations ($n_k$ is the number of $k$-points). Correlation functions $C(k,\tau)$ can even be measured directly in $k$-space, which also can be done in $\cal{O}$$(N_s\!\cdot\! n_k)$ operations. For the case $1\ll n_k \ll N_s$ we have implemented a new Fourier transform algorithm performing much better than FFT in this situation.[@dorn.01]
Unlike a Fourier transform a Laplace transform in general cannot be inverted . Therefore the last transition step from $\tau$ to $\omega$ is by far more complicated than the previous one from $r$ to $k$. We use [*Maximum Entropy*]{} techniques developed within the last years and refer to earlier publications.[@preu.97]
Example: spin correlations of the 2D Heisenberg Model
=====================================================
In this section we use our standard benchmark model – the sauqre lattice Heisenberg antiferromagnet – to test our method of Green’s functions measurements for correctness and numerical efficiency. To this purpose we calculate and compare the correlation functions $$\begin{aligned}
\label{Sz_Sz}
&&\langle S^z(\tau,k)\,S^z(0,k)\rangle \quad \mbox{and}\\
\label{S+_S-}
&&\langle S^+(\tau,k)\,S^-(0,k)\rangle.\end{aligned}$$ In the $S^z$-eigenbasis, which is normally used to span the model’s Hilbert space, expression (\[Sz\_Sz\]) is a time-dependent correlation function of two diagonal operators. Therefore the diagonal operator in Eq. (\[Sz\_Sz\]) can be measured using Eq. (\[time\_dep\_corr\]) without introducing changes in the world-lines and vertices defining the current state of the system. The Green’s function Eq. (\[S+\_S-\]), however, consists of two non-diagonal operators and can only be measured with our new method of recording general Green’s functions that was described in section \[sec\_GF\]. Furthermore, at zero field $h=0$ both correlation functions are related via $$\label{SzSz_S+S-}
\langle S^z(\tau,k)\,S^z(0,k)\rangle
= \frac{1}{2} \langle S^+(\tau,k)\,S^-(0,k)\rangle,$$ so that the correctness of both estimators can be checked by directly comparing these two quantities.
When working with the antiferromagnet wee need to keep in mind that in order to keep the exchange matrix elements $S^+S^-$ and $S^-S^+$ positive we need to perform a gauge transformation, multiplying $S^+$ and $S^-$ on one sublattice by $-1$. This gauge transformation does not affect any diagonal operator, but leads to a momentum shift of $Q=(\pi,\pi)$, and (\[SzSz\_S+S-\]) for the Green’s function and Eq. (\[SzSz\_S+S-\]) becomes $$\label{SzSz_S+S-2}
\langle S^z(\tau,k)\,S^z(0,k)\rangle
= \frac{1}{2} \langle \hat{S}^+(\tau,k+Q)\,\hat{S}^-(0,k+Q)\rangle.$$ The numerical data in Table \[tab\_SzSz\_S+S-\] perfectly fulfil this equality and hence demonstrate the correctness of our Green’s functions measurements.
--------------------------------------------- ----- ------------------- ------------------
\[-2.3ex\]
\[0.8ex\]
\[-2.3ex\] $(\frac{\pi}{2},\frac{\pi}{2})$ 0 $0.16(89\pm 20)$ $0.168(2\pm 4)$
0.1 $0.00(01\pm 17)$ $0.003(2\pm 3)$
0.5 $-0.00(36\pm 15)$ $-0.000(4\pm 3)$
\[1.2ex\] $(\frac{3\pi}{4},\frac{3\pi}{4})$ 0 $0.38(39\pm 21)$ $0.386(0\pm 6)$
0.1 $0.01(74\pm 20)$ $0.020(9\pm 5)$
0.5 $0.00(41\pm 19)$ $0.000(4\pm 4)$
\[1.2ex\] $(\pi,\pi)$ 0 $11.3(35\pm 17)$ $11.36(1\pm 9)$
0.1 $10.4(04\pm 17)$ $10.42(3\pm 9)$
0.5 $9.0(83\pm 17)$ $9.09(3\pm 9)$
\[1.2ex\] $(\frac{3\pi}{4},\pi)$ 0 $0.53(85\pm 23)$ $0.542(2\pm 7)$
0.1 $0.06(31\pm 20)$ $0.063(0\pm 6)$
0.5 $0.00(38\pm 17)$ $0.000(0\pm 5)$
\[1.2ex\] $(\frac{\pi}{2},\pi)$ 0 $0.28(92\pm 23)$ $0.287(6\pm 5)$
0.1 $0.01(08\pm 19)$ $0.008(5\pm 4)$
0.5 $0.00(06\pm 17)$ $0.000(3\pm 4)$
--------------------------------------------- ----- ------------------- ------------------
: Comparison of $\frac{1}{2} \langle \hat{S}^+(k+Q,\tau)\,\hat{S}^-(k+Q,0)\rangle$ and $\langle S^z(k,\tau)\,S^z(k,0)\rangle$ for the $16\times 16$-site 2D AF Heisenberg model at $\beta=16$ and zero magnetic field. The table shows some $k$-values around $(\pi,\pi)$.[]{data-label="tab_SzSz_S+S-"}
In the simulation recorded in Table \[tab\_SzSz\_S+S-\] we have calculated $\langle S^z\,S^z\rangle$ and $\langle S^+\,S^-\rangle$ for all allowed $k$-points on the path $(0,0)\to(\pi,0)\to(\pi,\pi)\to(0,0)$. Table \[tab\_SzSz\_S+S-\] shows a subset of these points in the vicinity of $(\pi,\pi)$. The three tasks “performing updates”, “measuring $\langle S^z\,S^z\rangle$” and “measuring $\langle S^+\,S^+\rangle$” contributed the following percentages to overall computation time:
------------------------------------- --- --------
performing updates : 18.8 %
measuring $\langle S^z\,S^z\rangle$ : 36.1 %
measuring $\langle S^+\,S^-\rangle$ : 45.1 %
------------------------------------- --- --------
From this list and the measurement accuracies in Table \[tab\_SzSz\_S+S-\] we conclude that the highly non-trivial Green’s functions measurements lead to a slightly better accuracy than the direct $\langle S^z({\bf r},\tau)S^z({\bf r}',\tau')\rangle$ measurements while consuming roughly the same amount of computer time as the latter. Measuring the Green’s function is thus the preferred method of determining also the diagonal real space dynamical correlation functions.
summary
=======
Stochastic Series Expansion (SSE) together with the implementation tricks and Green’s functions measurements described in this paper is a highly performant quantum Monte Carlo simulation technique allowing to access both static and dynamical properties of very large systems of thousands of sites and at very low temperatures. Compared to the loop-algorithm, which is slightly faster on big systems for some specific Hamiltonians, SSE has the advantages of not suffering from exponential slowing down in external fields; furthermore, SSE is more easily applicable to wide classes of Hamiltonians.
We thank A. Sandvik for valuable discussions. This work was supported by DFG HA 1537/16-1,2 and KONWIHR OOPCV ( A. D. ) and the Swiss National Science Foundation ( M. T. ). High performance calculations were performed at HLRZ Jülich and LRZ Munich.
M. Suzuki, Prog. Theor. Phys. [**65**]{}, 1454 (1976). J. E. Hirsch, R. L. Sugar, D. J. Scalapino and R. Blankenbecler, Phys. Rev. B [**26**]{}, 5033 (1982). M. Takasu, S. Miyashita and M. Suzuki, Prog. Theor. Phys. [**75**]{}, 1254 (1986). N. Hatano, J. Phys. Soc. Japan [**63**]{}, 1691 (1993). H. G. Evertz, G. Lana and M. Marcu, Phys. Rev. Lett. [**70**]{}, 875 (1993). H.G. Evertz and M. Marcu, in [*Lattice 92*]{}, Amsterdam 1992, ed. J. Smit et al., Nucl. Phys. B (Proc. Suppl.) [**30**]{} (1993) 277, and in “Quantum Monte Carlo Methods in Condensed Matter Physics”, ed. M. Suzuki, World Scientific 1994, p. 65. U.-J. Wiese and H.-P. Ying, Phys. Lett. A[**168**]{}, 143 (1992); Z. Phys. B[**93**]{} (1994) 147. N. Kawashima, J.E. Gubernatis, and H.G. Evertz, Phys. Rev. B[**50**]{} 136 (1994). N. Kawashima and J.E. Gubernatis, Phys. Rev. Lett. [**73**]{}, 1295 (1994). N. Kawashima, J. Stat. Phys. [**82**]{}, 131 (1996). N. Kawashima and J.E. Gubernatis, J. Stat. Phys. [**80**]{}, 169 (1995). H.G. Evertz, in [*“Numerical Methods for Lattice Quantum Many-Body Problems”*]{}, ed. D.J. Scalapino, Addison Wesley Longman, Frontiers in Physics (1998). M. Kohno and M. Takahashi, Phys. Rev. B [**56**]{}, 3212 (1997). A. W. Sandvik, Phys. Rev. B [**59**]{}, R14157 (1999). A. W. Sandvik and J. Kurkijärvi, Phys. Rev. B [**43**]{}, 5950 (1991); A. W. Sandvik, J. Phys. A [**25**]{}, 3667 (1992). A. W. Sandvik, Phys. Rev. B [**56**]{}, 11678 (1997). W. von der Linden, Phys. Rep. [**220**]{}, 53 (1992). B. Li, N. Madras and A. D. Sokal, J. Statist. Phys. [**80**]{}, 661-754 (1995). A. W. Sandvik and D.J. Scalapino, Phys. Rev. Lett. [**72**]{}, 2777 (1994). A. W. Sandvik, R. R. P. Singh and D. K. Campbell, Phys. Rev. B [**56**]{}, 14510 (1997). A detailed description of the new method will soon be submitted to Comp. Phys. Comm., and a C++ code package will be available from the CPC program library. W. von der Linden,R. Preuss and W. Hanke, J. Phys.: Condens. Matter [**8**]{}, 3881 (1996).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The foreseen CaLIPSO Positron Emission Tomography (PET) scanner is expected to yield simultaneously a fine image resolution, about 1 mm$^3$, and a high contrast. In this paper we present results of simulations for the full CaLIPSO PET scanner with a “cube" geometry. We quantify by simulations the expected image resolution and Noise Equivalent Count Rates and compare them to the performance of the most efficient clinically used PET scanner, the High-Resolution Research Tomograph by Siemens. We bring up the issues of the image reconstruction for a scanner with high spatial resolution. We also present simulated brain images for \[$^{18}$F\]-FDG and \[$^{11}$C\]-PE2I tracer distributions. Results demonstrate the high potential of the CaLIPSO PET scanner for small animal and brain imaging where combination of high spatial resolution and efficiency is essential.'
author:
- 'Olga Kochebina, Sébastien Jan, Simon Stute, Viatcheslav Sharyy, Patrice Verrecchia, Xavier Mancardi, and Dominique Yvon, '
title: Simulations and Image Reconstruction for the High Resolution CaLIPSO PET Scanner for Brain and Preclinical Studies
---
=1
[Kochebina : Simulations and Image Reconstruction for CaLIPSO PET Scanner]{}
Positron Emission Tomography, Monte Carlo simulations, high image resolution, PET image reconstruction.
Introduction
============
spatial resolution Positron Emission Tomography (PET) is an advanced method for diagnosis and investigate mainly in neurology studies to exam neurodegenerative diseases such as Alzheimer’s, Parkinson’s, Huntington’s diseases or multiple sclerosis. According to the radioactive tracer distribution it is possible to monitor the activity of cells and various processes such as the transport of substances, gene expression and the interaction between a ligand and a receptor. It is also of interest in oncology for the detection of small tumors and small metastases. High image resolution imager is also useful in preclinical researches on small rodents. Clinical Magnetic Resonance Imaging (MRI) scanners provide high image resolution but are less sensitive to monitor of biochemical cellular activity. Thus, high resolution PET imagers would be an excellent complement to MRI for diagnosis and medical research. The foreseen CaLIPSO[^1] PET scanner [@CaLIPSO] is a high resolution imager for human brain and preclinical studies. It combines simultaneously a high detection efficiency, a 3D spatial resolution of 1 mm$^3$ (Full Width at Half Maximum, FWHM) and a coincidence resolving time of about 150 ps or better, allowing for a high image contrast. High performances of the proposed PET scanner are possible thanks to the concept of double detection illustrated in Fig \[fig1\], An elementary cell of the PET imager is filled with an innovative liquid, trimethyl bismuth (TMBi) [@CaLIPSO:TMBi], It contains 82% by weight of Bismuth, the highest-Z radioactively stable element. Consequently, TMBi allows a photoelectric conversion yield of $\sim$47% [@CaLIPSO:TMBi], better than any other detector used or proposed for PET imaging in nuclear medicine. Photoelectric photon conversion or Compton scattering generate a so-called “primary" electron in the liquid TMBi, which produces two types of signal. The electron emits Cherenkov light while propagating in the TMBi. However, the number of photons is quite low and light collection with efficient photomultipliers (PMT) is required. We plan to work with Micro-Channel Plate PMT (MCP-PMT) which provide also an excellent time resolution. The light detection [@CaLIPSO:optic] is used for the precise measurements of the interaction time. The same primary converted electron ionizes the medium and generates free electrons which drift along a strong electric field, pass through a Frisch grid and are collected by a pixelated detector. The pixel size used to collect ionization is 1 mm$^2$ which ensures a high spatial resolution. The measurement of the ionization drift time allows to calculate the third coordinate or depth of interaction (DOI) with 1 mm precision [@CaLIPSO], This ionization signal is used not only for the position estimate but also for energy measurements with a precision of $\sim10\%$ (FWHM) [@CaLIPSO],
![Principle scheme of the CaLIPSO detector. The primary electron created by 511-keV $\gamma$ produces Cherenkov light and ionizes the medium. Both signals are used to measure arrival time, 3D position and energy.[]{data-label="fig1"}](fig1-eps-converted-to.pdf){width="3.5in"}
Moreover, due to fast Cherenkov light emission and detection an excellent time resolution, coincidence resolving time $\sim$150 ps (FWHM), is also expected. This makes possible to use time of flight technique (TOF) to improve signal to noise ratio in final images. However, in this paper we let aside this property of the CaLIPSO scanner in the simulations as the proof of concept for the detection methods are still under validation with ionization and optical prototypes.
The main aim of CaLIPSO project is to develop high resolution PET scanner with reconstruction of the DOI. This problematic has been the subject of many developments during last decades. The main attempts we present in the following for the illustration of the effort that is already made.
In order to increase the spatial resolution and obtain DOI PET it was proposed to pixelaze scintillation crystals. For example, in [@Vandenbroucke; @Yamaya; @Yoshida] the authors suggested to use array modules built out of scintillation crystal of the size of 1 mm$^3$ or less. The foreseen submillimetric 3D resolution was illustrated on prototype measurements and Monte Carlo simulations. However, the main limitation of this method is a drop of detection efficiency with crystal size.
To overcome this limitation, monolithic scintillator crystals, LaBr3:Ce, coupled with digital silicon photomultipliers was proposed. In the first experimental characterization of such a TOF PET detector 1 mm spatial resolution [@Seifert] and coincidence resolving time of $\sim$160-180 ps (FWHM) [@vanDam] were demonstrated.
Completely different technology was also proposed to gain in the spatial resolution: CdTe semiconductor detectors. It has been shown that a PET scanner based on these detectors can have in transaxial plane 2.3 mm–4.8 mm resolution [@Shiga], Time resolution of such machine is low, 6.8 ns (FWHM), and TOF cannot be used. Thus, the image contrast is enhanced with another technique: a high energy resolution, 4.1% for 511 keV $\gamma$, is used to reduce the scatter coincidences background.
Another idea how to enhance spatial resolution without drop of detection efficiency is to use a liquid instead of crystals. For example, a PET scanner based on a liquid Xenon detectors [@Xenon] was an inspiration for CaLIPSO. The double signal detection is also used: scintillation light and ionisation signals. Studies on detector prototypes showed a submillimeter 3D spatial resolution and energy resolution better than 10% (FWHM).
For the high spatial resolution scanner, it is important to have low statistical noise in each voxel. One of the possibility to enhance the signal-to-noise ratio of the image is to use the TOF technique. In [@Korpar] authors propose to use the Cherenkov radiation instead of scintillation light as in this case the time resolution is not limited by the scintillation decay time. Using PbF$_{2}$ crystal coupled to MCP-PMT they demonstrated the time resolution of about 70 ps (FWHM) for a single detector [@Korpar], The drawback of this approach is a low number of the optical photons produced in the Cherenkov radiator, and thus, low intrinsic detection efficiency ($<$10%). The CaLIPSO detector also uses the Cherenkov radiation as explained above. The first prototypes demonstrated the possibility to reach a much higher efficiency of the detection, of the order of 35%, by using the optimized light collection in the TMBi liquid [@Ramos],
Although new proposed technologies and methods are promising, an existing benchmark scanner, the HRRT (High Resolution Research Tomograph) by Siemens [@HRRT], is already used for clinical applications in $\sim$ 20 PET centres around the world. It has a spatial resolution of approximately 2.3 mm to 3.2 mm (FWHM) in the transaxial direction. This machine without TOF uses two layers of scintillation crystals, LSO and LYSO, connected to photomultiplier tubes. This scanner combines high image spatial resolution with high sensitivity [@HRRT], In this paper we use this advanced clinical scanner for a comparison with the simulated CaLIPSO PET system.
We previously simulated the CaLIPSO imager with a HRRT-like geometry [@Kochebina], In this paper, we introduce a geometry for the scanner with a larger solid angle. Through the use of Monte Carlo simulations, we measure its sensitivity and transaxial spatial resolution and perform a preliminary 2D image quality evaluation. It is important to mention that the current status of the actual detector cell design, which is still ongoing, did not justify the costly effort for developing a complete 3D reconstruction software taking TOF into account. This is further discussed throughout the paper.
In “Materials and Methods" section we present the simulation approaches and the evaluation setups. “Results" section gives the Noise Equivalent Count Rates calculation, estimations for image resolution using capillary sources, reconstructed images of the Derenzo phantom and simulated brain images for two tracers, \[$^{11}$C\]-PE2I and \[$^{18}$F\]-FDG. In “Discussion" and “Conclusion" sections we give overview of the results and the perspectives of the CaLIPSO project.
Materials and Methods
=====================
{width="\textwidth"}
The Monte Carlo simulation of the full size PET scanner was performed with the GATE 7.1 software [@gate; @gate:2010], We have chosen cubic shape (Fig. \[fig2\]) as it is possible thanks to the accurate reconstruction of the DOI inside a scanner cell, thus, exclusion of parallax effects. The main motivations for a cube shape of the scanner are the minimization of dead zones and the manufacture simplification. Moreover, study in [@cube] has shown that the rectangular system with DOI capability has a higher signal-to-noise ratio for detection tasks and a lower bias at a given noise level for quantitation tasks than the same system without DOI. The CaLIPSO simulated scanner consists of 4 sectors by 5$\times$6 elementary modules (Fig \[fig3\]). This determines the axial, 354 mm, and “radial"[^2], 307 mm, sizes of the imager field of view (FOV). In the current mechanical design the elementary detection module is made from the alumina ceramic body filled with the liquid TMBi. The volume is sealed with the transparent sapphire window using the metallic tightening from the external side. The MCP-PMT is coupled to the sapphire window with an optical gel. The read-out ionization pad structure is fabricated on the ceramic substrate and used to seal the volume from the inner side of the scanner. The elementary module is separated in four equal cells by 87% reflecting ceramic light guides (thickness 1 mm) that focus the Cherenkov light on the photomultiplier surface. The module size, 59$\times$59 mm, is defined by the size of the selected PLANACON$^{TM}$ MCP-PMT by Photonis [@Photonis], The active part of MCP-PMT is 53$\times$53 mm, thus, dead zones are about 20%. Laboratory test demonstrated that it is possible to reach 90% efficiency for the 511 keV photons converted by the photo-ionization effects in TMBi [@Ramos], The module depth, 50 mm, allows to obtain a total detection efficiency more than 40%.
![Geometry of an elementary module used in the simulation for CaLIPSO PET. The module body is made from the alumina ceramic, sealed with the sapphire window from one side and metallized ceramic plate for the ionization readout from other side. The volume is filled with TMBi liquid and separated in four square cells by the white ceramic light guides.[]{data-label="fig3"}](fig3-eps-converted-to.pdf){width="3.5in"}
The GATE digital detection model uses a dedicated parametrized description of the detector response for the ionization and light signal readout. These semi-analytic models are calibrated using detector prototypes when they are available and detailed Geant4 simulations in order to apply the following detector response functions:
- the efficiency of the optical signal detection as a function of energy obtained in Geant4 prototype simulations and published earlier [@Ramos] illustrated in Fig \[fig4\];
- 10% (FWHM) [@CaLIPSO] Gaussian blurring for the measured energy;
- 1 mm (FWHM) [@CaLIPSO] Gaussian blurring for the spatial coordinates.
![Optical signal detection efficiency. Optical signal detection efficiency as a function of primary electron energy for the CaLIPSO PET scanner from Geant4 simulations. The red arrow corresponds to the most probable electron energy from 511 keV gamma photoelectric conversion. []{data-label="fig4"}](fig4-eps-converted-to.pdf){width="3.5in"}
Coincidence events are selected within a 3 ns time window keeping multiple pairs of signles registered in the same time window if they are not detected in the same sector. The dead time is not yet taken into account but we expect that it would not significantly change the results. The first estimation shows that the non-paralyzable dead time $\tau_{dead}\sim$5 $\mathrm{\mu}$s, corresponding to drift time and not the pixelated detector dead time. The corresponding saturation rate is equal to $1/\tau_{dead}\sim$200 kHz per $1/4$ of scanner cell. Thus, saturation occupancy for the full scanner would be $480\times200$kHz$\sim$96 MHz while typical rate for a brain scan is at the level of 1 MHz. We also do not yet take into account the TOF potential of the CaLIPSO PET scanner but we discuss its influence on the Noise Equivalent Count Rate estimations later. This important advantage property of CaLIPSO PET scanner will be added to the simulations after conclusion on detector cell design.
In order to reduce the computation time, instead of positron source generation we simulate two 511 keV $\gamma$ photons emitted in opposite directions. In tests where it is relevant we take into account:
- *acollinearity angle* introduced as a randomly chosen angle correction from Gaussian distribution with $\sigma=0.5$ degrees;
- *positron range* set as a blurring of the source position according to distribution and values from [@PositronRange],
Image reconstruction
--------------------
The simulated data are reconstructed using the Customizable and Advanced Software for Tomographic Reconstruction (CASToR) [@castor], The CaLIPSO PET scanner is expected to have a spatial resolution of 1 mm$^3$, thus, a pseudo detection element would have the same size. This means that the full volume of TMBi liquid of CaLIPSO scanner contains $1.6\times10^7$ such elements. Thus, the number of lines of response (LOR) is $\sim2.6\times10^{14}$, which is 5 orders of magnitude higher than for HRRT scanner. This amount of LORs makes sinogram based image reconstruction impossible. Thus, the list-mode OSEM algorithm is used along with a simple line projector [@siddon] without including any resolution modeling. We perform 10 iterations with 16 subsets. The reconstructed images have a pixel size of 0.25 mm$\times$0.25 mm$\times$1 mm. In order to reduce voxel fluctuations we apply Gaussian smoothing with $\sigma=0.5$ mm on final brain images. This smoothing is not applied in the spatial resolution tests.
The first steps in the reconstruction are the scanner normalisation calculation and the sensitivity map creation. This stage remains difficult again due to the large data quantity. For example, to calculate the full scanner image normalization we need to generate $\sim2.6\times10^{16}$ events. Therefore, in this paper we use a simplified approach for the proof of concept and to illustrate the potential image spatial resolution: generate and reconstruct only one 2D axial slice. This approach allows to characterize image resolution specificities before having the complete and dedicated CaLIPSO image reconstruction package. For our first results we do not simulate and take into account the attenuation medium and we do not use the TOF information. Finally, only true coincidences are reconstructed as the reconstructed images are used for potential high resolution illustration.
Performance tests
-----------------
In our case, the main parameters of interest are the Noise Equivalent Count Rates (NECR) and image resolution.
### Noise Equivalent Count Rates {#sec:NECR}
In order to have an estimate of the expected image signal to noise ratio we use the following calculations of NECR parameter: $$\begin{aligned}
NECR=\frac{T^2}{T+S+2R},
\label{eq:NECR}\end{aligned}$$ where $T$, $S$ and $R$ are numbers of true, scatter and random coincidences respectively. The factor 2 corresponds to the on-line subtraction method for randoms correction, which is used in [@HRRT], reference in our comparisons. For consistency, we apply the same factor in the calculations. For the same reason, we also apply the same energy cut above 350 keV for coincidences selection.
The simulation is performed with a cylindrical phantom NEMA-1994 [@NEMA:1994] with diameter and height of 200 mm. The phantom placed in the center of the FOV is filled with water and uniformly distributed radioactivity. The scatter, random, true fractions and NECR values are obtained for different activity concentrations in a range from 0.6 kBq/cc to 50 kBq/cc.
The signal to noise ratio drops when spatial resolution increases as image voxel size becomes smaller and voxel variability increases. However, this ratio can be enhanced with time of flight method. To estimate TOF gain it is possible to use the the first order approximation [@Conti2006]: $$\begin{aligned}
NECR_{TOF}=\frac{D}{\Delta x}NECR,
\label{eq:NECRTOF}\end{aligned}$$ where $D$ is the size of phantom and $\Delta x$ is is the localization uncertainty $\Delta x=c\Delta t/2$ related to the time resolution $\Delta t$ and the speed of light $c$. However, from the various studies estimating the TOF gain in terms of signal to noise ratio (SNR) [@Tomitani; @Conti2006; @Conti2013; @Conti2015] depends on the way TOF is implemented in the reconstruction, the way random and scattered coincidences are handled and the actual count rates. We estimate the range of the NECR gain basing on approaches from these papers.
### Spatial resolution {#sec:reso}
For the spatial resolution tests the full process of image reconstruction from the simulated data is applied. We perform three tests with \[$^{18}$F\] tracer:
1. The image resolution from a sample with $\sim8\cdot10^{6}$ detected coincidences we simulate several capillary sources of 0.2 mm diameter placed at different radial positions of the FOV of the CaLIPSO scanner.
2. To find the separation power between two capillary sources of 0.2 mm diameter we simulate them close to each other, at 1 mm and 2 mm distances between centers of the sources. We reconstruct $\sim5.5\cdot10^{6}$ coincidences.
3. A qualitative test is done with the one slice from mini Derenzo phantom [@Derenzo] placed in a middle of FOV using $\sim1.5\cdot10^{8}$ detected coincidences. The phantom is a plastic cylinder of 45 mm diameter with hot rods of different diameters: 1.2 mm, 1.6 mm, 2.4 mm, 2.8 mm, 4.0 mm and 4.8 mm The distances between centers of the rods are two times their diameters.
These tests aim to illustrate the high spatial resolution of the scanner. Thus, we reconstruct images containing true coincidences without quantification scaling.
Examples of human brain scans {#sec:brain}
-----------------------------
The simulations for human brain scans are done using one slice of the brain Zubal phantom [@zubal] with a voxel size of 1.1 mm$\times$1.1 mm$\times$1.4 mm and with 18 different segmented brain structures. The activity distributions for two different tracers, \[$^{11}$C\]-PE2I and \[$^{18}$F\]-FDG are defined from activity measurements from real HRRT scans.
The aim of these tests is to illustrate the high image spatial resolution of the foreseen CaLIPSO PET scanner. Thus, we reconstruct high statistics images, i.e. containing $1.5\cdot10^{8}-1.6\cdot10^{8}$ true coincidences, and no quantification scaling is performed.
Results
=======
Noise Equivalent Count Rates {#noise-equivalent-count-rates}
----------------------------
Fig \[fig5\], left shows the curves for true, random and scatter coincidences as functions of activity and Fig \[fig5\], right corresponds to the NECR curves according to Eq \[eq:NECR\] in “Materials and Methods" section. One can see that the NECR values for the same activity are higher for the CaLIPSO system than for the HRRT. Thus, we expect to have an enhanced image contrast with the proposed CaLIPSO imaging system for the same voxel size. Additional increase of signal to noise ratio would be obtained with TOF technique. From Eq \[eq:NECRTOF\] a conservative estimate of $\Delta t=150$ ps gives $\Delta x=2.25$ cm and phantom size is $D$ = 20 cm. Thus, from the various studies estimating the TOF gain in terms of SNR [@Tomitani; @Conti2006; @Conti2013; @Conti2015], it is expected to be in the range 5 to 10. This represents a significant increase of the signal to noise ratio.
{width="\textwidth"}
Image resolution
----------------
### Test with capillary sources.
The results obtained as explained in “Materials and Methods" section for image resolution of capillary sources in radial and tangential directions are presented in Fig \[fig6\], We obtain image resolution of $\sim$1 mm (FWHM) with a slight degradation ($\sim$10% relative error) toward peripheral FOV regions. In Fig \[fig6\] we also compare the results with HRRT system [@HRRT] performance. One can observe that besides a lower image resolution the degradation toward peripheral FOV region is much higher for HRRT scanner.
![Image resolution. Image resolution for the foreseen CaLIPSO scanner and its comparison to measurements from the HRRT scanner (Dashed curves are redrawn from [@HRRT] by extracting data points with WebPlotDigitizer application [@WebPlotDigitizer]).[]{data-label="fig6"}](fig6-eps-converted-to.pdf){width="3.5in"}
{width="\textwidth"}
### Test with closely placed capillary sources
The results for the test proposed in “Materials and Methods" section with two capillaries placed closely one to each other are presented as a reconstructed image in Fig \[fig7\], left and as a 1D profile in Fig \[fig7\], right. One can notice that the two sources at 2 mm distance can be easily differentiated with a peak-valley ratio of $\sim$3 and a separation of $\sim$5 standard deviations. For the distance of 1 mm the difference between the peaks and the valley is about $\sim$1.2 and the separation is $\sim$2.4 standard deviations.
### Test with the Derenzo phantom
The generated distribution is illustrated in Fig \[fig8\], left, while in the right figure we show the reconstructed image. One can observe that the smallest rods of 1.2 mm diameter can be clearly separated (Fig \[fig8\], middle). Fig \[fig8\], right shows the HRRT image obtained using real measurements.
{width="\textwidth"}
Examples of human brain scans {#examples-of-human-brain-scans}
-----------------------------
The brain simulations for \[$^{11}$C\]-PE2I and \[$^{18}$F\]-FDG tracers are presented in Fig \[fig9\] and \[fig10\] respectively. Left figures correspond to the simulated activity distributions and right ones to the reconstructed images.
{width="\textwidth"}
{width="\textwidth"}
For the \[$^{11}$C\]-PE2I tracer (Fig \[fig9\]) one should keep in mind that the positron range of $^{11}$C is about 1 mm (FWHM), which is compatible with the scanner spatial resolution. This means that for such type of tracers the images can not be refined further by improving the machine spatial resolution.
For the \[$^{18}$F\]-FDG tracer the situation is different. The positron range of $^{18}$F is $\sim$0.6 mm (FWHM). Therefore, the high image resolution of the foreseen CaLIPSO scanner plays a key role, enabling a better separation of the different brain structures.
In Fig \[fig9\] and \[fig10\] the TOF techniques for image reconstruction are not taken into account and, thus, further improvement of image quality due to noise reduction would be possible. However, the simulation and reconstruction are performed without taking into account the attenuation, scatter and randoms contributions which would degrade the quality of the images. Nonetheless, from the results from NECR we expect that it would not be an issue.
Discussion
==========
In this article we present the simulation study of the performance for the foreseen CaLIPSO scanner. The results exhibit the high potential of such a system. However, the studies with reconstructed images are restricted to one 2D slice due to the large dimensionality of the problem. It is worth noting that this problem is becoming more recurrent for new PET scanners with a number of detecting elements few orders of magnitude higher than in state-of-the art clinical scanners (either for very high spatial resolution or for extended axial FOV). We believe that it should not be an issue with the development of appropriate methodologies (beyond the scope of this paper) and increasing computer performances.
Several other parameters let aside in this paper should be taken into account in future studies, such as dead time of detection modules, TOF information, attenuation and scatter corrections of images and quantification scaling including normalisation.
Another point to discuss is the complexity of the detection system and technology. The optical detection prototype is now operative [@Ramos], Currently we work on its improved version. The ionization detection prototype is under development and there are still several technological problems to be solved, including the ultrapurification of TMBi liquid and detection of low amplitude signals.
Conclusion
==========
In this paper we present simulation results of the foreseen CaLIPSO PET scanner. The simultaneous detection of light and charge signals in CaLIPSO detectors leads to promising performances for a PET imager. Thanks to this detection technology we expect an image resolution of 1 mm (FWHM) in the 3 dimensions as demonstrated in this work by simulated tests with capillary sources and the Derenzo phantom. The obtained NECR curve shows that the expected image signal to noise ratio is almost twice better than for the HRRT. This can be further increased by a factor 5-10 thanks to a TOF capability with a coincidence resolving time of 150 ps (FWHM).
Finally, reconstructed images of \[$^{11}$C\]-PE2I and \[$^{18}$F\]-FDG brain scans confirm the potential of the high spatial resolution expected with the CaLIPSO PET scanner.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported in part by the "Programme interdisciplinaire Technologie pour la Sante” of Commissariat á l’Energie Atomique et aux Energies Alternatives. Also authors would thank the support of The Neuropoole de Recherche Francilien and Laboratoire d’excellence Physique des deux infinis et des Origines.
[11]{}
Yvon et al. *CaLIPSO: An novel detector concept for PET imaging*, IEEE TNS 2014;61:60–66.
Verrecchia P, Ramos E, Yvon D, Tauzin G and Reithinger V *CaLIPSO: TMBi properties for particles detection*, IEEE-MIC conference record 2012;Nl-116:260–265.
Ramos et al. *Trimethyl Bismuth optical properties for particle detection and the CaLIPSO detector*, IEEE TNS 2015;62-3:1326–1335.
Vandenbroucke A, Foudray M K A,Olcott P D and Levin C S *Performance characterization of a new high resolution PET scintillation detector*, Phys. Med. Biol. 2010;55:5895–5911.
Yamaya T et al. *A SiPM-based isotropic-3D PET detector X’tal cube with a three-dimensional array of 1 mm3 crystals*, Phys. Med. Biol. 2011;56:6793–6807.
Yoshida E et al., *Spatial resolution limits for the isotropic-3D PET detector X’tal cube* Nucl. Instr. Meth. A 2013;728:107–111.
Seifert S, van der Lei G, van Dam HT and Schaart DR *First characterization of a digital SiPM based time-of-flight PET detector with 1 mm spatial resolution*, Phys. Med. Biol. 2013;58:3061–3074.
van Dam HT, Borghi G, Seifert S and Schaart DR *Sub-200 ps CRT in monolithic scintillator PET detectors using digital SiPM arrays and maximum likelihood interaction time estimation*, Phys. Med. Biol. 2013;58:3243–3257.
Shiga T et al. *A New PET Scanner with Semiconductor Detectors Enables Better Identification of Intratumoral Inhomogeneity*, J Nucl Med. 2009;50:148–155.
Miceli A et al. *Liquid Xenon Detectors for Positron Emission Tomography*, Journal of Physics: Conference Series 2011;312:062006.
Korpar S, Dolenec R, Krizan P, Pestotnik R, Stanovnik A *Study of TOF PET using Cherenkov light*, Physics Procedia 2012;37:1531–1536.
E. Ramos et al., *Efficient and fast 511-keV $\gamma$ detection through Cherenkov radiation: the CaLIPSO optical detector*, Journal of Instrumentation 2016;11:P11008.
de Jong HWAM, van Velden FHP, Kloet RW, Buijs FL, Boellaard R and Lammertsma AA *Performance evaluation of the ECAT HRRT: an LSO-LYSO double layer high resolution, high sensitivity scanner*, Phys. Med. Biol. 2007;52:1505–1526.
WebPlotDigitizer application,\
[`http://arohatgi.info/WebPlotDigitizer/`]{}
Kochebina et al. *Simulation for CaLIPSO PET Scanner*, The IEEE conference proceedings;2015:PaperID M6A2-3.
Jan S et al. *GATE: a simulation toolkit for PET and SPECT*, Phys. Med. Biol. 2004;49:4543–4561.
Jan S et al. *GATE V6: a major enhancement of the GATE simulation platform enabling modelling of CT and radiotherapy*, Phys. Med. Biol. 2010;56:881–901.
Qi J, Klein GJ, Huesman RH, Member S *Image Properties of List-Mode Likelihood Reconstruction for a Rectangular Positron Emission Mammograph With DOI Measurements*, IEEE Trans Nucl Sci. 2001;48(4):1343–1349.
Photonis `https://www.photonis.com/uploads/datasheet/`\
`pd/PLANACON-8x8-datasheet.pdf`
Cal-Gonzalez J, Herraiz JL, Espana S, Corzo PMG, Vaquero JJ, Desco M and Udias JM *[P]{}ositron range estimations with PeneloPET*, Phys. Med. Biol. 2013;58:5127–5152.
CASToR. Customizable and Advanced Software for Tomographic Reconstruction *Open source platform under development*, Available from: `http://www.castor-project.org`.
Siddon RL *Fast calculation of the exact radiological path for a three-dimensional CT array*, Med. Phys. 1985;12:252–255.
National Manufacturers Association NEMA Standards Publication NU 2-1994 (Washington, DC: National Electrical Manufacturers Association).
Conti M *Effect of Randoms on Signal-to-Noise Ratio in TOF PET*, IEEE TNS 2006;53-3:1188–1193.
Tomitani T *Image reconstruction and noise evaluation in photon time-of-flight assisted positron emission tomography*, IEEE TNS 1981;28-6:4582–4589.
Conti M, Eriksson L, and Westerwoudt V *Estimating Image Quality for Future Generations of TOF PET Scanners*, IEEE TNS 2013;60-1:87–94.
Eriksson L, and Conti M *Randoms and TOF gain revisited*, Phys. Med. Biol. 2015;60:1613–1623.
Data Spectrum Corporation *Mini and Micro Deluxe Phantom and Hot Spot Insert*, Available from: `http://www.spect.com/pub/`\
`Mini_and_Micro_Deluxe_Phantom_and_Hot_Spot_Insert.pdf`.
Zubal IG, Harrell CR, Smith EO, Smith AL and Krischlunas P *High resolution, MRI-based, segmented, computerized head phantom*, Available from: `http://noodle.med.yale.edu/zubal/info.htm`.
[^1]: French acronym for Liquid Ionization Calorimeter, Scintillation Position Organometallic
[^2]: defined by the side of the cube
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Multiplicity fluctuations in limited segments of momentum space are calculated for a classical pion gas within the statistical model. Results for the grand canonical, canonical, and micro-canonical ensemble are obtained, compared and discussed. We demonstrate that even in the large volume limit correlations between macroscopic subsystems due to energy and momentum conservation persist. Based on the micro-canonical formulation we make qualitative predictions for the rapidity and transverse momentum dependence of multiplicity fluctuations. The resulting effects are of similar magnitude as the predicted enhancement due to a phase transition from a quark-gluon plasma to a hadron gas phase, or due to the critical point of strongly interacting matter, and qualitatively agree with recently published preliminary multiplicity fluctuation data of the NA49 SPS experiment.'
author:
- Michael Hauer
title: |
Multiplicity Fluctuations in Limited Segments of Momentum Space\
in Statistical Models
---
Introduction {#Intro}
============
The statistical model has been, for a long time, successfully applied to fit experimental data on mean hadron multiplicities in heavy ion collision experiments over a wide range of beam energies and system sizes. For recent reviews see [@FOCley; @FOBeca; @FOPBM; @FORafe]. So naturally the question arises whether the statistical model is able to describe event-by-event fluctuations of these observables as well. And indeed, a first comparison suggests that this might be possible for the sample of most central events. Global conservation laws, imposed on a statistical system, lead, even in the large volume limit, to suppressed fluctuations. The multiplicity distributions of charged hadrons recently reported [@NA49_fluc] by the NA49 SPS experiment are systematically narrower than a Poissonian reference distribution. This could be interpreted [@MCEvsData] as effects due to energy and charge conservation in a relativistic hadronic gas.
Multiplicity fluctuations are usually quantified by the ratio of the variance of a multiplicity distribution to its mean value, the so-called scaled variance. In statistical models there is a qualitative difference in the properties of mean value and scaled variance. In the case of the mean multiplicity results obtained within the grand canonical ensemble (GCE), canonical ensemble (CE), and micro-canonical ensemble (MCE) approach each other in the large volume limit. One refers here to as the thermodynamic equivalence of these ensembles. It was recently found [@CEfluc1] that corresponding results for the scaled variance are different in different ensembles, and thus this observable is sensitive to conservation laws obeyed by a statistical system.
The growing interest in the experimental and theoretical study of fluctuations in strong interactions (see e.g., reviews [@reviewfluc]) is motivated by expectations of anomalies in the vicinity of the onset of deconfinement [@ood] and in the case when the expanding system goes through the transition line between a quark-gluon plasma and a hadron gas phase [@phasetrans]. In particular, a critical point of strongly interacting matter may be accompanied by a characteristic power-law pattern in fluctuations [@critpoint]. A non-monotonic dependence of event-by-event fluctuations on system size and/or center of mass energy in heavy ion collisions would therefore give valuable insight into the phase diagram of strongly interacting matter. Provided the signal survives the subsequent evolution and hadronization of the system (see also [@recomb]). Therefore, in order to asses the discriminating power of proposed measures, for a recent review see [@reviewfluc2], one should firstly study properties of equilibrated sources [@MCEvsData; @res; @VolDep; @vdw] and quantify ‘baseline‘ (or thermal/statistical) fluctuations. Apart from being an important tool in an effort to study a possible critical behavior, the study of fluctuations within the statistical model constitutes also a further test of its validity.
In this paper we make detailed predictions for the momentum space dependence of multiplicity fluctuations. We show that energy and momentum conservation lead to a non-trivial dependence of the scaled variance on the location and magnitude of the observed fraction of momentum space. These predictions can be tested against existing and future data from the heavy ion collision experiments at the CERN SPS and BNL RHIC facilities.
The paper is organized as follows: In section \[model\] we briefly introduce our model. In section \[GCECE\] we consider multiplicity distributions in a limited region of momentum space in GCE and CE. For the MCE we follow, in section \[MCE\], the procedure of Ref.[@clt] and show how to calculate the width of the corresponding distributions in the large volume limit. We revisit the so-called ‘acceptance scaling‘ previously suggested as an approximate implementation of experimental acceptance in section \[Results\]. Technical details of the calculations are presented in the Appendix. Concluding remarks and a summary in sections \[Remarks\] and \[Summary\] close the paper.
The Model {#model}
=========
The ideal Boltzmann $\pi^+$ $\pi^-$ $\pi^0$ gas serves as the standard example throughout this paper, while the main subject of investigation is the multiplicity distribution $P(N_{\Omega})$ of particles with momenta inside a certain segment $\Omega$ of momentum space. Calculations are done for the three standard ensembles GCE, CE, and MCE. For the sake of argument we will assume that we only want to measure $P(N_{\Omega}^-)$, i.e. the probability distribution of negatively charged pions in a limited segment $\Omega$ of momentum space. Hence $\pi^-$ with momenta inside $\Omega$ are observed, while $\pi^-$ inside the complementary segment $\bar \Omega$ are not observed. $\pi^+$ and $\pi^0$ are never detected. In GCE and CE the presence of $\pi^0$ as a degree of freedom is of no relevance, while in MCE it constitutes a heat bath for the remaining system. For consistency we use the same system throughout this discussion.
In order to keep the model simple, we assume a static homogenous fireball. Our considerations therefore exclude collective motion, i.e. flow, and resulting momentum spectra are purely thermal. We also omit resonance decay contributions in this work. The spectra presented in Fig. \[spectra\] are normalized to the total $\pi^-$ yield in GCE and CE. Thus they are the same in both ensembles. In MCE one expects in the large volume limit only small deviations from Boltzmann spectra. None of the forthcoming arguments are affected by this.
In the following we will use the transverse momentum and rapidity spectra presented in Fig. \[spectra\] to construct bins $\Omega_i = \Delta {p_T}_i =
\left[p_{T_i},p_{T_{i+1}} \right]$ (left), or $\Omega_i = \Delta y_i =
\left[y_i,y_{i+1} \right]$ (right), as indicates by the drop-lines.
In section \[GCECE\] we calculate the multiplicity distributions $P(N_{\Omega})$ for arbitrary segments $\Omega$ for the ideal Boltzmann GCE and CE. To characterize the distribution one can calculate its (raw) moments $\langle N_{\Omega}^n \rangle$ from: $$\label{Moments}
\langle N_{\Omega}^n \rangle ~=~ \sum \limits_{N_{\Omega}=0}^{\infty}
~N_{\Omega}^n~ P \left(N_{\Omega} \right)~.$$ A convenient measure for the width of a distribution is the scaled variance: $$\label{ScaledVar}
\omega_{\Omega} ~\equiv~ \frac{\langle N_{\Omega}^2 \rangle - \langle N_{\Omega}
\rangle^2}{\langle N_{\Omega} \rangle} ~.$$ In order to remove simple scaling effects, the bin sizes or segments are chosen such that each bin or segment contains the same fraction $q= \langle
N_{\Omega} \rangle ~/~ \langle N_{4\pi} \rangle$ of the total yield (compare Eq.(\[ScaledVar\])). Here $\langle N_{\Omega} \rangle$ denotes the average particle number in the momentum space segment $\Omega$, and $\langle N_{4\pi}
\rangle$ denotes the average total ($4\pi$ integrated) multiplicity. The effect of finite acceptance can approximately be taken into account by [@CEfluc1]: $$\label{accscaling}
\omega_{q} ~=~ 1 + q \left(\omega_{4\pi}-1 \right)~,$$ where $\omega_{4\pi}$ assumes the ideal situation when all particles are detected, while $\omega_{q}$ assumes that particles are detected with probability $q$ regardless of their momentum. Hence Eq.(\[accscaling\]) holds when particles are assumed to be uncorrelated in momentum space. In the limit $q\rightarrow 0$ one observes a random distribution with $\omega_q
\rightarrow 1$, i.e. a Poissonian, while when $q\rightarrow 1$ one sees the real distribution with width $\omega_q \rightarrow \omega_{4\pi}$. In this work we take explicitely correlations due to globally conserved charge (CE), and energy-momentum (MCE) into account and compare the results to Eq.(\[accscaling\]).
Grand Canonical and Canonical Ensemble {#GCECE}
======================================
Grand Canonical Ensemble
------------------------
In the GCE, both, heat and charge bath are assumed to be infinite. And thus neither charge, energy nor momentum are conserved exactly. Temperature $T$ and charge chemical potential $\mu$ regulate average energy and charge density in a system of volume $V$. Usually it is said that charge, energy and momentum are conserved in the average sense and fluctuations about an equilibrium value are allowed. Apart form Bose and Fermi effects [@Qstats] particles are therefore uncorrelated in momentum space. However this example serves as an illustration for the following CE and MCE calculations. We start by decomposing the Boltzmann single particle partition function $z^-\left(\phi_{N_{\Omega}}\right)$ of $\pi^-$ into two parts, $$\begin{aligned}
z^-\left(\phi_{N_{\Omega}}\right)= z^-_{\Omega} \left(\phi_{N_{\Omega}}
\right) + z^-_{\bar \Omega} &=& \frac{gV}{\left(
2\pi\right)^3} \int \limits_{\Omega} d^3 p ~
e^{-\frac{\varepsilon+\mu}{T}}~ e^{i \phi_{N_{\Omega}}} +
\frac{gV}{\left( 2\pi\right)^3} \int \limits_{\bar \Omega} d^3 p ~
e^{-\frac{\varepsilon+\mu}{T}} , \end{aligned}$$ where the single particle energy $\varepsilon = \sqrt{p^2+m^2}$, and $m$, and $g$ are mass and degeneracy factor of $\pi^-$ respectively. Only for momentum states inside the momentum space region $\Omega$ we introduce additionally a Wick-rotated fugacity $\exp \left(i \phi_{N_{\Omega}} \right)$. For the positive and neutral pion (which we do not want to detect in our example) we write: $$\begin{aligned}
z^+ ~=~ \frac{gV}{\left( 2\pi\right)^3} \int d^3 p ~
e^{-\frac{\varepsilon-\mu}{T}}~,
\qquad \qquad \textrm{and} \qquad \qquad
z^0 ~=~ \frac{gV}{\left( 2\pi\right)^3} \int d^3 p ~
e^{-\frac{\varepsilon}{T}}~.\end{aligned}$$ The value of the single particle partition function, for instance of the neutral pion, is given by: $$\label{aveN}
z^0=\langle N^0 \rangle = \frac{gV}{2\pi} m^2 T K_2 \left( \frac{m}{T}\right).$$ For the sake of simplicity we assume equal masses for all pions. To obtain the GCE multiplicity distribution for $N_{\Omega}$ in a momentum space segment $\Omega$ we use the Fourier integral over the generalized GCE partition function $ \mathcal{Z} \left( \phi_{N_{\Omega}}\right)=\exp \left[ z^-_{\Omega}
\left( \phi_{N_{\Omega}} \right) + z^-_{\bar{ \Omega}} + z^+ + z^0 \right] $, normalized by the GCE partition function: $$\begin{aligned}
\label{GCEPDF}
P_{gce} \left(N_{\Omega} \right) ~\equiv~ Z^{-1}_{gce} \times \int
\limits_{-\pi}^{\pi} \frac{d\phi_{N_{\Omega}}}{2\pi} ~ e^{-iN_{\Omega}
\phi_{N_{\Omega}} } ~ \mathcal{Z} \left( \phi_{N_{\Omega}}\right) ~=~
\frac{\left(z^-_{\Omega}\right)^{N_{\Omega}}}{N_{\Omega}!}
\exp \left[- z^-_{\Omega} \right]~,\end{aligned}$$ where the system partition function is given by $ Z_{gce} \equiv \mathcal{Z}
\left( \phi_{N_{\Omega}} = 0\right) $, and $z^-_{\Omega} =
z^-_{\Omega} \left(\phi_{N_{\Omega}}=0 \right)$. Independent of the shape or size of $\Omega$ we find a Poissonian for the multiplicity distribution Eq.(\[GCEPDF\]). Thus, using Eq.(\[ScaledVar\]), one finds for the scaled variance $\omega^{gce}_{\Omega} = 1$, since $\langle
N_{\Omega} \rangle = z^-_{\Omega}$, and $\langle N^2_{\Omega}
\rangle = \langle N_{\Omega} \rangle^2 + \langle N_{\Omega} \rangle$.
For Bose and Fermi statistics one does not expect a Poisson distribution and (in particular when the chemical potential is large) deviations from a Poissonian can be large. Thus one expects also deviations from Eq.(\[accscaling\]) when considering only finite acceptance.
Canonical Ensemble
------------------
In the CE the heat bath is still assumed to be infinite, while we remove the charge bath and drop the chemical potential. Thus, we introduce a further Wick-rotated fugacity $\mu/T \rightarrow i \phi_Q $ into the single particle partition functions to account for global (however not in the momentum space segment $\Omega$) conservation of electric charge $Q$. Particles in $\Omega$ are therefore correlated, due to the condition of fixed net-charge, with a finite charge bath composed of $\pi^+$ and unobserved $\pi^-$. We again split the single particle partition function for $\pi^-$ into an observed, $z^-_{\Omega}\left(\phi_{N_{\Omega}},\phi_Q\right)$, and an unobserved part, $z^-_{\bar \Omega} \left(\phi_Q\right)$, $$\begin{aligned}
z^-\left(\phi_{N_{\Omega}},\phi_Q\right) = z^-_{\Omega}
\left(\phi_{N_{\Omega}},\phi_Q\right) + z^-_{\bar \Omega}
\left(\phi_Q\right) = \frac{gV}{\left(
2\pi\right)^3} \int \limits_{\Omega} d^3 p ~
e^{-\frac{\varepsilon}{T}} e^{-i \phi_Q} e^{i \phi_{N_{\Omega}}} +
\frac{gV}{\left( 2\pi\right)^3}
\int \limits_{\bar \Omega} d^3 p ~ e^{-\frac{\varepsilon}{T}} e^{-i \phi_Q},\end{aligned}$$ while we do not want to measure $\pi^+$ and $\pi^0$, and thus: $$\begin{aligned}
z^+ \left(\phi_Q\right) ~=~ \frac{gV}{\left(
2\pi\right)^3} \int d^3 p ~ e^{-\frac{\varepsilon}{T}} e^{+ i \phi_Q}~,
\qquad \qquad \textrm{and} \qquad \qquad
z^0 ~=~ \frac{gV}{\left(
2\pi\right)^3} \int d^3 p ~ e^{-\frac{\varepsilon}{T}} .\end{aligned}$$ The normalization of the CE multiplicity distribution is given by the CE system partition function $Z_{ce}$, i.e. the number of all micro states with fixed charge Q, $Z^{ce} = I_Q\left(2z \right) \exp(z^0)$, where $I_Q$ is the modified Bessel function. The multiplicity distribution of $N_{\Omega}$ in a momentum space segment $\Omega$, while charge $Q$ is globally conserved, can be obtained from Fourier integration of the generalized GCE partition function $\mathcal{Z} \left( \phi_{N_{\Omega}}, \phi_Q \right) = \exp \left[ z^-_{\Omega}
\left( \phi_{N_{\Omega}}, \phi_Q \right)~+~ z^-_{\bar{ \Omega}}
\left( \phi_Q \right) + z^+ \left( \phi_Q \right) + z^0\right] $, over both angles $\phi_Q$ and $\phi_{N_{\Omega}}$: $$\begin{aligned}
P_{ce} \left(N_{\Omega}\right) &\equiv& Z_{ce}^{-1} \times
\int \limits_{-\pi}^{\pi} \frac{d\phi_{N_{\Omega}}}{2\pi} \int
\limits_{-\pi}^{\pi} \frac{d\phi_{Q}}{2\pi} ~ e^{-iN_{\Omega}
\phi_{N_{\Omega}} } ~ e^{-i Q \phi_{ Q} } ~\mathcal{Z} \left(
\phi_{N_{\Omega}}, \phi_Q \right) \\
&=& I_Q^{-1}\left(2z \right) \times
\frac{\left(z^-_{\Omega}\right)^{N_{\Omega}}}{N_{\Omega}!} ~ \sum
\limits_{a=0}^{\infty} ~ \frac{\left(z^-_{\bar \Omega}\right)^{a}}{a!} ~
\frac{z^{Q+N_{\Omega}+a}}{\left(Q+N_{\Omega}+a\right)!}~,\end{aligned}$$ where in CE $z^-_{\Omega} = z^-_{\Omega} \left(\phi_{N_{\Omega}}=\phi_Q=0
\right)$, $z^-_{\bar \Omega} = z^-_{\bar \Omega} \left(\phi_Q=0 \right)$, and $z= z^+\left(\phi_Q=0\right)=z^0$. For the respective first two moments one finds from Eq.(\[Moments\]): $$\begin{aligned}
\langle N_{\Omega} \rangle = z^-_{\Omega } ~ \frac{I_{Q+1} \left(2z
\right)}{I_Q \left(2z \right)}~, \qquad \textrm{and} \qquad
\langle N^2_{\Omega} \rangle = \left( z^-_{\Omega }\right)^2 ~\frac{I_{Q+2}
\left(2z \right)}{I_Q \left(2z \right)} + z^-_{\Omega }~ \frac{I_{Q+1}
\left(2z\right)}{I_Q \left(2z \right)}~. \end{aligned}$$ Thus, we obtain the well known canonical suppression of yields [@CEyield; @CEfits; @CEtransport; @RateEqYield] and fluctuations [@CEfluc1; @RateEqFluc]. The result, however, is completely independent of the position of the segment $\Omega$. And therefore the scaled variance, Eq.(\[ScaledVar\]), takes the form: $$\begin{aligned}
\omega^{ce}_{\Omega} = 1 + z^-_{\Omega} ~\left[ \frac{I_{Q+2}
\left(2z\right)}{I_{Q+1} \left(2z \right)} -\frac{I_{Q+1}
\left(2z\right)}{I_Q \left(2z \right)} \right]~, \qquad \textrm{and}
\qquad \omega^{ce}_{4\pi} = 1 + z ~ \left[ \frac{I_{Q+2}
\left(2z\right)}{I_{Q+1} \left(2z \right)} -\frac{I_{Q+1}
\left(2z\right)}{I_Q \left(2z \right)} \right]~,\end{aligned}$$ where $\omega_{\Omega}$ is the width of $P_{ce}(N_{\Omega})$, i.e. the multiplicity distribution of $\pi^-$ with momenta inside $\Omega$, while $\omega_{4\pi}$ is the width of the corresponding distribution when $\Omega$ is extended to the full momentum space. It can immediately be seen that this formula is consistent with acceptance scaling, Eq.(\[accscaling\]), $\omega_{\Omega} ~=~ 1 + q \left(\omega_{4\pi} -1 \right)$, if $q \equiv
z^-_{\Omega}/z$. Generally we find $\omega^{ce}_{4\pi} < \omega^{ce}_{\Omega}
< \omega^{gce}=1$. In the limit of $z^-_{\Omega}/z \rightarrow 0$ we approach the Poisson limit of a ‘random‘ distribution with $\omega = 1$, i.e. the observed part of the system is embedded into a much larger charge bath and the GCE is a valid description.
Micro-Canonical Ensemble {#MCE}
========================
For the MCE an analytical solution seems to be out of reach presently, so we use instead the asymptotic solution, applicable to large systems, derived in [@clt]. In order to avoid unnecessary repetition of calculations, we will only give a general outline here, and refer the reader for a detailed discussion to Ref.[@clt]. It should be mentioned that this method be would be also applicable to systems of finite spatial extention, provided the average particle number in a given momentum space bin exceeds roughly $\langle N_{\Omega}\rangle \gtrsim 5$. In this work we confine ourselves to large systems and try to asses the general trends.
The basic idea is to define the MCE multiplicity distribution in terms of a joint GCE distribution of multiplicity, charge, energy, momentum, etc. The MCE multiplicity distribution is then given by the (normalized) conditional probability in the GCE to find a number $N_{\Omega}$ of particles in a segment $\Omega$ of momentum space, while electric charge $Q$, energy $E$, and three momentum $\vec P$ are fixed. Therefore we will keep temperature and chemical potentials as parameters to describe our system. Effective temperature and effective chemical potential, i.e. Lagrange multipliers, can be determined by demanding that the GCE partition function is maximized for a certain equilibrium state $(Q,E,\vec P)$. This requirement is entirely consistent [@clt] with the usual textbook definitions of $T$ and $\mu$ in MCE and CE through differentiation of entropy and Helmholtz free energy with respect to conserved quantities. In principle we would have to treat all conservation laws on equal footing [@MCEmagic], and thus introduce Lagrange multipliers for momentum conservation as well. However here we are only interested in a static source, thus $\vec P = \vec 0$, and the relevant parameters are equal to zero.
In the large volume limit energy, charge, and particle density in the MCE will correspond to GCE values. This is required by the thermodynamic equivalence of ensembles for mean quantities. MCE and CE partition functions are generally obtained from their GCE counterpart by multiplication with delta-functions, which pick out a set of micro states consistent with a particular conservation law. Here it will be of considerable advantage to use Fourier representations of delta-functions, similar to the treatment in Section \[GCECE\]. This method could be considered to be a Fourier spectral analysis of the generalized GCE partition function [@clt].
The normalized conditional probability distribution of multiplicity $N_{\Omega}$ can be defined by the ratio of the values of two partition functions: $$\label{MCEprob}
P_{mce}(N_{\Omega})~\equiv~ \frac{\textrm{number of all states with $N_{\Omega}$,
$Q$, $E$, and $\vec P = \vec 0$}}{\textrm{number of all states with $Q$,
$E$, and $\vec P= \vec 0$} }~.$$ The real MCE partition function and our modified version are connected as $Z(V,N_{\Omega},Q,E,\vec P) \equiv \mathcal{Z}^{N_{\Omega},Q,E,\vec
P}(V,T,\mu) e^{+E/T} e^{-Q\mu/T}$. In either case the normalization in Eq.(\[MCEprob\]) is given by the partition functions with fixed values of $Q,E,\vec P$, but arbitrary particle number $N_{\Omega}$, hence $Z(V,Q,E,\vec
P) \equiv \sum_{N_{\Omega}=0}^{\infty} Z(V,N_{\Omega},Q,E,\vec P)$, or $\mathcal{Z}^{Q,E,\vec P}(V,T,\mu) \equiv \sum_{N_{\Omega}=0}^{\infty}
\mathcal{Z}^{N_{\Omega},Q,E,\vec P}(V,T,\mu)$. However when taking the ratio (\[MCEprob\]) auxiliary parameters chemical potential and temperature drop out: $$\label{MCEprob2}
P_{mce}(N_{\Omega})~ \equiv~ \frac{Z(V,N_{\Omega},Q,E,\vec P)}{Z(V,Q,E,\vec
P)} ~=~ \frac{\mathcal{Z}^{N_{\Omega},Q,E,\vec
P}(V,T,\mu)}{\mathcal{Z}^{Q,E,\vec P}(V,T,\mu)}~.$$ The main difference between the two versions of partition functions is that for $Z(V,N_{\Omega},Q,E,\vec P)$ one is confronted with a heavily oscillating (or even irregular) integrant, while for $\mathcal{Z}^{N_{\Omega},Q,E,\vec
P}(V,T,\mu)$ the integrant becomes ($T$,$\mu$ correctly chosen) very smooth. Thus, introduction of $T$ and $\mu$ allows to derive (and use) the asymptotic solution of Ref.[@clt].
We have a total number of 6 conserved ‘charges‘, and hence we need to solve the 6-dimensional Fourier integral for the numerator in Eq.(\[MCEprob2\])[^1]: $$\begin{aligned}
\label{MCEInt}
\mathcal{Z}^{N_{\Omega},Q,E, \vec P}&=&
\int \limits_{-\pi}^{\pi} \frac{d \phi_{N_{\Omega}}}{2\pi}
\int \limits_{-\pi}^{\pi} \frac{d \phi_Q}{2\pi}
\int \limits_{-\infty}^{\infty} \frac{d \phi_E}{2\pi}
\int \limits_{-\infty}^{\infty} \frac{d \phi_{P_x}}{2\pi}
\int \limits_{-\infty}^{\infty} \frac{d \phi_{P_y}}{2\pi}
\int \limits_{-\infty}^{\infty} \frac{d \phi_{P_z}}{2\pi} \nonumber \\
&\times& e^{-iN_{\Omega} \phi_{N_{\Omega}}}~ e^{-iQ\phi_Q} ~e^{- iE \phi_E}
~e^{-iP_x \phi_{P_x}}~ e^{-iP_y \phi_{P_y}}~ e^{-iP_z \phi_{P_z}} \nonumber \\
&\times& \exp \left[ V \sum_k \psi_k \left( \phi_{N_{\Omega}}\phi_Q, \phi_E,
\phi_{P_x},\phi_{P_y},\phi_{P_z} \right) \right].\end{aligned}$$ The summation in (\[MCEInt\]) should be taken over the single particle partitions $V \psi_k=z_k$ of all considered particle species $k$. The Wick-rotated fugacities $\phi_{Q}$, etc. are related to the individual conservation laws. The distinction between the Kronecker delta-function (limits of integration $\left[-\pi,\pi \right]$) for discrete quantities and the Dirac delta-function (limits of integration $\left[-\infty,\infty \right]$) for continuous quantities is important here, however for deriving an asymptotic solution it will not be. To simplify (\[MCEInt\]) we change to shorthand notation for $\phi_j =
(\phi_{N_{\Omega}}\phi_Q, \phi_E, \vec \phi_P)$ and the conserved ‘charge‘ vector $ Q^j = (N_{\Omega},Q,E,\vec P)$. We again split the single particle partition functions in two parts. The first part counts the number of momentum states observable to our detector, while the second part counts momentum states invisible to our detector: $$\begin{aligned}
\psi_{k} \left( \phi_j\right) &=& \frac{g_k}{\left( 2\pi\right)^3} \int
\limits_{ \Omega} d^3 p ~ e^{-\frac{\varepsilon_k - q_k \mu}{T}} ~e^{i q_{k,
\Omega}^j \phi_j}~ +~ \frac{g_k}{\left( 2\pi\right)^3}
\int \limits_{ \bar \Omega} d^3 p~ e^{-\frac{\varepsilon_k-q_k \mu}{T}} ~ e^{i
q_{k,\bar \Omega}^j \phi_j}~.\end{aligned}$$ For the ‘charge‘ vector of all measured particle species $k$ we write $q^j_{k,\Omega} = (1,q_k,\varepsilon_k,\vec p_k)$ for momenta inside $\Omega$, and $q^j_{k, \bar \Omega} = (0,q_k,\varepsilon_k,\vec p_k)$ for momenta outside of $\Omega$. For all unobserved particle species we write $q^j_{k,\Omega}=q^j_{k, \bar \Omega} =(0,q_k,\varepsilon_k,\vec p_k) $. Here $q_k$ is the electrical charge of particle species $k$, and $\varepsilon_k$ and $\vec p_k$ are its energy and momentum vector. In Ref.[@clt], where only multiplicity distributions in the full momentum space were considered, the general ‘charge‘ vector took the form $q^j_{k,4\pi} =
(n_k,q_k,\varepsilon_k,\vec p_k)$, where $n_k$ is the multiplicity of this particle. For stable particles $n_k=1$ in case they are observed, and $n_k=0$ if they are not measured, while for unstable particles $n_k$ could also denote the number of measurable decay products.
For large system volume the main contribution to the integral (\[MCEInt\]) comes from a small region around the origin [@VolDep]. Thus we proceed by Taylor expansion of the integrant of (\[MCEInt\]) around $\phi_j=\vec
0$. In this context $ \Psi \left( \phi_j\right) = \sum_k \psi_k \left(
\phi_j\right) $ would be called the cumulant generating function (CGF). Cumulants (expansion terms) are defined by differentiation of the CGF at the origin: $$\label{kappa_n}
\kappa_n^{j_1,j_2,\dots,j_n } ~\equiv~ \left(-i\right)^n\frac{\partial^n
\Psi \left( \phi_j \right) }{\partial \phi_{j_1}
\partial \phi_{j_2} \dots \partial \phi_{j_n} }
\Bigg|_{\phi_j = \vec 0}~.$$ Generally are cumulants tensors of rank $n$ and dimension defined by the number of conserved quantities. Here $\kappa_1$ is a 6 component vector, while $\kappa_2$ is a $6 \times 6$ matrix, etc.
The parts of the integrant related to discrete quantities, i.e. $N_{\Omega}$ and $Q$, are now not $2\pi$ periodic anymore (while in Eq.(\[MCEInt\]) they are), but superpositions of oscillating and decaying parts. Thus we extent the limits of integration to $\pm \infty$, what introduces a negligible error. Eq.(\[MCEInt\]) therefore simplifies to: $$\begin{aligned}
\label{MCEIntApprox}
\mathcal{Z}^{Q^j} &\simeq& \left[ \prod_{j=1}^{6} \int
\limits_{-\infty}^{\infty} \frac{d\phi_j}{\left( 2\pi \right)} \right] ~\exp
\Big[ -iQ^j\phi_j
~+~V \sum_{n=0}^{\infty} \frac{i^n}{n!} \; \kappa_n^{j_1,j_2,\dots,j_n } \;
\phi_{j_1} \phi_{j_2} \dots \phi_{j_n} \Big] ~.\end{aligned}$$ Summation over repeated indices is implied. Existence and finiteness of the first three cumulants provided, any such integral can be shown to converge to a multivariate normal distribution in the large volume limit: $$\label{MCE_MultNormal}
\mathcal{Z}^{Q^j} ~\simeq~ Z_{gce} \frac{\exp \left(-\frac{\xi^j \; \xi_{ j}}{2}
\right)}{\left(2\pi V \right)^{6/2} \det| \sigma| }~,$$ where $Z_{gce}\equiv \exp \left[V \kappa_0 \right]$ is the GCE partition function, $\kappa_0$ is the cumulant of $0^{th}$ order, $\xi^j=\left( Q^k - V
\kappa_1^k \right) \left( \sigma^{-1}\right)_{k}^{\;\;j} V^{-1/2}~$ is a measure for the distance of a particular macro state $Q^k$ to the peak $V
\kappa_1^k$ of the joint distribution, and $\sigma$ is the square root of the second rank tensor $\kappa_2$, see [@clt] for details.
The normalization in Eq.(\[MCEprob2\]) can essentially be found in two ways. The first way would be to integrate the distribution (\[MCE\_MultNormal\]) over all possible values of multiplicity $N_{\Omega}$, while all other variables are set to their peak values, e.g. $Q=V\kappa_1^Q$, $E=V\kappa_1^E$, $\vec P = \vec 0$. The second and more practical way is to use an approximation similar to Eq.(\[MCE\_MultNormal\]) to describe the macro state $Q^j = (Q,E,\vec P)$. The normalization in Eq.(\[MCEprob2\]), $\mathcal{Z}^{E,Q,\vec P}$, is then given by the 5-dimensional integral, similar to Eq.(\[MCEInt\]), without the integration over $\phi_{N_{\Omega}}$. The 1-dimensional slice along $N_{\Omega}$, i.e. the conditional distribution of particle number $N_{\Omega}$, while charge, energy and momentum are fixed to $Q,E,\vec P = \vec 0$, can then be shown [@clt] to converge to a Gaussian in the large volume limit: $$\begin{aligned}
\label{PMCE}
P_{mce}(N_{\Omega} ) ~ \simeq~ \frac{1}{\left(2\pi ~\omega^{mce}_{\Omega}
~ \langle N_{\Omega} \rangle \right)^{1/2}} ~ \exp \left(- \frac{ \left(
N_{\Omega} - \langle N_{\Omega} \rangle \right)^2}{2
~ \omega^{mce}_{\Omega} ~ \langle N_{\Omega} \rangle } \right) ~. \end{aligned}$$ The scaled variance $\omega^{mce}_{\Omega}$ is given by the ratio of the two determinants of the two relevant second rank cumulants, $\kappa_2$ and $\tilde
\kappa_2$, of the two partition functions $\mathcal{Z}^{N_{\Omega},E,Q,\vec
P}$ and $\mathcal{Z}^{E,Q,\vec P}$, hence[^2]: $$\label{SimpleOmega}
\omega^{mce}_{\Omega} = \frac{ \det | \kappa_2| }{
\kappa_1^{N_{\Omega}} \;\det| \tilde \kappa_2| }~.$$ The asymptotic ($V\rightarrow \infty$) scaled variance can therefore be written in the form of Eq.(28) in [@clt]. Considering only the asymptotic solution we need to investigate only the first two cumulants ($n=1,2$) in detail. We will first discuss the structure of $\kappa_1$ and $\kappa_2$, and then deduce a few properties of Eq.(\[SimpleOmega\]).
The first order cumulant $\kappa_1$ of $\mathcal{Z}^{N_{\Omega},Q,E,\vec
P}$ gives GCE expectation values for particle density $\kappa_1^{N_{\Omega}}$, charge density $\kappa_1^{Q}$, energy density $\kappa_1^{E}$, and expectation values of momentum $\kappa_1^{p_x}$, etc. Since we are only interested in a static source we find due to the antisymmetric momentum integral (see Appendix \[Calc\]) $\kappa_1^{p_x} = \kappa_1^{p_y} = \kappa_1^{p_z}=0$. The general form of the first cumulant $\kappa_1$ is then: $$\begin{aligned}
\kappa_1 =
\begin{pmatrix}
\kappa_1^{N_{\Omega}}, & \kappa_1^{Q}, & \kappa_1^{E}, & 0, & 0, & 0
\end{pmatrix}~. \label{vector}\end{aligned}$$
The second cumulant $\kappa_2$ of $\mathcal{Z}^{N_{\Omega},Q,E,\vec
P}$ contains information about correlations due to different conserved quantities. A detailed discussion of correlation terms only involving Abelian charges and/or energy, e.g. $\kappa_2^{Q,Q}$, $\kappa_2^{Q,E}$, and $\kappa_2^{E,E}$, can be found in [@clt]. Again, due to the antisymmetric nature of the momentum integral, all cumulant entries involving an odd order in one of the momenta, e.g. $\kappa_2^{E,p_x}$, $\kappa_2^{p_x,p_y}$, or $\kappa_2^{Q,p_x}$ are equal to zero. The general second order cumulant $\kappa_2$ thus reads: $$\begin{aligned}
\kappa_2 =
\begin{pmatrix}
\kappa_2^{N_{\Omega},N_{\Omega}} & \kappa_2^{N_{\Omega},Q} &
\kappa_2^{N_{\Omega},E} & \kappa_2^{N_{\Omega},p_x} &
\kappa_2^{N_{\Omega},p_y} & \kappa_2^{N_{\Omega},p_z} \\
\kappa_2^{Q,N_{\Omega}} & \kappa_2^{Q,Q} & \kappa_2^{Q,E} & 0 & 0 & 0 \\
\kappa_2^{E,N_{\Omega}} & \kappa_2^{E,Q} & \kappa_2^{E,E} & 0 & 0 & 0 \\
\kappa_2^{p_x,N_{\Omega}} & 0 & 0 & \kappa_2^{p_x,p_x} & 0 & 0 \\
\kappa_2^{p_y,N_{\Omega}} & 0 & 0 & 0 & \kappa_2^{p_y,p_y} & 0 \\
\kappa_2^{p_z,N_{\Omega}} & 0 & 0 & 0 & 0 & \kappa_2^{p_z,p_z}
\end{pmatrix}~. \label{matrix}\end{aligned}$$ Please note that by construction, Eq.(\[kappa\_n\]), the matrix (\[matrix\]) is symmetric, hence $\kappa_2^{N_{\Omega},Q} =
\kappa_2^{Q,N_{\Omega}}$, etc.
The second matrix $\tilde \kappa_2$, now related to the partition function $\mathcal{Z}^{Q,E,\vec P}$, is obtained from $\kappa_2$, Eq.(\[matrix\]), by crossing out the first row and first column. In the following we are going to make use of the fact that one can express the determinant of a matrix $A$ by: $$\label{calcdet}
\det |A| ~=~ \sum \limits_{j=1}^n \left( -1\right)^{j+k}
A_{j,k} M_{j,k} ~,$$ where $A_{j,k}$ is the matrix element $j,k$ of a general non-singular $n\times
n$ matrix $A$, and $ M_{j,k}$ is its complementary minor. A simple consequence of Eq.(\[calcdet\]) is: $$\label{normdet}
\det |\tilde \kappa_2| = \kappa_2^{p_x,p_x} ~\kappa_2^{p_y,p_y} ~\kappa_2^{p_z,p_z}
\left[ \kappa_2^{E,E}~\kappa_2^{Q,Q}- \left(\kappa_2^{E,Q}\right)^2 \right] ~=~
\left( \kappa_2^{p_x,p_x} \right)^3 \det |\hat \kappa_2|,$$ where $\kappa_2^{p_x,p_x} =\kappa_2^{p_y,p_y} =\kappa_2^{p_z,p_z} $, due to spherical symmetry in momentum space, and $\hat \kappa_2$ is just a $2\times2$ matrix involving only terms containing $E$ and $Q$. In case correlations between particle number and conserved momenta are vanishing, i.e. $\kappa_2^{N_{4\pi},p_x} = 0$, or $\kappa_2^{N_{\Omega},p_x} = 0$, then, similarly to Eq.(\[normdet\]), the determinant of $\kappa_2$ factorizes into a product of correlation terms $(\kappa_2^{p_x,p_x})^3$ and the determinant of a $3\times3$ sub-matrix involving only terms containing $E$, $Q$, and $N$. Hence in taking the ratio Eq.(\[SimpleOmega\]) one notes, that in this case momentum conservation will not affect multiplicity fluctuations in the large volume limit [@clt]. In this work, however we do not necessarily find $ \kappa_2^{N_{\Omega},p_x} = 0 $, as we only integrate over a limited segment $\Omega$ of momentum space, and taking momentum conservation into account may affect the result.
Finally it should be stressed that this procedure can be easily generalized to account for Bose or Fermi statistics. Also phenomenological phase space suppression (enhancement) factors $\gamma_q$ [@gammaQfirst] or $\gamma_s$ [@gammaSfirst] could be straightforwardly included. However, without proper implementation of the effect of additional correlations due to resonance decay and collective motion, i.e. flow, it seems of little value to do too strict calculations for experimentally measurable distributions. We thus return to the pion gas example from section \[GCECE\] and restrict the discussion to simple momentum space cuts in rapidity, transverse momentum, and azimuthal angle, see also the Appendix for details.
Results {#Results}
=======
Multiplicity fluctuations in the full momentum space
----------------------------------------------------
Let us firstly recall basic properties of multiplicity fluctuations of negative particles in the full momentum space ($4\pi$ fluctuations) in the three standard ensembles, of the Boltzmann pion gas considered here.
Multiplicity fluctuations in the CE are suppressed due to exact charge conservation. For a neutral ($Q=0$) system one finds in the large volume limit $\omega_{4\pi}^{ce} = 0.5$ [@CEfluc1]. Further suppression of fluctuations arise from additionally enforcing exact energy conservation in the MCE. Here one finds $\omega_{4\pi}^{mce} \approx 0.25$ for a Boltzmann pion gas at $T\approx 160MeV$. In the GCE, since no conservation laws are enforced, we always find a Poisson distribution with width $\omega_{4\pi}^{gce} =1$.
Since charge conservation in CE links the distributions of negatively charged particles to the one of their positive counterparts, i.e. $P(N_-) = P(N_+-Q)$, the relative width of $P(N_-)$ increases (decreases) as we move the electric charge density to positive (negative) values [@CEfluc2]. This can be easily be seen from Eq.(\[matrix\]) by crossing out all rows and columns containing energy and momentum and calculating the asymptotic scaled variance of negatively charged particles, $\omega^{ce}_{4\pi}$, from Eq.(\[SimpleOmega\]), $$\omega_{4\pi}^{ce}~=~ \frac{\kappa_2^{N_{4\pi},N_{4\pi}}\kappa_2^{Q,Q}-\left(
\kappa_2^{N_{4\pi},Q} \right)^2}{\kappa_1^{N_{4\pi}}~ \kappa_2^{Q,Q}} ~=~
\frac{\exp \left( \frac{\mu}{T} \right)}{2\cosh \left(
\frac{\mu}{T}\right)}~.$$ The same effect is present in the MCE, however the calculation is slightly longer.
Results for $4\pi$ multiplicity fluctuations of negatively charged particles in a Boltzmann pion gas at $T=160MeV$ and different charge densities are summarized in Table \[table\]. Additionally estimates, based on our previously employed ‘uncorrelated particle‘ approach, Eq(\[accscaling\]), for multiplicity fluctuations with limited acceptance are given.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\quad \omega^{gce}_{4\pi} \quad$ $\quad \omega^{ce}_{4\pi} \quad$ $ $ \quad \omega^{gce}_{q=1/9} \quad$ $\quad \omega^{ce}_{q=1/9} \quad$ $\quad \omega^{mce}_{q=1/9} \quad$
\quad \omega^{mce}_{4\pi} \quad$
-------------------- ----------------------------------- ---------------------------------- ---------------------------------- ------------------------------------- ----------------------------------- ------------------------------------
$\mu=0$ $1$ $0.5$ $0.235$ $1$ $0.944$ $0.915$
$\mu=-\frac{m}{2}$ $1$ $0.294$ $0.147$ $1$ $0.922$ $0.905$
$\mu=+\frac{m}{2}$ $1$ $0.706$ $0.353$ $1$ $0.967$ $0.928$
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Multiplicity fluctuation of $\pi^-$ in a classical pion gas in the large volume limit in the three standard ensembles at $T=160MeV$ for different charge densities. The index ‘$4\pi$‘ denotes fluctuations in the full momentum space, while the index ‘$q=1/9$‘ assumes acceptance scaling, Eq.(\[accscaling\]). The ratio $n_-/n_{tot}$ equals to $0.33$ for $\mu=0$, $0.48$ for $\mu=-m/2$, and $0.20$ for $\mu=+m/2$.[]{data-label="table"}
Despite the fact that $\omega_{4\pi}$ is very different in GCE, CE, or MCE and also rather sensitive to the charge density, the estimates for limited acceptance ($q=1/9$) based on Eq.(\[accscaling\]) vary only by a few %. In order to decisively distinguish predictions for different ensembles a large value of $q$ would be needed.
Multiplicity fluctuations in limited segments of momentum space
---------------------------------------------------------------
In Section \[GCECE\] we have seen that in the Boltzmann CE multiplicity fluctuations observed in a limited segment of phase space are insensitive to the position of this segment. The dependence on the size of the segment can thus be taken into account by use of acceptance scaling Eq.(\[accscaling\]). To balance charge a particle can be produced or annihilated anywhere in momentum space. And due to a infinitely large heat and momentum bath in the CE no momentum state is essentially preferred.
In the MCE this dependence is qualitatively different. When using the MCE formulation particles are correlated due to the constraints of exactly conserved energy and momentum, even in the large volume limit. Fluctuations in a macroscopic subsystem are strongly affected by correlations with the remainder of the system.
In Fig. \[dodp\] we show the scaled variance of multiplicity fluctuations for negatively charges particles in finite bins in transverse momentum (left), and rapidity (right). The bins are constructed such that each bin contains on average the same fraction $q$ of the total average yield. The width of each bin is indicated by the bars. Calculations are done for two values of acceptance ($q=1/5$, and $q=1/9$). The dashed and dotted lines correspond to acceptance scaling Eq.(\[accscaling\]), while the markers are calculated from Eq.(\[SimpleOmega\]). One finds that multiplicity fluctuations in bins with high transverse momentum and high values of rapidity are, due to energy and momentum conservation, essentially suppressed with respect to bins where individual particles carry less energy and momentum.
A intuitive explanation would probably look like this: Let us consider an event with an unusually large (small) number of particles at the most forward rapidity bin. In this bin we would find therefore a macroscopic state with unusually large (small) observed longitudinal momentum $P^{obs}_z$ and energy $E^{obs}$. The remainder of the system therefore has to have rather large (small) momentum $-P_z^{obs}$ and rather small (large) energy $E-E^{obs}$. Since both probability distributions, for the observed and the unobserved subsystems, do not factorize into independent probability distributions, but are correlated, this macro state would be rather unlikely. Fluctuations about the mean $\langle N_y \rangle$ at forward (backward) rapidities should therefore be suppressed. On the other hand can modest multiplicity fluctuations in a high $p_T$ bin induce stronger fluctuations in the lower $p_T$ bins, and fluctuations about $\langle N_{p_T}
\rangle$ in a low $p_T$ bin are enhanced. Even when detecting only a fraction of about 10% of the total system these correlations can have a sizeable effect.
Conservation laws
-----------------
It seems worthwhile to consider individual conservation laws and their impact on multiplicity fluctuations in more detail. On of the main advantages of the analytical procedure presented here, is certainly that one can easily ‘switch on‘ or ‘switch off‘ a particular conservation law. For illustrative purposes we show the result of $\omega^{mce}_{\Delta y}$ for MCE without longitudinal momentum conservation in Fig.\[dodywopz\].
In comparing Fig.\[dodp\], right, to Fig.\[dodywopz\] it becomes obvious that energy conservation alone cannot account for the strong suppression of multiplicity fluctuations at forward rapidities, but has to be explained by combined energy and longitudinal momentum conservation.
The relevant cumulants elements, which give information about the strenght of correlations between particle number and a particular conserved quantity, are $\kappa_2^{N_{\Omega},Q}$, $\kappa_2^{N_{\Omega},E}$, $\kappa_2^{N_{\Omega},p_x}$, etc. Whenever a element is vanishig, then the corresponding conservation law has no impact on multiplicity fluctuations. For details of the calculations please see the Appendix. Since for fluctuations of charged particles $\kappa_2^{N_{\Omega},Q}$ and $\kappa_2^{N_{\Omega},E}$ are generally non-zero, we will focus only on the effects of momentum conservation.
For multiplicity fluctuations in bins in transverse momentum momentum conservation does not affect the result, see Appendix \[App\_pt\], and the suppression effect is a result of energy conservation alone. When considering cuts in rapidity one finds in general $\kappa_2^{N_y,p_z} \not= 0$, but $\kappa_2^{N_y,p_x} = \kappa_2^{N_y,p_y} = 0$, and only longitudinal momentum conservation needs to be taken into account, see Appendix \[App\_y\]. In considering the third idealized case, where our detector observes only a segment in azimuthal angle $\phi$, but all rapidities $y$ and transverse momenta $p_T$, both global $P_x$, and $P_y$ conservation lead to non-trivial modifications of Eq.(\[accscaling\]), see Appendix \[App\_phi\].
To understand the difference between the strong suppression of fluctuations at high transverse momentum and the rather modest suppression at high rapidity when momentum conservation is not enforced, one should compare the elements $\kappa_2^{N_{p_T},E}$ in Eq.(\[omega\_pt\]), and $\kappa_2^{N_{y},E}$ in Eq.(\[omega\_y\]), which measure in Boltzmann approximation the average energy density carried by particles in a bin $\Delta p_T$ or $\Delta y$, to the total average energy density $\langle E_-\rangle = \kappa_2^{N_{4\pi},E}$ carried by $\pi^-$. (All other elements in Eqs.(\[omega\_pt\]) and (\[omega\_y\]) do not depend on the location of the segment.) In case of kinematical cuts in $\Delta p_T$ the fraction $\kappa_2^{N_{p_T},E} / \langle
E_- \rangle$ rises from about $5\%$ in the lowest to roughly $20\%$ in the highest $p_T$-bin. In contrast to that for the central $y$-bin this ratio is about $10\%$, while the most forward or backward bins it is roughly $12\%$. However in both cases the bins contain on average $q=1/9\approx 11\%$ of the total average $\pi^-$ yield. The effect of energy conservation is thus weaker for cuts in rapidity than for cut in transverse momentum, see also Appendices \[App\_pt\] and \[App\_y\].
Charged systems
---------------
In Figs. \[dodp\_chr\] the transverse momentum (left) and rapidity (right) dependence of the scaled variance is presented for two different values of charge density.
Similar to the CE, in MCE the effective size of the heat and charge bath matters. We find that in general MCE effects for negatively charged particles are stronger (weaker) when the electric charge density is negative (positive). In the limit of a strongly positively charged system, the $\pi^-$ subsystem could be considered as embedded in a large heat, charge, and momentum bath (provided by $\pi^+$ and $\pi^0$ particles) and MCE effects would cease. The GCE would here be the appropriate limit. In the opposite limit of a strongly negatively charged system, charge conservation essentially becomes equivalent to particle number conservation. This scenario might be more familiar from textbooks, where the CE is usually understood as the ensemble with fixed particle number. However here also the same arguments as above apply, except the effect would be much stronger, and $\omega^{mce}_{4\pi} = 0$.
In general one would expect that suppression effects in bins of high transverse momentum or high values of rapidity are stronger the more abundant the analyzed particle species is. In the context of heavy ion collision this implies that MCE effects should be stronger for positively charged particles than for negatively charged particles, due to the fact that the created system carries positive net-charge.
Previous work suggests that the asymptotic values for the scaled variance are indeed reached rather quickly [@VolDep] and above results are certainly applicable to large systems expected to be created in relativistic heavy ion collisions.
Remarks and Conclusion {#Remarks}
======================
Some concluding remarks seems to be in order. Although it might seem inappropriate to use the MCE formulation of a hadron resonance gas model for calculation of multiplicity fluctuations in heavy ion collisions, as energy and volume cannot be assumed to be the same in all events, it should be stressed that GCE and CE still imply a very particular type of heat (and momentum) bath, namely an infinite (and ideal) one. This assumption seems to us even less appropriate. Also the MCE is often understood as the ensemble with energy (and charge), however not momentum conservation. It is usually assumed that taking momentum conservation into account will not affect fluctuations in the large volume limit. We have shown [@clt] in a recent paper that this is indeed the case, when one assumes information about all produced particles. However for calculations of multiplicity fluctuations in arbitrary finite subsystems in momentum space all kinematic conservation laws need to be taken into account.
In a realistic heavy ion experiment it seems impossible to measure the entire final state of each collision. The observed subsystem could therefore be seen as effectively embedded into a (possibly much larger) heat, charge, and momentum bath. Sometimes it is therefore argued that, when investigating only a small part of a statistical system (canonical or micro-canonical), one can ignore correlations of the subsystem under investigation with the remaining system. This argument is often applied when considering yields and/or fluctuations in a limited segment of momentum space. More precisely, usually the GCE is thought to be the appropriate ensemble to model fluctuations of particle multiplicity or particle ratios found in some mid-rapidity interval [@GCEfluc]. In this work we have argued that this assumption should be checked carefully. The GCE is only the correct ensemble to choose, if heat and charge bath are assumed to be infinite, while the observed subsystem remains finite.
Based on our previous line of arguments, one would also expect that strong collective longitudinal and transverse flow would lead to a strong correlation of macroscopic subsystems. Longitudinal momentum conservation implies that when ‘observing‘ in an event a final state with a certain small (large) number of produced particles at very forward rapidity, a similarly small (large) number of particles should exist at backward rapidities. Particles in these bins carry substantial longitudinal momenta, and hence energy. Modest fluctuations in their numbers should therefore induce stronger fluctuations in the central rapidity region. The same line of arguments is applicable to the transverse momentum dependence. One would therefore expect a similar momentum space dependence of experimentally measured charged particle multiplicity fluctuations as shown in Figs. \[dodp\].
This argument is additionally supported by UrQMD simulations [@UrQMDfluc]. In transport calculations the produced systems stay far away from global or local equilibrium [@TransportEq] and other (dynamical) mechanisms might lead to similar effects. On the other hand could one also infer from [@UrQMDfluc] that even in non-equilibrium systems correlations due to exactly enforced conservation laws determine the general trend, although transport simulations show, for instance, a very different dependence of multiplicity fluctuations on beam energy [@HSDfluc1; @HSDfluc2] than statistical equilibrium models. This should be subject of further investigation.
Finally, and most importantly, we want to stress that recently presented preliminary NA49 analysis of multiplicity fluctuations in certain rapidity and transverse momentum windows [@BeniCoolData] shows qualitatively the very same trends as they are suggested by the MCE formulation of the statistical model. Data, UrQMD simulations, and the statistical model exhibit suppressed multiplicity fluctuations when bins with high transverse momentum (or high values of rapidity) are compared to bins of same mean multiplicity at lower transverse momentum (or lower values of rapidity). We are certainly tempted to interpret this rather unexpected common behavior as a manifestation of energy and momentum conservation effects.
Summary {#Summary}
=======
We have discussed the effect of momentum space cuts on multiplicity fluctuations in the framework of an ideal classical pion gas in the three standard ensembles, GCE, CE, and MCE. Only in the MCE we expect a momentum space dependence of multiplicity fluctuations, when comparing intervals of same average multiplicity. We have shown that even in the thermodynamic limit energy-momentum conservation can leave a sizable effect in the fluctuation pattern.
In a previous publication we have argued that despite the fact one may expect event-by-event fluctuations of the thermal energy, i.e. the part of the total energy which goes into thermal particle production rather than collective expansion, these event-by-event fluctuations remain small compared to energy fluctuations one would expect from grand canonical and canonical ensembles. In this work we have shown that energy and momentum conservation lead to a non-trivial momentum space dependence of the fluctuation pattern. This argument seems to be strongly supported by data.
Above results become all the more interesting when compared to models which seek to describe effects beyond our considerations. In fact our calculations suggest a similar strength of respective suppression or enhancement as they were predicted as signals for the critical point of strongly interacting matter, the onset of deconfinement, or generally a possible phase transition. One might also be tempted to argue, that enhanced fluctuations around mid-rapidity, when compared to a more forward rapidity slice, should be interpreted as a signal of a phase transition from a quark gluon plasma to a hadron gas phase, expected to be first realized in the presumably hotter and denser central rapidity region. However in this case there should be a non-monotonic variation as center of mass energy of colliding nuclei is changed. This seems not to be supported by preliminary NA49 data.
In summary, above results should be treated as a prediction for general trends of multiplicity fluctuations in limited segments of momentum space. The existence of this general behavior should be further tested by current experiments. Observation of effects similar to those of Figs. \[dodp\] in experimental data would, in our opinion, strongly speak in favor of our hypothesis that fluctuations of extensive observables are indeed dominated by material and motional conservation laws.
We would like to thank F. Becattini, V.V. Begun, M. Bleicher, E.L. Bratkovskaya, W. Broniowski, L. Ferroni, M.I. Gorenstein, M. Gaździcki, S. Häussler, V.P. Konchakovski, B. Lungwitz, and G. Torrieri for fruitful discussions.
Globally Conserved Quantities {#Calc}
=============================
Turning now to calculations of cumulants, Eq.(\[kappa\_n\]), we employ always coordinates most suitable to our problem. The invariant phase space element is given by: $$\varepsilon ~\frac{dN}{d^3p} ~=~ \frac{dN}{m_T~ dm_T ~dy ~d \phi} ~=~
\frac{dN}{p_T ~dp_T ~dy ~d \phi} ~=~ \varepsilon ~\frac{g}{\left( 2\pi\right)^3}
~\exp \left( -\frac{\varepsilon-\mu}{T}\right)~,$$ where the single particle energy $\varepsilon = m_T \cosh y$, its longitudinal momentum $p_z = m_T \sinh y$, transverse mass $m_T^2 = p_T^2 + m^2 $, transverse momentum $p_T^2 = p_x^2 + p_y^2$, and rapidity $y = \tanh \left(
p_z/\varepsilon\right)$. Additionally we employ spherical coordinates: $$\frac{dN}{d^3p} ~=~ \sin \theta ~ p^2 ~ \frac{dN}{d\phi~ d\theta ~dp}~.$$ For clarity we consider explicitely a few terms, not given in [@clt], here. The total energy density is given by the sum over individual contributions of all particle species $k$: $$\begin{aligned}
\label{kappa_1_E}
\kappa_1^E &=& \left(- i \frac{\partial }{\partial
\phi_{E}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0} ~=
\sum_k \int \limits_{0}^{+ \pi} ~d \theta \int
\limits_{-\pi}^{+\pi} d \phi \int \limits_{0}^{\infty} ~dp
~\varepsilon_k~ \frac{dN_k}{d\phi ~d\theta ~ dp} \nonumber \\
&=& \sum_k \frac{g_k ~e^{\frac{q_k\mu}{T}}}{2\pi^2} ~m_k^3 ~T~ \left[K_1
\left(\frac{m_k}{T} \right) + 3~ \frac{T}{m_k}~ K_2 \left(\frac{m_k}{T}
\right)\right]~= \sum_k \langle E_k \rangle~.\end{aligned}$$ The diagonal energy element $\kappa_2^{E,E}$ is given by: $$\begin{aligned}
\kappa_2^{E,E} &= & \left(- i \frac{\partial }{\partial
\phi_{E}}\right)^2 \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0} ~=
\sum_k \int \limits_{0}^{+ \pi} ~d \theta \int
\limits_{-\pi}^{+\pi} d \phi \int \limits_{0}^{\infty}~dp
~\varepsilon_k^2~ \frac{dN_k}{d\phi ~d\theta ~ dp}\nonumber \\
&=& \sum_k \frac{g_k ~e^{\frac{q_k\mu}{T}}}{2\pi^2} ~m_k^4 ~T~ \left[K_0
\left(\frac{m_k}{T} \right) + 5~ \frac{T}{m_k} ~K_1 \left(\frac{m_k}{T}
\right) + 12~\frac{T^2}{m_k^2}~ K_2 \left(\frac{m_k}{T} \right) \right] ~.\end{aligned}$$ Additionally we define the diagonal momentum correlation terms, with $p_z = p
\cos \theta$: $$\begin{aligned}
\label{kappa_p_p}
\kappa_2^{p_z,p_z} &=& \left(- i \frac{\partial }{\partial
\phi_{p_z}}\right)^2 ~\Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0}~=
\sum_k \int \limits_{0}^{2\pi} d\phi ~ \int \limits_{0}^{\pi} d\theta ~
\int \limits_{0}^{\infty} dp ~p_z^2 ~\frac{dN_k}{d\phi ~d\theta ~ dp}
\nonumber \\
&=&\sum_k \frac{g_k ~e^{\frac{q_k\mu}{T}}}{2\pi^2}~ m_k^4~T~
\Bigg[\frac{T}{m_k} ~K_1\left(\frac{m_k}{T} \right) +4~\frac{T^2}{m_k^2}~
K_2\left(\frac{m_k}{T} \right) \Bigg]~.\end{aligned}$$ Due to spherical symmetry in momentum space we find $\kappa_2^{p_x,p_x} =
\kappa_2^{p_y,p_y} = \kappa_2^{p_z,p_z}$. Correlation terms of odd order in one of the momenta are identical to zero. As an example we find for correlations between energy and longitudinal momentum: $$\begin{aligned}
\label{kappa_E_p}
\kappa_2^{E,p_z} &=& \left(- i \frac{\partial }{\partial
\phi_{E}}\right) \left(- i \frac{\partial } {\partial
\phi_{p_z}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0}
\nonumber \\
&=& \sum_k \int \limits_{0}^{2\pi} d\phi \int \limits_{0}^{\pi}
d\theta \int \limits_{0}^{\infty} dp ~\varepsilon ~p_z ~\frac{dN_k}{d\phi
~d\theta ~ dp} =0, \end{aligned}$$ since the integral over the polar angle $\int_{0}^{\pi} \sin \theta \cos
\theta=0$. Similarly we find $\kappa_2^{Q,p_x} = \kappa_2^{p_x,p_y} = 0$. Additionally $\kappa_1^{p_x}=0$, etc., since for a static source $\langle \vec
P \rangle = \vec 0$.
Transverse Momentum Segment {#App_pt}
===========================
The average particle number density of $\pi^-$ in a segment of transverse momentum ${\Delta p_T}$ is given by Eq.(\[kappa\_n\]), i.e. the first derivative of the CGF with respect to $\phi_{N_{\Omega}}=\phi_{N_{p_T}}$ at the origin: $$\begin{aligned}
\label{dNdpt}
\kappa_1^{N_{p_T}} &=& \left( -i \frac {\partial}
{\partial \phi_{N_{p_T}}} \right) \Psi\left( \phi_j \right)
\Bigg|_{\phi_j=\vec 0} ~=~
\int \limits_{\Delta p_T} dp_T \int
\limits_{0}^{2\pi} d\phi \int \limits_{-\infty}^{\infty}
dy ~ \frac{dN}{dp_T ~dy ~ d\phi} \nonumber \\
&=& \frac{g~e^{-\frac{\mu}{T}}}{2\pi^2} \int
\limits_{\Delta p_T} dp_T ~p_T \sqrt{p_T^2+m^2} ~K_1 \left(
\frac{\sqrt{p_T^2+m^2}}{T}\right)~.\end{aligned}$$ Please note that $\kappa_1^{N_{p_T}} = \int_{\Delta p_T} dp_T ~dN/dp_T =
\langle N_{p_T} \rangle$. Correlations of $\pi^-$ in a segment ${\Delta
p_T}$ with globally conserved energy are given by double differentiation of $\Psi\left( \phi_j \right)$ with respect to $\phi_{N_{p_T}}$ and $\phi_E$, thus: $$\begin{aligned}
\kappa_2^{N_{p_T},E} &=& \left(- i \frac{\partial } {\partial
\phi_{N_{p_T}}}\right) \left(- i \frac{\partial } {\partial
\phi_{E}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0}~= ~ \int
\limits_{\Delta p_T} dp_T ~ \int \limits_{0}^{2\pi} d\phi ~ \int
\limits_{-\infty}^{\infty} dy ~ \varepsilon ~ \frac{dN}{dp_T ~dy ~ d\phi}
\nonumber \\
&=& \frac{g~e^{-\frac{\mu}{T}}}{2\pi^2} \! \!\int \limits_{\Delta p_T} dp_T ~
p_T \left(p_T^2+m^2 \right) \left[ K_0 \left( \frac{\sqrt{p_T^2+m^2}}{T}
\right) + \frac{T}{\sqrt{p_T^2+m^2}} ~ K_1 \left( \frac{\sqrt{p_T^2+m^2}}{T}
\right) \right] ~.\end{aligned}$$ Correlations between conserved momenta and particles in $\Delta p_T$, given by the elements $\kappa_2^{N_{p_T},p_x}$,$\kappa_2^{N_{p_T},p_y}$, and $\kappa_2^{N_{p_T},p_z} $ are identical to zero, due to symmetry in azimuthal angle $\phi$ for the first two, and due to an antisymmetric rapidity integral for the last. Therefore, all elements involving an odd order in one of the momenta in Eq.(\[matrix\]) are equal to zero. The determinant of Eq.(\[matrix\]) thus factorizes, similar to Eq.(\[normdet\]), into a product of $(\kappa_2^{p_x,p_x})^3$ and the determinant of a $3
\times 3$ sub-matrix involving only terms containing $N_{p_T},E,Q$. Hence momentum conservation drops out when calculating Eq.(\[SimpleOmega\]). However the strength of correlations between particle number $N_{p_T}$ and globally conserved energy $E$ will depend on the position of the segment $\Delta p_T$. Thus using Eqs.(\[SimpleOmega\]) and (\[matrix\]), one can express the width of the MCE multiplicity distribution (\[PMCE\]) by: $$\label{omega_pt}
\omega^{mce}_{\Delta p_T}~=~
\frac{\kappa_2^{N_{p_T},N_{p_T}}}{\kappa_1^{N_{p_T}}} -
\frac{1}{\kappa_1^{N_{p_T}} \det |\hat \kappa_2|}\Bigg[
\left( \kappa_2^{N_{p_T},Q} \right)^2 \kappa_2^{E,E}+
\left( \kappa_2^{N_{p_T},E} \right)^2 \kappa_2^{Q,Q} - 2
\kappa_2^{N_{p_T},E} \kappa_2^{N_{p_T},Q} \kappa_2^{E,Q} \Bigg]$$ In Boltzmann approximation, we find from Eq.(\[kappa\_n\]), $\kappa_2^{N_{p_T},N_{p_T}} = \kappa_2^{N_{p_T},Q} = \kappa_1^{N_{p_T}} = q
\kappa_1^{N_{4\pi}}$, where we have defined the acceptance $q \equiv \kappa_1^{N_{p_T}} / \kappa_1^{N_{4\pi}} $. However, when observing a fraction $q$ of the particle density, one does not necessarily observe the same fraction $q$ of the energy density $\langle E_- \rangle$ carried by $\pi^-$, and thus $\kappa_2^{N_{p_T},E} \not= q \langle E_-
\rangle$. Therefore depending on the location of $\Delta p_T$, our detector sees a larger (smaller) fraction of the total energy, which leads to smaller (larger) particle number fluctuations, see Fig. \[dodp\], left panel. One can easily verify that setting $\kappa_2^{N_{p_T},E} = q \langle E_- \rangle $ in Eq.(\[omega\_pt\]), leads to acceptance scaling, Eq.(\[accscaling\]), $\omega^{mce}_{\Delta p_T} = 1 + q \left(\omega^{mce}_{4\pi}-1 \right)$.
Rapidity Segment {#App_y}
================
The average particle number density of $\pi^-$ in a rapidity interval ${\Delta y}$ is given by: $$\begin{aligned}
\label{dNdy}
\kappa_1^{N_y} &=& \left(- i \frac{\partial } {\partial
\phi_{N_{y}}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0}~=
~\int \limits_{m}^{\infty} dm_T ~ \int \limits_{0}^{2\pi} d\phi ~ \int
\limits_{\Delta y} dy ~ \frac{dN}{dm_T ~dy ~ d\phi} \nonumber \\
&=& \frac{g~e^{-\frac{\mu}{T}}}{\left(2\pi \right)^2}~ T^3~\int
\limits_{\Delta y} dy~ \exp \left( -\frac{m }{T} \cosh \left(y \right) \right)
\left[ \left(\frac{m}{T}\right)^2 + 2\frac{m}{T} \cosh^{-1} y +
2 \cosh^{-2} y \right]~.\end{aligned}$$ Please note that $\kappa_1^{N_{y}} = \int_{\Delta y} dy ~dN/dy =
\langle N_{y} \rangle$. Correlations of particles in ${\Delta y}$ with globally conserved energy are given by: $$\begin{aligned}
\kappa_2^{N_y,E} &=& \left(- i \frac{\partial } {\partial
\phi_{N_{y}}}\right) \left(- i \frac{\partial } {\partial
\phi_{E}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0}~= ~\int
\limits_{m}^{\infty} dm_T ~ \int \limits_{0}^{2\pi} d\phi ~ \int
\limits_{\Delta y} dy ~ \varepsilon ~ \frac{dN}{dm_T ~dy ~ d\phi} \nonumber \\
&=&\frac{g~e^{-\frac{\mu}{T}}}{\left(2\pi \right)^2} ~ T^4 ~ \int
\limits_{\Delta y} dy ~ \cosh y ~\exp \left(- \frac{m}{T} \cosh y \right)
\nonumber \\ && \times~ \left[\left(\frac{m}{T} \right)^3 + 3 \left(
\frac{m}{T} \right)^2 \cosh^{-1} y+ 6 ~\frac{m}{T}~ \cosh^{-2} y + 6
\cosh^{-3} y \right]~.\end{aligned}$$ The correlation term of particles in ${\Delta y}$ with globally conserved longitudinal momentum $P_z$ reads: $$\begin{aligned}
\kappa_2^{N_y,p_z} &=& \left(- i \frac{\partial } {\partial
\phi_{N_{y}}}\right) \left(- i \frac{\partial } {\partial
\phi_{p_z}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec 0}~= \int
\limits_{0}^{2\pi} d\phi ~ \int \limits_{m}^{\infty} dm_T ~ \int
\limits_{\Delta y} dy ~ p_z ~\frac{dN}{dm_T ~dy ~ d\phi} \nonumber \\
&=& \frac{g~e^{-\frac{\mu}{T}}}{\left(2\pi \right)^2} ~ T^4 ~ \int
\limits_{\Delta y} dy ~ \sinh y ~ \exp \left(- \frac{m}{T} \cosh y \right)
\nonumber \\ &&\times~
\left[\left(\frac{m}{T} \right)^3 + 3 \left( \frac{m}{T} \right)^2
\cosh^{-1} y + 6 ~\frac{m}{T}~ \cosh^{-2} y + 6 \cosh^{-3} y \right]~.\end{aligned}$$ Thus the element $\kappa_2^{N_y,p_z}$ in the matrix (\[matrix\]) is non-vanishing, and longitudinal momentum ($P_z$) conservation seems to affects correlations between particles in a segment $\Delta y$ and the remaining system. In contrast to that further elements are equal to zero, $\kappa_2^{N_y,p_x} =\kappa_2^{N_y,p_y} = 0$, and $P_x$ and $P_y$ conservation have no additional effect. When momentum conservation is taken into account the scaled variance (\[SimpleOmega\]) can be calculated from Eq.(\[calcdet\]): $$\begin{aligned}
\label{omega_y}
\omega^{mce}_{\Delta y} = \frac{\kappa_2^{N_{y},N_{y}}}{\kappa_1^{N_{y}}}
&-& \frac{1}{\kappa_1^{N_{y}} \kappa_2^{p_z,p_z} \det |\hat \kappa_2|}\Bigg[
\left( \kappa_2^{N_{y},Q} \right)^2 \kappa_2^{E,E} \kappa_2^{p_z,p_z}+
\left( \kappa_2^{N_{y},E} \right)^2 \kappa_2^{Q,Q} \kappa_2^{p_z,p_z}
\nonumber \\
&+&\left( \kappa_2^{N_{y},p_z} \right)^2 \left[ \kappa_2^{Q,Q}\kappa_2^{E,E}
- \left( \kappa_2^{E,Q} \right)^2 \right] -
2 \kappa_2^{p_z,p_z} \kappa_2^{N_{y},E} \kappa_2^{N_{y},Q}
\kappa_2^{E,Q}~\Bigg]~. \end{aligned}$$ Similarly to the previous section, we find a large (small) $\kappa_2^{N_{y},p_z}$ leads to small (large) fluctuations, see Fig. \[dodp\], right panel. When intervals symmetric in rapidity are assumed, e.g. $\Delta y = \left[-y_1,y_1 \right]$, or $ \Delta y =
\left[-y_2,-y_1\right] \cup \left[y_1,y_2 \right]$, correlations between particle number and momentum disappear, $\kappa_2^{N_{y},p_z}=0$, and Eq.(\[omega\_y\]) reduces to Eq.(\[omega\_pt\]), and momentum conservation does not play a role. Equally when disregarding longitudinal momentum conservation the same arguments as those of Appendix \[App\_pt\] apply and Eq.(\[omega\_pt\]) holds, however the effect is much weaker, see Fig. \[dodywopz\].
Azimuthal Angle Segment {#App_phi}
=======================
Th average particle number in $\Delta \phi $, while integrating over all $p_T$ and $y$ is simply a fraction $q = \Delta \phi / 2 \pi$ of the total yield $\langle N_{4\pi}\rangle$. Therefore $\kappa_1^{N_{\phi}} = q
\kappa_1^{N_{4\pi}}$. Equally, the energy carried by $\pi^-$ in this interval is $\kappa_2^{E,N_{\phi}}= q \langle E_-\rangle$. Due to symmetry around $y=0$, we find additionally $\kappa_2^{N_{\phi},p_z}= 0$. However for the transverse momenta $p_x= p_T \cos \phi$, and $p_y= p_T \sin \phi$ the correlation with $N_{\phi}$ is generally non-zero. $$\begin{aligned}
\kappa_2^{N_{\phi},p_x} &=& \left(- i \frac{\partial } {\partial
\phi_{N_{\phi}}}\right) \left(- i \frac{\partial } {\partial
\phi_{p_x}}\right) \Psi\left( \phi_j \right) \Bigg|_{\phi_j=\vec
0}~=\frac{g}{\left(2\pi \right)^3} \int
\limits_{\Delta \phi} d\phi ~ \int \limits_{0}^{\infty} dp_T ~ \int
\limits_{-\infty}^{\infty} dy ~ p_x ~\frac{dN}{dp_T ~dy ~ d\phi}
\nonumber \\ &=&
\int \limits_{\Delta \phi}d\phi \cos \phi ~
~\frac{2g~e^{-\frac{\mu}{T}}}{\left(2\pi
\right)^3}~m^2~T~\sqrt{\frac{\pi}{2}~m~T} ~ K_{5/2}
\left( \frac{m}{T} \right)~=~ \left( 2\pi
\right)^{-1} \Big[ \sin \phi \Big]_{\Delta \phi} ~\langle p_T \rangle ~.\end{aligned}$$ Similarly we find $\kappa_2^{N_{\phi},p_y} = - \left( 2\pi
\right)^{-1} \left[ \cos \phi \right]_{\Delta
\phi} ~\langle p_T \rangle $. Unlike in the previous sections there is no particular dependence of the position of the interval $\Delta \phi$. However in general there is a dependence. When momentum conservation is taken into account Eq.(\[SimpleOmega\]) can be calculated from Eq.(\[calcdet\]): $$\begin{aligned}
\label{omega_phi}
\omega^{mce}_{\Delta \phi} &=&
\frac{\kappa_2^{N_{\phi},N_{\phi}}}{\kappa_1^{N_{\phi}}} -
\frac{1}{\kappa_1^{N_{\phi}} \kappa_2^{p_x,p_x} \det |\hat
\kappa_2|} \Bigg[ \left( \kappa_2^{N_{\phi},Q} \right)^2 \kappa_2^{E,E}
\kappa_2^{p_x,p_x} + \left( \kappa_2^{N_{\phi},E} \right)^2
\kappa_2^{Q,Q} \kappa_2^{p_x,p_x}
\nonumber \\
&+&\left( \left( \kappa_2^{N_{\phi},p_x} \right)^2 + \left(
\kappa_2^{N_{\phi},p_y} \right)^2 \right) \left[
\kappa_2^{Q,Q}\kappa_2^{E,E}
- \left( \kappa_2^{E,Q} \right)^2 \right] -
2 \kappa_2^{p_x,p_x} \kappa_2^{N_{\phi},E} \kappa_2^{N_{\phi},Q}
\kappa_2^{E,Q}~\Bigg]~, \end{aligned}$$ where we have used $\kappa_2^{p_x,p_x}=\kappa_2^{p_y,p_y}$. As mentioned before there is no particular dependence of $\kappa_2^{N_{\phi},E}$ and $\kappa_2^{N_{\phi},Q}$ on the position of $\Delta
\phi$. However we have a term $ \left( \kappa_2^{N_{\phi},p_x} \right)^2 +
\left( \kappa_2^{N_{\phi},p_y} \right)^2 $. In case we assume a continuous interval $\Delta \phi_A = \left[\phi_1,\phi_2 \right]$ this terms reads: $$\left( \kappa_2^{N_{\phi},p_x} \right)^2 + \left( \kappa_2^{N_{\phi},p_y}
\right)^2 = \frac{\langle p_T \rangle^2}{\left( 2\pi \right)^2} ~\left[
1- \cos \left( \phi_1 - \phi_2 \right) \right]$$ This term is evidently positive, hence fluctuations are suppressed. One can easily verify that when one takes $\Delta \phi_B =\left[\phi_1,\phi_2\right]
\cup \left[\phi_1+\pi,\phi_2+\pi \right]$, i.e. two opposite slices in azimuthal angle, the correlation disappears, $\kappa_2^{N_{\phi},p_x}
=\kappa_2^{N_{\phi},p_y}= 0 $, and one returns to acceptance scaling, Eq.(\[accscaling\]).
[100]{}
J. Cleymans, H. Oeschler, K. Redlich and S. Wheaton, Phys. Rev. C [**73**]{} (2006) 034905. F. Becattini, J. Manninen and M. Gaździcki, Phys. Rev. C [**73**]{} (2006) 044905. A. Andronic, P. Braun-Munzinger and J. Stachel, Nucl. Phys. A [**772**]{} (2006) 167. J. Letessier and J. Rafelski, arXiv:nucl-th/0504028. B. Lungwitz [*et al.*]{} \[NA49 Collaboration\], PoS C [**FRNC2006**]{} (2006) 024. V. V. Begun, M. Gaździcki, M. I. Gorenstein, M. Hauer, V. P. Konchakovski and B. Lungwitz, Phys. Rev. C [**76**]{} (2007) 024902. V. V. Begun, M. Gaździcki, M. I. Gorenstein and O. S. Zozulya, Phys. Rev. C [**70**]{}, 034901 (2004). H. Heiselberg, Phys. Rept. [**351**]{} (2001) 161; S. Jeon and V. Koch, in [*Quark-Gluon Plasma*]{} 3, edited by R.C. Hwa and X.-N. Wang (World Scientific, Singapore, 2004), p.430. M. Gaździcki, M. I. Gorenstein and S. Mrowczynski, Phys. Lett. B [**585**]{} (2004) 115; M. I. Gorenstein, M. Gaździcki and O. S. Zozulya, Phys. Lett. B [**585**]{} (2004) 237. I. N. Mishustin, Phys. Rev. Lett. [**82**]{} (1999) 4779; H. Heiselberg and A. D. Jackson, Phys. Rev. C [**63**]{} (2001) 064904. M. A. Stephanov, K. Rajagopal and E. V. Shuryak, Phys. Rev. Lett. [**81**]{} (1998) 4816; Phys. Rev. D [**60**]{} (1999) 114028; M. Stephanov, Acta Phys. Polon. B [**35**]{} (2004) 2939. S. Häussler, S. Scherer and M. Bleicher, arXiv:hep-ph/0702188. T. K. Nayak, arXiv:0706.2708 \[nucl-ex\].
V. V. Begun, M. I. Gorenstein, M. Hauer, V. P. Konchakovski and O. S. Zozulya, Phys. Rev. C [**74**]{} (2006) 044903. F. Becattini, A. Keränen, L. Ferroni and T. Gabbriellini, Phys. Rev. C [**72**]{} (2005) 064904. M. I. Gorenstein, M. Hauer and D. O. Nikolajenko, Phys. Rev. C [**76**]{}, 024901 (2007). M. Hauer, V. V. Begun and M. I. Gorenstein, arXiv:0706.3290 \[nucl-th\]. V. V. Begun and M. I. Gorenstein, Phys. Rev. C [**73**]{} (2006) 054904; V. V. Begun, M. I. Gorenstein, A. P. Kostyuk and O. S. Zozulya, J. Phys. G [**32**]{} (2006) 935. K. Redlich and L. Turko, Z. Phys. C [**5**]{} (1980) 201; L. Turko, Phys. Lett. B [**104**]{} (1981) 153; R. Hagedorn and K. Redlich, Z. Phys. C [**27**]{} (1985) 541. J. Cleymans, K. Redlich and E. Suhonen, Z. Phys. C [**51**]{} (1991) 137; A. Keränen and F. Becattini, Phys. Rev. C [**65**]{} (2002) 044901; F. Becattini and U. W. Heinz, Z. Phys. C [**76**]{}, 269 (1997). O. Fochler, S. Vogel, M. Bleicher, C. Greiner, P. Koch-Steinheimer and Z. Xu, Phys. Rev. C [**74**]{}, 034902 (2006). C. M. Ko, V. Koch, Z. w. Lin, K. Redlich, M. A. Stephanov and X. N. Wang, Phys. Rev. Lett. [**86**]{}, 5438 (2001). S. Jeon, V. Koch, K. Redlich and X. N. Wang, Nucl. Phys. A [**697**]{}, 546 (2002). F. Becattini and L. Ferroni, Eur. Phys. J. C [**35**]{} (2004) 243; Eur. Phys. J. C [**38**]{} (2004) 225. J. Letessier, A. Tounsi and J. Rafelski, Phys. Lett. B [**475**]{} (2000) 213; J. Rafelski and J. Letessier, Phys. Rev. Lett. [**85**]{} (2000) 4695. P. Koch, B. Muller and J. Rafelski, Phys. Rept. [**142**]{} (1986) 167. V. V. Begun, M. I. Gorenstein and O. S. Zozulya, Phys. Rev. C [**72**]{} (2005) 014902. S. Jeon and V. Koch, Phys. Rev. Lett. [**83**]{}, 5435 (1999); G. Torrieri, S. Jeon and J. Rafelski, Phys. Rev. C [**74**]{} (2006) 024901; G. Torrieri, arXiv:nucl-th/0702062. B. Lungwitz and M. Bleicher, arXiv:0707.1788 \[nucl-th\]. E. L. Bratkovskaya, W. Cassing, C. Greiner, M. Effenberger, U. Mosel and A. Sibirtsev, Nucl. Phys. A [**681**]{} (2001) 84; L. V. Bravina [*et al.*]{}, Phys. Rev. C [**60**]{}, 024904 (1999). V. P. Konchakovski, M. I. Gorenstein and E. L. Bratkovskaya, Phys. Lett. B [**651**]{} (2007) 114. V. P. Konchakovski, M. I. Gorenstein and E. L. Bratkovskaya, arXiv:0704.1831 \[nucl-th\]. B. Lungwitz [*et al.*]{} \[NA49 Collaboration\], arXiv:0709.1646 \[nucl-ex\].
[^1]: We drop in the following the argument $(V,T,\mu)$ to simplify the notation.
[^2]: Please note, that in order to simplify formulas, the notation is slightly different from [@clt].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The MINOS experiment ran from 2003 until 2012 and collected a data sample including 10.71$\times10^{20}\,$ protons-on-target (POT) of beam neutrinos, 3.36$\times10^{20}\,$POT of beam antineutrinos and an atmospheric neutrino exposure of 37.88$\,$kt-yrs. The final measurement of the atmospheric neutrino oscillation parameters, [$\Delta m^2_{32}$]{} and [$\theta_{23}$]{}, came from a full three flavour oscillation analysis of the combined CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} beam and atmospheric samples and the CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} appearance samples. This analysis yielded the most precise measurement of the atmospheric mass splitting [$\Delta m^2_{32}$]{} performed to date. The results are $|$[$\Delta m^2_{32}$]{}$|=[2.28 - 2.46]\times10^{-3}\,$eV$^{2}$ (68%) and $\sin^{2}$[$\theta_{23}$]{}${}=0.35-0.65$ (90%) in the normal hierarchy, and $|$[$\Delta m^2_{32}$]{}$|=[2.32 - 2.53]\times10^{-3}\,$eV$^{2}$ (68%) and $\sin^{2}$[$\theta_{23}$]{}${}=0.34-0.67$ (90%) in the inverted hierarchy. The successor to MINOS in the NO$\nu$A era at FNAL, MINOS+, is now collecting data mostly in the $3-10\,$GeV region, and an analysis of [$\nu_\mu$]{} disappearance using the first 2.99$\times10^{20}\,$POT of data produced results very consistent with those from MINOS. Future data will further test the standard neutrino oscillation paradigm and allow for improved searches for exotic phenomena including sterile neutrinos, large extra dimensions and non-standard interactions.'
address: 'Department of Physics and Astronomy, University College London, Gower Street, London, WC1E 6BT, United Kingdom'
author:
- 'L. H. Whitehead'
bibliography:
- 'minosReview.bib'
title: Neutrino Oscillations with MINOS and MINOS+
---
Neutrino Oscillation,Long-baseline,MINOS,MINOS+
Introduction
============
Nearly two decades have passed since the first observation of neutrino oscillations by Super-Kamiokande [@superKPRL1998]. In that time it has become very clear from a number of experiments looking at neutrinos from the sun, the atmosphere, nuclear reactors and man-made neutrino beams that neutrinos can undergo oscillations from one flavour to another [@superKPRL2004; @superKPRD2010; @sno; @k2k; @minosPRL2006; @minosPRD2008; @t2kPRL2014NuE; @t2kPRL2014; @kamland; @borexino; @dayaBay2012; @reno2012; @doubleChooz2012], as described by the PMNS matrix [@mns1962; @ponte1968; @ponte1969]. The PMNS matrix, $U$, commonly parametrised by three mixing angles ([$\theta_{23}$]{}, [$\theta_{12}$]{} and [$\theta_{13}$]{}) and a *CP*-violating phase ([$\delta_{CP}$]{}), describes the mixing between the three weak flavour eigenstates, ${\ensuremath{| \nu_\alpha \rangle}} $, and mass eigenstates, [$| \nu_i \rangle$]{} in the following way:
$$\label{eq:flavMass}
{\ensuremath{| \nu_\alpha \rangle}} = \sum_{i=1}^3 U_{\alpha i} {\ensuremath{| \nu_i \rangle}}.$$
The three mixing angles have been measured to varying degrees of accuracy but the value of [$\delta_{CP}$]{} is still unknown. The oscillations arise from the quantum mechanical interference between the neutrino mass states and are driven by the mass-squared splittings between these mass states, $\Delta m^{2}_{ij}\equiv m^{2}_{i} - m^{2}_{j}$. It is possible to write down three mass-squared splittings, but only two are actually independent. One of the mass splittings, [$\Delta m^2_{21}$]{}, is considerably smaller than the others, meaning there are two scales at which oscillations can occur. The signs of the other mass-splittings, [$\Delta m^2_{32}$]{} and [$\Delta m^2_{31}$]{}, are currently unknown, meaning it is not known whether $m_3$ is the lightest or heaviest mass state. The case where it is the heaviest (lightest) is referred to as the normal (inverted) hierarchy. A final, important consequence of neutrino oscillations is the requirement that at least two of the neutrino mass states must be non-zero.
The two main oscillation channels of interest in long-baseline neutrino oscillation experiments are [$\nu_\mu$]{}$\rightarrow$[$\nu_\mu$]{} disappearance and [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} appearance. These channels were first probed using a man-made neutrino beam by the K2K experiment [@k2k; @k2kNuE2004]. The discovery of [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} oscillations was performed by T2K [@t2kPRL2014NuE] and [$\nu_\mu$]{}$\rightarrow$[$\nu_\tau$]{} appearance was discovered by the OPERA experiment [@operaPRL2015]. Oscillations in such experiments are driven by the two larger mass-splittings, [$\Delta m^2_{32}$]{} and [$\Delta m^2_{31}$]{}. Using a two neutrino approximation, with the parameters $\Delta m^{2}$ and $\sin^{2}2\theta$, the [$\nu_\mu$]{} disappearance probability for a neutrino with energy E and travelling over a distance L in the vacuum can be written as follows:
$$\label{eq:2flavDis}
P(\nu_\mu\rightarrow\nu_\mu) = 1 - \sin^{2}2\theta\sin^{2}\left(\frac{\Delta m^{2} L}{4E}\right).$$
However, [$\theta_{13}$]{} was measured by Daya Bay [@dayaBay2012] and later by RENO [@reno2012] and Double CHOOZ [@doubleChooz2012] and is hence known to be reasonably large. In addition, the uncertainty on measurements of [$\Delta m^2_{32}$]{} is of the same order as the size of [$\Delta m^2_{21}$]{}. The more accurate formalism requires the use of the full three flavour oscillation probabilities and the approximate parameters $\Delta m^{2}$ and $\sin^{2}2\theta$ in Eq. \[eq:2flavDis\] are modified in the following way [@3flavForm]:
$$\begin{aligned}
\label{eq:3flavDis}
\sin^{2}2\theta &= 4\cos^{2}\theta_{13}\sin^{2}\theta_{23}(1-\cos^{2}\theta_{13}\sin^{2}\theta_{23}), \nonumber \\
\Delta m^{2} = \Delta m^{2}_{32} + &\sin^{2}\theta_{12}\Delta m^{2}_{21} + \cos\delta_{CP}\sin\theta_{13}\sin2\theta_{12}\tan\theta_{23}\Delta m^{2}_{21}.\end{aligned}$$
The expressions given in Eq. \[eq:3flavDis\] illustrate how the interference between the two different mass-splitting terms causes the full oscillation probability to depend on all of the parameters of the PMNS matrix. It can be seen in Eq. \[eq:2flavDis\] that the two flavour oscillation probability is symmetric under the transformations of $\theta \rightarrow \frac{\pi}{2} - \theta$ and $\Delta m^{2} \rightarrow -\Delta m^{2}$. The equivalent parameter shifts for the three flavour case in Eq. \[eq:3flavDis\] are $\theta_{23} \rightarrow \frac{\pi}{2} - \theta_{23}$ and $\Delta m^{2}_{32} \rightarrow -\Delta m^{2}_{32}$, and it can be seen that the oscillation probability is not completely symmetric under these transformations, leading to approximate degeneracies instead of symmetries.
When neutrinos traverse matter, the Hamiltonian associated with the propagation is modified compared to that of the vacuum by interactions of the neutrinos with the matter. All three neutrino flavours can undergo neutral-current (NC) interactions with the matter via the exchange of a $Z$ boson but since this affects all neutrinos equally, it does not cause a change in the oscillations. However, only electron neutrinos can have charged-current (CC) interactions with the electrons in the matter via the exchange of a $W$ boson, giving rise to a change in the oscillations. This phenomenon is known as the MSW effect [@wolf1978; @ms1986]. In this case, $\theta_{13}$ is replaced by a modified mixing angle $\theta_M$ as defined by [@matterShift]
$$\label{eq:matterEffect}
\sin^{2}2\theta_{M}=\frac{\sin^{2}2\theta_{13}}{\sin^{2}2\theta_{13} + (\cos2\theta_{13}-A)^2},$$
where $A = 2\sqrt{2}G_{F}n_{e}E/\Delta m^{2}_{31}$, $G_{F}$ is the Fermi weak coupling constant and $n_{e}$ is the electron density. In the case of antineutrinos, the value of $A$ changes from $A\rightarrow-A$. It is clear to see that when $\cos2\theta_{13} = A$, the value of $\sin^{2}2\theta_{M}$ becomes maximal, producing a resonance in the oscillation probability for [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} oscillations, and hence modifies the [$\nu_\mu$]{} disappearance probability as well as the [$\nu_e$]{} appearance probability. This resonance occurs in multi-GeV atmospheric neutrino events that travel upwards through the earth’s mantle, and since $A$ is dependent on the sign of $\Delta m^{2}_{31}$, it provides a handle with which to study the neutrino mass hierarchy.
The [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} oscillation probability in matter, calculated up to second order in $\alpha = \Delta m^{2}_{21}/\Delta m^{2}_{32}$ is given by the following expression [@nueForm]:
$$\begin{aligned}
\label{eq:3flavApp}
P(\nu_\mu\rightarrow\nu_e) \approx \sin^{2}\theta_{23}\sin^{2}&2\theta_{13} \frac{\sin^{2}\Delta(1-A)}{(1-A)^2} +\alpha\tilde{J}\cos(\Delta\pm\delta_{CP})\frac{\sin\Delta A}{A}\frac{\sin\Delta(1-A)}{(1-A)} \nonumber \\
&+\alpha^2\cos^{2}\theta_{23}\sin^{2}2\theta_{12}\frac{\sin^{2}\Delta A}{A^{2}},\end{aligned}$$
with $\tilde{J}=\cos\theta_{13}\sin2\theta_{13}\sin2\theta_{12}\sin2\theta_{23}$ and $\Delta = \Delta m^2 _{31} L/4E$. The positive sign of [$\delta_{CP}$]{} in the second term refers to neutrinos, whilst the negative sign corresponds to antineutrinos. Equation \[eq:3flavApp\] shows that the [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} appearance channel is sensitive to: the octant of $\theta_{23}$ through the first term, the *CP*-violating phase through the presence of [$\delta_{CP}$]{} in the second term, and the mass hierarchy from the matter effect parameter, $A$.
The MINOS/MINOS+ Experiment
===========================
The Main Injector Neutrino Oscillation Search (MINOS) experiment was originally designed in order to accurately measure the atmospheric parameters of neutrino oscillations, namely [$\theta_{23}$]{} and [$\Delta m^2_{32}$]{}. MINOS began collecting atmospheric neutrino data in 2003 and beam data-taking began in 2005. The experiment ran until May 2012 when the beam was shut off in order to prepare for the NO$\nu$A experiment. At this point, MINOS transitioned into MINOS+, the name of the experiment going into the NO$\nu$A era. MINOS+ began collecting beam data in September 2013 when the beam switched back on.
The NuMI Beam
-------------
The Neutrinos at the Main Injector (NuMI) beam [@numiBeamPaper] is the neutrino beam at the Fermi National Accelerator Laboratory (FNAL) that supplied neutrinos for MINOS, and currently produces neutrinos for NO$\nu$A, MINER$\nu$A and MINOS+.
The main components of the NuMI beam are shown in Fig. \[fig:numi\]. Protons with an energy of 120$\,$GeV are extracted from the Main Injector (MI) proton accelerator and are steered onto a graphite target. The spray of hadrons, primarily pions and kaons, resultant from the collisions of the protons with the carbon nuclei are focussed by two current-pulsed magnetic horns and directed into the decay pipe.
![\[fig:numi\]Schematic diagram of the NuMI beam line showing the major components with their corresponding size (not to scale). The beam line is actually oriented downwards into the ground at an angle of 58$\,$mrad to the horizontal in order to point towards the Far Detector. The Near Detector cavern is downstream of the last section of rock on the right side of the figure.](figure1.pdf)
The focussed hadrons then decay at some point along the 675$\,$m decay pipe to produce the neutrino beam. Muon neutrinos are mostly produced by the following decays:
$$\label{eq:piDecay}
\pi^{+} \rightarrow \mu^{+} + \nu_\mu$$
$$K^{+} \rightarrow \mu^{+} + \nu_\mu$$
and electron neutrinos by the decays of kaons and tertiary muons:
$$\label{eq:muDecay}
\mu^{+} \rightarrow e^{+} + \nu_e + \overline{\nu}_\mu$$
$$K^{+} \rightarrow e^{+} + \nu_e + \pi^{0}$$
$$K^0_L \rightarrow e^{+} + \nu_e + \pi^{-}.$$
The charge conjugate processes also exist to produce the antineutrinos but are suppressed by the focussing of positively charged mesons in nominal neutrino beam mode.
Downstream of the decay pipe is the hadron monitor that measures the spatial distribution of any remaining hadrons. The absorber, formed from an aluminium core, with a steel and concrete surround, is located downstream of the hadron monitor and stops any remaining hadrons (mostly protons from the beam that did not interact and some mesons that did not decay in the decay pipe). It is only the muons and neutrinos that pass through the absorber, and the muons are then measured with three muon monitors interspersed with regions of the natural dolomite rock. The total 240$\,$m of rock upstream of the Near Detector (ND) cavern stops all of the muons, leaving a beam consisting only of neutrinos and antineutrinos.
It is possible to change the position of the target and the magnetic horns in order to change the energy distribution of the beam. The vast majority of the MINOS data were taken in the Low Energy (LE) beam configuration where the target was partially inserted into the first magnetic horn, giving a neutrino beam peaked at approximately 3$\,$GeV.
In standard neutrino mode operation the magnetic horns are set such that they focus positively charged mesons, resulting in a neutrino beam and known as Forward Horn Current (FHC) running. It is possible to reverse the current used to pulse the horns in order to focus the negatively charged mesons to produce a beam with an enhanced antineutrino component, a configuration known as Reverse Horn Current (RHC). In FHC (or [$\nu_\mu$]{}-dominated) mode, the beam consists of 91.7% [$\nu_\mu$]{}, 7.0% [$\overline{\nu}_\mu$]{}, and 1.3% [$\nu_e$]{} $+$ [$\overline{\nu}_e$]{} and in RHC (or [$\overline{\nu}_\mu$]{}-enhanced) mode, 58.1% [$\nu_\mu$]{}, 39.9% [$\overline{\nu}_\mu$]{}, and 2.0% [$\nu_e$]{} $+$ [$\overline{\nu}_e$]{} [@minosApp2012]. Short periods of data were taken in other configurations in order to study the beam.
The NuMI beam supplied a total of 10.71$\times10^{20}$ protons-on-target (POT) in FHC mode and 3.36$\times10^{20}$ POT in RHC mode to the MINOS experiment. Figure \[fig:numiPOT\] shows the number of protons delivered per week and the total accumulated POT as a function of time from May 2005 until May 2012. The POT in FHC mode is shown in green, and the orange shows the data in RHC mode. Short special runs, such as those with the magnetic horns turned off or at higher energy, are shown in red. As of September 2013, the NuMI beam is operated in Medium Energy (ME) mode to supply neutrinos to NO$\nu$A, an experiment that is off-axis from the beam, and MINOS+. The ME beam has a peak at about 6$\,$GeV on-axis for MINOS+.
![\[fig:numiPOT\]The total accumulated POT (blue line) and the number of protons per week (histograms) delivered by the NuMI beam from May 2005 until May 2012. The green regions show the POT delivered in standard LE mode in the FHC configuration. The orange shows the antineutrino running, and the red shows special short runs in different configurations. Figure from Ref. [@numiBeamPaper].](figure2.pdf)
Beam Flux Simulation {#sec:beamFluxSim}
--------------------
The neutrino beam flux is simulated using a combination of the GEANT4 [@geant4] geometry package and the FLUKA [@flugFluka] hadron production package known as FLUGG [@flugFluka]. Figure \[fig:fluxConfig\] shows the true energy distribution for simulated events in the ND for the LE (solid), ME (dashed) and pseudo high energy(pHE) (dotted) beam configurations. The beam simulation does not, however, provide a perfect description of the neutrino flux that is measured at the ND. As such, the ND is used to constrain the simulation, since there are underlying uncertainties, particularly in the hadron production in the target, that cause disagreement between data and simulation. This method is described in detail in Ref. [@minosPRD2008], but is outlined below.
![\[fig:fluxConfig\]The true energy distribution of neutrino interactions in the ND from the simulation of three beam configurations. The solid line corresponds to LE running, the mode in which most MINOS data were taken. The dashed line shows the ME spectrum, as used in MINOS+, and the dotted line corresponds to pHE. Figure from Ref. [@ashleyReview].](figure3.pdf)
Each bin of the reconstructed CC [$\nu_\mu$]{} energy spectrum in the ND contains neutrinos coming from the decay of meson parents that had given values of the transverse $(p_t)$ and longitudinal $(p_z)$ momentum components as they left the target. The different $(p_t,p_z)$ bins contribute to different energy bins, meaning that the ND data can be used to constrain the hadron production in the target. Special data samples were taken with different beam configurations to enhance the range of the $(p_t,p_z)$ space covered by the ND data to provide more accurate fits. This tuning procedure allows for the calculation of weights that are applied to the simulation, providing a more accurate description of the data.
The MINOS Detectors
-------------------
MINOS has two functionally equivalent detectors [@minosNIM] called the Near Detector and the Far Detector (FD). The ND, located at FNAL, lies 1.04$\,$km from the target and has a mass of about 1$\,$kt. The FD has a mass of 5.4$\,$kt, is located 705$\,$m (2070$\,$m water-equivalent) underground in the Soudan Mine, Minnesota, at a distance of 735$\,$km from the target.
The two detectors are magnetised steel/scintillator sampling calorimeters. They are formed from interleaved planes of 2.54$\,$cm steel to provide the interaction target mass and 1$\,$cm plastic scintillator to provide the active region of the detectors. The plastic scintillator planes are formed from bars that are 1$\,$cm $\times$ 4.1$\,$cm in cross-section and vary in length from $2.5\,$m to $8.0\,$m, and are read out via a wavelength shifting optical fibre that is embedded into the surface of the scintillator bars. The wavelength shifting fibres are read out using multi-anode PMTs, and are instrumented on one (both) ends for the ND (FD). The orientation of the scintillator bars on alternating planes are at $45^\circ$ and $-45^\circ$ to the vertical, to provide two orthogonal views, which along with the depth into the detector along the beam direction, provide 3D tracking information.
The magnetic field in each of the detectors is toroidal and provided by a current-carrying coil that passes through the middle of the detectors. The current direction can be reversed in order to change the charge of the particles that are bent inwards to the centre of the detector. The direction of the current is chosen to match the current in the magnetic horns such that negatively charged muons are focussed when running in FHC mode, and positively charged muons are focussed in RHC mode. The particles focussed into the detector generally have better energy resolution since range can be used to measure the momentum of contained particles, and for those that exit the detector, the longer path-length enables a better determination of the energy by curvature. The average magnetic field strength in the ND and FD is 1.3$\,$T and 1.4$\,$T, respectively [@minosNIM].
Neutrino Interactions in the Detectors {#sec:minosInt}
--------------------------------------
There are three main types of neutrino interactions expected in the MINOS detectors:
- CC [$\nu_\mu$]{} and [$\overline{\nu}_\mu$]{}: The muon neutrino interacts with a nucleus $X$ via the exchange of a $W$ boson in the process $\nu_\mu + X \rightarrow \mu^{-} + X'$. These events are characterised by the track-like energy deposits caused by the muon, in addition to a hadronic shower at the interaction vertex. The separation of CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} interactions is performed by using the curvature to measure the sign of the muon charge.
- NC $\nu$: A neutrino scatters off a nucleus $X$ via $Z$ boson exchange $\nu + X \rightarrow \nu + X'$. Neutral current interactions appear purely as a hadronic shower, for all three neutrino flavours, since the scattered neutrino is not detected. With no charged lepton resultant from the interaction, it is not possible to distinguish between NC events involving different neutrino flavours.
- CC [$\nu_e$]{} and [$\overline{\nu}_e$]{}: The electron neutrino interacts with a nucleus $X$ via $W$ boson exchange in the process $\nu_e + X \rightarrow e^{-} + X'$. These events appear as a small electromagnetic shower, and since the electron does not have a track-like topology, no separation between CC [$\nu_e$]{} and [$\overline{\nu}_e$]{} interactions can be made.
While a small number of CC [$\nu_\tau$]{} events occur in the FD at high energy, it is very difficult to distinguish them from the event types listed above, such that no event selection is attempted.
Muon Neutrino Disappearance
===========================
MINOS can measure the atmospheric oscillation parameters [$\Delta m^2_{32}$]{} and [$\theta_{23}$]{} by looking for the disappearance of muon neutrinos. MINOS is sensitive to CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} interactions from two sources: the NuMI beam and atmospheric neutrinos. This section outlines the methods used to select samples of these events from the two different sources.
Beam Neutrinos {#sec:beamDis}
--------------
The method employed by the beam neutrino analysis is to use the ND to predict the expected FD reconstructed neutrino energy distribution for a given set of oscillation parameters in order to find the best fit values of the parameters. Beam muon neutrinos are selected in the ND and FD by looking for the track-like signature of the muon in charged-current [$\nu_\mu$]{} interactions. The complete sample consists of selections of the following types of events:
- CC [$\nu_\mu$]{} interactions in the FHC beam.
- CC [$\overline{\nu}_\mu$]{} interactions in the FHC beam.
- CC [$\nu_\mu$]{} rock and anti-fiducial muons (RAF) in the FHC beam.
- CC [$\overline{\nu}_\mu$]{} interactions in the RHC beam.
All samples apart from the RAF selection require that the interaction vertex lies within the fiducial volume of the detector. The RAF selection aims to select those neutrino-induced muons that traverse the detector from neutrino interactions in the rock upstream of, and surrounding, the detector cavern, as well as those interactions that occur outside of the fiducial volume, close to the edge, of the detector. As such, RAF events consist only of a muon track that enters the detector from the outside, or a CC [$\nu_\mu$]{} interaction that occurred inside, but very close to the edge of the detector. In either case, only the muon is considered [@rafThesis].
Firstly, candidate events are considered only if they are in time with the beam and contain a track-like energy deposit. Four variables that describe the topology and energy deposition of the track are used as inputs to a k-Nearest-Neighbour (kNN) algorithm that produces as output a value between 0 and 1 that acts as the particle identification (PID) variable [@pidThesis]. This PID is used to preferentially select the CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} events over the NC background events, which rarely contain an extended track-like structure.
The neutrino energy is measured as the sum of the muon energy and the hadronic shower energy. The muon energy is measured using the range of the muon in the case that it is fully contained within the detector, and using curvature in the magnetic field if it exits the detector. The hadronic shower energy is measured using a kNN that looks at aspects of the shower profile to return the energy. This method was found to give an improved energy resolution over using a pure calorimetric energy measurement (as used in the first two MINOS analyses), reducing the energy resolution from 55% to 43% for 1.0-1.5$\,$GeV showers, for example [@kNNShwThesis].
The selected fiducial events are binned as a function of reconstructed neutrino energy, and those CC [$\nu_\mu$]{} events in the FHC beam are further divided by their estimated energy resolution to improve sensitivity [@kNNShwThesis; @jmThesis; @sjcThesis]. For the RAF events, only the muon energy is considered and hence the events are binned in reconstructed muon energy only.
The predicted FD CC [$\nu_\mu$]{} or CC [$\overline{\nu}_\mu$]{} energy spectrum is calculated using a combination of simulation and the ND data. The beam flux simulation and the method used to tune it based on the ND data was described in Section \[sec:beamFluxSim\]. The transport of the simulated particles through the detector simulation is performed by the GCALOR [@gcalor] and GEANT3 [@geant3] packages. The process to calculate the FD prediction, known as the extrapolation procedure, consists of the following steps. Firstly, the event selection is performed at the ND for both data and simulation. The simulation is used to produce a matrix that converts between the reconstructed and true neutrino energy. The selected ND data are then multiplied by this matrix to convert to a pseudo-true energy. At this stage a correction is also applied to account for the selection efficiency in the ND. The next step applies the beam matrix, a correction that accounts for the difference in acceptance of the neutrino beam between the two detectors (the beam appears as a point source for the FD, whereas the ND sees an extended source). With the energy spectrum now in pseudo-true energy, the neutrino oscillations are applied using the three flavour (or historically, two flavour) oscillation formalism. Finally, the FD selection efficiency is applied and the energy is converted back to reconstructed neutrino energy using the FD version of the reconstructed to true energy conversion matrix. The resultant energy spectrum from this process is the predicted spectrum for the FD for the given oscillation parameters used in the extrapolation procedure.
Atmospheric Neutrinos {#sec:atmosDis}
---------------------
The atmospheric neutrino selection aims to select those neutrinos produced in cosmic ray interactions in the upper atmosphere. These interactions give rise to both muon and electron type neutrinos from the decay of pions and muons, as shown previously for the production of the neutrino beam in Eqs \[eq:piDecay\] and \[eq:muDecay\]. The total exposure to atmospheric neutrinos over the lifetime of the MINOS experiment amounted to 37.88$\,$kt-yrs.
The atmospheric neutrino sample is collected exclusively by the FD since the location deep in the Soudan Mine gives a large reduction in the background events coming from cosmic rays. The atmospheric neutrino interactions are then separated from the remaining cosmic background by looking for events that have their interaction vertex inside the fiducial volume (contained-vertex sample), or by looking for muon-like, upward-going events entering the detector from the bottom region of the detector (non-fiducial muons) [@chThesis; @abThesis; @jdcThesis]. The cosmic ray background in the contained-vertex sample is further reduced by checking for activity in the cosmic ray veto shield associated with the main detector event.
The contained-vertex and non-fiducial muon selections are divided into candidate CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} samples depending on the measured charge of the muon in the event. In the two-flavour MINOS analyses, these data were binned as a function of $\log_{10}(L/E)$ but the binning scheme was changed for the three flavour analysis. In the three flavour analysis, the data are binned in two dimensions as a function of $\log_{10}(E)$ and the zenith angle $\cos\theta_z$. This scheme was chosen in order to maximise the sensitivity to the MSW effect, and hence the mass hierarchy, that modifies the oscillation probability as the neutrinos travel through the interior of the earth, where the distance travelled depends on the measured value of $\cos\theta_z$.
Lastly, a selection is made to identify contained-vertex shower events. This selection consists mainly of NC $\nu$, CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} interactions. These events have limited sensitivity to neutrino oscillations but are all included in the fit in a single bin to constrain the absolute atmospheric neutrino flux [@speakmanThesis].
The simulation of atmospheric neutrinos is based on the Bartol flux predictions [@bartolFlux]. Atmospheric neutrinos that interact inside the fiducial volume, the contained-vertex sample, are simulated using NEUGEN3 [@neugen] in the same way as for the beam neutrinos. NUANCE [@nuance] is used to simulate the interaction of the atmospheric neutrinos in the rock surrounding the cavern and to propagate the final state particles up to the edge of the detector. The simulation of the particles in the detector is then the same for both samples, using the GCALOR [@gcalor] and GEANT3 [@geant3] packages, in exactly the same way as for the beam neutrino simulation. Efforts are made to account for the change in atmospheric neutrino fluxes as a function of time due to variations in the solar cycle, an important consideration since the period over which data were collected covers nearly an entire solar cycle. It is predicted that the atmospheric neutrino flux can vary by up to 7% over this period [@minosAtmos2012].
Oscillations are applied to the FD predicted energy spectra using a reweighting technique for all of the data samples listed in Sections \[sec:beamDis\] and \[sec:atmosDis\]. The process includes the addition of the backgrounds from [$\nu_\tau$]{} and [$\overline{\nu}_\tau$]{} appearance. The oscillations applied were historically those derived from the two-flavour approximation, but in the final MINOS analysis described in this article the full three flavour formalism was used. The oscillation parameters are then varied during the fit in order to extract the parameters that provide the best fit to the data.
Two Flavour Oscillation Results
-------------------------------
Oscillations are applied to the FD predicted energy spectra using a reweighting technique for all of the data samples listed in Sections \[sec:beamDis\] and \[sec:atmosDis\]. The process includes the addition of the backgrounds from [$\nu_\tau$]{} and [$\overline{\nu}_\tau$]{} appearance. The early analyses performed by MINOS, as well as other experiments, considered the oscillations in terms of an approximate two neutrino case. In the limit that [$\theta_{13}$]{} tends to zero, all but one of the additional terms shown in Eq. \[eq:3flavDis\] that modified the two-flavour approximation vanish, leaving just $\Delta m^2 = \Delta m^2_{32} + \sin^{2}\theta_{12}\Delta m^2_{21}$. Since $\sin^{2}\theta_{12}\Delta m^2_{21} / \Delta m^2_{32} \approx 0.01$ then [$\Delta m^2_{21}$]{} could easily be ignored when measurements of $\Delta m^{2}$ were much less accurate than 1%. Thus the sector governing [$\nu_\mu$]{}$\rightarrow$[$\nu_\mu$]{} oscillations could be considered as decoupled from the solar scale oscillations.
The first measurement from MINOS of $\Delta m^2$ and $\sin^{2}2\theta$ was made in 2006 using 1.27$\times$10$^{20}\,$POT [@minosPRL2006] and was followed by updated analyses using exposures of 3.36$\times$10$^{20}\,$POT [@minosDis2008] and 7.25$\times$10$^{20}\,$POT [@minosDis2011]. The first measurement of the antineutrino oscillation parameters $\Delta \overline{m}^2$ and $\sin^{2}2\overline{\theta}$ was made in 2008 using 1.71$\times$10$^{20}\,$POT [@minosAntiDis2011] and was followed by a further analysis using an exposure of 2.95$\times$10$^{20}\,$POT [@minosAntiDis2012]. Measurements of both the neutrino and antineutrino oscillation parameters were also made using just the atmospheric neutrino oscillation sample, exploiting the complete MINOS atmospheric neutrino sample of 37.88$\,$kt-years [@minosAtmos2012].
The final two-flavour fit [@minosNuNubar] considered the full 10.71$\times$10$^{20}\,$POT in FHC mode, 3.36$\times$10$^{20}\,$POT in RHC mode and the 37.88$\,$kt-years of atmospheric neutrinos. This fit was performed both using four parameters (meaning that different oscillation parameters were used to fit the neutrinos and antineutrinos) and two parameters (where neutrinos and antineutrinos are assumed to oscillate in the same way). The values and 1$\sigma$ uncertainties of $\Delta m^2$ measured from these analyses are summarised in Fig. \[fig:minosHistory\], showing good agreement between measured values for neutrinos and anti-neutrinos with the full exposure. Figure \[fig:dmdmbar\] explicitly shows the agreement between the values of $\Delta m^2$ and $\Delta\overline{m}^{2}$ measured in the four parameter version of the final fit, and hence that the parameters that govern oscillations of neutrinos and antineutrinos are the same within the uncertainty of the measurement, allowing all of the samples to be considered together to fit the parameters of the PMNS matrix in the full three flavour fit described in Section \[sec:3flavResults\]. The values of $\Delta m^2$ and $\Delta\overline{m}^{2}$ were the most accurate measurements made of the two flavour mass-splitting, but the values of $\sin^{2}2\theta$ and $\sin^{2}2\overline{\theta}$ had a larger uncertainty compared to those measured by Super-K [@superKPRL2011].
![\[fig:minosHistory\]The values of $|\Delta m^{2}|$ and $|\Delta\overline{m}^{2}|$ as measured by MINOS using the two-flavour approximation throughout the lifetime of the experiment [@minosPRL2006; @minosDis2008; @minosDis2011; @minosNuNubar; @minosAntiDis2011; @minosAntiDis2012; @minosAtmos2012]. The black points show measurements made using CC [$\nu_\mu$]{} interactions, and those in green using CC [$\overline{\nu}_\mu$]{} interactions. The two red points show the combination of CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} events under the assumption that the oscillation parameters are identical between neutrinos and antineutrinos. The $x$-axis provides details of the beam and atmospheric neutrino exposure used to produce the measurement.](figure4.pdf)
![\[fig:dmdmbar\]A comparison of the MINOS measurements of the two-flavour oscillation mass-splitting governing long-baseline [$\nu_\mu$]{} and [$\overline{\nu}_\mu$]{} disappearance, $|\Delta m^{2}|$ and $|\Delta\overline{m}^{2}|$, showing both the 68% (red) and 90% (blue) confidence limit contours [@minosNuNubar]. ](figure5.pdf)
Electron Neutrino Appearance
============================
MINOS is also able to search for the subdominant appearance of electron neutrinos in the muon neutrino beam. This channel, being subdominant, must always been considered in the case of three neutrino flavours. The main aim of the search is to perform a measurement of [$\theta_{13}$]{}, with the main measurable being given by the first term in Eq. \[eq:3flavApp\] as $\sin^{2}\theta_{23}\sin^{2}2\theta_{13}$.
The preselection of candidate events begins with the requirement that the events must occur in time with the neutrino beam. Additionally, the events must be consistent in direction with the beam. Shower-like topology events are then selected by requiring that the event does not have a track-like object of at least 25 planes, or extending at least 15 planes from the edge of the shower. The events are also required to have at least five consecutive planes with energy deposits of at least one half of the energy deposit expected from a minimum ionising particle. Only the region in energy where the majority of [$\nu_e$]{} and [$\overline{\nu}_e$]{} appearance is expected is used in the analysis, limiting the allowed reconstructed neutrino energy to be within the range from 1 to 8$\,$GeV.
The candidate CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} interactions are then identified using the library-event-matching (LEM) method whereby each data event is compared on a hit-by-hit basis to a vast library of 20 million simulated signal (CC [$\nu_e$]{} for FHC beam data or CC [$\overline{\nu}_e$]{} for RHC beam data) and 30 million background (NC) events [@jporThesis; @rbtThesis; @apsThesis]. As discussed in Section \[sec:minosInt\], it is not possible separate CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} interactions in this analysis. The best 50 matches to the data event are used to calculate a series of variables that are combined to form a single PID variable, called $\alpha_{LEM}$, using an artificial neural network. All those events with values of $\alpha_{LEM}$ above 0.6 are selected as part of the analysis, a cut value defined to maximise the sensitivity to the [$\nu_e$]{} and [$\overline{\nu}_e$]{} appearance signal.
Selected events are binned in two dimensions as a function of reconstructed neutrino energy and $\alpha_{LEM}$. The bins with values of $\alpha_{LEM}$ closer to one have the most sensitivity to oscillations since they have the highest purity of CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} events.
FD Prediction
-------------
Due to the absence of [$\nu_e$]{} and [$\overline{\nu}_e$]{} appearance in the ND, different methods are used to predict the expected background and signal components of the FD energy spectrum. The three main backgrounds to the appearance signal come from NC interactions, CC [$\nu_\mu$]{} or CC [$\overline{\nu}_\mu$]{} events, and intrinsic beam CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} interactions. These three backgrounds are measured using the ND using the same binning scheme used in the main event selection for both data and simulation [@joaoThesis]. A selection is then performed using simulation at the FD, and for each bin in energy and $\alpha_{LEM}$ the bin content for each background component is multiplied by a correction factor from the ND equal to the ratio of the number of data to simulation events. The small [$\nu_\tau$]{} and [$\overline{\nu}_\tau$]{} appearance background must be calculated in a different way since, like the CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} appearance, it does not occur in the ND. This background is derived from the simulation and then corrected using the ND measurement of CC [$\nu_\mu$]{} or CC [$\overline{\nu}_\mu$]{} events.
The ND can not be used directly to measure the signal efficiency due to the lack of an appearance signal at such a short baseline. Instead, a sample of CC [$\nu_\mu$]{} interactions are selected from data. The energy deposits in these interactions arising from the muon are then removed from the event [@amhThesis] and replaced with energy deposits from a simulated electron shower [@jaabThesis]. The simulated electron vertex, direction and energy are set to match those of the reconstructed muon in the original data event. This procedure makes an effective sample of CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} data events that can be used to study the efficiency of selecting and identifying the signal events. This method was validated using those events that are not sensitive to the appearance signal, defined by $\alpha_{LEM} < 0.5$, to predict the number of events in the same region of the FD data and agreement was found within the 0.3(0.6)$\sigma$ of the statistical uncertainty for the CC [$\nu_e$]{}([$\overline{\nu}_e$]{}) sample [@minosApp2012].
Results
-------
MINOS has performed two [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} searches on data samples of 7.01$\times$10$^{20}\,$POT [@minosApp2010] and 8.20$\times$10$^{20}\,$POT [@minosApp2011], and a combined [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} and [$\overline{\nu}_\mu$]{}$\rightarrow$[$\overline{\nu}_e$]{} search based on a total exposure of 10.6$\times$10$^{20}\,$POT FHC and 3.3$\times$10$^{20}\,$POT RHC [@minosApp2012]. The POT of the final analysis does not agree exactly with those quoted for the muon neutrino disappearance analysis because the short high energy run is not included in the electron neutrino appearance analysis as it has no sensitivity to the appearance signal. The result of the combined [$\nu_e$]{} and [$\overline{\nu}_e$]{} appearance search is shown on the left of Fig. \[fig:nueAppContour\], excluding the null hypothesis of no appearance at approximately the 96% confidence level. The contours are shown for the normal (top) and inverted (bottom) hierarchy for the lower octant of [$\theta_{23}$]{}. The best fit is also shown for the upper octant, showing little sensitivity to the octant of [$\theta_{23}$]{}. This analysis found the value of [$\theta_{13}$]{} to be greater than zero with less significance than the T2K result from 2011 [@t2kPRL2011] and the reactor experiments from 2012 [@dayaBay2012; @reno2012; @doubleChooz2012].
![\[fig:nueAppContour\]Left: the [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} appearance contour shown as a function of $2\sin^{2}2\theta_{13}\sin^{2}\theta_{23}$ and [$\delta_{CP}$]{}. The normal hierarchy is shown in the top panel, and the inverted hierarchy below, with the 68% and 90% contours shown for the lower octant of [$\theta_{23}$]{}. The best fit curve is also shown for the upper octant, showing little sensitivity to the octant of [$\theta_{23}$]{}. Right: the likelihood shown as a function of [$\delta_{CP}$]{} for the four combinations of mass hierarchy and [$\theta_{23}$]{} octant. Those likelihood values above the horizontal lines are disfavoured at the 68% and 90% C.L. Figure from Ref. [@minosApp2012].](figure6Left.pdf "fig:") ![\[fig:nueAppContour\]Left: the [$\nu_\mu$]{}$\rightarrow$[$\nu_e$]{} appearance contour shown as a function of $2\sin^{2}2\theta_{13}\sin^{2}\theta_{23}$ and [$\delta_{CP}$]{}. The normal hierarchy is shown in the top panel, and the inverted hierarchy below, with the 68% and 90% contours shown for the lower octant of [$\theta_{23}$]{}. The best fit curve is also shown for the upper octant, showing little sensitivity to the octant of [$\theta_{23}$]{}. Right: the likelihood shown as a function of [$\delta_{CP}$]{} for the four combinations of mass hierarchy and [$\theta_{23}$]{} octant. Those likelihood values above the horizontal lines are disfavoured at the 68% and 90% C.L. Figure from Ref. [@minosApp2012].](figure6Right.pdf "fig:")
The data were also used to study the mass hierarchy, value of [$\delta_{CP}$]{} and the octant of [$\theta_{23}$]{} by using an external constraint from the reactor neutrino experiments $\sin^{2} 2\theta_{13} = 0.098 \pm 0.013$ [@dayaBay2012; @reno2012; @doubleChooz2012]. The likelihood is shown as a function of [$\delta_{CP}$]{} for the four combinations of mass hierarchy and [$\theta_{23}$]{} octant on the right of Fig. \[fig:nueAppContour\]. The sensitivity to these parameters is low, but this represents the first attempt from a long-baseline neutrino oscillation experiment to constrain these parameters and lays the foundation for future measurements.
Combined Three Flavour Analysis {#sec:3flavResults}
===============================
The analysis outlined here uses the full three flavour oscillation framework to perform a combined fit of the beam and atmospheric CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} disappearance samples along with the CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} appearance samples. The CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} event spectra from this analysis are shown in Fig. \[fig:eventSpec\]. All data are shown compared to both the null oscillations prediction (grey) and the best fit prediction with oscillations (red). The beam data (top row) also show the background from NC events (filled grey) and the atmospheric data is also compared to the background arising from cosmic-ray muons (filled blue).
The [$\nu_e$]{} and [$\overline{\nu}_e$]{} appearance data are shown in Fig. \[fig:nueSpec\], binned as a function of reconstructed energy and $\alpha_{LEM}$. The bins between 5 and 8$\,$GeV are shown for display purposes, but are combined into a single bin in the fitting procedure.
![\[fig:nueSpec\]The [$\nu_e$]{} (left) and [$\overline{\nu}_e$]{} (right) candidate energy spectra, each shown in three bins of $\alpha_{LEM}$, the PID parameter. The data in black are shown compared to the expected background (red line) and the expected three flavour appearance signal (purple for CC [$\nu_e$]{} and lilac for CC [$\overline{\nu}_e$]{}). The bins from 5-8$\,$GeV are combined for all samples in the fit. ](figure8.pdf)
A combined fit of all the MINOS data allows for the maximum extraction of information on the mass hierarchy, octant of [$\theta_{23}$]{} and the value of [$\delta_{CP}$]{}. The appearance sample is sensitive to [$\delta_{CP}$]{}, and it also provides some information on the mass hierarchy via matter effects as well as a small sensitivity to the octant of [$\theta_{23}$]{}. The atmospheric neutrino sample includes a resonance region in multi-GeV upward going events, again providing sensitivity to the mass-hierarchy. This resonance exists in the normal hierarchy for neutrinos and in the inverted hierarchy for antineutrinos.
The oscillation parameters that are free within the fit are [$\Delta m^2_{32}$]{}, $\sin^{2}\theta_{23}$, $\sin^{2}\theta_{13}$ and [$\delta_{CP}$]{}. The value of $\sin^{2}\theta_{13}$ is constrained using a Gaussian penalty term in the fit, using the central value and 1$\sigma$ uncertainty from the average of the results from the reactor neutrino experiments Daya Bay [@dayaBay2013], RENO [@reno2012] and Double Chooz [@doubleChooz2012]: $\sin^{2}\theta_{13} = 0.0242 \pm 0.0025$. The values of [$\Delta m^2_{21}$]{} and $\sin^{2}\theta_{12}$ are kept fixed in the fit at the following values: $\Delta m^{2}_{21}=7.54\times10^{-5}\,$eV$^2$ and $\sin^{2}\theta_{12}= 0.307$ [@fogli]. The effect of varying [$\Delta m^2_{21}$]{} and $\sin^{2}\theta_{12}$ was checked by shifting them by their quoted uncertainty and looking at the change in the fitted values of [$\Delta m^2_{32}$]{} and $\sin^{2}\theta_{23}$. The changes caused by varying these parameters were found to be negligible, hence no penalty terms are included for [$\Delta m^2_{21}$]{} and $\sin^{2}\theta_{12}$.
The oscillation probabilities used to perform the fit are calculated directly from the PMNS matrix without assumptions. The method takes advantage of matrix manipulation algorithms specially designed for high computational speed [@fastPMNS]. Matter effects are included using a four layer approximation of the PREM model [@PREM1981]. All of the systematic uncertainty parameters are included with the corresponding samples and treated as nuisance parameters with penalty terms in the fit. The systematic parameters are those that account for the main differences between the simulation and the data. The likelihood is calculated separately for the [$\nu_\mu$]{} disappearance and [$\nu_e$]{} appearance samples and the two contributing values are then summed together under the assumption that the systematic uncertainties in the two samples are uncorrelated.
The 2D confidence limits for [$\Delta m^2_{32}$]{} and $\sin^{2}$[$\theta_{23}$]{}, calculated by maximising the log-likelihood at each point in the 2D parameter space with respect to $\sin^{2}$[$\theta_{13}$]{}, [$\delta_{CP}$]{} and all of the systematic parameters, is shown in Fig. \[fig:minosCombi\]. The 68% contour is shown in red and the 90% contour is shown in blue. The overall best fit point is found to be in the inverted hierarchy, lower octant region. The results are $|$[$\Delta m^2_{32}$]{}$|=[2.28 - 2.46]\times10^{-3}\,$eV$^{2}$ (68%) and $\sin^{2}$[$\theta_{23}$]{}${}=0.35-0.65$ (90%) in the normal hierarchy $|$[$\Delta m^2_{32}$]{}$|=[2.32 - 2.53]\times10^{-3}\,$eV$^{2}$ (68%) and $\sin^{2}$[$\theta_{23}$]{}${}=0.34-0.67$ (90%) in the inverted hierarchy. These measurements of $|$[$\Delta m^2_{32}$]{}$|$ are the most precise at the time of writing, but Super-K [@superKPRD2010] and T2K [@t2kPRL2014] have higher accuracy measurements of [$\theta_{23}$]{}. The case known as maximal mixing in two flavour oscillations, namely that [$\theta_{23}$]{}${}=\pi/4$, is disfavoured at 1.54 units of $-2\Delta\log(\mathcal{L})$ [@minosCombined].
![\[fig:minosCombi\]Left: The 68% (red) and 90% (blue) confidence limit contours as a function of [$\Delta m^2_{32}$]{} and $\sin^{2}\theta_{23}$ from the combined analysis of 10.71$\times$10$^{20}\,$POT FHC data, 3.36$\times$10$^{20}\,$POT RHC data and 37.88$\,$kt-years of atmospheric neutrinos. Right: The profiled 1D likelihood for [$\Delta m^2_{32}$]{} (top) and $\sin^{2}\theta_{23}$ (bottom) assuming both normal hierarchy (red) and inverted hierarchy (blue). The best fit point lies in the inverted hierarchy, lower octant quadrant, and there is a slight tendency to disfavour the normal hierarchy, upper octant region.](figure9.pdf)
Figure \[fig:minosCP\] shows the 1D likelihood profile as a function of [$\delta_{CP}$]{}. This distribution shows an enhanced sensitivity compared to the [$\nu_e$]{}+[$\overline{\nu}_e$]{} appearance only result shown in the right plot in Fig. \[fig:nueAppContour\], but less than that of T2K [@t2kPRL2014NuE]. The best-fit oscillation parameters are shown in Table \[tab:fitResults\] for each combination of the mass hierarchy and octant of [$\theta_{23}$]{}. The data slightly disfavour the normal hierarchy, upper octant case across the whole range of [$\delta_{CP}$]{}, and above 90% for approximately half of the range of [$\delta_{CP}$]{}, with the best fit point in that case being disfavoured by a $-2\Delta\log(\mathcal{L})$ of 1.74. The other three choices of the mass hierarchy and octant have very similar values of $-2\Delta\log(\mathcal{L})$ and remain degenerate.
![\[fig:minosCP\]The likelihood surface profiled as a function of $\delta_{CP}$ under the four choices of mass hierarchy and $\sin^{2}\theta_{23}$ octant. The normal hierarchy upper octant case is disfavoured at 90% for half of the range of [$\delta_{CP}$]{}.](figure10.pdf)
[0.9]{}[ccccccc]{}\
Mass Hierarchy & [$\theta_{23}$]{} Octant & [$\Delta m^2_{32}$]{}${}/10^{-3}\,$eV$^{2}$ & $\sin^{2}\theta_{23}$ & $\sin^{2}\theta_{13}$ & [$\delta_{CP}$]{}${}/\pi$ & $-2\Delta\log(\mathcal{L})$\
\
[$\Delta m^2_{32}$]{}${}<0{}$ & [$\theta_{23}$]{}${}<\pi/4$ & -2.41 & 0.41 & 0.0243 & 0.62 & 0\
[$\Delta m^2_{32}$]{}${}<0{}$ & [$\theta_{23}$]{}${}>\pi/4$ & -2.41 & 0.61 & 0.0241 & 0.37 & 0.09\
[$\Delta m^2_{32}$]{}${}>0{}$ & [$\theta_{23}$]{}${}<\pi/4$ & 2.37 & 0.41 & 0.0242 & 0.44 & 0.23\
[$\Delta m^2_{32}$]{}${}>0{}$ & [$\theta_{23}$]{}${}>\pi/4$ & 2.35 & 0.61 & 0.0238 & 0.62 & 1.74\
\
Three Flavour Oscillations with MINOS+
======================================
The neutrino energy spectrum from the ME tune of the NuMI beam for MINOS+ is shown by the dashed line in Fig. \[fig:fluxConfig\], compared to MINOS LE configuration (solid) and the high energy configuration (dotted), and peaks between 3$\,$GeV to 10$\,$GeV. This means that MINOS+ probes the oscillation paradigm in the tail of the neutrino oscillation spectrum. It is in this region that more exotic phenomena such as sterile neutrinos or large extra dimensions are more easily seen from the distortion of the oscillation signal. These searches will not be discussed here, but the high statistics measurement of oscillations away from the oscillation maximum provide a stringent test of three flavour neutrino oscillations.
MINOS+ collected a total of 2.99$\times$10$^{20}\,$ POT in the first year of running from September 2013 until September 2014. The ND reconstruction software was rewritten in order to better cope with the higher rate of interactions produced by the upgraded NuMI beam, both in terms of minimising the impact of event pile-up and by increasing the speed of the algorithms to facilitate the prompt processing of data. A main focus of this effort was to prevent tracking failures, which provided a considerably improved ND efficiency.
In the first instance, a fit to just the MINOS+ data sample was performed such that the best fit oscillation parameters could be compared to those measured by MINOS. The reconstructed CC [$\nu_\mu$]{} energy spectrum is shown in the left plot of Fig. \[fig:minosPlusData\] and is also shown as a ratio to the no oscillations case on the right. The unoscillated prediction is shown in red along with two oscillated predictions: the blue shows the MINOS+ only best fit and the green shows the best fit using the parameters measured by the final combined MINOS analysis, as described in Section \[sec:3flavResults\]. The blue and green curves are very consistent, showing that the oscillation parameters measured by MINOS clearly provide a good description of the MINOS+ data.
![\[fig:minosPlusData\]The reconstructed CC [$\nu_\mu$]{} energy spectrum for beam neutrinos for MINOS+ (left) and the ratio of data to the unoscillated MC (right) for an exposure of 2.99$\times10^{20}\,$POT. The red curve shows the prediction in the case of no oscillations, the blue curve shows the best fit to the MINOS+ data alone, and the green curve shows the combined fit result from MINOS.](figure11Left.pdf "fig:") ![\[fig:minosPlusData\]The reconstructed CC [$\nu_\mu$]{} energy spectrum for beam neutrinos for MINOS+ (left) and the ratio of data to the unoscillated MC (right) for an exposure of 2.99$\times10^{20}\,$POT. The red curve shows the prediction in the case of no oscillations, the blue curve shows the best fit to the MINOS+ data alone, and the green curve shows the combined fit result from MINOS.](figure11Right.pdf "fig:")
A combined fit of all the data included in the MINOS combined analysis and the 2.99$\times10^{20}\,$POT of MINOS+ data was also performed. The combined reconstructed CC [$\nu_\mu$]{} energy spectrum is shown in Fig. \[fig:minosMinosPlusSpec\] compared to the best fit prediction from oscillations in blue. The MINOS and MINOS+ components of the best fit prediction are shown in the pink and blue filled histograms, respectively, showing that the statistics in the region around $6-8\,$GeV have doubled with only about a third of the expected exposure for MINOS+. The difference in the best fit point between the final MINOS result and this combined fit was $-2\Delta\log(\mathcal{L})=1.3$. The 2D contours in [$\Delta m^2_{32}$]{} and $\sin^{2}\theta_{23}$ are not shown as only a small improvement is seen compared to the combined MINOS analysis. This is expected since the MINOS+ energy distribution only provides a fairly small sensitivity to the oscillation parameters compared to MINOS. The result will be updated with the data from the full MINOS+ exposure.
![\[fig:minosMinosPlusSpec\]The reconstructed CC [$\nu_\mu$]{} energy spectrum for beam neutrinos in MINOS and MINOS+. The red curve shows the prediction in the case of no oscillations, and the blue curve shows the best fit prediction. The two filled histograms show the components of the best fit corresponding to MINOS (pink) and MINOS+ (blue).](figure12.pdf)
Conclusion
==========
The MINOS experiment collected data from the NuMI beam over a period spanning 2005 until 2012, and atmospheric neutrinos at the FD from 2003 until 2012, putting it at the forefront of neutrino oscillation physics for a decade. The combined analysis of the CC [$\nu_\mu$]{} and CC [$\overline{\nu}_\mu$]{} disappearance samples, coming from both beam and atmospheric neutrino sources, and the CC [$\nu_e$]{} and CC [$\overline{\nu}_e$]{} appearance samples using a full three flavour fit produced the world’s best measurement of the atmospheric mass splitting [$\Delta m^2_{32}$]{}. Whilst the sensitivity to the mass hierarchy, octant of [$\theta_{23}$]{} and the *CP*-violating phase [$\delta_{CP}$]{} is small, the data tend to disfavour the combination of normal mass hierarchy and upper octant of [$\theta_{23}$]{} at the 90% confidence level across half of the [$\delta_{CP}$]{} phase-space. The overall best fit point was measured to be in the inverted hierarchy, lower octant region.
The first year of data from the MINOS+ experiment, using the ME beam configuration, was analysed and shown to give very consistent results compared to the values of the neutrino oscillation parameters measured in the final MINOS analysis. Further data collected by MINOS+ will provide future stringent tests of the standard neutrino oscillation paradigm and allow for investigations of more exotic phenomena such as sterile neutrinos, non-standard interactions and large extra dimensions.
Acknowledgements
================
The work of the MINOS and MINOS+ collaborations is supported by the US DOE, the UK STFC, the US NSF, the State and University of Minnesota, the University of Athens in Greece, and Brazil’s FAPESP and CNPq. We are grateful to the Minnesota Department of Natural Resources, the crew of the Soudan Underground Laboratory, and the personnel of Fermilab, for their vital contributions.
Bibliography
============
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We detected a ring-like distribution of far-infrared emission in the direction of the center of the Virgo cluster. We studied this feature in the FIR, radio, and optical domains, and deduced that the dust within the feature reddens the galaxies in the direction of the Virgo cluster but does not affect stars within the Milky Way. This is likely to be a dusty feature in the foreground of the Virgo cluster, presumably in the galactic halo. The HI distribution follows the morphology of the FIR emission and shows peculiar kinematic behavior. We propose that a highly supersonic past collision between an HI cloud and the Galactic HI formed a shock that heated the interface gas to soft X-ray temperatures. HI remnants from the projectile and from the shocked Galactic HI rain down onto the disk as intermediate velocity gas.
Our finding emphasizes that extragalactic astronomy must consider the possibility of extinction by dust at high Galactic latitude and far from the Galactic plane, which may show structure on one-degree and smaller scales. This is particularly important for studies of the Virgo cluster, for example in the determination of the Hubble constant from Cepheids in cluster galaxies.
author:
- 'Noah Brosch & Elchanan Almoznino'
- 'Bogdan Wszolek & Konrad Rudnicki'
title: The Nature of a Dusty Ring in Virgo
---
\#1\#2\#3\#4\#5\#6\#7
to\#2
------------------------------------------------------------------------
Introduction
============
The nature of non-luminous matter that is not part of detected and catalogued galaxies remains unsolved by modern astrophysics. As mentioned in a recent thesis, low surface brightness (LSB) objects may prove to be the “icebergs” of the extragalactic world (de Blok 1997). Some searches for non-luminous matter have been successful, the detection of a giant HI ring around the small group of galaxies in Leo centered on M96 (Schneider 1983), extended HI emission in the M81 group (Lo & Sargent 1979), HI companions to dwarf galaxies (for $\sim$25% of the cases: Taylor 1996), and a large neutral hydrogen cloud in the southern outskirts of the Virgo cluster (HI 1225+01: Giovanelli & Haynes 1989).
Along with HI clouds, a few large LSB galaxies have been identified: Malin-1 (Bothun 1987), F568-6 (Bothun 1990), and 1226+0105 (Sprayberry 1993). Their typical star formation rates are $\sim$0.1 M$_{\odot}$/yr and the metallicities are $\sim$1/3 solar. The HI rotation curves, measured by de Block (1997) and by Pickering (1997), indicate that their gaseous component is dynamically significant at all radii and that the galaxies are fully dark-matter dominated; their detected baryonic component is less than 4% of the total mass. This last conclusion is valid at least as long as we do not accept any of the more exotic theories of gravitation. The LSB galaxies lack bulges, bars, and nuclear activity, as well as CO or IR emission (have no molecules or dust).
There have also been a few intriguing reports of presumably intergalactic dust clouds. A cloud with 0.5-1.2 mag of extinction was identified in Microscopium by Hoffmeister (1962). Three other similar objects were listed by Rudnicki (1986); they extinguish background objects by 0.57 to 1.2 mag. In all reports the main point of contention was the actual distance to the cloud, which could put it in extragalactic space but could also locate it in the halo of the Milky Way (MW). Sometimes, the argument for an extragalactic location was based on a comparison of the properties of objects whose distance could be estimated and which were located behind the cloud with those of similar objects clearly not within the cloud limits (RR Lyrae stars; Murawski 1983).
The extragalactic nature is only fairly confidently established for the Abadi-Edmunds cloud at $\sim$3 Mpc (Abadi & Edmunds 1978). HI 21 cm line emission was detected from this object, whereas in other cases it was not. However, in other cases far-infrared (FIR) emission was detected and could be identified (on morphological and positional criteria) with the obscuring clouds. FIR and HI emission were clearly detected in the case of the Okroy cloud (Wszolek 1988a, 1989). FIR emission was only marginally detected from the Rudnicki-Baranowska cloud (Wszolek 1988b). This indicates that the physical conditions in this kind of objects are far from being uniform. More such examples must be identified and their properties examined.
It is possible that the phenomenon of intergalactic hydrogen clouds could be related to the high-velocity cloud (HVC) complexes. These are HI structures whose radial velocities deviate by several 100 km s$^{-1}$ from the conventional galactic rotation. A recent review of HVCs is by Wakker & van Woerden (1997). Their Table 2 lists a few cloud complexes at distances $\geq$25 kpc; some of these may not belong at all to the MW. IRAS searches for FIR emission of HVC were negative (Wakker & Boulanger 1986), indicating that either the HVCs are dust-free or that their dust grains are much cooler than could be detected with IRAS. In this context we also mention the proposition by Blitz (1999) that the HVCs make up the missing mass by being essentially dark halos with low velocity dispersions.
We report here results from a study of a diffuse ring-like FIR feature at high galactic latitude, which we interpret as “local”, not extragalactic, despite first indications to the contrary. The region toward which this feature is located is the center of the Virgo cluster of galaxies. This part of the sky has been studied in exquisite detail, yet new studies always detect interesting features. For example, Katsiyannis (1998) produced a very deep image of the central regions of the cluster from a combination of 13 deep Kodak TechPan films obtained with the UK Schmidt telescope. The image shows large variations in the brightness of the intra-cluster medium, with the brightest regions north of the cluster center. M87 is fairly central in the region of enhanced brightness, close to the upper left corner of the “very high contrast image” in their Fig. 6. Previous deep imaging of the central VC region (Weil 1997) revealed a diffuse extension of (presumably stellar) material extending $\sim$100 kpc to the SE of M87. Intergalactic red giant stars were apparently discovered near M87 by Ferguson (1998). It is therefore relevant to search for, and to try and explain, any extended feature one may detect in the direction of the center of the cluster. In this context, we mention the study of Haikala (1995) who examined the UV emission detected in the direction of a dust globule close to the North Galactic Pole, slightly north of the Virgo cluster (VC).
Any material that could produce extinction needs to be accounted for. To the best of our knowledge, nobody attempted to study the obscuration and FIR emission by ISM or IGM in the direction of a rich, nearby cluster of galaxies. This is particularly important for the VC, which serves as one of the key stones in the distance ladder leading up to the determination of the Hubble constant (van den Bergh 1996). The HST Key Project on the Extragalactic Distance Scale, where the required accuracy of the determination of H$_0$ is 10%, could be affected significantly by unaccounted extinction. Until now, seven galaxies within 10$^{\circ}$ of the Virgo center have been observed for Cepheids in this context (Macri 1999).
The plan of the paper is as follows: we first describe the FIR observations, which revealed the feature, and present confirmatory evidence of its reality. We then attempt to derive additional properties of the feature, which has an approximate ring shape, using data in the optical and radio domains. We show that the dust in the feature does not seem to affect the stars in the Milky Way but that it apparently reddens galaxies in the VC and beyond. The full data set is discussed in the last section of the paper, in which we also derive some properties of the dust grains in the feature.
Observational data
==================
COBE/DIRBE
----------
Far infrared (FIR) observations from the COBE satellite, specifically with the DIRBE instrument, reveal non-uniform FIR emission from the center of the VC. The DIRBE instrument mapped the entire sky at ten wavelength bands from 1.25 to 240$\mu$m and operated from November 1989 to December 1993 (cryo-cooling was available only for ten months, restricting the availability of the FIR channels). An important feature of DIRBE was that the measurements were performed against an internal calibrator source, with proper accounting for instrumental offsets and interplanetary FIR emission. For the present analysis we used the Annual Average Sky Maps (AASM: Hauser 1997), which provide a single, ten-month averaged intensity value per pixel in each of the DIRBE bands. Note that the zodiacal light contribution was not subtracted from the DIRBE counts. This is because we do not estimate the zodiacal contribution to the FIR bands to be significant or to show features on the angular scales relevant here.
We conducted a number of studies of galaxies in the Virgo cluster in which we studied various photometric indices for entire objects as well as for localized regions in each galaxy (Almoznino & Brosch 1998, Heller 1998). The possibility that these programs could be affected by foreground dust imposed our selection of the Virgo Cluster as the initial target for combined FIR and other spectral band interpretations.
We detected a ring-like structure of FIR emission in COBE/DIRBE maps of the VC, which is centered approximately on M87. The ring is approximately centered on (1950) 12$^h$31$^m$; +13$^{\circ}$ (l=285$^{\circ}$.8, b=75$^{\circ}$; J2000) and its diameter is $\sim4^{\circ}$. The width of the FIR emission in the rim of the ring is $\sim1^{\circ}$. The detection was made originally on the COBE/DIRBE maps, but the existence of the feature was established also on IRAS maps (see below). The M87 galaxy (l$\approx282^{\circ}$.5, b$\approx+74^{\circ}$.4) is normally taken as the center of the Virgo Cluster (VC) and one could imagine scenarios by which some sort of FIR-emitting matter could be distributed around it. For this reason, we decided to follow the FIR detection of the feature, which we call here “the Virgo Ring” (VR), and investigate it further.
The detection was made on the AASM, which have noise levels of 3 10$^{-3}$ MJy sr$^{-1}$ at 100$\mu$m, 0.6 MJy sr$^{-1}$ at 140$\mu$m, and 0.3 MJy sr$^{-1}$ at 240$\mu$m (Kashlinsky 1999). The ring is visible even by superficial inspection of these COBE/DIRBE gray scale maps. No traces of the ring can be seen on 60$\mu$m or shorter wavelength maps. To obtain detailed insight into the structure of the VR we produced isophotal maps at $\lambda$=100 and 240$\mu$m using the original 0$^{\circ}$.3 square pixels, which are shown as isophote plots in Figure 1. The 100$\mu$m map shows a region of depressed FIR flux where F$_{100}\approx$8.2 MJy sr$^{-1}$. This is surrounded by regions of enhanced FIR emission, which reach F$_{100}\approx$10 MJy sr$^{-1}$. The 240$\mu$m map indicates that the region of reduced FIR emission has F$_{240}\approx$3.7 MJy sr$^{-1}$ while the surrounding regions have F$_{240}\approx$5 MJy sr$^{-1}$. It is clear that (a) the DIRBE data indicate a region of low FIR emission surrounded by enhanced emission, and (b) the feature is real, because it appears on more than one DIRBE map. The lowest values of the FIR flux originate presumably from the zodiacal light that was not subtracted from the AASMs and from the cosmic FIR background. As both these components are much smoother than the feature we describe here, there is no need to model them in detail.
IRAS
----
The peculiar FIR features detected by COBE/DIRBE are confirmed by IRAS measurements. The IRAS mission mapped the sky in four wavelength bands from January 1983 to November 1983. The primary goal of the IRAS survey was the detection of point sources, but a catalog of extended sources has also been produced, as well as sky brightness images in each of the four bands with 2’ pixels and 4’-6’ resolution (Beichman 1988).
IRAS 60 and 100$\mu$m Extended Emission Data in the 16$^{\circ}.5\times16^{\circ}.5$ square fields no. 83 and 84 were used to confirm the existence of the ring and to exclude the possibility of instrumental artefacts produced by the COBE/DIRBE instrument. We created maps at these two spectral bands with a 4’$\times$4’ beam. The VR is clearly visible on the 100$\mu$m map shown in Fig. 1. A similar 100$\mu$m map based on IRAS observations, and where this feature is also visible, was reproduced already by Leggett (1987) as their Plate 2. The enhanced IRAS resolution relative to COBE/DIRBE allows a good morphological evaluation of the FIR feature. In addition to the north-westerly extension of the FIR emission, along the IRAS scan direction, one sees an arc-like distribution of emission, which could be interpreted as forming an elliptical ring. Note that the feature is visible only on the 100$\mu$m map (shown in Figure 1) and is not seen on the 60$\mu$m map, or on those at even shorter wavelengths (not shown here).
Although the low resolution COBE/DIRBE maps seem to indicate that the FIR emission is arranged in a ring, with low FIR at the center and high emission on its perimeter, the higher resolution IRAS maps show that this is not the case. The FIR emission is distributed in an open configuration, with a region of low emission centered on $\sim12^h30^m$, +13$^{\circ}$.2. The FIR emission could best be described as a fork, or a two-arc shape limited to $\alpha$=185$^{\circ}-189^{\circ}$. The eastern side of the feature shows a small region of enhanced FIR emission centered on $\alpha$=185$^{\circ}$ and $\delta$=13$^{\circ}$.5 that stands out over its surroundings and to which we refer as the “main blob” (MB).
Optical information: stars
--------------------------
The dust revealed by the FIR observations may (a) extinguish and (b) redden stars behind it. The first effect is a consequence of the “total extinction” property, whereas the second is the result of “wavelength-selective extinction”. The relative importance of the two effects is linked through the parameter $R=\frac{A_V}{E(B-V)}$, which is determined to first order by the size of the dust grains.
We tested two assumptions, one of extinction within the Milky Way (MW) that would affect some of the stars but not others, and a second that the VR is extragalactic and is located between the MW and the VC. In the second case it would affect the VC galaxies, but none of the MW stars.
For testing the possibility that the dust is “local” one requires a large number of stars with magnitudes and colors. These were extracted from the USNO-A2.0 catalog, which includes blue and red magnitudes for each star. The USNO-A2.0 catalog contains $>$5 10$^8$ objects ($\sim$12,750 per square degree) and is based on scans of the Palomar Sky Survey (PSS) plates produced with the Precision Measuring Machine (PMM). The catalog is an improvement over the version 1.0 both in astrometric accuracy and in photometric precision. The photometric accuracy is probably not better than $\sim$0.15 mag, but the depth of the catalog is considerable, as it reaches 20-22 mag (color-dependent). It can, therefore, serve as a source of stellar objects with which one can test the assumption of foreground extinction.
We extracted objects in a number of $1^{\circ}\times1^{\circ}$ regions from the USNO-A2.0 catalog. The extraction locations are listed in Table 1 and correspond to some FIR-bright regions (where we expect a higher density of extinguishing dust) or to some FIR-faint regions (which should be $\sim$transparent). We produced Wolf diagrams for each location, and show these in Figure 2. The Wolf diagrams plot the cumulative distribution of stellar magnitudes against magnitude, and the signature of total extinction in such a plot is a step-like deviation, to fainter magnitudes of the cumulative star counts, from the pattern set by the brighter (and closer, on average) stars. The diagrams do not show such a step-like trend for regions in the direction of stronger FIR emission when compared with the behavior of the cumulative distribution in regions with lower FIR emission.
It is also possible to compare the measured behavior of the cumulative star counts with that “predicted” in absence of localized extinction effects by using a model for the stellar distribution in the Galaxy for the same Milky Way locations as sampled here. A very successful and intuitively simple stellar distribution model was produced by Bahcall & Soneira (1984) and is available on-line[^1]. We calculated predicted star counts for the locations of the extracted data from the USNO-A2.0 catalog using the version of the model retrieved in December 1998. The locations are listed in Table 1. We compared the predicted cumulative star counts with the actual star counts. The comparisons are shown in Figure 3 and show no significant deviations from the predicted behavior.
The exercises shown in Figs. 2 and 3 indicate that the stellar distributions are not influenced by the material producing the FIR emission. The conclusion is, therefore, that this material is either extremely nearby, so that all the stars are affected in the same manner, or that it is very distant, beyond the more distant stars listed in the USNO-A2.0 catalog.
Optical information: galaxies
-----------------------------
If the dust observed in the FIR does not affect stars in our galaxy, it may be located far from the MW and could affect only objects seen behind it. Testing the assumption of a dust cloud distant from the MW requires a sample of background objects with relatively high surface density, as well as brightness and color information. In the Virgo region, the “standard” extragalactic catalog has been for a number of years the Binggeli (1985) Virgo Cluster Catalog (VCC). The VCC covers $\sim$140 square degrees and contains 2096 galaxies. The surface density of galaxies is, therefore, $\sim$15 galaxies/square degree, on average. While this may appear sufficient, the photometry is not adequate because the galaxy magnitudes in the VCC are eye estimates and may have significant deviations. In addition, no colors are available for most VCC galaxies. We decided therefore to rely on a more recent galaxy compilation, which reaches deeper in brightness and is thus denser than the VCC, has better photometry, and contains color information for the objects.
Currie & Young (1998, hereafter VPC) produced an extensive three-color photometric catalog of galaxies in the central regions of the VC. The catalog is based on COSMOS scans of one U plate, two B$_J$ plates, and one R$_C$ plate, all obtained with the UK Schmidt telescope. The plates were photometrically calibrated and objects were extracted automatically, with stars and galaxies separated by an automatic algorithm. The VPC provides an impartial survey of galaxies in the region of interest for the present study, which reaches to B$_J\approx$19 mag, thus it is somewhat shallower but comparable in depth with the stars from the USNO-A2.0 catalog. The area covered by the VPC covers 23 square degrees and is centered on (1950) 12:26 +13:08. The average galaxy surface density is therefore 49 galaxies/square degree, considerably more than that of the VCC.
We attempted to detect total extinction effects on the VPC galaxies by limiting the analysis to regions with high FIR emission and comparing these with similar analyses in the direction of regions with lower FIR emission. Four parallerogram-shaped fields were selected, marked A, B, C, and D on Figure 4. Fields C and D are used to determine the nature of the galaxy population in the general region of the VR. Field A could also be used for this purpose, but we caution that a background cluster of galaxies (Abell 1552 at z$\approx$0.084) is located in this field and thus region A may not be representative. This galaxy cluster is presumably part of a background sheet-like complex, which includes also Abell 1526 at a very similar redshift. Field A was selected to offer insight on how the presence of background galaxies disturbs the results. The enhancements of the galaxy background may distort the Wolf diagrams of galaxies (Figure 5), and indeed some FIR enhancements could be in the direction of these galaxy clusters. However, searches for dust in clusters of galaxies have, so far, been negative (Maoz 1995). Thus, we may tentatively discount the FIR enhancements in the direction of background clusters of galaxies as chance superpositions. Field B is considered not to be affected by absorption/extinction and is in the direction of the low FIR emission of the VR.
Attempts to detect the presence of dust as a “total extinction” effect, which modifies the cumulative galaxy counts between the different regions, were not successful. The differences were not significant and indicate that if dust is present, it may cause at most a small amount of total extinction: A$_B\leq$0.5 mag. We therefore checked for the presence of color-dependent extinction by studying the distribution of the (U–R$_C$) color index in one square degree areas over the central part of the Virgo cluster. The data used for this test, and an extensive description of the method and results, are given in the Appendix to this paper. Here we emphasize that the results show that the galaxies in the direction of the Virgo Ring (VR) part with the lowest FIR emission appear slightly bluer than those in the direction of the two regions with higher FIR emission. The difference is significant to $\geq$95%. Interpreted as dust extinction, this difference in average (U–R$_C$) color index indicates a possible wavelength-dependent extinction of $\Delta$(U–R$_C$)$\simeq$0.3 mag between areas with high FIR emission and areas with less dust, a total extinction A$_V\simeq$0.33 for a typical Milky Way extinction law, although this was not checked here.
Radio information
-----------------
Here we show that X-ray observations of the region indicate a two-component makeup for the hot gas, and that the morphology and kinematics of the HI are peculiar. Böhringer (1995) mapped the X-ray emission from the immediate vicinity of M87; this is the region of interest of the present study. Their findings show the presence of thermal X-ray emission from cooler gas than the intracluster medium. A ROSAT map of the general region, larger than the one analyzed in the 1995 paper, was presented by Böhringer (1994) and shows a ridge of X-ray emission which approximately coincides with the FIR emission ridge to the west of M87. They mention, in particular, the sharp drop in X-ray intensity on the western side of M87. Böhringer (1994) subtracted a model distribution of X-rays from M87 from the ROSAT map and derived a residual map (their Fig. 2) which shows the background cluster A1552 at 12:30+11:30 and a long filament, which is elongated $\sim$north-south at $\alpha\approx$12$^h$30$^m$ and from $\delta\approx$+15 to +6. This filament curves around M87 on its westerly side and seems to follow the contours of the 100$\mu$m emission.
It is tempting to speculate on a possible link between the X-ray and FIR emission presented above, but we caution that this may not be real. One possible factor affecting the morphology of X-ray emiting gas is the amount of foreground HI, which modifies mainly the low energy end of the X-ray spectrum. Shadows in the X-ray background caused by foreground HI clouds have been detected mainly in soft X-rays by Egger & Aschenbach (1995). However, the feature detected in the ROSAT maps by Böhringer (1994) is seen in the hard energy band (0.4–2.4 keV), and is thus difficult to attribute it to gas absorption.
EUVE observations of the Virgo cluster (VC) center (Lieu 1996) show the presence of gas at $\sim$0.5 10$^6$ K near M87. This matter forms an additional component of the intra-cluster material (ICM) in Virgo, as follows from their analysis, and cannot be the same hot gas which is responsible for the X-ray emission detected by ROSAT. In order to confirm the existence of this second ICM component of the VC, Lieu (1996) performed HI 21 cm observations with the 43-m Green Bank telescope (angular resolution 21’). The region surveyed by them was centered on M87, had an extent of 2$^{\circ}\times1^{\circ}.6$, and the grid of HI measurements was spaced every 8’$\simeq$1/3 of a resolution element.
A comparison of the HI map of Lieu (1996) with the FIR distributions (see Figure 1) demonstrates that the FIR emission follows the total HI column density. Although Lieu do not mention the velocity range over which the Green Bank observations were performed, we assumed these to be at $\sim$0 km s$^{-1}$ because they are supposedly of “Galactic HI” origin. While some VC galaxies do have negative heliocentric velocities (c.f. Binggeli 1985), they mostly concentrate at 1,000–2,000 km s$^{-1}$. For this reason, we think it likely that the HI detected by Lieu (1996) does indeed belong to the MW and, by inference, so does the material producing the FIR emission. We note at this point that the center of the low N(HI) region, at 12:28+12:45 (1950) and only $\sim$half a degree away from M87, has N(HI)$\simeq$1.8 10$^{20}$ atoms cm$^{-2}$. This is coincident with the low FIR emission region. The ridges with the higher N(HI) values correspond to enhanced FIR emission regions.
We produced N(HI) plots for the region using data from the Leiden-Dwingeloo HI survey (Hartmann & Burton 1997, LDS) in order to confirm the HI distribution measured by Lieu (1996). The LDS was conducted with the 25-m radio telescope at Dwingeloo and the data we used cover the velocity range –459$<v_{lsr}<$+415 km s$^{-1}$ with a resolution of 1.03 km s$^{-1}$. The 25-m radio telescope has a 36’ half-power beam and the survey was performed with 0$^{\circ}$.5 spacings. We used the file TOTAL\_HI.FIT from the CD-ROM supplied with the printed atlas to extract the proper sky region. The data were transformed from Galactic to equatorial coordinates, accounting for the change of scale from one side of the image to the other. This was done by dividing each pixel value by its [*cos(b)*]{}, to yield consistent units over the field. The HI total column density from the LDS is shown in Figure 6 together with the IRAS 100$\mu$m map and confirms the general impression from the Lieu (1996) map. The HI distribution has a region with lower N(HI) at the center of the Virgo Ring (VR) and ridges of higher HI emission on both sides of the VR.
We also produced position-velocity (PV) plots using the channel data from the LDS, limiting these to l=276.0, the galactic longitude of the HI peak which coincides with the main FIR emission blob in area A of Fig. 6 (the more intense of the FIR peaks), to the center of the VR (l=283.0), and to the second highest FIR peak at l=290.5. The PV plots are shown in Figure 7 and indicate that at the position of the VR there is a significant disturbance of the HI, with a strong extension to negative velocities appearing in the PV plots of the high-FIR region. The sheet-like HI distribution, which links HI at low latitudes with gas near the Galactic Pole and has a slightly negative LSR velocity, appears disturbed at b$\approx$75$^{\circ}$.
The velocity plot through the peak emission at l=276.0 at this latitude (Figure 8) shows three peaks separated by $\sim$20 km s$^{-1}$. The strongest has the most negative velocity, approximately –30 km s$^{-1}$ and a FWHP of $\sim$11 km s$^{-1}$. The weakest peak at this location is near +4 km s$^{-1}$ (LSR). The PV in the low FIR region at the center of the VR (l=283$^{\circ}$, b$\approx$75$^{\circ}$) shows a single strong peak at $\sim$–7 km s$^{-1}$ (LSR), with a FWHP of 12.5 km s$^{-1}$ and a low shoulder extending to more negative velocities, down to the velocity of the strong peak at the location of the main blob (–30 km s$^{-1}$). The third PV, at (l=290$^{\circ}$.5, b$\approx$75$^{\circ}$) is narrow with a FWHP of $\sim$ 5km s$^{-1}$ and is centered at –7 km s$^{-1}$ (LSR).
Discussion
==========
We identified a ring-like feature of FIR emission at high galactic latitude, which is distant from the main body of the Galaxy and extinguishes light from galaxies in the central part of the Virgo cluster (VC). There is no way to establish a distance to the extinguishing cloud with the data we presented above, except to note that it is probably $>$1 kpc.
A nearby dust feature, observed by Haikala (1995) in the far-UV, has been located at $\sim$120 pc using the distribution of E$_{b-y}$ color excesses. This dust cloudlet produces a visual extinction A$_V\leq$0.4 mag and is located at (l=251.1, b=+73.3); this location is very similar to what we found for the Virgo Ring (VR) and may indicate that either our distance evaluation is wrong, or that the location technique of the Haikala feature did not use a sufficient number of more distant stars.
Indications that the dust cloud cannot be a nearby feature originate mainly from the lack of influence on the distribution of stars. Supporting evidence to the same comes from the reddening study of Knude (1996). He used uvbyH$\beta$ measurements of A3-G0 stars with B$\leq$11.5 mag and $\vert$b$\vert>70^{\circ}$ to determine the distribution of extinction. His results for E$_{b-y}$, broken by galactic latitude and by longitude quadrants, are of particular interest. The area of interest for our study is located between the 3rd and 4th quadrants at b$\approx75^{\circ}$; the reddening to this region is small, E$_{b-y}\leq0.017$, which translates into A$_B\leq$0.095. The stars studied by Knude (1996) are closer than 1.5 kpc (for main sequence A stars brighter than 11.5 mag), thus the color-dependent extinction of the VC galaxies we detected, which is equivalent to A$_B\approx$0.4 mag, should be produced by material more distant than 1.5 kpc.
If the cloud would be in the VC itself, its physical size would be $\sim$1.5 Mpc, very large indeed ! The issue of possible diffuse dust in clusters of galaxies has been studied by Ferguson (1993). He concluded, from the lack of a difference between cluster and field galaxies in the correlation of the Mg$_2$ index and (B–V), that dust is not present in the Virgo cluster (upper limit E(B–V)$<$0.06 mag.). A similar conclusion for a large number of Abell clusters, based on the (V–I) color indices of radio quasars seen in their background, was reached by Maoz (1995).
Not accounting for foreground dust may affect adversely some key observations. Our finding confirms the supposition of Zonn (1957) and Zonn & Stodolkiewicz (1958), that because of the patchy structure of the interstellar dust [*it is not enough to correct for extinction assuming that the dust is localized in a narrow slab near the Galactic equator, but the detailed distribution of dust must be investigated to account properly for the extinction.*]{} In particular, many observations of the Virgo cluster and of objects within (the HST Cepheid Key Project: Graham 1998, Macri 1999) may carry significant errors because of improper extinction corrections.
In this section we estimate the dust temperature and dust-to-gas ratio. To evaluate the temperature of the dust in VR we subtracted from the map intensities the minimal value for the central part of the VR, near M87, in the 100 and 240$\mu$m COBE/DIRBE bands (8 and 3.5 MJy sr$^{-1}$ respectively). To determine the color temperature for the dust in the VR cloud we assumed that the dust particles are in thermal equilibrium and that the foreground galactic IR radiation and IR emission from all point sources in the region have been subtracted accurately. These assumptions may not be necessarily fulfilled in our case. The subtraction of the intensity of the inner part of the ring would be accurate only if the distribution of the foreground galactic radiation is fairly smooth; this is not the case even at high galactic latitudes. Some galactic features may add non-negligible FIR contributions to the foreground. The radiation from very cold dust grains (T$\approx$3K) could not be detected by the means used here. We could not rule out the possibility that transient heating of dust grains takes place, and that only occasional excitation by energetic photons or particles causes them to emit brief pulses of the radiation measured by COBE/DIRBE and IRAS.
With all these caveats in mind, we calculated the temperature of two regions with maximal FIR intensities within regions delimited by: $\alpha$: 185$^{\circ}$–186$^{\circ}$, $\delta$: 13$^{\circ}$–14$^{\circ}$ (approximately the Main Blob=MB region) and $\alpha$: 189$^{\circ}$–190$^{\circ}$, $\delta$: 12$^{\circ}$–13$^{\circ}$ (slightly off the secondary FIR peak). The temperature was calculated with the relation: $$log_{10} T = 1.30274 + 0.26266(log_{10} R)+0.04935(log_{10} R)^2$$ from Schlegel (1998), where R=$\frac{I_{100}}{I_{240}}$ is the ratio of the FIR intensities in the COBE/DIRBE bands. The results are 22 and 20K for the first and the second region, respectively. These values can be accepted as upper limits for the dust temperature in VR due to both effects mentioned above. The temperatures do not differ significantly from those usually adopted for interstellar dust clouds and are at the high end of the range for the “warm” cirrus component (Lagache 1998). The difference between the two regions, with the MB being “hotter” than the secondary FIR blob, is marginally significant, because the formal error in the derivation of temperatures is 1-2K.
A study of the optical depth effects of high latitude clouds (Chlewicki & Laureijs 1988) found that typically $\frac{I_{100}}{N(HI)}$=0.7 MJy ster$^{-1}$/10$^{20}$ atoms cm$^2$. Deul & Burton (1990) studied the HI content and gas kinematics of seven cirrus clouds detected in the IRAS maps. They established that the FIR to HI ratio varies between 0.9 and 3.0 MJy ster$^{-1}$/10$^{20}$ atoms cm$^{-2}$. The color temperature varies much less, between 0.20 and 0.27 for $\frac{I_{60}}{I_{100}}$. Unfortunately, we did not detect 60$\mu$m emission from the VR, thus a direct comparison with the color temperature derived by Deul & Burton (1990) is not possible. However, the ratio $\frac{I_{100}}{N(HI)}$ we find for the FIR ridges ($\sim$1.5 MJy ster$^{-1}$/10$^{20}$ atoms cm$^{-2}$) is in the range of that measured for cirrus clouds. We determined N(HI) from the brightness temperature using data from Fig. 9 and applying the conversion $$N(HI)=1.8 \, 10^{18} \int{T_B(v) \, dv}$$ assuming that the HI is optically thin.
The VR is located close to the northernmost point of the North Polar Spur (NPS) and could, in principle, be part of it. The NPS is presumably the remnant of a $\sim$15 Myr old explosive event in the direction of the Galactic center, which released 10$^{56}$ ergs and could be the outcome of $\sim10^5$ supernovae produced in a small region within one million years. The giant loop is detected in a number of wavebands, from the X-ray to the radio continuum (Sofue 1994). The possibility that the VR is part of a small structure, extending to lower galactic longitudes from near the northernmost apex of the large NPS feature in Sofue’s Fig. 2 (Plate L11), cannot be excluded.
It is also clear that the Virgo Ring (VR) is not a known high velocity cloud, because the compilation of Deul & van Woerden (1990) lists as nearest object HVC 2 at 12:09+15:32 and v$_{lsr}$=100.7 km s$^{-1}$, covering 3.2 square degrees of the sky. Wakker & van Woerden (1997) list a possible small HVC near the VR location. This is an HI feature associated with optical absorption lines observed in SN 1994D, located at l=290$^{\circ}$.07; b=+70$^{\circ}$.13 and $\sim$240 km s$^{-1}$ LSR. This SN occured in NGC 4526, and the absorbing gas (with five absorption systems) was found by Ho & Filippenko (1996) at the same velocity as the HI measured by Kumar & Thonnard (1983). Ho & Filippenko mention other similar absorption features in the spectrum of SN 1991T (at l=292.59; b=65.18, in NGC 4527), which is $\sim5^{\circ}$ away from SN1994D. It seems that there is material in the general direction of the VR with LSR velocities a few 100 km s$^{-1}$, and that it contains at least Ca II and Na I. In the vicinity of the VR, velocities in excess of 100 km s$^{-1}$ are required in order to qualify an HI feature as a HVC.
Our analysis of the HI distribution and kinematics, using the LDS survey data, indicates that there is a significant disturbance of the hydrogen at the location of the VR. In particular, the three-peak structure of the velocity-column density is peculiar. There are a number of possible explanations for this HI peculiarity, starting from an expanding supernova (SN) shell, a collision between Galactic HI and a high-velocity cloud, an ejection of a number of shells from a red (super)giant, etc. Each of these possibilities imposes some constrains on the problem.
Of all these possibilities, we consider the most likely to be that of a past collision between a small HI cloud and Galactic hydrogen. The HI remnant of this event could be the feature seen at –30 km s$^{-1}$ in the high-FIR area and the Galactic HI would be the prevalent emission at –7 to –9 km s$^{-1}$, which shows up in all plots. At a velocity difference of $\sim$20 km s$^{-1}$ the collision would be highly supersonic. We note that the immediate vicinity of the VR has been studied in exactly this context by Stark (1994). The region most likely to be part of the same complex is the BX field (see Table 1 of Stark 1994), where the X feature peaks near –8 km s$^{-1}$ (LSR) but there are parts of the B feature where velocities up to –40 km s$^{-1}$ are observed.
Stark (1994) interpreted the “intermediate negative velocity” (INV) gas, with v$\leq$–20 km s$^{-1}$, as material at local velocity displaced by a shock following a collision with a high velocity cloud. The difference in LSR velocity between the BX features and candidate HVCs in this vicinity (at a few 100 km s$^{-1}$ positive) forced Stark to propose scenarios for the dissipation of the impinging HVC. They asserted that this cloud has evaporated and is now present as high-temperature gas. Collisions of HVCs with the Galactic disk have been studied by Kerp (1996). They show that a collision with a differential velocity of 25 km s$^{-1}$ may increase the temperature of the post-shock material from 10$^4$ to 10$^5$K; such temperatures are consistent with the EUVE measurements (Lieu 1996).
It is possible to estimate the parameters of the Main Blob (MB) using the HI information from the LDSS. With the T$_B$ value from –30 to –20 km s$^{-1}$ we find $$<N(HI)>\approx1.5 \, 10^{20} cm^{-2}$$ This implies a gas-to-dust ratio for the MB of $\frac{N(HI)}{E(B-V)}\approx$5 10$^{20}$ cm$^{-2}$ mag$^{-1}$, different from the canonical value for the Galaxy ($\frac{N(HI)}{E(B-V)}$=5.4 10$^{21}$ cm$^{-2}$ mag$^{-1}$; Bohlin 1975). The MB appears, therefore, to be dust-rich (or HI-poor). Its projected angular size is $\sim0^{\circ}.5$; this translates to a size of 10($\frac{d}{1 \, kpc}$) pc and a total hydrogen mass of $$M_{tot}(HI)\approx100 \frac{d}{1 \, kpc} M_{\odot}$$ with an average (volume) density of $\sim$4 $(\frac{d}{1 \, kpc})^{-1}$ cm$^{-3}$ assuming a spherical configuration. The cloudlet carries significant kinetic energy, considering its velocity relative to the Galactic HI (assumed to be $\sim-9$ km s$^{-1}$); E$_K\approx$3 10$^{47} \, (\frac{d}{1 \, kpc})^2$ ergs. Also, the HI profile observed in its direction (see top panel of Fig. 9) is significantly wider than that of the second blob (bottom panel of Fig. 9). This could be an indication of higher turbulence within the MB relative to other HI entities in the same vicinty but at less negative LSR velocities.
It is significant that the MB, an infalling HI gas cloudlet, contains dust grains as evidenced by the extinction it produces. This indicates that any shocks that the material in MB might have encountered did not heat up its material to temperatures high enough to completely destroy the grains. However, its highly supersonic interaction with the ambient HI distribution caused a wider 21 cm profile. The MB produces significant extinction and is fairly distant from the MW plane while having an intermediate LSR velocity. This is different from the findings of Knude & H$\o$g (1999), who concluded that intermediate-velocity HI clouds show no extinction.
The detection of HI condensations on sub-degree scales and high above the Galactic plane is relevant for the studies of galaxies in the Virgo cluster and beyond. A spectroscopic survey of objects in the background clusters at z$\approx$0.08 could reveal absorption lines produced by material within the MB cloudlet. This, along with a measure of the ionization presumably taking place at the interface between the MB and the ambient HI, could reveal its composition and location. The essential finding is, however, the small scale on which significant extinction variations are encountered at high $\vert$b$\vert$. This shows that the derivation of “average extinction dependences”, with galactic latitude (Knude 1996), is not very relevant to the determination of cosmologically-important parameters from studies of individual VC galaxies.
Our study shows that dust can affect the light of background galaxies in the VC and, through this, the determination of H$_0$ by using Cepheid photometry. This HST Key Project relies on a canned approach to deredden individual stars and is based on the “standard” MW extinction law with R=$\frac{A_V}{E(B-V)}$=3.3 (c.f. Madore & Freedman 1991). It remains to be proven that this relation applies for high galactic latitude dust clouds, such as those studied here.
Conclusions
===========
We detected a dusty HI cloud seen in projection against the central regions of the Virgo cluster. We showed that the cloud, which appears ring-like in low-resolution FIR and total HI maps, has a complex FIR morphology when examined with higher resolution. The HI kinematics indicate a substantial disturbance at this location.
The Virgo Ring (VR) is located beyond the main body of the Galaxy, because the distributions and colors of stars in its direction are not affected by it. The cloud could not be more distant than the VC, because it influences the colors of galaxies within the cluster. The HI evidence argues strongly for a relatively nearby location. The connection of the ring with the North Galactic Spur may be accidental, but the peculiar morphology and kinematics of the HI associated with the FIR emission make a connection, perhaps through a collision between an HI cloud and Galactic HI, more likely. Such a collision could perhaps explain also the extreme-UV (or soft X-ray) emission observed from the neighborhood of M87.
The importance of our finding lies in the possibility that many key observations done in the direction of the Virgo cluster may have been adversely affected by extinguishing clouds, which are not homogeneous and where the total V-band extinction may reach up to a few tenths of a magnitude.
Acknowledgements {#acknowledgements .unnumbered}
================
NB acknowledges support from the Israel Science Foundation. EA is supported by a grant from the Israel Ministry of Science to develop TAUVEX, a UV imaging telescope. KR and BW were supported by the Polish KBN grant number C76/98. We are grateful to Drs. C.K. Young and M.J. Currie for supplying electronic versions of the VPC data tables in advance of publication. NB is grateful to Mike Hauser for discussions on the reality of COBE/DIRBE features, to John Bahcall for allowing public use of his Galaxy model, and to Sara Beck and Sasha Kashlinsky for critical readings of one of the first drafts.
Appendix {#appendix .unnumbered}
========
Classical Wolf diagrams for galaxies (Zonn & Rudnicki 1965) were made for the regions labelled A through D in Fig. 4 and are shown in Figure 5. Only curve C could be interpreted as shaped by extinction, and this only by a large amount of leeway. Curve A, as already cautioned, has a completely different shape. Both A and D curves are higher than the comparison curve B, and are presumably shaped by the inhomogeneous distribution of galaxies in the VC area. If, however, we would adopt curve C as representative for the background galaxies (no VC subclusters or background galaxy clusters are evident in this area), then the obscuring cloud would have to be relatively nearby, as no parts of the C and B curves overlap.
To prevent distortion of the results by possible subclustering, we decided to study the color distribution of the background galaxies. The advantage of using color is that this parameter is distance-independent, provided the objects of study are not “too distant”, in the sense of requiring a k-correction. We assume that the objects in the VPC conform to this constraint and do not attempt to correct the color for redshift (which is not known, in most cases). We also caution about “edge effects”, which could arise because of a smaller population of the sub-areas located near the edges of the VPC coverage; such areas cover less than one square degree, have less than the average number of galaxies, and would be less representative.
We analyzed the color distribution of all galaxies in the entire VPC. We selected objects only by their position; specifically, we included all the objects listed in the catalog with their magnitudes and colors, without separating them by morphological type, apparent blue magnitude, angular size, being members or non-members of the cluster, etc. The area enclosed within 184$^{\circ}<\alpha<189^{\circ}$ and 10$^{\circ}<\delta<16^{\circ}$ was divided into square degree areas at integer $\alpha$ and $\delta$ values. In each sub-area we calculated the statistics of the color distribution for the galaxies. The total number of galaxies per cell does not appear to show a specific pattern, apart for the edge effects mentioned above.
We checked the distribution of photometric indices over the sub-areas of the VPC against the locations of the enhanced FIR emission. In each one square degree area we counted the galaxies, calculated their mean B$_J$ magnitude and the average (U–B$_J$), (B$_J$–R$_C$), and (U–R$_C$) color indices, and studied the statistical properties of the distributions of the color indices in each area. The color index (U–R$_C$) has been calculated as (U-B$_J$)-(B$_J$-R$_C$). The results are presented in Tables A1 and A2 for (U–R$_C$), where each cell contains these parameters for each one square degree region. The results for the other color indices were similar and are not shown.
Each sub-area in Table A1 is represented by one cell in the table, where the rows are the declination of the sub-areas and the columns are the right ascension, both in decimal degrees. The distribution of the (U–R$_C$) color is represented by the mean, the standard error of the mean (in round brackets), and the number of galaxies in the sub-area \[in square brackets\]. The standard error of the mean $SE_a$ is defined as: $$SE_a=\frac{SD_a}{\sqrt{N_a}}$$ Here $SD_a$ is the standard deviation of the values in sub-area [ *a*]{}, and $N_a$ is the number of galaxies in the same sub-area.
Table A2 shows the properties of the distribution of (U–R$_C$) in each of the one square degree areas. Specifically, we list the kurtosis with its standard error, and the skewness with its standard error, for the distribution of galaxies within the square degree area. These are listed in order to demonstrate how much are the distributions of (U–R$_C$) values in individual cells similar (or different).
The significance of possible differences among the cells was considered using the mean and its standard error. A Student’s [*t-test*]{} between the means, by which one estimates the variable $$t=\frac{\vert<X_1>-<X_2>\vert}{\sqrt{SE_1^2+SE_2^2}}$$ was performed. Here $<X_i>$ is the mean of variable [*i*]{}, and SE$_i$ is its standard error of the mean (Brandt 1970).
A cursory perusal of Table A1 shows that the only sub-areas where (U–R$_C$) is small, and which are not on the borders of the VPC (Currie & Young 1998), are the two areas at ($\alpha$=187.5, $\delta$=13.5), which we call “area GA”, and at (187.5, 14.5) which we call “area GB”. These regions are central in the Virgo Ring (VR) and this finding, if significant, could indicate less wavelength-dependent extinction in this direction (which shows reduced FIR emission) than in neighboring areas. The comparison is done with two regions located in the direction of enhanced FIR emission, at (186.5, 12.5) which we call “area GC”, and at (188.5, 12.5) which we call “area GD”. The latter area [**is**]{} located at the edge of the VPC but does not appear to be depleted in galaxies, thus the comparison with it is valid.
Using the t-test described above, the difference in means of the (U–R$_C$) color index between areas GA and GC yields t=1.97 (significant at the 95% level), and between areas GA and GD t=2.72 (signficant at the 99.5% level). A similar significance level was found for the difference between areas GB and GC, with t=2.76. The results of the t-test indicate that the galaxies in the VPC, which are located behind the low FIR emission area, are slightly bluer than others which are located behind areas of high FIR emission. This impression is supported by an examination of the higher moments of the distribution (Table A2). The skewness of the (U–R$_C$) values in areas GA and GB is very similar, but is very different from the skewness of the distribution in areas GC and GD.
References {#references .unnumbered}
==========
Abadi, H.J. & Edmunds, M.G. 1978, A&A, 70, 189
Almoznino, E. & Brosch, N. 1998, MNRAS, 298, 920
Bahcall, J.N. & Soneira, R. 1984, ApJS, 55, 67
Beichman, C.A., Neugebauer, G., Habing, H.J., Clegg, P.E. & Chester, T.J. 1988 [*IRAS Catalogs and Atlases: Vol. 1 (Explanatory Supplement)*]{}, NASA RP-1190
Binggeli, B., Sandage, A. & Tammann, G.A. 1985, AJ, 90, 1681
Blitz, L., Spergel, D.N., Teuben, P.J. & Hartmann, D. 1999, astro-ph/9901307
Bohlin, R.C. 1975, ApJ, 200, 402
Böhringer, H., Nulsen, P.E.J., Braun, R. & Fabian, A.C. 1995, MNRAS, 274, L67
Böhringer, H., Briel, U.G., Schwarz, A., Voges, V., Hartner, G. & Trumper, J. 1994, Nature, 368, 828
Bothun, G., Impey, C.D., Malin, D.F. & Mould, J.R. 1987, AJ, 94, 23
Bothun, G., Schombert, J.M., Impey, C.D. & Schneider, S.E. 1990, ApJ, 360, 427
Brandt, S. 1970 [*Statistical and Computational Methods in Data Analysis*]{}, Amsterdam: North-Holland Publishing Co., pp. 125 [*et seq.*]{}
Currie, C.K. & Young, M.J. 1998, A&AS, 127, 367
de Blok, W.J.G. 1997, PhD thesis, University of Groningen
de Blok, W.J.G., van der Hulst, J.M. & Bothun, G.D. 1995, MNRAS, 274, 235
Deul, F.R. & Burton, W.B. 1990, A&A, 230, 153
Egger, R.J. & Aschenbach, B. 1995, A&A, 294, L25
Ferguson, H.C. 1993, MNRAS, 263, 343
Ferguson, H.C., Tanvir, N.R. & von Hippel, T. 1998, Nature, 391, 461
Giovanelli, R. & Haynes, M. 1989, ApJ, 346, L5
Graham, J.A., Ferrarese, L., Freedman, W.L. 1998, BAAS, 192, 66.12
Haikala, L.K., Mattila, K., Bowyer, S., Sasseen, T.P., Lampton, M. & Knude, J. 1995, ApJ, 443, L33
Hartmann, D. & Burton, W.B. 1997, [*Atlas of Galactic Neutral Hydrogen*]{}, Cambridge (UK): Cambridge University Press
Hauser, M.G., Kelsall, T., Leisawitz, D. & Weiland, J. 1997, [*COBE Diffuse Infrared Background Experiment (DIRBE) Explanatory Supplement*]{}, COBE Reference Document Ref.No. 97A
Heller, A., Almoznino, E. & Brosch, N. 1998, MNRAS, in press
Ho, L.C. & Filippenko, A.V. 1996, ApJ, 463, 818
Hoffmeister, C. 1962, Z.f.Astrophys., 55, 40
Kashlinsky, A. 1999, private communication
Katsiyannis, A.C., Kemp, S.N., Berry, D.S. & Meaburn, J. 1998, A&AS, 132, 387
Kerp, J., Mack, K.-H., Egger, R., Pietz, J., Zimmer, F., Mebold, U., Burton, W.B. & Hartmann, D. 1996, A&A, 312, 67
Knude, J. & H$\o$g, E. 1999, A&A, 341, 451
Knude, J. 1996, A&A, 306, 108
Kumar, K.C. & Thonnard, N. 1983, AJ, 88, 260
Lagache, G., Abergel, A., Boulanger, F. & Puget, J.-L. 1998, A&A, 333, 709
Leggett, S.K., Clowes, R.G., Kalafi, M., MacGillivray, H.T., Puxley, P.J., Savage, A. & Wolstencroft, R.D. 1987, MNRAS, 227, 563
Lieu, R., Mittaz, J.P.D., Bowyer, S., Lockman, F.J., Hwang, C.-Y. & Schmitt, J.H.M.M. 1996, ApJL, 458, 5
Lo, K.Y. & Sargent, W.L.W. 1979, ApJ, 227, 756
Macri, L.M. 1999, astro-ph/9901332
Madore, B. & Freedman, W.L. 1991, PASP, 103, 933
Maoz, D. 1995, ApJ, 455, L115
Murawski, W. 1983, Acta Cosmologica, 12, 7
Okroy, R. 1965, Astron. Cirk., 320, 4
Pickering, T.E., Impey, C.D., Van Gorkom, J.H. & Bothun, G.D. 1997, AJ, 114, 1858
Rudnicki, K. 1986, [*Gamov Cosmology*]{} (F.Melchiorri & R.Ruffini, eds.), LXXXVI Corso of the Soc. It. Fisica, Bologna, p. 480
Rudnicki, K. & Baranowska, M. 1966, Acta Astron., 16, 65
Schneider, S.E., Helou, G., Salpeter, E.E. & Terzian, Y. 1983, ApJ, 273, L1
Schlegel, D.J., Finkbeiner, D.P. & Davis, M. 1998, ApJ, 500, 525
Sofue, Y. 1994, ApJ, 431, L91
Sprayberry, D., Impey, C.D., Irwin, M.J., McMahon, R.G. & Bothun, G.D. 1993, ApJ, 417, 114
Stark, R., Dickey, J.M., Burton, W.B. & Wennmacher, A. 1994, A&A, 281, 199
Taylor, C.L., Thomas, D.L., Brinks, E. & Skillman, E.D. 1996, ApJS, 107, 143
van den Bergh, S. 1996, PASP, 108, 1091
Wakker, B.P & Boulanger, F. 1986, A&A, 170, 84
Wakker, B.P. & van Woerden, H. 1991, A&A, 250, 509
Wakker, B.P. & van Woerden, H. 1997, ARAA, 35, 217
Weil, M.L., Bland-Hawthorn, J. & Malin, D.F. 1997, ApJ, 490, 664
Wszolek, B., Rudnicki, K., de Bernardis, P., Masi, S. & Salvi, A. 1988a, Ap.Space Sci. 152, 29
Wszolek, B., Rudnicki, K., de Bernardis, P. & Masi, S. 1988b, in [*Large Scale Structures in the Universe*]{}, (Seitter, W.C., Duerbeck, H.W. & Tacke, M., eds.) Lecture Notes in Physics 310, 223
Wszolek, B., Rudnicki, K., de Bernardis, P. & Masi, S. 1989, [*The World of Galaxies*]{} (Corvin, H.G., Jr. & Bottinelli, C., eds.) p. 499
Zonn, W. 1957, Bull. Acad. Polonaise des Sciences, vol. V, no. 1, 47
Zonn, W. & Stodolkiewicz, J. 1958, Bull. Acad. Polonaise des Sciences, ser. sci. math. astr. phys. vol. VI, no. 3, 185
Zonn, W. & Rudnicki, K., 1965, in [*Stellar Astronomy*]{}, Washington DC, 168
Figure captions {#figure-captions .unnumbered}
===============
Figure 1: FIR maps of the VR region. The top right panel shows the IRAS 100$\mu$m map. The COBE/DIRBE 100$\mu$m (top left panel) and 240$\mu$m emission (lower left panel) show the VR, but have a lower angular resolution than the IRAS 100$\mu$m map. The contours of the FIR emission are in MJy ster$^{-1}$. For orientation purposes, the location of M87 is (187.07, +12.67).
Figure 2: Wolf diagram for USNO-A2.0 stars in five selected regions, one corresponding to the lowest FIR-emitting region in Fig. 1 and the other four coinciding with peaks of the FIR emission. The regions are defined in Table 1 and the cumulative distributions are depicted as follows: USNO1222+1330 as a dotted line, USNO1226+1430 as a dashed line, USNO1230+1130 as a long-dashed line, USNO1230+1330 as a solid line with open squares, and USNO1234+1230 as a dot-dashed line.
Figure 3: Comparison of predicted star counts (using the Bahcall-Soneira model) and the actual star counts from the USNO-A2.0 catalog, for the five regions listed in Table 3. Each region is one degree square; the actual star counts are represented by open squares and the model prediction is shown as the solid line.
Figure 4: Selected areas to check for total extinction effects on galaxies. The contours refer to the IRAS 100$\mu$m map and the heavy-lined parallelograms marked A through D are the selected areas for testing the number density of galaxies.
Figure 5: Wolf diagram for galaxies. The three distribution curves A, C, D correspond to regions within higher FIR emission of Fig. 1 and curve B to the region of low FIR emission. For explanation of the designations A, B, C, D see Fig. 5. The data from each of the regions is depicted as follows: region A - solid line, region B - dotted line, region C - dashed line, and region D - long-dashed line.
Figure 6: Total HI column density from the LDS (left panel) compared with the IRAS 100$\mu$m emission (right panel). Darker shades in the HI plot correspond to higher column densities. Lighter shades in the FIR plot correspond to more intense emission.
Figure 7: Galactic latitude-velocity plots near b$\approx$75$^{\circ}$ for l=276.0 (top panel), l=283.0 (middle panel), and l=290.5 (bottom panel). The first plot corresponds to the main blob (more intense FIR emission), the second to the “hole”, center of the VR, and the third to the second blob (next intense FIR emission). The horizontal band marked with heavy lines in each panel indicates the region for which data were used to plot Fig. 8.
Figure 8: Cuts through the position-velocity plots of Fig. 7, averaging the brightness temperature for b=74$^{\circ}, \, 74^{\circ}$.5, and b=75$^{\circ}$ (three latitude bands). The order of the plots is like in Fig. 8. The average HI column density shows the velocity distribution of HI clouds in this region. Three peaks stand out clearly, with the most intense one at the highest negative velocities for the main blob.
[cccccc]{} USNO1222+1330 & 12:22:00 +13:30:00 & 185.5 +13.5 & 277.3 +74.7 & 1804 USNO1226+1430 & 12:26:00 +14:30:00 & 186.5 +14.5 & 279.3 +76.0 & 1593 USNO1230+1130 & 12:30:00 +11:30:00 & 187.5 +11.5 & 286.4 +73.5 & 2126 USNO1230+1330 & 12:30:00 +13:30:00 & 187.5 +13.5 & 284.4 +75.4 & 1696 USNO1234+1230 & 12:34:00 +12:30:00 & 188.5 +12.5 & 289.0 +74.7 & 1669
[ccccccc]{}
15.5 & 1.485(.109)\[08\] & 1.300(.103)\[04\] & 1.360(.186)\[06\] & 1.678(.162)\[08\] & 1.751(.118)\[12\] 14.5 & 1.347(.079)\[27\] & 1.476(.060)\[48\] & 1.547(.061)\[46\] & 1.380(.096)\[26\] & 1.526(.083)\[30\] 13.5 & 1.573(.053)\[68\] & 1.508(.057)\[56\] & 1.545(.074)\[39\] & 1.390(.094)\[26\] & 1.445(.090)\[22\] 12.5 & 1.363(.070)\[36\] & 1.602(.047)\[57\] & 1.608(.058)\[51\] & 1.593(.070)\[45\] & 1.693(.060)\[35\] 11.5 & 1.274(.075)\[36\] & 1.582(.068)\[32\] & 1.658(.054)\[72\] & 1.514(.066)\[45\] & 1.420(.101)\[26\] 10.5 & 1.296(.070)\[11\] & 1.347(.077)\[22\] & 1.530(.099)\[23\] & 1.596(.104)\[20\] & 1.054(.114)\[14\]
[ccccccc]{}
15.5 & -.37(1.48) .11(.75) & 1.89(2.62) 1.35(1.01) & .83(1.74) .16(.85) & .74(1.48) -.46(.75) & .18(1.23) -.64(.64) 14.5 & -.31(.87) .58(.45) & -1.38(.67) -.05(.34) & -1.10(.69) -.36(.35) & -.96(.89) .11(.46) & .35(.83) .00(.43) 13.5 & .83(.57) -.86(.29) & -.32(.63) -.07(.32) & -.03(.74) .45(.38) & -.09(.89) .53(.47) & 1.89(.95) -.91(.49) 12.5 & -.43(.77) -.33(.39) & 1.38(.62) -.71(.32) & .43(.67) -.47(.33) & .35(.70) -.64(.35) & -.93(.78) -.43(.40) 11.5 & -.47(.77) -.08(.39) & -1.06(.81) .10(.41) & -.21(.56) -.10(.28) & -.10(.70) .38(.35) & -.86(.89) -.02(.46) 10.5 & .59(1.28) -.56(.66) & -1.04(.95) -.55(.49) & -.58(.94) -.55(.48) & .14(.99) .57(.51) & -.41(1.15) .11(.60)
[^1]: http://www.sns.ias.edu/$\sim$jnb/Html/galaxy.html
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Uniform sampling of binary matrix with fixed margins is an important and difficult problem in statistics, computer science, ecology and so on. The well-known swap algorithm would be inefficient when the size of the matrix becomes large or when the matrix is too sparse/dense. Here we propose the Rectangle Loop algorithm, a Markov chain Monte Carlo algorithm to sample binary matrices with fixed margins uniformly. Theoretically the Rectangle Loop algorithm is better than the swap algorithm in Peskun’s order. Empirically studies also demonstrates the Rectangle Loop algorithm is remarkablely more efficient than the swap algorithm.'
author:
- 'Guanyang Wang[^1]'
bibliography:
- 'bibliography.bib'
title: A Fast MCMC for the Uniform Sampling of Binary Matrices with Fixed Margins
---
Introduction {#sec:Introduction}
============
The problem of sampling binary matrices with fixed row and column sums has attracted much attention in numerical ecology. In ecological studies, the binary matrix is called *occurrence matrix*. Rows usually corresponds to species, the columns, to locations. For example, the binary matrix shown on Table \[tab:finch\] is known as “Darwin’s Finch” dataset, which comes from Darwin’s studies of the finches on the Galapagos islands (an archipelago in the East Pacific). The matrix represents the presence/absence of 13 species of finches in 17 islands. A $``1"$ or $``0"$ in entry $(i,j)$ indicates the presence or absence of species $i$ at island $j$. It is clear from Table \[tab:finch\] that some pairs of species tend to occur together (for example, species 9 and 10) while some other pairs tend to be disjoint. Therefore, it is of our interest to investigate whether the cooperation/competition influences the distribution of species on islands, or the patterns found are just by chance.
---- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- ---
A B C D E F G H I J K L M N O P Q
1 0 0 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1
2 1 1 1 1 1 1 1 1 1 1 0 1 0 1 1 0 0
3 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 0 0
4 0 0 1 1 1 0 0 1 0 1 0 1 1 0 1 1 1
5 1 1 1 0 1 1 1 1 1 1 0 1 0 1 1 0 0
6 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0
7 0 0 1 1 1 1 1 1 1 0 0 1 0 1 1 0 0
8 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
9 0 0 1 1 1 1 1 1 1 1 0 1 0 0 1 0 0
10 0 0 1 1 1 1 1 1 1 1 0 1 0 1 1 0 0
11 0 0 1 1 1 0 1 1 0 1 0 0 0 0 0 0 0
12 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0
13 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
---- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- ---
: Occurrence Matrix occurrence matrix of the finches on the Galapagos islands.[]{data-label="tab:finch"}
Assuming different species have independent distributions on islands, then the observed binary matrix is simply a random sample from the uniform distribution of all the binary matrices with fixed margins. Table \[tab:fixed\_sum\] gives an example of all configurations of $3\times 3$ binary matrices with $[1,2,1]$ as both row and column sums. Ideally, if we could list all the binary matrices with arbitrary size, then we could compare the pattern found in the observed matrix with others, to conclude whether the observed matrix is simply by chance. However, enumerating matrices with fixed margins is often impractical both theoretically and computationally for moderate size of matrices. Therefore, sampling such random matrices becomes the natural choice.
[ c c c c c ]{} A & B & C & D & E\
\
$ \begin{bmatrix} 0 & 1 & 0 \\ 1 & 0 & 1 \\ 0 & 1 & 0 \end{bmatrix}$ & $ \begin{bmatrix} 0 & 1 & 0 \\ 1 & 1 & 0 \\ 0 & 0 & 1 \end{bmatrix}$ & $ \begin{bmatrix} 1 & 0 & 0 \\ 0 & 1 & 1 \\ 0 & 1 & 0 \end{bmatrix}$ & $ \begin{bmatrix} 0 & 1 & 0 \\ 0 & 1 & 1 \\ 1 & 0 & 0 \end{bmatrix}$ & $ \begin{bmatrix} 0 & 0 & 1 \\ 1 & 1 & 0 \\ 0 & 1 & 0 \end{bmatrix}$\
The problem of sampling such matrices also occurs in many other fields, with different names. For example, an equivalent formulation is uniformly sampling undirected bipartite graphs with given vertex degrees. A bipartite graph $G = (U,V,E)$ is a graph whose vertices are divided into two disjointed sets, denoted by $U = \{u_1, \cdots, u_m\}$, $V = \{v_1, \cdots, v_n\}$. $E$ is called the edge set where every edge connects one vertex in $U$ to one in $V$. The binary matrix $M = (m_{i,j})_{m\times n }$ is often called the bi-adjacency matrix of $G$ and is defined by $$m_{i,j} = \begin{cases}
1, \text{if there is an edge connecting}~ u_i ~\text{and}~ v_j,\\
0, \text{otherwise}.
\end{cases}$$ Bipartite graphs are often used in network studies to model the interaction between two objects, for example, customers and products. It is often required to sample graphs with preserved degree sequence in network analysis uniformly. Throughout this paper, we will use the term ‘binary matrix’ instead of ‘bipartite graph’ to avoid confusion, although they are equivalent.
The algorithms of sampling binary matrices with fixed margins are divided into two classes. The first class of algorithms relies on the *rejection sampling* or *importance sampling* techniques, see [@snijders1991enumeration], [@miklos2004randomization], [@chen2005sequential], [@harrison2013importance] [@holmes1996uniform] for examples. Importance sampling usually provides degree from non-uniform distribution, but it can be used to construct estimators to estimate the quantities of interest, such as the number of binary matrices with given margins. Chen et al. [@chen2005sequential] introduced a sequential importance sampling (SIS) scheme to test the hypothesis we mentioned at the beginning of this paper on “Darwin’s Finch" dataset.
The second class falls into the *Markov Chain Monte Carlo* (MCMC) category and will be our main focus in this paper. The well-known “swap algorithm" has been used for decades. To the author’s best knowledge, it is first introduced by Besag and Clifford [@besag1989generalized] in 1989 to solve a statistical testing problem. The swap algorithm has been formally proposed and analysed by Rao et al. [@rao1996markov] and [@kannan1999simple] in the 1990s. A similar question is to sample matrices with non-negative integer entries, fixed row and column sums. Diaconis and Gangolli have proposed a random walk Metropolis algorithm [@diaconis1995rectangular]. Many variations and extensions of this algorithm are described by Diaconis and Sturmfels [@diaconis1998algebraic].
The swap method attempts to make a *single* swap in each iteration, but when the matrix is large, or is mostly filled (or unfilled), the efficiency of swap algorithm can be relatively low. In 2008, Verhelst [@verhelst2008efficient] proposed a new MCMC algorithm based on the idea of performing multiple swaps per iteration. In 2014, Strona et al. [@strona2014fast] introduced the “Curveball algorithm", which uses a ‘fair trade’ operation to replace the ‘swap’ operation in the swap algorithm, aiming for a faster mixing. The mathematical formulation of Curveball algorithm is equivalent to Verhelst’s algorithm, but with different implementation and reasoning. A nice survey and numerical comparisons of the existing algorithms can be found in a recent dissertation [@rechner2018markov]. The class of ‘multiple swaps’ algorithms tends to improve the mixing time empiricially. However, each step of the ‘multiple swaps’ algorithm is slower than the classical swap algorithm. Meanwhile, it is hard to compare the ‘multiple swaps’ algorithms and classical swap algorithm theoretically, as the corresponding Markov-chains have complicated behaviors and are therefore hard to analyze mathematically. The only existing result can be found in [@carstens2018speeding].
In this paper, we introduce a novel algorithm called Rectangle Loop algorithm. The algorithm is based on the classical swap algorithm, with a careful utilization of the matrix structure given by margins. We have also proved the resulting Markov Chain dominates the classical chain used in the swap algorithm in the sense of Peskun’s partial ordering [@peskun1973optimum], and is easy to implement. Section \[sec:Existing Methods\] gives a review of swap algorithm and Curveball algorithm, including the details of both algorithms and a discussion. In Section \[sec:rectangle\_loop\] we introduce our new algorithm – Rectangle Loop algorithm. Section \[sec:theoretical\_results\] proves the theoretical properties of Rectangle Loop algorithm. Section \[sec: Simulation\_results\] gives numerical results.
Existing Methods {#sec:Existing Methods}
================
Swap Algorithm
--------------
The swap algorithm, or equivalently, swap chain is based on the idea of *swapping checkerboard units*. Here a checkerboard unit is a two by two matrix with one of the following forms: $$\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix},
\begin{pmatrix}
0 & 1 \\
1 & 0
\end{pmatrix}.$$ A swap means changing one checkerboard unit to the other.
Starting from an initial matrix, one chooses two rows and two columns uniformly at random among all rows and columns. If the resulting $2\times 2$ submatrix with entries in the intersection of these rows and columns is a checkerboard unit, it is swapped, otherwise, do nothing.
**Input:** initial binary matrix $A_0$, number of iterations $T$\
Choose two distinct rows and two distinct columns uniformly at random **If** the corresponding $2\times 2$ submatrix of $A_{t-1}$ is a checkerboard unit, i.e. $$\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix} \qquad \text{or} \qquad
\begin{pmatrix}
0 & 1 \\
1 & 0
\end{pmatrix} ,$$
swap the submatrix $\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix}
\leftarrow \begin{pmatrix}
0 & 1 \\
1 & 0
\end{pmatrix}$ or vice versa.
**Otherwise** $A_{t} \leftarrow A_{t-1} $
The swap algorithm is a Metropolis-type Markov chain Monte Carlo which converges to the uniform distribution.
Curveball Algorithm
-------------------
The swap algorithm can often be inefficient, taking Darwin’s Finch data for example, there are ${13\choose2} {17\choose2} = 10608$ submatrices with size $2\times 2$, however, only about $3\%$ of them are swappable. This means it requires a very large $T$ (the number of iterations) to ensure the generated degree is close to uniformly distributed. The Curveball algorithm provides another solution.
**Input:** initial binary matrix $A_0$, number of iterations $T$\
Choose two distinct rows $r_a, r_b$ uniformly at random Determine two disjoint sets $$\begin{aligned}
&S_{a-b} \doteq \{k: A_t(a,k) = 1, A_t(b,k) = 0\} \\
&S_{b-a} \doteq \{l: A_t(a,l) = 0, A_t(b,l) = 1\},\\\end{aligned}$$ here assuming $\lvert S_{i-j} \rvert \leq \lvert S_{j-i} \rvert$ Choose a subset $V\subset S_{j-i}$ uniformly at random Set $A_{t+1} = A_{t}$ except for row $a, b$.
For row $a$: $$A_{t+1}(a, l) = \begin{cases}
1 \quad \text{if}~ l\in V\\
0 \quad \text{if}~ l\in S_{a-b} \cup S_{b-a} \setminus V \\
A(a,l) \quad \text{otherwise}
\end{cases}$$
For row $b$ $$A_{t+1} (b, k ) = \begin{cases}
1 \quad \text{if}~ k\in S_{a-b} \cup S_{b-a} \setminus V\\
0 \quad \text{if}~ k\in V \\
A(b,k) \quad \text{otherwise}
\end{cases}$$
The Curveball algorithm uses ‘trade’ instead of ‘swap’ operation in each iteration. Steps 3-5 in Algorithm \[alg:Curveball\] gives an illustration of trading, it trades elements in column $V$ with elements in column $S_{a-b}$ for row $a$ and row $b$, preserving their row and column sums. Though seemingly complicated, there is a very intuitive explanation of the Curveball algorithm. We refer the readers to [@strona2014fast] for detailed illustrations.
Rectangle Loop Algorithm {#sec:rectangle_loop}
========================
The swap algorithm is proven to converge to uniform distribution. However, getting stuck at the same configuration is inefficient and thus the convergence could be very slow. For example, numerical experiments suggest that one would expect more than $30$ iterations before each successful swap using ‘Darwin’s Finch’ dataset. Assuming the randomly chosen row is mostly filled, such as row $1$ in Table \[tab:finch\], the two random chosen entries in this row would most likely be $[1,1]$ but the row of a ‘checkerboard unit’ has to be either $[1,0]$ or $[0,1]$. Therefore swapping rarely happens when the chosen row/column is mostly filled (or equivalently, mostly unfilled).
The Rectangle Loop algorithm is designed to increase the chance of swapping. The idea is illustrated in Figure $1$. In this example the target matrix is of size $5\times 5$, with row names $R_1, \cdots, R_5$ and column names $C_1, \cdots, C_5$. In each step, we choose one row and one column uniformly at random (Step A). Suppose $R_2$ and $C_2$ is chosen, with corresponding entry $1$ , the red number in the top middle plot of Figure \[fig:Rectangle\_loop\]. Then we randomly choose a $0$ among all the $0$s in $R_2$ (Step B). Since there is only one $0$ in $R_2$, which is at location $C_4$, this is our only choice. Again, we scan through all the entries in the same column with the $0$ just chosen ($C_4$) and randomly choose a $1$ among all $1$s (Step C). In our example, the $1$s of $C_4$ are located at $R_1$ and $R_4$. Suppose we have chosen $(R_4, C_4)$. Now the three locations $(R_2, C_2), (R_2, C_4), (R_4, C_4)$ altogether give us the fourth one $(R_4, C_2)$, making the four entries a rectangle (Step D). If the fourth entry equals $0$, then we swap the submatrix as we did in swap method (Step E). Otherewise the fourth entries equals $1$ and the original matrix is not changed. After Step A - E is iterated for many times, the resulting randomized matrices are used as representatives of uniformly distributed matrix with fixed margins.
The main difference between the Rectangle Loop algorithm and the swap algorithm is the sampling scheme. The Rectangle Loop algorithm is performing ‘conditional sampling’, making it more efficient than swap method, which is doing ‘unconditional choosing’. For example, suppose both the swap method and Rectangle Loop algorithm have chosen $R_2$ and $C_2$, an entry with value $1$. Then $R_2$ has only one $0$ which is in column $4$. For the swap algorithm, the probability of correctly choosing $C_4$ is only $\frac 14$, as it is uniformly choosing among all columns. The Rectangle Loop algorithm, however, as the mechanism guarantees we only sample from the zero entries, chooses $C_4$ with probability $1$. Therefore it significantly increase the swapping probability, leading to a faster convergence than the swap chain.
The details of the Rectangle Loop algorithm is described in Algorithm \[alg:Rectangle\]. Noteworthy, when finding a $0$ entry, we sample $1$ with the same column as the $0$. When finding a $1$, we sample $0$ with the same row as the $1$. This ‘symmetric’ design ensures the algorithm that converges to the correct distribution, as will be proved in Section \[sec:theoretical\_results\]. The paths of sampled entries in each iteration always form a rectangle, that is where the name ‘Rectangle Loop algorithm’ comes from.
**Input:** initial binary matrix $A_0$, number of iterations $T$\
Choose one row and one column $(r_1,c_1)$ uniformly at random Choose one column $c_2$ at random among all the $0$ entries in $r_1$ Choose one row $r_2$ at random among all the $1$ entries in $c_2$ [$A_{t-1}(r_1,c_1) = 0$]{} Choose one row $r_2$ at random among all the $1$ entries in $c_1$ Choose one column $c_2$ at random among all the $0$ entries in $r_2$ Swap the submatrix [$A_t \leftarrow A_{t-1}$]{}
Theoretical Results {#sec:theoretical_results}
===================
Given row sums $\textbf{r} = (r_1, r_2, \cdots, r_m)$ and column sums $\textbf{c} = (c_1, c_2, \cdots, c_n)$, we define $\Sigma_{\textbf r, \textbf c}$ be the set of all matrices with row sums $\textbf r$ and column sums $\textbf c$. The suffcient and necessary condition for $\Sigma_{\textbf r, \textbf c}$ not being zero is given by Gale [@gale1957theorem], Ryser [@ryser2009combinatorial] in 1957. We call two matrices $A$, $B$ ‘swappable’ if one can transform to the other via one step swap algorithm. Equivalently, $A$ and $B$ only differs in a $2\times 2$ ‘checkerboard unit’. For the sake of simplicity, we assume henceforth $0 < r_i < n, 0 < c_j < m$ for any $1\leq i \leq m, 1\leq j \leq n$, as otherwise we could simply delete that degenerate row/column. The following theorem characterizes the limit distribution and transition probability of the swap chain.
Given $\textbf{r}, \textbf{c}$ and an initial matrix $A_0 \in \Sigma_{\textbf r, \textbf c}$, the swap algorithm defines a Metropolis-type Markov chain with stationary distribution $\text{Unif}(\Sigma_{\textbf r, \textbf c})$, transition kernel: $${\mathbb{P}}( A , B) = \begin{cases}
\frac{1}{{m \choose 2} {n \choose 2}} \qquad &\text{If $A$ and $B$ are swappable},\\
1 - \frac{s(A)}{{m \choose 2} {n \choose 2}} \qquad &\text{If $B = A$}, s(A) \doteq \#\{\text{C: C and A are swappable}\}\\
0 \qquad &\text{Otherwise}
\end{cases}$$ and acceptance probability $1$.
When $A$ and $B$ are swappable, there exists two rows $i_1, i_2$ and two columns $j_1, j_2$ such that $A$ and $B$ only differs in the $2\times 2$ submatrix extracted by row $i_1, i_2$ and column $j_1, j_2$. Therefore the probability of swapping $A$ to $B$ equals $$\frac{1}{{m \choose 2} {n \choose 2}}.$$ Recall that in the setting of Metropolis-Hastings, the acceptance probability from $A$ to $B$ is given by $$\min\{1, \frac{\pi(B){\mathbb{P}}(B,A)}{\pi(A){\mathbb{P}}(A,B)},\}$$ notice here the stationary distribution $\pi$ is designed to be $\text{Unif}(\Sigma_{\textbf r, \textbf c})$, ${\mathbb{P}}(A,B) = {\mathbb{P}}(B,A) = \frac{1}{{m \choose 2} {n \choose 2}}$. Hence it is clear that the swap algorithm is a Metropolis-type Markov chain with $\text{Unif}(\Sigma_{\textbf r, \textbf c})$ as stationary distribtuion and acceptance probability $1$, which justifies the correctness of the swap algorithm.
The key point in swap algorithm is symmetry. When two different states $A,B$ are swappable, the associated transition probability is symmetric, i.e., $${\mathbb{P}}(A,B) = {\mathbb{P}}(B,A),$$ this ensures the chain has acceptance probability $1$.
In order to compare the efficiency of different Markov kernels with the same distribution, Peskun [@peskun1973optimum] first introduced the following partial-ordering.
Let ${\mathbb{P}}_1$, ${\mathbb{P}}_2$ be two Markov transition kernels on the same state space $\mathcal S$ with same stationary distribution $\pi$, then ${\mathbb{P}}_1$ *dominates* ${\mathbb{P}}_2$ *off the diagonal*, ${\mathbb{P}}_1 \succeq {\mathbb{P}}_2 $, if $${\mathbb{P}}_1(x, A) \geq {\mathbb{P}}_2 (x, A)$$ for all $x\in \mathcal S$ and $A$ measurable with $x\notin A $.
When the state space is finite, as in our case, ${\mathbb{P}}_1 \succeq {\mathbb{P}}_2$ iff all the off-diagonal entries of ${\mathbb{P}}_1$ are greater than or equal to the corresponding off-diagonal entries of ${\mathbb{P}}_2$. This indicates ${\mathbb{P}}_1$ has lower probability to get stuck in the same state, and is exploring the state space in a more efficient way. The following theorem shows the rectangale loop algorithm also has uniform distribution as stationary distribution, and the corresponding chain dominates the swap chain off the diagonal. For simplicity, we will use ${\mathbb{P}}_s$ and ${\mathbb{P}}_r$ to denote transition kernel for the swap chain and rectangle loop chain, respectively, omitting its dependency on $\textbf{r}, \textbf{c}, \text{Unif}(\Sigma_{\textbf r, \textbf c}).$
Given $\textbf{r}, \textbf{c}$ and an initial matrix $A_0 \in \Sigma_{\textbf r, \textbf c}$, the Rectangale loop algorithm defines a Metropolis-type Markov chain with stationary distribution $\text{Unif}(\Sigma_{\textbf r, \textbf c})$. The transition kernel ${\mathbb{P}}_r$ dominates ${\mathbb{P}}_s$ off the diagonal.
Given any two swappable configurations $A$ and $B$, we are aiming to show ${\mathbb{P}}_r(A, B) = {\mathbb{P}}_r(B,A) \geq {\mathbb{P}}_s(A, B)$.
As $A, B$ are swappable, there exists two rows $i_1, i_2$ and two columns $j_1, j_2$ such that $A$ and $B$ only differs in the $2\times 2$ submatrix extracted by row $i_1, i_2$ and column $j_1, j_2$. Without loss of generality, we assume the checkerboard unit corresponding to $A$ has the form $\begin{bmatrix}
1 & 0 \\ 0 & 1
\end{bmatrix}$, as shown in Figure \[sec:theoretical\_results\]. Notice that there are four vertices of the $2\times 2$ submatrix and the Rectangle Loop algorithm chooses one arbitray row and column at its first step. This suggests the probability of transforming $A$ to $B$ equals the summation of four probabilities, each one corresponds to choosing one specific vertex of the ‘checkerboard unit’ .
Figure \[fig:prob\_A\_B\] illustrates the calculation of one path, starting from the vertex $(i_1, j_1)$. The possibility of choosing row $i_1$ and column $j_1$ is $\frac 1 {mn}$. Then one chooses a $0$ among all the $0$s in row $i_1$, and there are $n - r_{i_1}$ of them. Therefore the possibility of choosing column $j_2$ equals $\frac{1}{n - r_{i_1}}$. Similarly, after choosing $j_2$, one chooses a $1$ among all $1$s in column $j_2$, and there are $c_{j_2}$ of them. Hence the possibility of choosing row $i_2$ equals $\frac 1 {c_{j_2}}$. The fourth entry is fixed after determining the first three entries thus the last step has probability $1$. Multipling the possibilities above altogether, the possibility of transforming $A$ to $B$, starting with $(i_1, j_1)$, equals $$\frac{1}{mn}\cdot \frac{1}{n - r_{i_1}} \cdot \frac 1 {c_{j_2}}.$$
The probability of transforming $A$ to $B$ with other starting vertex can be calculated accordingly. It turns out ${\mathbb{P}}_r(A,B)$ can be written as the following summation: $${\mathbb{P}}_r(A,B) = \frac{1}{mn} \bigg( \frac{1}{n - r_{i_1}} \cdot \frac 1 {c_{j_2}} + \frac 1 {c_{j_2}} \cdot \frac{1}{n - r_{i_2}} + \frac{1}{n - r_{i_2}} \cdot \frac 1 {c_{j_1}} + \frac 1 {c_{j_1}} \cdot \frac{1}{n - r_{i_1}} \bigg).$$
Following the same strategy, ${\mathbb{P}}_r(B,A)$ can also be calculated below. Figure \[fig:prob\_B\_A\] illustrates the calculation of ${\mathbb{P}}_r(B,A)$ starting from $(i_1, j_1)$. $${\mathbb{P}}_r(B,A) = \frac{1}{mn} \bigg( \frac 1 {c_{j_1}} \cdot \frac{1}{n - r_{i_2}} + \frac{1}{n - r_{i_2}} \cdot \frac 1 {c_{j_2}} + \frac 1 {c_{j_2}}\cdot \frac{1}{n - r_{i_1}} + \frac{1}{n - r_{i_1}} \cdot \frac 1 {c_{j_1}} \bigg).$$
After matching all the terms of ${\mathbb{P}}_r(B,A) $ with ${\mathbb{P}}_r(A,B)$, we conclude that ${\mathbb{P}}_r(B,A) = {\mathbb{P}}_r(A,B)$, which justifies the Rectangle Loop algorithm has Unif($\Sigma_{\textbf r, \textbf c}$) as stationary distribution.
To show ${\mathbb{P}}_r(A, B) \geq {\mathbb{P}}_s(A, B)$, notice that ${\mathbb{P}}_s(A, B) = \frac{1}{{m \choose 2} {n \choose 2}} = \frac 4 {m(m-1)n(n-1)}$. and $$\begin{aligned}
{\mathbb{P}}_r(A,B) = &\frac{1}{mn} \cdot \frac{1}{n - r_{i_1}} \cdot \frac 1 {c_{j_2}} + \frac{1}{mn} \cdot \frac 1 {c_{j_2}} \cdot \frac{1}{n - r_{i_2}} + \\
& \frac{1}{mn} \cdot \frac{1}{n - r_{i_2}} \cdot \frac 1 {c_{j_1}} + \frac{1}{mn} \cdot \frac 1 {c_{j_1}} \cdot \frac{1}{n - r_{i_1}} .\\\end{aligned}$$ It is clear that ${\mathbb{P}}_r(A,B)$ can be written as the summation of four terms. Each term in the summation is greater than or equal to $\frac 1 {m(m-1)n(n-1)}$. Therefore we conclude that ${\mathbb{P}}_r(A,B) \geq {\mathbb{P}}_s(A,B) $ for any swappable $A,B$. This indicates ${\mathbb{P}}_r \succeq {\mathbb{P}}_s$, the Rectangle Loop algorithm is exploring the state space in a more efficient way than the swap algorithm.
Simulation Results and Applications {#sec: Simulation_results}
===================================
A Concrete Example
------------------
Now we use the example used in [@strona2014fast] and [@miklos2004randomization] to compare the existing algorithms and the Rectangle Loop algorithm. The example below is concrete. The transition matrix can be calculated explicitly and convergence can be assessed analytically.
The five matrices shown in Table \[tab:fixed\_sum\] are all possible configurations of $3\times 3$ binary matrices with $[1,2,1]$ as both row and column sums. The transition matrices for swap algorithm, Curveball algorithm and Rectangale loop algorithm is shown in Table \[tab:transition\].
[ c c c ]{} Swap & Curveball & Rectangle Loop\
\
$ \begin{bmatrix} \frac5 9 & \frac19 & \frac19 & \frac19 &\frac19 \\ \frac 19& \frac23 & 0 &\frac19 & \frac 19 \\ \frac19& 0 & \frac 23 &\frac19 &\frac19 \\ \frac 19 & \frac 19 & \frac19 &\frac 23 & 0\\
\frac19 & \frac19 &\frac 19 &0 & \frac23
\end{bmatrix}$ & $ \begin{bmatrix}
\frac13 & \frac16 & \frac16 & \frac16 &\frac16 \\ \frac16 & \frac13 & 0 & \frac16 &\frac13\\ \frac16 & 0 & \frac13 &\frac13 & \frac 16\\
\frac 16 & \frac16 &\frac13 &\frac13 & 0 \\
\frac 16 &\frac 13 &\frac16 &0 &\frac 13 \end{bmatrix}$ & $ \begin{bmatrix} 0 & \frac14 & \frac14 &\frac14 &\frac14 \\ \frac14 & \frac14 & 0 &\frac13 &\frac16 \\ \frac14& 0 & \frac14 & \frac 16 &\frac 13\\
\frac 14 &\frac 13 &\frac16 &\frac 14 & 0\\
\frac14 &\frac16 &\frac 13 &0 &\frac14 \end{bmatrix}$\
Figure \[fig:comparison\_3by3\] shows the comparison of the three algorithms. Here we measure the distance between transition kernel ${\mathbb{P}}$ and the stationary distribution $\pi$ by total variation distance:
$$\max_{A\in \Sigma_{\textbf r, \textbf c}}\lVert {\mathbb{P}}^k(A,\cdot) - \pi \rVert_{\text{TV}} = \frac 12 \max_{A\in \Sigma_{\textbf r, \textbf c}} \sum_{B\in \Sigma_{\textbf r, \textbf c}}\lvert{\mathbb{P}}^k(A,B) - \pi(B) \rvert,$$ where $k$ denotes the power. It is clear that all the algorithms converge exponentially to uniform distribution, but the Rectangle Loop algorithm converges faster than swap algorithm and Curveball algorithm.
Experiments on Empirical Mixing Time
------------------------------------
For larger matrices, it is infeasible to calculate the transition matrix theoretically. To provide empirical justification for the advantage of the Rectangle Loop algorithm. We have designed the experiment as follows. For each $p = 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5$, a $100\times 100$ binary matrix is generated for which each entry has probability $p$ to be $1$. We ran each algorithm for $10000$ iterations and collected the corresponding number of swaps, as shown in Table \[tab:comparison\_rectangle\_swap\]. When the filled portion $p$ is small, the Rectangle Loop algorithm is extremely efficient, producing more than $73$ times more swaps than the swap algorithm. For large $p$, the advantage of Rectangle Loop algorithm is reduced, but still very significant. For $p = 0.5$, the Rectangle Loop still produces $4$ times more swaps than the swap algorithm. Noteworthy, the zeros and ones play the symmetric rule in a binary matrix, therefore it is not necessary to generate the random matrix for $p > 0.5$.
The result above justifies our theoretical result that the Rectangle Loop algorithm converges faster than the swap algorithm. However, the above experiments did not consider the running time for each iteration. In fact, one iteration of Rectangle Loop algorithm is computationally more expensive than that of swap algorithm. To investigate this issue, we also record the time per swap for both algorithms, as shown in the last column of Table \[tab:comparison\_rectangle\_swap\]. It turns out that the Rectangle Loop algorithm still has a significant advantage than the swap algorithm after the running time issue is taken into account. For $p = 0.01$, the Rectangle Loop is about $31$ times more efficient than the swap algorithm. Even for $p = 0.5$, Rectangle Loop algorithm is still about $4$ times more efficient than the swap algorithm.
Filled portion Number of swaps Time per swap (/s)
-- ---------------- ----------------- ----------------------
$586$ $1.18\times 10^{-5}$
$8$ $3.67\times 10^{-4}$
$977$ $5.30\times 10^{-6}$
$42$ $3.52\times 10^{-5}$
$1838$ $3.23\times 10^{-6}$
$156$ $1.25\times 10^{-5}$
$3271$ $2.64\times 10^{-6}$
$509$ $5.68\times 10^{-6}$
$4222$ $2.10\times 10^{-6}$
$803$ $5.06\times 10^{-6}$
$4794$ $1.27\times 10^{-6}$
$1160$ $4.98\times 10^{-6}$
$5080$ $1.37\times 10^{-6}$
$1271$ $5.36\times 10^{-6}$
: The comparison between swap algorithm and Rectangle Loop algorithm. Each algorithm is implemented $10000$ iterations on $100\times 100$ matrices with different filled portions. The third column records the number of successful swaps among the $10000$ iterations, the last column records the average time per swap, respectively.[]{data-label="tab:comparison_rectangle_swap"}
We have also used the pertubation score suggested by Strona et.al. [@strona2014fast] to access convergence for both algorithm. Pertubation score of a matrix is defined by the fraction of cells differing from the corresponding ones of the initial matrix. It takes several iterations for each algorithm to stabilize around its expectation. It is shown in Figure \[fig:Rectangle\_vs\_swap\] that it takes less iterations and less time for the Rectangle Loop algorithm to stabilize, suggesting a faster mixing than the swap algorithm.
Finch Data Applications
-----------------------
Going back to ‘Darwin’s Finch’ dataset, we use the test statistics $\bar{S}^2$ suggested by Roberts and Stone [@roberts1990island] to compare the three algorithms. $\bar{S}^2$ is defined by $$\bar{S}^2(A) = \frac{1}{m(m-1)}\sum_{i\neq j} s_{ij}^2,$$ where $m$ is the number of rows of matrix $A$, $S = (s_{ij}) = A A^T$. For the finch data, $\bar{S}^2 = 45.03$. Suppose this number is too large or too small, comparing with its expectation over all the matrices having the same margins as finch data. We would like to conclude that the cooperation/competition do influence the distribution of species. To investigate this, we implemented the swap algorithm, Curveball algorithm and Rectangle Loop algorithm on the same data for $20000$ iterations, using its average as an estimator for $\bE({{\bar{S}}}^2)$. The results are shown in Figure \[fig:comparison\_finch\]. After $20000$ iterations, Rectangle Loop algorithm gives an estimate of $42.135$ with standard deviation $0.537$, swap algorithm gives an estimate of $42.126$ with standard deviation $0.509$, Curveball algorithm gives an estimate of $42.191$, with standard deviation $0.590$. Therefore the observed data falls outside the three standard deviation boundaries for all three algorithms, suggesting strong evidence that the observed occurence matrix is not just by chance. Meanwhile, both swap algorithm and Rectangle algorithm gives similar estimations and lower standard deviations, which seem to be more accurate than the Curveball algorithm. Lastly, there is a significant pattern in Figure \[fig:comparison\_finch\] that for both $\bar S^2$ and standard deviation estimation, the Rectangle Loop algorithm becomes stabilized much earilier than the swap algorithm, indicating a faster mixing.
Discussion {#sec:discussion}
==========
There is a growing tendency to study the behavior of binary matrices with fixed margins in numerous scientific fields, ranging from mathematics to natural science to social science. For example, mathematicians and computer scientists are interested in the total number of configurations of given margin sums. Ecologists use the so-called *occurence matrix* to model the presence/absence of species in different locations. Biologists use the binary matrix to model neuronal networks. Social scientists use the binary matrix for studying social network features.
One of the central and difficult problems is uniformly sampling binary matrices with given margins. In this article we have developed the Rectangle Loop algorithm which is efficient, intuitive and easy to implement. Theoretically, the algorithm is superior to the classical swap algorithm in Peskun’s order. In practice, the Rectangle Loop algorithm is notably more efficient than the swap approach. For a fixed number of iterations, Rectangle Loop algorithm produces $4-73$ times more successful swaps than the swap algorithm. For a fixed amount of time, Rectangle Loop algorithm still produces $4-31$ times more successful swaps than the swap algorithm. This suggests the Rectangle Loop algorithm is efficient both statistically and computationally.
There are many other problems that remain. From a theoretical point of view, it is important to give sharp bounds on the convergence speed of a given Markov chain. However, giving a useful running time estimate is often challenging in practical problems. It would be very interesting if the swap algorithm, Curveball algorithm and the Rectangle Loop algorithm can be investigated analytically. From an applied point of view, there are many factors that influence the performance of algorithms, such as running time per step (swap algorithm is the fastest, while Curveball algorithm is the slowest), initialization of the matrix, size of the matrix, ratio between row number and column numbers, filled proportions. Our empirical studies suggest that all the factors have a significant impact on the convergence speed for all the algorithms. It would be beneficial if more numerical experiments are carried out, yielding a complete and comprehensive comparison between all the existing algorithms.
[^1]: Guanyang Wang is with the Department of Mathematics, Stanford University. Email: {guanyang}@stanford.edu.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Robert Raussendorf
- 'Tzu-Chieh Wei'
title: Quantum computation by local measurement
---
Introduction
============
Quantum computation is a promising approach to harness the laws of quantum mechanics for solving computational problems. A particular striking example is Shor’s efficient quantum algorithm for factoring large numbers [@Shor], which breaks the RSA crypto system. After this splendid start, the growing field has encountered numerous challenges, some of which it has mastered, some of which it still faces. As an example, decoherence was initially conceived as an insurmountable obstacle to scalable quantum computation [@Unruh]. However, the theory of quantum-error correction [@Shor2]-[@AGP] and, alternatively, the scheme of topological quantum computation [@Topol1; @Topol2], show that it can in principle be overcome. Also, impressive experimental progress has been made in recent years towards realizing quantum computers in the laboratory [@Fact15]-[@Adi]. Yet, building a large-scale device in the foreseeable future remains a great challenge [@roadmap].
At a fundamental level, we may ask “Which quantum mechanical property is responsible for the quantum speedup?” In spite of a number of candidates that have been proposed—such as entanglement, superposition and interference, and largeness of Hilbert space—we have no rigorous and generally applicable answer to this question yet. Making progress in this direction may, in addition to deepening our understanding of quantum computation, also lay the foundation for the design of novel quantum algorithms. In 1948, introducing the path integral formalism to quantum mechanics [@Feyn], Richard Feynman wrote: “[*[One feels like Cavalieri must have felt calculating the volume of a pyramid before the invention of calculus.]{}*]{}” Addressing the above questions in the theory of quantum computation feels like that, too.
The paradigm of measurement-based quantum computation (MBQC), with the teleportation-based schemes [@GoChua] and the one-way quantum computer [@oneway; @oneway2; @oneway3] as the most prominent examples, offers a new framework within which both theoretical and experimental challenges of quantum computation can be addressed. This article focuses on the one-way quantum computer, in which the measurements driving the computation are strictly local. We will discuss its prospects for experimental realization, and examine the roles that entanglement and quantum correlations play for it.
In the one-way MBQC, the process of computation is driven solely by local measurements, applied to a highly entangled resource state. This is in stark contrast to the (standard) circuit model, where the quantum is driven by elementary steps of unitary evolution, so-called quantum gates. In the MBQC, after a highly entangled resource state such as a 2D cluster state [@RB01a] has been created, the local systems, say qubits, are measured individually in certain bases and a prescribed temporal order. The choice of measurement bases specifies which quantum algorithm is being implemented. The measurement outcomes cannot be chosen; they are individually random. This randomness can be prevented from creeping into the logical processing by adjusting measurement bases according to previously obtained measurement outcomes. Finally, the computational output is produced by correlations of measurement outcomes.
The remainder of this article is dedicated to a few questions that arise at this point. (i) Is MBQC experimentally feasible? - We discuss pro’s and con’s for the experimental realization of MBQC in Section \[Exp\]. (ii) Why does MBQC work at all? - We provide an explanation of the inner workings of MBQC in Section \[UniPro\]. (iii) Do resource states for universal MBQC arise naturally in quantum systems? - In Section \[Cluster\], we describe how a particular universal resource state, the cluster state, can be realized via unitary evolution under an Ising Hamiltonian. In Section \[sec:1qubit\], we explain teleportation-based implementations of quantum gates. Building on that, in Section \[sec:MBQCuniversal\], we explain how a CNOT gate and general one-qubit rotations can be realized using cluster states, leading to universality of MBQC. Section \[GS\] describes how computational resources can arise as ground states of relatively simple Hamiltonians. A resource for universal MBQC is the Affleck-Kennedy-Lieb-Tasaki state on the honeycomb lattice; see Section \[AKLT2d\]. (iv) Which role does entanglement play for MBQC? - In MBQC, the result of the computation is obtained at the price of consuming all or most of the entanglement initially present in the resource state. Therefore, entanglement appears as a key resource for MBQC. In Section \[ent\], this intuition is (partially) corroborated, and in Section \[TETU\] its limits are shown. (v) Which role do quantum correlations play for MBQC? The computational power of MBQC hinges on strong correlations among the random measurement outcomes. These classical correlations derive from quantum correlations in the resource state. In Section \[QMF\], we illustrate this in a specific example, by turning the Greenberger-Horne-Zeilinger proof of Bell’s theorem into a measurement-based quantum computation.
![\[flow\] Quantum computation by measuring individual qubits initially prepared in a cluster state on a two-dimensional lattice, circuit simulator view. The choice of measurement bases specifies the sequence of simulated quantum gates. Circles symbolize measurements of $\sigma_z$, and arrows refer to measurement bases in the x-y plane.](flow){width="7cm"}
Prospects for realization of scalable MBQC {#Exp}
==========================================
Physical systems considered for realizing a quantum computer have to meet a set of requirements, known as the DiVincenzo criteria. Specifically, one requires (i) A scalable setup with well-defined qubits, (ii) The ability to initialize the qubits in a fiducial state, say $|00..0\rangle$, (iii) A universal set of quantum gates, (iv) The ability to measure individual qubits, and (v) Long coherence times. For various potential realizations of a quantum computer, such as trapped ions [@ITsum]-[@ITCM], lattices of cold atoms [@CCC; @OLW], photons [@Walther] and superconducting qubits [@Nakamura], at least a subset of the DiVincenzo criteria have been proven in the experiment. It can be expected that these physical systems will mature into medium-scale test beds for quantum computers over the next couple of years, but it is far from certain that which one of them will emerge as the quantum counterpart of the silicon chip.
The scheme of measurement-based quantum computation can simplify the architecture of a quantum computer since it reduces the requirements on the interaction between qubits. First, instead of tunable interactions between selected pairs of qubits, MBQC only requires a translation-invariant, nearest-neighbor Ising coupling. This interaction is highly scalable and parallelized, and requires no control other than an on/off switch.
Physical systems that can naturally make use of this advantage are optical lattices filled with cold atoms. In recent years, substantial experimental progress has been made in trapping, cooling and manipulating cold atoms in optical lattices. Large areas in such a lattice can be regularly filled with one atomic qubit per site, by driving a superfluid to Mott phase transition [@SuMo]. Furthermore, the Ising interaction can be realized by cold controlled collisions between the atoms [@CCC]. Finally, the extremely difficult single site readout has recently been experimentally demonstrated [@SiRea].
A second setting in which MBQC helps to overcome a limitation of the interaction are probabilistic heralded entangling gates. In this setting, the entangling gate sometimes (or mostly) fails but success is confirmed by a classical signal. The problem with using probabilistic gates in quantum circuits in the same way as deterministic ones is that a single failed gate ruins the entire computation. Instead, probabilistic heralded entangling gates may be used to grow a cluster state. A simple protocol for probabilistic growth of a linear cluster state is depicted in Fig. \[PC\]. It works whenever the success probability of the heralded entangling gate is greater than 2/3. This protocol can be refined. It turns out that cluster states of arbitrary size and geometry can be grown efficiently for any success probability $p>0$ of the entangling gate, enabling universal computation [@Kok; @DR].
![\[PC\] Simple method for probabilistic growth of cluster states. A probabilistic heralded conditional phase gate ($\mbox{cPhase}_{ij}:=\exp(i\pi |11\rangle_{ij}\langle 11|)$) is applied to an existing linear cluster state. If the gate succeeds, the length of the cluster is increased by one. If the gate fails, the cluster qubit next to the qubits affected by the failed gate is measured in the $\sigma_z$-eigenbasis, recovering the cluster to the left. The length of the cluster is reduced by two. On average, the cluster grows if the success probability of the entangling gate is $>2/3$.](ProbCluster){width="10cm"}
A physical setting where probabilistic heralded entangling gates arise naturally is the Knill-Laflamme-Milburn (KLM) scheme of linear optics quantum computation [@KLM]. Probabilistic growth of cluster states can be gainfully applied in this setting. Specifically, it reduces the operational overhead from a factor that grows with the size of the computation to a constant factor [@NiePrCl; @BrPrCl]. See [@YoRez] for a similar result within teleportation-based quantum computation [@GoChua]. Readers interested in knowing more detail about how MBQC helps to reduce the resource in KLM scheme beyond the explanation in Fig. \[PC\] should refer to the above cited works.
MBQC also has a disadvantage for the realization of quantum computation. As a glance at Fig. \[flow\] reveals, the number of qubits that need to be stored simultaneously is significantly increased as compared to the circuit model. Note, however, that the cluster state can be continuously created ‘on the fly’, mitigating this effect.
Finally, it shall be noted that MBQC has been realized on a small scale in an experiment with photons [@Walther]. There, a four-qubit cluster state was created and subsequently measured, allowing for the realization of Grover search [@Grover] on a four-item data base.
How MBQC works {#UniPro}
==============
In this section we show that a measurement-based quantum computer has the full power of quantum computation, by mapping to the (standard) circuit model.
To establish this result, we regard MBQC as a circuit simulator. In this view, the cluster state provides a ‘canvas’ upon which a quantum circuit is imprinted by local measurements. One spatial direction of the cluster, the vertical direction, say, labels the positions of logical qubits on a line. The perpendicular direction corresponds to the circuit time. As the measurements progress from left to right, the logical qubits are propagated across the cluster slice by slice, implementing quantum gates in the passing.
We assume familiarity with elementary notions of the circuit model, such as the quantum register, quantum gates and quantum circuits [@NC]. In short, a quantum circuit begins with the initialization of an $n$-qubit quantum register in a fixed state such as $|00...0\rangle$. Second, a sequence of unitary quantum gates is applied. Third, the qubits of the quantum register are individually measured in a fixed basis, and the readout of the computation is thereby obtained.
Cluster states {#Cluster}
--------------
Of central importance for MBQC is the choice of the initial resource state. For sure, this state has to be entangled, but not every entangled state will do. In fact, MBQC resource states that allow for universal quantum computation by local measurement are extremely rare [@GrossFlammiaEisert; @BremnerMoraWinter], and only a small number of such states are known explicitly.
The first universal resource state to be discovered was the two-dimensional cluster state, which we introduce here by a practical two-step procedure to create it. Consider a two-dimensional lattice ${\cal{L}}_2$, with one qubit located at each site $a \in V({\cal{L}}_2)$, the set of vertices, and with an edge set $E({\cal{L}}_2)$, where an $e$ in $E({\cal L}_2)$ denotes a pair of distinct vertices, e.g., $(i,j)$, that are interacting with each other. Then, a 2D cluster state $\CS$ is created by (i) preparing the qubits $a \in V({\cal{L}}_2)$ individually in the state $|+\rangle_a = 1/\sqrt{2}(|0\rangle_a + |1\rangle_a)$, and (ii) unitarily evolving this state under the Ising-like Hamiltonian $$H({\cal{L}}_2) =\hbar g\sum_{(i,j) \in E({\cal{L}}_2)}\frac{I^{(i)}-\sigma_z^{(i)}}{2}\otimes\frac{I^{(j)}-\sigma_z^{(j)}}{2}
= \hbar g \sum_{(i,j) \in E({\cal{L}}_2)}|1\rangle_i\langle 1|\otimes |1\rangle_j\langle 1|,$$ for a time $T=\pi/g$. Equivalently, $$\label{CScreate}
\CS = \prod_{(a,b) \in E({\cal{L}}_2)}\mbox{cPhase}_{a,b}\, \bigotimes_{a \in V({\cal{L}}_2)} |+\rangle_a.$$ Therein, $\mbox{cPhase}_{c,t}:=\exp\left(i\pi
|11\rangle_{c,t}\langle 11| \right)=\mbox{cPhase}_{t,c}$ is a unitary quantum gate, a so-called conditional phase gate, a.k.a. control-phase gate, which is symmetric w.r.t. to the control (c) and target (t) qubits. One can also rewrite the conditional phase gate in a useful representation $$\label{eqn:cp}
\mbox{cPhase}_{c,t}=|0\rangle_c\langle 0|\otimes I_t +
|1\rangle_c\langle1|\otimes Z_t,$$ which only flips the phase of the target qubit when the control qubit is in the state $|1\rangle$. The conditional phase gates on the r.h.s. of Eq. (\[CScreate\]) commute, so that their temporal ordering is immaterial.
The control-phase gate can be shown to satisfy the following properties $$\begin{array}{rcl}
\mbox{cPhase}_{a,b}\, X_a\, \mbox{cPhase}_{a,b}^\dagger &=& X_aZ_b,\\
\mbox{cPhase}_{a,b}\, X_b\, \mbox{cPhase}_{a,b}^\dagger &=& X_bZ_a,\\
\mbox{cPhase}_{a,b}\, Z_a\, \mbox{cPhase}_{a,b}^\dagger &=& Z_a,\\
\mbox{cPhase}_{a,b}\, Z_b\, \mbox{cPhase}_{a,b}^\dagger &=& Z_b,
\end{array}$$ where, for convenience, we have used $X$, $Y$, and $Z$ to denote the Pauli matrices $\sigma_x$, $\sigma_y$, and $\sigma_z$, respectively. Using the above equations, we can show that $$\prod_{ (i,j)\in E({\cal L}_2)} {\rm cPhase}_{i,j}\,X_u=\Big( X_u\prod_{v\in {\rm Nb}(u)}Z_v\Big) \prod_{(i,j)\in E({\cal L}_2)}
{\rm cPhase}_{i,j},$$ where ${\rm Nb}(u)$ denotes the set of vertices that are neighbors of $u$. Applying this relation to an initial state $| +\rangle| +\rangle\cdots|
+\rangle$, we arrive at $$\label{Cdef}
\Big(X_u \prod_{v\in {\rm Nb}(u)}Z_v \Big)|\phi\rangle_{{\cal
L}_2}=|\phi\rangle_{{\cal L}_2},\;\; \forall u\in V({\cal{L}}_2).$$ Eq. (\[Cdef\]) uniquely specifies the cluster state $|\phi\rangle$ given the graph ${\cal L}_2$, and is thus equivalent to Eq. (\[CScreate\]) as a definition for cluster states. It has the advantage of being independent of any particular creation procedure.
Cluster states are special cases of a slightly more general class of states, the graph states. We now define graph states in a way similar to Eq. (\[Cdef\]).
\[DC\]Consider a graph $G$ with vertex set $V(G)$ and edge set $E(G)$, and a set of qubits, one for each vertex $a \in V(G)$. The graph state $|G\rangle$ is the unique simultaneous eigenstate with eigenvalue 1 of the Pauli operators $$\label{CScorr}
K_a = X_a \bigotimes_{b \in {\rm Nb}(a) \in E(G)}Z_b,\;\;\;\; \forall a \in V(G),$$ i.e., $|G\rangle = K_a|G\rangle$ for all $a \in V(G)$. Therein, $X \equiv \sigma_x=|0\rangle\langle 1| + |1\rangle\langle 0|$ and $Z \equiv \sigma_z=|0\rangle\langle 0| - |1\rangle\langle 1|$.
A cluster state $|\phi\rangle_{\cal{L}}$ is a graph state with the corresponding graph being a lattice ${\cal{L}}$ of some dimension $d$, $|\phi\rangle_{\cal{L}}:= |{\cal{L}}\rangle$.
Let us now consider two examples of one-dimensional cluster states. Example I: The two-qubit cluster state $|\phi_2\rangle$. From the first definition Eq. (\[CScreate\]), using $\mbox{cPhase}_{i,j}:=\exp(i\pi|11\rangle_{ij}\langle 11|) =
|0\rangle_i\langle 0|\otimes I^{(j)}+ |1\rangle_i\langle 1|\otimes
Z^{(j)}$, we have $$\label{phi2} |\phi_2\rangle= {\rm
cPhase}_{12}|+\rangle_1|+\rangle_2=\frac{1}{\sqrt{2}} {\rm
cPhase}_{12}(|0\rangle_1+|1\rangle_1)|+\rangle_2=\frac{
|0\rangle_1|+\rangle_2+|1\rangle_1|-\rangle_2}{\sqrt{2}},$$ where we have expanded the first qubit in $|0/1\rangle$ basis and used the relation $Z|\pm\rangle=|\mp\rangle$ to flip the phase of the second qubit. Although both the initial state and the conditional phase gate are symmetric w.r.t. qubits 1 and 2, the final state does not explicitly show this symmetry in the given basis. One may as well use the second qubit as the control and the first qubit as the target and redo the calculation, $$\label{phi2a} |\phi_2\rangle= {\rm
cPhase}_{21}|+\rangle_1|+\rangle_2=\frac{1}{\sqrt{2}} {\rm
cPhase}_{21}|+\rangle_1(|0\rangle_2+|1\rangle_2)=\frac{|+\rangle_1|0\rangle_2+|-\rangle_1|1\rangle_2}{\sqrt{2}}.$$ The second definition Eq. (\[CScorr\]) of the cluster state $|\phi_2\rangle$ yields the relations $X_1Z_2|\phi_2\rangle=Z_1X_2|\phi_2\rangle=|\phi_2\rangle$. We can now explicitly verify that the state on the r.h.s. of Eq. (\[phi2\]), as well as Eq. (\[phi2a\]), satisfies these two relations.
Example II: The three-qubit cluster state $|\phi_3\rangle$ on a line. Defined through Eq. (\[CScreate\]), it takes the form $$\label{phi3}
|\phi_3\rangle= {\rm cPhase}_{23}\, {\rm
cPhase}_{21}|+\rangle_1|+\rangle_2 |+\rangle_3={\rm
cPhase}_{23}|\phi_2\rangle|+\rangle_3=\frac{|+\rangle_1|0\rangle_2|+\rangle_3+|-\rangle_1|1\rangle_2|-\rangle_3}{\sqrt{2}}$$ Notice that we have used the second form of $|\phi_2\rangle$ (\[phi2a\]), indicated by the labeling of the conditional phase gate ${\rm cPhase}_{21}$, as then the second qubit is in the $|0/1\rangle$ basis, convenient for the subsequent gate ${\rm cPhase}_{23}$. Eq. (\[Cdef\]) yields an implicit but equivalent definition, through the stabilizer equations $X_1Z_2\,|\phi_3\rangle = |\phi_3\rangle$, $Z_1X_2Z_3\,|\phi_3\rangle = |\phi_3\rangle$, $Z_2X_3\,|\phi_3\rangle = |\phi_3\rangle$. Again, these relations can be explicitly verified for the state on the r.h.s. of Eq. (\[phi3\]). Note that if we attach to $|\phi_3\rangle$ a fourth qubit in $|+\rangle_4$ and apply ${\rm cPhase}_{43}$ we obtain the linear four-qubit cluster state $|\phi_4\rangle$, and we can continue this procedure for any linear cluster state. The number of terms needed to describe the cluster state grows exponentially in the number of qubits. Indeed, it doubles upon adding two more qubits into the chain.
From the stabilizer relations follows, for example, that if qubits 1 and 3 are measured in the $Z$-basis and qubit 2 is measured in the $X$-basis, then the measured eigenvalues $\lambda_Z^{(1)},
\lambda_X^{(2)}, \lambda_Z^{(3)} \in \{1,-1\}$ are individually random but correlated, with $\lambda_Z^{(1)}
\lambda_X^{(2)}\lambda_Z^{(3)} =1$. In MBQC, output bits of the computation will be inferred from correlations like this one, but generally in more complicated bases. We note that $|\phi_3\rangle$ is locally equivalent to the so-called Greenberger-Horne-Zeilinger (GHZ) state, $|\mbox{GHZ}\rangle = (|000\rangle +
|111\rangle)/\sqrt{2}$. So-called one-qubit Hadamard gates have the property that $H|0\rangle = |+\rangle$ and $H|1\rangle = |-\rangle$, such that $H_1H_3|\mbox{GHZ}\rangle=|\phi_3\rangle$. We shall return to the $|\phi_3\rangle$/GHZ-example in Section \[QMF\], where we establish a connection between MBQC and the GHZ-version [@GHZ] of Bell’s theorem.
Basics of quantum gates by teleportation {#sec:1qubit}
----------------------------------------
In preparation for our demonstrating the universality of MBQC, we review a few techniques of performing gates by quantum teleportation. We follow the discussion of [@CLN]. Consider a quantum circuit $$\label{LocR1}
\parbox{4cm}{\includegraphics[width=4cm]{UniRotExpl1.pdf}}$$ which takes a two-qubit state $|\mbox{in}\rangle_1 \otimes
|+\rangle_2$ as input. Therein, the first qubit is free to choose, $|\mbox{in}\rangle = a|0\rangle + b |1\rangle$, and the second qubit is always fixed. A conditional phase gate is applied to these qubits and, subsequently, the first qubit is measured in the eigenbasis of $$\label{Obs1}
O(\varphi_1) = \cos \varphi_1 \, X_1 + \sin \varphi_1 \, 2,\;\; -\frac{\pi}{2} < \varphi \leq \frac{\pi}{2}.$$ The measured eigenvalues are $\pm 1=(-1)^{s_1}$, $s_1\in \{0,1\}$, and the corresponding eigenstates are $|\phi_1,\pm\rangle= (|0\rangle \pm e^{i\phi_1}|1\rangle)/\sqrt{2}$. If one obtains the measurement outcome $s_1=0$, then the second qubit is projected to a state ${}_1\langle\phi_1,+|\cdot |\psi\rangle_{12}\sim a|+\rangle_2
+ b e^{-i\phi_1} |+\rangle_2 = e^{-i\phi_1/2}(a\, e^{i\phi_1/2}
|+\rangle_2 + b\, e^{-i\phi_1/2} |-\rangle_2 )\sim H e^{i\phi_1 Z/2}
(a|0\rangle + b|1\rangle)$, up to an overall phase $e^{-i\phi_1/2}$, where $H=(X+Z)/\sqrt{2}$ is the Hadamard gate. In the computational basis $\{|0\rangle,|1\rangle\}$, the Hadamard gate takes the matrix form $$H=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}1 & 1 \cr
1 & -1\end{array}\right).$$ If one obtains the measurement outcome $s_1=1$, then the second qubit is projected to a state $\langle\phi_1,-|\cdot |\psi\rangle_{12}\sim a|+\rangle - b
e^{-i\phi_1} |+\rangle = e^{-i\phi_1/2}(a e^{i\phi_1/2} |+\rangle -
b e^{-i\phi_1/2} |-\rangle) \sim H e^{i\phi_1 Z/2} Z(a|0\rangle +
b|1\rangle)$, up to an overall phase $e^{-i\phi_1/2}$. The two outcomes can be summarized in one equation $$|{\rm out}\rangle = H e^{i\phi Z/2} Z^s |{\rm in}\rangle.$$ The outcome of the circuit Eq. (\[LocR1\]) is that an ‘in’ state which was initially residing on qubit 1 has been re-located to qubit 2, and been acted upon by a unitary gate $H e^{i\phi Z/2} Z^s$ in the passing.
We may now feed the state $|\mbox{out}\rangle$ into another circuit of type Eq. (\[LocR1\]). The new output state $|\psi_3\rangle$, located on a third qubit, will be related to the initial state ‘in’ on the first qubit by $$\label{XZr}
\begin{array}{rcl}
|\psi^{(3)}\rangle &=& H e^{i\phi_2 Z/2} Z^{s_2}
|\mbox{out}\rangle=\big(H e^{i\phi_2 Z/2} Z^{s_2}\big) \big(H
e^{i\phi_1 Z/2} Z^{s_1}\big) |\mbox{in}\rangle\\
& = & \big(Z^{s_2}X^{s_1}e^{i(-1)^{s_1}\phi_2 X/2}
e^{i\phi_1 Z/2} \big) |\mbox{in}\rangle.
\end{array}$$ We find that in the above iterated circuit we can implement rotations about both the $X$-and the $Z$-axis. Two more aspects are worth of note. First, the rotation angle $\phi_x$ of the $X$-rotation depends on the measurement outcome $s_1$ implementing the preceding $Z$-rotation, $\phi_x=(-1)^{s_1}\phi_2$. Therefore, in order to realize a rotation about a given angle $\phi_x$, the measurement angle $\phi_2$—specifying the measurement basis for qubit 2—must be adjusted according to them measurement outcome $s_1$ obtained from qubit 1. Second, we obtain the desired rotation $e^{i(-1)^{s_1}\phi_x X/2} e^{i\phi_1 Z/2}$ only up to a random Pauli operator $Z^{s_2}X^{s_1}$. Such operators are called ‘byproduct operators’ in MBQC. Since they are known from the measurement outcomes, in the above circuit they can be undone by active intervention. Alternatively, they may be propagated forward through the circuit, like $Z^{s_1}$ in the above example, flipping rotation angles and, potentially, readout measurements in a controlled and correctable fashion.
![\[LocRot\]Circuit diagrams for MBQC-Simulation of a general one-qubit unitary. The horizontal lines represent the time direction of qubits. ${\rm cPhase}$ gates are between two qubits is indicated by the solid vertical line joining the corresponding two solid dots. Solid boxes denote measurements, such as the observables shown in Eq. (\[Obs\]). Lhs: the circuit Eq. (\[LocR1\]) iterated four times simulates general one-qubit unitaries. Rhs: The procedure MBQC-GS for a 5-qubt cluster state. By moving conditional phase gates past commuting measurements, the circuits on lhs and rhs are shown to be equivalent. Thus, the procedure MQC-GS on a one-dimensional cluster of 5 qubits simulates a general one-qubit rotation.](UniRotExpl2){width="14cm"}
Can we build a general one qubit unitary by concatenating the circuit of Eq. (\[LocR1\])? This is indeed possible. As shown in Fig. \[LocRot\], we concatenate the circuit Eq. (\[LocR1\]) four times, whereby a fifth qubit is transformed into the state $|\psi^{(5)}\rangle=U|{\rm in}\rangle$, with the unitary gate $U$ given by [@CLN] $$\label{U1}
U(\{\phi,s\})=\big(H e^{i\phi_4 Z/2} Z^{s_4}\big) \big(H e^{i\phi_3
Z/2} Z^{s_3}\big) \big(H e^{i\phi_2 Z/2} Z^{s_2}\big) \big(H
e^{i\phi_1 Z/2} Z^{s_1}\big).$$ We set $\phi_1=0$. Then, by reordering operations in the same way as in Eq. (\[XZr\]), can rewrite the resulting unitary as $$\label{RotSim}
U(\{\phi_2,\phi_3,\phi_4,s\})
= Z^{s_1+s_3} X^{s_2+s_4} \exp\left(i(-1)^{s_1+s_3}
\frac{\varphi_4}{2}X\right) \exp\left(i(-1)^{s_2}\frac{\varphi_3}{2}Z\right)
\exp\left(i(-1)^{s_1}\frac{\varphi_2}{2}X\right).$$ Therein, we have used the identities $HZ=XH$ and $XZ=-ZX$, and dropped an overall phase factor. Thus, up to byproduct operator $U_\Sigma \equiv Z^{s_1+s_3} X^{s_2+s_4}$, a general one-qubit rotation $$U_{\rm rot}=
\exp\left(-i\frac{\zeta}{2}X_i\right)\exp\left(-i\frac{\eta}{2}Z_i\right)\exp\left(-i\frac{\xi}{2}X_i\right),$$ with Euler angles $\zeta$, $\eta$, $\xi$ can be realized by the choosing the measurement angles $$\label{infl}
\begin{array}{rcl}
\varphi_1 &=& 0,\\
\varphi_2 &=& -(-1)^{s_1}\xi,\\
\varphi_3 &=& -(-1)^{s_2}\eta,\\
\varphi_4 &=& -(-1)^{s_1+s_3}\zeta.
\end{array}$$ We find that measurement angles, and thus measurement bases, depend on measurement outcomes of other qubits. This is the origin of temporal order in measurement-based quantum computation.
The reader will have noted that we could have accomplished the same task of implementing a general one-qubit unitary by concatenating the circuit Eq. (\[LocR1\]) only three times rather than four, and not setting the first measurement angle to zero. However, then the Hadamard gates in the counterpart of Eq. (\[RotSim\]) would not cancel. We would still obtain a general one-qubit unitary, albeit not in Euler normal form.
MBQC is universal {#sec:MBQCuniversal}
-----------------
To prove universality of MBQC, we need to show that (i) A universal set of gates can be simulated, (ii) Gate simulations compose in the same way as the gates themselves, and (iii) Cluster qubits not required in a particular computation can be removed.
### Simulating a universal set of gates
We need to be able to simulate a so-called universal set of gates. Such gate sets have the property that [*[any]{}*]{} unitary transformation on an $n$-qubit Hilbert space, for any $n \in
\mathbb{N}$, can be arbitrarily closely approximated by gates from the set.
A standard universal gate set consists of all one-qubit rotations for each qubit and the controlled Not (CNOT) gate on any pair of qubits [@NC]. These gates are defined as $$\label{UGS}
\begin{array}{rcl}
U_i &=& \exp\left(-i\frac{\zeta}{2}X_i\right)\exp\left(-i\frac{\eta}{2}Z_i\right)\exp\left(-i\frac{\xi}{2}X_i\right),\\
\CNOT_{c,t} &= &|0\rangle_c\langle 0|\otimes I_t + |1\rangle_c\langle 1|\otimes X_t.\end{array}$$ Therein, the subscripts $i$, $c$ (control), $t$ (target) are qubit labels, and $\zeta$, $\eta$ and $\xi$ are the Euler angles specifying the one-qubit rotation $U \in SU(2)$. $X$, $Z$ are Pauli operators ($X \equiv \sigma_x,\, Y\equiv \sigma_y,\, Z\equiv
\sigma_z$). Note that in the above gate set, only the CNOT gate has the power to entangle. It is equivalent, up to local unitaries, to the cPhase gate introduced in Eq. (\[CScreate\]).
An MBQC can be split up into MBQC gate simulations. Each gate simulation is like a LEGO piece, with example patterns shown in Fig. \[UniGates\], which were explained in the previous section. Be ${\cal{C}}$ a set of qubits, with $I \subset {\cal{C}}$ a set of input qubits, $O \subset {\cal{C}}$ a set of output qubits, and ${\cal{C}}\backslash O$ the set of qubits which are in ${\cal C}$ but not in $O$. Then, an MBQC gate simulation on ${\cal{C}}$ is the following
**Procedure** MBQC-GS.
1. [Create a cluster state with input $|\text{in}\rangle$, $|\phi(\text{in})\rangle = \left(\prod_{(a,b) \in E}\mbox{cPhase}_{a,b}\right) |\text{in}\rangle_I \bigotimes_{c \in {\cal{C}}\backslash I}|+\rangle_c$.]{}
2. [Measure all qubits $a \in {\cal{C}}\backslash O$, keep the state $|\text{out}\rangle_O$ of the unmeasured qubits in $O$. ]{}
The transformation $|\text{in}\rangle \longrightarrow
|\text{out}\rangle$ is unitary if suitable local measurement bases and sets $I,O$ are chosen. Specifically, cluster qubits which are not measured in the eigenbasis of $Z$ are measured in a basis in the equator of the Bloch sphere. The measured observable on such a qubit $a \in {\cal{L}}_2$ is $$\label{Obs}
O_a(\varphi_a) = \cos \varphi_a\, X_a + \sin \varphi_a\, Y_a,\;\; -\frac{\pi}{2} < \varphi_a \leq \frac{\pi}{2}.$$ The angle $\varphi_a$ specifying the measured observable $O_a$ is
called the ‘measurement angle’ for qubit $a$.
![\[UniGates\] Measurement patterns for a universal set of gates. (a,b) CNOT, (c) general one-qubit rotation. The CNOT gate in (a) has the control input and output located on the same cluster qubit. The extended CNOT in (b) has separate locations for input and output qubits, for both control and target.](UniGates2){width="14cm"}
#### One-qubit rotations.
We return to the procedure of performing a general one-qubit unitary by iterating the circuit Eq. (\[LocR1\]) four times, c.f. Section \[sec:1qubit\]. By moving cPhase gates backwards in time past local measurements they commute with, we can rewrite this circuit as preparation of a cluster state with one qubit of input, followed by local measurements of four cluster qubits including the input. The necessary re-ordering is displayed in Fig. \[LocRot\]. The circuit on the r.h.s. of Fig. \[LocRot\] precisely matches the procedure MQC-GS for gate simulations, which completes the construction.
#### CNOT-gate.
A cluster state of four qubits allows for a simulation of a controlled-NOT gate. Consider the graph shown in Fig. \[UniGates\]a. Let qubits 1 and 2 be in the states $(a|0\rangle_1+b|1\rangle_1)$ and $(c|0\rangle_2+d|1\rangle_2)$, respectively, and qubits 3 and 4 in the state $|+\rangle$. Now, a control-phase gate is applied between pairs $(1,3)$, $(2,3)$, and $(3,4)$. The joint state becomes $$\begin{aligned}
|\psi\rangle_{1234}&=&{\rm cPhase}_{43} {\rm cPhase}_{13} {\rm
cPhase}_{23}
\Big[(a|0\rangle_1+b|1\rangle_1)(c|0\rangle_2+d|1\rangle_2)|+\rangle_3|+\rangle_4\Big], \nonumber\\
&=& \Big[a|0\rangle_1
(c|0\rangle_2|+\rangle_3+d|1\rangle_2|-\rangle_3)+ b|1\rangle_1
(c|0\rangle_2|-\rangle_3+d|1\rangle_2|+\rangle_3)\Big]|0\rangle_4
\nonumber \nonumber \\
&& +\Big[a|0\rangle_1
(c|0\rangle_2|-\rangle_3+d|1\rangle_2|+\rangle_3)+ b|1\rangle_1
(c|0\rangle_2|+\rangle_3+d|1\rangle_2|-\rangle_3)\Big]|0\rangle_4,\nonumber\end{aligned}$$ where we have ignored overall normalization and have chosen a specific sequence of applying the commuting cPhase gates. Let us measure qubits 2 and 3 in the basis $|\pm\rangle$ and record the respective eigenvalues by $(-1)^{s_2}$ and $(-1)^{s_3}$ respectively. Suppose the measurement outcomes are $+1$, $+1$, i.e., $s_2=s_3=0$, then the post-measurement joint state of qubits 1 and 4 is $$\begin{aligned}
|\psi'\rangle_{14}&=&{}_{23}\langle++|\cdot |\psi\rangle_{1234}\nonumber\\
&=& a|0\rangle_1 (c|0\rangle_4 + d|1\rangle_4) + b|1\rangle_1
(c|1\rangle_4 + d|0\rangle_4)\nonumber\\
&=& {\rm CNOT}_{14} (a|0\rangle_1+b|1\rangle_1) (c|0\rangle_4 +
d|1\rangle_4) \nonumber.\end{aligned}$$ We see two effects: (i) there is a control-NOT gate applied on initial qubits 1 and 2 and then (ii) the information on qubit 2 has been transferred to qubit 4. Analyzing the three other cases $(s_2,s_3) \in \{(0,1),(1,0),(1,1)\}$, we conclude that the output state $|\psi_{\rm
out}\rangle$ on qubits 1 and 4 is related to the input state $|\psi_{\rm in}\rangle$ on qubits 1 and 2 via the transformation $$|\psi_{\rm out}\rangle={\rm CNOT}\, X_t^{s_3} Z_t^{s_2} |\psi_{\rm
in}\rangle =Z_c^{s_2} X_t^{s_3} Z_t^{s_2}{\rm CNOT} |\psi_{\rm
in}\rangle,$$ with qubit 1 acting as control qubit. We remark that qubits 2 and 3 are always measured in the $X$-eigenbasis, independent of all measurement outcomes on the cluster qubits. Therefore, the measurements on those qubits can be performed first, and the MBQC simulation of CNOT gates entirely drops out of the temporal order of measurements. The same holds for all gate simulations requiring only $X$ and $Y$-measurements, such as the simulations of Hadamard gates and rotations $\exp(i\pi/4\,Z)$.
The minimal configuration Fig. \[UniGates\]a for a CNOT realized on a four qubit cluster has the property that control input and output are located on the same cluster qubit. This may be a disadvantage for the composition of gate simulations. The minimal configuration can be expanded such that the locations for target input, control input, target output and control output are all separate; See Fig. \[UniGates\]b.
#### Composition of gate simulations.
It remains to be shown that MBQC gate simulations compose like the simulated gates themselves. This proceeds by a reordering-of-commuting-operations argument [@oneway; @oneway2],[@CLN]; also see Fig. \[LocRot\] for an example. Finally, note that the composition of gate simulations allow for a variable input that the standard cluster state does not provide. Starting with a cluster state amounts to fixing the initial state of the simulated quantum register in the state $\otimes_i |+\rangle_i$, which is the fiducial state appearing in DiVincenzo’s second criterion.
#### Removing redundant cluster qubits.
If a two-dimensional cluster state is used to perform measurement-based quantum computation by simulating the universal gates as pattern of measurement (see Fig. \[UniGates\]), these patterns may not cover all of the qubits on the two-dimensional grid. Cluster qubits such as those that are neither covered by the gate-simulation measurement patterns are not needed in a particular computation can be removed by measuring them in the $Z$-eigenbasis. Then, the remaining qubits are still in a cluster state, with the $Z$-measured qubits removed from the cluster. This follows directly from the creation procedure Eq. (\[CScreate\]) for cluster states, and the identities $|0\rangle_a\langle 0|\,
\mbox{cPhase}_{a,b}=|0\rangle_a\langle 0| \otimes I_b$, $|1\rangle_a\langle 1|\, \mbox{cPhase}_{a,b}=|1\rangle_a\langle 1|
\otimes Z_b$. Hence, if one measures any qubit on a cluster state in the $Z$-basis, obtaining an outcome $0$, the remaining qubits are exactly in a cluster state (as they are acted on by identity operators), with the measured qubit removed from the cluster. If the measurement outcome is $1$, then the resulting state is equivalent to cluster state, up to local unitaries $Z$ on qubits neighboring the measured one. This completes the proof that the MBQC can efficiently simulate the circuit model of quantum computation.
Ground states as computational resources {#GS}
========================================
For $n$-qubit quantum states distributed according to the uniform Haar measure, it has been shown that only a tiny fraction $<\exp(-n^2)$ of states can possibly be universal resources for MBQC [@GrossFlammiaEisert; @BremnerMoraWinter]. We will review this result for a different reason in Section \[TETU\]. Universal resource states thus seem very rare, but is the uniform Haar measure the right criterion to apply? Do resource states for measurement-based quantum computation occur naturally in physical systems, say as ground states of Hamiltonians with two-body interactions?
Let us first drop the requirement of computational universality, and ask the more modest question of whether ground states of suitably simple Hamiltonians can be used as resource states for MBQC at all. This will lead us to one-dimensional spin systems, and provide hints for identifying a computationally universal ground state in a second step.
Spin chains
-----------
In 1983, Haldane argued that the spin-$S$ Heisenberg antiferromagnetic chain has different behaviors depending on $S$ is integer or half-integer [@Haldane]. In particular, he predicted that when $S$ is an integer, the spin chain has a unique disordered ground state with a finite spectral gap. This picture was supported by a construction that Affleck, Kennedy, Lieb and Tasaki (AKLT) proposed in a spin-1 valence-bond model [@AKLT; @AKLT2]. The spin-1 AKLT model and the antiferromagnetic chain are later found to be in the so-called Haldane phase of the following bilinear-biquadratic model, $$\label{Haldane}
H=\sum_{i} \left[ \cos\theta(\vec{S}_i\cdot
\vec{S}_{i+1})+\sin\theta (\vec{S}_i\cdot \vec{S}_{i+1})^2 \right],$$ for $\theta \in (-\pi/4,\pi/4)$, where $\vec{S}$ denotes the spin operators for the spin-1 particle. The one-dimensional Affleck-Kennedy-Lieb-Tasaki (AKLT) state is a special point in the Haldane phase with $\tan \theta = 1/3$. For periodic boundary conditions, the ground state in the Haldane phase is unique. For a linear chain, it is four-fold (near) degenerate, with the splitting in the degeneracy being exponentially small in the length of the chain. There, both edges carry an effective spin-1/2 particle. The resulting edge states turn out to be very important for our description of MBQC using ground states in the Haldane phase.
If the chain is terminated by a spin-1/2 particle at one end with the additional Hamiltonian term $\sim \vec{S}\cdot
\vec{s}$, then the degeneracy is reduced to 2, resulting from an effective spin-1/2 at the other end. The system thus carries total spin-1/2, i.e., $S_{\rm tot}=1/2$ with the effective two levels being $|G_0\rangle\equiv \Big| S_{\rm tot}=1/2, S_{\rm
tot}^z=1/2\Big\rangle$ and $|G_0\rangle\equiv \Big| S_{\rm tot}=1/2,
S_{\rm tot}^z=-1/2\Big\rangle$. The degenerate ground space can be used to encode the information of a qubit: $|\Psi\rangle=a_0|G_0\rangle + a_1|G_1\rangle$.
As a first result demonstrating the usefulness of Haldane ground states for MBQC, it has been shown that the AKLT state allows to simulate arbitrary single-qubit unitary gates by single-spin measurements [@Gross; @Gross2; @BrennenMiyake]; also see [@Chen10]. This is not sufficient for universal quantum computation, but it provides a first connection between MBQC and spin systems which have been studied in condensed matter physics for completely different reasons.
Furthermore, if we slightly extend our computational model to comprise of two primitives, namely (i) Measurement of individual spins (as before) and (ii) adiabatic turn-off of individual spin-spin couplings in the Hamiltonian Eq. (\[Haldane\]), then a more general result can be obtained: The usefulness of the ground state of Eq. (\[Haldane\]) as computational resource extends to the entire Haldane phase surrounding the AKLT point [@Miyake]! This is interesting in several ways. For example, away from the AKLT point, the ground state of the spin chain is not explicitly known. Quantum computation by local measurement on those states works perfectly nonetheless. Furthermore, whereas quantum computation is usually considered ‘high maintenance’, i.e. any imperfect control of the system Hamiltonian quickly leads the computation off track, here we observe a feature of robustness: $\theta$ may vary largely without affecting the functioning of the computational scheme. These features are consequences of a symmetry-protected topological order [@Symm1; @Symm2] which characterizes the Haldane phase.
We now describe quantum computation by local measurements in the Haldane phase, following the original discussion by Miyake [@Miyake]. Let us denote by $|G_0(j)\rangle$ and $|G_1(j)\rangle$ the two degenerate ground states for a system of spins from $j$ to $N$, and by $|\Psi(j)\rangle$ a qubit state encoded in the ground state space, $|\Psi(j)\rangle =
a_0|G_0(j)\rangle+ a_1|G_1(j)\rangle$. Suppose we start with a chain of $N$ spins in the state $|\Psi(j=1)\rangle$, for known $a_0$, $a_1$. We can ask how this state is transformed when we turn off adiabatically the coupling between the first and the second spin. The initial value of the total spin is 1/2. After turning off the coupling, the subsystem of spins $(2,..,N)$ by itself is effectively a spin-1/2, residing in the ground space spanned by $|G_0(j=2)\rangle$ and $|G_1(j=2)\rangle$. Composing it with the spin 1 at site 1, by angular moment addition $1/2\otimes 1=1/2
\oplus 3/2$, the final Hamiltonian has ground states with spin 1/2 and 3/2. However, while the coupling is being turned off, rotational symmetry is maintained and the total spin conserved. The final state is thus confined to the spin 1/2 sector. Using the Clebsch-Gordan decomposition, one finds $$|\Psi(1)\rangle \rightarrow |\Psi'(1)\rangle = \sqrt{\frac{2}{3}} \left[\big(\frac{a_0}{\sqrt{2}}|0\rangle_{1} + a_1|-1\rangle_{1}\big) \otimes|G_0(2)\rangle -\big(a_0|+1\rangle_{1} + \frac{a_1}{\sqrt{2}}|0\rangle_{1}\big) \otimes|G_1(2)\rangle\right].$$ After the adiabatic turning-off of the coupling, one can measure the first spin in any orthonormal basis spanned by $|\pm 1\rangle$ and $|0\rangle$. For example, we consider the basis $\big\{|x\rangle\equiv(|-1\rangle -|1\rangle)/\sqrt{2},
|y\rangle\equiv(|-1\rangle +|1\rangle)/\sqrt{2},
|z\rangle\equiv|0\rangle\big\}$. After measuring spin 1 in the state $|\beta\rangle$, the post-measurement state of the spins $(2,..,N)$ is ${}_1\langle \beta|\cdot |\Psi'(1)\rangle$. This state behaves as if a quantum gate had acted on the system $(2,..,N)$ with the initial state $|\Psi(j=2)\rangle=a_0|G_0(2)\rangle+a_1|G_1(2)\rangle$. Depending on the measurement outcome ($|\beta =|x\rangle, |y\rangle,
|z\rangle$, respectively), the resulting state is $$X |\Psi(2)\rangle, \ XZ|\Psi(2)\rangle, \ Z|\Psi(2)\rangle.$$ Therein, $X$ and $Z$ are the effective Pauli X and Z operators, with $X=|G_0\rangle\langle G_1|+ |G_1\rangle\langle G_0|$ and $Z=|G_0\rangle\langle G_0|- |G_1\rangle\langle G_1|$. One may proceed to adiabatically turn off the subsequent couplings one by one and measure the decoupled qubits in the basis $\{|x\rangle,
|y\rangle, |z\rangle\}$. Up to the above Pauli rotations, this results in a quantum wire in which a logical qubit is propagated forward along the spin chain.
This process is easily generalized to induce an arbitrary rotation on the effective qubit state. For example, the basis $$\label{MB}
{\cal{B}}(\alpha) = \Big\{\frac{(1\pm
e^{-i\alpha})}{2}|x\rangle_j +\frac{(1\mp
e^{-i\alpha})}{2}|y\rangle_j, |z\rangle_j\Big\}$$ gives rise to a rotation about $z$-axis $R^z(\alpha)=|G_0\rangle\langle G_0| +e
^{i\alpha} |G_1\rangle\langle G_1|$, up to possible Pauli corrections. Measurement in another basis can induce rotation about other axes, such as $x$-axis. Therefore, arbitrary single-qubit unitary evolution can be simulated in a chain residing in the Haldane phase.
Let us briefly comment on the role of symmetries which protect the Haldane phase [@Symm1; @Symm2] and computation on its edge states. It has been shown that the Haldane phase is protected even in the presence of perturbations, as long as they possess certain symmetries such as time reversal [@Symm2], $S_k^{x,y,z}
\longrightarrow - S_k^{x,y,z}$. Now, the Hamiltonian Eq. (\[Haldane\]), even with individual couplings turned off one by one, has this symmetry. The measured observables with eigenbases Eq. (\[MB\]) possess it as well, such that the Haldane phase remains protected throughout the course of computation.
It needs to be pointed out that in the above construction, the adiabatic switching-off of couplings is not merely a means to extract individual spins from a ‘ground-state memory’. Away from the AKLT point, turning off a coupling does real work for the computational scheme by modifying the correlation with the edge states. Thus, the question arises of whether the same computational power can be obtained without the adiabatic part. I.e., are ground states of spin chains in the Haldane phase resources for [*[1-qubit]{}*]{} MBQC?
The answer is again affirmative. As shown by Bartlett and collaborators [@BartlettBrennen], local measurements on a Haldane ground state can be used to mimic a renormalization group transformation. The resulting state is again in the Haldane phase, with the length of the spin chain cut by a factor of three. The AKLT state is a fixed point of this transformation to which the whole Haldane phase is attracted. Thus, to simulate a one-qubit universal MBQC on a ground state in the Haldane phase, a first set of local measurements is used to bring the initial state as close as needed to an AKLT state, and the remaining measurements simulate the unitary gate.
Universal MBQC with AKLT states in two dimensions {#AKLT2d}
-------------------------------------------------
In the previous section we have identified an entire phase of ground states which can serve as resources for restricted measurement-based quantum computations. Beautiful and unexpected connections between measurement-based quantum computation and condensed matter physics have been found as a bonus. But a central question is so far unanswered: Are there ground states of two-body Hamiltonians which are [*[universal]{}*]{} resources for MBQC?
The answer again is ‘yes’. This was first established for spin 5/2 particles on a honeycomb lattice [@Chen], with a suitably tailored Hamiltonian. This result is important, because it had previously been proven that cluster states, the standard resource for universal MBQC, cannot arise as the ground state of a Hamiltonian with only two-body interactions [@Nielsen]. They can nonetheless be closely approximated by ground states of such Hamiltonians [@RudoBart].
What remains to be explored is whether Hamiltonians with universal resources for MBQC as ground states can look simpler, more natural. In this regard, it was first shown that the unique ground state of an AKLT-like Hamiltonian for spins 3/2 on a two-dimensional lattice Hamiltonian yields a universal resource for MBQC [@QMagn]. Here, AKLT-like means that within the two-dimensional lattice, 1D quasi-chains are coupled via the AKLT Hamiltonian, and the coupling between the chains is of a different type, but still two-body.
Finally, it has been shown by Miyake [@HoneycombPI] and by Wei, Affleck and Raussendorf [@HoneycombUBC] that the AKLT state on the honeycomb lattice is a universal computational resource. To briefly review this result, let us first recall the definition of the AKLT state on a honeycomb lattice.
The AKLT state [@AKLT] on the honeycomb lattice ${\cal{L}}$ has one spin-3/2 per site of ${\cal{L}}$. The state space of each spin 3/2 can be viewed as the symmetric subspace of three virtual spin-1/2’s, i.e., qubits. In terms of these virtual qubits, the AKLT state on ${\cal{L}}$ is $$\label{AKLT2}
|\Phi_{\rm AKLT}\rangle\equiv\bigotimes_{v \in V({\cal{L}})}P_{S,v}
\bigotimes_{e \in E({\cal{L}})} |\phi\rangle_e,$$ where $V({\cal L})$ and $E({\cal L})$ to denote the set of vertices and edges of ${\cal L}$, respectively. $P_{S,v}$ is the projection onto the symmetric (equivalently, spin 3/2) subspace at site $v$ of ${\cal{L}}$. For an edge $e=(v,w)$, $|\phi\rangle_{e}$ denotes a singlet state, with one spin 1/2 at vertex $v$ and the other at $w$.
The first step in the process of MBQC with the AKLT state is to apply a suitable generalized measurement [@NC], also called positive-operator-value measure (POVM), locally on every site $v$ on the honeycomb lattice ${\cal{L}}$; See Fig. \[AKLT\]. Specifically, the POVM consists of three rank-two elements
\[POVM2\] $$\begin{aligned}
{F}_{v,z}&=&\sqrt{\frac{2}{3}}
\left(\left|\frac{3}{2},\frac{3}{2},z\right\rangle \left\langle \frac{3}{2},\frac{3}{2},z\right| + \left|\frac{3}{2},-\frac{3}{2},z\right\rangle \left\langle \frac{3}{2},-\frac{3}{2},z\right|\right) = \frac{1}{\sqrt{6}}\left(S_z^2-\frac{1}{4}\right),\\
{F}_{v,x}&=&\sqrt{\frac{2}{3}}\left(\left|\frac{3}{2},\frac{3}{2},x\right\rangle \left\langle \frac{3}{2},\frac{3}{2},x\right| + \left|\frac{3}{2},-\frac{3}{2},x\right\rangle \left\langle \frac{3}{2},-\frac{3}{2},x\right|\right)= \frac{1}{\sqrt{6}}\left(S_x^2-\frac{1}{4}\right),\\
{F}_{v,y}&=&\sqrt{\frac{2}{3}}\left(\left|\frac{3}{2},\frac{3}{2},y\right\rangle \left\langle \frac{3}{2},\frac{3}{2},y\right| + \left|\frac{3}{2},-\frac{3}{2},y\right\rangle \left\langle \frac{3}{2},-\frac{3}{2},y\right|\right)= \frac{1}{\sqrt{6}}\left(S_y^2-\frac{1}{4}\right),\end{aligned}$$
where $\textbf{r} =x,y,z$ in $|s,m_s,\textbf{r}\rangle$ specifies the quantization axis. The above POVM elements obey the relation $\sum_{\nu \in \{x,y,z\}}F^\dagger_{v,\nu} F_{v,\nu} = I_{S,v}$, which is the identity on the spin-3/2 Hilbert space as well as on the symmetric subspace of three qubits, as required. Physically, $F_{v,a_v}$ is proportional to a projector onto the two-dimensional subspace within the $S_a=\pm 3/2$ space, which is the origin of logical qubits in the present construction. The outcomes $x$, $y$ or $z$ of the POVM at the individual sites of ${\cal{L}}$ are random, but short-range correlated. After the results of the POVM at all sites, the post-POVM state becomes $\bigotimes_{v \in
V({\cal{L}})}F_{v,a_v}\, |\Phi_{AKLT}\rangle$, with $a_v = x,y,z$ denoting the POVM outcome at site $v$.
![\[AKLT\]AKLT states. (a) Spin singlets of two virtual spins 1/2 are located on the edges of the honeycomb lattice. $P_{S,v}$ is a projection at each lattice site $v$ onto the symmetric subspace. (b) The first step in MBQC with AKLT states is a local POVM at each lattice site, with three possible outcomes $x$, $y$, $z$. Edges with the same POVM outcome at the end points (\*) are treated differently from edges where the POVM outcomes differ (\*\*).](AKLT){width="10cm"}
After using the same initial POVM, the two arguments proceed differently. In [@HoneycombPI] a mapping to quantum circuits is pursued, whereas in [@HoneycombUBC] the AKLT state is mapped to a two-dimensional cluster state which is already known to be universal. Let us very briefly summarize the two arguments.
In [@HoneycombPI], it is noted that whenever the POVM outcomes on two neighboring sites differ, then the link in between can be used to implement, along the remaining two orthogonal links, either an elementary computational wire or an entangling gate between two wires. Such links are sufficiently frequent to form a giant connected component [@GT], from which a ‘backbone’ is chiseled out by further local measurements. The backbone is a net composed of computational wires and bridges between them, similar to the one displayed in Fig. \[flow\]a.
The links between neighboring sites with the same POVM outcome cannot be used for entangling gates between wires in the backbone, and represent a (manageable) complication. A connected set of such links ranging from one wire to another compromises those wires, and therefore has to be avoided. Fortunately, identical POVM outcomes on the opposite ends of a link occur only with a probability of approximately 1/3, which is sufficiently infrequent for connected sets of such links to remain microscopic. They can then be dealt with by choosing a sufficiently large-meshed backbone, on which a general quantum circuit may be implemented.
In [@HoneycombUBC], proof of computational universality proceeds by reduction to a 2D cluster state, via an intermediate step. Namely, it is first shown that applying the local POVMs to the AKLT state results in a graph state. The corresponding graph depends on the random POVM outcomes but is always planar. It is obtained from the graph representing the honeycomb lattice ${\cal{L}}$ using the following rules: (i) If the POVM outcomes on two neighboring vertices agree, the edge in between is contracted, and (ii) in the resulting multigraph, edges of even multiplicity are deleted and edges of odd multiplicity are replaced by standard edges.
Then, typical graphs resulting from this procedure are shown by numerical simulation to reside in the supercritical phase of percolation, i.e., they have traversing paths. Such graph states can then be used to implement universal quantum computation by local measurements, which is demonstrated by reduction to 2D cluster states.
The role of entanglement {#Ent}
========================
Hardwired into the very foundations of quantum information science is the assumption that local quantum operations and classical communication (LOCC) are easily accomplished whereas non-local operations, i.e., quantum-mechanical interactions between different parts of a quantum system, are difficult. Consequently, a fundamental distinction is made between these two classes of operations. From this perspective, entanglement [@Entanglement; @Entanglement2] is a quintessential property of quantum systems. It measures the degree to which quantum states require non-local operation for their creation or to which they can enable non-local operation. It is also a key resource for many protocols of quantum information processing, such as teleportation, quantum cryptography, and quantum error-correction. The defining property of entanglement monotones [@VEM], which measure the ‘amount’ of entanglement contained in quantum states, is that they do not increase under LOCC.
Since MBQC is driven entirely by operations in the LOCC class, entanglement decreases as the computation proceeds. This provides our intuition that entanglement is a key resource for MBQC. A closer examination shows that, for quantum computation with pure resource states, significant entanglement is indeed necessary to achieve a quantum speedup, but more is not necessarily better.
Entanglement and quantum speedup {#ent}
--------------------------------
In this section, we demonstrate that any MBQC with a pure state that only contains a small amount of entanglement can be efficiently classically simulated, preventing a significant speedup (See [@Vidal03] for an analogous result in the circuit model). To do so, we must first overcome an obstacle.
Consider cluster state $|\phi_{\cal{C}}\rangle$ on a one-dimensional cluster ${\cal{C}}$. As illustrated by the example of simulating a general one-qubit rotation, MBQC on one-dimensional cluster states maps to the circuit model with a single qubit. It can thus be efficiently simulated classically. And yet, $|\phi_{\cal{C}}\rangle$ is highly entangled. For suitable bi-partitions ${\cal{C}}=A \cup B$ (e.g. odd vs. even-numbered qubits), the von-Neumann entropy $E_{A:B}(|\phi\rangle):=S(\rho_A)=-\Tr(\rho_A\log
\rho_A)=S(\rho_B)$, which is a valid entanglement measure for pure states [@Bennett], takes the large value of $E_{A:B}(|\phi_{\cal{C}}\rangle)=\lfloor |{\cal{C}}|/2 \rfloor$. To rescue the asserted connection between entanglement and speedup, we need to look for a different entanglement measure.
To this end, for general pure states $|\psi\rangle$ on a set $V$ of qubits, consider a subcubic tree $T$ (a tree graph with vertices of degrees between 1 and 3) whose leaves (vertices of degree 1) are associated with the qubits in $V$. For any edge $e$ of $T$, $T\backslash e$ consists of two components, inducing a bi-partition of $V$ into two sets $A_T^e$ and $B_T^e$. It can be shown [@universal; @universal2] that the quantity $$\label{Ewidth}
E_{\rm wd}(|\psi\rangle)\equiv \min_T \max_{e\in T} E_{A_T^e,B_T^e}(|\psi\rangle).$$ is an entanglement monotone. It is called ‘entanglement width’.
Returning to our above example of the one-dimensional cluster state $|\phi_{\cal{C}}\rangle$, it turns out that $E_{\rm
wd}(|\phi_{\cal{C}}\rangle)=1$. To see this, note that for the tree $\tau$ displayed in Fig. \[C6\]b, the von Neumann entropy with respect to the bi-partition $A_\tau^e:B_\tau^e$ is $E_{A_\tau^e,B_\tau^e}(|\phi_{\cal{C}}\rangle) =1$, for any $e \in
E(\tau)$. Thus, the entanglement width does at least remove the above counterexample towards establishing a connection between entanglement and hardness of classical simulation in MBQC. But is it of more general use? Can this entanglement measure, at least for broad classes of interest, be efficiently calculated?
These questions both have affirmative answers. First, calculating the entanglement width is in general hard, due to minimization over all subcubic trees. However, if the state in question is a graph state, then a close upper bound can be obtained efficiently [@eSpeed], using graph theoretic techniques [@Oum]. Furthermore, the following general result [@eSpeed] establishes entanglement width as the critical complexity parameter for the classical simulation of MBQC on graph states,
\[EffSim\] Let $|G\rangle$ be a graph state on $n$ qubits. Then, MBQC on $|G\rangle$ can be classically simulated in $\mbox{poly}(n, 2^{E_{\rm wd}(|G\rangle)})$ time.
A similar theorem can be established for general $n$-qubit quantum states instead of graph states only, but it requires extra conditions relating to the efficient computability of the entanglement width [@eSpeed].
Theorem \[EffSim\] shows that a substantial amount of entanglement, as measured by the entanglement width, is [*[necessary]{}*]{} for a quantum speedup in MBQC with graph states. However, it is not sufficient. MBQC with so-called surface code states [@SCS] on a $k\times k$ lattice, which are local unitary equivalent to graph states, can be efficiently simulated classically in $\mbox{poly}(k)$ time, but their entanglement width is linear in $k$ [@BR06].
A related yet separate question is whether substantial entanglement is required for [*[universality]{}*]{} of MBQC. If we consider quantum computation as a universal state preparator, so-called $CQ$-universality, then the answer is affirmative. A family of resource states can only be $CQ$-universal if the entanglement in the belonging states is unbounded [@universal; @universal2]. In the preceding discussion of the relation between entanglement and speedup in MBQC, we assumed classical input and output (all qubits were required to be measured), so-called $CC$-universality.
A $CQ$-universal quantum computer presumably has more power than a $CC$-universal quantum computer. It was thus conjectured that $CC$-universal resource states exist which cannot be LOCC-converted to any $CQ$-universal state. Some examples for $CC$-universal states were proposed by Gross and Eisert in the new framework of measurement-based quantum computation with Projected-Entangled-Pair-States (PEPS) [@Gross; @Gross2; @Verstraete]. However, it was shown later by Cai et al. [@Cai] that these $CC$-universal states can be locally converted to cluster states, which are $CQ$-universal. Whether or not the notions of $CC$ and $CQ$-universality are equivalent remains an open question.
 Linear cluster state C6. (b) Bipartitions induced by a sub-cubic tree, and optimal tree $\tau$.](lc6){width="8cm"}
In the next section, we present a different facet of the relation between entanglement and MBQC which provides a counterpoint to the result just discussed.
Too entangled to be useful {#TETU}
--------------------------
It came as a surprise when Gross, Flammia and Eisert [@GrossFlammiaEisert], and independently, Bremner, Mora and Winter [@BremnerMoraWinter], showed that if a quantum state is [*[too]{}*]{} entangled then it becomes use[*[less]{}*]{} for MBQC. They measured the entanglement content in terms of the geometric entanglement (GE) [@WeiGoldbart03]. In [@GrossFlammiaEisert], it is shown that if an $n$-qubit state $|\Psi_n\rangle$ is the resource state for a measurement-based quantum computation that succeeds with high probability, and furthermore $E_G(|\Psi_n\rangle)> n -\delta$, where $\delta$ is a small constant, then this quantum computation can be efficiently classically simulated. No quantum speedup is provided, and $|\Psi_n\rangle$ is thus not useful as a resource state.
Let us recall the definition of geometric measure of entanglement [@WeiGoldbart03]. It is motivated by the mean-field approximation. The idea is to find, among the set of product states, $\{\ket{\Phi} = \ket{\phi^{[1]}}\otimes \ket{\phi^{[2]}} \otimes
\cdots \otimes \ket{\phi^{[n]}}\}$, the one closest to $\ket{\psi}$. This is achieved by maximizing their overlap, $\Lambda_{\max}({\psi})\equiv \max_{\Phi}|\ipr{\Phi}{\psi}|$. The GE of $\ket{\psi}$ is then defined as [@WEGM04] $$E_G(\psi)\equiv -\log_2\Lambda_{\max}({\psi})^2.$$ The maximum value of $E_G$ for an $n$-qubit state is $n$.
Now, to prove the above claim, consider MBQC on a resource state $|\Psi\rangle$ of $n$ qubits which succeeds with high probability, 1/2 say. Furthermore, assume that $|\Psi\rangle$ has close to maximal geometric entanglement, $E_G(|\Psi\rangle)>n-\delta$. Then, there are $2^n$ possible measurement outcomes, and we are interested in the fraction of good outcomes $G$ which cause the computation to succeed. The probability of each individual outcome is bounded by $|\langle \alpha|\Psi\rangle|^2 \le 2^{-E_G(\Psi)} < 2^{-n+\delta}$. The success probability $\sum_{\alpha \in G}p(|\alpha\rangle)$ is $\geq 1/2$ by assumption, and thus in turn the fraction of good outcomes is $|G|/2^n \ge 2^{-\delta-1}$. The above quantum computation can therefore be efficiently simulated by a classical computer selecting the measurement outcome at random. Since the fraction of good outcomes is large, with high probability after a few trials, a good outcome will be selected. Thus, the above MBQC using a (too) entangled resource state does not perform better than a classical computer.
The conclusion that ‘too much entanglement renders a MBQC resource state useless’ may depend on the entanglement measure chosen. How much so, is presently unknown. However, the GE is a lower bound on other entanglement measures [@WEGM04; @HMMOV06], such as the relative entropy of entanglement [@ER] and the logarithmic robustness of entanglement [@Robustness]. The above result holds for those measures as well.
The role of quantum correlations {#QMF}
================================
We have shown in Section \[UniPro\] that the computational output in MBQC bitwise consists of correlations among certain measurement outcomes, and that these correlations derive from the quantum correlations Eq. (\[CScorr\]) defining the cluster state. Surely, the correlations Eq. (\[CScorr\]) uniquely define a highly entangled quantum state. But is there another sense in which these correlations reveal their quantumness? Yes.—They can in general not be described by a local hidden variable model. Anders and Browne [@AB] have demonstrated a connection, at least in a specific example, between measurement-based quantum computation and Bell’s theorem [@Bell]. The correlations Eq. (\[CScorr\]) feature prominently in this correspondence.
Hidden variable models (HVM) were spurred by Einstein, Podolski and Rosen’s famous paper [@EPR] entitled “Can Quantum Mechanics be considered complete?”. With no additional assumptions made, such theories cannot be ruled out as valid descriptions of physical reality. Bohm’s wave mechanics [@Bohm] is a prominent example. However, if the hidden-variable model is required to be local, then—as Bell’s theorem [@Bell] shows—it cannot reproduce all predictions of quantum theory.
For our purpose of relating Bell’s theorem to measurement-based quantum computation, we will revisit Mermin’s proof [@Merm] of Bell’s theorem, using the Greenberger-Horne-Zeilinger (GHZ) states [@GHZ]. Consider the quantum state $ |GHZ\rangle =
\big(|000\rangle + |111\rangle\big)/{\sqrt{2}}$, which is local unitary equivalent to the three-qubit cluster state $|\phi_3\rangle$ (\[phi3\]), as $|GHZ\rangle=H_1H_3|\phi_3\rangle$. It is straightforward to check that $|GHZ\rangle$ is the simultaneous eigenstate with eigenvalue +1 of the four Pauli operators $$\label{GHZcorr}
X_1X_2X_3,\, - X_1Y_2Y_3, \, - Y_1X_2Y_3,\, - Y_1Y_2X_3.$$ Now, a local HVM would assign ‘pre-existing values’ $v(X_1)$, $v(X_2)$, $v(X_3)$, $v(Y_1)$, $v(Y_2)$, $v(Y_3)$ to the observables $X_1$, $X_2$, $X_3$, $Y_1$, $Y_2$, $Y_3$ which are merely revealed by measurement. Can those values be consistently assigned such that the predictions of quantum mechanics are reproduced?
Suppose this is the case. Then, $v(X_1),v(X_2), .. , v(Y_3) \in \{1,-1\}$. Furthermore, $$\label{GHZconstr}
\begin{array}{rcr}
v(X_1)\,v(X_2)\,v(X_3) &=& 1,\\
v(X_1)\,v(Y_2)\,v(Y_3) &=& -1,\\
v(Y_1)\,v(X_2)\,v(Y_3) &=& -1,\\
v(Y_1)\,v(Y_2)\,v(X_3) &=& -1.
\end{array}$$ To see why these constraints need to be enforced, consider the first one as an example. The four Pauli operators $X_1$, $X_2$, $X_3$ and $X_1X_2X_3$ obey the identity $X_1\cdot X_2 \cdot X_3 = X_1X_2X_3$. Furthermore, they mutually commute and hence can be simultaneously diagonalized. The above identity therefore also holds for their simultaneous eigenvalues, which, according to quantum mechanics, are the possible simultaneous measurement outcomes. Since the HVM is required to reproduce the predictions of quantum mechanics, the same relation must hold for $v(X_1)$, $v(X_2)$, $v(X_3)$ and $v(X_1X_2X_3)$. Finally, $v(X_1X_2X_3)=1$ by Eq. (\[GHZcorr\]).
But Eq. (\[GHZconstr\]) cannot be satisfied! Multiplying all four equations in (\[GHZconstr\]) we obtain $$v(X_1)^2\,v(Y_1)^2\, v(X_2)^2\,v(Y_2)^2\, v(X_3)^2\,v(Y_3)^2\, = -1,$$ Since $v(X_1)^2 = v(X_2)^2 = v(Y_3)^2 =1$, this is a contradiction. Hence, no consistent assignment of pre-existing values $v(X_1),\, ..
\, , v(Y_3)$ exists. The correlations Eq. (\[GHZcorr\]) of the GHZ state cannot be reproduced by a local HVM, and are genuinely quantum mechanical.
As it turns out, the very same correlations power a simple yet illuminating example of MBQC [@AB]. Consider the task of carrying out a single OR-gate via MBQC, using a GHZ-state as quantum resource and a classical control computer for the pre-processing of measurement bases and post-processing of measurement outcomes. This classical processing has an important constraint, namely that—as usual in MBQC—it can only involve addition mod 2. This kind of computation by itself is very limited.
The computation proceeds as follows. Denote the two input bits to the computation by $a$ and $b$, and the output bit by $o$ of the desired OR gate, $o = a \vee b$. The observable measured on qubit $i$ is $X_i$ if $q_i=0$, and $Y_i$ if $q_i=1$. The measurement bases are related to the input $a,b$ via $$q_1 = a,\, q_2= b,\, q_3 = a + b \mod 2.$$ The output bit $o$ is related to the measurement outcomes $s_i\in
\{0,1\}$ of the qubits $i$ via the computation by the classical computer which can only performs AND gates (i.e., binary addition), $$o = s_1 + s_2 + s_3 \mod 2.$$ Note that the relations specifying the classical pre and post-processing are all linear mod 2, as required.
We now discuss the functioning of the MBQC-OR computer input by input. For example, if $a=b=0$ then the measured observables are $X_1$, $X_2$ and $X_3$. Since $X_1X_2X_3|GHZ\rangle = |GHZ\rangle$ by Eq. (\[GHZcorr\]), $s_1+s_2+s_3 \mod 2 =0$ for this input. Thus, $o=0=0 \vee 0$ as required by the logical table.
As a second example, consider $a=0$ and $b=1$. Then, the measured observables are $X_1$, $Y_2$ and $Y_3$. By Eq. (\[GHZcorr\]), $X_1Y_2Y_3|GHZ\rangle = -|GHZ\rangle$, and therefore $o = s_1+ s_2 +
s_3 \mod 2 = 1=0 \vee 1$. The remaining two cases of inputs are analogous, and the logical table of the OR-gate, $o=a\vee b$, is established.
To put this result into perspective, it surely does not take a quantum computer to execute an OR-gate. The present example is therefore of no practical relevance. However, it makes a fundamental point. The classical control computer alone, which is only able to perform addition mod 2, has almost no computational power. In contrast, if mod 2-addition is supplemented by the capability of performing OR-gates, the resulting computational device becomes classically universal. Thus, a supply of GHZ states and the ability to measure them locally leads to a vast increase in the computational power.
What is more, the very same quantum correlations upon which Mermin’s proof of Bell’s theorem rests turn out to power the above measurement-based quantum computation. This result [@AB] hints at a link between MBQC and non-locality of quantum mechanics. How general this connection is remains to be explored.
Conclusion
==========
We have given an introduction to the one-way quantum computer, a scheme of universal quantum computation driven by local measurements on an entangled resource state. After a short explanation of how this scheme of computation works, we have described its underlying computational model, and identified universal resources among ground states of relatively simple Hamiltonians—such as the AKLT state on the honeycomb lattice. Further, we have discussed the roles of entanglement and quantum correlations for this computational model. It should be noted that our knowledge in either of these areas is very incomplete, and the research highlights presented here should be understood as base camps for further exploration.
We would like to end with three questions of varying degree of generality that seem particularly close to condensed matter physics: Is there, similar to the one-dimensional case, a Haldane-like phase around the AKLT state on the honeycomb lattice; and if so, does computational universality extend from the AKLT state to all regions of that phase? Can universal resource states be classified? Can a general theory of quantum correlations for measurement-based quantum computation be established?
[**Acknowledgment**]{}. The authors thank Maarten van den Nest, Dan Browne, Akimasa Miyake, Wolfgang D[ü]{}r, Hans Briegel, Ian Affleck and Dietrich Leibfried for discussions. This work is supported by NSERC, Cifar and the Sloan Foundation.
[99]{}
Shor PW. 1997. [*SIAM J. Sci. Statist. Comput.* ]{} 26:1484
Unruh WG. 1995. [*Phys. Rev. A*]{} 51:992-97
Shor PW. 1995. [*Phys. Rev. A*]{} 52:R2493-96
Steane AM. 1996. [*Proc. R. Soc. A*]{} 452:2551-77
Gottesman D. 1998. [*Phys. Rev. A*]{} 57:127-37
Aharonov D, Ben-Or M. 1999. arXiv:quant-ph/9906129
Aliferis P, Gottesman D, Preskill J. 2006. [*Quant. Inf. Comput.* ]{} 6:97-165
Sarma SD, Freedman M, Nayak C. 2006. [*Physics Today*]{} 7:32-38 (2006).
Nayak C, Simon SH, Stern A, Freedman M, Das Sarma S. 2008. [*Rev. Mod. Phys.*]{} 80:1083-159
Vandersypen LMK, Steffen M, Breyta G, Yannoni CS, Sherwood MH, Chuang IL. 2001. [*Nature*]{} 414:883-87
Greiner M, Mandel O, H[ä]{}nsch TW, Bloch I. 2002 [*Nature*]{} 419:51-54
Nelson KD, Li X, Weiss DS. 2007. [*Nature Phys.*]{} 3:556-60
Blatt R, Wineland, D. 2008. [*Nature*]{} 453:1008-15
Leibfried D, Knill E, Seidelin S, Britton J, Blakestad RB, et al. 2005. [*Nature*]{} 438:639-42
H[ä]{}ffner H, Hänsel W, Roos CF, Benhelm J, Chek-al-kar D, et al. [*Nature*]{} 438:643-46
Moehring DL, Maunz P, Olmschenk S, Younge KC, Matsukevich DN, et al. 2007. [*Nature*]{} 449:68-71
Walther P, Resch KJ, Rudolph T, Schenck E, Weinfurter H, et al. 2005. [*Nature*]{} 434:169-76
Niskanen AO, Harrabi K, Yoshihara F, Nakamura Y, Lloyd S, et al. 2007. [*Science*]{} 316:723-26
Harris R, Johnson MW, Lanting T, Berkley AJ, Johansson J, et al. 2010. [*Phys. Rev. B*]{} 82:024511
Hughes R, Doolen G, Awschalom DA, Chapman M, Clark R, et al. 2004. [A Quantum Information and Technology Roadmap, Part I: Quantum Computation]{} in [*Report of the Quantum Information Science and Technology Experts Panel to ARDA*]{}
Feynman RP. 1948. [*Rev. Mod. Phys.*]{} 20:367-87
Gottesman D, Chuang IL. 1999. [*Nature*]{} 402:390-93 (1999)
Raussendorf R, Briegel HJ. 2001. [*Phys. Rev. Lett.*]{} [86]{}:5188-91
Raussendorf R, Browne DE, Briegel HJ. 2003. [*Phys. Rev. A*]{} [68]{}:022312 Briegel HJ, Browne DE, Dür W, Raussendorf R, and Van den Nest M. 2009. [*Nature Phys.*]{} [5]{}:19-26
Briegel HJ, Raussendorf R. 2001. [*Phys. Rev. Lett.*]{} 86:910-13
Greiner M, Mandel O, Esslinger T, H[ä]{}nsch TW, Bloch I. 2002. [*Nature*]{} 415:39-44
Bakr WS, Peng A, Tai ME, Ma R, Simon J, et al. 2010. [*Science*]{} 329:547-50
Barrett SD, Kok P. 2005. [*Phys. Rev. A*]{} 71:060310(R)
Duan LM, Raussendorf R. 2005. [*Phys. Rev. Lett.*]{} 95:080503
Knill E, Laflamme R, Milburn GJ. 2001. [*Nature*]{} 409:46-52
Nielsen MA. 204. [*[Phys. Rev. Lett.]{}*]{} 93:040503
Browne DE, Rudolph T. 2005. [*Phys. Rev. Lett.*]{} 95:010501
Yoran N, Reznik B. 2003. [*Phys. Rev. Lett.*]{} 91:037903
Grover LK. 1996. A fast quantum mechanical Algorithm for database search. In [*Proc. 28 Annual ACM Symp. on the Theory of Computing*]{}:212
Gross D, Flammia ST, Eisert J. 2009. [*Phys. Rev. Lett.*]{} [102]{}:190501
Bremner MJ, Mora C, Winter A. 2009. [*Phys. Rev. Lett.*]{} 102:190502
Greenberger DM, Horne MA, Zeilinger A. 1989. In [*Bell’s Theorem, Quantum Theory, and Conceptions of the Universe*]{}, ed. Kafatos M. 69. Dordrecht: Kluwer
Nielsen MA, Chuang IL. 2000. [*Quantum Computation and Quantum Information*]{}. Cambridge: Cambridge University Press
Childs AM, Leung DW, Nielsen MA. 2005. [*Phys. Rev. A*]{} 71:032318
Raussendorf R, Briegel HJ. 2002. [*Quant. Inf. Comp.*]{} 2:443-86
Browne DE, Kashefi E, Mhalla M, Perdrix S. 2007. [*New J. Phys.*]{} 9:250
Haldane FDM. 1983. [*Phys. Lett. A*]{} 93:464-68; Haldane FDM. 1983. [*Phys. Rev. Lett.*]{} 50:1153-56
Affleck I, Kennedy T, Lieb EH, Tasaki H. 1987. [*Phys. Rev. Lett.*]{} [59]{}:799-802
Affleck I, Kennedy T, Lieb EH, Tasaki H. 1988. [*Comm. Math. Phys.*]{} [115]{}:477-528
Gross D, Eisert J. 2007. [*Phys. Rev. Lett.*]{} [98]{}:220503
Gross D, Eisert J, Schuch N, and Perez-Garcia D. 2007. [*Phys. Rev. A*]{} [76]{}:052315
Brennen GK, Miyake A. 2008. [*Phys. Rev. Lett.*]{} [101]{}:010502.
Chen X, Duan R, Ji Z, Zeng B. 2010. [*Phys. Rev. Lett.*]{} [105]{}:020502.
Miyake A. 2010. [*Phys. Rev. Lett.*]{} [105]{}:040501
Gu ZC, Wen XG. 2009. [*Phys. Rev. B*]{} 80:155131
Pollmann F, Turner AM, Berg E, Oshikawa M. 2010. [*Phys. Rev. B*]{} 81:064439
Bartlett SD, Brennen GK, Miyake A, Renes JM. 2010. [*Phys. Rev. Lett.*]{} [105]{}:110502
Chen X, Zeng B, Gu ZC, Yoshida B, Chuang IL. 2009. [*Phys. Rev. Lett.*]{} 102:220501
Nielsen MA. 2006. [*Rep. Math. Phys.* ]{} 57:147-61
Rodolph T, Bartlett SD. 2006. [*Phys. Rev. A* ]{} 74:040302(R)
Cai J, Miyake A, Dür W, Briegel HJ. 2010. [*Phys. Rev. A*]{} 82:052309
Miyake A. 2010. e-print arXiv:1009.3491
Wei TC, Affleck I, Raussendorf R. 2011. [*Phys. Rev. Lett.*]{} 106:070501
Durrett R. 2007. [*[Random Graph Dynamics]{}*]{} pp. 27 - 69. Cambridge/New York: Cambridge University Press
Plenio MB, Virmani S. 2007. [*Quantum Inf. Comput.*]{} [7]{}:1-51
Horodecki R, Horodecki P, Horodecki M, Horodecki K. 2009. [*Rev. Mod. Phys.*]{} [81]{}:865
Vidal G. 2000. [*J. Mod. Opt.*]{} 47: 355-76
Vidal G. 2003. [*Phys. Rev. Lett.*]{} [91]{}:147902.
Bennett CH, Bernstein HJ, Popescu S, Schumacher. 1996. [*Phys. Rev. A* ]{} 53:2046–52
Van den Nest M, Miyake A, Dür W, Briegel HJ. 2006. [*Phys. Rev. Lett.*]{} [97]{}:150504
Van den Nest M, Dür W, Miyake A, Briegel HJ. 2007. [*New J. of Phys.*]{} [9]{}:204
Van den Nest M, Dür W, Vidal, G, Briegel HJ. 2007. [*Phys. Rev. A.*]{} [75]{}: 012337
Oum SI. 2005. [*Lect. Notes Comput. Sci.*]{} 3787:49-58
Kitaev A. 2003. [*Ann. Phys. (N.Y.)*]{} 303:2-30
Bravyi S, Raussendorf R. 2007. [*Phys. Rev. A*]{} [76]{}:022304
Verstraete F, Cirac JI. 2004. [*Phys. Rev. A*]{} [70]{}:060302(R)
Cai JM, Dür W, Van den Nest M, Miyake A, Briegel HJ, [*Phys. Rev. Lett.*]{} 103:050503
We TC, Goldbart PM. 2003. [*Phys. Rev. A*]{} 68:042307
Wei TC, Ericsson M, Goldbart PM, Munro WJ. 2004. [*Quantum Inf. Comput.*]{} [4]{}: 252-72
Hayashi M, Markham D, Murao M, Owari M, Virmani S. 2006. [*Phys. Rev. Lett.*]{} [96]{}:040501
Vedral V, Plenio MB, Rippin MA, Knight PL. 1997. [*Phys. Rev. Lett.*]{} [78]{}: 2275-78
Vidal G, Tarrach R. 1999. [*Phys. Rev. A*]{} [59]{}:141-55
Anders J, Browne DE. 2009. [*Phys. Rev. Lett.*]{} 102:50502
Bell JS. 1964. [*Physics*]{} 1:195
Einstein A, Podolsky B, Rosen N. 1935. [*Phys. Rev.*]{} 47:777-80
Bohm D. 1952. [*Phys. Rev.*]{} 85:166-79 and 180-93
Mermin ND. 1993. [*Rev. Mod. Phys.*]{} 65:803-15
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We study a continuous time random walk $X$ in an environment of dynamic random conductances in $\mathbb{Z}^d$. We assume that the conductances are stationary ergodic, uniformly bounded and bounded away from zero and polynomially mixing in space and time. We prove a quenched invariance principle for $X$, and obtain Green’s functions bounds and a local limit theorem. We also discuss a connection to stochastic interface models. .2cm [*Keywords:*]{} Random conductance model, dynamic environment, invariance principle, ergodic, corrector, point of view of the particle, stochastic interface model.
.2cm [*Subject Classification: 60K37, 60F17, 82C41*]{}
author:
- 'Sebastian Andres [^1]'
title: '**Invariance Principle for the Random Conductance Model with dynamic bounded Conductances**'
---
Introduction {#sec:intro}
============
We consider the Euclidean lattice $\bZ^d$ equipped with the set $E_d$ of non oriented nearest neighbour bonds: $E_d=\{e=\{x,y\}: x,y\in\bZ^d, |x-y|=1\}$. We will also write $x\sim y$ when $\{x,y\} \in E_d$. Denote by $\hat\Omega=[0,\infty)^{E_d}$ and by $\Omega$ the set of all measurable functions from $\bR$ to $\hat \Omega$. We equip $\Omega$ with a $\sigma$-algebra $\mathcal{F}$ and a probability measure $\bP$ so that $(\Omega, \mathcal{F},\mathbb{P})$ becomes a probability space. The random environment is given by the coordinate maps $\mu_{e}^\om(t)=\om_e(t)$, $t\in \mathbb{R}, e\in E_d$. We will refer to $\mu_e(t)$ as the *conductance* of the edge $e$ at time $t$. Further, write $\mu^\om_{xy}(t)=\mu_{\{x,y\}}(t)=\mu_{yx}(t)$, and $\mu_{xy}(t)=0$ if $\{x, y\} \not\in E_d$, and set $$\begin{aligned}
\label{jumpP}
\mu_x(t) = \sum_{y \in\bZ^d} \mu_{xy}(t)=\sum_{y \sim x} \mu_{xy}(t). $$
We denote by $D(\bR,\bZ^d)$ the space of $\bZ^d$-valued càdlàg functions on $\bR$. For a given $\om\in\Omega$ and for $s\in \bR$ and $x\in\bZ^d$, let $P_{s,x}^\omega$ be the probability measure on $D(\bR,\bZ^d)$, under which the coordinate process $(X_t)_{t\in\bR}$ is the continuous-time Markov chain on $\bZ^d$ starting in $x$ at time $t=s$ with time-dependent generator given by: $$\begin{aligned}
\label{e-LV}
\sL^\om_t f(x) = \sum_{y\sim x} \mu^\om_{xy}(t) (f(y)-f(x)).\end{aligned}$$ That is, $X$ is the time-inhomogeneous random walk, whose time-dependent jump rates are given by the conductances. Note that the counting measure, independent of $t$, is an invariant measure for $X$. Further, we denote by $p^\om(s,x;t,y)$, $x,y\in \bZ^d$, $s\leq t$, the transition densities of the time-inhomogeneous random walk $X$. This model of a random walk in a random environment is known in the literature – at least in the case of time-independent conductances – as the [*Random Conductance Model*]{} or RCM. Note that the total jump rate out of any site $x$ is not normalized, in particular the sojourn time at site $x$ depends on $x$. Therefore, the random walk $X$ is sometimes called the [*variable speed random walk (VSRW)*]{}. However, for the purpose of this paper it would also be possible to consider the [*constant speed random walk (CSRW)*]{} with total jump rates normalized to one (cf. Remark \[rem:csrw\] below).
On $(\Omega, \mathcal{F},\mathbb{P})$ we define a $d+1$ parameter group of transformations $(\tau_{t,x})_{(t,x)\in \mathbb{R} \times \bZ^d}$ by $$\tau_{t,x}: \, \Omega\rightarrow\Omega \quad (\mu_e(s))_{s\in \mathbb{R}, e\in E_d}\mapsto (\mu_{x+e}(t+s))_{s\in \mathbb{R}, e\in E_d}$$ so that obviously $\tau_{s+t, x+y}=\tau_{s,x}\circ \tau_{t,y}$. Notice that $$\begin{aligned}
\label{eq:shiftedmu}
p^{\tau_{h,z}\om}(s,x;t,y)=p^\om(s+h,x+z;t+h, y+z), \qquad \mu_{xy}^{\tau_{h,z}\om}(t)=\mu^\om_{x+z,y+z}(t+h).\end{aligned}$$
We are interested in the $\bP$ almost sure or quenched long range behavior, in particular in obtaining a quenched functional limit theorem (QFCLT) or invariance principle for the process $X$ starting in $0$ at time $0$. To that aim we need to state some assumptions on the environment measure $\bP$.
\[ass:ergodic\] $\tau_{t,x}(A)\in \mathcal{F}$ for all $A\in \mathcal{F}$, and the measure $\mathbb{P}$ is invariant and ergodic w.r.t. $(\tau_{t,x})$, i.e. $\mathbb{P}[A]\in\{0,1\}$ for any event $A$ such that $\tau_{t,x}(A)=A$ for all $t\in \bR$ and $x\in \bZ^d$.
\[Stochastic Continuity\]\[ass:stoch\_cont\] For any $\delta>0$ and $f\in L^2(\bP)$ we have $$\lim_{h\to 0} \bP[ |f(\tau_{h,0}\om)-f(\om)|\geq \delta ]=0.$$
Thanks to Assumption \[ass:ergodic\] and \[ass:stoch\_cont\] the family of operators $(T_{t})_{t\in\bR}$ acting on $L^2(\bP)$, defined by $T_{t}f=f\circ \tau_{t,0}$, forms a strongly continuous group of unitary operators. Its $L^2(\bP)$-generator will be denoted by $D_t: \, \mathcal{D}(D_t)\rightarrow L^2(\bP)$, defined by $$D_tf(\om)=\frac{\partial}{\partial t} T_tf_{|t=0}(\om)=\frac{\partial}{\partial t}_{|t=0} f(\tau_{t,0}\om).$$ By Corollary 1.1.6 in [@EK] the generator is closed and densely defined. Note that $D_t$ is an anti-selfadjoint operator in $L^2(\bP)$, i.e.$$\langle D_tf,g\rangle_{\bP}=-\langle f, D_t g\rangle_{\bP}, \qquad f,g\in \mathcal{D}(D_t),$$ in particular $$\begin{aligned}
\label{eq:antisymD}
\langle D_tf,f\rangle_{\bP}=0, \qquad f\in \mathcal{D}(D_t).\end{aligned}$$
\[Ellipticity\] \[ass:ellipticity\] There exist positive constants $C_l$ and $C_u$ such that $$\begin{aligned}
\label{ass_elliptic}
\mathbb{P}\big[ C_l \leq \mu_e(t) \leq C_u, \,\forall e\in E_d, t\in\mathbb{R} \big]=1.\end{aligned}$$
We recall that under Assumption \[ass:ellipticity\] the following heat kernel estimates have been proven in [@DD] (see also [@GOS Appendix B] for similar bounds).
\[prop:hke\] There exist constants $c_1,\ldots, c_5$ such that for $\bP$-a.e. $\om$ and for every $t\geq s\geq 0$ the following holds:
1. If $x,y\in \bZ^d$ and $D=|x-y|\leq c_1 (t-s)$, then $$\begin{aligned}
p^\om(s,x;t,y)\leq \frac{c_2} {(t-s)^{d/2}} \exp(-c_3 D^2/(t-s))\quad \text{(Gaussian regime)}.\end{aligned}$$
2. If $x,y\in \bZ^d$ and $D=|x-y|\geq c_1 (t-s)$, then $$\begin{aligned}
p^\om(s,x;t,y)\leq \frac{c_4}{1\vee (t-s)^{d/2}} \exp(-c_5 D(1+ \log(D /(t-s))) \quad \text{(Poisson regime)}.\end{aligned}$$
Our first result is the following averaged or annealed FCLT. Let $\bP\otimes P_{s,x}^\om$ be the joint law of the environment and the random walk, and the annealed law is defined to be the marginal $\bP^*_{s,x}=\int_\Omega P_{s,x}^\om \, d\bP(\om)$. Further, let $$X^{(\eps)}_t = \eps X_{t/\eps^2}, \q t \ge 0.$$
\[annealed-ip\] Let $d\geq 1$ and suppose that Assumptions \[ass:ergodic\]-\[ass:ellipticity\] hold. Then, the law of $X^{(\eps)}$ converges under $\bP^*_{0,0}$ to the law of a Brownian motion on $\bR^d$ with a deterministic non-degenerate covariance matrix $\Sigma$.
To prove a QFCLT we will need some mixing assumptions on the environment. We denote by $B(\Omega)$ the set of bounded and measurable functions on $\Omega$ and $C^1_{b,\mathrm{loc}}(\hat \Omega)$ the set of differentiable functions on $\hat \Omega=[0,\infty)^{E_d}$ with bounded derivatives depending only on a finite number of variables.
\[ass:time\_mixing\] There exists $p_1>1$ such that for every $m\in\bN$ the following holds: For each $\varphi,\psi\in B(\Omega)$ of the form $\varphi(\om)=\tilde\varphi(\om(t_1))$ and $\psi(\om)=\tilde\psi(\om(t_2))$ with $|t_1-t_2|\geq 1$ for some $\tilde\varphi, \tilde\psi\in C^1_{b,\mathrm{loc}}(\hat\Omega)$ depending on $m$ variables we have $$\left| \bE[\varphi \psi ] - \bE[\varphi] \bE[\psi] \right| \leq c_m |t_1-t_2|^{-p_1} \|\varphi\|_{L^\infty(\bP)} \|\psi\|_{L^\infty(\bP)}.$$
\[ass:space\_mixing\] Let $d\geq 3$. There exists $p_2>2d/(d-2)$ such that for every $m\in\bN$ and for every $x\in \bZ^d$ the following holds: For each $\varphi,\psi\in B(\Omega)$ of the form $\varphi(\om)=\tilde\varphi(\om(t_0))$ and $\psi(\om)=\tilde\psi(\om(t_0))$ for some $\tilde\varphi, \tilde\psi\in C^1_{b,\mathrm{loc}}(\hat\Omega)$ depending on $m$ variables we have $$\left| \bE[\varphi(\om) \psi(\tau_{0,x}\om) ] - \bE[\varphi] \bE[\psi] \right| \leq c_m |x|^{-p_2} \|\varphi\|_{L^\infty(\bP)} \|\psi\|_{L^\infty(\bP)}.$$
We are now ready to state the following QFCLT as our main result.
\[main-ip\] Let $d\geq 3$ and suppose that Assumptions \[ass:ergodic\]-\[ass:space\_mixing\] hold. Then, $\bP$-a.s. $X^{(\eps)}$ converges (under $P_{0,0}^\om$) in law to a Brownian motion on $\bR^d$ with a deterministic non-degenerate covariance matrix $\Sigma$.
Notice that Theorem \[main-ip\] only covers the transient lattice dimensions $d\geq 3$. In order to get an invariance principle for $X$ also in dimensions $d\leq 2$, we need to modify the mixing assumptions as follows.
\[ass:time\_mixing2\] Assumption \[ass:time\_mixing\] holds with $p_1>d+1$ if $d\geq 2$ and $p_1>4$ if $d=1$.
\[ass:space\_mixing2\] There exists $p_2>1$ such that for every $m\in\bN$ and for every $L>0$ the following holds: For each $\varphi,\psi\in B(\Omega)$ of the form $\varphi(\om)=\tilde\varphi(\om(t_1))$ and $\psi(\om)=\tilde\psi(\om(t_2))$, where $|t_1-t_2|\leq L$ and $\tilde\varphi, \tilde\psi\in C^1_{b,\mathrm{loc}}(\hat\Omega)$ depend on variables contained in two subsets $A_\varphi$ and $A_\psi$ of $\bZ^d$ with diameter at most $m$ and $\dist (A_\varphi, A_\psi)\geq L$, $$\left| \bE[\varphi \psi ] - \bE[\varphi] \bE[\psi] \right| \leq c_m L^{-p_2} \|\varphi\|_{L^\infty(\bP)} \|\psi\|_{L^\infty(\bP)}.$$
\[main-ip2\] Let $d\geq 1$ and suppose that Assumptions \[ass:ergodic\]-\[ass:ellipticity\], \[ass:time\_mixing2\] and \[ass:space\_mixing2\] hold. Then, $\bP$-a.s. $X^{(\eps)}$ converges (under $P_{0,0}^\om$) in law to a Brownian motion on $\bR^d$ with a deterministic non-degenerate covariance matrix $\Sigma$.
\[rem:csrw\] One can also consider the time-inhomogeneous [*constant speed random walk or CSRW*]{} $Y=(Y_t, t\in \bR, P_{s,x}^\omega, (s,x)\in \bR \times\bZ^d)$ with generator given by: $$\begin{aligned}
\sL^{Y}_t f(x) = \sum_{y\sim x} \frac{\mu_{xy}(t)}{\mu_x(t)} (f(y)-f(x)).\end{aligned}$$ In contrast to the VSRW $X$, whose waiting time at any site $x\in \bZ^d$ depends on $x$, the CSRW waits at each site an exponential time with mean one. Since the CSRW is a time change of the VSRW, an invariance principle for $Y$ follows from an invariance principle for $X$ by the same arguments as in [@ABDH Section 6.2]. In this case the limiting object is a Brownian motion in $\mathbb{R}^d$ with covariance matrix $\Sigma_C=(1/\bE \mu_0(0)) \Sigma_V$, where $\Sigma_V$ denotes the covariance matrix of the limiting Brownian motion in the invariance principle for $X$.
Next we state some consequences of our results, which follow from arguments in [@BH] by combining the invariance principle for $X$ and the Gaussian bound for the heat kernel. First, we have a local limit theorem for the heat kernel. Write $$\begin{aligned}
\label{def_k}
k_t(x) =k_t^{(\Sigma)}(x)= \frac 1 {\sqrt{(2\pi t)^d \det \Sigma } } \exp(- x\cdot \Sigma^{-1}x /2t)
\end{aligned}$$ for the Gaussian heat kernel with diffusion matrix $\Sigma$.
\[thm:llt\] Let $T>0$. For $x \in \bR^d$ write $\lfloor x \rfloor =( \lfloor x_1 \rfloor, \dots \lfloor x_d \rfloor)$.
1. Suppose that Assumptions \[ass:ergodic\]-\[ass:ellipticity\] hold. Then, $$\begin{aligned}
\lim_{n\to\infty}
\sup_{x\in \bR^d} \sup_{t \ge T}
\Big|n^{d/2}\bE[ p^\om(0,0; nt, \lfloor n^{1/2}x \rfloor)] - k_t(x)\Big| =0.\end{aligned}$$
2. Under the assumptions of Theorem \[main-ip\] or Theorem \[main-ip2\] we have $$\begin{aligned}
\lim_{n\to\infty}
\sup_{x\in \bR^d} \sup_{t \ge T}
\Big|n^{d/2}p^\om(0,0; nt, \lfloor n^{1/2}x \rfloor)- k_t(x)\Big| =0, \q
\hbox{$\bP$-a.s.} \end{aligned}$$
Given the annealed or quenched invariance principle and the heat kernel bounds in Proposition \[prop:hke\] this can be proven as in Section 4 of [@BH].
When $d\geq 3$ the calculations in Section 6 of [@BH] then give the following bound on the Green kernel $g^\om(x,y)$ defined by $$g^\om (x,y)=\int_0^\infty p^\om (0,x;t,y) \, dt.$$
Let $d\geq 3$ and suppose that the assumptions of Theorem \[main-ip\] or Theorem \[main-ip2\] hold.
1. There exist constants $c_1$ and $c_2$ such that for $x\not= y$ $$\frac{c_1}{|x-y|^{d-2}} \leq g^\om (x,y) \leq \frac{c_2}{|x-y|^{d-2}}.$$
2. Let $C=\Gamma(\frac d 2 -1)/2 \pi^{d/2} \det \Sigma$. For any $\varepsilon >0$ there exists $M=M(\varepsilon, \om)$ with $\bP[M<\infty]=1$ such that $$\begin{aligned}
\frac{(1-\varepsilon) C}{|x|^{d-2}} \leq g^\om (0,x) \leq \frac{(1+\varepsilon)C}{|x|^{d-2}} \qquad \text{for $|x|>M(\om)$.}\end{aligned}$$
3. We have, $\bP$-a.s., $$\lim_{|x|\to\infty} |x|^{2-d} g^\om(0,x)= \lim_{|x|\to\infty} |x|^{2-d} \bE [g^\om(0,x)]=C.$$
In the case of static conductances, quenched invariance principles for the random conductance model have been proven by a number of different authors under various restrictions on the law of the conductances, see [@SS; @BP; @Ma; @BD]. Recently, these results have been unified in [@ABDH], where a QFCLT has been obtained for the RCM with general nonnegative i.i.d. conductances. We also refer the reader to [@Bi] for a recent survey on this topic.
On the other hand, to our knowledge the present paper is the first one proving an invariance principle for the RCM with a time-dynamic environment. However, quenched invariance principles have been proven for several other discrete-time random walks in a dynamic random environment. In [@BMP1] a QFCLT is obtained for random walks in space-time product environments by using Fourier-analytic methods. This result has been improved in [@BMP2] to environments satisfying an exponential spatial mixing assumption and in [@BZ] to Markovian environments by using more probabilistic techniques. Another very successful approach is the well-established Kipnis-Varadhan technique based on the process of the environment as seen from the particle. In [@RS] this approach has been used to get a QFCLT for the random walk in space-time product environments. Moreover, it has been applied in [@DL2] to random walks in a dynamic enviroment, which forms a Gibbsian Markov chain in time with spatial mixing, and in [@JR] to random walks on $\mathbb{R}^d$, where the environment is i.i.d. in time and polynomially mixing in space. Recently, a general class of random walks in an ergodic Markovian environment satisfying some coupling conditions has been studied in [@RV].
Also in this paper we will follow the approach in [@RS], so we use the process of the environment as seen from the particle and the method of the ’corrector’, that is we decompose the random walk $X$ into a martingale and a time-dependent corrector function. Due to the time-inhomogeneity and the resulting lack of reversibility we need to apply the adaptions of the Kipnis-Varadhan method to non-reversible situations in [@MW] and [@KLO]. In particular, in order to construct the corrector we show that the generator of the environment seen from the particle is a perturbation of a normal operator in the sense of [@KLO Section 2.7.5]. This is done in Section \[sec:corr\]. As a byproduct this will already imply the annealed FCLT in Theorem \[annealed-ip\].
Once the corrector is constructed, the QFCLT for the martingale part is standard, so it remains to control the corrector. To that aim we still follow [@RS] and apply the theory of ’fractional coboundaries’ of Derriennic and Lin in [@DL]. The main step in this approach is to establish a subdiffusive bound on the corrector (see Proposition \[prop:asymp\_corr\] below), which is done in Section \[sec:asymp\_corr\]. To obtain this bound we establish so-called two-walk estimates, i.e. we consider the difference of two independent copies of $X$ evolving in the same fixed environment $\om$ (cf. e.g. [@JR] or Appendix A in [@RS2]). In $d\geq 3$, following [@Mou] we show that the variance decay of the environment viewed from the particle is strong enough for our purposes by using the mixing assumption \[ass:time\_mixing\] and \[ass:space\_mixing\] (see Lemma \[lem:asymp\_Pk\]). In the recurrent lattice dimensions $d\leq 2$ the estimate for the variance decay is not good enough, so we give a different argument here involving the modified mixing asumptions \[ass:time\_mixing2\] and \[ass:space\_mixing2\].
In Section \[sec:pf\_main\] we prove the main result, i.e. we state a tightness result, which is a direct consequence from the heat kernel bounds in Proposition \[prop:hke\], and show the QFCLT for the martingale part. To control the corrector we apply the results in [@DL], which are stated in the discrete-time setting. Since it is not clear to us, how to apply them directly in the continuous-time setting, we first prove the QFCLT for the discretized process as in [@BD]. More precisely, we define $\wh X_n = X_n$, $n \in \bN$, and consider the process $$\wh X^{(\eps)}_t = \eps \wh X_{ \lfloor t/\eps^2 \rfloor}.$$ We can control $\sup_{t \le T} | X^{(\eps)}_t - \wh X^{(\eps)}_t|$ – see Lemma \[disc-approx\] – so an invariance principle for $X^{(\eps)}$ will follow from one for $\wh X^{(\eps)}$.
Finally, in Section \[sec:interface\] we point out a link to stochastic interface models (see [@F]). Namely, a local limit theorem for the RCM with dynamic conductances can be used to obtain scaling limits for the space-time covariation of the Ginzburg-Landau interface model via Helffer-Sjöstrand representation.
Throughout the paper we write $c$ to denote a positive constant which may change on each appearance. Constants denoted $c_i$ will be the same through each argument.
**Acknowledgement.** I thank Martin Barlow, Jean-Dominique Deuschel and Martin Slowik for helpful discussions and useful comments.
Construction of the Corrector {#sec:corr}
=============================
Throughout this section we suppose that Assumptions \[ass:ergodic\]-\[ass:ellipticity\] hold. We define the process of the environment seen from the particle by $$\eta_t(\om)=\tau_{t,X_t}\om, \qquad \om \in \Omega, \, t\geq 0.$$
\[prop:semigroup\_eta\]
1. The process $(\eta_t)_{t\geq 0}$ is Markovian with transition semigroup $$P_tf(\om)=\sum_{y \in\bZ^d} p^\om(0,0;t,y) f(\tau_{t,y}\om) \qquad \mbox{for all $f\in B(\Omega)$}.$$ The semigroup $(P_t)$ extends uniquely to a strongly continuous semigroup of contractions on $L^2(\bP)$, whose generator $L: \mathcal{D}(L)\rightarrow L^2(\bP)$ is given by $$\begin{aligned}
Lf(\om)=D_tf(\om)+\sum_{y\sim 0} \mu^\om_{0y}(0) (f(\tau_{0,y}\om)-f(\om))\end{aligned}$$ with domain $\mathcal{D}(L)=\mathcal{D}(D_t)$.
2. The measure $\bP$ is invariant and ergodic for $\eta$.
i\) The Markov property as well as the representation of the semigroup follow from by similar arguments as in Lemma 3.1 in [@KLO]. For every bounded $f\in\mathcal{D}(D_t)$ we have $$\begin{aligned}
\frac {P_tf(\om)-f(\om)} t=\sum_{y\in\bZ^d} \frac{p^\om(0,0;t,y)}{t} (f(\tau_{0,y}\om)-f(\om))+\sum_{y\in\bZ^d} p^\om(0,0;t,y) \frac{f(\tau_{t,y}\om)-f(\tau_{0,y}\om)}{t}.\end{aligned}$$ Taking limits for $t\downarrow 0$, using the fact that $ p^\om(0,0;t,y)\to \delta_{0y}$, we obtain the formula for $Lf$. Obviously, the operators $L$ and $D_t$ have the same domain.
ii\) Let $f\in\mathcal{D}(L)$. Since the operator $D_t$ is anti-selfadjoint we have $\langle D_t f\rangle_{\bP}=0$. Hence, $$\begin{aligned}
\langle L f\rangle_{\bP}&=\sum_{y\in\bZ^d} \langle \mu_{0y}^\om(0) f(\tau_{0,y}\om) \rangle_{\bP}-\langle \mu_{0y}^\om(0) f(\om) \rangle_{\bP}
=\sum_{y\in\bZ^d} \langle \mu_{0y}^{\tau_{0,-y}\om}(0) f(\om) \rangle_{\bP}-\langle \mu_{0y}^\om(0) f(\om) \rangle_{\bP} \\
&=\sum_{y\in\bZ^d} \langle \mu_{0,-y}^\om(0) f(\om) \rangle_{\bP}-\langle \mu_{0y}^\om(0) f(\om) \rangle_{\bP}=0,\end{aligned}$$ where we have used the invariance of $\bP$ w.r.t. $\tau_{t,x}$ and . Thus, $\bP$ is an invariant measure for $\eta$. To prove that $\bP$ is also ergodic, let now $A\in \mathcal{F}$ with $P_t \indicator_A=\indicator_A$. Then, $$\begin{aligned}
0=\indicator_{A^c}(\om) \cdot P_t \indicator_A(\om)=\sum_{y\in\bZ^d} \indicator_{A^c}(\om) p^\om(0,0;t,y) \indicator_A(\tau_{t,y}\om).\end{aligned}$$ Since for all $t>0$ and $y\in \bZ^d$ there is a stricly positive lower bound for $p^\om(0,0;t,y)$ independent of $\om$ (see Proposition 4.3 in [@DD]) we get $$\indicator_{A^c}(\om) \cdot \indicator_A(\tau_{t,y}\om)=0.$$ Thus, the set $A$ is invariant under $\tau_{t,x}$. Since $\bP$ is ergodic w.r.t. $\tau_{t,x}$ we conclude that $A$ is $\bP$-trivial and the claim follows.
\[lem:DF\_eta\] For $f\in \mathcal{D}(L)$, $$\langle f,(-L)f\rangle_{\bP}=\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ \mu_{0y}^\om(0) (f(\tau_{0,y}\om) -f(\om))^2 \right].$$
Recall that $\langle f, D_tf \rangle_{\bP}=0$. Therefore, $$\begin{aligned}
&\langle f,(-L)f \rangle_{\bP}
=-\sum_{y\in\bZ^d} \bE \left[ f(\om) \mu_{0y}^\om (0) (f(\tau_{0,y}\om)-f(\om)) \right] \\
=& -\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ f(\om) \mu_{0y}^\om (0) (f(\tau_{0,y}\om)-f(\om)) \right]-\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ f(\om) \mu_{0,-y}^\om (0) (f(\tau_{0,-y}\om)-f(\om)) \right] \\
=& -\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ f(\om) \mu_{0y}^\om (0) (f(\tau_{0,y}\om)-f(\om))\right]-\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ f(\tau_{0,y}\om) \mu_{0,-y}^{\tau_{0,y}\om} (0) (f(\om)-f(\tau_{0,y}\om)) \right] \\
=&\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ \mu_{0y}^\om(0) (f(\tau_{0,y}\om) -f(\om))^2 \right],\end{aligned}$$ where we have used again the invariance of $\bP$ w.r.t. $\tau_{t,x}$ and .
Let $P_t^*$ and $L^*$ denote the $L^2(\bP)$-adjoint operators of $P_t$ and $L$, respectively.
\[prop:adjoint\] We have $$P_t^*f(\om)=\sum_{y\in\bZ^d} \hat p^\om(0,0;t,y) f(\tau_{-t,y}\om), \qquad \mbox{$f\in L^2(\bP)$},$$ with $\hat p^\om (s,x;t,y):=p^\om(-t,y;-s,x)$ and for $f\in \mathcal{D}(L)$ $$L^*f(\om)=-D_tf(\om)+\sum_{y\sim 0} \mu^\om_{0y}(0) (f(\tau_{0,y}\om)-f(\om)).$$
Using we compute the adjoint of $P_t$ as $$\begin{aligned}
\langle P_t f, g\rangle_{\bP}&= \sum_{y\in\bZ^d} \bE \left[ p^\om(0,0;t,y) f(\tau_{t,y}\om) g(\om)\right]= \sum_{y\in\bZ^d} \bE \left[ p^{\tau_{-t,-y}\om}(0,0;t,y) f(\om) g(\tau_{-t,-y}\om)\right] \\
&=\sum_{y\in\bZ^d} \bE \left[ p^{\om}(-t,-y;0,0) f(\om) g(\tau_{-t,-y}\om)\right]=\sum_{y\in\bZ^d} \bE \left[ \hat p^{\om}(0,0;t,y) g(\tau_{-t,y}\om) f(\om) \right],\end{aligned}$$ and the representation for $P_t^*$ follows. To compute $L^*$ we use a similar procedure as in Lemma \[lem:DF\_eta\] and get $$\begin{aligned}
\langle Lf,g \rangle_{\bP}&=\langle D_t f,g \rangle_{\bP}+\sum_{y\in\bZ^d} \bE\left[ \mu_{0y}^\om(0) (f(\tau_{0,y}\om)-f(\om)) g(\om) \right] \\
&=-\langle f,D_tg \rangle_{\bP}-\tfrac 1 2 \sum_{y\in\bZ^d} \bE\left[ \mu_{0y}^\om(0) (f(\tau_{0,y}\om)-f(\om)) (g(\tau_{0,y}\om)-g(\om)) \right] \\
&=-\langle f,D_tg \rangle_{\bP}+\sum_{y\in\bZ^d} \bE\left[ \mu_{0y}^\om(0) (g(\tau_{0,y}\om)-g(\om)) f(\om) \right],\end{aligned}$$ which gives the claim.
Next we introduce the Hilbert spaces $\mathcal{H}_1$ and $\mathcal{H}_{-1}$. Let $\mathcal{C}$ be a common core of the operators $L$ and $L^*$. On $\mathcal{C}$ we define the seminorm $$\| f\|^2_{\mathcal{H}_1}=\langle f,(-L)f\rangle_{\bP}, \qquad f\in \mathcal{C}.$$ Let $\mathcal{H}_1$ be the completion of $\mathcal{C}$ (or more precisely the completion of equivalence classes of elements in $\mathcal{C}$ w.r.t. the equivalence relation $f\sim g$ if $\|f-g\|_{\mathcal{H}_1}=0$) w.r.t. $\|.\|_{\mathcal{H}_1}$. Then, $\mathcal{H}_1$ is a Hilbert space with inner product $\langle .,. \rangle_{\mathcal{H}_1}$ given by polarization: $$\langle f,g \rangle_{\mathcal{H}_1}=\frac 1 4 \left( \| f+g \|^2_{\mathcal{H}_1}- \| f-g\|^2_{\mathcal{H}_1} \right).$$ Associated with $\mathcal{H}_1$ we define the dual space $\mathcal{H}_{-1}$ as follows. For $f\in L^2(\bP)$ let $$\| f\|^2_{\mathcal{H}_{-1}}=\sup_{g\in \mathcal{C}}\left( 2 \langle f ,g \rangle_{\mathbb{P}} - \| g\|^2_{\mathcal{H}_1} \right).$$ The Hilbert space $\mathcal{H}_{-1}$ is then defined as the $\| .\|_{\mathcal{H}_{-1}}$-completion of (equivalence classes of) elements in $\mathcal{C}$ with finite $\| .\|_{\mathcal{H}_{-1}}$-norm. As before the inner product $\langle .,. \rangle_{\mathcal{H}_{-1}}$ is defined through polarization. We refer to Section 2.2 in [@KLO] for more details.
Next we define the local drift $$V_j(\om)=\sum_{y\sim 0} \mu_{0y}^\om (0) y^j =\sL^\om_0 f_j(0), \qquad j=1,\ldots,d,$$ where $f_j(x)=x^j$, $x^j$ and $y^j$ denoting the $j$-th component of $x$ and $y$. Since $\mu^\om_{0y}(0)=0$ unless $y\sim 0$, we have $V_j(\om)=\mu^\om_{0,e_j}(0)-\mu^\om_{0,-e_j}(0)$.
\[lem:V\_H-1\] For every $j=1,\ldots,d$, $V_j\in L^2(\bP)\cap \mathcal{H}_{-1}$.
It suffices to show that $$\begin{aligned}
\label{eq:estH-1}
\big| \langle V_j,f\rangle_{\bP}\big|^2 \leq c \langle f, (-L)f \rangle_{\bP} \qquad \text{for all $f\in\mathcal{H}_1$}\end{aligned}$$ (cf. equation (2.12) in [@KLO])). By definition of $V_j$ we have $$\begin{aligned}
\langle V_j,f\rangle_{\bP}&=\bE \left[ \mu^\om_{0,e_j}(0) f(\om) \right]- \bE \left[\mu^\om_{0,-e_j}(0) f(\om)\right]=\bE \left[ \mu^\om_{0,e_j}(0) f(\om) \right]- \bE \left[ \mu^\om_{0,e_j}(0) f(\tau_{0,e_j}\om) \right] \\
&=-\bE \left[ \mu^\om_{0,e_j}(0) (f(\tau_{0,e_j}\om)-f(\om)) \right].\end{aligned}$$ Hence, using Cauchy Schwarz and Lemma \[lem:DF\_eta\] $$\begin{aligned}
\big| \langle V_j,f\rangle_{\bP}\big|^2 &\leq \bE[ \mu^\om_{0,e_j}(0)] \, \bE[ \mu^\om_{0,e_j}(0) (f(\tau_{0,e_j}\om)-f(\om))^2]
\leq C_u \sum_{y\in\bZ^d} \bE \left[ \mu^\om_{0y}(0) (f(\tau_{0,y}\om)-f(\om))^2 \right] \\
&=2 C_u \langle f, (-L)f \rangle_{\bP},\end{aligned}$$ and we obtain .
For $\lambda>0$, we consider for each $j$ the solution $u_\lambda^{j}$ of the resolvent equation $$\label{eq:res}
(\lambda-L) u_\lambda^{j}=V_j.$$
\[prop:conv\_u\] For every $j=1,\ldots,d$, there exists $u^{j} \in \mathcal{H}_1$ such that $$\begin{aligned}
\lim_{\lambda\to 0} \lambda \| u_\lambda^{j} \|^2_{L^2(\bP)}=0 \quad \text{and} \quad
\lim_{\lambda\to 0}u_\lambda^{j}=u^{j} \, \text{strongly in $\mathcal{H}_1$}.\end{aligned}$$
The proof of Proposition \[prop:conv\_u\] will be based on the following statement proven in [@KLO].
\[prop:pert\_op\] Suppose we have the decomposition $L=L^0+B$ of the operator $L$ such that
1. The operator $L^0$ is normal, i.e. $L^0 (L^0)^*=(L^0)^*L^0$.
2. The Dirichlet forms of $L$ and $L^0$ are equivalent, i.e. there exist positive constants $c_1$ and $c_2$ such that $$c_1 \langle f, (-L)f \rangle_{\bP}\leq \langle f, (-L^0)f \rangle_{\bP}\leq c_2 \langle f, (-L)f \rangle_{\bP}, \qquad \text{for all $f\in \mathcal{D}(L)$}.$$
3. $B$ satisfies a sector condition w.r.t. $L^0$, i.e. there exists a positive constant $c$ such that $$\begin{aligned}
\langle f,Bg\rangle^2_{\bP} \leq c \langle f, (-L^0)f\rangle_{\bP} \langle g, (-L^0)g\rangle_{\bP}, \qquad f,g\in \mathcal{D}(L).\end{aligned}$$
Then, for any fixed $V\in L^2(\bP)\cap \mathcal{H}_{-1}$ the solution $f_\lambda$ of the resolvent equation $(\lambda-L)f_\lambda=V$ satisfies $$\begin{aligned}
\lim_{\lambda\to 0} \lambda \| f_\lambda\|^2_{L^2(\bP)}=0 \quad \text{and} \quad
\lim_{\lambda\to 0}f_\lambda=f \, \text{strongly in $\mathcal{H}_1$},\end{aligned}$$ for some $f \in \mathcal{H}_1$.
By Proposition 2.25 in [@KLO] the assumptions imply that $$\sup_{0<\lambda \leq 1} \|Lf_\lambda \|_{\mathcal{H}_{-1}}<\infty.$$ The claim follows then from Lemma 2.16 in [@KLO].
**Proof of Proposition \[prop:conv\_u\].** We decompose the operator $L=L^0+B$ with $$\begin{aligned}
L^0 f&:=D_t f+\sum_{y\sim 0} C_l (f(\tau_{0,y}\om)-f(\om)), \qquad f\in \mathcal{D}(L), \intertext{and}
Bf&:= \sum_{y\sim 0} (\mu^\om_{0y}(0)-C_l) (f(\tau_{0,y}\om)-f(\om)), \qquad f\in \mathcal{D}(L).\end{aligned}$$ A similar calculation as in the proof of Lemma \[lem:DF\_eta\] shows that $$\begin{aligned}
\label{eq:DFL0}
\langle f, (-L^0)f \rangle_{\bP}&=\tfrac 1 2 \sum_{y\sim 0 } \bE \left[ C_l (f(\tau_{0,y}\om)-f(\om))^2 \right],\\ \label{eq:DFB}
\langle f, (-B)g \rangle_{\bP}&=\tfrac 1 2 \sum_{y\sim 0 } \bE \left[ (\mu^\om_{0y}(0)-C_l) (f(\tau_{0,y}\om)-f(\om)) (g(\tau_{0,y}\om)-g(\om))\right].\end{aligned}$$ The claim will follow from Proposition \[prop:pert\_op\] and Lemma \[lem:V\_H-1\] once we have verified conditions i)-iii) in Proposition \[prop:pert\_op\]. To show i), note that the closure of $L^0$ is the generator of a semigroup $(P^0_t)$ that corresponds to a process seen from the particle associated with a simple random walk on $\bZ^d$ with constant jump rates $C_l$. In particular, the associated process is time-homogeneous, i.e. the corresponding transition probabilities satisfy $p_0^\om(s,x;t,y)=p_0^\om(t-s,x,y)$ and $\hat p_0^\om(s,x;t,y)=\hat p_0^\om(t-s,x,y)$, where $p_0^\om(t,x,y)=p_0^\om(0,x;t,y)$ and $\hat p_0^\om(t,x,y)=\hat p_0^\om(0,x;t,y)$. Since this random walk is obviously reversible w.r.t. the counting measure, we have $p_0^\om(t,x,y)=\hat p_0^\om(t,x,y)$. Then, since we have similar representations for $P^0_t$ and $(P^0_t)^*$ as for the semigroups in Proposition \[prop:semigroup\_eta\] and Proposition \[prop:adjoint\], we get $$(P^0_t)^*P^0_t=P^0_t (P^0_t)^*, \qquad t\geq 0,$$ which implies that the closure of $L^0$ is normal (see Theorem 13.37 in [@Ru]).
Condition ii) is immediate from Lemma \[lem:DF\_eta\], and the ellipticity condition . To prove iii) we use , Cauchy Schwarz and the ellipticity condition , which gives $$\begin{aligned}
\langle f,Bg\rangle^2_{\bP} &\leq
\tfrac 1 2 C_u^2 d \, \bE \bigl[ \sum_{y\sim 0 } (f(\tau_{0,y}\om)-f(\om))^2 \bigr] \times \bE \bigl[ \sum_{y\sim 0 }(g(\tau_{0,y}\om)-g(\om))^2 \bigr] \\
&\leq \frac{ C^2_u d}{2 C^2_l} \langle f, (-L^0)f\rangle_{\bP} \langle g, (-L^0)g\rangle_{\bP},\end{aligned}$$ and the claim follows.
For abbreviation we write $u_\lambda=(u_\lambda^{1},\ldots,u_\lambda^{d})$ and $$\chi_{\lambda}(t,x,\om):= u_\lambda \circ \tau_{t,x}-u_\lambda.$$
\[prop:constr\_corr\] For all non-negative $t\in \mathbb{Q}$ and $x\in \bZ^d$ the limit $$\begin{aligned}
\label{def_chi}
\lim_{\lambda'\to 0}\chi_{\lambda'}(t,x,\om) =:\chi(t,x,\om)\end{aligned}$$ exists along a subfamily $(\lambda')$ for $\bP$-a.e. $\om$. Moreover, the mapping $t\mapsto \chi(t,X_t,\om)$ can be extended to a right-continuous function on $[0,\infty)$ such that $$\begin{aligned}
\label{eq:decompX}
M_t=X_t+\chi(t,X_t,\om), \qquad t\geq 0,\end{aligned}$$ is a $P_{0,0}^\om$-martingale.
For every $j=1,\ldots,d$ and every $\lambda>0$ we have that for $\bP$-a.e. $\om$ the processes $$\begin{aligned}
\label{def_Nk}
N^{j,\lambda}_t=u_\lambda^{j}(\eta_t)-u_\lambda^{j}(\om)-\int_0^t Lu_\lambda^{j}(\eta_s) \, ds\end{aligned}$$ and $$\begin{aligned}
\label{def_tildeM}
\tilde M^j_t=X^j_t-\int_0^t \mathcal{L}^\om_s f_j(X_s) \, ds\end{aligned}$$ are both $P_{0,0}^\om$-martingales, where as before $f_j(x)=x^j$. Then, using the definition of $V_j$ and the fact that $u_\lambda^{j}$ solves the resolvent equation we get $$\begin{aligned}
\label{eq:mart_decomp}
X_t^j&=\tilde M^j_t+\int_0^t \mathcal{L}^\om_s f_j(X_s) \, ds=\tilde M^j_t+\int_0^t V_j(\eta_s) \, ds=\tilde M^j_t+\int_0^t (\lambda-L)u_\lambda^{j}(\eta_s) \, ds \nonumber \\
&=\tilde M^j_t+N^{j,\lambda}_t-\left(u_\lambda^{j}(\eta_t)-u_\lambda^{j}(\om)\right)+\lambda \int_0^t u_\lambda^{j}(\eta_s) \, ds.\end{aligned}$$
In a first step we show that the martingale $N^{j,\lambda}_t$ converges in $L^2(\bP\otimes P_{0,0}^\om)$ as $\lambda \downarrow 0$ to a martingale $N^j_t$. To that aim it is enough to prove that $N^{j,\lambda}_t$ is a Cauchy sequence in $L^2(\bP\otimes P_{0,0}^\om)$. Since $\bP$ is an invariant measure for $\eta$ we use Lemma \[lem:DF\_eta\] to obtain $$\begin{aligned}
\bE E_{0,0}^\om \langle N^{j,\lambda}-N^{j,\lambda'}\rangle_t&= \int_0^t \bE E_{0,0}^\om \left[ L (u_\lambda^{j}-u_{\lambda'}^{j})^2-2 (u_\lambda^{j}-u_{\lambda'}^{j}) L(u_\lambda^{j}-u_{\lambda'}^{j}) \right](\eta_s) \, ds \\
&=2 t \langle (u_\lambda^{j}-u_{\lambda'}^{j}), (-L)(u_\lambda^{j}-u_{\lambda'}^{j})\rangle_{\bP} \\
&=2t \| u_\lambda^{j}-u_{\lambda'}^{j} \|^2_{\mathcal{H}_1},\end{aligned}$$ which implies that $N^{j,\lambda}_t$ is a Cauchy sequence in $L^2(\bP\otimes P_{0,0}^\om)$ by Proposition \[prop:conv\_u\]. Thus, the martingale $M_t^{j,\lambda}=\tilde M_t^j+N_t^{j,\lambda}$ converges to a martingale, whose right-continuous modification we denote by $M^j_t$. We define $M_t=(M_t^1,\ldots, M_t^d)$.
The next step is to show that the last term in converges to zero in $L^2(\bP\otimes P_{0,0}^\om)$ as $\lambda \downarrow 0$. Since $V_j\in \mathcal{H}_{-1}$ we have that $\lim_\lambda \lambda u_\lambda^{j}=0$ in $L^2(\bP)$ (cf. equation (2.15) in [@KLO]). Thus, for every $j=1,\ldots,d$, $$\begin{aligned}
\big\| \lambda \int_0^t u_\lambda^{j}(\eta_s) \, ds \big\|_{L^2(\bP\otimes P_{0,0}^\om)} \leq \lambda \int_0^t \| u_\lambda^{j}(\eta_s) \|_{L^2(\bP\otimes P_{0,0}^\om)}\, ds=t \, \lambda \| u_\lambda^{j} \|_{L^2(\bP)} \rightarrow 0.\end{aligned}$$
Thus, by taking $L^2(\bP\otimes P_{0,0}^\om)$-limits in we get that $\chi_{\lambda}(t,X_t,\cdot)$ converges in $L^2(\bP\otimes P_{0,0}^\om)$ as $\lambda\downarrow 0$ for every $t\geq 0$. By a diagonal procedure we can extract a suitable subsequence $\lambda'$ such that for $\bP$-a.e. $\om$ we have that $\chi_{\lambda}(t,X_t,\om)$ has a limit in $L^2( P_{0,0}^\om)$ and $P_{0,0}^\om$-a.s. along $\lambda'$ for all non-negative $t\in\mathbb{Q}$. In particular, the limit is $\sigma(X_t)$-measurable and will therefore be denoted by $\chi(t,X_t,\om)$. Hence, $$\begin{aligned}
\label{eq:Mpre}
X_t=M_t-\chi(t,X_t,\om).\end{aligned}$$ Moreover, for $\bP$-a.e. $\om$, $$\begin{aligned}
\sum_{y\in\bZ^d} p^\om(0,0;t,y) \left| \chi_{\lambda}(t,y,\om) - \chi(t,y,\om) \right|^2=E^0_\om \left| ( \chi_{\lambda}(t,X_t,\om) ) - \chi(t,X_t,\om)\right|^2 \rightarrow 0\end{aligned}$$ along $\lambda'$. Since $p^\om(0,0;t,y)>0$ for all $t>0$ and $x\in \bZ^d$, we conclude that for $\bP$-a.e. $\om$ the limit in exists for every non-negative $t\in \mathbb{Q}$ and every $y\in \mathbb{Z}^d$. Finally, using and the fact that $X_t$ and $M_t$ have right-continuous trajectories, we can extend $\chi(t,X_t,\om)$ to a right-continuous function on $[0,\infty)$ and follows.
\[rem:defh\] Note that for all non-negative $s,t \in \mathbb{Q}$ and $x,y\in\bZ^d$, $$\begin{aligned}
u_\lambda(\tau_{t,y}\om)-u_\lambda(\tau_{s,x}\om)=&\left(u_\lambda(\tau_{t,y}\om)-u_\lambda(\om) \right)- \left(u_\lambda(\tau_{s,x}\om)-u_\lambda(\om)\right) \\
\rightarrow & \chi(t,y,\om)-\chi(s,x,\om)\end{aligned}$$ along the chosen subsequence for $\bP$-a.e. $\om$. The function $h_\lambda(\om_0,\om_1):=u_\lambda(\om_1)-u_\lambda(\om_0)$ on $\Omega \times \Omega$ converges in $L^2(\Omega\times\Omega, \bP\circ (\tau_{s,x},\tau_{t,y})^{-1})$ to a function $h$. In particular, for $\bP$-a.e. $\om$, $$h(\tau_{s,x}\om,\tau_{t,y}\om)=\chi(t,y,\om)-\chi(s,x,\om).$$
\[cor:cocycle\] For $\bP$-a.e. $\om$ the corrector satisfies the cocycle property $$\chi(s+t,x+y,\om)=\chi(s,x,\om)+\chi(t,y,\tau_{s,x}\om)$$ for non-negative $s,t\in \mathbb{Q}$.
We have $$\begin{aligned}
u_\lambda\circ\tau_{s+t,x+y}-u_\lambda=(u_\lambda\circ\tau_{s,x}-u_\lambda)+(u_\lambda\circ\tau_{t,y}-u_\lambda)\circ \tau_{s,x},\end{aligned}$$ and the claim follows by taking the $L^2(\bP)$-limit along $\lambda'$ on both sides.
In the following, for any $G:\bZ^d\times\Omega \rightarrow \bR$ we shall write $$\| G\|_\om^2:=\sum_{y\sim 0} \mu_{0y}^\om(0) G(y,\om)^2 .$$
\[cor:Mbracket\] For every $v\in \mathbb{R}^d$ the covariation process of the martingale $M^v:=v\cdot M$ is given by $$\begin{aligned}
\label{eq:Mbracket}
\langle v \cdot M\rangle_t=\int_0^t \| v\cdot \Phi \|_{\eta_s}^2 ds,
\intertext{where} \label{def:Phi}
\Phi(x,\om):=x+\chi(0,x,\om).
\end{aligned}$$
First we compute the covariation process of the martingale $M_t^{j,\lambda}$ defined as in the proof of Proposition \[prop:constr\_corr\]. To that aim we define $z_{j,\lambda}(t,x,\om):=x^j+ u_\lambda^{j}(\tau_{t,x}\om)$. Then, by adding and we get $$\begin{aligned}
M_t^{j,\lambda}=z_{j,\lambda}(t,X_t,\om)-z_{j,\lambda}(0,0,\om)-\int_0^t \bar L z_{j,\lambda}(s,X_s,\om) \, ds,
\intertext{where}
\bar Lz_{j,\lambda}(t,x,\om):=D_t z_{j,\lambda}(t,x,\om)+\sum_{y\sim x} \mu_{xy}^\om (t) (z_{j,\lambda}(t,y,\om)-z_{j,\lambda}(t,x,\om)).\end{aligned}$$ In particular, $$\begin{aligned}
\bar Lz^2_{j,\lambda}-2 z_{j,\lambda} \bar L z_{j,\lambda}(t,x,\om)&=\sum_{y\in\bZ^d} \mu_{xy}^\om (t) (z_{j,\lambda}(t,y,\om)-z_{j,\lambda}(t,x,\om))^2 \\
&= \sum_{y\in\bZ^d} \mu_{0,y-x}^{\tau_{t,x}\om}(0) \left( y^j-x^j+ [u_\lambda^{j}\circ\tau_{0,y-x}-u_\lambda^{j}]\circ \tau_{t,x} (\om) \right)^2 \\
&=\sum_{y\in\bZ^d} \mu_{0,y}^{\tau_{t,x}\om}(0) \left( y^j+ [u_\lambda^{j}\circ\tau_{0,y}-u_\lambda^{j}]\circ \tau_{t,x} (\om) \right)^2,\end{aligned}$$ so that $$\begin{aligned}
\langle M_t^{j,\lambda}\rangle_t=\int_0^t \sum_{y\in\bZ^d} \mu_{0,y}^{\eta_s} (0) \left( y^j+ [u_\lambda^{j}\circ\tau_{0,y}-u_\lambda^{j}] (\eta_s) \right)^2 ds,\end{aligned}$$ and by taking limits on both sides along $\lambda'$, we obtain $$\begin{aligned}
\langle M^j\rangle_t=\int_0^t \sum_{y\in\bZ^d} \mu_{0y}^{\eta_s}(0) [y^j+\chi^j(0,y,\eta_s)]^2 ds=\int_0^t \| \Phi^j \|_{\eta_s}^2 \, ds.
\end{aligned}$$ For an arbitrary $v\in \bR^d$ a similar computation gives .
We conclude this section with a convergence result, which will imply the annealed invariance principle. Nevertheless, it will be convenient to complete the proof of Theorem \[annealed-ip\] in Section \[sec:pf\_main\] below.
\[prop:conv\_corr1\] We have $t^{-1/2} \chi(t,X_t,\om) \to 0$ in $L^2(\bP^*_{0,0})$ as $t\to \infty$.
Consider an arbitrary fixed $j\in\{1,\ldots,d\}$. Still using the notation in the proof of Proposition \[prop:constr\_corr\] we have for every $t$ and any $\lambda>0$, $$\begin{aligned}
\chi^j(t,X_t,\om)=M_t^j-X_t^j=M_t^j-M_t^{j,\lambda}+u_\lambda^{j}(\eta_t)-u_\lambda^{j}(\om)-\lambda \int_0^t u_\lambda^{j}(\eta_s) \, ds,\end{aligned}$$ and by Cauchy-Schwarz we get $$|\chi^j(t,X_t,\om)|^2\leq 3 |M_t^j-M_t^{j,\lambda}|^2+3 |u_\lambda^{j}(\eta_t)-u_\lambda^{j}(\om)|^2 + 3 \lambda^2 \int_0^t |u_\lambda^{j}(\eta_s)|^2 \, ds.$$ We argue similarly as in the proof of Proposition \[prop:constr\_corr\]. Using the fact that $\bP$ is an invariant measure for the environment process $\eta$ we obtain $$\begin{aligned}
\bE E_{0,0}^\om |M_t^j-M_t^{j,\lambda}|^2&\leq 2t \|u_\lambda^{j}-u^{j}\|^2_{\mathcal{H}_1}, \label{est_term1_res} \\
\bE E_{0,0}^\om |u_\lambda^{j}(\eta_t)-u_\lambda^{j}(\om)|^2 &\leq 4 \|u_\lambda^{j}\|^2_{L^2(\bP)}, \\
\bE E_{0,0}^\om \lambda^2 \int_0^t |u_\lambda^{j}(\eta_s)|^2 \, ds &\leq t \lambda^2 \|u_\lambda^{j}\|^2_{L^2(\bP)} \label{est_term3_res}.\end{aligned}$$ Choosing $\lambda=t^{-1}$ the claim follows by Proposition \[prop:conv\_u\].
Subdiffusive Bound on the Corrector {#sec:asymp_corr}
===================================
In this section we shall prove the following
\[prop:asymp\_corr\] Under the assumptions of Theorem \[main-ip\] or Theorem \[main-ip2\], there exists an $\alpha<1/2$ such that $$\bE E_{0,0}^\om\left[|\chi(n,X_n,\om)|^2\right]=O(n^{2\alpha}).$$
Convergence of the Resolvents
-----------------------------
\[prop:ulambda\] Under the assumptions of Theorem \[main-ip\] or Theorem \[main-ip2\], there exists an $\alpha<1/2$ such that for every $j=1,\ldots,d$, $$\|u_\lambda^{j}\|_{L^2(\bP)}=O(\lambda^{-\alpha}).$$
Note that while for the annealed FCLT the convergence in Proposition \[prop:conv\_corr1\] is sufficient, we will need the stronger statement in Proposition \[prop:asymp\_corr\] for the QFCLT. This difference also appears in the corresponding results on the resolvents $u_\lambda$, (cf. Proposition \[prop:conv\_u\] and Proposition \[prop:ulambda\]). Before we prove Proposition \[prop:ulambda\] we will first show how it implies Proposition \[prop:asymp\_corr\].
Similarly to the proof of Proposition \[prop:conv\_corr1\] we show that for a certain $\lambda$ chosen below depending on $n$ the terms in the right hand side of - are in $O(n^{2\alpha})$. We shall use similar arguments as in [@MW], in particular cf. Lemma 2 and Corollary 4 in [@MW]. In a first step we will show that $$\begin{aligned}
\label{est_normH1}
\|u_\lambda^{j}-u_{\lambda'}^{j}\|_{\mathcal{H}_1}^2\leq \frac {(\sqrt{\lambda}+\sqrt{\lambda'})^2}{2} \left( \|u_\lambda^{j}\|^2_{L^2(\bP)}+ \|u_{\lambda'}^{j}\|^2_{L^2(\bP)}\right).\end{aligned}$$ Indeed, using the fact that $u_\lambda^{j}$ solves the resolvent equation we have $$\begin{aligned}
\|u_\lambda^{j}-u_{\lambda'}^{j}\|_{\mathcal{H}_1}^2
&=\langle u_\lambda^{j}-u_{\lambda'}^{j}, (-L)(u_\lambda^{j}-u_{\lambda'}^{j}) \rangle_{L^2(\bP)}
=\langle u_\lambda^{j}-u_{\lambda'}^{j}, -(\lambda u_\lambda^{j}-\lambda' u_{\lambda'}^{j}) \rangle_{L^2(\bP)} \\
&=-\lambda \|u_\lambda^{j}\|^2_{L^2(\bP)} -\lambda' \|u_{\lambda'}^{j}\|^2_{L^2(\bP)}+ (\lambda+\lambda') \langle u_{\lambda}^{j},u_{\lambda'}^{j}\rangle_{L^2(\bP)} \\
&\leq 2\sqrt{\lambda \lambda'} \|u_{\lambda}^{j}\|_{L^2(\bP)} \|u_{\lambda'}^{j}\|_{L^2(\bP)}+(\lambda+\lambda') \langle u_{\lambda}^{j},u_{\lambda'}^{j}\rangle_{L^2(\bP)} \\
&\leq (\sqrt{\lambda}+\sqrt{\lambda'})^2\|u_{\lambda}^{j}\|_{L^2(\bP)} \|u_{\lambda'}^{j}\|_{L^2(\bP)},\end{aligned}$$ which gives . In particular, choosing $\lambda_k=2^{-k}$, we get $$\begin{aligned}
\|u_{\lambda_k}^{j}-u_{\lambda_{k-1}}^{j}\|_{\mathcal{H}_1}^2&\leq \frac {(\sqrt{2^{-k}}+\sqrt{2^{-k+1}})^2}{2} \left( \|u_{\lambda_k}^{j}\|^2_{L^2(\bP)}+ \|u_{\lambda_{k-1}}^{j}\|^2_{L^2(\bP)}\right) \\
&= \frac {(\sqrt{2}+1)^2}{2} \lambda_k \left( \|u_{\lambda_k}^{j}\|^2_{L^2(\bP)}+ \|u_{\lambda_{k-1}}^{j}\|^2_{L^2(\bP)}\right).\end{aligned}$$ Let now $k_n$ be the integer $k$ such that $2^{k-1}\leq n <2^k$. Then, we use the elementary estimate $\sqrt{a+b}\leq \sqrt{a}+\sqrt{b}$ for any $a,b\geq 0$ to obtain $$\begin{aligned}
\|u_{\lambda_{k_n}}^{j}-u^{j}\|_{\mathcal{H}_1}\leq \sum_{m=k_n+1}^\infty \|u_{\lambda_m}^{j}-u_{\lambda_{m-1}}^{j}\|_{\mathcal{H}_1}
\leq c \sum_{m=k_n+1}^\infty \sqrt{\lambda_m} \left( \|u_{\lambda_m}^{j}\|_{L^2(\bP)}+ \|u_{\lambda_{m-1}}^{j}\|_{L^2(\bP)}\right).\end{aligned}$$ Recall that $ \|u_{\lambda_m}^{j}\|_{L^2(\bP)}=O(\lambda_m^{-\alpha})$ by Proposition \[prop:ulambda\]. Therefore, for $n$ large enough $$\begin{aligned}
\|u_{\lambda_{k_n}}^{j}-u^{j}\|_{\mathcal{H}_1} \leq c \sum_{m=k_n+1}^\infty \lambda_m^{1/2-\alpha}=c \lambda_{k_n}^{1/2-\alpha}=c n^{\alpha-1/2}.\end{aligned}$$ Thus, the claim follows by choosing $\lambda_{k_n}$ for $\lambda$ in equation -.
Recall that $(P_t)_{t\geq 0}$ denotes the transition semigroup of the environment process $\eta$.
\[lem:asymp\_Pk\] Under the assumptions of Theorem \[main-ip\] or Theorem \[main-ip2\], there exists $\alpha<1/2$ such that for every $j=1,\ldots, d$, $$\begin{aligned}
\big\| \int_{0}^t P_s V_j \, ds \big\|_{L^2(\bP)} \leq c (1\vee t)^\alpha.\end{aligned}$$
Lemma \[lem:asymp\_Pk\], which will be proven in the next subsection, immediately implies Proposition \[prop:ulambda\].
Since $u_\lambda^{j}$ is the solution of the resolvent equation , $$\begin{aligned}
u_\lambda^{j} = \int_0^\infty e^{-\lambda s} P_sV_j \, ds= \lambda \int_0^\infty \int_s^\infty e^{-\lambda t} P_s V_j \, dt \, ds=\lambda \int_0^\infty e^{-\lambda t} \int_0^t P_sV_j \, ds \, dt.\end{aligned}$$ Hence, by Lemma \[lem:asymp\_Pk\] we get that $$\| u_\lambda^{j} \|_{L^2(\bP)} \leq c_1 \lambda \int_0^\infty e^{-\lambda t} (1\vee t)^\alpha \, dt
\leq c_1+ c_1 \lambda \int_1^\infty e^{-\lambda t} t^\alpha \, dt \leq c_1 +c_1 \Gamma(\alpha+1) \lambda^{-\alpha},$$ which is the claim.
A Two-Walk Estimate
-------------------
In this subsection we prove Lemma \[lem:asymp\_Pk\]. We shall use techniques from [@JR Section 3], [@RS2 Appendix A] and [@Mou]. Denote by $(X_t)_t$ and $(\tilde X_t)_t$ two independent random walks evolving in the same environment $\om$ both starting from zero. We will write $\bP_{2,x,\tilde x}$ in short for the averaged law of $(X,\tilde X)$ starting in $(x,\tilde x)\in \bZ^d\times \bZ^d$ and $\bE_{2,x,\tilde x}$ for the corresponding expectation, i.e. $\bE_{2,x, \tilde x}=\bE\otimes E^\om_{0,x}\otimes E^\om_{0,\tilde x}$. For abbreviation we will write $\bP_{2,x}=\bP_{2,x,0}$ and $\bE_{2,x}=\bE_{2,x,0}$ as well as $\bP_{2}=\bP_{2,0,0}$ and $\bE_{2}=\bE_{2,0,0}$. Furthermore, let $(Y_t)_{t\geq 0}$ be the continuous time Markov chain evolving in an environment $\om$ with transition probabilities given by $$\pi^\om_{s,x}[Y_t\in A]=P^\om[Y_t\in A | Y_s=x]=\sum_{u,v\in \bZ^d} \indicator_{\{v-u\in A\}} p^\om(s,0;t,u) p^\om(s,x;t,v).$$ The corresponding expectation will be denoted by $E^{\pi,\om}_{s,x}$. In particular, note that for every $\om$ the law of $ X_t-\tilde X_t$ induced by $E^\om_x\otimes E^\om_0$ is the same as that of $Y_t$.
\[lem:probY\] For any $0\leq s \leq t$ with $t-s\geq 1$, $y\in \bZ^d$ and any ball $B(x,r)$ we have $$\begin{aligned}
\pi^\om_{s,y} [Y_t\in B(x,r)]\leq c (t-s)^{-d/2} r^d.\end{aligned}$$
By the heat kernel estimates in Proposition \[prop:hke\] we have $$\begin{aligned}
\pi^\om_{s,y} [Y_t\in B(x,r)]&= \sum_{u,v\in \bZ^d} \indicator_{\{v-u\in B(x,r)\}} p^\om(s,0;t,u) p^\om(s,y;t,v)\\
&=\sum_{u\in \bZ^d} p^\om(s,0;t,u) \sum_{v\in \bZ^d} \indicator_{\{v\in B(x+u,r)\}} p^\om (s,y;t,v) \\
&\leq c (t-s)^{-d/2} r^d,\end{aligned}$$ which is the claim.
### Proof of Lemma \[lem:asymp\_Pk\] under Assumptions \[ass:time\_mixing\] and \[ass:space\_mixing\]
Let $d\geq 3$ and assume that \[ass:ergodic\]-\[ass:space\_mixing\] hold. It is enough to show that there exists $\beta>1/2$ such that for every $j=1,\ldots, d$, $$\begin{aligned}
\label{eq:asymp_Pk}
\big\| P_t V_j \big\|_{L^2(\bP)} \leq c (1\vee t )^{-\beta}.\end{aligned}$$ First note that by definition $V_j(\om)=\mu_{0,e_j}^\om(0)-\mu_{0,-e_j}^\om(0)$, so by Assumptions \[ass:ergodic\] and \[ass:ellipticity\] we have $\bE [ V_j]=0$ and $\|V_j\|_{L^\infty(\bP)}\leq 2C_u$, respectively. In particular, it suffices to prove for $t\geq 1$. Setting $$S_n(f):=\sum_{x\in B(0,n)} f(\tau_{0,x}\om),$$ we have by the translation invariance of $\bP$ $$\begin{aligned}
\bE[ (S_n(P_tV_j))^2] = \sum_{x,y\in B(0,n)} \bE[P_t V_j(\tau_{0,x-y}\om) P_tV_j(\om)]= \sum_{x,y\in B(0,n)} \bE_{2,x-y}[V_j(\tau_{t,X_t}\om)V_j(\tau_{t,\tilde X_t}\om)].\end{aligned}$$ Let $\kappa>0$ to be chosen below. Then, for every $z\in \bZ^d$ we use Lemma \[lem:probY\] and Assumption \[ass:space\_mixing\] and obtain $$\begin{aligned}
\label{est:X_Xtilde}
\bE_{2,z}[V_j(\tau_{t,X_t}\om)V_j(\tau_{t,\tilde X_t}\om)]&\leq \bE_{2,z}[V_j(\tau_{t,X_t}\om)V_j(\tau_{t,\tilde X_t}\om)\indicator_{\{|\tilde X_t - X_t|>n^\kappa \}}]+ c \bP_{2,z}[ |Y_t| \leq n^\kappa ] \nonumber \\
&\leq\bE_{2,z} \left[ \bE_{2,z} \big[V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) \big| \,X_t, \tilde X_t \big] \indicator_{\{|\tilde X_t - X_t|>n^\kappa \}} \right] + c t^{-d/2} n^{\kappa d} \nonumber \\
&=\bE_{2,z} \left[ \bE \big[V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om)] \indicator_{\{|\tilde X_t - X_t|>n^\kappa \}} \right] + c t^{-d/2} n^{\kappa d}\nonumber \\
&\leq c\left( n^{-\kappa p_2} + t^{-d/2} n^{\kappa d} \right).\end{aligned}$$ Hence, $$\begin{aligned}
\label{est_Sn}
n^{-2d}\, \bE[ (S_n(P_tV_j))^2] \leq c\left( n^{-\kappa p_2} + t^{-d/2} n^{\kappa d}\right).\end{aligned}$$ Next we rewrite the Dirichlet form of the process $\eta$ as $$\begin{aligned}
\label{eq:DF_rewritten}
&\langle P_t V_j,P_t V_j \rangle_{\mathcal{H}_1}=\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ \mu_{0y}^\om(0) (P_tV_j(\tau_{0,y}\om) -P_tV_j(\om))^2 \right]\nonumber \\
=&\tfrac 1 2 \sum_{y\in\bZ^d} \bE \left[ \mu_{0y}^\om(0) \left( E^\om_{0,y}[V_j(\tau_{t,X_t}\om)]-E^\om_{0,0}[V_j(\tau_{t,X_t}\om)] \right)\left( E^\om_{0,y}[V_j(\tau_{t,\tilde X_t}\om)]-E^\om_{0,0}[V_j(\tau_{t,\tilde X_t}\om)] \right) \right]\nonumber \\
=& \tfrac 1 2 \sum_{y\in\bZ^d} \left( \bE_{2,y,y} [ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) ] - \bE_{2,y,0} [ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) ] \right. \nonumber \\
& \left.
- \bE_{2,0,y} [ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) ] + \bE_{2,0,0} [ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) ] \right).\end{aligned}$$ Then, by the time mixing in Assumption \[ass:time\_mixing\] we have $$\begin{aligned}
\bE_{2,y,y} \left[ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) \right]&= \bE_{2,y,y}\left[ \bE_{2,y,y} \big[ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) \big| \,X_t, \tilde X_t \big] \right] \\
&=\bE_{2,y,y}\left[ \bE \big[ \mu_{0y}^\om(0) V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) \big] \right] \\
&\leq \bE_{2,y,y} \left[ \bE \big[ \mu_{0y}^\om(0) \big] \bE\big[ V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) \big] \right] + c t^{-p_1} \\
& =\bE \big[ \mu_{0y}^\om(0) \big] \cdot \bE_{2,0,0}\left[ \bE\big[ V_j(\tau_{t,X_t}\om) V_j(\tau_{t,\tilde X_t}\om) \big] \right]+ c t^{-p_1} \\
&\leq c \left( n^{-\kappa p_2} + t^{-d/2} n^{\kappa d} + t^{-p_1} \right),\end{aligned}$$ where we also used Assumption \[ass:ergodic\] in the fourth step and in the last step. The other three terms in can be treated similarly, and we obtain that $$\begin{aligned}
\label{est:DF}
\langle P_t V_j,P_t V_j \rangle_{\mathcal{H}_1}\leq c \left( n^{-\kappa p_2} + t^{-d/2} n^{\kappa d} + t^{-p_1} \right).\end{aligned}$$ Note that by the ellipticity in Assumption \[ass:ellipticity\] for any $f\in \mathcal{D}(L)$ the Dirichlet form $\langle f,f \rangle_{\mathcal{H}_1}$ is comparable with the Dirichlet form of the environment process associated with a simple random walk on $\bZ^d$. Thus, by Proposition 3.2 in [@Mou], which is a simple consequence of the local Poincaré inequality on $\bZ^d$, there exists $C_S>0$ such that for any $f\in \mathcal{D}(L)$ and $n\in \mathbb{N}$, $$\bE[f(\om)^2]\leq C_S n^2 \langle f,f \rangle_{\mathcal{H}_1} + \frac 2 {|B(0,n)|^2} \bE[S_n(f)^2].$$ Combining this with and we get $$\begin{aligned}
\label{est_PtVj}
\bE[P_tV_j(\om)^2]&\leq C_S n^2 \langle P_tV_j,P_t V_j \rangle_{\mathcal{H}_1} + \frac 2 {|B(0,n)|^2} \bE[S_n(P_t V_j)^2]\nonumber \\
&\leq c \left( n^{2-\kappa p_2} + t^{-d/2} n^{2+\kappa d} + t^{-p_1} n^2 \right).\end{aligned}$$ By Assumption \[ass:space\_mixing\] we have $p_2> 2d/(d-2)$, so there exists $\delta>0$ such that $p_2>(1+\delta) 2d/(d-2)$. Now let $$\begin{aligned}
\kappa> \max \left(\frac {2+\frac{4(1+\delta)}{d-2}} {p_2-\frac{2(1+\delta)d}{d-2}} , \frac 1 d \left( \frac{d-2}{p_1-1}-2 \right) \right)
\intertext{and}
\varrho\in \left( \frac 1 {1+\delta} \frac{d/2-1}{\kappa d+2},\frac{d/2-1}{\kappa d+2} \right) .\end{aligned}$$ Finally, choosing $n=t^\varrho$ in gives .
### Proof of Lemma \[lem:asymp\_Pk\] under Assumptions \[ass:time\_mixing2\] and \[ass:space\_mixing2\]
Let $d\geq 1$ and assume that \[ass:ergodic\]-\[ass:ellipticity\], \[ass:time\_mixing2\] and \[ass:space\_mixing2\] hold. Notice first that $$\begin{aligned}
\begin{split}
\big\| \int_0^t P_s V_j \, ds \big\|_{L^2(\bP)}^2&=\int_0^t \int_0^t \bE \bigl[ \sum_{x,y\in \bZ^d} V_j(\tau_{r,x}\om)V_j(\tau_{s,y}\om) p^\om(0,0;r,x) p^\om(0,0;s,y) \bigr] \, dr \, ds\\
&=
\int_0^t \int_0^t \bE \left[ E^\om_{0,0}[V_j(\tau_{r,X_r}\om)] E^\om_{0,0}[V_j(\tau_{s,\tilde X_s}\om)]\right] \, dr \, ds \\
&=2 \int_0^t \int_0^t \indicator_{\{ r\leq s \}} \bE \left[ E^\om_{0,0}[V_j(\tau_{r,X_r}\om)] E^\om_{0,0}[V_j(\tau_{s,\tilde X_s}\om)]\right] \, dr \, ds
\end{split}\end{aligned}$$ and that by definition $V_j(\om)=\mu_{0,e_j}^\om(0)-\mu_{0,-e_j}^\om(0)$, so by Assumption \[ass:ergodic\] and \[ass:ellipticity\] we have $\bE [ V_j]=0$ and $\|V_j\|_{L^\infty(\bP)}\leq 2C_u$, respectively. Again it suffices to consider $t\geq 1$.
For any $0\leq r <s \leq t$ with $s-r \geq 1$ we have by Assumption \[ass:time\_mixing2\] $$\begin{aligned}
\bE E^\om_{0,0}[V_j(\tau_{r,X_r}\om)] E^\om_{0,0}[V_j(\tau_{s,\tilde X_s}\om)]
&=\bE_2 \left[ \bE_2 \big[V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \big| \,X_r, \tilde X_s \big] \right] \\
&=\bE_2 \bE \left[ V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \right] \\
&\leq c (s-r)^{-p_1}\end{aligned}$$ with $p_1>d+1$ if $d\geq 2$ and $p_1>4$ if $d=1$. We fix $\delta\in (1/p_1,1/(d+1))$ if $d\geq 2$ and $\delta\in (1/p_1,1/4)$ if $d=1$ and set $T_t:= t^\delta$ and $L_t:=(1\vee c_1) T_t$ (with constant $c_1$ as in Proposition \[prop:hke\]). Then, $$\begin{aligned}
\label{eq:kl_apart} \begin{split}
\int_0^t \int_0^t \indicator_{\{ r\leq s \}} \indicator_{\{s-r\geq T_t \}} \bE\left[ E_{0,0}^\om[V_j(\tau_{r,X_r}\om)] E_{0,0}^\om[V_j(\tau_{s,\tilde X_s}\om)] \right]\, dr \, ds&\leq c t^2 T_t^{-p_1} .
\end{split}\end{aligned}$$ Now we shall consider pairs of times $r$ and $s$ with distance less than $T_t$. We decompose the integral as follows. $$\begin{aligned}
\label{eq:kl_close}
&\int_0^t \int_0^t \indicator_{\{ r\leq s \}} \indicator_{\{ s-r< T_t \}} \bE \left[ E_{0,0}^\om[V_j(\tau_{r,X_r}\om)] E_{0,0}^\om[V_j(\tau_{s,\tilde X_s}\om)] \right]\, dr \, ds \nonumber \\
\leq &
\int_0^t \int_0^t \indicator_{\{ r\leq s \}} \indicator_{\{ s-r< T_t \}} \bE_2 \bigl[V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \indicator_{\{|X_r-\tilde X_r| >2 L_t\}} \bigr] \, dr \, ds \nonumber \\
&+ c T_t\int_0^t \bP_2[|X_r-\tilde X_r| \leq 2 L_t] \, dr \nonumber \\
\leq & \int_0^t \int_0^t \indicator_{\{ r\leq s \}} \indicator_{\{ s-r< T_t \}} \bE_2 \bigl[ V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \indicator_{\{|X_r-\tilde X_s| >L_t\}} \indicator_{\{|X_r-\tilde X_r| >2 L_t\}}\bigr] \, dr \, ds\nonumber \\
&+ c \int_0^t \int_0^t \indicator_{\{ r\leq s \}} \indicator_{\{ s-r< T_t \}} \bP_2[|X_r-\tilde X_s| \leq L_t, \, |X_r-\tilde X_r|> 2 L_t] \, dr \, ds \nonumber \\
& + c T_t \int_0^t \bE \pi^\om_{0,0} [|Y_r| \leq 2 L_t] \, dr.\end{aligned}$$ To estimate the first term in note that conditioned on the event $\{ |X_r-\tilde X_s| >L_t\}$ we have that $V_j(\tau_{r,X_r}\om)$ and $V_j(\tau_{s,\tilde X_s}\om)$ depend only on variables contained in two subsets of $\bZ^d$ with distance $L_t$. Thus, by Assumption \[ass:space\_mixing2\] we obtain $$\begin{aligned}
\label{est_term1}
&\bE_2 \left[ V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \indicator_{\{|X_r-\tilde X_s| >L_t\}} \indicator_{\{|X_r-\tilde X_r| >2 L_t\}} \right] \nonumber \\
=& \bE_2 \left[ \bE_2 \big[V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \big| \,X_r, \tilde X_r , \tilde X_s \big] \indicator_{\{|X_r-\tilde X_s| >L_t\}} \indicator_{\{|X_r-\tilde X_r| >2 L_t\}} \right] \nonumber \\
=&\bE_2 \left[ \bE \big[V_j(\tau_{r,X_r}\om) V_j(\tau_{s,\tilde X_s}\om) \big] \indicator_{\{|X_r-\tilde X_s| >L_t\}} \indicator_{\{|X_r-\tilde X_r| >2 L_t\}}\right] \nonumber \\
\leq& c L_t^{-p_2} .\end{aligned}$$
Next we estimate the second term in . First we use the Markov property to get $$\begin{aligned}
\bP_2[|X_r-\tilde X_s| \leq L_t, \, |Y_r|> 2 L_t]\leq \bP_2[|\tilde X_s-\tilde X_r | >L_t]=\bP P_{0,0}^\om P^\om_{r,\tilde X_r}[|\tilde X_s-\tilde X_r | >L_t].\end{aligned}$$ Set $D_i:=\{ y\in \bZ^d:\, 2^i L_t \leq |y-\tilde X_r | \leq 2^{i+1} L_t\}$, $i\geq 0$. Then, noting that $s-r \leq T_t\leq c_1^{-1} L_t$, we use the heat kernel estimates in Proposition \[prop:hke\] to obtain $$\begin{aligned}
P^\om_{r,\tilde X_r}[|\tilde X_s-\tilde X_r | >L_t]&=\sum_{y\in B(\tilde X_r,L_t)^c} p^\om (r,\tilde X_r;s,y)=\sum_{i=0}^\infty \sum_{y\in D_i} p^\om (r,\tilde X_r;s,y) \\
&\leq c \sum_{i=0}^\infty \sum_{y\in D_i} \exp\big(-c |y-\tilde X_r| \log(|y-\tilde X_r| / (s-r))\big) \\
&\leq c \sum_{i=0}^\infty (2^i L_t)^d \exp\big(-c 2^i L_t \log(2^iL_t/T_t)\big)\\
&\leq c \sum_{i=0}^\infty (2^i L_t)^d \exp\big(-c 2^i L_t \big).\end{aligned}$$ An elementary computation now gives $$\begin{aligned}
\label{est_term2}
\bP_2[|X_r-\tilde X_s| \leq L_t, \, |Y_r| > 2 L_t] &\leq c L_t^d \int_1^\infty \exp(-c L_t u )\, du \leq c L_t^{d-1} \exp\left(-c L_t \right). \end{aligned}$$ To estimate the last term in we use Lemma \[lem:probY\] to obtain in the case $d\geq 2$ $$\begin{aligned}
\label{est_term3}
\int_0^t \pi^\om_{0,0}[|Y_r| \leq 2 L_t] \, dr \leq c \int_0^t (1\vee r)^{-d/2} L_t^d \, dr \leq c L_t^d \int_0^t (1\vee r)^{-1} \, dr \leq c \log t \, L_t^d,\end{aligned}$$ and if $d=1$ $$\begin{aligned}
\label{est_term3d1}
\int_0^t \pi^\om_{0,0}[|Y_r| \leq 2 L_t] \, dr \leq c \int_0^t (1\vee r)^{-1/2} L_t^d \, dr \leq c t^{1/2} \, L_t.\end{aligned}$$
Finally, combining and we get in the case $d\geq 2$ by , and $$\begin{aligned}
\big\| \int_0^t P_s V_j \big\|_{L^2(\bP)}^2 \, ds \leq c \left(t^2 T_t^{-p_1} +t T_t L_t^{-p_2} + t T_t L_t^{d-1} \exp(-c {L_t})+T_t \log t \, L_t^d \right).\end{aligned}$$ Analogously, if $d=1$ we obtain by , and that $$\begin{aligned}
\big\| \int_0^t P_s V_j \big\|_{L^2(\bP)}^2 \, ds \leq c \left(t^2 T_t^{-p_1} +t T_t L_t^{-p_2} + t T_t \exp(-c {L_t})+ t^{1/2} T_t \, L_t \right).\end{aligned}$$ The claim follows by our choice of $\delta$, $L_t$ and $T_t$.
Invariance Principle for $X$ {#sec:pf_main}
============================
In this section we prove the annealed FCLT in Theorem \[annealed-ip\] and the quenched FCLT in Theorem \[main-ip\] and Theorem \[main-ip2\], respectively. Throughout this section we suppose that Assumptions \[ass:ergodic\]-\[ass:ellipticity\] hold. The first step to prove a quenched invariance principle for the random walk $X$ is to show that the processes $X^{(\eps)}$ are tight.
\[t-tight\] Let $T>0$, $r>0$. Then $$\begin{aligned}
& \lim_{R \to \infty} \sup_{0<\eps\leq 1}
P_{0,0}^\om( \sup_{s \le T} | X^{(\eps)}_s| > R ) \to 0, \\
& \lim_{\delta \to 0} \limsup_{\eps \to 0}
P_{0,0}^\om( \sup_{|s_1-s_2| \le \delta, s_i \le T }
| X^{(\eps)}_{s_2}- X^{(\eps)}_{s_1}| > r )= 0.\end{aligned}$$
From the heat kernel estimates in Proposition \[prop:hke\] one can derive tail estimates for the exit times of $X$ from balls (see e.g. [@ABDH Proposition 4.7]). Then tightness follows by the same arguments as in [@ABDH Proposition 5.13].
For $n \in \bN$ let $\wh X_n=X_n$, and set $$\wh X^{(\eps)}_t = \eps \wh X_{\lfloor t/\eps^2 \rfloor}, \q 0< \eps \le 1.$$
\[disc-approx\] For any $u>0$, $$\lim_{\eps \to 0} P_{0,0}^\om( \sup_{0\le s \le T}
| \wh X^{(\eps)}_s - X^{(\eps)}_s| > u ) =0.$$
This follows from the proof of Theorem \[t-tight\] by the same arguments as in [@BD Lemma 4.12].
We will first establish the convergence of the processes $\wh X^{(\eps)}$; using Lemma \[disc-approx\] will then give the convergence of $X^{(\eps)}$ to the same limit. We define $$\label{whM-def}
\wh M_n =M_n, \qq
\wh M^{(\eps)}_t = \eps \wh M_{ \lfloor t/\eps^2 \rfloor}, \q t \ge 0,$$ so that $$\label{xmchi}
\wh X^{(\eps)}_t = \eps \wh X_{ \lfloor t/\eps^2 \rfloor}=
\wh M^{(\eps)}_t + \eps \chi(\lfloor t/\eps^2 \rfloor, \eps^{-1} \wh X^{(\eps)}_t ,\om).$$ Thus it is sufficient to prove that the martingale $\wh M^{(\eps)}$ converges to a Brownian motion with a certain covariance matrix, and that the second term in converges to zero in $P_{0,0}^\om$-probability for $\bP$-a.a. $\om$ (resp. in $\bP^*_{0,0}$-probability) to get the quenched FCLT (resp. the annealed FCLT). For any $G:\bZ^d\times\Omega \rightarrow \bR$ we define $$\ol \bE [G]=\sum_{y\sim 0} \bE \left[ \mu^\om_{0y}(0) G(y,\om) \right].$$
\[P:mconv\] For $\bP$-a.e. $\om$, the sequence of processes $(\wh M^{(\eps)})$ converges in law in the Skorohod topology to a Brownian motion with a non-degenerate covariance matrix $\Sigma$ given by $\Sigma_{ij} = \obE \Phi_i \Phi_j$.
We proceed as in [@BB]. Let $v \in \bR^d$ be a unit vector, write as before $\wh M^v_n= v\cdot M_n$, and let $$F_K(\om) = E^0_\om (|\wh M^v_1|^2; |\wh M^v_1| \ge K ).$$ Then $F_K$ is decreasing in $K$, in particular $ \bE [F_K ]\leq \bE [F_0]$. In the notation of Corollary \[cor:Mbracket\] $F_0(\om)=\|v \cdot \Phi \|^2_\om$, and so by the covariance process of $\wh M^v$ is $$\langle \wh M^v \rangle_n = \int_0^n F_0( \eta_s) \, ds.$$ So by the ergodicity of the environment process $\eta$ w.r.t. $\bP$ we have $n^{-1} \langle \wh M^v \rangle_n \to \bE [F_0]$ as $n\to \infty$, $P_{0,0}^\om$ a.s., for $\bP$-a.a. $\om$.
Using the same arguments as in [@BB Theorem 6.2] it is straightforward to check the conditions of the Lindeberg-Feller FCLT for martingales (see for example [@Du Theorem 3.4.5]), and deduce that $v \cdot \wh M^{(\eps)}$ converges to a real-valued Brownian motion with non-random covariance $\bE[ \| v\cdot \Phi \|^2_\om ]$, which can be written as $v\cdot \Sigma v$, where $\Sigma$ is the matrix with coefficients given by $\Sigma_{ij} = \obE [\Phi_i \Phi_j]$. By the Cramer-Wold Theorem (see e.g. Theorem 3.9.5 in [@Du]) we get that $\wh M^{(\eps)}$ converges in law to an $\bR^d$-valued Brownian motion with covariance matrix $\Sigma$.
It remains to show that $\Sigma$ is non-degenerate. By the uniform lower bound on the conductances in Assumption \[ass:ellipticity\] we have for every unit vector $v\in\bR^d$ that $v\cdot \Sigma v\geq v\cdot \Sigma_{C_l} v$, where $\Sigma_{C_l}$ denotes the non-degenerate covariance matrix of the limiting Brownian motion in the invariance principle for the simple random walk on $\bZ^d$ with constant jump rate $C_l$. Thus, $v\cdot \Sigma v>0$, which implies that $\Sigma$ is positive-definite.
To conclude the proof of the invariance principles we need to control the corrector function. First we complete the proof of the annealed FCLT.
Setting $R_n:=\chi(n,X_n,\om)$ we need to show that $$\begin{aligned}
\label{conv_Rn_annealed}
n^{-1/2} \max_{k\leq n} |R_k| \rightarrow 0 \qquad \mbox{in $\bP^*_{0,0}$-probability as $n\to \infty$.}\end{aligned}$$ By Proposition \[prop:conv\_corr1\] we have that $n^{-1/2} R_n$ converges to $0$ in $L^2(\bP^*_{0,0})$ and thus in $\bP^*_{0,0}$-probability. By an elementary property of real convergent sequences, we get .
Finally, to complete the proof of the quenched invariance principle we prove
Let $T>0$. Under the assumptions of Theorem \[main-ip\] or Theorem \[main-ip2\], for $\bP$-a.e. $\om$, $$\sup_{t\leq T}\eps \chi(\lfloor t/\eps^2 \rfloor, \eps^{-1} \wh X^{(\eps)}_t ,\om)\rightarrow 0 \qquad \mbox{in $P^\om_{0,0}$-probability.}$$
We will proceed as in [@RS] applying the theory of “fractional coboundaries” of Derriennic and Lin [@DL]. Setting $R_n:=\chi(n,X_n,\om)$ we need to show that $$\begin{aligned}
\label{conv_Rn}
n^{-1/2} \max_{k\leq n} |R_k| \rightarrow 0 \qquad \mbox{in $P_{0,0}^\om$-probability as $n\to \infty$.}\end{aligned}$$
Let $\tilde \bP$ denote the path measure on $\Omega^{\bN}$ of the random sequence $(\tau_{n,X_n}\om)_{n\geq 0}$ with initial distribution $\bP$, and let $\theta$ be the shift map on the sequence space $\Omega^{\bN}$. By the cocycle property in Corollary \[cor:cocycle\] we have $\chi(0,0,\om)=0$ and hence $$\begin{aligned}
R_n&=\sum_{k=0}^{n-1} \chi(k+1,X_{k+1},\om)-\chi(k,X_k,\om) =\sum_{k=0}^{n-1}h(\tau_{k,X_k}\om, \tau_{k+1,X_{k+1}}\om)\end{aligned}$$ with $h$ defined as in Remark \[rem:defh\]. For sequences $\bar \om=(\om^{(i)})_{i\in \bN}$ define $H(\bar \om)=h(\om^{(0)},\om^{(1)})$ and $$\tilde R_n=\sum_{k=0}^{n-1} H\circ \theta^k.$$ Then $H\in L^2(\tilde\bP)$ and the process $(\tilde R_n)$ has the same distribution under $\tilde \bP$ as the process $(R_n)$ under $\bP\otimes P_{0,0}^\om$.
By Proposition \[prop:asymp\_corr\] the assumptions of Theorem 2.17 in [@DL] are satisfied. We conclude that $H\in (I-\theta)^\gamma L^2(\tilde \bP)$ for any $\gamma\in (0,1-\alpha)$. Since $\alpha<1/2$ there exists such a $\gamma \in (1/2,1-\alpha)$. Then, (i) in Theorem 3.2 in [@DL] implies that $n^{-1/2} \tilde R_n$ converges to $0$, $\tilde\bP$-a.s. Hence, $n^{-1/2} R_n$ converges to $0$, $\bP\otimes P_{0,0}^\om$-a.s. In other words, $n^{-1/2} R_n$ converges to $0$, $P_{0,0}^\om$-a.s., for $\bP$-a.e. $\om$, which implies .
Application to Stochastic Interface Models {#sec:interface}
==========================================
In this section we point out a relation between our results and the stochastic dynamic of an interface describing the separation of two pure thermodynamical phases, known as the Ginzburg Landau $\nabla\phi$ model. We refer to [@F] for a survey on these models. The interface is described by a field of height variables $\phi_t(x)$, $x\in \mathbb{Z}^d$, $t\geq 0$, whose stochastic dynamics are given by the following infinite system of stochastic differential equations involving nearest neighbour interaction: $$\begin{aligned}
\label{phi_dyn}
\phi_t(x)=\phi_x-\int_0^t \sum_{y:|x-y|=1} V'(\phi_t(x)-\phi_t(y)) \, dt + \sqrt{2} w_t(x), \qquad x\in \mathbb{Z}^d.\end{aligned}$$ Here $\phi$ is the height of the interface at time $t=0$, $\{w(x), x\in \mathbb{Z}^d\}$ is a collection of independent Brownian motions and the potential $V\in C^2(\mathbb{R},\mathbb{R}_+)$ is even and strictly convex, i.e.$$\begin{aligned}
\label{Vconvex}
c_-\leq V''\leq c_+,\end{aligned}$$ for some $0<c_-<c_+<\infty$. Let for each $r>0$ $$\sE_r:=\{ \phi\in\bR^{\bZ^d}:\, \sum_x | \phi_x|^2 e^{-2 r |x|} <\infty\}$$ denote the set of tempered configurations. Then, for every initial value $\phi\in \sE_r$ the SDE admits a unique strong solution $\phi_t \in \sE_r$, $t\geq 0$, see [@FS]. Let $H$ be the formal Hamiltonian given by $$H(\phi)=\tfrac 1 2 \sum_{y:|x-y|=1} V(\phi_x-\phi_y),$$ then the formal equilibrium measure for the dynamic is given by the Gibbs measure $$\frac 1 Z \exp(-H(\phi)) \prod_x d\phi_x.$$ This can be made rigorous for the corresponding dynamic on a finite box. In dimension $d\geq 3$ Gibbs measures for the $\phi$-field on the whole lattice can be constructed by taking the thermodynamical limit, cf. Section 4.5 in [@F]. More precisely, for every $h\in \bR$ there exists a shift-invariant and ergodic $\phi$-Gibbs measure $m_h$ with mean $h$, i.e.$$\int \phi_x \, m_h(d\phi)=h, \qquad x\in\bZ^d.$$ These measures are also reversible and ergodic for the SDE . We denote by $\bP_{m_h}$ the law of the process $\phi_t$ started under the equilibrium distribution $m_h$ (and by $\bE_{m_h}$ the corresponding expectation).
Next we consider discrete gradients, i.e. height differences of the form $\nabla_b \phi=\phi_{y_b}-\phi_{x_b}$ for any bond $b=\{x_b,y_b\} \in E_d$. Then, as a vector field $\nabla \phi$ has zero curl in the sense that $$\sum_{b\in\sC}\nabla_b\phi=0$$ for every closed loop $\sC$, i.e. the bonds $\{x_i,x_{i+1}\}$ of a sequence of $x_0,\ldots,x_n$ in $\bZ^d$ satisfying $x_0=x_n$ and $|x_i-x_{i-1}|=1$ for $i\in\{1,\ldots n\}$. Let $\sX$ be the subset of $\bR^{E_d}$, whose elements have zero curl, and let for $r>0$ $$\sX_r=\{ \eta\in \bR^{E_d}: \, \eta_b=\nabla_b \phi \mbox{ for some } \phi\in\sE_r \}$$ be the subset of tempered gradients. Note that the drift term in the SDE can be rewritten as $$-\sum_{y:|x-y|=1} V'(\phi(x)-\phi(y))=\sum_{b:\, x_b=x} V'(\nabla_b \phi).$$ Then, for each initial $\nabla\phi \in \sX_r$, the gradient process $(\nabla_b\phi_t, b\in E_d, t\geq 0)$ is the unique strong solution of the SDE $$\begin{aligned}
\nabla_b\phi_t=\nabla_b\phi-\int_0^t \left( \sum_{b': x_{b'}=x_b} V'(\nabla_{b'}\phi_s)-\sum_{b': x_{b'}=y_b} V'(\nabla_{b'}\phi_s) \right) \, ds+\sqrt 2 \nabla_b w_t, \qquad b\in E_d, \end{aligned}$$ where $ \nabla_b w_t=w_t(y_b)-w_t(x_b)$, see again [@FS]. Also it has been shown in [@FS Theorems 3.1 and 3.2] that in any lattice dimension $d\geq 1$, given any $u\in \bR^d$, there exists a unique shift invariant ergodic $\nabla \phi$-Gibbs measure $\tilde m_u$ on $\sX_r$ satisfying $$\int_{\sX} \eta_{0,e_i} \, d\tilde m_u=u_i,$$ for every $i=1,\ldots, d$. Here $u$ is the tilt and $\tilde m_u$ the $u$-tilted measure. Moreover, $\tilde m_u$ is known to be an invariant reversible and ergodic measure for the gradient process $\nabla\phi_t$ ([@FS Proposition 3.1]).
Our aim is to investigate the decay of the space-time correlation functions under the equilibrium Gibbs measures. The idea – originally from Helffer and Sjöstrand [@HS] – is to describe the correlation functions in terms of a certain random walk in dynamic random environment (cf. also [@DD; @GOS; @BG]). Let $(X_t)_{t\geq 0}$ be the random walk on $\bZ^d$ with jump rates given by the random dynamic conductances $$\begin{aligned}
\mu^{ \nabla \phi}_{b}(t):=V''(\nabla_b \phi(t))=V''(\phi_{y_b}(t)-\phi_{x_b}(t)), \qquad b=\{x_b,y_b\}\in E_d.\end{aligned}$$ Since $V$ is even, the jump rates are symmetric, i.e. $\mu^{\nabla\phi}_{x_b,y_b}(t)=\mu^{\nabla\phi}_{y_b,x_b}(t)$. Further, let $p^{\nabla\phi}(s,x;t,y)$, $x,y\in \bZ^d$, $s\leq t$, denote the transition densities of the random walk $X$.
1. Let $d\geq 3$ and let $m_h$ be any ergodic $\phi$-Gibbs measure. Then, the environment $\mu^{\nabla\phi}$ started under $m_h$ satisfies Assumptions \[ass:ergodic\]-\[ass:ellipticity\]. Moreover, $\mu^{\nabla\phi}$ also satisfies Assumptions \[ass:time\_mixing\] and \[ass:space\_mixing\] if $d\geq 6$.
2. Let $d\geq 1$ and let $\tilde m_u$ be any ergodic $\nabla\phi$-Gibbs measure. Then, the environment $\mu^{\nabla\phi}$ started under $\tilde m_u$ satisfies Assumptions \[ass:ergodic\]-\[ass:ellipticity\]. Moreover, $\mu^{\nabla\phi}$ also satisfies Assumptions \[ass:time\_mixing\] and \[ass:space\_mixing\] if $d\geq 5$.
Assumption \[ass:ergodic\] is immediate from the ergodicity of the Gibbs measures $m_h$ and $\tilde m_u$, respectively. Assumption \[ass:stoch\_cont\] is clear from the pathwise continuity of $\phi_t$ and $\nabla \phi_t$ and the strict convexity of $V$ in guarantees the ellipticity in Assumption \[ass:ellipticity\].
By Theorem 6.1 in [@DD] the time-covariance under the $\phi$-Gibbs measure $m_h$ decays polynomially with order $d/2-1$ and the space-covariance decays polynomially with order $d-2$. Hence, Assumptions \[ass:time\_mixing\] and \[ass:space\_mixing\] hold if $d/2-1>1$ and $d-2>2d/(d-2)$ which is the case for $d\geq 6$.
On the other hand, by Theorem 6.2 in [@DD] the time-covariance for $\nabla \phi$ has polynomial decay of order $d/2$ and the space-covariance has polynomial decay of order $d$ . We have $d/2>1$ and $d>2d/(d-2)$ if $d\geq 5$.
We combine now the Helffer-Sjöstrand representation and the local limit theorem in Theorem \[thm:llt\] to get a scaling limit for the space-time covariation of the $\phi$-field.
Let $d\geq 3$ and let $m_h$ be any ergodic $\phi$-Gibbs measure. Then, there exist a non-degenerate covariance matrix $\Sigma$ such that $$\begin{aligned}
N^{d+2} \cov_{m_h}(\phi_0(0), \phi_{N^2t}(\lfloor Ny\rfloor)) \rightarrow \int_0^\infty k^{(\Sigma)}_{t+s}(y)\, ds, \qquad \text{as $N\to \infty$,}\end{aligned}$$ where $k_t$ is the Gaussian kernel with diffusion matrix $\Sigma$ in .
By the Helffer-Sjöstrand representation (cf. equation (6.10) in [@DD]) we have $$\cov_{m_h}(\phi_0(0), \phi_t(y))=\int_0^\infty \mathbb{E}_{m_h} \left[ p^{\nabla\phi}(0,0;t+s,y) \right] ds.$$ Using the annealed local limit theorem in Theorem \[thm:llt\] i) we obtain $$\begin{aligned}
N^{d+2} \cov_{m_h}(\phi_0(0), \phi_{N^2t}\left(\lfloor Ny \rfloor)\right)=&N^d \int_0^\infty \mathbb{E}_{m_h}\left[ p^{\nabla\phi}\big(0,0;N^2(t+s),\lfloor Ny\rfloor\big) \right] ds \\
\rightarrow & \int_0^\infty k^{(\Sigma)}_{t+s}(y)\, ds\end{aligned}$$ as $N\to \infty$, which is the claim.
Ultimately, we would like to derive an analogous scaling limit for the space-time covariance of the gradient process $\nabla \phi_t$, see also the discussion in [@BG Section 6]. However, what is still missing until now is a local limit theorem for the gradient of the heat kernel.
S. Andres, M.T. Barlow, J.-D. Deuschel and B. Hambly. Invariance principle for the random conductance model, Preprint to appear in [*Probab. Theory Rel. Fields*]{}, available online DOI 10.1007/s00440-012-0435-2.
A. Bandyopadhyay and O. Zeitouni. Random walk in dynamic [M]{}arkovian random environment. [*ALEA, Lat. Am. J. Probab. Math. Stat.*]{} [**1**]{} (2006), 205–224.
M.T. Barlow and J.-D. Deuschel. Invariance principle for the random conductance model with unbounded conductances. [*Ann. Probab.*]{} [**38**]{} (2010), 234-276
M.T. Barlow and B.M. Hambly. Parabolic Harnack inequality and local limit theorem for percolation clusters. [*Elec. J. Prob.*]{} [**14**]{} (2009), Paper 1, 1-16.
N. Berger and M. Biskup. Quenched invariance principle for simple random walk on percolation clusters. [*Probab. Theory Rel. Fields*]{} [**137**]{} (2007), no. 1-2, 83–120.
M. Biskup. Recent progress on the random concuctance model. [*Probab. Surv.*]{} [**8**]{} (2011), 294-–373.
M. Biskup and T.M. Prescott. Functional CLT for random walk among bounded random conductances. [*Elec. J. Prob.*]{} [**12**]{} (2007), paper 49, 1323–1348.
T. Bodineau and B. Graham. Helffer-Sjöstrand representation for conservative dynamics. [*Markov Proc. Rel. Fields*]{} [**18**]{} (2012), 71–88.
C. Boldrighini, R. A. Minlos and A. Pellegrinotti. Random walks in quenched i.i.d. space-time random environment are always a.s. diffusive. [*Probab. Theory Related Fields*]{} [**129**]{} (2004), no. 1, 133–156.
C. Boldrighini, R. A. Minlos and A. Pellegrinotti. Discrete-time random motion in a continuous random medium. [*Stochastic Process. Appl.*]{} [**119**]{} (2009), no. 10, 3285–3299.
T. Delmotte and J.-D. Deuschel. On estimating the derivatives of symmetric diffusions in stationary random environment, with applications to [$\nabla\phi$]{} interface model. [*Probab. Theory Related Fields*]{} [**133**]{} (2005), no. 3, 358–390.
Y. Derriennic and M. Lin. Fractional Poisson equations and ergodic theorems for fractional coboundaries. [*Israel J. Math.*]{} [**123**]{} (2001), 93–130.
D. Dolgopyat and C. Liverani. Non-perturbative approach to random walk in [M]{}arkovian environment. [*Electron. Commun. Probab. *]{}[**14**]{} (2009), 245–251.
R. Durrett. Probability: Theory and Examples (4th Edition), Cambridge University Press, 2010.
S. Ethier and T. Kurtz. Markov processes, Wiley Series in Probability and Mathematical Statistics, New York, 1986.
T. Funaki. Stochastic Interface Models. Ecole d’ été de probabilités de Saint Flour 2003. Lecture Notes in Mathematics, 1869. Springer-Verlag, Berlin, 2005.
T. Funaki, H. Spohn. Motion by mean curvature from the Ginzburg-Landau $\nabla \phi$ interface models. [*Commun. Math. Phys.*]{} [**185**]{} (1997), 1–36.
G. Giacomin, S. Olla and H. Spohn. Equilibrium fluctuations for $\nabla \varphi$ interface model. [*Ann. Probab. *]{}[**29**]{} (2001), 1138–1172.
B. Helffer and J. Sjöstrand. On the correlation for Kac-like models in the convex case. [*J. Stat. Phys.*]{} [**74**]{} (1994), 349–409.
M. Joseph and F. Rassoul-Agha. Almost Sure Invariance Principle for Continuous-Space Random Walk in Dynamic Random Environment. [*ALEA, Lat. Am. J. Probab. Math. Stat.*]{} [**8**]{} (2011), 43–57.
T. Komorowski, C. Landim and S. Olla. Fluctuations in Markov processes. Time symmetry and martingale approximation. Grundlehren der Mathematischen Wissenschaften, 345. Springer, Heidelberg, 2012.
P. Mathieu. Quenched invariance principles for random walks with random conductances. [*J. Stat. Phys.*]{} [**130**]{} (2008), no. 5, 1025–1046.
M. Maxwell and M. Woodroofe. Central limit theorems for additive functionals of Markov chains. [*Ann. Probab.*]{} [**28**]{} (2000), 713–724.
J.-C. Mourrat. Variance decay for functionals of the environment viewed by the particle. [*Ann. Inst. Henri Poincaré Probab. Stat.*]{} [**47**]{} (2011), no. 1, 294–327.
F. Rassoul-Agha and T. Sepp[ä]{}l[ä]{}inen. An almost sure invariance principle for random walks in a space-time random environment. [*Probab. Theory Related Fields*]{} [**133**]{} (2005), no. 3, 299–314.
F. Rassoul-Agha and T. Sepp[ä]{}l[ä]{}inen. Almost sure functional central limit theorem for ballistic random walk in random environment. [*Ann. Inst. H. Poincaré Probab. Statist.*]{} [**45**]{} (2009), no. 2, 373–420.
F. Redig and F. Völlering, Limit theorems for random walks in dynamic random environment. Preprint, available at arXiv:1106.4181v2.
W. Rudin. Functional Analysis. McGraw-Hill Series in Higher Mathematics, 1973.
V. Sidoravicius and A.-S. Sznitman. Quenched invariance principles for walks on clusters of percolation or among random conductances. [*Probab. Theory Rel. Fields*]{} [**129**]{} (2004), no. 2, 219–244.
SA: Institut für Angewandte Mathematik\
Rheinische Friedrich-Wilhelms-Universität Bonn\
Endenicher Allee 60, 53115 Bonn, Germany.\
[email protected]
[^1]: Research partially supported by NSERC (Canada)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a simple algorithm to implement the generalized derivative expansion introduced previously by L-H. Chan, and apply it to the calculation of the one-loop mass correction to the classical soliton mass in the 1+1 dimensional Jacobi model. We then show how this derivative expansion approach implies that the total (bosonic plus fermionic) mass correction in an N=1 supersymmetric soliton model is determined solely by the asymptotic values (and derivatives) of the fermionic background potential. For a static soliton the total mass correction is $-m/(2\pi)$, in agreement with recent analyses using phase-shift methods.'
address: 'Department of Physics, University of Connecticut, Storrs CT 06269, USA'
author:
- 'Gerald V. Dunne'
title: Derivative Expansion and Soliton Masses
---
=8.5in
.5cm
The calculation of quantum corrections to classical soliton masses is a key ingredient in the semiclassical approach to quantum field theory [@rajaraman; @rebbi]. This question has been re-addressed recently [@top; @jaffe] for $1+1$ dimensions using topological boundary conditions and phase-shift methods. In this Letter I compute the quantum mass correction for solitons in the recently introduced Jacobi model [@dunne], using a simple algorithm based on the derivative expansion [@aitchison; @chan]; and I apply this derivative expansion approach to the total mass correction in an $N=1$ SUSY model in $1+1$ dimensions.
Consider the following Lagrangian for a real scalar field $\phi$ in 1+1 dimensions: $$\begin{aligned}
{\cal L}=\frac{1}{2}(\partial_\mu \phi)(\partial^\mu\phi)
- V(\phi)
\label{blag}\end{aligned}$$ We shall consider the examples: $$\begin{aligned}
V(\phi)=\cases{\frac{m^4}{8\lambda}\left[\left(\frac{\sqrt{\lambda}}{m}
\phi\right)^2 -1\right]^2\qquad ; \phi^4 {\rm model}\cr
\frac{m^4}{\lambda}\left[1-\cos\left(\frac{\sqrt{\lambda}}{m}\phi\right)
\right]\qquad ; {\rm SineGordon\,\, model}\cr
\frac{2 m^4}{\lambda}\, {\rm sn}^2\left(\frac{\sqrt{\lambda}}{2m}\phi| \nu\right) \qquad ; {\rm Jacobi \,\,model}}
\label{pots}\end{aligned}$$ The $\phi^4$ and Sine-Gordon models are text-book cases [@rajaraman], while the Jacobi potential has recently been studied in the context of an exactly solvable model of quantum mechanical instantons [@dunne]. Since quantum mechanical instantons have the same functional form as solitons in 1+1 dimensions, some of the results of [@dunne] carry over here. The function ${\rm sn}(z|\nu)$ is one of the Jacobi elliptic functions [@ww], and $0\leq\nu\leq 1$ is the (real) elliptic parameter. This parameter $\nu$ controls the shape and period of the Jacobi potential, as illustrated in Figure 1. The Jacobi model is a deformation of the Sine-Gordon model, reducing smoothly to Sine-Gordon as $\nu\to 0$ \[note that ${\rm sn}(\frac{\phi}{2}|\nu=0)=\sin(\frac{\phi}{2})$\].
As is well known [@rajaraman], the mass $m$ and coupling $\lambda$ may be scaled out, so that the semiclassical loop expansion is an expansion in powers of $\frac{\hbar\lambda}{m^2}$. With this understood, we can set $m=\lambda=1$ in our one-loop calculations.
These models have classical static soliton/antisoliton solutions $\phi_c(x)$ satisfying $$\begin{aligned}
\phi_c^\prime(x)=\pm\sqrt{2 V(\phi_c(x))}
\label{soliton}\end{aligned}$$ For the potentials in (\[pots\]) the classical solitons are $$\begin{aligned}
\phi_c(x)=\cases{{\rm tanh}(\frac{x}{2})\qquad ; \phi^4 {\rm model}\cr
4\, {\rm arctan}(e^x)\qquad ; {\rm SineGordon\,\, model}\cr
2K(\nu)+2\,{\rm sn}^{-1}({\rm tanh}(x))\qquad ; {\rm Jacobi \,\,model}}
\label{sols}\end{aligned}$$ Here $K(\nu)=\int_0^{\pi/2}d\theta/\sqrt{1-\nu\sin^2\theta}$, is the elliptic quarter period [@ww]. Note that $K(0)=\pi/2$, and $K(\nu)\sim\frac{1}{2}\log(\frac{16}{1-\nu})$ as $\nu\to 1$.
-1cm -2.5in
The classical mass of the soliton is given, in units of $\frac{m^3}{\lambda}$, by $$\begin{aligned}
M_{\rm classical}=\int_{\rm min}d\phi\sqrt{2 V(\phi)}
=\cases{\frac{2}{3}\qquad ;\phi^4 {\rm model}\cr
8 \qquad ; {\rm SineGordon\,\, model}\cr
\frac{4}{\sqrt{\nu}}\log\left(\frac{1+\sqrt{\nu}}{1-\sqrt{\nu}}\right)
\qquad ; {\rm Jacobi \,\,model}}
\label{cmasses}\end{aligned}$$ where the integration is between the two neighboring minima of $V(\phi)$ between which the soliton interpolates.
-1cm -2.5in
In units of $m$, the one-loop quantum correction to this classical soliton mass is $$\begin{aligned}
M=\frac{1}{2}\int \frac{d\omega}{2\pi}\,{\rm tr}\,
\log\left[\omega^2-\frac{d^2}{dx^2}+V^{\prime\prime}(\phi_c(x))\right]
\label{oneloop}\end{aligned}$$ This expression must be suitably renormalized, but once this is done $M$ is a pure number. Here $V^{\prime\prime}(\phi_c(x))$ means $\frac{d^2}{d\phi^2}V(\phi_c(x))$. For the cases listed above, the bosonic fluctuation potential $V^{\prime\prime}(\phi_c(x))$ is $$\begin{aligned}
V^{\prime\prime}(\phi_c(x))=\cases{1-\frac{3}{2}\, {\rm
sech}^2(\frac{x}{2})\qquad ;\phi^4 {\rm model}\cr 1-2\,{\rm
sech}^2(x)\qquad ; {\rm SineGordon\,\, model}\cr
(1-\nu)\left({-1+2(1-\nu){\rm tanh}^2(x)+\nu\,{\rm tanh}^4(x)\over (1-\nu\,
{\rm tanh}^2(x))^2}\right)
\qquad ; {\rm Jacobi \,\,model}}
\label{fl}\end{aligned}$$ For the $\phi^4$ and Sine-Gordon cases, the fluctuation potentials in (\[fl\]) are of the exactly solvable Pöschl-Teller form, which means that the quantum mass correction in (\[oneloop\]) can be computed exactly. The fluctuation potential for the Jacobi model is shown in Figure 2; its spectral properties are not known exactly, and so we need an approximate method to compute the mass correction (\[oneloop\]). Here we use the derivative expansion [@aitchison; @chan].
To implement the derivative expansion we first note that $V^{\prime\prime}(\phi_c(x))\to 1$ as $x\to\pm\infty$, for the fluctuation potentials appearing in (\[oneloop\]). Moreover, this background value of $1$ (equal to $m^2$ with mass scales reinstated) is approached exponentially fast. So it is more natural to expand in terms of the difference $$\begin{aligned}
W(x)=V^{\prime\prime}(\phi_c(x))- 1
\label{exp}\end{aligned}$$ and its $x$ derivatives. This leads to a generalized derivative expansion [@chan] in which (\[oneloop\]) is rewritten and expanded as $$\begin{aligned}
M[W]&=&-\int_{-\infty}^\infty\frac{d\omega}{2\pi}\,
\omega^2\, {\rm tr}\left(
\frac{1}{\omega^2+1-\frac{d^2}{dx^2}+W(x)}\right)\nonumber\\ &=&
\int_{-\infty}^\infty\frac{d\omega}{2\pi}\, \omega^2\, z^{\frac{3}{2}}\,
A(z,W]
\label{deriv}\end{aligned}$$ where $z\equiv 1/(\omega^2+1)$. In the derivative expansion approach [@aitchison; @chan], the expression (\[deriv\]) is renormalized by dropping a divergent term that is independent of $W$ (this cancels in taking the difference with the $W=0$ case), and a divergent term linear in $W$ (this corresponds to making the bosonic tadpole graph vanish, as in [@top; @jaffe]). Then $A(z,W]$ is defined by the expansion $$\begin{aligned}
A(z,W]=-\frac{1}{2}\,z\, \sum_{n=0}^\infty z^n\, a_n[W]
\label{func}\end{aligned}$$ where the expansion coefficients $a_n[W]$ are functionals of $W(x)$, the first few of which are [@chan]: $$\begin{aligned}
a_0[W]&=&\frac{3}{8}\int_{-\infty}^\infty W^2\,
dx\nonumber\\ a_1[W]&=&-\frac{5}{32}\int_{-\infty}^\infty \left\{
2W^3+(W^\prime)^2\right\}\, dx\nonumber\\
a_2[W]&=&\frac{7}{128}\int_{-\infty}^\infty \left\{
5W^4+10W(W^\prime)^2+(W^{\prime\prime})^2\right\} \, dx\nonumber\\
a_3[W]&=&-\frac{9}{512}\int_{-\infty}^\infty \left\{14W^5+70
W^2(W^\prime)^2+14W(W^{\prime\prime})^2+(W^{\prime\prime\prime})^2\right\}
\, dx
\label{as}\end{aligned}$$ Given the expansion (\[func\]) for $A(z,W]$, the $\omega$ integrals in (\[deriv\]) can be performed to yield a simple formula for the soliton mass correction: $$\begin{aligned}
M[W]=-\frac{1}{8\sqrt{\pi}}\,\sum_{n=0}^\infty
\frac{\Gamma(n+1)}{\Gamma(n+5/2)}\, a_n[W]
\label{formula}\end{aligned}$$ Thus, the derivative expansion computation of the soliton mass correction reduces to the calculation of the $a_n[W]$ in (\[as\]) for the given $W(x)$ in (\[exp\]). We stress again that each of these $a_n[W]$ is a pure number [@comment].
The functionals $a_n[W]$ can be obtained by a direct expansion of the resolvent in (\[deriv\]), as in [@chan]. However, it is easier computationally to note that these functionals have simple recurrence relations (these functionals also arise in asymptotic and WKB expansions of discriminants [@braden; @bd], and in zeta function analysis [@konoplich]). In fact, $$\begin{aligned}
a_n[W]=(-1)^{n+1} (2n+3)\, \int_{-\infty}^\infty
r_{2n+4}(x)\, dx
\label{rs}\end{aligned}$$ where the functions $r_n(x)$ are defined by the following recurrence relation: $$\begin{aligned}
r_0(x)&=&1\quad,\quad r_1(x)=0\quad,\quad
r_2(x)=-\frac{1}{2}W(x)\quad, \nonumber\\ r_n(x)&=&\frac{i}{2}\,
r_{n-1}^\prime(x)-\frac{1}{2} \sum_{j=1}^{n-1}r_j(x)\, r_{n-j}(x)\quad,\quad n\geq 3
\label{recurrence}\end{aligned}$$ This algorithm for generating the $r_{2n+4}(x)$ functions is simple to implement using, for example, Mathematica or Maple. Furthermore, all the integrals in (\[rs\]) can be performed analytically for each of the three soliton models listed in (\[pots\]).
So far, this algorithm could be applied to any suitable potential $W(x)$ in (\[deriv\]), with $W(x)\to 0$ sufficiently quickly as $x\to\pm\infty$. If the background is that of a soliton $\phi_c(x)$ satisfying (\[soliton\]), so that $W(x)$ is given by (\[exp\]), it is more convenient to change variables in the integration in (\[rs\]) from $x$ to $\phi_c$, using (\[soliton\]). For example: $$\begin{aligned}
a_0[W]&=&\frac{3}{8}\int_{min}\frac{d\phi}{\sqrt{2V(\phi)}}\, W^2
\nonumber\\
a_1[W]&=&-\frac{5}{32}\int_{min}\frac{d\phi}{\sqrt{2V(\phi)}}\, \left\{
2W^3+2V(V^{(3)})^2\right\}\nonumber\\
a_2[W]&=&\frac{7}{128}\int_{min}\frac{d\phi}{\sqrt{2V(\phi)}}\, \left\{
5W^4+20WV(V^{(3)})^2+(V^{\prime}V^{(3)}+2 VV^{(4)})^2\right\}\nonumber\\
a_3[W]&=&-\frac{9}{512}\int_{min}\frac{d\phi}{\sqrt{2V(\phi)}}\,
\left\{14W^5+140W^2V(V^{(3)})^2+14W(V^\prime V^{(3)}+2V V^{(4)})^2\right. \nonumber\\
&&\hskip 3cm\left.+2V(V^{\prime\prime}V^{(3)} +3V^\prime V^{(4)}+2VV^{(5)})^2\right\}
\label{revas}\end{aligned}$$ These modified expressions for $a_n[W]$ are straightforward to generate, either by converting the expressions in (\[as\]), or by a new recurrence relation based on (\[recurrence\]). In (\[revas\]), both $V$ and $W=V^{\prime\prime}-1$ are treated as functions of $\phi$, and the integration is between the neighboring minima of $V(\phi)$ between which the soliton interpolates, just as in the standard calculation (\[cmasses\]) of the classical mass. The advantages of this form (\[revas\]) are : (i) the only input is the original potential $V(\phi)$; (ii) the integrals are easier than those in (\[as\]); (iii) it organizes the expansion (\[formula\]) as a derivative expansion, as each $a_n$ in (\[revas\]) involves only terms with $(2n+4)$ derivatives of $V$ with respect to $\phi$.
The derivative expansion calculation of the soliton mass correction (\[formula\]) can now be programmed as a simple algorithm: generate the expressions in (\[as\]) or (\[revas\]) using recursion relations, do the integrals, and then insert the $a_n$ into (\[formula\]). We first test this algorithm on the well-known solvable cases. For the Sine-Gordon model one finds $a_n^{\rm SG}=2$ for all $n$, confirming a result in [@chan]. Then, using our formula (\[formula\]), we immediately obtain the standard result [@rajaraman; @rebbi] $$\begin{aligned}
M^{\rm SG}= -\frac{1}{4\sqrt{\pi}}\,\sum_{n=0}^\infty
\frac{\Gamma(n+1)}{\Gamma(n+5/2)}=-\frac{1}{\pi}
\label{sgmass}\end{aligned}$$ For the $\phi^4$ model, one finds $a_n^{\phi^4}=2+(1/4)^{n+1}$ for all $n$, again confirming a result in [@chan]. Then, from (\[formula\]) we find the standard result [@rajaraman; @rebbi] $$\begin{aligned}
M^{\rm \phi^4}=-\frac{1}{\pi}
-\frac{1}{32\sqrt{\pi}}\,\sum_{n=0}^\infty
\frac{\Gamma(n+1)}{4^n\Gamma(n+5/2)}=\frac{1}{4\sqrt{3}}-\frac{3}{2\pi}
\label{ffmass}\end{aligned}$$ One can speed up the convergence of the derivative expansion by separating out the contribution from the zero mode of the fluctuation potential [@chan]. This is achieved by writing $A(z,W]=(A(z,W]+\frac{1}{\omega^2})-\frac{1}{\omega^2}$ in (\[deriv\]) and (\[func\]), which separates out a $-\frac{1}{\pi}$ term, so that the mass correction (\[formula\]) becomes $$\begin{aligned}
M[W]=-\frac{1}{\pi}-\frac{1}{8\sqrt{\pi}}\,\sum_{n=0}^\infty
\frac{\Gamma(n+1)}{\Gamma(n+5/2)}\, \left(a_n[W]-2\right)
\label{modformula}\end{aligned}$$ In this form (\[modformula\]), already the leading $(n=0)$ derivative expansion term gives the exact answer for the Sine-Gordon case. For the $\phi^4$ model, the leading term gives $10\%$ accuracy, the first correction gives better than $2\%$ accuracy, two corrections give $0.3\%$ accuracy, and three corrections give $0.05\%$ accuracy. This is impressive accuracy, especially considering $a_0, \dots, a_3$ can easily be calculated analytically.
For the Jacobi model there does not appear to be a simple closed formula for the $a_n$. This is presumably a reflection of the fact that in this case, unlike in the $\phi^4$ and Sine-Gordon cases, $V^{\prime\prime}(\phi_c(x))$ is not an exactly solvable potential. Nevertheless, we can compute various orders $a_n$ of the derivative expansion expression with ease:
$$\begin{aligned}
a_0^{\rm Lame}&=& - \frac{3}{128\nu^{3/2}}\left[2\sqrt{\nu}
(5 - 38\nu + 5\nu^2) -(5 + 3\nu + 3\nu^2 + 5\nu^3)
\log\left(\frac{1 + \sqrt{\nu}}{1-\sqrt{\nu}}\right)\right]
\nonumber\\
a_1^{\rm Lame}&=& \frac{1}{12288\nu^{5/2}}\left[-2\sqrt{\nu}(1785 + 520\nu -11746\nu^2 + 520\nu^3 + 1785\nu^4) \right. \nonumber\\
&&\hskip 2cm\left.+ 15(119 - 5\nu + 14\nu^2 + 14\nu^3 - 5\nu^4 + 119\nu^5)\log\left(\frac{1+\sqrt{\nu}}{1-\sqrt{\nu}}\right)\right] \nonumber\\
a_2^{\rm Lame}&=& \frac{1}{3932160 \nu^{7/2}}
\left[-2\sqrt{\nu}(1323945 - 683970\nu + 146839\nu^2 -3800572\nu^3 +
146839\nu^4 - 683970\nu^5 + 1323945\nu^6) \right.\nonumber\\
&&\left. + 105(12609 - 10717\nu + 2449\nu^2 + 779\nu^3 + 779\nu^4 +
2449\nu^5 - 10717\nu^6 + 12609\nu^7) \log\left(\frac{1+\sqrt{\nu}}{1-\sqrt{\nu}}\right)\right] \nonumber\\
a_3^{\rm Lame}&=& -\frac{3}{587202560\nu^{9/2}}\left[2\sqrt{\nu}
(258321525 - 355550860\nu + 143327940\nu^2 - 681332\nu^3 \right.\nonumber\\
&&\hskip 2cm\left. - 194557554\nu^4 - 681332\nu^5 + 143327940\nu^6 - 355550860\nu^7 + 258321525\nu^8)\right.\nonumber\\
&&\hskip 2cm\left. - 105(2460205 - 4206267\nu + 2275076\nu^2 - 275052\nu^3 - 24586\nu^4 - 24586\nu^5\right.\nonumber\\
&&\hskip 2cm \left. - 275052\nu^6 + 2275076\nu^7 - 4206267\nu^8 + 2460205\nu^9) \log\left(\frac{1+\sqrt{\nu}}{1-\sqrt{\nu}}\right)\right]
\label{lameas}\end{aligned}$$
Notice the appearance of the classical soliton mass $\frac{4}{\sqrt{\nu}}\log\left(\frac{1+\sqrt{\nu}}{1-\sqrt{\nu}}\right)$ in these expressions for the one-loop quantum correction.
Figure 3 shows the successive derivative expansion approximations to the quantum mass correction, as a function of the elliptic parameter $\nu$. The expansion appears to be convergent for each $\nu$. Furthermore, a nontrivial test of the results in (\[lameas\]) is that they must reduce to the Sine-Gordon results as $\nu\to 0$. Indeed, from (\[lameas\]) we verify that for each $n$ $$\begin{aligned}
a_n^{\rm Lame}\to 2 \quad {\rm as}\quad \nu\to 0
\label{red}\end{aligned}$$ On the other hand, as $\nu\to 1$ the quantum mass correction diverges, just as does the classical Jacobi soliton mass in (\[cmasses\]). This singular behavior is due to the fact that as $\nu\to 1$ the period of the Jacobi potential diverges logarithmically, so neighboring vacua become infinitely separated, and the asymptotic limits of the classical soliton $\phi_c(x)$ in (\[sols\]) diverge.
We now consider the $N=1$ supersymmetric extension of the bosonic soliton model (\[blag\]). Consider $$\begin{aligned}
{\cal L}_{\rm SUSY}=\frac{1}{2}(\partial_\mu\phi)
(\partial^\mu\phi)-\frac{1}{2}U^2(\phi)
+\frac{i}{2}\bar{\psi}\partial \hskip-6pt /\psi
-\frac{1}{2}U^\prime(\phi) \bar{\psi}\psi
\label{susylag}\end{aligned}$$ where $\psi$ is a Majorana fermion and $$\begin{aligned}
U(\phi)=\cases{\frac{1}{2}\left(\phi^2-1\right)\qquad ; \phi^4 {\rm model}\cr
-2\, {\rm sin}(\frac{\phi}{2})\qquad ; {\rm SineGordon\,\, model}\cr
-2\,{\rm sn}\left(\frac{\phi}{2}|\nu\right)\qquad ; {\rm Jacobi \,\,model}}
\label{upots}\end{aligned}$$ The bosonic potential is $V(\phi)=\frac{1}{2}U^2(\phi)$, solitons are given by $\frac{d}{dx}\phi_c=-U(\phi_c(x))$, and the boson-fermion coupling is $$\begin{aligned}
V_F(\phi)=U^\prime(\phi)
\label{fpot}\end{aligned}$$
-1cm -2.5in
The fermionic contribution to the one-loop soliton mass correction is $$\begin{aligned}
M_F&=&i\,{\rm tr}\,\log\left[i\partial \hskip-6pt
/ -V_F(\phi_c(x))\right]\nonumber\\
&=&-\frac{1}{2}\left\{M[W_+]+M[W_-]\right\}
\label{fmass}\end{aligned}$$ where $M[W]$ is the bosonic mass correction in (\[deriv\]) and $$\begin{aligned}
W_\pm(x)\equiv \left[V_F(\phi_c(x))\right]^2\pm\frac{d}{dx}V_F(\phi_c(x)) -1
\label{susyq}\end{aligned}$$ But with the SUSY coupling (\[fpot\]), $W_-(x)$ is equal to the bosonic potential $W(x)$: $$\begin{aligned}
W_-(x)=W(x)\equiv \frac{d^2}{d\phi^2}V(\phi_c(x))-1
\label{eq}\end{aligned}$$ Thus the [*total*]{} one-loop mass correction, including both bosonic and fermionic contributions, is $$\begin{aligned}
M_{\rm total}=\frac{1}{2}\left\{M[W_-]-M[W_+]\right\}
\label{diff}\end{aligned}$$ That is, the net mass correction is (half) the difference of two bosonic mass corrections, for $W_\mp(x)$ respectively.
Now notice the remarkable fact that if we compute this difference using the derivative expansion (\[formula\]), then term by term each of the functionals $a_n[W_-]-a_n[W_+]$ obtained from (\[as\]) is an integral of a total derivative. For example, $$\begin{aligned}
a_0[W_-]-a_0[W_+]&=&\frac{3}{8}\int_{-\infty}^\infty
\frac{4}{3}\frac{d}{dx} \left\{3V_F-V_F^3\right\}\, dx\nonumber\\
a_1[W_-]-a_1[W_+]&=&-\frac{5}{32}\int_{-\infty}^\infty
\frac{12}{5}\frac{d}{dx} \left\{ -5V_F+ \frac{10}{3}V_F^3-V_F^5-\frac{5}{3}V_F(V_F^\prime)^2\right\}\, dx \nonumber\\
a_2[W_-]-a_2[W_+]&=&\frac{7}{128}\int_{-\infty}^\infty
\frac{40}{7}\frac{d}{dx} \left\{ 7V_F-7V_F^3+\frac{21}{5}V_F^5-V_F^7+7(V_F-V_F^3)(V_F^\prime)^2-
\frac{7}{5}V_F^{\prime\prime} (V_F^\prime)^2 -\frac{7}{10}V_F(V_F^{\prime\prime})^2\right\} \, dx\nonumber\\
a_3[W_-]-a_3[W_+]&=&-\frac{9}{512}\int_{-\infty}^\infty
\frac{140}{9}\frac{d}{dx} \left\{ -9V_F+12V_F^3-\frac{54}{5}V_F^5+\frac{36}{7}V_F^7-V_F^9 -18V_F(1-V_F^2)^2(V_F^\prime)^2 -9V_F(V_F^\prime)^4\right.\nonumber\\
&&\left. +\frac{36}{5}(1-V_F^2)(V_F^\prime)^2V_F^{\prime\prime} +\frac{18}{5}V_F(1-V_F^2)(V_F^{\prime\prime})^2+
\frac{18}{35}(V_F^{\prime\prime})^3 -\frac{54}{35}V_F^\prime V_F^{\prime\prime} V_F^{\prime\prime\prime} -\frac{9}{35}V_F(V_F^{\prime\prime\prime})^2\right\}
\, dx\nonumber\\
\label{diffas}\end{aligned}$$ In (\[diffas\]), $V_F^\prime$ means $\frac{d}{dx}V_F(\phi_c(x))$. It is straightforward to generate these expressions to a desired order using Mathematica or Maple.
Due to this dramatic simplification, the differences $a_n[W_-]-a_n[W_+]$ are determined solely by the asymptotic values (at $x=\pm\infty$) of $V_F(x)$ and its $x$ derivatives. But for a solitonic background, $V_F(\phi_c(x))$ approaches $\pm 1$ exponentially fast, so $$\begin{aligned}
V_F(\phi_c(x))&=&\pm 1 \quad {\rm at} \quad x=\pm\infty\nonumber\\
V_F^{(n)}(\phi_c(x))&=&0 \quad {\rm at} \quad
x=\pm\infty\quad, \quad n\geq 1
\label{asymptotic}\end{aligned}$$ For example, these conditions are clearly satisfied for the $\phi^4$, Sine-Gordon and Jacobi models, $$\begin{aligned}
V_F(\phi_c(x))=\cases{{\rm tanh}(\frac{x}{2})\qquad ;\phi^4 {\rm model}\cr
{\rm tanh}(x)\qquad ; {\rm SineGordon\,\,model}\cr
\frac{(1-\nu)\, {\rm tanh}(x)}{1-\nu\,{\rm tanh}^2(x)}
\qquad ; {\rm Jacobi \,\,model}}
\label{fpotegs}\end{aligned}$$ Inserting the asymptotic conditions (\[asymptotic\]) into (\[diffas\]), the derivatives of $V_F$ do not contribute, and we discover that [*the differences are independent of n*]{}: $$\begin{aligned}
a_n[W_-]-a_n[W_+]=2\quad, \quad {\rm for\,\, all\,\, n}
\label{ans}\end{aligned}$$ Thus, combining the derivative expansion formula (\[formula\]) with the expression (\[diff\]) for the total mass correction, we find (in units of $m$) $$\begin{aligned}
M_{\rm total}=-\frac{1}{8\sqrt{\pi}}\,\sum_{n=0}^\infty
\frac{\Gamma(n+1)}{\Gamma(n+5/2)}=-\frac{1}{2\pi}
\label{totalmass}\end{aligned}$$ This is in agreement with recent analyses [@top; @jaffe] using topological boundary conditions and phase shift methods. We note that in the derivative expansion approach, this result is quite general, relying only on the asymptotic values of $V_F(\phi_c(x))$ and its $x$ derivatives.
This work has been supported in part by the U.S. Department of Energy grant DE-FG02-92ER40716.00.
R. Rajaraman, [*Solitons and Instantons*]{} (North Holland, Amsterdam, 1987).
C. Rebbi and G. Soliani, [*Solitons and Particles*]{}, (World Scientific, Singapore, 1984).
H. Nastase, M. Stephanov, P. van Nieuwenhuizen and A. Rebhan, “Topological boundary conditions, the BPS bound, and elimination of ambiguities in the quantum mass of solitons”, Nucl. Phys. B [**542**]{}, 471 (1999).
N. Graham and R. L. Jaffe, “Unambiguous one loop quantum energies of (1+1)-dimensional bosonic field configurations”, Phys. Lett. B [**435**]{}, 145 (1998), “Energy, central charge, and the BPS bound for 1+1 dimensional supersymmetric solitons”, Nucl. Phys. B [**544**]{}, 432 (1999), “Fermionic one-loop corrections to soliton energies in 1+1 dimensions”, Nucl. Phys. B [**549**]{}, 516 (1999).
G. Dunne and K. Rao, “Lamé Instantons”, hep-th/9906113.
I. J. R. Aitchison and C. M. Fraser, “Derivative expansions of fermion determinants: anomaly-induced vertices, Goldstone-Wilczek currents, and Skyrme terms”, Phys. Rev. D [**31**]{}, 2605 (1985).
L-H. Chan, “Effective action expansion in perturbation theory”, Phys. Rev. Lett. [**54**]{}, 1222 (1985), (erratum) [**56**]{}, 404 (1986), “Generalized derivative expansion and one-loop corrections to the vacuum energy of static background fields”, Phys. Rev. D [**55**]{}, 6223 (1997).
E. Whittaker and G. Watson, [*A Course of Modern Analysis*]{} (Cambridge, 1927).
H. W. Braden, “Mass corrections to periodic solitons”, J. Math. Phys. [**24**]{}, 929 (1987).
C. M. Bender, G. Dunne and P. Meisinger, “Complex periodic potentials with real band spectra”, Phys. Lett. A [**252**]{}, 272 (1999).
R. V. Konoplich, “Calculation of quantum corrections to nontrivial classical solutions by means of the zeta function”, Theor. Math. Phys. [**73**]{}, 1286 (1987).
This is in contrast to the case of the derivative expansion for the QED effective action in an inhomogeneous background field, which yields a (typically divergent) series in powers of the inhomogeneity scale; see for example: G. Dunne and T. Hall, “Borel Summation of the Derivative Expansion and Effective Actions”, hep-th/9902064, Phys. Rev. D in press. Here, the fluctuation potentials () contain no scale, so the determinant gives a number, not a series.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The diversity of structures in the Universe (from the smallest galaxies to the largest superclusters) has formed under the pull of gravity from the tiny primordial perturbations that we see imprinted in the cosmic microwave background. A quantitative description of this process would require description of motion of zillions of dark matter particles. This impossible task is usually circumvented by *coarse-graining* the problem: one either considers a Newtonian dynamics of “particles” with macroscopically large masses *or* approximates the dark matter distribution with a continuous density field. There is no closed system of equations for the evolution of the matter density field alone and instead it should still be discretized at each timestep. In this work we describe a method of solving the full 6-dimensional Vlasov-Poisson equation via a system of auxiliary Schrödinger-like equations. The complexity of the problem gets shifted into the choice of the number and shape of the initial wavefunctions that should only be specified at the beginning of the computation (we stress that these wavefunctions have nothing to do with quantum nature of the actual dark matter particles). We discuss different prescriptions to generate the initial wave functions from the initial conditions and demonstrate the validity of the technique on two simple test cases. This new simulation algorithm can in principle be used on an arbitrary distribution function, enabling the simulation of warm and hot dark matter structure formation scenarios.'
author:
- |
Matthieu Schaller$^1$[^1], Claude Becker$^2$, Oleg Ruchayskiy$^2$, Alexey Boyarsky$^{3,4}$ and Mikhail Shaposhnikov$^2$\
$^1$Institute for Computational Cosmology, Durham University, South Road, Durham, UK, DH1 3LE\
$^2$Institut de Théorie des Phénomènes Physiques, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland\
$^3$Instituut-Lorentz for Theoretical Physics, Universiteit Leiden, Niels Bohrweg 2, Leiden, The Netherlands\
$^4$Bogolyubov Institute of Theoretical Physics, Kyiv, Ukraine
bibliography:
- './bibliography.bib'
date: 'Accepted 2014 May 29. Received 2014 May 21; in original form 2013 October 27'
nocite: '[@*]'
title: A new framework for numerical simulations of structure formation
---
\[firstpage\]
cosmology: theory, dark matter, large-scale structure of Universe – methods: N-body, numerical
Introduction {#sec:introduction}
============
The Lambda Cold Dark Matter ($\Lambda$CDM) cosmological model is the current theoretical framework to describe the formation and evolution of large scale structures in the Universe. In this model, the growth of structures occurs through the hierarchical collapse of a collisionless fluid of cold dark matter (CDM). Small initial perturbations grow through merging to create more and more massive halos and complex sub-structures (e.g. @Davis1985 [@Bertschinger1998; @Springel2005]). These initial perturbations are thought to be (almost) Gaussian, created from quantum fluctuations during the inflation epoch and are the origin of all the objects seen in the Universe. The knowledge of the precise initial conditions and a comprehensive understanding of the underlying physical laws should, in principle, enable us to evolve these fluctuations forward in time and provide a test of the current models.
Most of the important features observable in the Universe today have grown via non-linear evolution from tiny primordial density perturbations. This makes the whole process of understanding their evolution complex and requires the use of techniques well beyond the linear perturbation theory [@Bernardeau2002]. Indeed, at scales below roughly $\unit[10]{Mpc}$ the evolution of structures had already entered the non-linear stage (i.e. the density contrast $\delta \rho$ is of order (or much greater) than the background density $\bar\rho$). The main resource available to cosmologists is the use of bigger and bigger cosmological simulations, most of them using the particle technique known as [$N$-body ]{}simulation [@Hockney1988; @Dehnen2011]. Numerical simulations may, for instance, help shed some light on the unknown nature of dark matter.
Clearly, the number of dark matter particles is way too large to track individually each of them on a computer. Therefore most of the cosmological [$N$-body ]{}simulations use macroscopically large simulation “particles” (with their masses ranging from masses much larger than DM particles up to the size of a small galaxy, $10^8-10^9 {M_\odot}$).
The problem of dark matter evolution in the Universe can be formulated as an evolution of a collisionless self-gravitating fluid.
The main tool used to describe this dark matter fluid is the *phase space density distribution* $f(x,v,t)$, defined such that $f(x,v,t)dqdv$ represents the mass of material at position $x$ moving at velocity $v$ at time $t$. This function is usually normalized such that its integral over all positions and velocities gives the total mass
$$\int d^3x\int d^3v f(x,v,t) = M_{\rmn{tot}}.$$
Notice that one could also normalize this integral to one or to the total number of particles in the system. When integrating over velocity space only, one gets the usual mass density $\rho(x)$, whereas integrating over all space returns the velocity distribution $d_v(v)$:
$$\rho(x) = \int f(x,v,t)d^3v, \qquad d_v(v) = \int f(x,v,t) d^3x.$$
This distribution function obeys the Liouville theorem [@Binney2008] and if the only force acting on the particle is the gravitational potential $U(x)$, we can write a closed system of equations for the formation of structures [@Bertschinger1995; @Bernardeau2002]:
$$\begin{aligned}
{\frac{\partial f}{\partial \tau}} + \frac{v}{a(\tau)}{\frac{\partial f}{\partial x}} - a(\tau)\nabla U {\frac{\partial f}{\partial v}} &= 0, \\
\nabla^2 U &= 4\pi Ga^2(\tau) \delta\rho,
\end{aligned}
\label{eq:VP}$$
where $x$ and $v$ are comoving coordinates and velocities, $a(\tau)$ is the scale factor and $\tau$ is the conformal time (We will use this convention throughout this paper). This Vlasov-Poisson system has no solution in the general case and the only way to handle it is to use numerical techniques. For completeness, we also give the expressions for the density and density contrast:
$$\begin{aligned}
\rho(x,\tau) &=& \frac{1}{a^3(\tau)}\int f(x,v,\tau)d^3v, \\
\delta\rho(x,\tau) &=& \frac{1}{a^3(\tau)}\left(\int f(x,v,\tau)d^3v - \frac{M_{\rmn{tot}}}{V_{\rmn{tot}}} \right),\end{aligned}$$
where $V_{\rmn{tot}}$ is the total comoving volume over which we average.
Structure formation simulations {#sec:theory}
===============================
The numerical analysis of the Vlasov-Poisson system of equations (\[eq:VP\]) is very challenging. The first reason is that the system is six-dimensional. Recent simulations can only handle up to $64$ resolution elements in each spatial and velocity space direction [@Yoshikawa2013] due to memory restrictions. Even the use of the biggest supercomputers would not allow to go much beyond this figure.
The second shortcoming of such technique is the development of fine-grained structures that are very difficult to follow numerically. These become very important in structure formation scenarios as clusters typically present many matter streams and shell crossings.
Those two main shortcomings make the search for more advanced numerical scheme important. The problem of high-dimensionality could be removed if there were a way to use the density field $\rho(x)$ instead of the probability distribution function $f(x,v)$. This can be done by integrating the first few moments of the Vlasov equation and then use techniques known for hydrodynamical simulations (see e.g. @Hockney1988). This technique is limited by the formal need to integrate all moments and not just the first few ones to obtain an exact solution. Instead of a 6D space, there is now a (formally) infinite number of variables obeying an infinite series of equations. [@Peebles1987], for instance, truncates the series and uses the first two moments (mass conservation, Euler equation) of the collisionless Boltzmann equation to evolve in time the initial perturbations. The framework reaches its limits whenever the velocity dispersion of the fluid becomes important or when shell crossing occurs.
[$N$-body ]{}simulations {#ssec:simulations}
------------------------
The other option to solve the system of equations (\[eq:VP\]) is to use a particle method in which the distribution function is sampled by a finite number $N$ of particles such that
$$f(x,v) \cong \frac{1}{N} \sum_{i=1}^{N} m_i \delta\left(x - x_i\right)\delta\left(v - v_i\right).
\label{eq:nbody}$$
Each particle or body is then evolved according to Newton’s law under the influence of the gravitational potential created by all the others as described by Poisson’s equation. In other words, [$N$-body ]{}simulations solve the Vlasov equation via its characteristics by sampling the initial phase space distribution with a discrete number of particles. The number of bodies is typically chosen as large as computationally feasible. The [$N$-body ]{}formalism is thus a Monte-Carlo approximation of the Vlasov-Poisson system. The advent of large supercomputers combined with the development of more efficient numerical algorithms has enabled the field of cosmological simulations to make considerable progress over the last decades. Simulations such as the *Millennium run* [@Springel2005] or *Bolshoi simulation* [@Klypin2011] are able to follow as many as a few billion particles.
The complicated part of the [$N$-body ]{}simulation is the evaluation of the forces between pairs of particles. Over the years, many ingenious techniques (see [@Dehnen2011] for a review) have been invented to reduce the algorithms complexity for the force integration to $\mathcal{O}(N\log N)$ or even better [@Dehnen2000]. All these techniques (tree-code, particle-mesh, P$^3$M, AMR, tree-PM,...) do however rely on particles and do, hence, share the same initial assumptions leading to the two following challenges.
Firstly, since the dark matter fluid is supposed to be collisionless, one has to manually suppress artificial two-body collisions arising between the pseudo-particles introduced to sample the phase space distribution. This is usually done by introducing an ad-hoc softening length and suppressing the gravitational force at scales below it [@Dehnen2001]. [$N$-body ]{}simulations are run under the assumption that for a suitable choice of the smoothing, the evolution of the $N$ pseudo-particles under the softened force should be the same as the gravitational evolution of the elementary dark matter particles.
The second challenge is to relate the particle distribution to the theoretical Vlasov-Poisson the particles are supposed to model. Despite its obvious relevance, it seems that the question of the precise quantitative importance of the discretization (\[eq:nbody\]) and its effects is still not settled [@Joyce2008].
As a matter of fact, there are no alternative tools to study the cosmic structure formation with the same resolution as [$N$-body ]{}simulations. This is of course not a limitation of the [$N$-body ]{}method itself, but makes it more complicated to evaluate the possible errors of [$N$-body ]{}simulations quantitatively, as there are basically no independent results to compare with. For instance [@Ludlow2011] find a non-negligible fraction of halos in CDM simulations that cannot be matched to peaks in the initial density distribution and are possible artefacts of the [$N$-body ]{}method. The different techniques used to calculate the forces are, of course, different and can lead to marginally different results for the same initial sampling of the field when the resolution limit is reached. They do, however, all share the decomposition of $f(x,p)$ in a set of $N$ macroscopic particles and will, hence, share the consequences of this Ansatz.
Spurious effects due to the discretization become more apparent when looking at simulations of warm dark matter (WDM) or hot dark matter (HDM) cosmologies. The initial matter power-spectrum entering such simulations is truncated below a certain free-streaming scale related to the dark matter particle rest mass. Those particles having a small mass, they also have a finite velocity distribution function at every point in space, making the problem effectively 6 dimensional. In practice, these velocities are neglected and the DM fluid is treated in the cold fluid limit. These simulations are run using the same [$N$-body ]{}framework but with an initial density and velocity power spectrum truncated below the scale of interest. This should lead to a suppression of small halos below a characteristic mass and the simulations ought to be able to reproduce all structures with a mass above this limit. They could thus quickly converge towards a solution. [@Colin2000; @Wang2007; @Colin2008] did, however, demonstrate that this is not the case and that spurious halos form and merge to form structures below the theoretical mass threshold. Various techniques are used in the literature to cure this problem. [@Lovell2012], for instance, filter their halo catalogues during the post-processing of their simulations. The end results are thus free from spurious halos but it does not solve the intrinsic discreteness problem of the [$N$-body ]{}technique.
More details about these challenges and a comprehensive review of the topic can be found in [@Dehnen2011]. Notice that this formalism is still a very active and lively area of research with alternative more advanced formulations being proposed frequently. Some authors [@Abell2012; @Shandarin2012] proposed recently to use tessellations of the 3D matter sheet in 6D space to track some of the phase space information. This may allow them to solve the coarse graining problem and reduce the impact of non-physical two body relaxations between the macroscopical particles. This formalism has lead to promising results in the study of WDM cosmology and the differences between the CDM and WDM halo mass functions [@Angulo2013].
All the potential shortcomings of the [$N$-body ]{}formalism and the difficulty to evaluate their impact on the simulation results make it important to develop another framework not based on a particle approach.
An alternative framework
------------------------
Our framework resembles the attempt by [@Peebles1987] to use only the density field $\rho(x)$ and potential $U(x)$. The main problem of such an approach is that there is no closed system of equations that includes only the density and gravitational potential.
The situation is different when looking at quantum physics. In this realm, all the phase-space information can be encoded in a single function, the wavefunction $\psi(x)$ which does not depend on the velocity $v$. It is thus possible to write a closed Schrödinger-Poisson system that would replace the Vlasov-Poisson one and that would only depend on the spatial variable $x$ (See also [@Short2006] for a similar idea). This would effectively be a 3D system of equations but would allow to simulate the full 6D phase space and hence allow simulation of alternative cosmologies, such as the one including WDM or free-streaming neutrino contributions. The principal difficulty is then to find a good mapping between the distribution function $f(x,v)$ of interest and its “quantum” equivalent $\psi(x)$ and vice-versa. This is achieved by using the so-called *Wigner distribution function*
$$f(x,v) \simeq \int e^{\frac{i}{\hbar}vy}\psi^*\left(x+\frac{y}{2}\right)\psi\left(x-\frac{y}{2}\right)
d^3y.$$
which obeys an equation similar to the Vlasov equation but is constructed from wave functions. The main feature of this mapping is that the density field can simply be expressed as
$$\label{eq:1}
\rho(x) = |\psi(x)|^2.$$
However, the limitation of this approach is that one single wavefunction is in general not sufficient to encode all the complexity of the distribution function and we would then use the more general version:
$$\label{eq:2}
\rho(x) = \sum_n |\psi_n(x)|^2.$$
The summation index $n$ can, as a first thought, be understood as a sum over the velocities $v$ that appear in the distribution function $f(x,v)$. We somehow trade a 6D function for a (finite) set of 3D (complex valued) functions. We will, however, demonstrate that the number of wavefunctions required can be very low (of order unity in some cases), making the whole framework effectively 3D.
It is important to stress from the onset that we are not trying to solve the evolution of structure formation at the quantum level. Although we make use of quantum mechanics concepts, we merely use it as mathematical “trick” to solve the Vlasov-Poisson system (\[eq:VP\]). For this reason, the constant $\hbar$ appearing in our equations has to be understood as a computational parameter whose value bares no relation to the actual Planck constant $\hbar_{\rmn{phys}}= 1.0545 \cdot10^{-34}~\rmn{m}^2~\rmn{kg}~\rmn{s}^{-1}$.
Once the wavefunctions are built, they are evolved forward in time using [Schrödinger]{}’s equation. The density sourcing Poisson’s equation is obtained through equation (\[eq:2\]) and one can then solve for the potential at each time step using standard techniques. This potential enters the [Schrödinger]{}equation, closing the loop. We have, hence, built a closed system of equations using only a set of wave functions (which serve as a proxy for density) and the gravitational potential. $$\label{eq:SP}
\begin{array}{rcl}
i\hbar \partial_t \psi_n &=& - \displaystyle\frac{\hbar^2}{2} \nabla^2_x \psi_n +
U\psi_n, \Big.\\
\nabla^2_x U &=& 4\pi G\delta\rho. \Big.
\end{array}$$
The study of structure formation then becomes an exercise in solving $n$ copies of the [Schrödinger]{}equation on a computer, which is a well-studied problem. The velocity distribution can be recovered by Fourier transforming the wave functions and if one is interested in the phase space distribution, one can apply the Wigner transform. This is, however, not part of the algorithm itself. This can be done in post-processing if necessary. The entire evolution of the system can be done at the “quantum level”, i.e. using the wave functions alone.
We stress that this is another approximation of the true underlying physical problem (equation \[eq:VP\]) and that this framework, as any other, will have limitations. Some of these limitations and their relevance to the case of structure formation studies will be discussed in this paper. We will address those in the context of the science we are interested in and demonstrate how alternative cosmologies, including non cold dark matter scenarios, could effectively be simulated.
The development of this framework has been pioneered by [@Widrow1993] and [@Davies1997] with the important difference that these authors use a single wavefunction and another way to map the distribution function in the quantum world. Their general procedure is very similar to ours: sample the wavefunction from the initial phase space distribution, evolve in time using the Schrödinger-Poisson equations and recover the final phase space distribution from the wavefunction.
Note also that another possible route, where the Hartree equation is used instead of the Schrödinger equation, has been explored by [@Aschbacher2001] and [@Frohlich2010].
The time evolution of the Schrödinger-Poisson system is done using an explicit finite-differences scheme for the wave function and a FFT algorithm to solve the Poisson equation. We try to improve upon their algorithm for the time evolution as will be described below. [@Widrow1993] have made several simulations using this Schrödinger method obtaining results in agreement with usual [$N$-body ]{}simulations. They also claim that their method is computationally comparable to [$N$-body ]{}simulations making it a promising tool for cosmological purposes.
These authors choose to use one single wave function to represent the distribution function. This has important consequences on the validity of equation (\[eq:1\]). By using one single wave function, the phase space distribution built from it can not be everywhere positive and the authors have thus to add an additional Gaussian smoothing. We alleviate this shortcoming by using more than one wave function and a different transformation from wave functions to phase space distribution.
Their choice of Gaussian smoothed density also led them to a simple technique to generate the initial wave function. They use a set of particles sampling the phase space distribution function exactly as in the case of [$N$-body ]{}simulations. They can then turn each particle into a Gaussian in phase space by smoothing it and use this set of wave packets as their initial wave function.
In our approach, we depart from this need of an initial [$N$-body ]{}sampling by considering other techniques to generate the set of wave functions. By doing so, we allow for a completely generic distribution function and should, in principle, not experience the consequences of an *a priori* artificial Monte-Carlo sampling of $f(x,v)$.
The second feature of our framework is the replacement of the Poisson equation by a Klein-Gordon equation for the potential $U(x)$: $$-\frac{1}{c^2}\frac{\partial^2}{\partial t^2}U + \nabla^2_xU = 4\pi G\delta\rho,$$
where $c$ is the numerical speed of gravity. This scalar gravity equation is, once again, purely a mathematical trick to reduce the complexity of the original system (\[eq:VP\]) and not an attempt to modify Newton’s gravity. Such a replacement makes the framework entirely local and does not require complicated integration methods for the Poisson equation. The complexity of the scheme is then formally reduced to $\mathcal{O}(M)$, where $M$ is the number of mesh points in real space used in the simulation. In this respect our approach also differs from the original work by [@Widrow1993], who stick to the classical Poisson form of gravity.
We stress that this step is not formally necessary. The well-known techniques used to solve Poisson’s equation on a mesh (FFT, Gauss-Seidel relaxation, etc.) can also be used in our framework. This change of equation for gravity does just make the computations slightly faster in the cases where our approximation is valid. However, in the case of cosmological simulations with vastly different scales interacting, it is unclear how the Klein-Gordon equation for gravity would behave and defaulting to standard mesh techniques might be required.
The algorithm in brief {#ssec:algorithm}
======================
Here we present the main algorithm of our framework, decomposed in a few simple steps. A formal derivation and a discussion of the convergence and accuracy of the method will be presented in the next section.
Choose the parameters of your simulation. The precision and speed of the method is governed by three parameters $\hbar$, $c$ and $N$. The algorithm of choosing them is the following:\
The parameters $c$ and $\hbar$ are linked to the time and space resolution ($\Delta x$ and $\Delta\tau$ respectively) of the simulation via the Courant condition: $$\frac{\Delta x}{\Delta\tau} > c$$ and the condition on the stability of the discretized [Schrödinger]{}equation: $$\frac{\Delta x^2}{\Delta\tau} > \hbar$$ The number of wavefunctions $N$ is chosen depending on the number of relevant modes of the decomposition in wavefunctions of the initial distribution function. The optimal value of $N$ is problem dependent and is also influenced by the algorithm chosen to discretize the distribution function. The details of this procedure will be given in section \[sec:IC\]. The precision of the original accuracy is also dictated by the choice of $\hbar$. The “quantum” nature of the formalism imposes limitations on the precision of the description of position and velocity at the same point following the equivalent of Heisenberg’s uncertainty principle.
Take an initial phase-space distribution function of the matter fields $f(x,v)$ (in the case of cosmological simulations, it is expressed via the power spectrum $P(k)$). Decompose the distribution function in $N$ complex-valued $\psi_n(x)$ such that $$\label{eq:initial_distribution}
f(x,v) \approx \sum_{n=1}^N \int
e^{\frac{i}{\hbar}v\,y}\psi_n^*\left(x+\frac{y}{2}\right)\psi_n\left(x-\frac{y}{2}\right)
d^3y.$$ The number $N$ of wavefunctions is chosen such as to minimize the error introduced by the decomposition and will, in practice, be as big as computationally feasible. Various ways to generate this initial set of wavefunctions for a given $f(x,v)$ are presented in section \[sec:IC\].
*At this stage the precision of the approximation $f(x,v)\to
\{\psi_n(x)\}$ is controlled by two parameters, $\hbar$ and $N$.*
The wavefunctions $\psi_n(x)$ are now evolved forward in time using the coupled [Schrödinger]{}-Klein-Gordon system of equations $$\begin{array}{rcl}
i \hbar \partial_t \psi_n &=& - \displaystyle\frac{\hbar^2}{2} \nabla^2_x \psi_n +
U\psi_n, \Big.\\
-\displaystyle\frac{1}{c^2}\displaystyle\frac{\partial^2}{\partial t^2}U + \nabla^2_xU &=& 4\pi
G\left(\displaystyle\sum_{n=1}^N |\psi_n(x)|^2 - \bar\rho\right). \Big.
\end{array}$$
The integration in time of the [Schrödinger]{}-Klein-Gordon system can be done explicitly using finite differences on a regular grid, as will be described in section \[sec:numerics\].
Controlling your simulation. As the simulation is running you should monitor the following quantities in order to see that the choice of the method does not introduce artefacts. The correction terms $$\label{eq:correction_term}
\sum_{\shortstack{$\scriptstyle r\geq 3$\\$\scriptstyle r~\rm {odd}$}}
\frac{1}{r!}\left(\frac{\hbar}{2i}\right)^{r-1} {\frac{\partial ^r}{\partial x^r}}V {\frac{\partial ^r}{\partial v^r}}f(x,v)$$
should be small when compared to the ones ($v{\frac{\partial f}{\partial x}}$ and ${\frac{\partial U}{\partial x}}{\frac{\partial f}{\partial v}}$) entering the Vlasov equation. Thanks to the $1/r!$ decrease and the smoothness of the gravitational potential $V$ in most cases of interest, computing the first term of this series is generally sufficient. If this term grows above the value of the other terms in the Vlasov equation, then the approximation introduced in this paper is not valid any more. Reducing the value of $\hbar$ or increasing the number $N$ of wavefunctions used in the initial discretization will decrease the contribution of the correction terms but this will lead to a higher computational cost. The correction terms as well as the terms entering the Vlasov equation are expensive to compute but need not be computed at each time step.
Once the final time has been reached, the distribution function can be recovered by computing the integral (\[eq:initial\_distribution\]) or if one is only interested in the density field only, then equation (\[eq:2\]) is sufficient and straightforward to compute.
Formal derivation {#sec:framework}
=================
In the previous section, we described the problem we were interested and the usual schemes used in the literature. We also presented a brief description of the route we intend to follow in order to tackle the issues outlined. In this section, we describe the whole formulation in detail, derive its main equations and discuss its limits. For completeness, we start with a review of a formulation of quantum mechanics and show how its main ingredient, the *Wigner Distribution Function*, will play the role of an approximate distribution function for our problem. Readers interested only in the end results can jump directly to Section \[ssec:local\].
Phase-space quantum mechanics {#ssec:QM}
-----------------------------
Quantum mechanics is usually presented as emerging from the Hamiltonian formulation of classical mechanics through canonical quantization (See for instance [@Sakurai]). In this procedure, variables are promoted to Hermitian operators and the Poisson bracket is replaced by a commutator. Alternatively, one can also use Feynman’s propagator and the path integral formalism to move from classical to quantum mechanics.
Alongside these well-known quantization procedures, there exist other equivalent formulations which try to emphasize more clearly certain aspects. The Moyal (or phase-space) formulation is among those and tries to find a quantum equivalent to the classical phase-space and distribution functions [@Ercolessi2007; @Hillery1984]. The quantization procedure tries to find a correspondence between classical functions (called symbols) of the phase space variables and quantum operators in Hilbert space:
$$\mbox{Operators in Hilbert space}\leftrightarrow \mbox{Phase space symbols}$$
As the position and momentum operators do not commute, this mapping can not be unique. Different operator orderings will be mapped to different phase-space symbols. Hermann Weyl proposed a systematic way to associate a quantum operator to a classical distribution function, which is now referred to as *Weyl quantization*. This complex procedure will not be discussed further here but its inverse, the *Wigner transform* will be useful for our formalism. This transformation associates to every quantum operator $\hat{A}$ a real phase-space function $A(x,v)$:
$$A(x,v) = {\mbox{sym}}(\hat{A}) := \int e^{\frac{i}{\hbar}vy}\left\langle x-\frac{y}{2}
\right|\hat{A}\left|x+\frac{y}{2}\right\rangle dy,
\label{eq:Wigner_transform}$$
where $\langle \cdot|\cdot\rangle$ is the usual Bra-ket notation for quantum states. The transformation of a product of operators is given by
$${\mbox{sym}}(\hat{A}\hat{B}) :={\mbox{sym}}(\hat{A}) \star {\mbox{sym}}(\hat{B}),$$
where the *Moyal star product* $\star$ contains the quantum mixing of the operators. This product of functions in phase space is defined as $$A(x,v)\star B(x,v) := A(x,v)~
e^{\frac{i\hbar}{2}\left(\overleftarrow{\partial_x}\overrightarrow{\partial_v} -
\overleftarrow{\partial_v}\overrightarrow{\partial_x} \right)}~B(x,v)$$ and is a central element in this formulation of quantum mechanics. Defining the *Moyal bracket* [@Moyal1949] by
$$\left\lbrace A, B\right\rbrace_M := A\star B - B \star A$$
the commutator of operators is associated to the Moyal brackets of two symbols in the following way:
$${\mbox{sym}}\left(\left[\hat{A},\hat{B}\right]\right)
=\left\lbrace{\mbox{sym}}\left(\hat{A}\right),{\mbox{sym}}\left(\hat{B}\right)\right\rbrace_M.$$
The dynamical equation in this formulation can be written in a simple way using these brackets and reads
$$i\hbar \partial_t f = \left\lbrace \hat{H}, f \right\rbrace_M,$$
where $\hat{H}$ is the Hamiltonian of the system. The interesting property of this formulation of quantum mechanics is that in the semi-classical limit $\hbar\rightarrow 0$, the dynamical equation reduces to the classical equation of motion expressed in terms of Poisson brackets
$$\partial_t f = \left\lbrace H,f\right\rbrace_P = H \left(\overleftarrow{\partial_x}\overrightarrow{\partial_v} -
\overleftarrow{\partial_v}\overrightarrow{\partial_x} \right) f.$$
This illustrates how the algebraic structures of classical and quantum mechanics are related through the continuous changing of the parameter $\hbar$. This is the reason why such an approach to quantum mechanics is known as *deformation quantization* [@Hirshfeld2002].
Let’s now stop this overview and move to the part of this formalism which will be useful for the construction of our new simulation framework.
Wigner distribution function {#ssec:WDF}
----------------------------
The Wigner transform (equation \[eq:Wigner\_transform\]) maps a quantum operator $\hat{A}$ to a classical function in phase space. Wigner used this to associate a real phase space function to a quantum system [@Wigner1932], now called the *Wigner distribution function* (WDF). It is defined as the symbol in phase space associated to the density operator $\hat\rho$:
$$P_W(x,v) := {\mbox{sym}}\left(\hat\rho\right) = \int e^{\frac{i}{\hbar}vy} \left\langle x-\frac{y}{2}
\right|\hat{\rho}\left|x+\frac{y}{2}\right\rangle d^3y.$$
As usual, the density operator can be expressed as the combination of pure state wavefunctions $\psi_n$:
$$\hat\rho = \sum_n \lambda_n \left|\psi_n\right\rangle\left\langle\psi_n\right|, \quad \lambda_n \geq 0, \quad \sum_n
\lambda_n = 1.$$
For mixed states, the WDF is thus
$$P_W(x,v) = \int e^{\frac{i}{\hbar}vy}\sum_n \lambda_n\psi^*_n\left(x+\frac{y}{2}\right)\psi_n\left(x-\frac{y}{2}\right)
d^3y,
\label{eq:WDF}$$
while for a pure state, it reads
$$P_W(x,v) = \int e^{\frac{i}{\hbar}vy}\psi^*\left(x+\frac{y}{2}\right)\psi\left(x-\frac{y}{2}\right)
d^3y.$$
To simplify the expressions, we will use the notation $x_\pm = x \pm \frac{y}{2}$ and $\psi_{n\pm} = \psi_n(x_\pm)$ in what follows. The WDF has many similarities to the classical distribution function: $P_W(x,v)$ is a real function, as can be seen by taking the conjugate and performing the change of variable $y\rightarrow-y$. It is also normalized to $1$ in the following sense
$$\int d^3x\int \frac{d^3v}{(2\pi \hbar)^3} P_W(x,v) = 1.$$
It has similar marginal distributions as can be seen by integrating over all velocities: $$\begin{aligned}
\int\frac{d^3v}{(2\pi \hbar)^3} P_W &=& \sum_n
\lambda_n\int
\delta^3\left(y\right)\psi^*_n\left(x_+\right)\psi_n\left(x_-\right)d^3y\nonumber \\
&=& \sum_n \lambda_n \left| \psi_n\left(x\right) \right|^2,
\label{eq:density}\end{aligned}$$ or over all space
$$\begin{aligned}
\int d^3x P_W &=& \sum_n \lambda_n \int d^3x_-
d^3x_+e^{\frac{i}{\hbar}vx_+}e^{-\frac{i}{\hbar}vx_-}\psi^*_{n+}\psi_{n-} \nonumber \\
&=& \sum_n \lambda_n\left|\tilde\psi_n\left(\frac{p}{\hbar}\right) \right|^2.\end{aligned}$$
In both cases the non-negative property of these marginal distributions is a property of the wavefunctions in quantum mechanics.
The Wigner distribution function does, however, have the peculiar property that it may assume negative values. For this reason, it is called a *quasi-probability distribution* and cannot be interpreted as a phase-space probability density in the sense of classical mechanics. The non-positivity of the WDF can be seen by integrating over all phase-space the product of two distributions built from different states $\psi$ and $\Phi$:
$$\int dx \int dv~ P_W[\psi](x,v) P_W[\Phi](x,v) \propto \left| \left\langle \psi |\Phi\right\rangle\right|^2.$$
The right-hand side vanishes if the two states $\psi$, $\Phi$ are orthogonal which implies that the WDF cannot be positive everywhere. According to the Hudson theorem [@Hudson1974], the WDF of a pure state is point wise non-negative if and only if the state is Gaussian. If $\hat\rho$ is not a pure state, it can be represented as a convex combination of pure state operators, $\hat\rho =
\sum_n \lambda_n | \psi_n \rangle\langle \psi_n|$, in infinitely many ways. The WDF satisfies the so-called mixture property [@Ballentine1998], which is the requirement that the phase space distribution should depend only on the density operator $\hat\rho$ and not on the particular way it is represented as a mixture of some set of pure states $\left\lbrace \left|\psi_n\right\rangle\right\rbrace$. To summarize, the Wigner distribution function has many properties similar to the classical phase space distribution. Nevertheless it has been realized from the early days, that the concept of a joint probability at a phase space point is limited in quantum mechanics because the Heisenberg uncertainty principle makes it impossible to simultaneously specify the position and velocity of a particle. Therefore, the best one can hope to do is to define a function that has a maximum of properties analogous to those of the classical distribution function. Many different variants of distribution functions - Husimi, Kirkwood-Rihaczek, Glauber - have been studied over the decades, all with their own advantages and shortcomings (See [@Lee1995] for a review). The WDF is despite its non-positivity considered to be a useful calculational tool and finds applications in various domains outside of quantum physics, like signal processing or optics [@Bastiaans1997].
[@Widrow1993] use a Husimi distribution [@Husimi1940] to recover the phase space information from the wavefunction. The Husimi distribution is essentially equal to the Wigner distribution with an additional Gaussian smoothing of width $\eta$
$$P_H(x,v) = \frac{1}{(2\pi\hbar)^3}\frac{1}{(\pi\eta^2)^{3/2}}\left|\int d^3y
e^{-\frac{(x-y)^2}{2\eta^2}-\frac{i}{\hbar}vy} \psi(y) \right|.$$
Compared to the WDF it has the advantage of yielding a phase space distribution that is positive-definite at every point. This comes at the price of the marginal distributions not being equal to the usual position and velocity distributions, but rather Gaussian broadened versions of it
$$\rho_H(x) = \frac{1}{(\pi\eta^2)^{3/2}}\int d^3 y e^{-\frac{(x-y)^2}{2\eta^2}}|\psi(y)|^2.$$
Only in the limit $\eta\rightarrow0$ does it reduce to the usual probability distribution. Similarly one can show that the other marginal distribution reduces to the standard velocity distribution only when $\eta\rightarrow\infty$ This complementarity is of course related to Heisenberg’s uncertainty principle. Note that it is in principle this smoothed distribution that enters Poisson equation instead of $|\psi(x)|^2$. Since this would requiring an additional space integration at each time step, Widrow and Davies approximate it with the usual distribution $|\psi(x)|^2$ in the Poisson equation.
Actually, $P_H(x,p) \simeq f(x,p)$ only when averaged on scales $\Delta x \geq \eta$, $\Delta p \geq \frac{\hbar}{\eta}$ . Note that there is no a priori reason why the non-linear time evolution should yield an answer that is again, in average, close to the real distribution function. Let us stress that we allow for several wavefunctions to have an initial phase space representation that is arbitrary close to the classical distribution function at every point, not only when averaged.
Let us recall that our goal is not to interpret the Wigner distribution function as a fully-fledged phase space distribution, but rather as a convenient mathematical tool.
Dynamical equation for the WDF {#ssec:EOM}
------------------------------
We now want to derive the dynamical equation satisfied by the WDF. A derivation starting from Liouville’s equation for the density matrix can be found in [@Ballentine1998]. Another possibility is to start by taking the time derivative of the Wigner distribution function and use the fact that the wavefunctions satisfy Schrödinger’s equation.
Suppose each of the wavefunctions satisfies Schrödinger equation
$$i\hbar \partial_t \psi_n = -\frac{\hbar^2}{2}\nabla^2 \psi_n + V\psi_n,
\label{eq:schroedinger}$$
then the time derivative of the WDF becomes
$$\begin{aligned}
\partial_t P_W =\int e^{\frac{i}{\hbar}vy} \sum_n \lambda_n \left[
-\frac{i\hbar}{2}\right. \left. \left(\nabla^2_+\psi^*_{n+}\psi_{n-} \right. \right.
\nonumber \\
\left.- \psi^*_{n+}\nabla^2_-\psi_{n-} \right) - \left.\frac{1}{i\hbar}\left(V_+-V_-\right)\psi^*_{n+}\psi_{n-}\bigg.
\right] d^3y,\end{aligned}$$
where, once again, the subscripts $+$,$-$ denote the dependence on $x_\pm = x \pm \frac{y}{2}$. The terms containing a Laplacian can be rewritten in terms of spatial derivatives of $P_W$ only and the previous equation becomes
$$\begin{aligned}
0 &=&\partial_t P_W + \vec{v}\cdot\vec{\nabla}_x P_W \nonumber \\
&& - \frac{1}{i\hbar}\int e^{\frac{i}{\hbar}py}\left(V_+-V_-\right)\sum_n\lambda_n\psi^*_{n+}\psi_{n-}~d^3y.\end{aligned}$$
This is the dynamical equation for the WDF, that we will refer to as the *Wigner equation*. This dynamical equation depends on both $P_W$ and the wavefunctions which implies that we might have to define initial conditions for both. Let’s now demonstrate that one can get rid of the dependency on the $\psi_n$. Let’s expand the potential in Taylor series
$$V(x_+) - V(x_-) = y {\frac{\partial }{\partial x}}V(x) + 2\sum_{\shortstack{$\scriptstyle r\geq 3$\\$\scriptstyle r~\rm {odd}$}}
\frac{1}{r!}{\frac{\partial ^r}{\partial x}}V(x)\left(\frac{y}{2}\right)^2$$
and use this result in the dynamical equation: $$\begin{aligned}
0 &=& {\frac{\partial }{\partial t}} P_W + \vec{v}\cdot{\frac{\partial }{\partial x}} P_W - {\frac{\partial V}{\partial x}}{\frac{\partial }{\partial v}}P_W \nonumber\\
& & + \sum_{\shortstack{$\scriptstyle r\geq 3$\\$\scriptstyle r~\rm {odd}$}}
\frac{1}{r!}\left(\frac{\hbar}{2i}\right)^{r-1} {\frac{\partial ^r}{\partial x^r}}V {\frac{\partial ^r}{\partial v^r}}P_W.
\label{eq:Wigner}\end{aligned}$$ One can notice that the first three terms correspond to the classical Vlasov equation.
In three cases, the Wigner equation exactly coincides with the classical Vlasov equation: for a free particle ($V=0$), for a uniform field ($V\propto x$) and for a harmonic oscillator ($V \propto x^2$). In general, there are additional terms that can be interpreted as quantum corrections[^2] or simply higher-order corrections. In any other case, corrections in the form of a power series in $\hbar$ will appear and modify the dynamic. Note that in this derivation, the only assumption made on the $\lambda_n$ is that they be constant. In principle any value is acceptable and it can even be negative or complex. As we are not using these equations to solve a quantum mechanics problem, where $\lambda_n>0$, we can use this fact to create more general sets of wavefunctions to approximate a given $f(x,v)$.
Note that the mass $m$ does not appear in the [Schrödinger]{}equation in the same way that it does not appear in the Vlasov system. This, once again, illustrates that we are not solving the quantum mechanics evolution of the individual DM particles but rather find an approximation of the DM fluid evolution equation.
Let us recap what we have derived so far. By inspecting the Moyal formulation of quantum mechanics, we found a quantity, the Wigner distribution function $P_W$. This quasi-probability density function obeys the Wigner equation, an equation similar to the Vlasov equation but with additional terms in the form of a power series in $\hbar$.
Semi-classical limit {#ssec:limit}
--------------------
The Wigner equation (\[eq:Wigner\]) reduces to the classical Vlasov equation in the limit $\hbar\rightarrow 0$. Even though the quantum correction is formally $\mathcal{O}\left(\hbar^2\right)$, the derivatives of $P_W$ could generate additional inverse powers of $\hbar$, making the semi-classical limit more involved[^3].
The properties of the semi-classical limit depend of course on the potential $V(x)$. In this paragraph we present some results concerning the case of interest to us, where the potential satisfies Poisson’s equation. In particular, different authors investigated the semi-classical limit of the Wigner-Poisson (W-P) system to the Vlasov-Poisson (V-P) system for the Coulomb potential.
The mathematically rigorous classical limit from W-P to V-P has been solved first in 1993 independently by [@Lions1993] and [@Markovitch1993]. Both references consider a so-called completely mixed state; i.e. an infinite number of pure states with a strong additional constraint on the occupation probabilities:
$$\mbox{Tr}\hat\rho^2 = \sum_n \lambda_n^2 \leq C\hbar^3,$$
where $C$ is a constant. Under this assumption, the classical limit of the solution to the 3D W-P system converges to the solution of the V-P system. Note that the Wigner distribution function can also have negative values, whereas the semi-classical limit is a true, non-negative distribution function. In both references, this was overcome by using a Gaussian-smoothed Wigner function.
The situation for a pure state is completely different [@Zhang2002]. According to these authors, it appears that a density operator which has the above property that the trace of its square tends to zero with the third power of the Planck constant seems to be closer to classical mechanics than a pure state. For a pure state in 1D, the semi-classical limit is not unique: examples have been constructed where different regularization schemes give different limits [@Majda1994]. The question whether there exists a selection principle to pick the correct classical solution has also been investigated but is not yet settled [@Jin2008]. No proof of the semi-classical limit from W-P to V-P is known for the pure state case in 2D or 3D.
For more details the reader is referred to the original papers or the review [@Mauser2002]. See also [@Frohlich2007] for an alternative approach to the semi-classical limit.
Finally, let us stress once again, that we seek to use our knowledge of quantum mechanics to simplify the resolution of the mathematical problem presented in the Introduction. We are not trying to describe the physics of structure formation at the quantum level nor trying to find a wavefunction for the entire Universe.
Local interaction framework {#ssec:local}
---------------------------
In Newtonian gravity, much like in classical electrodynamics, each body moves in the potential generated by all the others. As both forces are long-ranged, the total force acting on each of the $N$ particles will be given by the sum of the contributions from all the other particles, no matter how far away. In gravitational [$N$-body ]{}problems, the $N$ sampling bodies also receive a contribution from all the other bodies and a naive algorithm would require $\mathcal{O}(N^2)$ operations for the force calculation at each time step. But it is well-known that this long-ranged interaction through the potential can be replaced by a purely local interaction with a gauge boson or a spin-zero boson. In this approach, each particle only interacts locally with the bosonic field.
We propose to reformulate the cosmological Vlasov-Poisson problem system (\[eq:VP\])
$$\begin{aligned}
{\frac{\partial f}{\partial t}} + \frac{v}{a} {\frac{\partial f}{\partial x}} - a {\frac{\partial U}{\partial x}}{\frac{\partial f}{\partial v}} &=&0, \nonumber\\
\nabla^2 U &=& 4\pi Ga^2 \delta\rho,\end{aligned}$$
as a purely local problem. To achieve spatial locality, we shall trade the real-valued phase space distribution function $f(x,v)$ for a finite set of complex-valued wavefunctions $\left\lbrace \psi_n(x)\right\rbrace$. For this we shall assume that the classical distribution function can be approximated by the Wigner distribution function of some auxiliary mixed states:
$$f(x,v) \simeq P_W(x,v) = \int e^{\frac{i}{\hbar}vy}\sum_n \lambda_n \psi^*_n(x_+) \psi_n(x_-)d^3 y$$
The details of how this approximation is to be understood, and how we construct in practice the set of wavefunctions $\left\lbrace \psi_n(x)\right\rbrace$ for any given $f(x,v)$ will be discussed in Section \[sec:IC\]. For the time being, let us assume that we have determined a set of wavefunctions such that the above approximation holds.
The dynamical evolution of the WDF is given by the quantum-corrected Vlasov equation (the Wigner equation (\[eq:Wigner\])), or equivalently, by the Schrödinger equation (\[eq:schroedinger\]) of the wavefunctions interacting in a self-consistent way with a potential obeying the Poisson equation. The cosmological Vlasov equation in an expanding Universe and expressed using conformal time $\tau$ is very similar to the classical one, up to the replacements
$$v \mapsto \frac{v}{a(\tau)}, \qquad V \mapsto a(\tau)U.$$
Therefore the Schrödinger-Poisson system in the expanding universe becomes
$$\label{eq:SP_cosmo}
\begin{array}{rcl}
i\hbar \partial_\tau \psi_n &=& - \displaystyle\frac{\hbar^2}{2a} \nabla^2_x \psi_n +
aU\psi_n, \Big.\\
\nabla^2_x U &=& 4\pi Ga^2\delta\rho, \Big.
\end{array}$$
where $\delta\rho$ is the cosmological density contrast. The mass density $\left[\rmn{kg}~\rmn{m}^{-3} \right]$ relates to the wavefunctions $\left[\rmn{kg}^{1/2}~\rmn{m}^{-3/2} \right]$ by
$$\rho = \frac{1}{a^3} \int \frac{d^3v}{(2\pi\hbar)^3}f(x,v) = \frac{1}{a^3} \sum_n \lambda_n \left| \psi_n\right|^2.$$
The normalization is chosen such that the phase space density integrates to the total mass
$$\int d^3 x\int \frac{d^3v}{(2\pi\hbar)^3}f(x,v) = \int d^3 x \sum_n \lambda_n \left| \psi_n\right|^2 = M_{\rmn{tot}},$$
implying for the background density
$$\bar\rho = \langle\rho\rangle = \frac{1}{V_\rmn{tot}} \frac{1}{a^3}\int d^3 x \sum_n \lambda_n \left| \psi_n\right|^2 =
\frac{1}{a^3}\frac{M_\rmn{tot}}{V_\rmn{tot}},$$
where $V_\rmn{tot}$ denotes the total comoving volume. Therefore the density contrast $\delta\rho$ reads
$$\delta\rho = \frac{1}{a^3}\left(\sum_n \lambda_n \left| \psi_n\right|^2 - \frac{M_\rmn{tot}}{V_\rmn{tot}} \right).$$
In the semi-classical limit ($\hbar\rightarrow 0$), the Schrödinger-Poisson system (\[eq:SP\_cosmo\]) formally reduces to the original Vlasov-Poisson system describing gravitational structure formation.
Notice that the total mass is conserved by construction as the normalization of the wavefunctions is a constant of motion of the Schrödinger equation.
So far, we achieved locality in the sense that our set of equations does not explicitly depend on the velocity variable $v$. We traded our 6 dimensional phase-space density function for a (possibly infinite) set of complex-valued functions that depend on the space coordinate $x$ only. The numerical complexity of the problem has thus been drastically reduced as long as the number of wavefunctions remains small. Before addressing this question, let us go one step further and discuss the second equation of our system (\[eq:SP\_cosmo\]).
The Poisson equation is a non-local equation as the Laplacian operator couples the contributions from the whole space. This can, however, be changed by replacing the Laplacian by a d’Alembertian operator. With this change, the Poisson equation becomes a Klein-Gordon equation and our transformed cosmological problem now reads
$$\label{eq:S_KG}
\begin{array}{rcl}
i\hbar \partial_\tau \psi_n &=& - \displaystyle\frac{\hbar^2}{2a} \nabla^2_x \psi_n +
aU\psi_n, \Big.\\
-\displaystyle\frac{1}{c^2}\partial^2_{\tau\tau}U + \nabla^2_x U &=& 4\pi Ga^2\delta\rho. \Big.
\end{array}$$
This system is entirely local, meaning that it can be numerically evolved in time on a grid by summing contributions of local sampling points only. If the contribution of the term $-\frac{1}{c^2}\partial^2_{\tau\tau}U$ becomes small, then this system reduces to the Schrödinger-Poisson system discussed previously. This is in particular true in the non-relativistic limit $c\rightarrow \infty$. It is important to understand that the speed $c$ does not necessarily have to take the value of the physical speed of light (or of gravity) $c_{\rmn{phys}} = 299792458~\rmn{m}~\rmn{s}^{-1}$. It must simply be understood as a parameter that we can use to approach the physical problem we are interested in (equation \[eq:VP\]). As for $\hbar$, we are free to choose this parameter in a way that is convenient for our simulations, as long as we remain in the non-relativistic limit, meaning that the gravitational field $U$ propagates much faster than the matter fields $\psi_n$.
Note, however, that using a non-infinite speed for the mediator of gravity in cosmological simulations may also be of some physical interest as the Poisson equation is, formally, only a weak-field approximation of the underlying Einstein equations from which a finite speed for the gravity emerges. Thus, modifying this parameter may also yield interesting physical results.
Let us summarize what we achieved so far. Using the formalism derived in the Sections \[ssec:QM\] to \[ssec:limit\], we have been able to construct a completely local system of equations (\[eq:S\_KG\]) which in the non-relativistic classical limit $\hbar\rightarrow 0$, $c\rightarrow\infty$ reduces to the problem of cosmological structure formation. The probability density function can be computed at any time using the definition of the WDF (equation \[eq:WDF\]) but we stress that this operation is in general not necessary as one is usually interested in the evolution of the mass density (equation \[eq:density\]) only.
Let us finally say that replacing the Poisson equation by a scalar field is not strictly necessary as the algorithmic complexity of the problem has already been drastically reduced by the introduction of the WDF. Having a [Schrödinger]{}-Poisson system to solve instead of equation (\[eq:VP\]) is more accurate than our final system (\[eq:S\_KG\]). It does, however, simplify a lot the numerical algorithms in some cases and does not seem to impact heavily the results as long as the parameter $c$ is chosen wisely. The effect of this choice on the evolution of highly clustered matter fields found in the low redshift Universe has not, however, not been explored.
Lagrangian formulation {#ssec:lagrangian}
----------------------
The system of equations (\[eq:S\_KG\]) can be derived from a Lagrangian density using the Euler-Lagrange equations. We consider a real scalar field $U$ interacting with the complex scalar matter fields $\psi_n$. The Lagrangian for this system reads
$$\begin{aligned}
\mathcal{L} &=& \frac{1}{2c^2}\dot U ^2 - \frac{1}{2}\left(\vec{\nabla}U\right)^2 + \zeta\bar\rho U \nonumber \\
& & + \zeta \sum_n\lambda_n\left[\frac{i\hbar}{2}\left(\psi_n^*\dot\psi_n - \dot\psi_n^*\psi_n\right) \right.
\nonumber\\
& & \qquad\qquad\quad- \left.\frac{\hbar^2}{2a}\vec{\nabla}\psi_n^*\cdot\vec\nabla\psi_n - |\psi_n|^2U\right].
\label{eq:lagrangian}\end{aligned}$$
The equations of motion are found to be
$$\begin{aligned}
i\hbar \dot\psi_n &=& - \frac{\hbar^2}{2a}\nabla^2 \psi_n + U\psi_n, \\
-\displaystyle\frac{1}{c^2}\ddot U + \nabla^2 U &=& \zeta \left(\sum_n \lambda_n\left|\psi_n\right|^2 -
\bar\rho\right),\end{aligned}$$
which is the system we derived in the previous section if we set $\zeta =
\frac{4\pi G}{a}$. The Hamiltonian density corresponding to the Lagrangian (\[eq:lagrangian\]) is given by
$$\begin{aligned}
\mathcal{H} &=& \frac{1}{2c^2}\dot U ^2 + \frac{1}{2}\left(\vec{\nabla}U\right)^2\nonumber \\
& & + \zeta \sum_n\lambda_n \frac{\hbar^2}{2a}\left|\vec{\nabla}\psi_n\right|^2 \nonumber \\
& & + \zeta \left(\sum_n \lambda_n\left|\psi_n\right|^2 - \bar\rho \right)U,
\label{eq:hamiltonian}\end{aligned}$$
which has a positive definite kinetic energy term for the scalar potential, as expected from a well-behaved theory.
One can also decompose this Hamiltonian in its various energy components. Doing so allows us to control the impact of the dynamic term for the field $U$ and consider it a valid approximation of the underlying Vlasov-Poisson problem when its value is much lower than the other energy components. Together with the computation of the higher order terms of the Wigner equation (\[eq:Wigner\]), this measure of the impact of $c\neq\infty$ gives us a measure of the approximations we made and can thus help us assess the validity of the outcome of our simulations.
Generating Initial Conditions {#sec:IC}
=============================
In the previous section, we showed how one can trade the Vlasov equation for the phase space distribution function for Schrödinger’s equation for the wavefunctions, as this allows for the introduction of a scalar field as the mediator of the gravitational force. Of course we do not require the wavefunctions to have any intrinsic physical interpretation. We rather consider them, just like the WDF, as a mathematical tool and not as fundamental entities. Still we are faced with the problem of how to determine a set of wavefunctions such that their WDF corresponds to the initial classical phase space distribution.
One possible approach is to start from a set of $N$ particles sampling the phase-space distribution function and build Gaussians centred on each point with a certain width $\eta$
$$|\eta(x_i,v_i)\rangle\propto e^{-\frac{(x-x_i)^2}{2\eta^2}-\frac{i}{\hbar}v_i \cdot x}.$$
The wavefunction is then obtained from the incoherent superposition of these wave-packets for each “particle”
$$|\psi\rangle = \frac{1}{\sqrt{N}}\sum_{i=1}^N e^{i\phi_i}|\eta(x_i,v_i)\rangle,$$
where $e^{i\phi_i}$ is a random phase. This sampling procedure relies on the assumption that each “particle” has a well-defined velocity. It is unclear how it could be generalized to the case of warm dark matter, where the velocity dispersion is important. We remove the need for this assumption by allowing for several wavefunctions. At the same time this allows us to represent any initial phase space distribution without relying on [$N$-body ]{}sampling. Such an approach was used by Widrow & Kaiser and is well-suited for Husimi distributions as they contain an extra Gaussian smoothing. We will, instead, try to work directly with the distribution function without sampling it in particles and hence taking the risk of facing the coarse graining and discreteness effects (Section \[ssec:simulations\]) that we are trying to avoid in our framework.
Since the wavefunctions encode both, the position and velocity information, a single wavefunction (pure state) can in general not be sufficient to describe a generic $f(x,v)$. One should rather look for a set of wavefunctions (mixed state). The more wavefunctions we allow for, the more freedom we have and the more accurately the WDF should represent any given distribution. At the same time the total number of wavefunctions should be as small as possible because this will reduce the computational complexity of our numerical simulations.
Given the classical distribution function $f(x,v)$, we want to expand it using the WDF Ansatz
$$f(x,v) = \sum_{n=1}^{N} \lambda_n \int e^{\frac{i}{\hbar}vy} {\psi_n^*\left(x+\frac{y}{2}\right)} {\psi_n\left(x-\frac{y}{2}\right)} d^3y.$$
Fourier transforming from $v$-space to $\eta$-space we get
$$f(x,\eta) = \sum_{n=1}^{N} \lambda_n {\psi_n^*\left(x+\frac{\eta}{2}\right)} {\psi_n\left(x-\frac{\eta}{2}\right)}.
\label{eq:Fourier_WDF}$$
Finding the wavefunctions is now a simpler problem provided one can easily compute the Fourier transform of the distribution function one is interested in. We will discuss different approaches to tackle this problem of determining the set of wavefunctions $\psi_n$ and weights $\lambda_n$ representing a given initial phase space distribution $f(x,v)$. Let us stress from the outset that these procedures need only to be used once at the beginning of a numerical simulation, to set up the initial conditions.
Last but not least, we need to emphasize that the number of wavefunctions is preserved by the quantum mechanical evolution. There is no evolution equation for $\lambda_n$. Only the shape of the $\psi_n$ will change. This shows that it is the complexity of the initial conditions that dictates the number of wavefunctions required. In a setup where only a restricted number of harmonics are present in the initial probability distribution, already relatively few wavefunctions would be sufficient to represent the system and its time evolution.
Brute-force minimization {#ssec:minimization}
------------------------
The first and obvious method we present to choose the initial wavefunctions is a brute-force minimization. The underlying idea is to define a functional measuring the total absolute error made by approximating the phase space distribution by the WDF Ansatz
$$\Phi := \int d^3q \int d^3\eta \left|f(x,\eta) - \sum_{n=1}^{N}\lambda_n \psi_{n+}\psi_{n-}
\right|^2 ,$$
where, once again, $\psi_{n\pm} = \psi_n\left(x\pm \frac{\eta}{2}\right)$. We can then determine a set of wave functions that minimizes this error. In practice, the minimization is most easily done via discretization on a lattice. The problem is then cast into a minimization of the scalar error function with a large number of variables corresponding to the values of the wavefunctions at the lattice points. For different $N=1,2,\ldots$, we can determine the set of wavefunctions $\psi_n$ and corresponding weights $\lambda_n$ which minimizes the error. One can then compare the results for different $N$ to find an optimal approximation with a high enough accuracy and a minimal number of wavefunctions.
Since we are not seeking a true quantum mechanical interpretation, let us consider the most general case of complex-valued weights. A naive minimization will not yield wavefunctions normalized to unity. Instead of adding this normalization as a constraint to the minimization, we remove the amplitude of the complex weights $\lambda_n$, and only keep their phases $e^{i\phi_n}$.The amplitudes of the weights are taken to be the norm of the wavefunctions, thereby normalizing them to unity. If we simply minimize the error functional, we will in general obtain wavefunctions that are not smooth enough on the lattice to be evolved numerically. For this purpose it is useful to add a term of the form of a kinetic term to the functional that will allow us to enforce a certain degree of smoothness. We construct the kinetic term from the square of the derivatives with a certain overall factor $\chi$ to tune the smoothness:
$$\mathcal{K} = \chi \int d^3x \sum_{n=1}^N \left|{\frac{\partial \psi_n}{\partial x}}\right|^2.$$
Finally, we minimize this kinetic term with the total error summed over all lattice points
$$\mathcal{E} = \mathcal{K} + \Phi.$$
We have applied the method to cosmic initial conditions of cold dark matter in the Zel’dovich approximation, for simplicity in a one-dimensional case. The results confirm the expectation that, increasing the number of wavefunctions, the total error is reduced. In the case we studied, it turned out that already a relatively small number of wavefunctions (compared for instance to the number of lattice points) was enough to achieve a reasonable accuracy. As usual with minimization procedures, there is no guarantee that the algorithm converges to a global minimum. This would for instance mean that one has to repeat the minimization with different initial random seeds and compare their outcomes. Also, even though this minimization was shown to work for a given phase space distribution $f(x, v)$, in practice it becomes computationally challenging even for rather small 3D lattice sizes, as the number of variables in the minimization procedure grows quickly. Despite its applicability to any distribution function, the brute-force minimization might not be the best method to determine the initial wavefunctions.
Eigenvalue problem for Hermitian operator {#ssec:operator}
-----------------------------------------
We now turn our attention to obtaining an analytic solution to the problem of determining the initial wavefunctions. More precisely we will show how the Wigner Ansatz can be reformulated as an eigenvalue problem, which we can then solve analytically in some specific cases.
Since $f(x,\eta)$ is the Fourier transform of a real function $f(x,v)$, it satisfies the condition $f^*(x,-\eta)=f(x,\eta)$. Introducing the coordinates $x_\pm := x \pm \frac{\eta}{2}$, we can define $$g(x_-,x_+) := f\left(\frac{x_++x_-}{2}, x_+-x_-\right),$$ which is then Hermitian
$$g^*(x_+,x_-) = g(x_-,x_+).$$
Hilbert-Schmidt’s theorem states that any square-integrable Hermitian kernel can be expressed in terms of its spectral decomposition
$$g(x_-,x_+) = \sum_n \lambda_n\psi^*_n(x_-)\psi_n(x_+),
\label{eq:spectral}$$
where the $\lambda_n$ are real eigenvalues and ${\psi_n}$ the set of orthonormal eigenfunctions with respect to the standard scalar product on $L^2(\mathbf{C}^3)$
$$\left\langle \psi_n | \psi_m\right\rangle := \int \psi^*_n(x)\psi_m(x) d^3x = \delta_{nm}.
\label{eq:scalar}$$
The Fourier space WDF (equation \[eq:Fourier\_WDF\]) has exactly the same form as the spectral decomposition (equation \[eq:spectral\]). Therefore we conclude that *any* given phase space distribution function $f (x, v)$ can be written exactly as a WDF, if need be with an infinite number of wavefunctions. The wavefunctions are the eigenfunction of the Hermitian operator $g(x_- , x_+ )$ and its real eigenvalues correspond to the weights of the wavefunctions in the mixed state. Notice though, that they can in general take negative values, implying that we cannot give a full quantum-mechanical interpretation to the mixed state, as the corresponding density operator is not positive-definite. Let us emphasize once more that we consider the wavefunctions as a mere mathematical tool.
Multiplying both sides of (\[eq:spectral\]) by $\psi_\alpha(x_-)$ and integrating over $x_-$ , the orthonormality of the eigenfunctions implies the following integral equation
$$\int g(x_-,x_+) \psi_\alpha(x_-) d^3x = \lambda_\alpha \psi_\alpha(x_+).$$
This equation shows that the determination of the wavefunctions reduces to finding the eigenfunctions of the Hermitian kernel $g$. Unfortunately, for a completely general phase space distribution function, the above equation might not allow an analytic solution.
This procedure can be generalized by allowing for a more general scalar product containing a non-trivial weight function $w(x)$:
$$\left\langle \psi| \phi \right\rangle_w := \int \psi^*(x) \phi(x) w(x) d^3x.$$
For such a scalar product, the eigenvalue decomposition of $g(x_-,x_+)$ still exists but the eigenfunctions are now orthonormal with respect to the weighed scalar product. The eigenvalue problem thus reads
$$\int g(x_-,x_+) \psi_\alpha(x_-) w(x_-)d^3x = \lambda_\alpha \psi_\alpha(x_+).
\label{eq:scalar_weight}$$
Let us emphasize that the weighted scalar product is only used to determine the wavefunctions whose WDF equals the classical distribution function. The choice of $w(x)$ is completely arbitrary and does not affect the properties of the WDF or the Schrödinger evolution of the wavefunctions. Clearly the spectrum will depend on the choice of weight function. The additional freedom of choosing $w(x)$ could allow to reduce the number of wavefunctions needed in the Wigner Ansatz. Furthermore the arbitrariness of the weight function also reflects the freedom we have to choose wavefunctions representing the initial state.
Fourier-series decomposition {#ssec:fourier}
----------------------------
Let us study the eigenvalue problem for a phase space distribution of the form[^4] $f(x, v) = \rho(x)\delta(v)$, meaning the product of a generic distribution in space with a delta function in velocity space. This choice corresponds to the case of CDM at early times, when the velocities are negligible.
In such a case, the integral operator $g(x_-,x_+)$ becomes real and symmetric
$$g(x_-,x_+) = \rho\left(\frac{x_++x_-}{2} \right).$$
We choose the trivial weight function $w(x) = 1$, which might not be the optimal choice for a minimal number of wavefunctions, but yields a working example of the method. We now assume a periodic distribution of matter in $[0,L]$ and expand the density as a Fourier series over the interval
$$\rho(x) = \rho_0 + \sum_{n=1}^\infty a_n \cos\left(\frac{2\pi n}{L}x \right) + \sum_{n=1}^\infty b_n
\sin\left(\frac{2\pi n}{L}x \right).$$
The term $\rho_0$ can be dropped without loss of generality as it can trivially be represented in the WDF using a constant wavefunction.
The eigenvalue problem is easier to solve on the doubled interval $[0,2L]$. The standard scalar product for real functions on this interval is simply
$$\left\langle\psi|\phi \right\rangle := \frac{1}{L} \int_0^{2L} \phi(x) \psi(x) dx,$$
which means that the eigenvalue problem reads
$$\frac{1}{L} \int_0^{2L} \rho\left(\frac{x+y}{2} \right) \psi(y) dy = \lambda \psi(x).$$
We now have to choose an orthonormal basis for the wavefunctions $\psi$. As we work with a periodic interval, it is natural to use harmonic functions over $[0,2L]$. The most general case is thus
$$\psi(x) = \sum_{n=1}^\infty \left[\alpha_n \cos\left(\frac{\pi n}{L}x \right) + \beta_n \sin\left(\frac{\pi n}{L}x
\right)\right].$$
Using trigonometric identities and the orthonormality relations between the sine and cosine functions of different modes, the problem can be recast in a matrix problem for the coefficients of the Fourier series:
$$\left(\begin{array}{cc}
a_n & b_n \\
b_n & -a_n
\end{array}\right)
\left(\begin{array}{c}
\alpha_n \\
\beta_n
\end{array}\right)
= \lambda
\left(\begin{array}{c}
\alpha_n \\
\beta_n
\end{array}\right).$$
Therefore, the normalized eigenfunctions and eigenfunctions of the integral operator are finally given by
$$\psi_n^\pm(x) = \mathcal{N}\left[ \left(a_n + \lambda_n^\pm\right) \cos\left(\frac{\pi n}{L}x \right)+
b_n \sin\left(\frac{\pi n}{L}x \right)\right], \label{eq:Fourier_decomposition}$$
$$\lambda_n^\pm = \pm \sqrt{a^2_n + b^2_n},$$
where $\mathcal{N} = \left[\left(a_n \pm \sqrt{a_n^2 + b_n^2} \right)^2 + b_n^2\right]^{-1/2}$ normalizes the eigenfunctions to unity. It can be checked explicitly that these eigen-vectors satisfy the condition of orthonormality and yield the correct spectral representation
$$\begin{aligned}
\rho\left(\frac{x_+ + x_-}{2}\right) = \sum_{n=1}^{\infty} \left[\lambda_n^+ \psi_n^+(x_+)\psi_n^+(x_-)\right.
\nonumber \\
\quad~+ \left.\lambda_n^- \psi_n^-(x_+)\psi_n^-(x_-)\right].\end{aligned}$$
corresponding to the WDF
$$\begin{aligned}
P_W(x,v) = \sum_{n=1}^{\infty} \int e^{\frac{i}{\hbar}vy}\left[\lambda_n^+ \psi_n^+\left(x + \frac{y}{2}
\right)\psi_n^+\left(x - \frac{y}{2}\right)\right.
\nonumber \\
+ \left.\lambda_n^- \psi_n^-\left(x+\frac{y}{2}\right)\psi_n^-\left(x-\frac{y}{2}\right)\right] d^3y.\end{aligned}$$
As a conclusion we have been able to solve the eigenvalue problem on the finite interval and use it to find the wavefunctions for the WDF Ansatz. This applies for a generic density profile $\rho(x)$ periodic on $[0,L]$ and a phase space distribution of the form $f(x, v) = \rho(x)\delta(v)$. The wavefunctions are harmonic functions with increasing velocity. In general we would need an infinite number of wavefunctions to avoid smoothing the smallest scales of the power spectrum. For many applications a finite or even small number of wavefunctions may be sufficient.
In this procedure, we used the geometry of the problem to decide which orthonormal basis to use. The periodicity of the density distribution naturally led us towards the use of harmonic functions. In cases were the density is not periodic, one could use Chebyshev polynomials or any other basis whose geometry helps reduce the number of modes.
As already mentioned, the technique presented in this section holds for any power spectrum and in particular is well suited to the case of WDM without initial velocities as is usually done in numerical simulations. This truncated CDM power-spectrum can easily be decomposed in a Fourier series and hence used in our framework. If the thermal velocities of the WDM particles have to be included, then another technique has to be used (see sections \[ssec:operator\] and \[ssec:SVD\]).
Cosmological initial conditions {#ssec:cosmo_IC}
-------------------------------
Observations of structure in the universe are perfectly compatible with the simplest possible statistical description, namely a Gaussian distribution. More precisely, each Fourier mode of the density contrast $\delta(\vec{k})$ (not to be confused with the Dirac delta distribution) satisfies an isotropic Gaussian distribution, entirely described by the power spectrum $P(k):= \langle|\delta(\vec(k))|^2\rangle$, which is a function of the modulus $k$ only, not of the direction. From the knowledge of the power spectrum one can then generate a realization with the desired statistical properties
$$\begin{aligned}
\delta(\vec{x}) &=& \sum_{\vec{k}}\left[\sqrt{P(k)}\mathcal{N}(0,1)\cos(\vec{k}\cdot\vec{x})\right. \nonumber\\
&&\left.\qquad+\sqrt{P(k)}\mathcal{N}(0,1)\sin(\vec{k}\cdot\vec{x})\right],\end{aligned}$$
where $\mathcal{N}(0,1)$ denotes a Gaussian random number with zero mean and unit dispersion. This shows that the density contrast for cosmological initial conditions is in a form for which we know how to construct the WDF, provided that we start our simulation at times, when the Zel’dovich velocities of the particles are negligible. Compared to [$N$-body ]{}simulations we do not need to first perform a FFT to compute $\delta(\vec{x})$ but can find the initial wavefunctions directly from the power spectrum. Additionally we do not need any glassy pre-initial conditions to model the constant background.
There is, however, a little caveat when generating initial conditions for CDM. Such an initial spectrum is formally made of a Dirac distribution in $v$-space which means that even an infinite number of continuous wavefunction can not reproduce exactly this singularity. This can also be explained by the quantum aspect of our formalism. Heisenberg’s uncertainty principle forbids us to have at the same time an infinitely precise description of position and velocity of our wavefunction. There will be some necessary spread in velocity space proportional to the value of $\hbar$ chosen in the simulation. The spectrum obtained will thus formally not exactly be the CDM one but will contain some intrinsic velocities for the DM. These would vanish in the limit $\hbar\rightarrow0$.
We would in principle require as many wavefunctions as Fourier modes are relevant in the power spectrum, which may lead to a prohibitive computational cost. Expanding the power-spectrum in an other basis or using a non-trivial weight $w(x)$ in the scalar product (\[eq:scalar\_weight\]) may help reduce the number of wavefunctions required. On the other hand, we may turn this as an advantage as this new formalism can allow us to probe some parts of the power spectrum only without having to use the full range of $\vec{k}$.
Matrix formulation {#ssec:SVD}
------------------
Given that the WDF Ansatz can be thought of as spectral decomposition of an Hermitian operator, we can now analyse the solution in the discrete case, where the problem reduces to a matrix problem. Let us again restrict the analysis to one dimension. Working on a lattice $(x_1, x_2, \ldots,x_M)$, we can think of any function $f(x)$ as a vector $(f(x_1),
f(x_2), \ldots, f(x_M))^T$ and of any function of two variables as a matrix. We can thus reinterpret the functional relationship
$$g(x_-,x_+) = \sum_{n=1}^N \lambda_n\psi^*_n(x_-)\psi_n(x_+)$$
in terms of matrices $$\hat G_{ij} = \sum_{n=1}^N \lambda_n {\Psi}_{jn}^*{\Psi}_{in} = \sum_{n=1}^N\sum_{k=1}^N
{\Psi}_{in}\lambda_n\delta_{nk}{\Psi}_{kj}^\dagger.$$ The property $g(x_+,x_-) = g^*(x_-,x_+)$ translates into the fact that $\hat G \in\mathbf{C}^{M\times M}$ is a Hermitian matrix $\hat G^\dagger = \hat G$ which we can diagonalize by means of a unitary transformation
$$\hat G_{ij} = \left({\Psi}\cdot\hat{\Lambda}\cdot{\Psi}^\dagger\right)_{ij},$$
where $${\Psi}\in\mathbf{C}^{M\times N}, \qquad\hat{\Lambda} =
\mbox{diag}(\lambda_1,\ldots,\lambda_N)\in \mathbf{R}^{N\times N}.$$ The columns of $\Psi$ are the wavefunctions $\psi_n$ sampled on the lattice. The property that $\Psi$ is unitary $\Psi^\dagger\Psi = \mathbf{1}$ implies that the normalization of the wavefunctions on the lattice. This matrix formulation has the advantage, that it is straightforward to compute the spectrum of any given Hermitian matrix. The shortcomings of this approach are two-fold: firstly we would need as many wavefunctions as lattice points, which comes at a big computational cost, and secondly the eigen-vectors have no a priori reason to be smooth enough to be used as initial conditions for our numerical scheme. Note, however, that in all the cases we tested, the eigenvalue decomposition has yield smooth enough functions.
Moreover it has to be noted that we would need to compute the eigen-vectors for a matrix containing the full 3D lattice. Computing the eigen-vectors of a $n \times n$ matrix is in general a problem of complexity $\mathcal{O}(n^3)$. Since the size of the matrix is related to the number of lattice points $M^3$, one quickly reaches such lattice sizes making the solution of the eigenvalue problem impossible. This issue can be solved by combining this technique with the minimization procedure. One can first use an eigenvalue decomposition on a coarse grid and use this as an input of the brute-force minimization algorithm on a finer grid. A technique using multiple grids at the same time could also be used in the same way that Gauss-Seidel relaxation is done in some particle-mesh gravity solvers.
There are multiple known algorithms available to decompose a matrix in eigen-vectors. We chose to use the singular value decomposition (SVD) as the publicly available implementations return the eigenvalues sorted in decreasing order. This allows us to choose only the wavefunctions whose eigenvalues are above a certain (arbitrarily chosen) level.
Discussion and remarks {#ssec:remarks}
----------------------
For numerical simulations in a finite box with periodic boundary conditions, the spatial lattice resolution also dictates the resolution in velocity space. The size of the box is related to the lattice size in $v$-space since the wave-vectors take discrete values $\vec{v} = \frac{2\pi}{L}\vec{n}$. The maximal wave-vector is linked to the lattice spacing in real space. This illustrates the relationship between the number of wavefunctions and the spatial resolution of the simulation. If we keep all the modes, we need $\mathcal{O}(M^3)$ wavefunctions, where $M$ is the number of lattice points in one direction. Note that this corresponds, in order of magnitude, to the number of particles in [$N$-body ]{}simulations. So even if we keep the maximal number of wavefunctions needed to accurately represent the initial conditions, the complexity of our numerical scheme will still be comparable to the naive $\mathcal{O}(N^2)$ complexity of [$N$-body ]{}simulations. As we will generally use much less wavefunctions, the complexity is much lower and may even trump the usual $\mathcal{O}(N\log N)$ complexity offered by tree-codes or FFT schemes to solve Poisson equation.
An other advantage of working with harmonic wavefunctions to represent the initial conditions is that we have an intuitive picture of what happens if we remove some modes. In analogy with the Fourier series, the density will not be represented exactly at every point, but the approximation becomes closer and closer as we include more and more modes. Knowing some of the properties of the system we want to model may help to get a deeper insight into which modes are really needed. The same is true when the density is expanded in another basis even if it may be more difficult to get an intuitive mental picture of the impact of high-order modes when dealing with Chebyshev polynomial say.
In many simulations one does not necessarily need the same resolution on all scales. Instead one could work with an adaptive grid [@Plewa2005] and have higher resolution in the scales of interest. This would allow to reach better precisions while keeping the number of wavefunctions constant. A similar technique is used in [$N$-body ]{}solvers such as <span style="font-variant:small-caps;">RAMSES</span> [@Teyssier2002] or <span style="font-variant:small-caps;">ART</span> [@Kravstov1999].
In the special case of simulations of cosmic structure formation, the concept of cosmic variance could help to further reduce the number of wavefunctions required. Indeed, given that we can only observe one universe, the statistical fluctuation in large angular patches is high, as not many statistically independent patches are available in our sky. This is a well-known fact when studying the CMB radiation. This means that the statistical error is anyway large on these scales, so we do not need to work with a very high precision. Let us also recall that the freedom of choosing the weight function in the scalar product (\[eq:scalar\_weight\]) of the eigenvalue problem may help to considerably reduce the number of wavefunctions. Even though this seems to be a promising route to take, we did not investigate it any further in this work.
Another area of interest could be the derivation of a scheme to generate initial wavefunctions analytically in the case of warm dark matter (see for instance [@Boyarsky2009]) or for any initial distribution with non-zero initial velocity spread.
Implementation & Numerical results {#sec:numerics}
==================================
In the previous two sections, we showed how the cosmological Vlasov-Poisson problem (\[eq:VP\]) can be approximated by the Schrödinger-Klein-Gordon system (\[eq:S\_KG\]). We showed that this approximation is valid in the limit $\hbar\rightarrow 0$, $c\rightarrow\infty$, $N\rightarrow\infty$. We also demonstrated how the wavefunctions can be built and that in general they can approximate the true density distribution in the limit $N\rightarrow\infty$. For some specific cases or for smart choices of eigenfunction basis, the exact $f(x,v)$ can even be ensured with a finite or low $N$. But let us keep the general case in mind.
Contrary to the [$N$-body ]{}framework, where the convergence towards the exact solution is not granted in general, we propose a method where we have a handle on the behavior of the simulation and where we are able to easily test the dependency of the result on the parameters $\hbar,c$ and $N$. This allows us to truly speak about converged results and understand the limits of our model.
Let us now present how this scheme can be discretized and implemented on a computer. We will present the implementation we used, which is probably the simplest version of what can be done.
Implementation {#ssec:implementation}
--------------
The simplest possible numerical scheme to solve partial differential equations is to use an explicit scheme in time. An implicit scheme would be more precise but would require more computing time and memory, the latter quantity being, as we will show, a rather scarce resource. This explains the choice of an explicit scheme, even if this imposes the use of a Courant-like condition for our time steps. For the same reasons a scheme accurate up to order $(\Delta \tau)^2$ in time has been chosen. As going to a precision of order $(\Delta \tau)^4$ would require almost twice as much memory, this choice can reasonably not be made. Using a symplectic integrator may, however, be useful in future studies as they do not cost more in terms of memory but conserve the energy of Hamiltonian systems exactly.
Regarding the spatial derivatives, there are no constraints coming from the memory requirements. One could in principle go to an arbitrary level of accuracy. But as the time derivatives only have a limited precision, it is not worth going to a precision higher than $(\Delta x)^4$ , using the usual five-point stencil.
With these two points being set, the system of equations (\[eq:S\_KG\]) can be written on a lattice as follows:
$$\begin{aligned}
\psi_n(x,\tau + \Delta \tau) &=& \psi_n(x,\tau-\Delta\tau) + i\frac{\hbar\Delta \tau}{a(\tau)} \nabla^2_{\rmn{dis}}
\psi_n(x,\tau) \nonumber\\
&& - i\frac{2\Delta\tau}{\hbar}U(x,\tau)\psi_n(x,\tau) \nonumber \\
U(x,\tau+\Delta\tau) &=& 2U(x,\tau) - U(x,\tau-\Delta\tau) \nonumber\\
&& +c^2\Delta\tau^2\nabla^2_{\rmn{dis}}U(x,\tau) \nonumber\\
&&-\frac{4\pi Gc^2\Delta\tau^2}{a(\tau)}\left(\sum_n^N\lambda_n|\psi_n(x,\tau)|^2-\bar\rho\right),\nonumber\end{aligned}$$
where the discretized divergence operator is given by
$$\begin{aligned}
\nabla^2_{\rmn{dis}} f(x) &=& \frac{1}{12 \Delta x^s} \left[\big.-f(x+2\Delta x) +16f(x+\Delta x)\right. \nonumber \\
&& \qquad\left. -30 f(x)+16f(x-\Delta x) -f(x-2\Delta x)\big.\right]. \nonumber\end{aligned}$$
In the non-cosmological case, the factors $a(\tau)$ can be dropped and one can use time $t$ instead of conformal time $\tau$. One can show that this scheme is unitary and conserves the norm of each wavefunction. Since the iterative solution contains the fields at neighbouring lattice sites, care has to be taken that the boundary conditions are implemented correctly. This is most easily done by augmenting the arrays containing the values of the fields on the lattice by so-called ghost points to store the periodic boundary conditions.
The last important point regarding the numerics is the choice of $c$ and $\hbar$. It is clear that the Klein-Gordon equation reduces to the Poisson equation in the limit $c\rightarrow \infty$ and that the higher order terms of (\[eq:Wigner\]) vanish in the limit $\hbar \rightarrow 0$. But numerical stability imposes more conditions on these values. An explicit scheme can only converge if there is no information propagating of a distance of one cell during one time step. The scalar field propagates at speed of $c$, which gives us the following condition:
$$\frac{\Delta x}{\Delta \tau} > c,$$
which is the usual *Courant condition*. In practice, the right-hand side is multiplied by a constant ($10 - 10^2$) in order to avoid any instability and to remain far from the actual condition. This condition gives a clear relation between those three quantities and shows that one cannot arbitrarily improve the spatial discretization without changing the time step size. It is not surprising to have to introduce such a condition. Indeed, if we were to truly use a value of $c=\infty$ in our simulations, we would have to use smaller and smaller time steps for a fixed grid spacing. At some point, solving the Poisson equation would become algorithmically cheaper. The Courant condition is thus the price to pay to avoid solving the usual Poisson $\mathcal{O}(M\log M)$ problem.
The evolution of the Schrödinger equation also imposes conditions on the time and space slicing. It can be shown that the following relation
$$\frac{\Delta x^2}{\Delta \tau} > \hbar$$
must hold, encouraging us, once again, to choose $\hbar$ as small as possible. At this stage, no lower bound has been analytically derived for $\hbar$. The full dependence on $\hbar$ of the simulation results is still an open question left for further investigation of this framework.
Complexity and memory requirements {#ssec:complexity}
----------------------------------
Having presented the algorithm of the time evolution, let us estimate its computational complexity and memory requirements. Consider a three-dimensional spatial grid made of $M^3$ lattice points. Let $N_\psi$ be the number of wavefunctions we evolve. Adding the spatial components of the scalar field, $N_f
=
N_\psi + 1$ is the total number of fields we evolve in time. At each time step, we need to compute each of the fields at every lattice point, making the algorithm of complexity
$$\mathcal{O}(M^3\cdot N_f).$$
This has to be compared with [$N$-body ]{}simulations, which have a naive complexity of $\mathcal{O}(N^2)$, that can be reduced to $\mathcal{O}(N \log N )$ using optimized algorithms. The more particles are tracked, the better becomes the spatial resolution. Roughly, for a total of $N$ particles, $\Delta x_{\mbox{resol}} \sim L_{\mbox{box}}/N^{1/3}$. In our case, the spatial resolution is defined by the lattice spacing $\Delta x_{\mbox{resol}} \sim L_{\mbox{box}}/M^3$. Thus for comparable spatial resolution, we would need $M \sim N^{1/3}$. From this we conclude that the complexity of our algorithm scales as $\mathcal{O}(M\cdot N_f)$. In the ideal situation where we only need a few wavefunctions, $N \sim \mathcal{O}(1)$, our new framework provides an $\mathcal{O}(M)$ algorithm to study structure formation. It seems that in the worst case we would need as many wavefunctions as there are Fourier modes on the lattice, $N_\psi \sim
\mathcal{O}(M^3)$ implying a complexity $\mathcal{O}(M^2)$, which is the same as the naive force summation in [$N$-body ]{}simulations.
These estimates illustrate that our algorithm can indeed compete with the complexity of [$N$-body ]{}simulations. It also shows how crucial it is to reduce the number of wavefunctions as much as possible.
Let us next have a look at the memory requirements of our approach. Given that our time evolution relies on a two-level explicit scheme, we need to keep the field configurations at two time steps in memory. For $N_\psi$ complex wavefunctions and one real scalar field components on the whole lattice, we need $2\cdot M^3(2N_\psi+1)$ variables. Assuming that each is stored as a `double` of 8 bytes, we can estimate the minimal memory needed by our numerical simulation to be
$$\geq 2 \cdot M^3(2N_\psi+1)\cdot 8~\mbox{bytes}.$$
Let us look once more at the worst case scenario $N_\psi \sim\mathcal{O}(M^3) \sim \mathcal{O}(N_\psi)$. Hence, the memory required now raises to
$$\geq 32 \cdot N_\psi^2~\mbox{bytes}.$$
This has to be compared with [$N$-body ]{}simulations, which have to store at least the position and velocity of each particle at every time step leading to a memory consumption of $$\geq 2\cdot 6 \cdot N \cdot 8~\mbox{bytes}.$$ As an example we may give the Millennium simulation [@Springel2005], which needed about 400 GB to store the information of their $2160^3 \simeq 10^{10}$ particles, in agreement with the above estimate. We have to conclude that our approach can be strongly constrained by its memory requirements. The gain in computational complexity seems to have come at a considerable cost in memory. If we consider the 1 TB of memory available to the Millennium simulation, we could only have $\sim 57^3$ lattice points! However, if we were to use as many wavefunctions as spatial lattice points, we could as well directly simulate the Vlasov-Poisson system without introducing any approximation. The whole point of the framework we introduced is to simulate a realistic probability distribution with a low number of wavefunctions, in which case the memory requirements are not prohibitive any more and scale with $N$ as in the [$N$-body ]{}case. We also mentioned the idea of using an adaptive mesh to improve the (spatial) resolution without having to increase the number of wavefunctions.
We now turn to two cases we simulated and show that this new framework is able to reproduce the known solutions. We also show how the solution depends on the parameters $c$, $\hbar$, $N$ and $\Delta x$.
Spherical collapse of a DM sphere {#ssec:collapse}
---------------------------------
There are few known non-trivial analytical solutions to the Vlasov-Poisson system (\[eq:VP\]) even in the static Universe ($a(\tau) = 1$) case. The collapse of a uniform sphere is among these and is of particular interest for cosmological applications. A comprehensive treatment of the case, known as *Tolman solution* [@Tolman1934], can, for instance, be found in the textbook [@Weinberg1972]. A uniform sphere of initial density $\rho_0$ and radius $R_0$ is collapsing under its own gravitational potential. Gauss’s law for gravity states that the evolution of a sphere is not influenced by the matter lying outside itself. This means that the density inside the sphere will remain constant with the radius at every time $t$. In other words, all matter will reach the centre at the same time which will lead to an infinite density. At this stage, the Newtonian description becomes invalid and one would have to use GR in order to take into account all the effects. In the framework of Newtonian gravity, the matter will simply cross the centre and oscillates around the centre. Due to the discretization, the simulated central density cannot become infinite and these oscillations cannot be reproduced exactly. The same shortcomings are present in [$N$-body ]{}codes.
The evolution of the radius $R$ with time is a quantity which can be easily tracked. In parametric form, the Tolman solution reads $(0\leq\beta\leq\pi)$: $$\begin{aligned}
t &=& \frac{\beta + \sin\beta}{2\sqrt{\frac{8\pi G}{3}\rho_0}}, \\
R &=& \frac{1}{2}(1+ \cos\beta).\end{aligned}$$ The density inside the sphere will evolve following the relation
$$\rho(r,t) = \frac{\rho_0R_0^3}{R^3(t)}.
\label{eq:Tolman}$$
For simplicity in what follows, we set $R_0=1$, $G=1$ and $\rho_0 = \pi$. The final collapse time (in arbitrary units) is then reached when $t_c \approx 0.306$. We will work on the periodic interval $[-5,5]$ which should be big enough to avoid any unwanted effects from the boundaries.
This problem possesses an obvious spherical symmetry and in order to be able to explore a wide resolution range it is interesting to re-derive the whole framework presented in the Section \[sec:framework\] and \[sec:IC\] using this assumption. A careful derivation can be found in appendix \[sec:spheric\] and the end result is that the Vlasov-Poisson system with spherical symmetry can be re-cast in the one dimensional Schrödinger-Klein-Gordon system
$$\begin{aligned}
i\hbar {\frac{\partial \psi_n}{\partial \tau}} &=& \frac{-\hbar^2}{2a(\tau)}{\frac{\partial ^2\psi_n}{\partial r^2}} + \frac{V}{r}\psi_n,\\
- \frac{1}{c^2}{\frac{\partial ^2V}{\partial \tau^2}}+ {\frac{\partial ^2V}{\partial r^2}} &=& 4\pi G r\left(\frac{2\pi}{r^2}\sum_n \lambda_n
|\psi_n|^2 - \frac{4\pi^2\Xi}{V_{\rm{tot}}} \right),\end{aligned}$$
where the potential $V = Ur$ and $\Xi$ is the normalization of the wavefunctions (see equation \[eq:normalisation\]). The main difference with the framework in presented earlier is the explicit dependency on the position coordinate $r$. The algorithms developed to find the wavefunctions corresponding to a given distribution function are identical.
To generate the initial set of wavefunctions and eigenvalues we chose to use the matrix formulation (Section \[ssec:SVD\]). The initial density profile being discontinuous, it is obvious that it cannot be recovered exactly with a finite set of continuous functions. There will be some noticeable differences between the exact density profile and its approximation appearing at the discontinuity points, that is at the edge of the sphere. It is thus better to use a approximately correct but continuous density profile. In the case at hand, we used the following initial setup:
$$\rho(r,t=0) = \frac{\pi}{2}\tanh\left(\xi(r+1)\right) - \frac{\pi}{2}\tanh\left(\xi(r-1)\right),
\label{eq:smoothed_IC}$$
with $\xi=20$. The value of $\xi$ is somewhat arbitrary and has been chosen in order to be as close as possible to the perfect sphere (i.e. high $\xi$) and avoid any Gibbs oscillation at the edge of the sphere (i.e. low $\xi$). The results presented here are not really dependent on $\xi$. This parameter has just been introduced for convenience and to avoid having to analyse the effects of these unwanted and unrealistic oscillations. In fact, even a value of $\xi=\infty$ yields comparable results to what is shown below once the Gibbs oscillations have been smoothed out manually from the output.
Once discretized on a lattice, the eigenvalue decomposition is straightforward to obtain, for instance using the SVD function implemented in the usual scientific software packages. Recall that there is no guarantee that the obtained functions will be periodic on the interval of interest or even that these function will be smooth. It is a pure matrix operation without any relation between the matrix elements representing the wavefunctions. The interval $[−5, 5]$ has been uniformly discretized regularly in $5000$ line elements in order to get a high enough spatial accuracy. This means that we want to perform the SVD decomposition of a $5000 \times 5000$ matrix and that we can use up to $N = 5000$ wavefunctions in the simulation. The matrix reads
$$\hat G_{ij} = \rho\left(\frac{r_i+r_j}{2}\right),$$
where the $r_i$’s are the uniformly distributed lattice points. This matrix is by construction symmetric and positive definite, meaning that its eigenvalues will be positive or null. Most of the SVD routines in scientific packages sort the eigenvalues $\lambda_n$ according to their magnitude which allows us to classify the most important contributions and discard the negligible terms in equation \[eq:WDF\] if one does not want to use all the $N$ functions. The first four wavefunctions are shown on figure \[fig:wavefunctions\].
The wavefunctions obtained through this procedure are smooth (at the lattice level at least) and real but are not periodic nor anti-periodic, which leads to spurious diffusion at the boundaries of the box. For this reason, we decided to multiply them by a square-box like compact function going to zero close at the box boundaries. The first four wavefunctions before and after applying this window filter are also shown on figure \[fig:wavefunctions\]. This procedure does not modify the distribution function obtained through the WDF. This reflects the fact that there is infinitely many ways to decompose the same $f(r,v)$ in wavefunctions. Notice that this procedure of adding a window function can only be done if the density vanishes at the boundaries.
![The first four wavefunctions contributing to the WDF of the approximate uniform sphere before (dotted lines) and after (superimposed solid lines) having applied the smooth window function to make them vanish at the boundaries of the box. These functions are different from zero almost everywhere but their combination in a WDF corresponds to the density profile (dashed black line, equation \[eq:smoothed\_IC\]), which is zero on most of the interval.[]{data-label="fig:wavefunctions"}](Figures/psi.eps){width="84mm"}
Apart from the wavefunction, the eigenvalue associated to each mode also enters the WDF (equation \[eq:WDF\]). These are obtained at the time than the discretized wavefunctions and their values are represented on figure \[fig:eigenvalues\]. The actual normalization of the eigenvalues does not really matter as any common factor can be absorbed as normalization in front of the WDF. But the ratio of the values plays a role. All the different wavefunctions (modes) entering the decomposition of $f(r,v)$ may not play an important role exactly as in the case of a Fourier series decomposition where some of the modes can safely be neglected. As can be seen on figure \[fig:eigenvalues\], the values of the various $\lambda_n$ decrease rapidly and for $n>100$, they represent less than $10^{-3}$ of the most important mode. As the eigenvalues are constants of motions, we can hope that neglecting modes with a high $n$ (and hence a small $\lambda_n$) will not affect the simulation too much. In fact, unless the magnitude of the wavefunction corresponding to one of the neglected mode grows significantly over the course of the simulation, this mode should remain small at all time and can thus be safely ignored.
![The first 1000 eigenvalues $\lambda_n$ corresponding to the SVD decomposition of the spherical collapse problem. The values decrease rapidly and become negligible (when compared to the first one) for $n > 100$. They even reach a minimum close to the machine epsilon for $n>700$. Our fiducial run uses all the wavefunctions up to $n=79$ which corresponds to $\frac{\lambda_n}{\lambda_0} > 10^{-3}$. This limit is shown as the red solid line on the figure. The small panel presents a zoomed-in region of the eigenvalues with $n<25$. The decrease on this small subset is already of more than an order of magnitude.[]{data-label="fig:eigenvalues"}](Figures/lambda.eps){width="84mm"}
In our main run, we used all eigenfunctions $\Psi_n$ whose eigenvalue fulfils $\lambda_n>10^{-3} \lambda_0$, which left us with only $N=79$ functions to evolve. The other numerical parameters we chose in our fiducial run are $c=10$ and $\hbar=0.005$. We do not expect $\hbar$ to have a big impact on the results in this case as the potential is a combination of a second and third order polynomial for which the higher order corrections in the Wigner equation (\[eq:Wigner\]) should be small.
{width="\textwidth"}
Figure \[fig:rho\_evolution\] shows four density profiles at different time steps in the simulation together with the analytical solution (equation \[eq:Tolman\]). Until $t\approx 0.2$, the behaviour of the density profiles remains close to the exact solution apart from the very edges of the sphere that are slightly smoothed. The centre of the density profile is almost flat as expected and has almost the correct value. When coming closer to the collapse time $t_c\approx0.306$, the profiles starts to deviate more and more from the expected profile. This can be seen on the last two panels of figure \[fig:rho\_evolution\] where the density inside the sphere is clearly different from a square box function. The very centre of the sphere still remains close to the analytical solution but the edges are not sharp any more and are smoothed over many lattice elements. This strongly suggests that the estimation of the derivatives of both the potential and the wavefunctions are getting poor or that the number of wavefunctions used in the run is not high enough. Our scheme uses a fourth order accurate derivative stencil but this does not necessarily help recovering sharp features such as the one present at the edge of the sphere. Increasing $N$ and reducing $\Delta r$ may help recover the right density profile everywhere in the sphere.
The results on figure \[fig:rho\_evolution\] have been obtained using $N=79$ wavefunctions corresponding to all eigenvalues $\lambda_n > 10^{-3}\lambda_0$. This should be sufficient as the eigenvalues are constants of motion and we do not expect any of the neglected wavefunctions to grow by a huge factor over the course of the simulation. In order to assess this, we run the same simulation with $N=155$, corresponding to all wavefunctions whose eigenvalues $\lambda_n > 10^{-5} \lambda_0$. Notice here that decreasing the minimal eigenvalue entering the WDF by two orders of magnitude only increases $N$ by a factor of $2$. We are thus far from the worst case scenario (see Section \[ssec:complexity\]) where the same number ($N=5000$) of wavefunctions than lattice points have to be used.
![Comparison of the density profiles at the initial time for $N=79$ (green solid line) and $N=155$ (blue dashed line) wavefunctions. The figure zooms in the central regions where the difference can be spotted. The $N=79$ line presents a lot of oscillations that are suppressed if more wavefunctions are used. The $N=155$ line almost perfectly matches the density profile given by equation \[eq:smoothed\_IC\]. Notice, however, that this differs from the perfect sphere profile (red dashed-dotted line), which cannot be represented by a finite set of continuous functions.[]{data-label="fig:comparison_N"}](Figures/comparison_N.eps){width="84mm"}
Figure \[fig:comparison\_N\] shows a comparison at $t=0$ of those two initial setups. The figure only shows a zoomed-in view focused on the sphere itself as the difference are less visible in the outer regions of the simulation domain. As can be seen, the $N=155$ initial setup (dashed blue line) is a much better representation of the smoothed density profile (equation \[eq:smoothed\_IC\]). At this resolution, the two are indistinguishable. The $N=79$ initial conditions (green solid line) presents some oscillations inside the sphere that are very similar to the Gibbs phenomenon that appears when computing the Fourier series of the square box function. Using a smoothed density profile and decomposing in eigenvalues using the matrix formulation thus yields a result which is very similar to generating the ICs through the Fourier decomposition (Section \[ssec:fourier\]). This could have been anticipated by looking at the wavefunctions (figure \[fig:wavefunctions\]), where the different $\psi$’s resemble sines and cosines functions at least qualitatively. As can be seen, the relative error introduced by using only $N=79$ wavefunctions is of the order $10^{-3}$, whereas the error computed when using $N=155$ is smaller than $10^{-6}$, showing once again that increasing the number of eigenfunctions used by a factor of $2$ increases the simulation by more than $2$ orders of magnitude. However, it should be noticed that using another basis or weighting function for the eigenvalue decomposition (\[eq:scalar\_weight\]) may yield another $N$ with the same or different accuracy. Comparing the number of wavefunctions only makes sense when using a similar decomposition technique. Let us also mention that we also tried using harmonic functions and Chebyshev polynomials for this test case and obtained similar results.
At later times, the simulation snapshots are identical to the ones presented earlier on figure \[fig:rho\_evolution\]. The relative difference between the two runs is of order $10^{-3}$ as in the initial conditions. This implies that the difference between our simulation results and the analytical solution can not be reduced by using more and more wavefunctions. The additional modes that have been discarded when using only $79$ eigenfunctions do not contribute significantly to the final results. This could have been expected as their weightings ($\lambda_n$) are very small compared to the main modes. We can thus gain confidence in the way we generate ICs, discarding higher order modes may not be an issue and we may be able to run our algorithm in a near linear regime even when a violent collapse of matter is studied.
In conclusion, increasing $N$ does make the initial conditions and the simulation outputs converge towards a solution at a high rate. However, the discrepancy between the solution and the simulation does apparently not come from the wrong choice of the parameter $N$. Let us now explore the dependency on the grid resolution.
![The output at $t=0.26$ for different lattice resolution using $N=79$ wavefunctions. The green dash-dotted line corresponds to a low resolution run with $\Delta r=5\cdot10^{-3}$, the blue dashed line corresponds to our fiducial run at $\Delta r = 2\cdot 10^{-3}$ and the black solid line is the output of a high resolution run using $\Delta r=10^{-3}$. The quality of the output is clearly improved by using a higher resolution lattice. This can be directly related to the problem of estimating sharp derivatives on a grid, where the only solution is to increase the resolution.[]{data-label="fig:comparison_dr"}](Figures/comparison_dr.eps){width="84mm"}
On figure \[fig:comparison\_dr\], we show the results of three runs at different grid resolutions leaving the number of wavefunctions and all the other parameters fixed. The blue dashed line corresponds to the fiducial run ($\Delta
r=2\cdot10^{-3}$), the green dotted line to a lower resolution run using $\Delta r=5\cdot10^{-3}$ and the black solid line corresponds to the high resolution run with $\Delta r=10^{-3}$. As can be seen, increasing the resolution has a huge impact on the quality of the result. As anticipated, the sharp features can only be resolved correctly when enough grid points are used. Notice that the high resolution run almost matches exactly a rescaled version of the initial density profile (equation \[eq:smoothed\_IC\]), but does break down at later times in the same way that the fiducial run did between $t=0.20$ and $t=0.26$ (figure \[fig:rho\_evolution\]). Increasing the resolution is thus important to be able to retrieve all features of this somewhat artificial test case. This test case presents a strong density gradient at the edge of the sphere which does not spread over many cells. This makes it difficult to resolve for a grid code but in a cosmological simulation such sharp gradients should not arise as the density profiles usually follow power laws and do not have infinite gradients. As already mentioned, using an adaptive mesh would help in such a case as more resolution elements could be used at the edge of the sphere without having to slow down the simulation due to an unnecessary oversampling of the steady regions.
This demonstrates that our framework converges towards the analytical solution once the spatial resolution is high enough and once the number of wavefunctions has been carefully chosen to represent the distribution function of interest.
![Evolution of the density at the centre of the sphere ($r=0$) for different values of the numerical speed of light $c$. The red dashed line corresponds to the analytical solution (\[eq:Tolman\]), the vertical dash-dotted line represents the final collapse time and the different solid lines correspond to the different values of $c$. The higher the value of $c$, the closer the line lies to the exact solution. The line with $c=\infty$ has been obtained by solving Poisson’s equation on the grid instead of evolving gravity using Klein-Gordon’s equation. The quality of the simulation outcome clearly depends on the value of $c$ but if the value is high enough (compared to the velocity of the matter), the difference with the $c=\infty$ becomes very small. Once the density peak has been reached, the value of $\rho(r=0)$ decreases as is expected after the different matter shells have crossed. The simulations have not been carried on much beyond this point as the departure from the analytical solution is already significant. Moreover, the peak can not be represented accurately by any numerical mean and any subsequent event would be erroneous.[]{data-label="fig:comparison_c"}](Figures/comparison_c2.eps){width="84mm"}
This new framework should converge towards the solution in the limit $N\rightarrow\infty$, $\Delta x\rightarrow 0$, $c\rightarrow\infty$ and $\hbar\rightarrow0$, the last two being, despite their physical origin, only numerical parameters. Figure \[fig:comparison\_c\] presents the evolution of the density at the centre of the sphere for our fiducial run and for higher values of $c$. The simulation with $c=\infty$ has been obtained by solving Poisson’s equation on the grid at every time step instead of using Klein-Gordon’s equation. Increasing $c$ improves the quality of the result and even relatively small values ($c=50$) of this parameter lead to a behaviour close to the limiting case. Poisson’s equation can thus safely be replaced by Klein-Gordon’s equation.The maximal speed reached by matter shells in our fiducial run is $p\approx10$ before the very end of the collapse, which can anyway not be studied by a simulation. Using a value of $c=10$ is thus intuitively too low and this plot confirms this. The speed of gravity must be at least a few times bigger than the matter velocity.
Once the peak has been reached, the different matter shells should cross the centre and the density at $r=0$ has to decrease. The start of this behaviour can also be seen in figure \[fig:comparison\_c\]. The main issue with this analysis is that is happening after the moment where the density at the centre becomes infinite and hence not representable on a computer. In practice, all the wavefunctions should become infinite at this precise point and zero elsewhere. This is obviously impossible on a lattice and does anyway lead to inaccurate derivatives. To get closer and closer to the singularity requires a finer and finer mesh. The smaller the mesh size, the better the shell crossing can be followed.
Notice, however, that this is an issue present in this ideal sphere case only. In a realistic scenario, where the matter has a non-zero radial velocity and in an expanding background, the usual NFW profiles [@Navarro1996] should be recovered without singularity problems. This would, however, require a truly 3D simulation and not just a spherically symmetric 1D setup.
Increasing $c$ has a big impact on the simulation run time as the time step size varies as $c^{-1}$, making the total simulation wall clock time proportional to $c$. An option that has not been explored here is to change the value of $c$ to be always a (small) multiple of the maximal matter speed in the simulation. This would allow us to choose bigger time steps in the early stage of the simulation when all the matter moves slowly. It would also avoid making an initial guess for the value of $c$ without knowing how fast the matter will move during the run.
As discussed earlier, the dependency on $\hbar$ is difficult to test in this case as the analytical potential only presents first order corrections in the Wigner equation. We did run some simulations with various values of this parameter without noticing important differences in the behaviour of the matter distribution. Understanding the exact dependency on $\hbar$ of the framework is left to a future work.
This simple spherical collapse test showed that we were able to reproduce the analytical solution in the limit $N\rightarrow\infty$, $\Delta x\rightarrow0$ and $c\rightarrow\infty$ as expected. We investigated the different deviations from the exact solution and could explain them through our choices of numerical parameters. We also discussed how the implementation could be improved by using a mesh-refinement and adaptive $c$ values. The results obtained so far show that this new framework can reproduce known solutions and give us confidence to use it on more involved cases.
Going beyond the first collapse {#ssec:nonLinear}
-------------------------------
With the previous test case, we showed how our framework was able to reproduce the collapse of a matter distribution in the linear regime and studied the dependency on the model parameters. However, in most cases of interest, the systems considered in simulations are way past the linear regime. They also present multiple matter streams, i.e. at a given position $x$, there are multiple velocities $v$ and the distribution function is “wound up”. It is hence important to explore whether this behaviour can be recovered by our framework. Note that tracking precisely these multiple matter streams is extremely difficult in the case of [$N$-body ]{}simulations unless advanced phase-space tessellation techniques are used [@Abell2012; @Shandarin2012].
The test case presented in the previous section exhibits a nice analytical solution but, as discussed, the matter distribution becomes infinitely thin at the time of the collapse which makes all attempts at taking derivatives difficult. To alleviate this issue, we use a simpler one dimensional test case with a much smoother density distribution. In this section, we study the evolution in one dimension of the cold distribution function
$$f(x,v) = \rho(x)\delta(v), \qquad \rho(x) = \rho_0\exp(-x^2 / 2s^2),$$
with $s$ the scale size of the matter distribution. This test case has already been studied by [@Widrow1993] in the context of their framework which makes use of a Husimi distribution instead of the Wigner one. We will use a periodic domain of size $L_{\rm{box}} \gg s$.
The first step in the algorithm is to decompose the initial condition into a series of wavefunctions. There are many ways to do this and one could easily use either a decomposition in terms of sine waves or using the matrix decomposition used in the previous test case. The decomposition in Fourier modes is straightforward and the initial distribution function can be recovered in a satisfactory way with less than $10$ wavefunctions. However, to demonstrate the fact that the number $N$ of wavefunctions is only a relevant quantity once a decomposition scheme has been chosen, we will use a simpler single wavefunction to represent $f(x,v)$:
$$\Psi(x,t=0) = \sqrt{\rho(x)}.\label{eq:3}$$
Using this simple decomposition leads to a an initial Wigner distribution of the form
$$f(x,v) = \rho(x)\exp(-v^2/2\hbar^2),$$
once equation \[eq:initial\_distribution\] has been applied. This example also explicitly shows how $\hbar$ enters the framework and the effect this quantity has on the initial conditions and hence on the subsequent evolution of the distribution function. The previous test case gave us some insights into how to choose the value of the speed of light $c$. We could use similar considerations here to choose an appropriate value, however, to simplify the discussion, we choose to set $c=\infty$ and solve Poisson’s equation for gravity at every time step using the fast Fourier transform algorithm. In what follows, we set $\rho_0=1$, $s=10^{-2}$, $\hbar=10^{-3}$, $L_{\rm{box}}=1$ and discretize our volume in $M=100$ intervals.
![Evolution of the density at the centre of the domain ($x=0$) beyond the first collapse ($t>t_c$). The multiple collapses and re-expansions of the matter distribution can be tracked by the framework. The appearance of multiple matter streams during the evolution of the collapse can be resolved by the simulation even with one single wavefunction.[]{data-label="fig:nonLinearEvolution"}](Figures/evolution){width="84mm"}
To trace the non-linear evolution of the system, we trace the value of the density field at $x=0$. This is in essence similar to figure \[fig:comparison\_c\] for the previous case but we now let the simulation run past the initial collapse time. The result of this evolution is shown on figure \[fig:nonLinearEvolution\]. As can be seen, the first peak is followed by a relaxation of the system and then by a series of additional regularly spaced collapses that occur every time the matter distribution crosses the spatial origin. The first four peaks can be well followed despite the relatively low spatial resolution and the single wavefunction used in this example. Using a higher value of $M$ leads to more peak being resolved and less oscillations in the value of the central density. This, once again, highlights the key importance of the spatial resolution over the raw number of wavefunctions. This is also true in standard [$N$-body ]{} simulations that use meshes to solve Poisson’s equation. The quality of the solution is mostly driven by the high number of grid elements and less by the pure number of particles used in the simulation. Our framework is hence able to track the collapse of a matter distribution when multiple shell crossings occur and in the presence of multiple matter streams.\
It is interesting to discuss what would happen if more wavefunctions were used to represent $f(x,v)$. Obviously, one cannot add more $\Psi$ to the decomposition given by equation \[eq:3\] as it already provides a exact match to the density profile; any addition would reduce that agreement. Note that one might want to consider doing so as it could reduce the spread in velocity and hence give a better set of initial conditions but there does not seem to be a simple way to do so. Alternatively, one might consider using a Fouried decomposition of $\rho(x)$ (section \[ssec:fourier\]) with a high enough number of cosine waves to reproduce $\rho(x)$. A small number of waves will be sufficient as the case is smooth enough and by doing so, the spread in momentum can be reduced. The more wavefunctions are used, the smaller the initial spread in velocity, allowing us to get rid of the explicit dependence on $\hbar$ in the initial Wigner distribution. This obviously comes at a higher numerical cost but might be necessary in some situations. The freedom of getting a spread in velocity space smaller than $\hbar$ is a fundamental difference between or framework and earlier work based on the Husimi function [@Widrow1993; @Davies1997].
Linear structure growth in [$\Lambda$CDM ]{} {#ssec:growth}
--------------------------------------------
We now apply this new framework to a simple example of cosmic perturbation growth. We will consider the simplest possible case of a constant background $\bar\rho$ in a cold dark matter Universe and a small perturbation $\epsilon \ll
1$ with a single Fourier mode $k_p$ taken along the $x$ direction:
$$\rho(\vec{x},t) = \bar\rho + \epsilon\left[\cos(k_p x_x) + \sin(k_p x_x) \right].$$
This basic setup should be sufficient to study the behaviour of the framework in an expanding Universe case.
Generating the wavefunctions corresponding to this initial distribution function was discussed in Section \[ssec:fourier\]. The equations (\[eq:Fourier\_decomposition\]) define a representation of the density in terms of wavefunctions. As we only have one single mode, we only need one wavefunction for the constant background ($\psi_0$) and two for the perturbation. We run the simulation on a $30^3$ spatial lattice corresponding to a physical box size of $60~\rm{Mpc}$. It is important to notice here the low number of wavefunctions $N = 3 \ll 30^3$, allowing us to run our algorithm in a near linear regime. We choose the scale of the perturbation to be larger than the Nyquist frequency and small compared to the box size to avoid unwanted effects due to the limited box size.
In a purely matter dominated (Einstein-de Sitter) Universe, the scale factor $a(\tau)$ will grow as the square of the conformal time. Without loss of generality, we can normalise it such that it is equal to one at the start of the simulation $a(\tau_{\rm{ini}})=1$, implying
$$H^2_{\rm{ini}} = \frac{8\pi G}{3} \bar\rho_{\rm{com}} a(\tau_{\rm{ini}})^{-3} = \frac{8\pi G}{3} |\psi_0|^2.$$
The above relation fixes the value of this wavefunction in terms of the initial Hubble parameter, which can be computed by rescaling today’s value $H_0$ to the redshift corresponding to the beginning or our simulation
$$H_{\rm{ini}} = H_0(1+z_{\rm{ini}})^{3/2}.$$
We ran our simulations for the choice $z_{\rm{ini}} = 1000$ and using today’s Hubble parameter $H_0 \simeq
70~\rm{km}~\rm{s}^{-1}~\rm{Mpc}^{-1}$. The initial conditions with a density contrast of $\delta_{\rm{ini}} = 10^{-6}$ where evolved up to a redshift of $z_{\rm{fin}} = 200$. We use a normalised time line such that $z_{\rm{ini}}$ corresponds to $ \tau=0$ and $z_{\rm{fin}}$ corresponds to $\tau=1$ using $3\cdot 10^4$ time steps. The same initial perturbations were evolved in a matter-dominated, expanding universe and in a static universe without expansion.
The parameter $c$ has been chosen in accordance with the results of the previous test by making it bigger than the speed of the matter in the simulation and small enough to avoid drastically pulling down the time step. In what follows, $c=10$. The parameter $\hbar$ has been, once again, chosen small enough for the quantum corrections to be negligible. More specifically, this means that the first quantum correction in the Wigner equation (\[eq:Wigner\]) has to be small compared to the contribution to the classical Vlasov equation:
$${\frac{\partial V}{\partial x}}{\frac{\partial P_W}{\partial v}} > \frac{1}{24}\hbar^2{\frac{\partial ^3V}{\partial x^3}}{\frac{\partial ^3P_W}{\partial v^3}}.$$
We verified that this indeed the case in our simulations when using $\hbar=0.005$. We could, in principle, also verify that the higher-order corrections are also suppressed but computing the fifth derivative of the potential will lead to a very noisy estimate and may not lead to useful results.
Figure \[fig:density\_growth\] shows the time evolution of the density. The initial amplitude of the harmonic density increases with time, without distortion of the shape, as expected from the linear regime of structure formation. The growth of structure seems thus to be well reproduced by our framework even with such a low number of lattice points and wavefunctions. The simulation could, in principle, be carried on to a much lower redshift than $z=200$ but at some point, the spatial resolution issues highlighted in the previous test would appear here as well. Recall that we have only 30 grid points in our $60~\rm{Mpc}$ box. As soon as the variation of the density becomes important on a scale of order a few $\rm{Mpc}$, the discretized derivatives will cease to approximate the analytical ones and our formalism will break down as would any uniform grid code with the same resolution. We, hence, decided to restrict ourselves to the regime where our density field and the wavefunctions are well behaved in order to make a useful analysis of the results.
To analyse the growth of the perturbation in more detail, we performed a Fourier transform on the density contrast to obtain $|\delta_k|^2$. In this way we could also check that no other Fourier modes than the one initially present were excited during the simulation. This is a cross-check for the linearity of the evolution of the small density perturbation. The figure \[fig:growth\_comparison\] compares the growth $|\delta_k(\tau)|^2/|\delta_k(\tau_{\rm{ini}})|^2$ for our mode in the expanding and non-expanding universes. Clearly, the growth of the perturbation is suppressed in presence of expansion.
![Time evolution of the density field in an expanding universe. The different lines correspond to various time steps in normalised units. The values are taken along one line parallel to the $x$-axis in the box but all lines yield the same results. The initial amplitude of the harmonic density increases with time, without distortion of the shape, as expected from structure formation in the linear regime.[]{data-label="fig:density_growth"}](Figures/density_growth.eps){width="84mm"}
![Comparison of the growth of the perturbation $|\delta_k(\tau)|^2/|\delta_k(\tau_{\rm{ini}})|^2$ in a non-expanding universe (red squares, upper line) and in a matter-dominated, expanding universe (blue circles, lower line) as a function of the conformal time $\tau$. As expected, the growth is clearly suppressed in the presence of expansion.[]{data-label="fig:growth_comparison"}](Figures/growth_comparison.eps){width="84mm"}
These results clearly show that our framework is able to follow the growth of a single-mode density perturbation in an expanding background. The main features are recovered even when a low number of lattice points and wavefunctions is used. By taking advantage of the ease of decomposition in orthonormal Fourier modes of the cosmological power spectrum (Section \[ssec:cosmo\_IC\]) more complex cases can be studied by superposing the different modes. The results obtained here give us confidence about the behaviour of the framework in the non-linear regime of cosmic growth. The main features of [$\Lambda$CDM ]{}can probably be recovered in a higher-resolution run with more wave functions and a longer run time.
As in the previous test case, one could track the matter distribution into the non-linear regime and track the appearance of multiple matter streams. This is of course of crucial importance for realistic simulations of structure growth in the Universe. It is, however, obvious that the addition of the scale factor $a(\tau)$ in the simulation will not alter the behaviour seen in section \[ssec:nonLinear\] and we are confident that multiple streams would also appear and be correctly tracked by the evolution of the wavefunctions. A more detailed study of the framework in the context of cold or warm dark matter cosmologies is left for future work.
Conclusion {#sec:conclusion}
==========
We introduced a new alternative framework for simulation of structure formation which is not based on the usual discretization of the density field in a set of particles. We made use of the Wigner distribution function to recast the distribution function in a set of wavefunctions. We could thus replace the 6-dimensional Vlasov equation by a set of Schrödinger equations acting on the wavefunctions. The Poisson equation for gravity has been transformed into a Klein-Gordon equation making the system of equations completely local. We demonstrated how this system of equation could be derived from a Lagrangian and how the total energy and mass are conserved by the equations of motion.
We presented different methods to generate the initial conditions depending on the distribution function of interest and described how a cosmological power spectrum can be discretised in a low number of wavefunctions. The framework has then be tested on two simple models to assess its validity and the dependency of the outcome on the numerical parameters has been sketched. The results obtained thus far show that this framework is viable and may become a possible alternative to the [$N$-body ]{}method.
The important new features introduced in this framework are the possibility to simulate a generic distribution function and not only cold dark matter. Although finding an easy and generic way to generate initial conditions for warm or hot dark matter remains an open question, there are no intrinsic limitations in the framework that could prevent such simulations. It also provides an alternative to [$N$-body ]{}codes and could thus help assess the validity of simulations. Our technique can be shown to converge towards the solution in the limit $c\rightarrow\infty$, $\hbar\rightarrow0$ and $N\rightarrow\infty$ making the formal convergence studies possible. The computational complexity of the algorithm grows as $\mathcal{O}(N\cdot M)$ where $M$ is the number of lattice points. This demonstrates the importance of finding the appropriate decomposition of the distribution function in wavefunctions. The complexity can hence be anything between linear and quadratic in the number of points. The case of structure formation may be close to the ideal case thanks to the possibility to discretise the power-spectrum in a low number of modes.
This scheme is especially aimed at tackling the fundamental challenges that the [$N$-body ]{}method faces when dealing with non-CDM cosmologies. This includes simulation of a WDM Universe but also neutrino components in a standard [$\Lambda$CDM ]{}model or any other particle with non-negligible thermal velocities. At the same time, exploring CDM through this framework might help understand more precisely the limitations of the [$N$-body ]{}method by comparing results in the same way that various hydrodynamic solvers help understand the behaviour of the codes and their limits.
One could also argue [@Sikivie2010] that such an approach may be appropriate to simulate axions which remain quantum during the entire cosmological evolution. In such a case, the real value of $\hbar$ and particle mass would have to be used, which would, however, probably lead to very high computational costs.
In this paper, we presented the validity of the method but many promising and interesting options have not yet been explored. The first obvious domain to investigate is the dependency on $\hbar$ of the results. Early results tend to show that it may not be a crucial issue thanks to the universal gravitational profiles being low-degree power laws and hence generating only small quantum corrections to the Vlasov equation. It still remains an open question.
The other important area of investigation is the generation of initial conditions for more general cases than simple CDM. The procedures presented here can not be applied without making some educated guess on the best shape of harmonic functions or without having to solve gigantic matrix eigenvalue problems. Combining some of these procedures or using interpolation techniques between lattice points are possible improvements worth exploring.
Finally, on the implementation side, lot of work can be done to make the codes more efficient. We already discussed the possibility of using an adaptive mesh to refine the grid in the regions of interest. It may also be possible to use an adaptive value of $c$ and of the time step in the same way that [$N$-body ]{}codes use different time bins for different particles. The locality of the interactions is an important feature as it makes the parallelisation of the code straightforward. Running such a simulation on big clusters could thus be easily achieved without having to worry too much about complex communications and scalability issues.
Let us conclude by stating that our approach has a number of attractive features. Most importantly, the full phase space information is encoded in the wavefunctions. Working with many wavefunctions, we are in principle able to represent any given phase space distribution, including those where the velocity dispersion is important. Potentially, this would allow for numerical simulations of structure formation in presence of warm dark matter.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by the Swiss National Science Foundation and by the Tomalla Foundation. We would like to thank S. Cole, A. Maccio, J. Read and T. Theuns, for useful comments and discussions. O.R. acknowledges the support in part by the National Science Foundation under Grant No. PHYS-1066293 and the hospitality of the Aspen Center for Physics.
Spherically symmetric case {#sec:spheric}
==========================
The framework presented in section \[sec:framework\] can be simplified in the case of (spatially) spherically symmetric distribution functions. The dimensionality of the problem is then reduced and allows more comprehensive convergence studies thanks to the lower number of discretization points needed.
If we consider only radial motion, then the distribution function can only depend on the distance to the centre $r$, the radial velocity $v_r$ and the angle between those two vectors. We choose to use the cosine of this angle as our coordinate, denoted as $y$ in what follows. The gravitational potential does only depend on the distance to the centre. We thus have $f\equiv f(r,v_r,y)$ and $U\equiv U(r)$. The density at a given $r$ and total mass can be expressed using these new coordinates and read
$$\begin{aligned}
\rho(r) &=& \frac{2\pi}{a^3(\tau)} \int_0^\infty v_r^2 dp_r \int_{-1}^1 dyf(r,v_r,y),\\
M &=& 4\pi \int_0^\infty r^2 dr \rho(r).\end{aligned}$$
It can be shown that the total mass is a conserved quantity under the equations of motion for $f$. The Vlasov-Poisson system using those coordinates and assuming spherical symmetry becomes
$$\begin{aligned}
\frac{\partial f}{\partial\tau} + \frac{yv_r}{a(\tau)}\frac{\partial f}{\partial r} -
a(\tau)\frac{\partial U}{\partial r}\left[y\frac{\partial f}{\partial v_r}+\frac{(1-y^2)}{v_r}\frac{\partial
f}{\partial y}\right] \nonumber\\
+ \frac{(1-y^2)v_r}{ra(\tau)}\frac{\partial f}{\partial y} = 0,
\Bigg.\\
\frac{1}{r^2}\frac{\partial}{\partial r}\left(r^2 \frac{\partial U}{\partial r}\right) =
4\pi Ga(\tau)^2 \left( \rho(r) - \bar\rho \right).\end{aligned}$$
It may, in principle, be possible to find a Wigner-like distribution function for which the Wigner equation corresponds to this Vlasov equation. The wavefunctions entering such a distribution would probably obey a spherically symmetric version of Schrödinger’s equation. This is, however, not the only way to handle this system.\
The distribution function can be decomposed in two parts, one for each sign of the coordinate $y$:
$$f(r,v_r,y) = f_-(r,v_r)\delta_-(y+1) + f_+(r,v_r,\tau)\delta_+(y-1),$$
where $\delta_\pm(x)$ are Dirac distributions defined on the interval $[-1,1]$ only. We can then integrate over $y$ and obtain two equations, one for $f_+$ and another identical up to the signs for $f_-$ together with a boundary condition ensuring that the two distributions match when they reach $r=0$ or $v_r=0$. The next step in the procedure is to rescale these distribution functions by introducing $g_\pm(r,v_r) = f_\pm(r,v_r)r^2v_r^2$ and define a combined distribution $h(r,v_r)$ such that
$$h(r,v_r) = \left\lbrace
\begin{array}{rcl}
g_+(|r|,|v_r|) & \rm{if} & rv_r > 0\\
g_-(|r|,|v_r|) & \rm{if} & rv_r < 0\\
\end{array}
\right.\label{eq:h}$$
This new distribution function will obey the following Vlasov equation
$${\frac{\partial h}{\partial \tau}} + \frac{v_r}{a(\tau)} {\frac{\partial h}{\partial r}} - a(\tau){\frac{\partial U}{\partial r}}{\frac{\partial h}{\partial v_r}} = 0,$$
which is identical to the 1D Vlasov equation (\[eq:VP\]). The difference being in the definition of density and mass that now read $$\begin{aligned}
\rho(r) &=& \frac{2\pi}{r^2R^3(\tau)} \int_{-\infty}^\infty dv_r h(r,v_r, \eta),\\
M &=& \frac{4\pi^2}{a^3(\tau)} \int_{-\infty}^\infty dr \int_{-\infty}^\infty h(r,v_r,\eta)db.\end{aligned}$$ As we are back to the well-known case of Cartesian coordinates (at least for the Vlasov equation), we can introduce the same decomposition in terms of wave functions than in Section \[ssec:WDF\]. We will thus solve a set of 1D Cartesian Schrödinger equations alongside a 3D spherically symmetric Poisson equation with a slightly odd density definition. Using the usual trick $V(r) = U(r)ra(\tau)$, the laplacian term in Poisson’s equation can be simplified and the system we want to evolve reads
$$\begin{aligned}
i\hbar {\frac{\partial \psi_n}{\partial t}} &=& - \frac{\hbar^2}{2a(\tau)}{\frac{\partial \psi_n^2}{\partial r^2}} + m\frac{V}{r}\psi_n
\Bigg.,\\
{\frac{\partial ^2V}{\partial r^2}} &=& 4\pi Gr\left(
\frac{2\pi}{r^2} \displaystyle\sum_n \lambda_n |\psi_n(r)|^2 - \frac{4\pi^2\Xi}{V_{\rm{tot}}}\right),
\label{eq:spheric_Poisson}\end{aligned}$$
where $\Xi$ is the normalization of the wavefunctions that can be related to the total mass of the system through
$$M = \frac{4\pi^2}{a^3(\tau)}\sum_n \lambda_n \int_{-\infty}^{\infty} |\psi_n(r)|^2 dr =
\frac{4\pi^2\Xi}{a^3(\tau)}.
\label{eq:normalisation}$$
A dynamical term can then be added to equation \[eq:spheric\_Poisson\] to make the framework entirely local as discussed in Section \[ssec:local\]. The system can eventually be evolved as if it was a purely one-dimensional problem. The only difference being the more complicated density terms sourcing Klein-Gordon’s (or Poisson’s) equation and the $1/r$ term in the potential of Schrödinger’s equation.
The generation of initial conditions can be done in exactly the same way than outlined in Section \[sec:IC\]. The only difference being the use of the modified distribution $h(r,v_r)$ (equation \[eq:h\]) instead of $f(r,v_r,y)$ as the starting point of the procedure.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: It may sound surprising that the equation for the harmonic oscillator reduces exactly to the classical Vlasov equation, even though we know that the quantum mechanical treatment introduces discrete energy levels. In this case the quantum information is encoded purely in the initial conditions.
[^3]: This formulation of the statement is not fully satisfying, as the true semi-classical limit is also a statement about the properties of the wavefunction, and not identical to sending $\hbar\rightarrow 0$ which is anyway a dimensional parameter.
[^4]: For the sake of simplicity we restrict the analysis of this section to the one dimensional case, but the generalization to the 3D case is straightforward.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Doris Folini
- Jean Heyvaerts
- Rolf Walder
bibliography:
- 'folini.bib'
date: 'Received 22 August 2002 / Accepted 1 October 2003'
title: Structuring and support by Alfvén waves around prestellar cores
---
Introduction {#sec:intro}
============
Magnetic fields are observed in at least some molecular clouds [@1999ApJ...520..706C; @bourke-et-al:01]. Whether all molecular clouds are threaded by magnetic fields is still under debate. @2000ApJ...537L.135W observe ordered magnetic fields on small scales of about 0.05 pc in, adopting their terminology, prestellar cores ($N \approx
10^{5}$ cm$^{-3}$). Also on somewhat larger scales, in star forming regions, ordered magnetic fields are reported [@2002ApJ...571..356M; @2002ApJ...569..304M]. Coherent velocities in prestellar cores are observed on scales of about 0.01 pc [@1998ApJ...504..207B]. On larger scales, observed line widths indicate supersonic motions. Taken together, these observations suggest dense condensates, threaded by magnetic fields, to be embedded in a supersonically turbulent environment. The generation of magnetic waves under such conditions is inevitable.
On larger scales, such magnetic waves are likely to be strongly damped (e.g. by ion-neutral friction or instabilities) or dominated by other processes (e.g. ISM-turbulence or incoming magnetic waves as studied by @elmegreen:99). This finding is in agreement with molecular cloud theories and observations. While some years ago it was thought that molecular clouds had to be supported against their self-gravity for at least $10^{8}$ years, new results are much more in agreement with a picture in which molecular clouds form, stars are born, and the clouds are dispersed, all within some $10^{6}$ years. Observations of molecular clouds in the solar neighborhood show that most clouds do form stars [@2001ApJ...562..852H], from which it is concluded that star formation begins essentially as soon as a molecular cloud forms. Using stellar evolutionary tracks leads to the further conclusion that star formation in a molecular cloud takes place rapidly, once it has started [@palla-stahler:00]. For stellar populations with an average age larger than about 3 Myr, no more molecular material can be detected [@2001ApJ...562..852H and references therein], indicating that star formation also ceases rapidly. Numerical simulations also support such a dynamical scenario [@1999ApJ...527..285B; @2000ApJ...530..277E; @maclow:02].
For smaller spatial scales, on the other hand, recent observations and simulations support the idea that magnetic fields and waves play an important role in the structuring of the environment of – possibly only transient – high density molecular clumps and the inhibition of accretion onto such clumps. Observations of the starless dense core L1512 and its immediate vicinity by @2001ApJ...555..178F show six dense filaments pointing towards the core and extending up to about 1 pc. The matter within each filament is observed to move towards the core while probably describing circular motions in the direction transverse to the filament. This motion and the orientation of the filaments make it likely that they are not merely part of the turbulent cascade within the cloud.
@2001ApJ...547..280H performed grid studies for 3D hydrodynamical and MHD simulations of the formation of persistent cores. They find that increasing spatial resolution leads to increased accretion in the hydrodynamical case, but to a decreased one in the MHD case. The authors ascribe this difference to better resolution of MHD waves, which then counteract accretion. For the case of a plane-parallel slab, at the boundary of which a monochromatic Alfvén wave is injected, 2D simulations with constant gravity by @pruneti-velli:97, as well as 2D and 3D simulations without gravity by @del-zanna-et-al:01, show the development of high-density filaments parallel to the direction of propagation of the Alfvén wave. Note that filamentation is observed only when open, not periodic, boundaries are used at the planes perpendicular to the direction of wave-propagation. @2002ApJ...566L..49C have shown that magnetic fields can have rich structures well below the viscous dissipation scale, possibly affecting the density structure as well.
In this paper, we study the effect of magnetic waves in the framework of a simplified model, a 1D plane-parallel, self-gravitating slab with one central source of monochromatic, circularly polarized Alfvén waves. Related models have been investigated by other authors before [@gammie-ostriker:96; @martin-et-al:97; @1999ApJ...517..226F; @2002MNRAS.329..195F; @kudoh][^1]. The work here differs from previous 1D simulations in that we focus on the highly nonlinear, long-term evolution. Also, we consider only one central source of Alfvén waves, instead of injecting energy at each grid point, and we use open boundaries, not periodic ones. For this setting, we present a dimensional analysis as well as a parameter study based on numerical simulations.
Our results show that already one source of waves is sufficient to structure and support a turbulent slab in a quasi-static manner. The WKB solution for an Alfvén-wave supported, self-gravitating 1D slab by @martin-et-al:97 gives a good order of magnitude estimate for the average spatial extent of the turbulent slab, but fails to account for the rich interior structure. And while the Poynting-flux is constant in the WKB solution, there is substantial loss of Poynting-flux in the solution of the full equations. As governing parameters for deviations of the numerical solution from the analytical WKB solution we identify the initial, central Alfvén wave-length and the initial, central plasma beta. This is remarkable in view of the highly nonlinear, turbulent nature of the slab, where, for example, the true central Alfvén wave-length loses, even on average, any connection with its initial value after a fraction of a free-fall time. To observe both the structuring and support of the slab by magnetic waves, we find a good spatial resolution and high order of integration to be decisive.
The paper is organized as follows. In Sect. \[sec:model\] we describe our physical model and the numerical method we use. We give a dimensional analysis of the problem in Sect. \[sec:dimensionless\] before proceeding to the numerical results in Sect. \[sec:num\_results\]. A discussion of our results follows in Sect. \[sec:discussion\], conclusions are given in Sect. \[sec:conclusions\].
The model {#sec:model}
=========
Physical model problem {#sec:phys_mod}
----------------------
We consider a 1D (x-direction), plane-parallel, self-gravitating slab which we assume to be symmetric with respect to a central plane at $x=0$ (yz-plane, infinitely extended), where an Alfvén-wave is injected. In this geometry, all variables are functions of distance $x$ to the central plane and time $t$ only. Velocities and magnetic fields perpendicular to the x-direction are allowed, but gradients can occur only in the x-direction. To describe the time evolution of this slab, we use the ideal, isothermal MHD equations, including a source term to account for self-gravity. In their conservative formulation, these equations read: $$\begin{aligned}
\label{eq:div}
\frac{\partial \rho}{\partial t} +
\nabla \cdot (\rho \vec{v}) & = & 0, \\
\label{eq:mom}
\frac{\partial (\rho \vec{v})}{\partial t} +
\nabla \cdot \left[\rho \vec{v} \vec{v} +
p_\mathrm{tot}\vec{I} -
\frac{\vec{B}\vec{B}}{\mu_\mathrm{0}}
\right] & = & \rho \vec{g}, \\
\label{eq:mag}
\frac{\partial \vec{B}}{\partial t} +
\nabla \cdot (\vec{v}\vec{B} - \vec{B}\vec{v}) & = & 0. \end{aligned}$$ Here, $\rho$ denotes the mass density, $\vec{v} =
(v_\mathrm{x},v_\mathrm{y},v_\mathrm{z})$ the velocity, and $\vec{B} = (B_\mathrm{x},B_\mathrm{y},B_\mathrm{z})$ the magnetic field. We will often use the notation $B_\mathrm{\parallel}$ for the constant background magnetic field $B_\mathrm{x}$, as well as $B_\mathrm{\perp}$ for the magnitude of the transverse magnetic field, and $B_\mathrm{\perp 0}$ for $B_\mathrm{\perp}$ at $x=0$ and $t=0$. Generally, a subscript $0$ to a variable refers to the initial ($t=0$) and, if space dependent, central ($x=0$) value of a variable. $\mu_\mathrm{0}$ is the magnetic permeability of free space, and $\vec{I}$ is the identity tensor. $p_\mathrm{tot} =
p_\mathrm{th} + (1/2\mu_\mathrm{0})\vec{B}^{2}$ denotes the total pressure, where the thermal pressure is given by the isothermal equation of state $p_\mathrm{th} = \rho c_{\mathrm{s}}^{2}$, with $c_{\mathrm{s}} =
\sqrt{RT}$ the isothermal sound speed. $T$ is the temperature of the gas and $R$ is the gas constant. Within the framework of our 1D model, the force exerted by self-gravity in the x-direction is given by $\vec{g} = (g(x,t),0,0)$ with $$g(x,t) = - 4 \pi G \int_{0}^{x} \rho(x',t) dx'.
\label{eq:sg}$$ $G$ denotes the gravitational constant. We consider only one sort of particle and consequently have no wave damping due to ion-neutral friction in our model. The average mass per particle is the mass $m_\mathrm{H}$ of a hydrogen atom.
For special cases, analytical solutions exist. For an infinitely extended slab with $\vec{B}=\vec{0}$, the stationary solution of the above model problem is given by the hydrostatic density distribution [@spitzer:68], $$\rho(x) = \rho_\mathrm{0} \cdot \frac{1}{\cosh^{2}(x/H)},
\label{eq:hydstat}$$ where $\rho_\mathrm{0}$ is the mass density at the central plane of the slab, $H = c_{\mathrm{s}}^{2} / (2 \pi G {\cal
M}^{\mathrm{hs}})$ is the hydrostatic scale height, and ${\cal M}^{\mathrm{hs}} = \int_{0}^{\infty}
\rho(x) dx$ is the column density of the hydrostatic slab. $\rho_{\mathrm{0}}$ and $c_{\mathrm{s}}$ are free parameters of the solution.
For the case where the magnetic background field $B_\mathrm{\parallel}
\ne 0$, and when a monochromatic, circularly polarized Alfvén wave is injected at the central plan of the slab, @martin-et-al:97 derived a stationary, analytical solution of the above model problem in the framework of a WKB approximation. In contrast to the hydrostatic solution, the WKB solution has three free parameters: $\rho_{\mathrm{0}}$, $c_{\mathrm{s}}$, and $B_{\mathrm{\perp}}$.
Model parameters and naming conventions {#sec:params}
---------------------------------------
The full model problem as formulated in Sect. \[sec:phys\_mod\] has five free parameters. They could be specified in a dimensionless form, as we are going to discuss in Sect. \[sec:dimensionless\]. Guided by observations of molecular clouds, we choose, however, the following set of dimensional parameters: the background magnetic field $B_{\mathrm{\parallel}}$, the temperature $T$ of the slab, the initial, central mass density $\rho_{\mathrm{0}}$, the amplitude of the Alfvén wave, specified by either $B_{\mathrm{\perp} 0}$ or $v_{\mathrm{\perp} 0}=B_{\mathrm{\perp} 0}/\sqrt{\mu_{\mathrm{0}}
\rho_{\mathrm{0}}}$, and the frequency $\omega$ of the wave.
We varied parameters within limits that correspond roughly to observed parameters in molecular clouds: magnetic fields of 10-100 $\mu$G, temperatures between 5 K and 40 K, and central particle densities ranging from 250 to 2000cm$^{-3}$. For the wave frequency $\omega$ we have assumed values in the range $ 10^{4} \mbox{yr} \le 2\pi/\omega
\le 2.5 \cdot 10^{5}
\mbox{yr}$. With this choice of parameters we are in a low-beta regime, the initial, central plasma beta $\beta_{0} = 2
c_{\mathrm{s}}^{2} / c_\mathrm{A0}^{2}$ lying in a range between 0.003 and 0.7. Here, $c_{\mathrm{A0}} =
B_{\mathrm{\parallel}}/\sqrt{\mu_{\mathrm{0}}\rho_{\mathrm{0}}}$ denotes the initial, central Alfvén-speed. The corresponding Alfvén wave-length $\lambda_{\mathrm{A0}} = c_\mathrm{A0}
\cdot 2\pi/\omega$, lies in a range between about 0.07 pc and 0.35 pc. The magnetic field $B_{\mathrm{\perp} 0}$ corresponds to transverse velocities $v_{\mathrm{\perp} 0}$ in the range between $2.5
\cdot 10^{4}$cm/s and $3.2 \cdot 10^{5} $cm/s. The detailed parameters for each of the performed simulations are given in Table \[tab:mod\_params\] of the appendix. They are also reflected in the name of each simulation. For example, R20.10.25.4.40 is the simulation with $B_{\mathrm{\parallel}}=20\mu$G, $B_{\mathrm{\perp}}=10\mu$G, $2\pi/\omega= 25 \times 10^{4}$ years, a central density of $N_{\mathrm{0}} = 4 \times 250$ particles per cm$^{3}$, and $T=40$K.
Numerical solution {#sec:numerical_method}
------------------
### Numerical method
In our simulations, we consider a slab of finite (not infinite) spatial extension ${\cal D}$. We use a finite volume method on an equidistant spatial grid to solve the ideal, isothermal MHD equations, Eqs. \[eq:div\]–\[eq:mag\]. Fluxes are computed using a second order in time and third order in space stabilized Lax-Friedrichs solver[^2] as described in @barmin-et-al:96. As is shown in @barmin-et-al:96, the accuracy of this solver is comparable to that of a second order Riemann solver. Self-gravity is taken into account using a Strang-splitting [@strang:68].
### Initial conditions {#sec:initial_conditions}
At time $t=0$ we assume the slab to have a density distribution according to the analytical WKB solution of @martin-et-al:97 for a given set of parameters as specified in Table \[tab:mod\_params\]. The velocity in the x-direction, as well as the transverse components of the velocity and the magnetic field, we set to zero. The magnetic field in x-direction, $B_{\mathrm{\parallel}}$, and the temperature $T$, we set to the values given in Table \[tab:mod\_params\]. The mass column density of the initial WKB solution, within the computational domain and to infinity, we denote by ${\cal M}_{\mathrm{0}}$ and ${\cal
M}_{\mathrm{0}}^{\mathrm{\infty}}$ respectively.
### Boundary conditions {#sec:boundary_conditions}
Boundary conditions are implemented using four boundary cells at each of the two domain boundaries. Note that in this way we merely control the physical variables set in these cells. The fluxes entering and leaving the domain are determined by the interaction of the solution as set in the boundary cells and the numerical solution within the computational domain.
In the boundary cells at the inner domain boundary $(x=0)$, the physical variables are set in accordance with a left-handed, circularly polarized Alfvén-wave, whose velocity amplitude $v_\mathrm{\perp}$ we keep fixed in time, $v_{\mathrm{\perp}} = v_{\mathrm{\perp
0}}$. Allowing $v_{\mathrm{x}}$ and $\rho_{\mathrm{0}}$, and thus $B_{\mathrm{\perp}}$, to vary according to the solution in the domain, this yields: $$\begin{aligned}
\rho(x) & = & \rho(-x), \nonumber \\
v_\mathrm{x}(x) & = & - v_\mathrm{x}(-x), \nonumber \\
v_\mathrm{y}(x) & = & v_\mathrm{\perp} \cos(\omega(t - \Delta t)), \nonumber \\
v_\mathrm{z}(x) & = & v_\mathrm{\perp} \sin(\omega(t - \Delta t)), \nonumber \\
B_\mathrm{x}(x) & = & B_\mathrm{\parallel}, \nonumber \\
B_\mathrm{y}(x) & = & v_\mathrm{y}(x) \cdot \sqrt{\mu_\mathrm{0}\rho(x)}, \nonumber \\
B_\mathrm{z}(x) & = & v_\mathrm{z}(x) \cdot \sqrt{\mu_\mathrm{0}\rho(x)}. \end{aligned}$$ For the calculation of the time retardation $\Delta t =
x/c_{\mathrm{A}}$ we take into account that $c_\mathrm{A}$ varies with $x$ and $t$.
At the outer boundary, we distinguish two cases. If $v_{x}({\cal
D},t)>0$ we use a zeroth order extrapolation for all conserved variables. If $v_{x}({\cal D},t)<0$ we use a zeroth order extrapolation for $\vec{v}$ and $\vec{B}$, but restrict the density to $10^{-4}
\rho_\mathrm{0}$. Note that these outer boundary conditions allow for both accretion or loss of matter, energy, and momentum. Associated changes in the mass column density over the domain, ${\cal M}(t)$, are, however, mostly less than 2% (see Table \[tab:mod\_params\]).
We have chosen open boundaries as these match best with our intention to investigate the effect of only one source of Alfvén waves. Using periodic boundary conditions instead would implicitly introduce several sources of Alfvén waves, separated from each other by a distance $2{\cal D}$.
### Choice of domain size, discretization, and integration time {#sec:choice}
For the simulations we chose a domain of size ${\cal D}=6$ pc, or about 20 hydrostatical scale heights, covered by 5000 cells. With this choice, we fulfill the following four basic requirements. 1) The initial WKB solution fits well on the domain, 90% of its column density ${\cal M}_{\mathrm{0}}$ occupy less than 60% of the domain. 2) The numerical solution fits well on the domain, 90% of its column density ${\cal M}(t)$ are contained in the inner half of the domain and ${\cal M}(t)$ barely changes with time. 3) The hydrostatic solution is covered by sufficiently many cells. 4) The numerical solution does not depend on the discretization. Several grid studies show 5000 cells to be both necessary and sufficient (see Sect. \[sec:disc\_num\] and Fig. \[fig:spaceres\]). We followed all our simulations for $2
\cdot 10^{7}$ years, or about 20 sound crossing times of the hydrostatic scale height.
Dimensional analysis {#sec:dimensionless}
====================
Before coming to the numerical results in Sect. \[sec:num\_results\], we present a dimensional analysis of the model problem formulated in Sect. \[sec:phys\_mod\]. Dimensional parameters which completely specify the problem are the initial mass column density ${\cal
M}_{\mathrm{0}}^{\mathrm{\infty}}$, the magnetic field $B_{\mathrm{\parallel}}$, the sound speed $c_{\mathrm{s}}$, the wave frequency $\omega$ of the imposed oscillations, and their velocity amplitude $v_{\mathrm{\perp} 0}$ at the central plane. Since Alfvén wave propagation and self-gravity are essential parts of the problem, the dimensional constants $\mu_{\mathrm{0}}$ and $G$ also determine the solution. The solution in infinite space is determined by these seven dimensional input parameters. For our finite computational domain this is no longer strictly true, but we shall neglect this complication for the moment.
From the seven dimensional input parameters we can build four natural reference values for length, time, mass density, and magnetic field. The three extra dimensional parameters would then be associated with three dimensionless quantities that can be constructed from the seven parameters. In infinite space, the (suitably normalized) physical quantities in the solution would be functions of the normalized time and space variables and of these three dimensionless input parameters.
Natural dimensional reference values {#sec:naturaldimles}
------------------------------------
The four natural dimensional reference values (subscript $\mathrm{u}$ for ’unity’ in the following) should be defined such that none of them would approach zero or infinity when some parameters of the problem take expectedly large or small values. For example, we should allow $\omega$ to become very large, as it might be in the WKB limit and as it is possibly met in actual situations. Similarly, $c_{\mathrm{s}}$ may, in some clouds, become small enough to be neglected. This means that sensible natural references should be constructed from ${\cal
M}_{\mathrm{0}}^{\mathrm{\infty}}$, $B_{\mathrm{\parallel}}$, $v_{\mathrm{\perp} 0}$, $\mu_{\mathrm{0}}$ and $G$ alone. Obviously $B_{\mathrm{\parallel}}$ provides a reference magnetic field, $B_{\mathrm{u}}$, while the references of mass density, length, and time, $\rho_{\mathrm{u}}$, $d_{\mathrm{u}}$ and $t_{\mathrm{u}}$, can be defined from ${\cal M}_{\mathrm{0}}$, $v_{\mathrm{\perp} 0}$ and $G$ alone. Straightforward dimensional analysis shows that these four reference scales can be taken as: $$\begin{aligned}
B_{\mathrm{u}} & = & B_{\mathrm{\parallel}}, \nonumber \\
\rho_{\mathrm{u}} & = & \frac{ 2 \pi G ({\cal M}_{\mathrm{0}}^{\mathrm{\infty}})^2}
{v_{\mathrm{\perp} 0}^2}, \nonumber \\
d_{\mathrm{u}} & = & \frac{ v_{\mathrm{\perp} 0}^2}
{2 \pi G {\cal M}_{\mathrm{0}}^{\mathrm{\infty}}}, \nonumber \\
t_{\mathrm{u}} & = & \frac{ v_{\mathrm{\perp} 0}}
{2 \pi G {\cal M}_{\mathrm{0}}^{\mathrm{\infty}}}.
\label{eq:nondim1}\end{aligned}$$ The factors $2\pi$ were inserted for convenience, such that $\rho_{\mathrm{u}}$ and $d_{\mathrm{u}}$ are the central density and scale height of a self-gravitating isothermal sheet with sound speed $v_{\mathrm{\perp}}$ that would have, in infinite space, a column density ${\cal M}_{\mathrm{0}}^{\mathrm{\infty}}$. The reference time $t_{\mathrm{u}} = d_{\mathrm{u}}/v_{\mathrm{\perp} 0}$ is of the order of the Jeans period associated with $\rho_{\mathrm{u}}$.
From the three remaining dimensional input parameters $\mu_{\mathrm{0}}$, $c_{\mathrm{s}}$ and $\omega$ we can form, given these references, 3 natural dimensionless numbers, for example: $$\begin{aligned}
\alpha_{\mathrm{u}} & = &
\frac{ v_{\mathrm{\perp} 0}^2}
{\left( B_{\mathrm{u}}^2/(\mu_{\mathrm{0}} \rho_{\mathrm{u}})\right)}, \nonumber \\
\beta_{\mathrm{u}} & = &
\frac{ 2 c_S^2}
{ \left( B_{\mathrm{u}}^2/(\mu_{\mathrm{0}} \rho_{\mathrm{u}})\right)}, \nonumber \\
W_{\mathrm{u}} & = &
\frac{\omega }{ \sqrt{ 2 \pi G \rho_{\mathrm{u}}} }.
\label{eq:nondim2}\end{aligned}$$ The parameter $\alpha_{\mathrm{u}}$ compares the imposed velocity amplitude $v_{\mathrm{\perp}0}$ to the reference Alfvén velocity $c_{\mathrm{Au}}$, defined by $c_{\mathrm{Au}}^2 = B_{\mathrm{u}}^2
/(\mu_{\mathrm{0}} \rho_{\mathrm{u}})$. The parameter $
\beta_{\mathrm{u}}$ is the ratio of the reference gas pressure to the magnetic pressure of $B_{\mathrm{u}}$. The parameter $W_{\mathrm{u}}$ is a WKB parameter. Clearly, neither the set of the four reference dimensional values nor the set of the three input dimensionless numbers is uniquely defined. Other dimensional reference values could be obtained by multiplying the chosen ones by any function of the dimensionless input numbers $\alpha_{\mathrm{u}}$, $\beta_{\mathrm{u}}$, and $W_{\mathrm{u}}$. Similarly, the above set of three such numbers could be replaced by three other arbitrary functions of them.
Dimensional reference values from WKB solution {#sec:wkbdimles}
----------------------------------------------
The reference quantities defined in Eq. \[eq:nondim1\] need not be very close to the actual values of the physical quantities in the solution, even near the central plane. It may be felt desirable, however, that the reference values be as close as possible to actual ones, even at the price of defining the reference values in a more sophisticated way than $\rho_{\mathrm{u}}$, $d_{\mathrm{u}}$ and $t_{\mathrm{u}}$. If the WKB solution is to be a guide, such estimates could be obtained from the solution of Martin et al. (1997). From this solution, it is possible to relate the central density $\rho_{\mathrm{0}}$ of a solution to its column density ${\cal
M}_{\mathrm{0}}$ in the finite computational domain of thickness ${\cal D}$ and to the total column density to infinity, ${\cal
M}_{0}^{\mathrm{\infty}}$. The limited extent of the computational domain is reflected in the fact that the ratio $({\cal
M}_{\mathrm{0}}/{\cal M}_{\mathrm{0}}^{\mathrm{\infty}})$ is less than unity. A scale length $d^{\mathrm{wkb}}_{\mathrm{0}}$ may then be defined as the distance from the central plane of the slab where 90% of the column density ${\cal M}_{\mathrm{0}}$ are reached. The time scale $t_{\mathrm{0}}$ can be defined as $t_{\mathrm{0}}= (2 \pi G
\rho_{\mathrm{0}})^{-1/2}$. The reference value of the magnetic field may be chosen as before, $B_{\mathrm{0}} = B_{\mathrm{\parallel}}$.
In analogy with Eq. \[eq:nondim2\], dimensionless numbers $\alpha_{\mathrm{0}}$, $\beta_{\mathrm{0}}$, and $W_{\mathrm{0}}$ can be introduced. Note that $\beta_{\mathrm{0}}$ is, in fact, identical to $\beta_{\mathrm{0}}$ as given in Sect. \[sec:params\]. An Alfvén-velocity can be defined by $c_{\mathrm{A0}}^{2} =
B_{\mathrm{\parallel}}^2 / (\mu_{\mathrm{0}}\rho_{\mathrm{0}})$, and Alfvén wave-length by $\lambda_{\mathrm{A0}} = 2 \pi
c_{\mathrm{A0}}/\omega$.
Aliases to the WKB parameter $W_{\mathrm{0}}$ could be used as well. In particular, one could prefer the ratio of the WKB length scale to the Alfvén wavelength associated with $B{\mathrm{\parallel}}$ and $\rho_{\mathrm{0}}$, $W_{\mathrm{\lambda
A0}} = \lambda_{\mathrm{A0}}/d^{\mathrm{wkb}}_{\mathrm{0}}$. For our simulations, the values for a number of these dimensionless parameters are listed in Table \[tab:mod\_params\].
Reference values for substructure scale {#sec:refsub}
---------------------------------------
Unlike the WKB solution, the numerical solution of the model problem from Sect. \[sec:phys\_mod\] is far from smooth (see e.g. Fig. \[fig:r20.20.25.4\]). And while the lengths $d_{\mathrm{u}}$ or $d^{\mathrm{wkb}}_{\mathrm{0}}$ are natural estimates for the global size of the mass distribution, the natural scale length associated with substructures induced by the wave is, of course, quite different. It is either of the order of the reference Alfvén wavelength $\lambda_{\mathrm{Au}} = 2 \pi
c_{\mathrm{Au}}/\omega$ or of the Alfvén wavelength associated with the WKB reference quantities, $\lambda_{\mathrm{A0}} = 2 \pi
c_{\mathrm{A0}}/\omega$. These two scales are of the same order of magnitude for a wave-supported cloud. The actual characteristic scale of cloud substructure is the product of any one of them with a function of the dimensionless input parameters. On physical grounds, we expect, however, this function to be of the order of unity, since no other small length scale is likely to play an important role. In the low-$\beta$ regime we consider, the sonic wavelength $2 \pi c_S
/\omega$ is usually much shorter and sound perturbations nonlinearly evolve into shocks anyway, so that this sonic wavelength is not expected to show up in the spectrum of the fluctuations, except perhaps at the level of the very small thickness of the dense sheets that form.
In the next section, we present the numerical solutions for the model problem from Sect. \[sec:phys\_mod\] for various parameter sets. Note that our study, inspired by molecular clouds, covers only a small part of the entire parameter space. For this part of the parameter space, we identify and discuss some of the dependences indicated by the dimensional analysis.
Numerical results {#sec:num_results}
=================
The injection of magnetic waves at the central plane of the self-gravitating slab has two major consequences: the slab becomes supersonically turbulent and its spatial extension is clearly larger than in hydrostatic equilibrium.
Neither of these consequences is surprising. The energy provided by the injected wave must result in additional support of the slab against its self-gravity. Density inhomogeneities are to be expected since within the frame of linear analysis a parametric instability of the injected Alfvén wave exists [@1978ApJ...224.1013D; @1978ApJ...219..700G; @turkmani]. Early on in our simulations we observe the associated growth of high density sheets, which is accompanied by the destruction of the transverse magnetic field (see Fig. \[fig:param\]). Note also that @malara-velli:96 and @malara-et-al:00 demonstrated that even a non-monochromatic spectrum of Alfvén waves is subject to parametric instability with linear and nonlinear growth rates of the same order of magnitude as in the monochromatic case.
So far the turbulent, nonlinear evolution of such a system is, however, not well investigated. The results we present in the following demonstrate that the same parameters which govern the parametric instability, $\lambda_\mathrm{A0}$ and $\beta_\mathrm{0}$, are also of crucial importance for the nonlinear, turbulent solution. This despite the fact that the true, time averaged values of these quantities at $x=0$ (and close by) deviate substantially from $\lambda_\mathrm{A0}$ and $\beta_\mathrm{0}$.
![Early on in all our simulations, high density sheets develop under the influence of the parametric instability of the Alfvén-wave, while the transverse magnetic field is partly destroyed. The figure is a snapshot of simulation R20.20.5.4.10 after $2.1 \cdot 10^{5}$ years. Shown are the density (solid line) and the energy density of the transverse magnetic field (dashed line), normalized to their initial, central values, as functions of distance to the central plane (x-axis, in units of $\langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}}$, the time averaged slab extension). The dotted curves denote $N/N_{\mathrm{0}}$ (lower curve) and $E^{\mathrm{\perp}}_{\mathrm{mag}} /
E^{\mathrm{\perp}}_{\mathrm{mag0}}$ (upper curve) of the initial WKB solution.[]{data-label="fig:param"}](f1.ps){width="8.5cm"}
{width="8.5cm" height="6.9cm"} {width="8.1cm" height="6.9cm"}
Spatial extension of turbulent slab {#sec:extension}
-----------------------------------
We find that the time averaged spatial extension of the solution roughly agrees with the spatial extension of the corresponding WKB solution. As a function of time, the spatial extension $d^{\mathrm{sol}}(t)$ can oscillate but does not have to. $d^{\mathrm{sol}}(t)$ we define in analogy with $d_{\mathrm{0}}^{\mathrm{wkb}}$ (see Sect. \[sec:dimensionless\]) as the distance where the column density reaches 90% of ${\cal M}(t)$. We similarly define the extension $d^{\mathrm{hs}}(t)$ and $d^{\mathrm{wkb}}(t)$ of the corresponding hydrostatical and WKB solutions, where corresponding means that at time $t$ the three solutions have the same ${\cal M}(t)$. Finally, we denote by $\left <
. \right >_{\mathrm{t}}$ the time average between $5\cdot 10^{6}$ and $2\cdot 10^{7}$ years.
### Dependence on system parameters {#sec:extav}
In all our simulations, the time averaged extension $\langle
d^{\mathrm{sol}}\rangle _\mathrm{t}$ of the slab agrees to within a factor of three with the spatial extension of the WKB solution $d^{\mathrm{wkb}}_{\mathrm{0}}$. The dominant parameters governing the extension of the slab are, therefore, the parameters governing the WKB solution, i.e. $\rho_{0}$, $B_{\mathrm{\perp}}$, and $T$.
Deviations from the spatial extension of the WKB solution we find to depend linearly on the dimensionless WKB parameter $W_{\mathrm{\lambda A0}} = \lambda_\mathrm{A0} /
d^{\mathrm{wkb}}_{\mathrm{0}}$. Fig. \[fig:lamdep\]a shows the ratio of the spatial extensions $\langle d^{\mathrm{sol}}\rangle
_\mathrm{t} / d^{\mathrm{wkb}}_{\mathrm{0}}$ as a function of $\lambda_\mathrm{A0} / d^{\mathrm{wkb}}_{\mathrm{0}}$ for the different runs. A linear least square fit, also shown in the figure, gives a dependence $\langle d^{\mathrm{sol}} \rangle _\mathrm{t} =
\alpha \lambda_\mathrm{A0} + \beta d^{\mathrm{wkb}}_{\mathrm{0}}$ with $\alpha = 4.06$ and $\beta = 0.36$. Linear fits of similar quality are obtained if instead of $d^{\mathrm{wkb}}_{\mathrm{0}}$ one uses as scaling parameter $\langle d^{\mathrm{hs}}\rangle _\mathrm{t}$, or $\langle d^{\mathrm{wkb}}\rangle _\mathrm{t}$. For the last case, the fitting parameters are $\alpha = 3.86$ and $\beta = 0.37$. If instead of $\lambda_\mathrm{A0}$ we consider the time average of the true Alfvén wave-length $ \langle \lambda_\mathrm{A} \rangle
_\mathrm{t} $ at or close to the slab center, we cannot identify any such clear dependence. $ \langle \lambda_\mathrm{A} \rangle
_\mathrm{t} $ is usually larger than $\lambda_\mathrm{A0}$ by a factor of about 2 to 8.
From Fig. \[fig:lamdep\]a it can be seen that best agreement between $\langle d^{\mathrm{sol}}\rangle _\mathrm{t}$ and $d^{\mathrm{wkb}}_{\mathrm{0}}$ is obtained around $\lambda_\mathrm{A0}/ d^{\mathrm{wkb}}_{\mathrm{0}} = 0.17$. The deviations at larger values of $\lambda_\mathrm{A0} /
d^{\mathrm{wkb}}_{\mathrm{0}}$ are not too surprising. A basic assumption for the validity of WKB theory is that the density changes only on scales much larger than the Alfvén wave-length. This assumption fails to hold as $\lambda_\mathrm{A0} /
d^{\mathrm{wkb}}_{\mathrm{0}}$ increases. For the deviations at small values of $\lambda_\mathrm{A0} / d^{\mathrm{wkb}}_{\mathrm{0}}$ we have checked that they are not caused by a too coarse spatial discretization, which would cause artificial wave-damping and thus reduced support against self-gravity. Increasing the spatial resolution by a factor of four left the nonlinear solution unchanged. Instead, the true reason again lies in the failure of the WKB approach to be valid. The mass distribution in the system still consists of thin, dense sheets separated by more diffuse medium. As the sheet thickness is much less than the Alfvén wavelength in the tenuous medium, the conditions for a WKB description are not met. We come back to this point in more detail in Sect. \[sec:analext\].
Besides the dominant effect of $\lambda_\mathrm{A0}/
d^{\mathrm{wkb}}_{\mathrm{0}}$, Fig. \[fig:lamdep\]a suggests $\beta_{\mathrm{0}}$ to have a second order effect. Simulations with identical ratio $\lambda_{\mathrm{A0}}
/d^{\mathrm{wkb}}_{\mathrm{0}}$ usually have a larger extension for smaller $\beta_{\mathrm{0}}$.
In view of what was said in Sect. \[sec:dimensionless\], the average slab thickness must equal the reference scale $d_{\mathrm{0}}^{\mathrm{wkb}}$ multiplied by a function of $\alpha_{\mathrm{0}}$, $\beta_{\mathrm{0}}$ and $W_{\mathrm{\lambda A
0}}$. In the limit of very small pressure the dependence on $\beta_{\mathrm{0}}$ disappears. In the limit of very large $W_{\mathrm{\lambda A 0}}$, that is very small $\lambda_{\mathrm{A0}}
/ d^{\mathrm{wkb}}_{\mathrm{0}}$, the dependence on $W_{\mathrm{\lambda A 0}}$ also disappears. There may remain some dependence on $\alpha_{\mathrm{0}}$, which Fig. \[fig:lamdep\]a indicates to be weak. The ratio $\langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}} / d^{\mathrm{wkb}}_{\mathrm{0}}$ then approaches a value which is apparently close to $(1/2)$. For higher values of $\lambda_{\mathrm{A0}} / d^{\mathrm{wkb}}_{\mathrm{0}}$, we expect some dependence of $\langle d^{\mathrm{sol}} \rangle_{\mathrm{t}} /
d^{\mathrm{wkb}}_{\mathrm{0}}$ on this parameter, which may be represented by a Taylor expansion to first order for not too large $\lambda_{\mathrm{A0}} / d^{\mathrm{wkb}}_{\mathrm{0}}$. This is in rough agreement with what is seen in Fig. \[fig:lamdep\]a, and is consistent with the idea that the larger the Alfvén wave-length, the larger the spacing between dense sheets, and the thicker the system.
### Time variability of slab extension {#sec:timevaria}
As a function of time, the spatial extension $d^{\mathrm{sol}}(t)$ of the slab can be very variable but does not have to be. We mention already here that the substructure of the slab, voids and high density sheets, always shows oscillatory motions (see Sect. \[sec:densstru\] and Fig. \[fig:r20.20.25.4\]). In Fig. \[fig:lamdep\]b, the ratio of $d^{\mathrm{sol}}(t) / d^{\mathrm{wkb}}_{\mathrm{0}}$ is shown as a function of time for five runs, representative of the total of our simulations. As can be seen, only some of the runs show a strong variability while others have a more or less constant extension. A measure for the strength of this time variability is the standard deviation $\sigma_{\mathrm{d}}$ of $d^{\mathrm{sol}}(t) /
d^{\mathrm{wkb}}_{\mathrm{0}}$. Using for the calculation of $\sigma_{\mathrm{d}}$ the same time interval as for the time averages, we find the relative standard deviation $\sigma^{\mathrm{rel}}_{\mathrm{d}} =
\sigma_{\mathrm{d}}/(\langle d^{\mathrm{sol}} \rangle_{\mathrm{t}} /
d^{\mathrm{wkb}}_{\mathrm{0}})$ to lie in an interval of $ 0.1
\;\rlap{\lower 2.5pt \hbox{$\sim$}}\raise 1.5pt\hbox{$<$}\;
\sigma^{\mathrm{rel}}_{\mathrm{d}}
\;\rlap{\lower 2.5pt \hbox{$\sim$}}\raise 1.5pt\hbox{$<$}\;
0.9$ (see Table \[tab:mod\_params\]).
In about half of our simulations, the variability takes the form of a roughly periodic oscillation. The period of the oscillation is, however, not always well defined. Twice the sound crossing time of the average slab thickness, $T_{\mathrm{s}} = 2 \langle
d^{\mathrm{sol}} \rangle_{\mathrm{t}} / c_{\mathrm{s}}$, is often close to the observed period $T_{\mathrm{obs}}$. The ratio $T_{\mathrm{obs}}/T_{\mathrm{s}}$ averages over the different runs to a value close to 0.95. However, this ratio, normalized to its average value, has significant scatter of the order of 0.8. Moreover, the fact that our slabs are more wave-supported than gas pressure supported makes it difficult to understand a scaling like $T_{\mathrm{obs}} \sim
T_{\mathrm{s}}$ on physical grounds. A more natural scaling, as we will show in Sect. \[sec:oscanal\], would be that the oscillation period is proportional to twice the crossing time of the average slab thickness at velocity $v_{\mathrm{\perp 0}}$. For $T_{\mathrm{v\perp 0}}=2
\langle d^{\mathrm{sol}} \rangle_{\mathrm{t}} / v_{\mathrm{\perp0}}$ the ratio $T_{\mathrm{obs}} \sim T_{\mathrm{v\perp 0}}$ averages, over our different runs, to 3.4 with a scatter of 0.4. For $T_{\mathrm{v\perp 0}}=2 d^{\mathrm{wkb}}_{\mathrm{0}} /
v_{\mathrm{\perp0}}$ the average is 3.6 with a scatter of only 0.2. Thus, to within a factor of order unity, but somewhat larger than unity, our simulations are also consistent with $T_{\mathrm{obs}}$ being proportional to $T_{\mathrm{v\perp 0}}$.
Inhomogeneous structure of turbulent slab {#sec:densstru}
-----------------------------------------
Under the influence of both the injected wave and self-gravity, the interior structure of the slab becomes very inhomogeneous, substructure develops. This substructure is the result of time-dependent, nonlinear effects and as such is beyond the reach of WKB theory. Again, we find $\lambda_{\mathrm{A0}}$ to play a critical role, but now the temperature of the slab has an effect as well.
The density distribution is characterized by extended, low density voids and narrow, high density sheets. In these low density voids, the other variables, $\vec{v}$ and $\vec{B}$, remain nearly constant. Therefore, the density substructure is a good mirror for their substructure as well, and we restrict ourselves to density in the following. To characterize the density substructure in our gravitationally stratified slab, we represent the time series of 1D density distributions as a 2D grey-scale plot. Low density regions then appear as more or less extended (dark) patches, while high density sheets take the form of thin (bright) lines.
Fig. \[fig:r20.20.25.4\] shows such a representation of the density for two runs. Apparent also in this representation is the presence (run R100.10.5.4.10) or absence (run R20.20.25.4.10) of a global oscillation of the slab. However, it can also be seen that, independent of the existence of a global oscillation, the high density sheets within the slab undergo oscillatory motions. The motion of a single sheet is roughly parabolic, but may be interrupted at any time as two sheets collide. Different sheets can have widely different oscillation periods, which are mostly much smaller than the period of the global oscillation of the slab.
This motion of the high density sheets, together with the number of sheets, determines the size of the voids (dark patches in Fig. \[fig:r20.20.25.4\]) in the 2D representation of the data. We call this the scale of the substructure, a larger scale substructure thus referring to an overall large size of the voids. Based on the only qualitative measure of the grey-scale plots, we find that this scale of the substructure is, in essence, determined by $\lambda_{\mathrm{A0}}$ and temperature $T$. The size of the substructure in the two runs shown in Fig. \[fig:r20.20.25.4\], which have the same $\lambda_{\mathrm{A0}}$ and temperature $T$, but otherwise widely different parameters, is very similar. This becomes even more apparent if Fig. \[fig:r20.20.25.4\] is compared with Fig. \[fig:d3dtl\], where simulations with only half the Alfvén wave-length (Fig. \[fig:d3dtl\]a) and four times the slab temperature $T$ (Fig. \[fig:d3dtl\]b) are shown. On the other hand, the scale of the density distribution for the two simulations shown in Fig. \[fig:d3dtl\] is again similar. This suggest that augmenting the temperature by a factor of four has the same effect as reducing the Alfvén wave-length $\lambda_{\mathrm{A0}}$ by a factor of two. Or, put otherwise, that the scale of the density substructure in space and time depends on the ratio $ \sqrt{T} / \lambda_{\mathrm{A0}}$ or, equivalently, on $\sqrt{\beta_{\mathrm{0}}}\omega$. We currently have, however, no quantitative measure to further corroborate this speculation.
The dominant role of $\lambda_{\mathrm{A0}}$ for the scale of the density distribution may have been anticipated from what was said towards the end of Sect. \[sec:dimensionless\] as well as from parametric instability theory [@1978ApJ...224.1013D; @1978ApJ...219..700G]. Analytical results for the linear regime indeed show the separation of density disturbances to increase with increasing Alfvén wave-length. On the other hand, the system we consider is highly nonlinear and $\lambda_{\mathrm{A0}}$ soon loses its meaning even on average.
![The energy $E^{\mathrm{\perp}}_{\mathrm{mag}}$ of the transverse magnetic field is very inhomogeneous. It is generally smaller than in the corresponding WKB solution and not correlated with the density substructure. Shown are the density (solid line) and the energy density of the transverse magnetic field (dashed line), normalized to their initial, central values, as functions of distance to the central plane (x-axis, in units $\langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}}$) at the example of simulation R20.20.5.4.10 at a time of $2.07 \cdot 10^{7}$ years. The dotted curves denote $N/N_{\mathrm{0}}$ (lower curve) and $E^{\mathrm{\perp}}_{\mathrm{mag}} /
E^{\mathrm{\perp}}_{\mathrm{mag0}}$ (upper curve) of the corresponding WKB solution.[]{data-label="fig:dens_emag"}](f5.ps){width="8.5cm"}
{width="8.1cm" height="6.3cm"} {width="8.1cm" height="6.3cm"}
Energy of turbulent slab
------------------------
At the central plane, we constantly feed energy into the slab. The energy provided by this ’one point forcing’ penetrates far into the slab, despite the gradual destruction through isothermal shocks and viscous dissipation. Note that the viscous dissipation is merely given by the numerical method and the discretization, and may not accurately mimic real viscosity, diffusion, or resistivity. The energy density of the transverse magnetic field is far from smooth and is generally smaller than its initial WKB value. A typical situation at later times is shown in Fig. \[fig:dens\_emag\]. The density substructure leads to wave reflection and shock formation. Occasionally, the WKB value of $E^{\mathrm{\perp}}_{\mathrm{mag}}$ can be exceeded, as two high density sheets move against each other, thus compressing the transverse magnetic field between them.
### Loss of Poynting-flux
Unlike in the WKB solution, the Poynting-flux in our numerical solution is neither constant in space nor time. Let us denote by $P_{\mathrm{x0}}$ the Poynting-flux of the initial WKB solution and by $\langle P_{\mathrm{x}} \rangle
_\mathrm{t}$ the time averaged x-component of the true Poynting-flux. We find that the change $\Delta \langle P^{\mathrm{sol}}_{\mathrm{x}} \rangle _\mathrm{t}$ of the Poynting-flux over a distance $\langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}}$ from the central plan of the slab lies in a range $ 0.03 \;\rlap{\lower 2.5pt \hbox{$\sim$}}\raise
1.5pt\hbox{$<$}\; \Delta \langle P^{\mathrm{sol}}_{\mathrm{x}} \rangle
_\mathrm{t}/P_{\mathrm{x0}}
\;\rlap{\lower 2.5pt \hbox{$\sim$}}\raise 1.5pt\hbox{$<$}\; 0.4$. The ratio depends about linearly on $\sqrt{\beta_{\mathrm{0}}}$, as can be seen in Fig. \[fig:dpfx\]a. In view of the results in Sect. \[sec:densstru\], this dependence suggests a connection between the loss of Poynting-flux and the amount of density structure we have in the slab. There we have seen that larger values of $\sqrt{T}/\lambda_{\mathrm{A0}}$, which implies larger values of $\sqrt{\beta_{\mathrm{0}}}$, correlate with a finer scale density structure.
Largest losses of Poynting-flux occur in the innermost region of the slab, and more than 90% of the total loss of Poynting-flux over the entire computational domain occurs within a distance $\langle
d^{\mathrm{sol}} \rangle_{\mathrm{t}}$ of the central plane of the slab. Fig. \[fig:dpfx\]b shows the spatial variation of the loss of Poynting-flux. There, $\langle P_{\mathrm{x}} \rangle _\mathrm{t}(x) -
\langle P_{\mathrm{x}} \rangle _\mathrm{t}(x={\cal D})$ is shown, in units of $P_{\mathrm{x0}}$, as a function of increasing, relative column density ${\cal M}_{\mathrm{0}}(x)/{\cal M}_{\mathrm{0}}$ for five runs. Except for the value of $\omega$, the runs have identical parameters with a value of $\beta_{\mathrm{0}}=0.087$. As can be seen, the decrease of Poynting-flux in this representation is fairly similar for all five simulations. Note, however, that the actual spatial extension of the five runs is very different ($\langle
d^{\mathrm{sol}} \rangle_{\mathrm{t}} = 0.11$${\cal D}$ for R20.20.5.4.10 but $\langle d^{\mathrm{sol}} \rangle_{\mathrm{t}} =
0.34$${\cal D}$ for R20.20.25.4.10).
The change of Poynting-flux is part of the change of the total energy flux. In fact, we find $\Delta
\langle P^{sol}_{\mathrm{x}} \rangle _\mathrm{t}$ to amount to between 60% and 90% of the change of the total energy flux between the central plane of the slab and $\langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}}$. The remaining amount is made up from the fluxes of gravitational and kinetic energy, the latter usually contributing considerably more. As the slab is, in time average, in a quasi-stationary state, this change in energy flux between the central plane and $\langle d^{\mathrm{sol}}
\rangle _\mathrm{t}$ must be radiated.
### Energy equipartition
In all our simulations, the ratio of the time averaged energy density of the transverse magnetic field and of the energy density associated with the transverse velocity, $\langle
E^{\mathrm{\perp}}_{\mathrm{mag}}
\rangle_{\mathrm{t}} / \langle E^{\mathrm{\perp}}_{\mathrm{kin}}
\rangle_{\mathrm{t}}$, shows some spatial variation. Within about the innermost 10% of $\langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}}$, we find for $\langle
E^{\mathrm{\perp}}_{\mathrm{mag}} \rangle_{\mathrm{t}} /
\langle E^{\mathrm{\perp}}_{\mathrm{kin}} \rangle_{\mathrm{t}}$ typical values of about 2, for a few simulations even somewhat higher. Further out, to distances around $\langle
d^{\mathrm{sol}}\rangle_{\mathrm{t}}$, we observe approximate equipartition between $\langle
E^{\mathrm{\perp}}_{\mathrm{mag}}
\rangle_{\mathrm{t}}$ and $\langle E^{\mathrm{\perp}}_{\mathrm{kin}}
\rangle_{\mathrm{t}}$. The time averaged kinetic and magnetic energy densities themselves decrease with increasing distance from the central plane of the slab.
If we look at the time and space averaged energy quantities, $\langle E^{\mathrm{\perp}}_{\mathrm{mag}}
\rangle_{\mathrm{t,x}}$ and $\langle E^{\mathrm{\perp}}_{\mathrm{kin}}
\rangle_{\mathrm{t,x}}$, we find approximate equipartition in most of our simulations. Here, $\langle \cdot \rangle_{\mathrm{t,x}}$ denotes time average and subsequent spatial average, the later taken over $x \in [0,\langle d^{\mathrm{sol}} \rangle_{\mathrm{t}}]$. $\langle E^{\mathrm{\perp}}_{\mathrm{mag}} \rangle_{\mathrm{t,x}} /
\langle E^{\mathrm{\perp}}_{\mathrm{kin}} \rangle_{\mathrm{t,x}}$ lies in a range between 0.8 and 1.2 except for five runs (R100.45.5.8.10, R20.20.5.1.10, R20.10.2.1.10, R20.20.5.1.20 with a ratio between 0.6 and 0.8 and R100.45.1.4.10 with a ratio of 0.4). For the time and space average of the total kinetic energy density $\langle
E_{\mathrm{kin}}\rangle_{\mathrm{t,x}}$, we find the ratio $\langle
E^{\mathrm{\perp}}_{\mathrm{kin}}\rangle_{\mathrm{t,x}} / \langle
E_{\mathrm{kin}} \rangle_{\mathrm{t,x}}$ to lie in a range between 0.7 and 0.9, except for run R100.20.5.4.10 where this ratio is only 0.6.
Discussion {#sec:discussion}
==========
Spatial extension of a non-homogeneous slab {#sec:analext}
-------------------------------------------
In Sect. \[sec:num\_results\] we have seen that the WKB extension provides only a first approximation to the spatial extension of the fragmented slab. In the following, we give a qualitative explanation of why the extension of a fragmented slab differs from its WKB extension, in particular also for small values of $\lambda_{\mathrm{A0}} / d^{\mathrm{wkb}}_{\mathrm{0}}$, the seeming WKB limit. The WKB picture considers a diffuse mass distribution. The thickness of the layer of material between column density 0 (the central plane) and column density ${\cal M}$ is $$H({\cal M}) = \int_{0}^{\cal M} \frac{d{\cal M}'}{\rho({\cal M}')}
\label{eq:hm}$$ where $\rho({\cal M})$ is the density of the matter sitting on top of an amount of material of total column density ${\cal M}$. If a slab is fragmented and contains dense sheets separated by tenuous regions, the contribution of the sheets to the integral in Eq. \[eq:hm\] is very small. Thus the system’s thickness is essentially due to the tenuous intersheet regions. This indicates that an entirely diffuse slab would be thicker than one which has a substantial fraction of its mass in the form of dense sheets, provided that the tenuous regions of the diffuse slab have drastically smaller density than the intersheet regions of the fragmented slab.
In a limit where the WKB approximation is close to being valid, however, $\rho({\cal M})$ should be quite similar in both cases. To see this, let us denote by $B_{\perp}({\cal M})$ the magnetic amplitude of the wave at a point sitting on top of an amount of material of total column density ${\cal M}$. Suppose the slab contains a large number of low-column-density thin sheets levitating in equilibrium under the pressure of the Alfvénic flux. Assuming equilibrium yields a relation between the jump $dB_{\perp}$ of $B_{\perp}$ at the crossing of one such sheet and the corresponding jump in column density $d{\cal M}$: $$d \left(\frac{B_{\mathrm{\perp}}^2}{4 \mu_{\mathrm{0}}} \right )=
- 4 \pi G {\cal M} \mbox{\,\,} d{\cal M}.$$ The same relation also holds for any infinitesimal slab of tenuous material in equilibrium. Thus, irrespective of whether the matter has an entirely diffuse distribution or is partially fragmented in a large number of thin sheets, the same relation between $B_{\perp}^2({\cal M})$ and ${\cal M}$ holds. This means that the distribution of wave support as a function of column density is independent of the details of the density distribution. If the intersheet medium can also be described by a WKB approach, it would result that the density distribution $\rho({\cal M})$ would be the same in the entirely tenuous slab as it is in the tenuous regions of the fragmented slab. A glance at Eq. \[eq:hm\] then shows that the fragmented slab would be less extended, by a factor which depends on which fraction of its mass is in the form of dense sheets.
Oscillation of slab extension {#sec:oscanal}
-----------------------------
In Sect. \[sec:timevaria\] we have seen that the spatial extension of the slab can oscillate on a time scale of a few Myr. That a period on the order of $T_{\mathrm{v\perp0}} = 2 \langle d^{\mathrm{sol}}
\rangle_{\mathrm{t}}/ v_{\mathrm{\perp0}}$ is to be expected on physical grounds shows the following analysis.
We discuss the fundamental period of deviations from equilibrium of an isothermal slab, supported against self-gravity by gas pressure and a flux of WKB Alfvén waves. Such a model is, admittedly, far from our highly structured slabs, but it should be sufficient to understand the main parameters which control the global oscillation. In this model, the force exerted by Alfvén waves takes the form of the gradient of the wave energy density $U$. The wave energy flux $ U c_{\mathrm{A\parallel}}$ we assume to be constant in space at any time, with $c_{\mathrm{A\parallel}}$ the Alfvén velocity associated with $B_{\mathrm{\parallel}}$ and the local mass density. The equilibrium is described by functions $\rho_{\mathrm{e}}(x)$, $U_{\mathrm{e}}(x)$ and ${\cal M}_{\mathrm{e}}(x)$, the latter being the column density between 0 and $x$. An additional subscript 0 denotes the values of these functions at $ x = 0$. In particular, $U_{\mathrm{e0}} =
\rho_{\mathrm{e0}} v_{\mathrm{e0\perp}}^2 /2$. As a boundary condition for the perturbation, we impose, in analogy with our simulations, that the velocity amplitude of the wave injected at $x=0$ is fixed in time. This has the important consequence that the wave flux forced into the slab at the central plane is modulated by the varying gas density there. Indeed, for fixed $B_{\mathrm{\parallel}}$ and $v_{\mathrm{\perp}}$ the sum of the kinetic and Poynting Alfvén wave energy flux, $P = \rho v_{\mathrm{\perp}}^2
c_{\mathrm{A\parallel}}/2$, is proportional to $\sqrt{\rho}$.
In about half of our simulations, we indeed observe such a modulation of the Poynting-flux and density close to the central plane of the slab, after applying a running mean in time to filter out the strong short term variability stemming from individual high density sheets (see Fig. \[fig:rhopfxdsol\]). The modulation of the Poynting-flux then is in phase with the density perturbation and in antiphase with the slab thickness, as it should be.
Linearizing the equations of motion about the equilibrium and assuming a monochromatic perturbation, ${\cal M}_{\mathrm{1}}(x) e^{-i
\omega t}$ of the column density, we find that ${\cal M}_{\mathrm{1}}(x)$ is a solution of the following linear homogeneous equation: $$\begin{aligned}
& & \rho_{\mathrm{e}} c_{\mathrm{s}}^2
\frac{d}{dx}
\left( \frac{1}{\rho_{\mathrm{e}}} \frac{d {\cal M}_{\mathrm{1}}}{dx} \right) +
\frac{U_{\mathrm{e}}}{\rho_{\mathrm{e}}} \rho_{\mathrm{e}}^{3/2}
\frac{d}{dx} \left( \frac{1}{\rho_{\mathrm{e}}^{3/2}}
\frac{d {\cal M}_{\mathrm{1}}}{dx} \right) + \nonumber \\
& & (\omega^2 + 4 \pi G \rho_{\mathrm{e}})
{\cal M}_{\mathrm{1}} +
\frac{{\cal M}'_{\mathrm{1}}(0)}{ 2 \rho_{\mathrm{e}}}
\frac{dU_{\mathrm{e}}}{dx} = 0.
\label{eq:modeq}\end{aligned}$$ The last term in Eq. \[eq:modeq\] represents the modulation in time of the Poynting-flux entering the slab at $x = 0$. It is proportional to the mass density perturbation at this point, ${\cal
M}'_{\mathrm{1}}(0) = d{\cal M} /dx = \rho_{\mathrm{1}}(0)$. Had we assumed that the slab suffers perturbations under a constant wave energy flux, this term would have been absent from Eq. \[eq:modeq\]. The spectrum associated with Eq. \[eq:modeq\] may only be found numerically, but its gross features may be anticipated from a simple dimensional analysis, substituting $(ik{\cal M}_{\mathrm{1}})$ for $(d{\cal M}_{\mathrm{1}} /dx)$ and $(- U_{\mathrm{0}}/H_{\mathrm{e}})$ for $(dU_{\mathrm{0}}/dx)$. $H_{\mathrm{e}}$ is the thickness of the equilibrium slab, approximately given by the relation $$4 \pi G
\rho_{\mathrm{e0}}^2 H_{\mathrm{e}}^2 = (U_{\mathrm{e0}} +
\rho_{\mathrm{e0}} c_{\mathrm{s}}^2).
\label{eq:eqrel}$$
![The global oscillation of the slab, if existent, is in antiphase with the modulation of the running mean in time (time window 1.5 Myr) of the Poynting-flux and the density at the central plane of the slab. Shown are, as a function of time in $10^{6}$ years, the normalized spatial extension of the slab (dashed line), and the normalized running mean in time of the Poynting-flux (dotted line) and of the density (solid line). Normalized means that the quantities were divided by the maximum value they assume in the time interval shown.[]{data-label="fig:rhopfxdsol"}](f7.ps){width="8.5cm"}
This rough procedure does not make justice of the specific profile of $\rho_{\mathrm{e}}(x)$. To take care of this, we introduce numerical coefficients, $\alpha_{\mathrm{1}}$, $\alpha_{\mathrm{2}}$, $\alpha_{\mathrm{3}}$, $\alpha_{\mathrm{4}}$, all positive and of order unity. From Eq. \[eq:modeq\], the dispersion relation is expected to be of the form: $$\omega^2 =
\alpha_{\mathrm{1}} \ c_{\mathrm{s}}^2 k^2 +
\alpha_{\mathrm{2}} \ \frac{v_{\mathrm{\perp} e0}^2}{2} k^2 -
\alpha_{\mathrm{3}} \ 4 \pi G \rho_{\mathrm{e0}} +
\alpha_{\mathrm{4}} \frac{i k v_{\mathrm{\perp} e0}^2}{2 H_{\mathrm{e}}}.
\label{eq:disprel}$$ The last, imaginary, term of Eq. \[eq:disprel\] represents the last term of Eq. \[eq:modeq\], associated with the modulation of the wave energy flux. For the fundamental mode, $k \approx
H_{\mathrm{e}}^{-1}$. Its frequency should then be given by an expression of the form: $$\omega^2 =
4 \pi G \rho_{\mathrm{e0}}
\left( \alpha_{\mathrm{1}} \
\frac{c_{\mathrm{s}}^2}{c^2} +
\alpha_{\mathrm{2}} \
\frac{ v_{\mathrm{\perp}e0}^2/2 }{ c^2} -
\alpha_{\mathrm{3}} + i \alpha_{\mathrm{4}} \
\frac{ v_{\mathrm{\perp}e0}^2/2 }{ c^2}
\right),
\label{eq:fundmond}$$ where $c^{2}=c_{\mathrm{s}}^2 + v_{\mathrm{\perp}e0}^2/2$. This shows that the fundamental time must be, to within a factor of order unity but presumably smaller due to the presence of the negative term $- \alpha_{\mathrm{3}}$, the Jeans frequency $(4 \pi G
\rho_{\mathrm{e0}})^{1/2}$. From Eq. \[eq:eqrel\], the associated time is, for $c_{\mathrm{s}} << v_{\mathrm{e0\perp}}$, of order of $\sqrt{2}H_{\mathrm{e}}/v_{\mathrm{e0\perp}}$. Given the uncertainty left by this analysis on the unknown factors $\alpha_{\mathrm{1}}$ - $\alpha_{\mathrm{4}}$, this is in rough agreement with the results of our simulations. The modulation of the Alfvén flux injected in the slab, caused by the modulation of the central density perturbation, $\rho_{\mathrm{1}}(0)$, constitutes a feedback which endows $\omega^2$ with an imaginary part, and thus induces instability, even when the system is stable under constant Alfvén wave energy flux. It is interesting to note that such a modulation is indeed present in our simulations. The coherent oscillations observed in our calculations may represent a nonlinear state of development of such an instability. At high amplitude, the oscillations of the different fluid elements would cease to be isochronic and their initial coherence may be lost.
Driving of turbulence and size of structure
-------------------------------------------
The one-point forcing we apply in our simulations is in contrast to most other investigation of highly compressible turbulence, where the forcing is applied at each grid point (see e.g. @maclow:99 for a description of such forcing and @gammie-ostriker:96 for the particular case of a 1D-slab). Our results show that also such one point forcing is perfectly capable of supporting and structuring a slab over large distances. The injected waves are very efficient in distributing the driving energy over a large spatial range.
For the distance over which the injected energy is spread, as well as for the scale of the density structure that forms, the driving wave-length $\lambda_{\mathrm{A0}}$ is decisive. This is not too surprising. In Sect. \[sec:dimensionless\] we have seen that there is no other small length scale likely to play a role in the problem we consider. It then appears quite natural that $\lambda_{\mathrm{A0}}$ governs the structure size. Remembering that smaller scale structure means more inhomogeneity on a smaller spatial scale, a faster destruction of the injected energy for smaller $\lambda_{\mathrm{A0}}$ also seems natural.
That local energy injection can be sufficient to drive turbulence in a whole volume has also been reported for the case of the interaction zone of hypersonically colliding flows in 2D plane-parallel hydrodynamics . Although energy there is injected only at the oblique shocks confining the interaction zone, supersonic turbulence persists in the whole volume. A correlation between driving wave-length and structure size exists as well (Folini & Walder, in preparation). A correlation between structures size and the wave-length of the driving was also reported by @maclow:99 for hydrodynamic turbulence in a 3D periodic box, which was monochromatically forced at each grid point.
Numerical resolution {#sec:disc_num}
--------------------
We have experimented with different spatial resolutions and different orders of integration for the MHD equations. Not surprising, we find that simulations with shorter wave-lengths of the Alfvén waves are more delicate in terms of spatial resolution and order of integration. For run R20.20.5.4.10 we have verified that our results remain essentially unchanged (mean quantities differ by a few percent at most) if we further increase the spatial resolution. For coarser discretizations (using 1000 cells instead of 5000) or lower orders of integration, strong damping of the Alfvén wave due to numerical diffusion occurs and both the density substructure as well as the additional support against self-gravity are essentially lost, as can be seen in Fig. \[fig:spaceres\]. For run R20.20.5.4.10, for example, $\langle d^{\mathrm{sol}}/d_{\mathrm{0}}^{\mathrm{wkb}}
\rangle_{\mathrm{t}}$ is reduced from 0.46 to 0.32, for run R20.20.10.4.10 the value is reduced from 0.75 to 0.48. With regard to numerical simulations, these results show that the problem we investigated in this paper is currently out of reach for 3D simulations.
Connection to high density cores
--------------------------------
We, finally, would like to return to the larger frame, molecular clouds and the formation of dense cores. As mentioned earlier, both theoretical and observational results today indicate that the immediate vicinity of a forming dense core is highly structured. Most striking are the observations by @2001ApJ...555..178F of the filaments around the starless dense core L1512. Various explanations for their existence, all employing magnetic fields, seem plausible . It seems most likely that such structures, or rather the responsible physical processes, affect the formation and accretion rate of dense cores.
Detailed studies of this structuring and its consequences are mandatory to better link turbulence in molecular clouds with star formation. Such studies are, however, only at their beginning. The 1D slab we studied here is by no means an accurate model of a molecular cloud. Nevertheless, it shows that magnetic waves can play a crucial role in counteracting self-gravity. The 3D grid studies by @2001ApJ...547..280H, mentioned in the introduction, point in the same direction. And while the 2D and 3D MHD simulations by @del-zanna-et-al:01 of a plan parallel slab focus on structure formation rather than on accretion, they show that density filaments form, more or less aligned with the background magnetic field.
With regard to future simulations in this field, existing simulations show spatial resolution to be decisive. Another crucial issue is dimensionality. The ’fingering’ observed by @del-zanna-et-al:01 - and by @2001ApJ...555..178F - clearly cannot be studied in a 1D model. And, as the authors showed as well, it can only be observed if open boundaries are used, not periodic ones. Finally, and probably most crucial of all, a future model of dense core formation will have to account for the generation of magnetic waves self-consistently. This may be the strongest reason to aim, in a next step, at multidimensional models.
Conclusions {#sec:conclusions}
===========
For the case of a 1D plane-parallel, self-gravitating slab we have shown by means of numerical simulations and dimensional analysis the following:
1\) One source of energy injection is sufficient to sustain turbulence throughout the slab.
2\) The time-averaged spatial extension of the turbulent slab is comparable to the extension of the corresponding WKB solution. Deviations are at most a factor of three and depend about linearly on the ratio of the initial central Alfvén wave-length and the extension of the WKB solution.
3\) The scale of the substructure is governed by the initial central Alfvén wave-length and the temperature. Larger wave-lengths and smaller temperatures lead to larger scale structure.
4\) The energy loss, and thus the energy radiated by the slab, is dominated by the loss of Poynting-flux, which increases almost linearly with $\sqrt{\beta_{\mathrm{0}}}$.
5\) Within the slab, the energy density of the transverse magnetic field and the kinetic energy density of the transverse velocities are in approximate equilibrium. The latter accounts for between 70% and 90% of the total kinetic energy density.
\(6) A too coarse mesh or a low order scheme leads to substantial wave-damping, thus to loss of structuring and support.
D.F. was supported by a grant of the Swiss National Science Foundation, grant number 8220-056553. R.W. was supported by the Scientific Discovery through Advanced Computing (SciDAC) program of the US-DOE, grant number DE-FC02-01ER41184. R.W. thanks the Institut für Astronomie at ETH Zürich for its hospitality and providing a part-time office. J.H. thanks the EC Platon program HTRN-CT-2000-00153 and the Platon group for helpful discussions. The calculations were done on the NEC SX5 at IDRIS, Paris, France, and on the CRAY SV1B at ETH Zürich, Switzerland. The authors would like to thank the referee Dr. Hanawa for his valuable comments which helped to improve the paper.
Simulation parameters
=====================
[|l||c|c||c|c|c|c|c|c|c||c|c|c|c|c|]{} 6.5cm
------------------------------------------------------------------------
run & $\lambda_\mathrm{A0}$ & $\frac{{\cal M}_\mathrm{0}}{m_\mathrm{H}{\cal D}}$ & $\frac{{\cal M}_{\mathrm{0}}}{{\cal M}^{\mathrm{\infty}}_{\mathrm{0}}}$ & $\frac{{d}^{\mathrm{wkb}}_{\mathrm{0}}}{{d}^{\mathrm{hs}}_{\mathrm{0}}}$ & $\frac{{d}^{\mathrm{wkb}}_{\mathrm{0}}}{\cal D}$ & $\alpha_{\mathrm{0}}$ & $\beta_{\mathrm{0}}$ & $W_{\mathrm{0}}$ & $W_{\mathrm{\lambda A 0}}$ & $\frac{\langle {d}^{\mathrm{sol}} \rangle_{t} }{{\cal D}}$ & $\frac{\langle {d}^{\mathrm{sol}} \rangle_{t}}{{d}^{\mathrm{wkb}}_{\mathrm{0}}}$ & $\sigma_{\mathrm{d}}^{\mathrm{rel}}$ & $\frac{{\cal M}_\mathrm{20}}{{\cal M}_\mathrm{0}}$ & $\Delta P$\
R100.45.1.4.10 & 1 & 195 & 0.30 & 14.3 & 0.37 & 0.20 & 0.0035 & 750. & 0.032 & 0.12 & 0.32 & 0.24 & 1.02 & 6.9
------------------------------------------------------------------------
\
R100.45.5.4.10 & 5 & 195 & 0.30 & 14.3 & 0.37 & 0.20 & 0.0035 & 150. & 0.16 & 0.38 & 1.02 & 0.74 & 0.89 & 5.3\
R100.45.5.8.10 & 3.5 & 220 & 0.34 & 9.9 & 0.23 & 0.20 & 0.0069 & 110. & 0.18 & 0.33 & 1.43 & 0.76 & 0.96 & 7.3\
R100.20.1.4.10 & 1 & 116 & 0.38 & 5.5 & 0.24 & 0.04 & 0.0035 & 750. & 0.049 & 0.12 & 0.50 & 0.22 & 1.02 & 5.3\
R100.20.5.4.10 & 5 & 116 & 0.38 & 5.5 & 0.24 & 0.04 & 0.0035 & 150. & 0.24 & 0.41 & 1.71 & 0.90 & 0.92 & 4.2\
R100.10.5.4.10 & 5 & 82.1 & 0.45 & 2.6 & 0.16 & 0.01 & 0.0035 & 150. & 0.37 & 0.30 & 1.88 & 0.77 & 1.02 & 3.6\
R50.45.10.8.10 & 3.5 & 220 & 0.34 & 9.9 & 0.23 & 0.81 & 0.028 & 53. & 0.18 & 0.32 & 1.39 & 0.52 & 0.93 & 2.6\
R20.20.5.4.10 & 1 & 116 & 0.38 & 5.5 & 0.24 & 1.00 & 0.087 & 150. & 0.045 & 0.11 & 0.46 & 0.14 & 1.02 & 21.\
R20.20.10.4.10 & 2 & 116 & 0.38 & 5.5 & 0.24 & 1.00 & 0.087 & 75. & 0.098 & 0.18 & 0.75 & 0.20 & 1.00 & 19.\
R20.20.12.4.10 & 2.5 & 116 & 0.38 & 5.5 & 0.24 & 1.00 & 0.087 & 60. & 0.12 & 0.18 & 0.75 & 0.12 & 1.02 & 20.\
R20.20.17.4.10 & 3.5 & 116 & 0.38 & 5.5 & 0.24 & 1.00 & 0.087 & 43. & 0.17 & 0.25 & 1.04 & 0.24 & 0.99 & 17.\
R20.20.25.4.10 & 5 & 116 & 0.38 & 5.5 & 0.24 & 1.00 & 0.087 & 30. & 0.24 & 0.34 & 1.42 & 0.30 & 0.98 & 19.\
R20.20.25.8.10 & 3.5 & 136 & 0.42 & 3.8 & 0.14 & 1.00 & 0.17 & 21. & 0.30 & 0.18 & 1.39 & 0.88 & 1.02 & 22.\
R20.20.5.1.10 & 2 & 85.7 & 0.30 & 9.7 & 0.57 & 1.00 & 0.022 & 300. & 0.041 & 0.23 & 0.40 & 0.85 & 0.99 & 13.\
R20.10.5.4.10 & 1 & 82.1 & 0.45 & 2.6 & 0.16 & 0.25 & 0.087 & 150. & 0.073 & 0.11 & 0.68 & 0.09 & 1.02 & 19.\
R20.10.25.4.10 & 5 & 82.1 & 0.45 & 2.6 & 0.16 & 0.25 & 0.087 & 30. & 0.37 & 0.24 & 1.50 & 0.42 & 1.01 & 13.\
R20.10.2.1.10 & 1 & 56.4 & 0.37 & 4.8 & 0.43 & 0.25 & 0.022 & 600. & 0.027 & 0.20 & 0.47 & 0.19 & 0.99 & 12.\
R20.10.5.1.10 & 2 & 56.4 & 0.37 & 4.8 & 0.43 & 0.25 & 0.022 & 300. & 0.055 & 0.24 & 0.56 & 0.13 & 0.99 & 11.\
R20.10.12.1.10 & 5 & 56.4 & 0.37 & 4.8 & 0.43 & 0.25 & 0.022 & 120. & 0.14 & 0.42 & 0.98 & 0.15 & 0.93 & 10.\
R20.20.5.4.5 & 1 & 102. & 0.35 & 8.7 & 0.22 & 1.00 & 0.043 & 150. & 0.053 & 0.09 & 0.41 & 0.15 & 1.03 & 19.\
R20.20.5.4.20 & 1 & 135. & 0.41 & 3.6 & 0.27 & 1.00 & 0.17 & 150. & 0.044 & 0.13 & 0.48 & 0.13 & 1.01 & 25.\
R20.20.5.4.40 & 1 & 162. & 0.44 & 2.4 & 0.30 & 1.00 & 0.35 & 150. & 0.039 & 0.18 & 0.60 & 0.14 & 1.02 & 32.\
R20.20.25.4.5 & 5 & 102. & 0.35 & 8.7 & 0.22 & 1.00 & 0.043 & 30. & 0.27 & 0.31 & 1.41 & 0.42 & 0.98 & 17.\
R20.20.25.4.15 & 5 & 126. & 0.40 & 4.3 & 0.26 & 1.00 & 0.13 & 30. & 0.23 & 0.28 & 1.08 & 0.22 & 0.97 & 17.\
R20.20.25.4.20 & 5 & 135. & 0.41 & 3.6 & 0.27 & 1.00 & 0.17 & 30. & 0.22 & 0.26 & 0.96 & 0.17 & 0.96 & 20.\
R20.20.25.4.40 & 5 & 162. & 0.44 & 2.4 & 0.30 & 1.00 & 0.35 & 30. & 0.20 & 0.27 & 0.90 & 0.15 & 0.97 & 24.\
R20.20.25.8.5 & 3.5 & 117. & 0.38 & 5.9 & 0.13 & 1.00 & 0.087 & 21. & 0.32 & 0.20 & 1.54 & 0.93 & 1.05 & 18.\
R20.20.25.8.20 & 3.5 & 164. & 0.45 & 2.6 & 0.16 & 1.00 & 0.35 & 21. & 0.26 & 0.17 & 1.06 & 0.27 & 1.02 & 29.\
R20.20.25.8.40 & 3.5 & 205. & 0.47 & 1.8 & 0.18 & 1.00 & 0.69 & 21. & 0.23 & 0.19 & 1.06 & 0.32 & 1.02 & 39.\
R20.20.5.1.20 & 2 & 93.9 & 0.32 & 5.7 & 0.61 & 1.00 & 0.043 & 300. & 0.039 & 0.28 & 0.46 & 0.14 & 0.99 & 14.\
R20.10.5.4.5 & 1 & 68.2 & 0.42 & 3.8 & 0.14 & 0.25 & 0.043 & 150. & 0.084 & 0.10 & 0.71 & 0.18 & 1.02 & 15.\
R20.10.5.4.20 & 1 & 102. & 0.47 & 1.8 & 0.18 & 0.25 & 0.17 & 150. & 0.065 & 0.13 & 0.73 & 0.12 & 1.02 & 26.\
R20.10.25.4.20 & 5 & 102. & 0.47 & 1.8 & 0.18 & 0.25 & 0.17 & 30. & 0.33 & 0.25 & 1.39 & 0.45 & 0.98 & 17.\
R20.10.25.4.40 & 5 & 133. & 0.49 & 1.4 & 0.22 & 0.25 & 0.35 & 30. & 0.27 & 0.28 & 1.27 & 0.14 & 0.99 & 25.\
R20.10.12.1.5 & 5 & 49.9 & 0.34 & 7.6 & 0.39 & 0.25 & 0.011 & 120. & 0.15 & 0.46 & 1.18 & 0.29 & 0.94 & 8.8\
R20.10.12.1.20 & 5 & 65.5 & 0.40 & 3.0 & 0.47 & 0.25 & 0.043 & 120. & 0.13 & 0.46 & 0.98 & 0.11 & 0.95 & 13.\
\[tab:mod\_params\]
[^1]: The paper by Kudoh & Basu was submitted during the refereeing process of this paper.
[^2]: The code is part of the A-MAZE code package [@amaze:00], comprising 3D adaptive mesh codes for magneto-hydrodynamics and radiative transfer. The codes are available at http://www.astro.phys.ethz.ch/staff/folini.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently, a generalized test approach is proposed by [@Sadooghi-ja-ma-16] and a fiducial approach is proposed by [@xu-li-18] to test the equality of coefficients in several regression models with unequal variances. In this paper, it is shown that the considered quantities in these approaches are identically distributed and therefore, these approaches are same. Also, this result satisfies for the one-way ANOVA problem.'
author:
- |
A. A. Jafari$\thanks{[email protected]}$\
[Department of Statistics, Yazd University, Yazd, Iran]{}
date:
-
-
title: |
Comparing Two Approaches in Heteroscedastic\
Regression Models
---
[**Keywords:**]{} Fiducial approach; Generalized test variable; One-Way ANOVA; Regression.
Introduction
============
Consider the $k$ regression models $${{\boldsymbol Y}}_i =X_i{{\boldsymbol \beta }}_i+{{\boldsymbol \varepsilon }}_i\ \ \ \ {{\boldsymbol \varepsilon }}_i~\sim N\left({\boldsymbol 0}{\boldsymbol ,\ }{\sigma }^2_iI_{n_i}\right),\ \ \ \ \ \ \ i=1,\dots ,k,$$ where $X_i$ is $n_i\times p$ design matrix with rank $p$, ${{\boldsymbol Y}}_i{\boldsymbol =(}Y_{i1},\dots ,Y_{{in}_i})'$ is $n_i\times 1$ ($n_i>p$ for all $i$) observation vector, ${{\boldsymbol \beta }}_i=({\beta }_{i1},{\beta }_{i2},\dots ,{\beta }_{ip})'$ is the vector of parameters with dimension of $p$, ${\boldsymbol \varepsilon}_i$ is the $n_i\times 1$ disturbance vector, and $I_{n_i}$ is $n_i\times n_i$ identity matrix. Furthermore, all the ${{\boldsymbol \varepsilon }}_i$ are independent.
It is well-known that the unbiased estimations for ${{\boldsymbol \beta }}_i$ and ${\sigma }^2_i$ are ${\hat{{\boldsymbol \beta }}}_i ={\left(X'_iX_i\right)}^{-1}X'_i{{\boldsymbol Y}}_i$ and $S^2_i=\boldsymbol Y'_i\left(I_p-X_i{\left(X'_iX_i\right)}^{-1}X'_i\right){\boldsymbol Y}_i/(n_i-p)$, respectively, such that they are independent. When the variances ${\sigma }^2_i$’s are unknown, an usual test statistic to test $$\label{eq.H0}
H_0:{{\boldsymbol \beta }}_1={{\boldsymbol \beta }}_2=\dots ={{\boldsymbol \beta }}_k,$$ is [see @tian-ma-ve-09; @Sadooghi-ja-ma-16; @xu-li-18] $$Q_0=\sum^k_{i=1}S^{-2}_i\hat{\boldsymbol \beta}'_i\left(X'_iX_i\right) \hat{\boldsymbol \beta}_i-
\left[\sum^k_{{\boldsymbol i}=1}{S^{-2}_i\hat{\boldsymbol \beta }'_i\left(X'_iX_i\right)}\right]{\left[\sum^k_{i=1}{S^{-2}_i\left(X'_iX_i\right)}\right]}^{-1}\left[\sum^k_
{i=1}S^{-2}_i\left(X'_iX_i\right)\hat{\boldsymbol \beta }_i\right].$$
Under null distribution, $Q_0$ can be approximated by the chi-square distribution with $p (k-1)$ degrees of freedom for large sample sizes. However, this approximation does not work well for small samples [see @tian-ma-ve-09; @Sadooghi-ja-ma-16; @xu-li-18]. Therefore, some other approaches are proposed to test equality of regression models with unequal variances for example a parametric bootstrap approach by [@tian-ma-ve-09], a generalized approach by [@Sadooghi-ja-ma-16] and a fiducial approach by [@xu-li-18]. In this paper, we compare these generalized and fiducial approaches. We will see that although these approaches are proposed in different ways and are not the same in appearance but they are identical.
In Section \[sec.com\], the generalized and fiducial approaches are reviewed and compared. In Section \[sec.anova\], one way-ANOVA is discussed as a special case.
Comparing two approaches {#sec.com}
========================
For testing $H_0$ in , a generalized approach is proposed by [@Sadooghi-ja-ma-16] and a fiducial approach is proposed by [@xu-li-18]. In this section, these approaches are reviewed, briefly. Then, it is shown that they are identical.
Consider ${\boldsymbol b}_i$ and $s^2_i$, $i=1,\dots,k$, are the observed values of $\hat{\boldsymbol \beta}_i$ and $S^2_i$, respectively. For given $\left(\boldsymbol b_i, s^2_i\right)$, [@xu-li-18] derived a fiducial quantity to test as $$\label{eq.QF}
Q_F=\sum^k_{i=1}{\boldsymbol t}'_i{\boldsymbol t}_i-\left[\sum^k_{i=1}{s^{-1}_i{{\boldsymbol t}}'_i{\left(X'_iX_i\right)}^{\frac{1}{2}}}\right]{\left[\sum^k_{i=1}{s^{-2}_i\left(X'_iX_i\right)}\right]}^{-1}
\left[\sum^k_{i=1}{s^{-1}_i{\left(X'_iX_i\right)}^{\frac{1}{2}}{\boldsymbol t}_i}\right],$$ where ${\boldsymbol t}_i$ follows a multivariate student’s t-distribution $t_p(n_i-p,{\boldsymbol 0},I_p)$, $i=1,\dots ,k$, [see @kotz-na-04multi] and they are mutually independent. Therefore, the p-value to test is given by $P(Q_F>Q_0)$.
Let $H=C \bigotimes I_p$ where $\bigotimes $ denotes Kronecker product, $C=\left[I_{k-1}:{\boldsymbol 1}\right]$, and ${\boldsymbol 1}=(1,\dots ,1)'$. Also, consider $W'W={\left[HSH'\right]}^{-1}$ where $S=\left[{\rm diag}\left(s^2_i{\left(X'_jX_j\right)}^{-1}\right)\right]$. [@Sadooghi-ja-ma-16] defined a generalized test variable to test the hypothesis in as $$\label{eq.QG}
Q_G={{\boldsymbol Z}}'WH\left[{\rm diag}\left(\frac{\left(n_i-p\right)s^2_i}{U_i}{\left(X'_jX_j\right)}^{-1}\right)\right]H'W'{\boldsymbol Z},$$ where ${\boldsymbol Z} \sim N ({\boldsymbol 0},I_{p(k-1)} )$ and $U_i\sim{\chi }^2_{(n_i-p)}$, $i=1,\dots,k,$ such that ${\boldsymbol Z}$ and $U_i$’s are mutually independent. Therefore, the generalized p-value is given by $P(Q_G>Q_0)$.
$Q_F$ and $Q_G$ are identically distributed.
Consider $\boldsymbol V_i \sim N ({\boldsymbol 0},I_p )$ and $U_i\sim {\chi }^2_{(n_i-p)}$, $i=1,\dots ,k,$ such that $\boldsymbol V_i$ and $U_i$’s are mutually independent. Also, consider $D =\left[{\rm diag}\left(\frac{n_i-p}{U_i}\right)\right]\bigotimes I_p$. Based on [@Sadooghi-ja-ma-16], $Q_G$ has the same distribution as $$Q^*_G= {\boldsymbol V}'\left(I_{pk}-qq'\right)D\left(I_{pk}-qq'\right){\boldsymbol V},$$ where $${\boldsymbol V}=\left[ \begin{array}{c}
{{\boldsymbol V}}_1 \\
{{\boldsymbol V}}_2 \\
\vdots \\
{{\boldsymbol V}}_k \end{array}
\right],\ \ \ q=\left[ \begin{array}{c}
q_1 \\
q_2 \\
\vdots \\
q_k \end{array}
\right],\ \ \ \ q_i=\left[s^{-1}_i{\left(X'_iX_i\right)}^{1/2}\right]{\left[\sum^k_{j=1}{s^{-2}_j\left(X'_jX_j\right)}\right]}^{-1/2}.$$ Consider $P=\left(I_{pk}-qq'\right)$ and $D^*=\left[{\rm diag}\left(\sqrt{\frac{n_i-p}{U_i}}\right)\right]\bigotimes I_p$. Then $$\begin{aligned}
Q^*_G&=&{{\boldsymbol V}}'PD^*D^*P{\boldsymbol V} = {{\boldsymbol V}}'D^*PPD^*{\boldsymbol V}
\\
&=&{{\boldsymbol V}}'D^*PD^*{\boldsymbol V}\\
&=&{{\boldsymbol V}}'D^*\left(I_{pk}-qq'\right)D^*{\boldsymbol V}\\
&=&{{\boldsymbol V}}'D{\boldsymbol V}-{{\boldsymbol V}}'D^*qq'D^*{\boldsymbol V}\\
&=&\sum^k_{i=1}{\frac{n_i-p}{U_i}{{\boldsymbol V}}'_i{{\boldsymbol V}}_i}-\left[\sum^k_{i=1}{\sqrt{\frac{n_i-p}{U_i}}s^{-1}_i
{{\boldsymbol V}}'_i {\left(X'_iX_i\right)}^{\frac{1}{2}}}\right]{\left[\sum^k_{i=1}{s^{-2}_i\left(X'_iX_i\right)}\right]}^{-1}\\
&&\times
\left[\sum^k_{i=1}{\sqrt{\frac{n_i-p}{U_i}}s^{-1}_i{\left(X'_iX_i\right)}^{\frac{1}{2}}{{\boldsymbol V}}_i}\right]\\
&=&\sum^k_{i=1}{{{\boldsymbol T}}'_i{{\boldsymbol T}}_i}-\left[\sum^k_{i=1}{s^{-1}_i{{\boldsymbol T}}'_i{\left(X'_iX_i\right)}^{\frac{1}{2}}}\right]{\left[\sum^k_{i=1}{s^{-2}_i\left(X'_iX_i\right)}\right]}^{-1}
\left[\sum^k_{i=1}{s^{-1}_i{\left(X'_iX_i\right)}^{\frac{1}{2}}{{\boldsymbol T}}_i}\right],\end{aligned}$$ where ${{\boldsymbol T}}_i=\sqrt{\frac{n_i-p}{U_i}}{{\boldsymbol V}}_i\sim t_p(n_i-p,{\boldsymbol 0},I_p)$. Therefore, $Q_F$ and $Q_G$ are identically distributed.
One Way ANOVA problem {#sec.anova}
======================
Let $Y_{i1},Y_{i2},\dots ,Y_{in_i}$ is a random sample from a normal distribution with mean ${\mu }_i$ and variance ${\sigma }^2_i$, $i=1,\dots ,k$. The problem of testing equality of means of these $k$ distributions, i.e. $$H^*_0:{\mu }_1=\dots ={\mu }_k,$$ is well-known to one-way ANOVA. It is a special case of $H_0$ in with $p=1$ and $X_i={\boldsymbol 1}$. Therefore, the fiducial quantity in becomes to $$\label{eq.QF2}
Q_F=\sum^k_{i=1}{t^2_i}-\frac{{\left(\sum^k_{i=1}{\sqrt{\frac{n_i}{s_i}}t_i}\right)}^2}{\sum^k_{i=1}{\frac{n_i}{s^2_i}}},$$ where $t_i$ has a t distribution with $n_i-1$ degrees of freedom, which is the fiducial quantity introduced by [@li-wa-li-11]. When $p=1$ and $X_i={\boldsymbol 1}$, the generalized test variable in becomes to $$\label{eq.QG2}
Q_G={{\boldsymbol Z}}'WC\left[{\rm diag}\left(\frac{\left(n_i-p\right)s^2_i}{n_iU_i}\right)\right]C'W'{\boldsymbol Z},$$ where $W'W={\left[CSC'\right]}^{-1}$, $S=\left[{\rm diag}\left(s^2_i/n_i\right)\right]$, ${\boldsymbol Z}\sim N ({\boldsymbol 0},I_{(k-1)} )$ and $U_i\sim {\chi }^2_{(n_i-p)}$, $i=1,\ \dots ,k$. This generalized test variable is introduced by [@sa-ja-ma-12]. This generalized test variable is also proposed by [@xu-wa-08-anova] in another form [see @sa-ja-ma-12]. Based on Lemma, $Q_F$ in and $Q_G$ in are identically distributed. Therefore, the all proposed approaches by [@xu-wa-08-anova], [@li-wa-li-11] and [@sa-ja-ma-12] for the one-way ANOVA problem with unequal variances are same.
Kotz, S. and Nadarajah, S. (2004). . Cambridge University Press, Cambridge.
Li, X., Wang, J., and Liang, H. (2011). Comparison of several means: A fiducial based approach. , 55(5):1993–2002.
Sadooghi-Alvandi, S. M., Jafari, A. A., and Mardani-Fard, H. A. (2012). One-way [ANOVA]{} with unequal variances. , 41(22):4200–4221.
Sadooghi-alvandi, S. M., Jafari, A. A., and Mardani-Fard, H. A. (2016). Comparing several regression models with unequal variances. , 45(9):3190–3216.
Tian, L., Ma, C., and Vexler, A. (2009). A parametric bootstrap test for comparing heteroscedastic regression models. , 38(5):1026–1036.
Xu, J. and Li, X. (2018). A fiducial p-value approach for comparing heteroscedastic regression models. , 47(2):420–431.
Xu, L.-W. and Wang, S.-G. (2008). A new generalized p-value for anova under heteroscedasticity. , 78(8):963–969.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we introduce dynamic time-frequency-division duplex (D-TFDD), which is a novel duplexing scheme that combines time-division duplex (TDD) and frequency-division duplex (FDD). In D-TFDD, a user receives from the base station (BS) on the downlink in one frequency band and transmits to the BS on the uplink in another frequency band, as in FDD. Next, the user shares its uplink transmission (downlink reception) on the corresponding frequency band with the uplink transmission or the downlink reception of another user in a D-TDD fashion. Hence, in a given frequency band, the BS communicates with user 1 (U1) and user 2 (U2) in a D-TDD fashion. The proposed D-TFDD scheme does not require inter-cell interference (ICI) knowledge and only requires channel state information (CSI) of the local BS-U1 and BS-U2 channels. Thereby, it is practical for implementation. The proposed D-TFDD scheme increases the throughput region between the BS and the two users in a given frequency band, and significantly decreases the outage probabilities on the corresponding BS-U1 and BS-U2 channels. Most importantly, the proposed D-TFDD scheme doubles the diversity gain on both the corresponding BS-U1 and the BS-U2 channels compared to the diversity gain of existing duplexing schemes, which results in very large performance gains.'
author:
- 'Mohsen Mohammadkhani Razlighi, Nikola Zlatanov, and Petar Popovski [^1] [^2] [^3]'
bibliography:
- 'litdab.bib'
title: 'Dynamic Time-Frequency Division Duplex'
---
Introduction {#Sec-Intro}
============
Traditionally, a half-duplex (HD) base station (BS) operates in either the time-division duplex (TDD) mode or the frequency-division duplex (FDD) mode in order to receive and transmit information from/to its users. In the TDD mode, a user uses the same frequency band for uplink and downlink, while uplink and downlink transmissions occur in different time slots [@holma2011lte], see Fig. \[Fig:sys-mod-TDD-FDD\]. On the other hand, in the FDD mode, a user is allocated two frequency bands, one dedicated for uplink and the other one for downlink, where the uplink and downlink transmissions occur simultaneously [@holma2011lte], see Fig. \[Fig:sys-mod-TDD-FDD\]. In this paper, we introduce time-frequency-division duplex (TFDD), which is a novel duplexing scheme that combines TDD and FDD, and yields significant performance gains compared to TDD and FDD.
Background on the Different Types of Duplexing Schemes
------------------------------------------------------
### Static vs. Dynamic Duplexing
In general, the duplexing method between a BS and its users can be static or dynamic. In static duplexing, the time-frequency resources in which the BS performs uplink receptions and downlink transmissions from/to the user are prefixed and unchangeable over time[@TechNoteLET]. On the other hand, in dynamic duplexing schemes, each time-frequency resource unit can be dynamically allocated for communications based on the instantaneous channel state information (CSI). As a result, dynamic duplexing schemes achieve a much better performance compared to static duplexing schemes [@7511412; @6353682], and thereby have attracted significant research interest [@7511412; @6353682; @1386526; @7876862; @8004461].
### Centralized Dynamic Duplexing vs. Distributed Dynamic Duplexing
A dynamic duplexing scheme can be implemented in either centralized or distributed fashion [@1386526]. In centralized dynamic duplexing schemes, the decision for allocating the time-frequency resources for communication is performed at a central node, which then informs all BSs about the decision. In this way, the communication between neighbouring cells can be synchronized in order to minimize inter-cell interference[^4] (ICI) [@957300; @7208806; @7000558; @1336656; @4769393; @7876862; @7636855; @8016428; @7491359; @8004461; @7003998]. However, centralized dynamic duplexing schemes require at the central node full CSI from all links in all cells in order for the central node to make an optimal decision for allocating the time-frequency resources for each BS. In addition, the central node also needs to inform all other network nodes about the scheduling decisions. This requires a large amount of signalling information to be exchanged between the central node and all other network nodes. As a result, implementation of centralized dynamic duplexing schemes, in most cases, is infeasible in practice.
On the other hand, in distributed dynamic duplexing schemes, each BS allocates the time-frequency resources for its users without any synchronization with other BSs [@4769393; @6666413; @7070655]. To this end, only local CSI is needed at each BS. As a result, distributed dynamic duplexing schemes are much more appropriate for practical implementation compared to the centralized dynamic duplexing scheme. However, distributed dynamic duplexing schemes have to cope with higher ICI than centralized dynamic duplexing schemes.
The proposed TFDD scheme can be characterized as a distributed dynamic duplexing scheme, which means it is suitable for practical implementation.
Contribution
------------
In this paper, we introduce dynamic time-frequency-division duplex (D-TFDD), which is a novel duplexing scheme that combines D-TDD and FDD. In D-TFDD, a user receives from the BS on the downlink in one frequency band and transmits to the BS on the uplink in another frequency band, as in FDD. Next, the user shares its uplink transmission (downlink reception) on the corresponding frequency band with the uplink transmission or the downlink reception of another user in a D-TDD fashion. Hence, in a given frequency band, the BS communicates with user 1 (U1) and user 2 (U2) in a D-TDD fashion. The proposed D-TFDD scheme does not require ICI knowledge and only requires CSI of the local BS-U1 and BS-U2 channels. Thereby, it is practical for implementation. The proposed D-TFDD scheme increases the throughput region between the BS and the two users in a given frequency band, and significantly decreases the outage probabilities on the corresponding BS-U1 and BS-U2 channels. Most importantly, the proposed D-TFDD scheme doubles the diversity gain on both the corresponding BS-U1 and the BS-U2 channels compared to the diversity gain of existing duplexing schemes, which results in very large performance gains.
Relevance of D-TFDD to 5G and Beyond
------------------------------------
One of the prominent aspects of fifth generation (5G) mobile networks is having a flexible physical layer design. In one hand, this capability facilitates implementing challenging physical layer protocols, and on the other hand opens the door for unconventional schemes to be implemented on the physical layer. Such a flexible hardware-software design can easily accommodate our D-TFDD scheme and thereby improve the performance of 5G networks[@wong_schober_ng_wang_2017]. In addition, distributed resource allocation for dense heterogeneous wireless networks is in one of the main scopes of 5G [@wong_schober_ng_wang_2017], which also fits well with our D-TFDD scheme. Moreover, proposed scheme is applicable to multi-tier multi-cell systems, which is another feature of 5G networks.
Related Works on D-TDD and D-FDD
--------------------------------
Distributed D-TDD schemes for BSs have been investigated in [@6666413; @7070655; @1705939; @4556648; @1638665; @7136469; @4524858] and references therein. In particular, [@6666413] proposed cooperation among cells that resembles a centralized D-TDD scheme. The authors in [@7070655] proposed a D-TDD scheme which alleviates the ICI impairment by splitting the uplink and downlink frequency. Authors in [@1705939] and [@4556648] investigate a distributed D-TDD scheme designed for multiple-antennas. The work in [@1638665] proposes a distributed multi-user D-TDD scheduling scheme, where the ICI is not taken into account, which may lead to poor performance in practice. The authors in [@7136469] investigated an identical network as in [@1638665], but with ICI taken into account. However, the solution in [@7136469] is based on a brute-force search algorithm for allocating the time slots. Authors in [@4524858] proposed a D-TDD scheme that performs optimal power, rate, and user allocation. However, the ICI level in [@4524858] is assumed to be fixed during all time slots, which may not be an accurate model of ICI in practice, since due to the fading and the power-allocations at neighbouring BSs, the ICI varies with time.
On the other hand, D-FDD has been introduced in [@4769393], where the authors proposed a scheme for adapting the downlink to uplink bandwidth ratio.
We note that [@4769393; @6666413; @7070655; @1705939; @4556648; @1638665; @7136469; @4524858] require full knowledge of the ICI, which may not be practical, as discussed in Sec. \[Estimation\_discuss\]. We also note that the schemes in [@4769393; @6666413; @7070655; @1705939; @4556648; @1638665; @7136469; @4524858] transform to the static-TDD and/or static-FDD scheme when ICI is not known. Hence, the static-TDD and/or static-FDD are much more practical for implementation than the D-FDD duplexing scheme since they do not require ICI knowledge.
The rest of this paper is organized as follows. In Section \[Sec-Sys\], we present the system and channel model. In Sections \[Sec-FR\] and \[subSec-PRL\], we present the D-TFDD schemes for the cases when the ICI is known and unknown, respectively. Simulation and numerical results are provided in Section \[Sec-Num\], and the conclusions are drawn in Section \[Sec-Conc\].
System and Channel Model {#Sec-Sys}
========================
In the following, we consider a cellular network consisting of cells, where each cell has a single BS and users that the BS serves.
Frequency and Time Allocation in D-TFDD
---------------------------------------
In the proposed D-TFDD scheme, we have two possible frequency allocation schemes at each BS, Frequency Allocation Scheme 1 shown in Fig. \[Fig:sys-mod-Type23\] and Frequency Allocation Scheme 2 shown in Fig. \[Fig:sys-mod-Type1\]. In both frequency allocation schemes, each user is allocated two distinct frequency bands whithin the cell, one for uplink transmission and the other for downlink reception, identical as in FDD, see Figs. \[Fig:sys-mod-Type23\] and \[Fig:sys-mod-Type1\]. In Frequency Allocation Scheme 1, the frequency band of a user allocated for uplink transmission (downlink reception) is shared in a D-TDD fashion with the uplink transmission (downlink reception) of another user, as shown in Figs. \[Fig:sys-mod-Type23\]. Whereas, in Frequency Allocation Scheme 2, the frequency band of a user allocated for uplink transmission (downlink reception) is shared in a D-TDD fashion with the downlink reception (uplink transmission) of another user, as shown in Figs. \[Fig:sys-mod-Type1\]. Hence, for $N$ users, the BS needs to allocate $N$ frequency bands in D-TFDD, same as in TDD.
Frequency Allocation Scheme 1 is more appropriate for cellular communication networks, where the transmit powers of BSs are much higher than the transmit powers of the users. Specifically, when Frequency Allocation Scheme 1 is applied to every BS in the cellular network such that the uplink frequency bands are used only for uplink and the downlink frequency bands are used only for downlink at all BSs, then all uplink links receive inter-cell interference only from other uplink transmissions and all downlink links receive inter-cell interference only from other downlink transmissions. As a result, the problem in existing D-TDD schemes of strong downlink transmission from one BS interfering with the weak uplink reception at another BS is avoided in D-TFDD with Frequency Allocation Scheme 1. On the other hand, Frequency Allocation Scheme 2 might be more suitable for networks where the uplink and downlink transmissions have equal powers.
The Three-Node Subnetwork
-------------------------
We assume that there is no interference between different frequency bands. The only interference at a user/BS in a given frequency band is a result of the transmission in the same frequency band from the users and the BSs in other cells. As a result, for a given frequency band, the considered cellular network employing the D-TFDD scheme can be divided into three-node subnetworks, where each subnetwork consists of a BS and two users working in the same frequency band that are impaired by ICI coming from the rest of the subnetworks working in the same frequency band, as shown in Fig. \[Fig:sys-mod-final\]. Depending on whether the downlink reception (uplink transmission) in a given frequency band is shared with the downlink reception or the uplink transmission of another user, there can be three types of three-node subnetworks, as shown in Fig. \[Fig:sys-mod-final\]. Type 1 is when both users perform downlink receptions in the given frequency band. Type 2 is when both users perform uplink transmissions in the given frequency band. And Type 3 is when one of the users performs uplink transmission and the other user performs downlink reception in the given frequency band. Note that the three types of three-node subnetworks differ only in the direction of the transmission on the BS-U1 and BS-U2 channels.
Now, in order for the D-TFDD to be a distributed duplexing scheme, each of the three-node subnetworks must perform D-TFDD independently from the rest of the subnetworks in the cellular network. As a result, without loss of generality, the equivalent system model that needs to be investigated is comprised of a BS communicating with U1 and U2 in different time slots but in the same frequency band, where receivers are impaired by ICI, as shown in Fig. \[Fig:sys-mod-final\].
Inter-Cell Interference
-----------------------
The receiving nodes of a given three-node subnetwork are impaired by interference from all other nodes in the network that transmit on the frequency band used for reception at the BS, and/or U1, and/or U2, also referred to as ICI, see Fig. \[Fig:sys-mod-final\]. Let the power of the ICI at the receiving nodes on the BS-U1 and BS-U2 channels in time slot[^5] $t$ be denoted[^6] by $\gamma_{I1}(t)$ and $\gamma_{I2}(t)$, respectively. Then, we can obtain $\gamma_{I1}(t)$ and $\gamma_{I2}(t)$ as[^7] $$\begin{aligned}
\gamma_{I1}(t)&= \sum_{k \in \mathcal{K}} P_k \gamma_{k1}(t) ,\label{eq_Ic1}\\
\gamma_{I2}(t)&= \sum_{k \in \mathcal{K}} P_k \gamma_{k2}(t) ,\label{eq_Ic2}\end{aligned}$$ where $\mathcal{K}$ is the set of interfering nodes, $P_k$ is the power of interfering node $k$, and $\gamma_{k1}(t)$ and $\gamma_{k2}(t)$ are the square of the channel gains between interfering node $k$ and the receiver on the BS-U1 channel, and interfering node $k$ and the receiver one the BS-U2 channel, in time slot $t$, respectively.
Inter-Cell Interference Estimation Overhead {#Estimation_discuss}
-------------------------------------------
A network comprised of $K$ HD nodes, all operating in the same frequency band, requires at least $K$ estimation periods in order for the ICI to be estimated at all $K$ nodes. To see this, note that a HD node can either receive or transmit in a given frequency band. As a result, in order for a HD node to estimate the interference from the remaining $K-1$ HD nodes, an estimation period must be dedicated for this purpose in which the considered HD node receives and the rest of the $K-1$ HD transmit. Since this process has to be repeated for each of the $K$ HD nodes, it follows that a network comprised of $K$ HD nodes, all operating in the same frequency band, must dedicate $K$ time periods for ICI estimation at the $K$ HD nodes. Hence, ICI estimation at $K$ HD nodes entails an overhead of $K$ estimation periods. In addition, since the transmission schedule of the different HD nodes is not known in advance, the estimated ICI may differ significantly than the real one, which means that the overhead of $K$ time periods is a lower bound of the actual number of time periods needed for estimation of the actual ICI. In fact, this is a key point. The only realistic way to have the transmission schedule of the HD nodes known in advance is to have a central controller that gathers all the channels, makes a scheduling decision for each link in each time slot and forwards that decision to the nodes. This is not feasible in current systems and will likely not be feasible in future systems as long as the coherence time equals a time slot during which the CSI acquisition, the transmission of the scheduling decisions, and the actual transmission of data need to take place.
The overhead needed for ICI estimation requires resources that may prohibit ICI estimation in practice. As a result, in this paper, we investigate the practical case without ICI knowledge, and propose a distributed D-TFDD scheme for this practical scenario. In addition, in order to obtain an upper bound on the performance of the D-TFDD in terms of outage probability and throughput rate for unknown ICI, we will also investigate the case where the ICI is known at the nodes. Consequently, we will propose distributed D-TFDD schemes for the cases with and without ICI knowledge, and show that the proposed distributed D-TFDD scheme without ICI knowledge has performance which is close to its upper bound achieved when the ICI is known.
Channel Model
-------------
In a given subnetwork, we assume that the BS-U1 and BS-U2 channels are complex-valued additive white Gaussian noise (AWGN) channels impaired by slow fading and ICI. Next, we assume that the transmission time is divided into $ T\to\infty $ time slots. Furthermore, we assume that the fading is constant during one time slot and changes from one time slot to the next. In time slot $t$, let the complex-valued fading gains of the BS-U1 and BS-U2 channels be denoted by $h_{1}(t)$ and $h_{2}(t)$, respectively. Moreover, let the variances of the complex-valued AWGNs at receiving nodes of the the BS-U1 and BS-U2 channels be denoted by $\sigma_1^2$ and $\sigma_2^2$, respectively[^8]. For convenience, we define normalized magnitude-squared fading gains of the BS-U1 and BS-U2 channels as $\gamma_{1}(t)=|h_{1}(t)|^2/\sigma_1^2$ and $\gamma_{2}(t)=|h_{2}(t)|^2/\sigma_2^2$, respectively. Furthermore, let the transmit powers of the transmit nodes on the BS-U1 and BS-U2 channels in time slot $t$ be denoted by $P_1$ and $P_2$, $\forall t$, respectively.
Using the above notation, and taking into account the AWGNs and the ICIs given by (\[eq\_Ic1\]) and (\[eq\_Ic2\]), the capacities of the BS-U1 and BS-U2 channels in time slot $t$, denoted by $C_{1}(t)$ and $C_{2}(t)$, respectively, are obtained as $$\begin{aligned}
C_{1}(t)& \hspace{-0.75mm} = \hspace{-0.75mm} {\log _2}\hspace{-0.75mm} \left( { \hspace{-1mm}1 \hspace{-1mm}+ \hspace{-1mm}\frac{{P_1}{\gamma _{1}}(t)}{1+\gamma _{I1}(t)}} \right)
\hspace{-1.5mm},\hspace{-0.5mm}\label{eq_c1}\\
C_{2}(t)& \hspace{-0.75mm} = \hspace{-0.75mm} {\log _2}\hspace{-0.75mm} \left( { \hspace{-1mm}1 \hspace{-1mm}+ \hspace{-1mm}\frac{{P_2}{\gamma _{2}}(t)}{1+\gamma _{I2}(t)}} \right)
\hspace{-1.5mm}.\hspace{-0.5mm}\label{eq_c2}\end{aligned}$$
Discrete-Rate Transmission
--------------------------
We assume that the transmit nodes on the BS-U1 or the BS-U2 channels transmit their codewords with rates which are selected from discrete finite sets of data rates, denoted by $\mathcal{R}_1 = \{R_1^1,R_1^2,...,R_1^M\}$ and $\mathcal{R}_2 = \{R_2^1,R_2^2,...,R_2^L\}$, respectively, where $M$ and $L$ denote the total number of non-zero data rates available for transmission at the transmit nodes on the BS-U1 and BS-U2 channels, respectively. This allows us to have a transmission model used in practice which also converges to continuous transmission rates model when $M\to\infty$ and $L\to\infty$, and to the single fixed-rate transmission model when $M=1$ and $L=1$.
D-TFDD For Known ICI {#Sec-FR}
====================
In this section, we assume that the ICI is known at the nodes at the start of each time slot. Although this assumption is not practical as discussed in Sec. \[Estimation\_discuss\], it will enable us to obtain an upper bound on the practical D-TFDD without ICI knowledge at the nodes.
BS-U1 and BS-U2 Throughput Region {#subSec-PR}
-----------------------------------
In a given time slot, depending on whether we are communicating on the BS-U1 and BS-U2 channels, the considered three-node subnetwork, shown in Fig. \[Fig:sys-mod-final\], can be in one of the following three states\
*State 0:* No transmission occurs on both BS-U1 and BS-U2 channels.\
*State 1:* Channel BS-U1 is selected for transmission and channel BS-U2 is inactive/silent.\
*State 2:* Channel BS-U2 is selected for transmission and channel BS-U1 is inactive/silent.
In State 1, the transmitting node on the BS-U1 channel can choose to transmit with any rate in the set $\mathcal{R}_1$. Similarly, in State 2, the transmitting node on the BS-U2 channel can choose to transmit with any rate in the set $\mathcal{R}_2$. In order to model these states for time slot $t$, we introduce the binary variables $q_1^m(t)$, $m=1,2...,M$ and $q_2^l(t)$, for $l=1,...,L$, defined as $$\begin{aligned}
q_1^m(t) =& \left\{
\begin{array}{ll}
1& \textrm{if channel BS-U1 is selected for the transmission of a codeword with rate $R_1^m$} \\
&\textrm{ and power $P_1$ in time slot $t$}\\
0& \textrm{otherwise},
\end{array}\label{eq_q1}
\right.\\
q_2^l(t) =& \left\{
\begin{array}{ll}
1& \textrm{if channel BS-U2 is selected for the transmission of a codeword with rate $R_2^m$} \\
&\textrm{ and power $P_2$ in time slot $t$}\\
0& \textrm{otherwise},
\end{array}
\right.\label{eq_q2}\end{aligned}$$ respectively. In addition, since the considered network can be in one and only one state in time slot $t$, the following has to hold $$\begin{aligned}
\label{eq_PR_0}
\sum_{m=1}^M q_1^m(t)+\sum_{l=1}^L q_2^l(t)\in\{0,1\},\end{aligned}$$ where if $\sum_{m=1}^M q_1^m(t)+\sum_{l=1}^L q_2^l(t)=0$ holds, then both the BS-U1 and the BS-U2 channels are inactive in time slot $t$. Condition (\[eq\_PR\_0\]) results from the HD constraint of the BS, i.e., the BS can either receive or transmit in a given time slot on the same frequency band.
Since the available transmission rates are discrete, outages can occur. An outage occurs if the data rate of the transmitted codeword is larger than the capacity of the underlying channel. To model the outages on the BS-U1 and the BS-U2 channels, we introduce the following auxiliary binary variables, $O_1^m(t)$, for $m=1,...,M$, and $O_2^l(t)$, for $l=1,...,L$, respectively, defined as $$\begin{aligned}
O_1^m(t)&=\left\{
\begin{array}{ll}
1 & \textrm{if } C_{1}(t)\geq R_1^m\\
0 & \textrm{if } C_{1}(t)< R_1^m,
\end{array}
\right. \label{eq_PR_fr1}\\
O_2^l(t)&=\left\{
\begin{array}{ll}
1 & \textrm{if } C_{2}(t)\geq R_2^l\\
0 & \textrm{if } C_{2}(t)< R_2^l.
\end{array}
\right. \label{eq_PR_fr2}\end{aligned}$$
Using $O_1^m(t)$, we can obtain that in time slot $t$ a codeword transmitted on the BS-U1 channel with rate $R_1^m$ can be decoded correctly at the receiver if and only if (iff) $q_1^m(t)O_1^m(t)>0$ holds. Similarly, using $O_2^l(t)$, we can obtain that in time slot $t$ a codeword ctransmitted on the BS-U2 channel with rate $R_2^l$ can be decoded correctly at receiver iff $q_2^l(t) O_2^l(t)>0$ holds. Thereby, the average achieved throughputs during $T\to\infty$ time slots on the BS-U1 and BS-U2 channels, denoted by $\bar R_{1}$ and $\bar R_{2} $, respectively, are given by $$\begin{aligned}
\bar R_{1} &=\lim_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{m=1}^M R_1^m q_1^m(t) O_1^m (t),\label{eq_PR_fr3}
\\
\bar R_{2} &=\lim_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{l=1}^L R_2^l q_2^l(t) O_2^l (t).\label{eq_PR_fr4}\end{aligned}$$
The throughput pair $(\bar R_1, \bar R_2)$, defined by (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), for some fixed vectors $[q_1^m(1), q_1^m(2),...,q_1^m(T)]$ and $[q_2^l(1), q_2^l(2),...,q_2^l(T)]$ gives one point on the graph where $\bar R_1$ and $\bar R_2$ are axis. All possible combinations of $[q_1^m(1), q_1^m(2),...,q_1^m(T)]$ and $[q_2^l(1), q_2^l(2),...,q_2^l(T)]$ give a region of points that is bounded by a maximum boundary line of the BS-U1 and BS-U2 throughput region. Our task now is to find the maximum boundary line of this BS-U1 and BS-U2 throughput region, $(\bar R_{1} ,\bar R_{2})$, by selecting the optimal values of $q_1^m(t)$, $q_2^l(t)$, $\forall m,l,t$, respectively.
The maximum boundary line of the BS-U1 and BS-U2 throughput region $(\bar R_{1},\bar R_{2})$, given by (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), can be found from the following maximization problem $$\begin{aligned}
\label{eq_op_PR}
& {\underset{q_1^m(t), q_2^l(t), \;\forall l,m,t.\;}{\textrm{Maximize: }}}\; \mu \bar R_{1}+\big(1 - {\mu} \big) \bar R_{2} \nonumber\\
& {\rm{Subject\;\; to \; :}} \nonumber\\
& \qquad {\rm C1:} \; \ q_1^m(t)\in\{0,1\}, \forall m \nonumber\\
& \qquad {\rm C2:} \; \ q_2^l(t)\in\{0,1\}, \forall l \nonumber\\
& \qquad {\rm C3:}\; \sum_{m=1}^M q_1^m(t)+\sum_{l=1}^L q_2^l(t)\in\{0,1\},\end{aligned}$$ where $\mu$ is a priori given constant which satisfies $0\leq \mu\leq 1$. A specific value of $\mu$ provides one point on the boundary line of the BS-U1 and BS-U2 throughput region[^9] $(\bar R_{1},\bar R_{2})$. By varying $\mu$ from zero to one, the entire boundary line of the BS-U1 and BS-U2 throughput region $(\bar R_{1},\bar R_{2})$ can be obtained. The solution of (\[eq\_op\_PR\]) is given in the following theorem.
\[Theo\_PR\] The optimal state and rate selection variables, $q_1^m(t)$ and $q_2^l(t)$, of the D-TFDD scheme for known ICI that maximize the BS-U1 and BS-U2 throughput region of the considered subnetwork, which are found as the solution of (\[eq\_op\_PR\]), are given as $$\begin{aligned}
\textrm{BS-U1 transmission } \blacktriangleright \;&\ q_1^{m^*}(t)=1, q_1^{m}(t)=0, \; \forall m \ne m^* \textrm{ and } q_2^l(t)=0, \; \forall l \nonumber\\
&\textrm{ if}\; \left[ \Lambda_1^{m^*}(t) \geq \Lambda_2^{l^*}(t) \; \textrm{and}\; \Lambda_1^{m^*}(t)> 0 \right], \nonumber\\
\textrm{BS-U2 transmission } \blacktriangleright\; &q_2^{l^*}(t)=1, q_2^{l}(t)=0, \; \forall l \ne l^* \; \textrm{and}\; q_1^m(t)=0, \; \forall m \nonumber\\
&\textrm{ if}\; \left[ \Lambda_2^{l^*}(t) > \Lambda_1^{m^*}(t) \; \textrm{and}\; \Lambda_2^{l^*}(t) > 0 \right], \nonumber\\
\textrm{Silence } \blacktriangleright \;& q_1^{m}(t)=0, \; \forall m \;\textrm{and} \; q_2^l(t)=0, \; \forall l \nonumber\\
&\textrm{ if}\; \left[ \Lambda_1^{m^*}(t) =0 \; \textrm{and}\; \Lambda_2^{l^*}(t) = 0 \right], \label{eq_scheme_AP_G}\end{aligned}$$ where $\Lambda_1^{m}(t)$, $\Lambda_2^{l}(t)$, $m^*$ and $l^*$ are defined as $$\begin{aligned}
\Lambda_1^{m}(t) &=\mu R_1^{m} O_1^{m} (t), \label{eq_PR_1}\\
\Lambda_2^{l}(t) &=(1-\mu) R_2^{l} O_2^{l} (t) \label{eq_PR_2},\\
m^*&=\arg {\underset{m }{\textrm{max} }}\{\Lambda_1^{m}(t) \},\\
l^*&=\arg{\underset{l }{\textrm{max} }}\{ \Lambda_2^{l}(t) \}.\end{aligned}$$
Please refer to Appendix \[app\_PR\] for the proof.
Note that for the proposed D-TFDD scheme in Theorem \[Theo\_PR\] to operate, the receivers of the BS-U1 and BS-U2 channels need to know $O_1^m (t)$ and $O_2^l (t)$, respectively, $\forall m,l$, at the start of time slot $t$, which requires knowledge of the ICI.
Diversity Gain of the Proposed D-TFDD for Known ICI {#sec_op}
---------------------------------------------------
It is quite interesting to investigate the diversity gain achieved with the D-TFDD scheme for known ICI proposed in Theorem \[Theo\_PR\]. In the literature, the asymptotic outage probability, from which the diversity gain is obtained, is derived assuming only a single available transmission rate at the transmitter, see [@1362898]. Following this convention, in the following, we derive the asymptotic outage probabilities of the BS-U1 and the BS-U2 channels, denoted by $P_{\rm out}$, achieved with the D-TFDD scheme for known ICI proposed in Theorem \[Theo\_PR\] for $\mu=\frac{1}{2}$, $M=L=1$, $ P_1= P_2$, and $R_1^1=R_2^1=R_0$. For simplicity, we only investigate the case of Rayleigh fading, and also assume that the BS-U1 and BS-U2 channels are affected by independent and identically distributed (i.i.d) fading.
\[Theo\_PR\_Outage\_G\] The asymptotic outage probability of the D-TFDD scheme for known ICI proposed in Theorem \[Theo\_PR\] for the case of Rayleigh fading and when $\mu=\frac{1}{2}$, $M=L=1$, $ P_1= P_2=P$, and $R_1^1=R_2^1=R_0$ hold, is given by $$\begin{aligned}
\label{Out_U6GG}
P_{\rm out}& \to \frac{ \gamma_{\rm th}^2 \hat \Omega_I }{\Omega_0^2}, \; \textrm{as} \;{P\to\infty},\end{aligned}$$ where $\gamma_{\rm th}=\frac{2^{R_0}-1}{P} $, $\Omega_0=E\left \{\frac{|h_{1}(t)|^2}{\sigma_1^2 }\right \}=E\left\{\frac{|h_{2}(t)|^2}{\sigma_2^2}\right\}$, and $\hat \Omega_I=E \Big \{ (1+\gamma _{I1}(t)) (1+\gamma _{I2}(t)) \Big \}$.
As can be seen from (\[Out\_U6GG\]), the outage probability $ P_{\rm out}$ has a diversity gain of two.
Please refer to Appendix \[app\_PR\_Outage\_G\] for the proof.
Note that existing D-TDD and D-FDD schemes achieve a diversity gain of one, which leads to the conclusion that the proposed D-TFDD scheme doubles the diversity gain compared to existing duplexing schemes, which in turn leads to very large performance gains, cf. Sec. \[Sec-Num\].
D-TFDD For Unknown ICI {#subSec-PRL}
======================
The D-TFDD scheme proposed in Section \[Sec-FR\] requires the receivers of the BS-U1 and BS-U2 channels to know $O_1^m (t)$ and $O_2^l (t)$, respectively, $\forall m,l$, at the start of time slot $t$, for $t=1, 2, ... T$, which requires ICI knowledge. However, as discussed in Sec. \[Estimation\_discuss\], the estimation of the ICI entails huge cost for a cellular network comprised of $K$ HD nodes which may not be practical. Motivated by this problem, in the following, we propose a D-TFDD scheme where the nodes do not have knowledge of the ICI, and as a result, the network nodes do not have to waste huge resources for estimating the ICI. We only assume that the CSI of the BS-U1 and BS-U2 channels are known at the BS, i.e., we assume local CSI knowledge at the BS. Specifically, the BS knows ${\gamma _{1}}(t)$ and ${\gamma _{2}}(t)$ at the start of time slot $t$, which can be acquired by allocating two estimation periods; one for the BS-U1 channel and the other for the BS-U2 channel, which is a huge improvement compared to the $K$ estimation periods that need to be allocated when the ICI needs to be estimated, see Sec. \[Estimation\_discuss\] .
Proposed D-TFDD For Unknown ICI {#SubSec_PRL_Proposed}
--------------------------------
The throughput region of the BS-U1 and BS-U2 channels employing the D-TFDD scheme for unknown ICI is also given by (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), where $O_1^m(t)$ and $O_2^l(t)$ are defined in (\[eq\_PR\_fr1\]) and (\[eq\_PR\_fr2\]), respectively. The only difference now is that $q_1^m(t)$ and $q_2^l(t)$ in (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]) are different when the D-TFDD for unknown ICI is applied. The optimal $q_1^m(t)$ and $q_2^l(t)$, which maximize the throughput region, defined by (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), of the D-TFDD scheme for unknown ICI are given in the following.
For the case when the ICI is unknown at the nodes, first we define $$\begin{aligned}
& m^* \buildrel \Delta \over = \arg {\underset{m }{\textrm{max} }} \{ R_1^m O_{1,e}^m (t)\},\label{eq_AP_S134}\\
& l^* \buildrel \Delta \over =\arg {\underset{l }{\textrm{max} }} \{ R_2^l O_{2,e}^l (t)\}\label{eq_AP_S135},\end{aligned}$$ where $ O_{1,e}^m (t)$ and $O_{2,e}^l (t)$ are defined as $$\begin{aligned}
O_{1,e}^m(t)&=\left\{
\begin{array}{ll}
1 & \textrm{if } C_{1}^e(t)\geq R_1^m\\
0 & \textrm{if } C_{1}^e(t)< R_1^m,
\end{array}
\right. \label{eq_PR_fr11l}\\
O_{2,e}^l(t)&=\left\{
\begin{array}{ll}
1 & \textrm{if } C_{2}^e(t)\geq R_2^l\\
0 & \textrm{if } C_{2}^e(t)< R_2^l.
\end{array}
\right. \label{eq_PR_fr22l}\end{aligned}$$ The variables, $C_{1}^e(t)$ and $C_{2}^e(t)$, used in (\[eq\_PR\_fr11l\]) and (\[eq\_PR\_fr22l\]), are defined as $$\begin{aligned}
C_{1}^e(t)&={\log _2}\left( {1 + \frac{{P_1}{\gamma _{1}}(t)}{1+\gamma_{I1}^e(t)}} \right)
,\label{eq_c11}\\
C_{2}^e(t)&={\log _2}\left( {1 + \frac{{P_2}{\gamma _{2}}(t)}{1+\gamma_{I2}^e(t)}} \right)
,\label{eq_c22}\end{aligned}$$ where $\gamma_{I1}^e(t)$ and $\gamma_{I2}^e(t)$ are given in Proposition \[Theo\_PR\_L0cal\] in the following, and they can be thought of as estimates of the ICI in time slot $t$.
The optimal $q_1^m(t)$ and $q_2^l(t)$, which maximize the throughput region, defined by (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), of the D-TFDD for unknown ICI are as follows $$\begin{aligned}
\textrm{BS-U1 transmission } \blacktriangleright &\; q_1^{m^*}(t)=1, q_1^{m}(t)=0, \; \forall m \ne m^* \textrm{ and } q_2^l(t)=0, \; \forall l, \nonumber\\
&\textrm{ if}\; \left[ \Lambda_1(t) \geq \Lambda_2(t) \; \textrm{and}\; \Lambda_1(t) > 0 \right], \nonumber\\
\textrm{BS-U2 transmission } \blacktriangleright &\;q_2^{l^*}(t)=1, q_2^{l}(t)=0, ; \forall l \ne l^*, \textrm{ and } q_1^m(t)=0, \; \forall m, \nonumber\\
&\textrm{ if}\; \left[ \Lambda_2(t) \geq \Lambda_1(t) \; \textrm{and}\; \Lambda_2(t) > 0 \right], \nonumber\\
\textrm{Silence } \blacktriangleright &\; q_1^{m}(t)=0, \forall m \;\textrm{and} \; q_2^l(t)=0, \; \; \forall l, \nonumber\\
&\textrm{ if}\; \left[ \Lambda_1(t) =0 \; \textrm{and}\; \Lambda_2(t) = 0 \right], \label{eq_scheme_AP}\end{aligned}$$ where $\Lambda_1(t)$ and $\Lambda_2(t)$ are given by $$\begin{aligned}
\Lambda_1(t) &= \mu C_{1}^e(t),\label{eq_AP_S1111}\\
\Lambda_2(t) &= (1-\mu) C_{2}^e(t),\label{eq_AP_S2222}\end{aligned}$$ and $m^*$ and $l^*$ are given by (\[eq\_AP\_S134\]) and (\[eq\_AP\_S135\]), respectively. In (\[eq\_AP\_S1111\]) and (\[eq\_AP\_S2222\]), $\mu$ is a priori given constant which satisfies $0\leq \mu\leq 1$. By varying $\mu$ from zero to one, the entire boundary line of the BS-U1 and BS-U2 throughput region can be obtained.
\[Theo\_PR\_L0cal\] The variables $\gamma_{I1}^e(t)$ and $\gamma_{I2}^e(t)$, found in the expressions in (\[eq\_c11\]) and (\[eq\_c22\]), which maximize the BS-U1 and BS-U2 throughput region of the D-TFDD scheme for unknown ICI proposed in (\[eq\_scheme\_AP\]) are found as follow $$\begin{aligned}
&\gamma_{I1}^e(t+1)=\gamma_{I1}^e(t) -\delta_1(t) \Phi_1(t),\label{eq_z1_Rsr1}\\
&\gamma_{I2}^e(t+1)=\gamma_{I2}^e(t) -\delta_2(t) \Phi_2(t),\label{eq_z1_Rsr2}\end{aligned}$$ where $\delta_{k}( t)$, for $k\in\{1,2\}$, can be some properly chosen monotonically decaying function of $t$ with $\delta_{k}( 1)<1$, such as $\frac{1}{2t}$. Furthermore, $\Phi_1(t)$ and $\Phi_2(t)$ in (\[eq\_z1\_Rsr1\]) and (\[eq\_z1\_Rsr2\]) are obtained as $$\begin{aligned}
\Phi_1(t)&\hspace{-0.5mm}=\hspace{-0.5mm}\frac{t\hspace{-0.5mm}-\hspace{-0.5mm}1}{t} \Phi_1(t\hspace{-0.5mm}-\hspace{-0.5mm}1) \hspace{-0.5mm}-\hspace{-0.5mm}\frac{1}{t} \frac{ P_1 \gamma_{1}(t) \delta_1^{m^*}(t) q_1^{m^*}(t)\bigg ( \hspace{-0.5mm} 2\chi_1^e(t) \big [ O_{1,e}^{m^*}(t)-O_{1}^{m^*}(t) \big]-\mu R_1^{m^*} \hspace{-0.5mm} \bigg)}{\ln(2) \Big (1+\gamma_{I1}^e(t)+P_1 \gamma _{1}(t)\Big) \Big (1+\gamma_{I1}^e(t)\Big)},\label{eq_zeta_Rsr1131}\\
\Phi_2(t)&\hspace{-0.5mm}=\hspace{-0.5mm}\frac{t\hspace{-0.5mm}-\hspace{-1.5mm}1}{t} \Phi_2(t\hspace{-1mm}-\hspace{-0.5mm}1) \hspace{-0.5mm}-\hspace{-0.5mm}\frac{1}{t} \frac{ P_2 \gamma_{2}(t) \delta_2^{l^*}(t) q_2^{l^*}(t) \hspace{-0mm} \bigg ( \hspace{-1mm} 2 \chi_2^e(t) \big [ O_{2,e}^{l^*}(t)\hspace{-0.5mm}-\hspace{-1mm} O_{2}^{l^*}(t) \big] -(1\hspace{-1mm}-\hspace{-1mm}\mu) R_2^{l^*} \bigg)}{\ln(2) \Big (1+\gamma_{I2}^e(t)+P_2 \gamma _{2}(t)\Big) \Big (1+\gamma_{I2}^e(t)\Big)},\label{eq_zeta_Rsr1231}\\end{aligned}$$ respectively, where $\delta_1^{m^*}(t)$ and $\delta_2^{l^*}(t)$, are defined as $$\begin{aligned}
\label{eq_2elta_fr1}
\delta_1^{m^*}(t)=\left\{
\begin{array}{ll}
1 & \textrm{if } \left(R_1^{m^*}-C_{1}^e(t-1) \right) \left(R_1^{m^*}-C_{1}^e(t)\right) \leq 0\\
0 & \textrm{if } \left(R_1^{m^*}-C_{1}^e(t-1) \right) \left(R_1^{m^*}- C_{1}^e(t)\right)>0,
\end{array}
\right.\end{aligned}$$ and $$\begin{aligned}
\label{eq_2elta_fr2}
\delta_2^{l^*}(t)=\left\{
\begin{array}{ll}
1 & \textrm{if } \left(R_2^{l^*}-C_{2}^e(t-1) \right) \left(R_2^{l^*}-C_{2}^e(t)\right) \leq 0\\
0 & \textrm{if } \left(R_2^{l^*}-C_{2}^e(t-1) \right) \left(R_2^{l^*}- C_{2}^e(t)\right)>0,
\end{array}
\right.\end{aligned}$$ respectively. On the other hand, the variables $\chi_1^e(t)$ and $\chi_2^e(t)$ in (\[eq\_zeta\_Rsr1131\]) and (\[eq\_zeta\_Rsr1231\]) are calculated as $$\begin{aligned}
&\chi_k^e(t+1)=\chi_k^e(t) +\delta_k^\chi( t) \left[ {\bar \epsilon}_k(t) - \epsilon\right]^+, \, k\in\{1,2\},\label{eq_z1_Rsr1234}\end{aligned}$$ where $\delta_k^\chi( t)$ for $k\in\{1,2\}$ can be some properly chosen monotonically decaying function of $t$ with $\delta_k^\chi( 1)<1$, such as $\frac{1}{2t}$, $\epsilon$ is a small constant which can be configured at the system level, and $[.]^+$ denotes only positive values. Moreover, in (\[eq\_z1\_Rsr1234\]), ${\bar \epsilon}_k(t)$, for $k\in\{1,2\}$, are obtained as $$\begin{aligned}
{\bar \epsilon}_1(t)&=\frac{t-1}{t} {\bar \epsilon}_U(t-1) +\frac{1}{t} q_1^{m^*}(t) \bigg( O_{1}^{m^*}(t)-O_{1,e}^{m^*}(t) \bigg )^2\hspace{-2mm},\label{eq_zeta_Rsr1122}\\
{\bar \epsilon}_2(t)&=\frac{t-1}{t} {\bar \epsilon}_D(t-1) +\frac{1}{t} q_2^{l^*}(t) \bigg( O_{2}^{l^*}(t)-O_{2,e}^{l^*}(t) \bigg )^2\hspace{-2mm}.\label{eq_zeta_Rsr1133}\end{aligned}$$ Finally, we note that, ${\bar \epsilon}_k(0)$, $\Phi_k(0)$, $\chi_k^e(0)$, and $\gamma_{Ik}^e(0)$, $k \in\{1,2\}$, found in (\[eq\_z1\_Rsr1\])-(\[eq\_zeta\_Rsr1133\]), are initialized to zero. The variables $\Phi_1(t)$, $\Phi_2(t)$, $\chi_1^e(t+1)$, $\chi_2^e(t+1)$, $\delta_1^{m^*}(t)$, $\delta_2^{l^*}(t)$, ${\bar \epsilon}_1(t)$ and ${\bar \epsilon}_2(t)$ are auxiliary variables used for real-time estimation of the ICIs $\gamma_{I1}^e(t)$ and $\gamma_{I2}^e(t)$.
The D-TFDD scheme for unknown ICI is provided in an algorithmic form in Algorithm \[Lyp\_algorithm\].
Initiate ${\bar \epsilon}_k(0)$, $\Phi_k(0)$, $\chi_k^e(1)$, and $\gamma_{Ik}^e(1)$, $k \in\{1,2\}$ to zero \[lyp\_proc\] compute $C_{1}^e(t)$ and $C_{2}^e(t)$ with (\[eq\_c11\]) and (\[eq\_c22\]); compute $O_{1,e}^m(t)$ and $O_{2,e}^l(t)$ with (\[eq\_PR\_fr11l\]) and (\[eq\_PR\_fr22l\]); compute $m^* $ and $l^* $ with (\[eq\_AP\_S134\]) and (\[eq\_AP\_S135\]); compute $\Lambda_1(t)$ and $\Lambda_2(t)$ with (\[eq\_AP\_S1111\]) and (\[eq\_AP\_S2222\]); compute $q_1^{m},\forall m$ and $q_2^l(t),\forall l$ with (\[eq\_scheme\_AP\]); update ${\bar \epsilon}_1(t)$ and ${\bar \epsilon}_2(t)$ with (\[eq\_zeta\_Rsr1122\]) and (\[eq\_zeta\_Rsr1133\]); compute $\delta_1^{m^*}(t)$ and $\delta_2^{l^*}(t)$ with (\[eq\_2elta\_fr1\]) and (\[eq\_2elta\_fr2\]); update $\Phi_1(t)$ and $\Phi_2(t)$ with (\[eq\_zeta\_Rsr1131\]) and (\[eq\_zeta\_Rsr1231\]); update $\chi_1^e(t+1)$ and $\chi_2^e(t+1)$ with (\[eq\_z1\_Rsr1234\]); update $\gamma_{I1}^e(t+1)$ and $\gamma_{I2}^e(t+1)$ with (\[eq\_z1\_Rsr1\]) and (\[eq\_z1\_Rsr2\]);
Please refer to Appendix \[app\_PR\_Local\] for the proof.
Diversity Gain of the Proposed D-TFDD for Unknown ICI {#subSec-OUT_L}
------------------------------------------------------
It is quite interesting to investigate the diversity gain that the D-TFDD scheme for unknown ICI, proposed in Section \[SubSec\_PRL\_Proposed\], achieves. In the following, we derive the asymptotic outage probabilities of the BS-U1 and the BS-U2 channels, achieved with the D-TFDD scheme for unknown ICI proposed in Sec. \[SubSec\_PRL\_Proposed\] for the case when $\mu=\frac{1}{2}$, $M=L=1$, and $R_1^1=R_2^1=R_0$.
\[Theo\_PR\_Outage\_L\] The asymptotic outage probability of the D-TFDD scheme for unknown ICI, proposed in Section \[SubSec\_PRL\_Proposed\], for the case of Rayleigh fading and when $\mu=\frac{1}{2}$, $M=L=1$, $ P_1= P_2=P$, and $R_1^1=R_2^1=R_0$ hold, is given by $$\begin{aligned}
\label{eq_out1asl}
P_{\rm out}& \to \frac{\gamma_{\rm th}^2}{\Omega_0^2} \left[ \hat \Omega_{I1} + \hat \Omega_{I2} + \hat \Omega_{IS} \right], \; \textrm{as} \;P \to \infty,\end{aligned}$$ where $\gamma_{\rm th}=\frac{2^{R_0}-1}{P} $, $\Omega_0=E\left\{\frac{|h_{1}(t)|^2}{\sigma_1^2 (1+\gamma_{I1})}\right\}=E\left\{\frac{|h_{2}(t)|^2}{\sigma_2^2 (1+\gamma_{I2})}\right\}$, $$\begin{aligned}
&{\hat \Omega_{I1}} = \Omega_I \left(\frac{1+\gamma _{I2}}{1+\gamma _{I1}} \right) \left( \sum_{n=0}^{K-1} [1+(n+1) \Omega_I ]e^{-{\frac{\gamma _{I1}}{\Omega_I}}} \sum_{t=0}^{n} \frac{({\frac{\gamma _{I1}}{\Omega_I}})^i}{t!} \right), \label{Otage_U_final}\\
&{\hat \Omega_{I2}} = \Omega_I \left(\frac{1+\gamma _{I1}}{1+\gamma _{I2}} \right) \left( \sum_{n=0}^{K-1} [1+(n+1) \Omega_I ]e^{-{\frac{\gamma _{I2}}{\Omega_I}}} \sum_{t=0}^{n} \frac{({\frac{\gamma _{I2}}{\Omega_I}})^i}{t!} \right), \label{Otage_D_final}\\
&{\hat \Omega_{IS}} = ({1+\gamma _{I1}}) ({1+\gamma _{I2}}),\label{Otage_S_final} \\
&\Omega_I=E\{\gamma _{I1}(t)\}=E\{\gamma _{I2}(t)\}.\end{aligned}$$ In the above expression, $\gamma_{I1}$ and $\gamma_{I2}$ are constant values obtained as $\mathop {\lim }\limits_{t \to \infty } \gamma_{I1}^e(t) =\gamma_{I1}$ and $\mathop {\lim }\limits_{t \to \infty } \gamma_{I2}^e(t) =\gamma_{I2}$. It is clear from (\[eq\_out1asl\]) that $ P_{\rm out}$ has a diversity gain of two.
Please refer to Appendix \[app\_PR\_Outage\_L\] for the proof.
The result in Theorem \[Theo\_PR\_Outage\_L\] shows that the D-TFDD scheme for unknown ICI proposed in Sec. \[SubSec\_PRL\_Proposed\] achieves double the diversity gain compared to existing duplexing schemes, which leads to very large performance gains. Moreover, Theorem \[Theo\_PR\_Outage\_L\] shows that doubling of the diversity gain on both the BS-U1 and BS-U2 channels is achievable even when there is no ICI knowledge at the nodes, which is a very interesting result that shows the superior performance of the D-TFDD scheme for unknown ICI proposed in Sec. \[subSec-PR\] compared to existing duplexing schemes.
Simulation And Numerical Results {#Sec-Num}
================================
In this section, we evaluate the performance of the proposed D-TFDD scheme for unknown ICI and its upper bound, the proposed D-TFDD scheme for known ICI. Then, we compare its performance to the performance achieved with a static-TDD scheme and to the state-of-the-art D-TDD and D-FDD schemes. To this end, we first introduce the benchmark schemes and then present the numerical results.
Benchmark Schemes
-----------------
### Static-TDD
In the static-TDD scheme, see [@holma2011lte], the BS receives and transmits in prefixed time slots. Assuming single transmission rates at the transmitting nodes of the BS-U1 and BS-U2 channels, and assuming that that the fractions of the total number of time slots, $T$, allocated on the BS-U1 and BS-U2 channels are $\mu$ and $1-\mu$, respectively, e.g., channel BS-U1 is active in the first $\mu T$ and channel BS-U2 is active in following $(1-\mu) T$ time slots, the BS-U1 and BS-U2 throughput during $T\to\infty$ time slots, is given by $$\begin{aligned}
\bar R_{k}&=\lim_{T\to\infty}\frac {\mu_k}{T} \sum_{t=1}^T O_k^1(t) R_k^1, k\in\{1,2\},\label{eq_ben_1}\end{aligned}$$ and the outage probability is given by $$\begin{aligned}
P_{\rm out}&=1- \lim_{T\to\infty}\left ( \frac {1}{T} \sum_{t=1}^{\mu T} O_1^1(t) +\sum_{t=\mu T+1}^{T} O_2^1(t) \right ),\label{eq_ben_123}\end{aligned}$$ where $O_k^1(t)$, $k\in\{1,2\}$, is defined as $$\begin{aligned}
O_k^1(t)&=\left\{
\begin{array}{ll}
1 & \textrm{if } \log_2\bigg(1+\frac{ P_k}{\mu_k} \gamma_{k}(t)\bigg)\geq R_k^1\\
0 & \textrm{if } \log_2\bigg(1+\frac{ P_k}{\mu_k} \gamma_{k}(t)\bigg)< R_k^1.
\end{array}\label{eq_ben_fr1}
\right. \end{aligned}$$
### D-TDD Scheme
The distributed D-TDD scheme proposed in [@4524858] is considered as a benchmark for comparison. We note that this scheme requires full knowledge of the ICI, which, as argued in Section \[Estimation\_discuss\], is not practical. In addition, we note that this scheme without ICI knowledge transforms to the static-TDD. Hence, the practical distributed TDD scheme is the static-TDD since it does not need ICI knowledge. The distributed D-TDD scheme in [@4524858] can serve only as an upper bound to the practical static-TDD.
Note that D-TDD and D-FDD schemes only different in how they share the time and frequency resources, but achieve the same performance. Therefore, for comparison purposes, we can use either of them. In the following, we choose the D-TDD scheme.
Numerical Results
-----------------
All of the results presented in this section have been performed by numerical evaluation of the derived results and are confirmed by Monte Carlo simulations. Moreover, Rayleigh fading for the BS-U1 and BS-U2 channels, and Chi-square distribution for the ICI at the receiving nodes of the BS-U1 and BS-U2 channels are assumed. In all the numerical examples in Figs. \[Outage\_Discrete\_1\]-\[discrete\_rate\_SNR\], we assume $M = L$ and $R_1^k=R_2^k=kR$, for $k = 1, 2, ..., M$, where $R$ is defined differently depending on the corresponding example. The signal to interference plus noise ratio (SINR) is defined as the ratio of the average received signal power to interference power plus the noise power.
### Constraint on the Average Transmit Power
In the numerical examples, we select the fixed powers $P_1$ and $P_2$ such that the following long-term power constraints hold $$\begin{aligned}
\lim_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{m=1}^M q_1^m(t)P_1 \leq \bar P_1\;\textrm{and}\; \lim_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{l=1}^L q_2^l(t)P_2 \leq \bar P_2.\label{eq_PR_p2}\end{aligned}$$ This enables all duplexing schemes to use identical average transmit powers, and thereby enable fair comparison between the different schemes.
### Outage Probability {#Sec-Out}
In Fig. \[Outage\_Discrete\_1\], we illustrate the outage probabilities of the proposed D-TFDD schemes for unknown ICI and its upper bound, the D-TFDD for known ICI, as well as the benchmark schemes as a function of the SINR for $M =1$, $\mu=\frac{1}{2}$, where $R$ is set to $R = 1$ bits/symb. As predicted by Theorems \[Theo\_PR\_Outage\_G\] and \[Theo\_PR\_Outage\_L\], Fig. \[Outage\_Discrete\_1\] shows that the proposed D-TFDD schemes for unknown and known ICI achieve double the diversity gain compared to the benchmark schemes. Intuitively, the doubling of the diversity gain occurs since the proposed D-TFDD schemes can select for transmission between two independent channels, in each time slot, compared to the existing D-TDD and D-FDD system, which can select between two dependent channels, in each time slot. The doubling of the diversity gain leads to very large performance gains in terms of SINR. For example, SINR gains of $10 \; \textrm{dB}$ and $15$ dB can be achieved for outage probabilities of $10^{-2}$ and $10^{-3}$, respectively. On the other hand, the proposed D-TFDD schemes for unknown ICI has around 3 dB penalty loss compared its upper bound, the proposed D-TFDD schemes for known ICI. This example shows the large performance gains of the proposed D-TFDD scheme with unknown ICI compared to existing D-TDD and/or static-TDD schemes.
![Outage probability of the proposed D-TFDD schemes. Local-CSI and Full-CSI labels highlight that the corresponding schemes are without and with ICI knowledge, respectively.[]{data-label="Outage_Discrete_1"}](Outage_Discrete_1.eps){width="6in"}
### Throughput Region {#Sec-RTR}
For the example in Fig. \[RateRegion\_700m\] users have ominidirectional antenna with unity gain and the BS has a directional antenna with gain of $16$ dBi. The power at the users is set to $24$ dBm and the power at BS is set to $46$ dBm, and we use the proposed scheme with Type 1 and Type 2 for this example. The distances between U1 and BS, as well as BS and U2, are assumed to be fixed and set to $700$m. The noise figure of BS and U2 are set to $2$ dB and $7$ dB, respectively. The above parameters reflect the parameters used in practice. In addition, the mean power of the channel gains of the BS-U1 and BS-U2 channels are calculated using the standard path-loss model as [@6847175; @7105650; @7463025] $$\begin{aligned}
E\{|h_k(t)|^2\} = \left(\frac{c}{{4\pi {f_c}}}\right)^2d_k^{ - \beta }\ \textrm{, for } k\in\{1,2\},\end{aligned}$$ where $c$ is the speed of light, $f_c$ is the carrier frequency, $d_k$ is the distance between the transmitter and the receiver of link $k$, and $\beta$ is the path loss exponent. Moreover, the carrier frequency is set to $f_c=1.9$ GHz, and we assume $\beta=3.6$ for the BS-U1 and BS-U2 channels.
In Fig. \[RateRegion\_700m\], we show the throughput region achieved with the proposed D-TFDD schemes for unknown and known ICI with $M=1$, for two different scenarios, one with high interference, SINR=10 dB, and the other one with low interference, SINR=20 dB. Furthermore, we show the throughput regions achieved with the benchmark schemes. For the proposed and the benchmark schemes the value of $R$ is optimized numerically for a given $\mu$ such that the throughput is maximized. As can be seen from Fig. \[RateRegion\_700m\], the proposed D-TDD scheme without ICI knowledge achieves almost the exact throughput region as its upper bound achieved with D-TFDD with full ICI knowledge for SINR=20 dB. Also, in the relatively high interference region, i.e., SINR=10 dB, the proposed D-TFDD scheme without ICI knowledge achieves a throughput region which is very close to its upper bound achieved with the D-TFDD scheme with ICI knowledge. On the other hand, the throughput that the proposed D-TFDD scheme for unknown ICI achieves is close or higher than the throughput achieved with the benchmark D-TDD scheme with ICI knowledge, which is an interesting result since the proposed scheme without ICI knowledge wastes only two time slots compared to the $K$ time slots that the D-TDD and D-FDD schemes with ICI knowledge waste. More importantly, the gains that the proposed D-TFDD scheme for unknown ICI achieves compared to the benchmark schemes without ICI knowledge are considerable. For example, the proposed D-TFDD scheme for unknown ICI has a BS-U1 throughput gain of about $66\%$, $80\%$ and $100\%$, for SINR=10 dB, and about $17\%$, $18\%$ and $38\%$, for SINR=20 dB, compared to the existing D-TDD and to the static-TDD schemes without ICI knowledge for a BS-U2 throughput of 2, 3, and 4 bits/symb, respectively.
![Throughput regions of the proposed D-TFDD schemes: left for SINR=10 dB, and right for SINR=20 dB.[]{data-label="RateRegion_700m"}](RateRegion_Interference_700m.eps){width="6in"}
### Sum Throughput {#Sec-RTR1}
In Fig. \[discrete\_rate\_SNR\], we illustrate the sum of the BS-U1 and BS-U2 throughputs achieved with the proposed D-TFDD scheme for unknown ICI and the static-TDD as a function of the SINR for $M =1, 4, 16 , \infty$, where $R$ is set to $R = 10/M$ bits/symb. From Fig. \[discrete\_rate\_SNR\] we can see that by increasing $M$ from 1 to 4 in the proposed D-TFDD scheme for unknown ICI, we can gain more than 10 dBs in SINR for around $4$ bits/symb sum throughput. Whereas, by increasing $M$ from 4 to 16 we can gain an additional 1 dB in SINR for around $4$ bits/symb of the sum throughput. Moreover, the proposed D-TFDD scheme for unknown ICI achieves substantial gains compared to the static-TDD. For example, about 3 dB and 10 dB SINR gain is achieved for $M=1$ and $M=4$, respectively, for around $4$ bits/symb sum throughput. Finally, Fig. \[discrete\_rate\_SNR\] shows that the proposed scheme for $M=4$ performs very close to the case when $M=\infty$. Note that in Fig. \[discrete\_rate\_SNR\], the throughputs saturate since as a function of $M$, the transmission rates are $10/M$ bits/symb. Hence, the maximum possible transmission rate is $10$ bits/symb, even if the channels are error-free, which is the case in the high SINR.
The above numerical examples show that the proposed D-TFDD scheme provides double the diversity gain compared to existing TDD/FDD schemes, which improves that reliability of the communication. Moreover, since the proposed D-TFDD scheme works in a distributed fashion, it does not need any coordination of the BSs. Another strength is that is does not require ICI estimation, which makes it practical for implementation. Finally, the proposed D-TFDD scheme fits well into the scope of 5G. On the other hand, a weakness of the proposed D-TFDD scheme is the requirement of local CSI of the U1-BS and U2-BS channels at the BS, which entails signalling overhead.
![Throughputs vs. SINR of the proposed D-TFDD schemes with and without ICI knowledge with different discrete-rates quantization level.[]{data-label="discrete_rate_SNR"}](discrete_rate_1.eps){width="6in"}
Conclusion {#Sec-Conc}
==========
In this paper, we proposed a distributed D-TFDD scheme for unknown ICI. Using the proposed D-TFDD scheme, in a given frequency band, the BS adaptively selects to either communicate with U1 or with U2 in a given time slot based on the qualities of the BS-U1 and BS-U2 channels without ICI knowledge such that the BS-U1 and BS-U2 throughput region is maximized. We have shown that the proposed D-TFDD scheme provides significant throughput and outage probability gains compared to the conventional static-TDD scheme, as well as to the D-TDD and D-FDD schemes. Moreover, we observed the the proposed D-TFDD scheme doubles the diversity gain on both the BS-U1 and BS-U2 channels compared to existing duplexing schemes, even when the ICI is unknown, which leads to very large performance gains.
Proof of Theorem \[Theo\_PR\] {#app_PR}
-----------------------------
Constraints C1, C2 and C3 in (\[eq\_op\_PR\]) make the problem non-convex. To solve (\[eq\_op\_PR\]), we first we relax these constraints to $0 \leq q_1^m(t) \leq 1$, $0 \leq q_2^l(t) \leq 1$, and $0 \leq \sum_{m=1}^M q_1^m(t)+\sum_{l=1}^L q_2^l(t) \leq 1$, thereby making the relaxed problem convex. The solutions of the relaxed convex problem is then shown to be such that $q_1^m(t)$ and $q_2^l(t)$ take the limiting values 0 or 1, and not the values between 0 and 1. As a result, the relaxed convex problem is equivalent to the original problem. To solve the relaxed problem, we use the Lagrangian. Thereby, we obtain $$\begin{aligned}
\label{eq_op_PR_append}
&\begin{array}{ll}
{\cal L} =& - \lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{m=1}^M \mu R_1^m q_1^m(t) O_1^m (t)
- \lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{l=1}^L (1-\mu)R_2^l q_2^l(t) O_2^l \\
&- { \sum_{m=1}^M {\lambda _1^m(t)} q_1^m(t)} - \left(1 - {\sum_{m=1}^M {\lambda _2^m(t)}q_1^m(t)}\right)
- { \sum_{l=1}^L {\lambda _3^l(t)} q_2^l(t)}- \left(1 - {\sum_{l=1}^L {\lambda _4^l(t)}q_2^l(t)}\right) \\
\end{array}\nonumber\\[-0.5em]
& \quad\quad\,\,\,\,\,-{\lambda _5(t)}\left({\sum_{m=1}^M q_1^m(t)} + \sum_{l=1}^L{ q_2^l(t)}\right) - {\lambda _6(t)}\left(1 - \sum_{m=1}^M{ q_1^m(t)} - \sum_{l=1}^L{ q_2^l(t)}\right), \end{aligned}$$ where $\lambda _1^m(t)\geq0$, $\lambda _2^m(t)\geq0$, $\lambda _3^l(t)\geq0$, $\lambda _4^l(t)\geq0$, $\lambda _5(t)\geq0$, and $\lambda_6(t)\geq0$, $\forall m,l,i$, are the Lagrangian multipliers. Next, we rewrite (\[eq\_op\_PR\_append\]) as $$\begin{aligned}
\label{eq_oprew_PR_append}
&\begin{array}{ll}
{\cal L} =& -\lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \mu q_{1}(t) {\underset{m }{\textrm{max} }}\{R_1^m O_1^m (t)\}
- \lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T (1-\mu) q_{2}(t) {\underset{l }{\textrm{max} }}\{R_2^l O_2^l (t)\} \\
&- { \sum_{m=1}^M {\lambda _1^m(t)} q_1^m(t)} - \left(1 - {\sum_{m=1}^M {\lambda _2^m(t)}q_1^m(t)}\right)
- { \sum_{l=1}^L {\lambda _3^l(t)} q_2^l(t)}- \left(1 - {\sum_{l=1}^L {\lambda _4^l(t)}q_2^l(t)}\right) \\
\end{array}\nonumber\\[-0.5em]
& \quad\quad\,\,\,\,\,-{\lambda _5(t)}\left({ q_{1}(t)} + {q_{2}(t)}\right) - {\lambda _6(t)}\left(1 - { q_{1}(t)} -{ q_{2}(t)}\right). \end{aligned}$$
Now, using (\[eq\_oprew\_PR\_append\]) and defining $-\lambda_{5}(t)+\lambda_{6}(t)\triangleq\beta(t)$, we can find the optimal state-selection variables $q_1^m(t)$ and $q_2^l(t)$ as follows. The conditions which maximize (\[eq\_op\_PR\_append\]), in the cases when the transmit node on the BS-U1 channel transmits with $R_1^m$ and the transmit node on the BS-U2 channel is silent, are $$\begin{aligned}
\label{eq_PR_power_q1}
[ \mu {\underset{m }{\textrm{max} }}\{R_1^m O_1^m (t)\} - \beta(t) > 0 ] \; \textrm{and} \; [ (1 - \mu ) {\underset{l }{\textrm{max} }}\{R_2^l O_2^l (t)\} - \beta(t) < 0 ].\end{aligned}$$ On the other hand, the conditions for maximizing (\[eq\_oprew\_PR\_append\]) in the case when the transmit node on the BS-U2 channel transmits with $R_2^l$ and the transmit node on the BS-U1 channel is silent, are $$\begin{aligned}
\label{eq_PR_power_q2}
[ \mu {\underset{m }{\textrm{max} }}\{R_1^m O_1^m (t)\} - \beta(t) < 0 ] \; \textrm{and} \; [ (1 - \mu ) {\underset{l }{\textrm{max} }}\{R_2^l O_2^l (t)\} - \beta(t) > 0 ].\end{aligned}$$ In (\[eq\_PR\_power\_q1\]) and (\[eq\_PR\_power\_q2\]), we can substitute $\mu R_1^m O_1^m (t) $ with $\Lambda_1^m(t)$ and $ (1 - \mu ) R_2^l O_2^l (t) $ with $\Lambda_2^l(t)$, and thereby obtain $q_1^m(t)$ and $q_2^l(t)$ as in (\[eq\_scheme\_AP\_G\]). This completes the proof.
Proof of Theorem \[Theo\_PR\_Outage\_G\] {#app_PR_Outage_G}
----------------------------------------
In time slot $t$, an outage occurs if the BS-U1 channel is selected for transmission and the BS-U1 channel is too weak to support the rate $R_0$, i.e., $q_1^1(t)=1$ and $O_{1}^1(t)=0$, or if the BS-U2 channel is selected to transmit and the BS-U2 channel is too weak to support the rate $R_0$, i.e., $q_2^1(t)=1$ and $O_{2}^1(t)=0$, or if both the BS-U1 the BS-U2 channels are not selected for transmission in time slot $t$, i.e., if $q_1^1(t)=q_2^1(t)=0$, since in that case the time slot $t$ is wasted. Hence, the outage probability $P_{\rm out}$ can be obtained as $$\begin{aligned}
\label{eq_out12_G}
P_{\rm out}&={\rm Pr}\Big\{[q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0] \textrm{ OR } [q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0] \nonumber\\
& \quad \quad \quad \textrm{ OR } [q_1^1(t)=q_2^1(t)=0]\Big\}\nonumber\\
& \stackrel{(a)}{=}{\rm Pr}\big\{ q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \big\} + {\rm Pr}\big\{ q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0 \big\}\nonumber\\ &+{\rm Pr}\big\{q_1^1(t)=q_2^1(t)=0\big\},\end{aligned}$$ where $(a)$ follows since the events $q_1^1(t)=1$ and $q_1^1(t)=0$, and also the events $q_2^1(t)=1$ and $q_2^1(t)=0$ are mutually exclusive. Since $\mu=\frac{1}{2}$, $\Lambda_1^1(t)$ and $\Lambda_2^1(t)$ in (\[eq\_PR\_1\]) and (\[eq\_PR\_2\]) simplify to $$\begin{aligned}
\Lambda_k^1(t) &=\frac{1}{2} R_0 O_k^1 (t) ,\; k\in\{1,2\}.\label{eq_PR_1a}\end{aligned}$$ Now, inserting $\Lambda_1^1(t)$ and $\Lambda_2^1(t)$ from (\[eq\_PR\_1a\]) into (\[eq\_scheme\_AP\_G\]), we obtain that $q_1^1(t)=1$ if $O_1^1 (t)\geq O_2^1 (t)$ and $O_1^1 (t)>0$, which means that $q_1^1(t)=1$ occurs if $O_1^1 (t)=1$. Hence, the event $q_1^1(t)=1$ and $O_1^1 (t)=0$ is an impossible event, thereby leading to ${\rm Pr}\big\{ q_2^1(t)=1 \textrm { AND } O_2^1(t)=0 \big\}=0$ in (\[eq\_out12\_G\]). Similarly, we can conclude that ${\rm Pr}\big\{ q_1^1(t)=1 \textrm { AND } O_1^1(t)=0 \big\}=0$ in (\[eq\_out12\_G\]). Next, we obtain that $q_1^1(t)=q_2^1(t)=0$ occurs iff $O_1^1 (t)= O_2^1 (t)=0$ holds, thereby leading to ${\rm Pr}\big\{q_1^1(t)=q_2^1(t)=0\big\}={\rm Pr}\big\{O_1^1 (t)=0 \textrm{ AND } O_2^1 (t)=0\big\}$ in (\[eq\_out12\_G\]). Inserting this into (\[eq\_out12\_G\]), we obtain $$\begin{aligned}
\label{eq_out1_2a}
P_{\rm out}&={\rm Pr}\big\{O_1^1 (t)=0 \textrm{ AND } O_2^1 (t)=0\big\} \nonumber\\
&={\rm Pr}\left\{\log_2\left(1 + P\frac{|\gamma_{1}(t)|^2}{ 1+\gamma _{I1}(t)}\right)< R_0 \textrm{ AND } \log_2\left(1 + P\frac{|\gamma_{2}(t)|^2}{ 1+\gamma _{I2}(t)}\right)< R_0\right\}\nonumber\\
& ={\rm Pr}\left\{ \frac{|\gamma_{1}(t)|^2}{ 1+\gamma _{I1}(t)}<\gamma_{\rm th} \textrm{ AND } \frac{|\gamma_{2}(t)|^2}{ 1+\gamma _{I2}(t)}<\gamma_{\rm th} \right\},\end{aligned}$$ where $\gamma_{\rm th}=\frac{2^{R_0}-1}{P} $. The variables $\gamma_{1}(t)$ and $\gamma_{2}(t)$ have identical and independent exponential distributions with PDFs denoted by $f_{\gamma_{1}}(\gamma_{1})$ and $f_{\gamma_{2}}(\gamma_{2})$, respectively, both with mean $\Omega_0=E\{\frac{|h_{1}(t)|^2}{\sigma_1^2}\}=E\{\frac{|h_{2}(t)|^2}{\sigma_2^2 }\}$. On the other hand, the variables $\gamma_{I1}(t)$ and $\gamma_{I2}(t)$ have identical yet dependent exponential distributions with joint PDF denoted by $f_{\gamma_{I1},\gamma_{I2}}\big(Z_1,Z_2\big)$. As a result, (\[eq\_out1\_2a\]) can be obtained as $$\begin{aligned}
P_{\rm out}= \int\limits_{0}^{\infty} \int\limits_{0}^{\infty} \int\limits_{0}^{\gamma_{\rm th} \big(1+Z_1\big)} {} \int\limits_{0}^{\gamma_{\rm th} \big(1+Z_2\big)} & f_{\gamma_{1},\gamma_{2},\gamma_{I1},\gamma_{I2}}\big(\gamma_{1},\gamma_{2},Z_1,Z_2\big)
\; d{\gamma_{1}} d{\gamma_{2}} d{Z_1} d{Z_2} .\label{Out_U5GG1}\end{aligned}$$ We can rewrite $f_{\gamma_{1},\gamma_{2},\gamma_{I1},\gamma_{I2}}\big(\gamma_{1},\gamma_{2},Z_1,Z_2\big)$ in (\[Out\_U5GG1\]) as $$\begin{aligned}
\label{Out_U5GG1123}
&f_{\gamma_{1},\gamma_{2},\gamma_{I1},\gamma_{I2}}\big(\gamma_{1},\gamma_{2},Z_1,Z_2\big)
=f_{\gamma_{1}|\gamma_{I1},\gamma_{I2}}\big(\gamma_{1}\big) f_{\gamma_{2}|\gamma_{I1},\gamma_{I2}}\big(\gamma_{2}\big) f_{\gamma_{I1},\gamma_{I2}}\big(Z_1,Z_2\big),\end{aligned}$$ since $\gamma_{1}$ and $\gamma_{2}$ have independent distributions. By substituting the PDFs of $\gamma_{1}$ and $\gamma_{2}$ into (\[Out\_U5GG1123\]), then inserting (\[Out\_U5GG1123\]) into (\[Out\_U5GG1\]), we can obtain (\[Out\_U5GG1\]) as $$\begin{aligned}
P_{\rm out} = \int\limits_{0}^{\infty} & \int\limits_{0}^ {\infty} \frac{1}{\Omega_0} \int\limits_{0}^{\gamma_{\rm th} (1+Z_2)} e^{-(\frac{\gamma_{2}}{\Omega_0})} d{\gamma_{2}}
\times \frac{1}{\Omega_0} \int\limits_{0}^{\gamma_{\rm th} (1+Z_1)} e^{-(\frac{\gamma_{1}}{\Omega_0})} d{\gamma_{1}} \nonumber\\
& \times f_{\gamma_{I1},\gamma_{I2}}(Z_1,Z_2) d{Z_1} d{Z_2} .\label{Out_U5GG2}\end{aligned}$$
In (\[Out\_U5GG2\]), by integrating over $\gamma_{1}$ and $\gamma_{2}$, and letting $P \to \infty$ (consequently $\gamma_{\rm th} \to 0$), we obtain $$\begin{aligned}
P_{\rm out} \to \int\limits_{0}^{\infty} \int\limits_{0}^ {\infty} & \frac{\gamma_{\rm th} (1+Z_2)}{\Omega_0} \times \frac{\gamma_{\rm th} (1+Z_1)}{\Omega_0}
\times f_{\gamma_{I1},\gamma_{I2}}(Z_1,Z_2) d{Z_1} d{Z_2} \; \textrm{as} \;{P\to\infty} .\label{Out_U5GG3}\end{aligned}$$
Finally, by replacing $$\begin{aligned}
E \Big \{ (1+\gamma _{I1}(t)) (1+\gamma _{I2}(t)) \Big \}= \int\limits_{0}^{\infty} \int\limits_{0}^ {\infty} (1+Z_2)\times (1+Z_1) \times f_{\gamma_{I1},\gamma_{I2}}(Z_1,Z_2) d{Z_1} d{Z_2},\end{aligned}$$ into (\[Out\_U5GG3\]), the outage is obtained as in (\[Out\_U6GG\]). This completes the proof.
Proof of Proposition \[Theo\_PR\_L0cal\] {#app_PR_Local}
----------------------------------------
The BS-U1 and BS-U2 throughputs are given in (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), respectively. However, since we can not compute $O_{1}^{m}(t)$ and $O_{2}^{l}(t)$, instead of (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), we use estimates for (\[eq\_PR\_fr3\]) and (\[eq\_PR\_fr4\]), given by $$\begin{aligned}
\bar R_{1,e} &=\lim_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{m=1}^M R_1^m q_1^m(t) O_{1,e}^m (t),\label{eq_PR_fr3123}
\\
\bar R_{2,e} &=\lim_{T\to\infty}\frac 1T \sum_{t=1}^T \sum_{l=1}^L R_2^l q_2^l(t) O_{2,e}^l (t).\label{eq_PR_fr4123}\end{aligned}$$ The accuracy of the estimates $\bar R_{1,e}$ and $\bar R_{2,e}$, depends on the following expressions $$\begin{aligned}
\delta_1=&\lim_{T\to\infty}\frac 1T \sum_{t=1}^T q_1^m(t) \bigg ( O_{1}^{m}(t)-O_{1,e}^{m}(t) \bigg )^2. \label{eq_PR_fr334}\\
\delta_2=&\lim_{T\to\infty}\frac 1T \sum_{t=1}^T q_2^l(t) \bigg ( O_{2}^{l}(t)-O_{2,e}^{l}(t) \bigg )^2.\label{eq_PR_fr434}\end{aligned}$$ Hence, (\[eq\_PR\_fr334\]) and (\[eq\_PR\_fr434\]) express the average of the difference squared between the outages when the ICI is known and the estimation of the outages when the ICI is unknown. The smaller (\[eq\_PR\_fr334\]) and (\[eq\_PR\_fr434\]) are, the more accurate the estimates $\bar R_{1,e}$ and $\bar R_{2,e}$ become. In fact, when $\delta_1 \to 0$, and $\delta_2 \to 0$, $\bar R_{1,e} \to \bar R_{1}$ and $\bar R_{2,e} \to \bar R_{2}$.
Now, if $\delta_1 < \epsilon$ and $\delta_2 < \epsilon$ hold, the constants $\gamma _{I1}$ and $\gamma _{I2}$ that maximize the estimated BS-U1 and BS-U2 throughput region, defined in (\[eq\_PR\_fr3123\]) and (\[eq\_PR\_fr4123\]), can be found from the following maximization problem $$\begin{aligned}
\label{eq_op_PRlt}
& {\underset{\gamma _{I1}, \gamma _{I2}.\;}{\textrm{Maximize: }}}\; \lim_{T\to\infty}\frac 1T \sum_{t=1}^T \mu R_1^{m^*} q_1^{m^*}(t) O_{1,e}^{m^*}(t)
+ \lim_{T\to\infty}\frac 1T \sum_{t=1}^T \big(1 - {\mu} \big) R_2^{l^*} q_2^{l^*}(t) O_{2,e}^{l^*}(t) \nonumber\\
& {\rm{Subject\;\; to \; :}} \nonumber\\
& \qquad {\rm C1:} \; \lim_{T\to\infty}\frac 1T \sum_{t=1}^T q_1^{m^*}(t)\bigg ( O_{1}^{m^*}(t)-O_{1,e}^{m^*}(t) \bigg )^2\leq \epsilon, \nonumber\\
& \qquad {\rm C2:} \; \lim_{T\to\infty}\frac 1T \sum_{t=1}^T q_2^{l^*}(t)\bigg ( O_{2}^{l^*}(t)-O_{2,e}^{l^*}(t) \bigg )^2 \leq \epsilon,\end{aligned}$$ where $m^*=\arg {\underset{m }{\textrm{max} }}\{R_1^m O_{1,e}^m (t)\}$ and $l^*=\arg{\underset{l }{\textrm{max} }}\{R_2^l O_{2,e}^l (t)\}$.
By applying the Lagrangian function on (\[eq\_op\_PRlt\]), we obtain $$\begin{aligned}
\label{eq_op_PR_appendl}
&\begin{array}{ll}
{\cal L} =& - \lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \mu R_1^{m^*} q_1^{m^*}(t) O_{1,e}^{m^*}(t)
- \lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T (1-\mu)R_2^{l^*} q_2^{l^*}(t) O_{2,e}^{l^*}(t) \\
\end{array}\nonumber\\[-0.5em]
& +\lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \chi_1 q_1^{m^*}(t) \bigg ( O_{1}^{m^*}(t)-O_{1,e}^{m^*}(t) \bigg )^2+ \lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \chi_2 q_2^{l^*}(t)\bigg ( O_{2}^{l^*}(t)-O_{2,e}^{l^*}(t) \bigg)^2,\end{aligned}$$ where $\chi_1 \geq 0$ and $\chi_2 \geq 0$ are the Lagrangian multipliers, found such that C1 and C2 in (\[eq\_op\_PRlt\]) hold. By differentiating ${\cal L}$ in (\[eq\_op\_PR\_appendl\]) with respect to $\gamma _{I1}$ and $\gamma _{I2}$, and equivalenting the results to zero, we obtain $$\begin{aligned}
\label{gamaU_l11}
\lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \frac{ - P_1 \gamma_{1}(t) \delta_1^{m^*}(t) q_1^{m^*}(t) \bigg ( -\mu R_1^{m^*} + 2\chi_1 \big [ O_{1,e}^{m^*}(t)-O_{1}^{m^*}(t) \big] \bigg)}{\ln(2) \Big (1+\gamma _{I1}+P_1 \gamma _{1}(t)\Big) \Big (1+\gamma _{I1}\Big)}=0 ,\end{aligned}$$ $$\begin{aligned}
\label{gamaD_l11}
\lim\limits_{T\to\infty}\frac 1T \sum_{t=1}^T \frac{ - P_2 \gamma_{2}(t) \delta_2^{l^*}(t) q_2^{l^*}(t)\bigg ( -(1-\mu) R_2^{l^*} + 2\chi_2 \big [ O_{2,e}^{l^*}(t)-O_{2}^{l^*}(t) \big] \bigg)}{\ln(2) \Big (1+\gamma _{I2}+P_2 \gamma _{2}(t)\Big) \Big (1+\gamma _{I2}\Big)}=0.\end{aligned}$$ Due to the law of large numbers, (\[gamaU\_l11\]) and (\[gamaD\_l11\]) can be written as $$\begin{aligned}
\label{gamaU_l11a}
E \left ( \frac{ - P_1 \gamma_{1}(t) \delta_1^{m^*}(t) q_1^{m^*}(t)\bigg ( -\mu R_1^{m^*} + 2\chi_1 \big [ O_{1,e}^{m^*}(t)-O_{1}^{m^*}(t) \big] \bigg)}{\ln(2) \Big (1+\gamma _{I1}+P_1 \gamma _{1}(t)\Big) \Big (1+\gamma _{I1}\Big)} \right ) =0 ,\end{aligned}$$ $$\begin{aligned}
\label{gamaD_l11a}
E \left (\frac{ - P_2 \gamma_{2}(t) \delta_2^{l^*}(t) q_2^{l^*}(t)\bigg ( -(1-\mu) R_2^{l^*} + 2\chi_2 \big [ O_{2,e}^{l^*}(t)-O_{2}^{l^*}(t) \big] \bigg)}{\ln(2) \Big (1+\gamma _{I2}+P_2 \gamma _{2}(t)\Big) \Big (1+\gamma _{I2}\Big)} \right )=0.\end{aligned}$$
Calculating the constants $\gamma _{I1}$ and $\gamma _{I2}$ from (\[gamaU\_l11a\]) and (\[gamaD\_l11a\]) requires the derivation of the above expectations. Instead, we use a more practical approach, where the constants $\gamma _{I1}$ and $\gamma _{I2}$ are estimated as $\gamma_{I1}^e(t)$ and $\gamma_{I2}^e(t)$ in time slot $t$. To this end, we apply the gradient descent method [@Boyd_CO] on (\[gamaU\_l11\]) and (\[gamaD\_l11\]) to obtain $\gamma_{I1}^e(t)$ and $\gamma_{I2}^e(t)$ as in (\[eq\_z1\_Rsr1\]) and (\[eq\_z1\_Rsr2\]), where $\delta_{k}( t)$ for $k\in\{1,2\}$ is an adaptive step size which controls the speed of convergence of $\gamma_{Ik}^e(t)$ to $\gamma_{Ik}$, for $k\in\{1,2\}$, which can be some properly chosen monotonically decaying function of $t$ with $\delta_{k}( 1)<1$. Note that $\lim\limits_{t\to\infty} \gamma_{I1}^e(t) = \gamma_{I1}$ and $\lim\limits_{t\to\infty} \gamma_{I2}^e(t) = \gamma_{I2}$. This completes the proof.
Proof of Theorem \[Theo\_PR\_Outage\_L\] {#app_PR_Outage_L}
----------------------------------------
In time slot $t$, an outage occurs if the BS-U1 channel is selected for transmission and the BS-U1 channel is too weak to support the rate $R_0$, i.e., $q_1^1(t)=1$ and $O_{1}^1(t)=0$, or if the BS-U2 channel is selected to transmission and the BS-U2 channel is too weak to support the rate $R_0$, i.e., $q_2^1(t)=1$ and $O_{2}^1(t)=0$, or if both the BS-U1 and BS-U2 channels are not selected for transmission in time slot $t$, i.e., if $q_1^1(t)=q_2^1(t)=0$, since in that case the time slot $t$ is wasted. Assuming that $\gamma_{I1}^e(t)$ and $\gamma_{I2}^e(t)$ have converged to their steady states given by $\mathop {\lim }\limits_{t \to \infty } \gamma_{I1}^e(t) =\gamma_{I1}$ and $\mathop {\lim }\limits_{t \to \infty } \gamma_{I2}^e(t) =\gamma_{I2}$, the outage probability, $P_{\rm out}$, can be found as $$\begin{aligned}
\label{eq_out12}
P_{\rm out}&={\rm Pr}\Big\{[q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0] \textrm{ OR } [q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0] \nonumber\\
& \quad \quad \quad \textrm{ OR } [q_1^1(t)=q_2^1(t)=0]\Big\}\nonumber\\
& \stackrel{(a)}{=}{\rm Pr}\big\{ q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \big\} + {\rm Pr}\big\{ q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0 \big\}\nonumber\\ &+{\rm Pr}\big\{q_1^1(t)=q_2^1(t)=0\big\},\end{aligned}$$ where $(a)$ follows since the events $q_1^1(t)=1$ and $q_1^1(t)=0$, and also the events $q_2^1(t)=1$ and $q_2^1(t)=0$ are mutually exclusive.
We divide (\[eq\_out12\]) into three events; BS-U1 communication event, $[q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0] $, BS-U2 communication event $[q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0]$, and the silent event $[q_1^1(t)=q_2^1(t)=0]$. In the following, we calculate the probability of these three events.
For the BS-U1 communication event, we have $q_1^1(t)=1$ when either of the two following events occur
- $O_{1,e}^1 (t)=1$ and $O_{2,e}^1(t)=0$. This event occurs when $\frac{\gamma_{1}(t)}{1+\gamma _{I1}} \geq \gamma_{\rm th}$ and $\frac{\gamma_{2}(t)}{1+\gamma _{I2}} < \gamma_{\rm th}$, where $\gamma_{\rm th}=\frac{2^{R_0}-1}{P} $.
- $O_{1,e}^1 (t)=O_{2,e}^1(t)=1$ and $\gamma_{U}^e(t) \geq \gamma_{D}^e(t)$. This event occurs when $\frac{\gamma_{1}(t)}{1+\gamma _{I1}} > \gamma_{\rm th}$, $\frac{\gamma_{2}(t)}{1+\gamma _{I2}} > \gamma_{\rm th}$, and $\frac{\gamma_{1}(t)}{1+\gamma _{I1}} \geq \frac{\gamma_{2}(t)}{1+\gamma _{I2}} $.
On the other hand, the event $O_{1}^1(t)=0$ occurs with the following probability $$\begin{aligned}
\label{Otage_Z}
{\rm Pr}\Big\{O_{1}^1(t)=0 \Big\}= {\rm Pr} \left({\frac{\gamma_{1}(t)}{1+\gamma_{I1}(t)} < \gamma_{\rm th} }\right),\end{aligned}$$ where $\gamma_{I1}(t)$ is the sum of $K$ identical and independent random variables with mean (identical mean is assumed for simplicity) $\Omega_I$, which has the following PDF $$\begin{aligned}
\label{PDF_erl}
f_{\gamma_{I1}}(z)=\frac{z^{K-1} e^{-\frac{z}{\Omega_I}}}{\Omega_I^K (K-1)!}\end{aligned}$$ and the following cumulative distribution function (CDF) $$\begin{aligned}
\label{CDF_erl}
F_{\gamma_{I1}}(z)=1- \sum_{n=0}^{K-1} \frac{1}{n!} e^{-\frac{z}{\Omega_I}} {\bigg(\frac{z}{\Omega_I} \bigg)^n}.\end{aligned}$$
In addition, the variables $\gamma_{1}(t)$ and $\gamma_{2}(t)$ have i.i.d. exponential distributions with PDFs, $f_{\gamma_{1} }(\gamma_{1} )$ and $f_{\gamma_{2}}(\gamma_{2})$, respectively, that have mean $\Omega_0=E\{\frac{|h_{1}(t)|^2}{\sigma_1^2 }\}=E\{\frac{|h_{2}(t)|^2}{\sigma_2^2 }\}$.
Using the above, we can rewrite ${\rm Pr}\Big\{q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \Big\}$ in an integral form as $$\begin{aligned}
\label{Out_U1}
&{\rm Pr}\Big\{q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \Big\}\nonumber\\
&= {\int\limits_{\gamma_{\rm th} ({1+\gamma _{I2}})}^{\infty} \int\limits_{0}^{\gamma_{\rm th}({1+\gamma _{I1}}) } {\rm Pr} \left({ Z >\frac{\gamma_{1}}{\gamma_{\rm th} } -1 }\right) f_{\gamma_{1}}(\gamma_{1}) f_{\gamma_{2}}(\gamma_{2}) d{\gamma_{2}} d{\gamma_{1}} }\nonumber\\
&+ { \int\limits_{\gamma_{\rm th} ({1+\gamma _{I2}})}^{\infty} \int\limits_{\gamma_{2} \big( \frac{1+\gamma _{I1}}{1+\gamma _{I2}} \big)}^{\infty} {\rm Pr} \left({ Z >\frac{\gamma_{1}}{\gamma_{\rm th} } -1 }\right)f_{\gamma_{1}}(\gamma_{1}) f_{\gamma_{2}}(\gamma_{2}) d{\gamma_{1}} d{\gamma_{2}}} \nonumber\\
&\mathop = \limits^{(a)} \int\limits_{\gamma_{\rm th}({1+\gamma _{I1}})}^{\infty} {\int\limits_{0}^{\gamma_{1} \big( \frac{1+\gamma _{I2}}{1+\gamma _{I1}} \big)} \sum_{n=0}^{K-1} \frac{1}{n!} e^{-\left(\frac{[\frac{\gamma_{1}}{\gamma_{\rm th}}-1]}{\Omega_I} \right )} {\left(\frac{[\frac{\gamma_{1}}{\gamma_{\rm th}}-1]}{\Omega_I} \right)^n} f_{\gamma_{1}}(\gamma_{1}) f_{\gamma_{2}}(\gamma_{2}) d{\gamma_{2}} d{\gamma_{1}} },\end{aligned}$$ where (a) follows from (\[CDF\_erl\]). Now, performing the integration with respect to $\gamma_{2}$, we obtain $$\begin{aligned}
\label{Out_U2}
&{\rm Pr}\Big\{q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \Big\}\nonumber\\
&= \frac{\gamma_{\rm th} \Omega_I e^{- (\frac{\gamma_{\rm th}}{\Omega_0})}}{\Omega_0} \int\limits_{\frac{\gamma _{I1}}{\Omega_I}}^{\infty} { \sum_{n=0}^{K-1} \frac{1}{n!} e^{-U'} {(U')^n} e^{-\left(\frac{\gamma_{\rm th} \Omega_I U'}{\Omega_0 } \right)} \left(1-e^{- \left[(\frac{\gamma_{\rm th} }{\Omega_0 }) (1+\Omega_I U') (\frac{1+\gamma _{I2}}{1+\gamma _{I1}} ) \right ]} \right) dU'},\end{aligned}$$ where $U'=\frac{ \left(\frac{\gamma_{1}}{\gamma_{\rm th}}-1 \right)}{\Omega_I}$. In (\[Out\_U2\]), when $P \to \infty$, and consequently $\gamma_{\rm th} \to 0$, we can approximate $(1-e^{-[(\frac{\gamma_{\rm th} }{\Omega_0 }) (1+\Omega_I U') (\frac{1+\gamma _{I2}}{1+\gamma _{I1}} ) ]})$ by $({[(\frac{\gamma_{\rm th} }{\Omega_0 }) (1+\Omega_I U') (\frac{1+\gamma _{I2}}{1+\gamma _{I1}} ) ]})$, $e^{-(\frac{\gamma_{\rm th}}{\Omega_0})}$ by 1, and $e^{-(\frac{\gamma_{\rm th} \Omega_I U'}{\Omega_0 })}$ by 1. As a result, (\[Out\_U2\]) can be rewritten as $$\begin{aligned}
& {\rm Pr}\Big\{q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \Big\}\label{Out_U3}\\
&\to \frac{ \gamma_{\rm th}^2 \Omega_I }{\Omega_0^2} \times \frac{1+\gamma _{I2}}{1+\gamma _{I1}} \hspace{-2mm}\left[\sum_{n=0}^{K-1} \frac{1}{n!} \int\limits_{\frac{\gamma _{I1}}{\Omega_I}}^{\infty} { e^{-U'} {(U')^n} dU'} + \Omega_I \sum_{n=0}^{K-1} \frac{1}{n!} \int\limits_{\frac{\gamma _{I1}}{\Omega_I}}^{\infty} { e^{-U'} {(U')^{n+1}} dU'}\right] \hspace{-2mm} \; \textrm{as} \;P \to \infty.\nonumber\end{aligned}$$
By calculating the integrals and the summations in (\[Out\_U3\]), we obtain $$\begin{aligned}
\label{Out_U4}
{\rm Pr}\Big\{q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \Big\}\to \frac{ \gamma_{\rm th}^2 {\hat \Omega_{I1}} }{\Omega_0^2} \; \textrm{as} \;P \to \infty,\end{aligned}$$ where ${\hat \Omega_{I1}}$ is given in (\[Otage\_U\_final\]).
For the BS-U2 communication event, we obtain that $q_2^1(t)=1$ if $O_{2,e}^1 (t)=1$ and $O_{1,e}^1(t)=0$, or $O_{2,e}^1 (t)=O_{1,e}^1(t)=1$ and $\frac{\gamma_{1}(t)}{1+\gamma _{I1}}<\frac{\gamma_{2}(t)}{1+\gamma _{I2}}$. The event $O_{2,e}^1 (t)=1$ and $O_{1,e}^1(t)=0$ occurs when $\frac{\gamma_{2}(t)}{1+\gamma _{I2}} > \gamma_{\rm th}$ and $\frac{\gamma_{1}(t)}{1+\gamma _{I1}} < \gamma_{\rm th}$. The event $O_{2,e}^1 (t)=O_{1,e}^1(t)=1$ and $\frac{\gamma_{1}(t)}{1+\gamma _{I1}}<\frac{\gamma_{2}(t)}{1+\gamma _{I2}}$ occurs when $\frac{\gamma_{2}(t)}{1+\gamma _{I2}} > \gamma_{\rm th}$, $\frac{\gamma_{1}(t)}{1+\gamma _{I1}} > \gamma_{\rm th}$, and $\frac{\gamma_{1}(t)}{1+\gamma _{I1}}<\frac{\gamma_{2}(t)}{1+\gamma _{I2}}$. Using this, we can derive ${\rm Pr}\Big\{q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0 \Big\}$ with a similar approach as the calculation of ${\rm Pr}\Big\{q_1^1(t)=1 \textrm { AND } O_{1}^1(t)=0 \Big\}$, which results in $$\begin{aligned}
\label{Out_U44}
{\rm Pr}\Big\{q_2^1(t)=1 \textrm { AND } O_{2}^1(t)=0 \Big\}\to \frac{ \gamma_{\rm th}^2 {\hat \Omega_{I2}} }{\Omega_0^2} \; \textrm{as} \;P \to \infty,\end{aligned}$$ where ${\hat \Omega_{I2}}$ is given in (\[Otage\_D\_final\]).
Finally, for the silent event, we obtain that $q_1^1(t)=q_2^1(t)=0$ occurs iff $O_{1,e}^1 (t)= O_{2,e}^1 (t)=0$ holds, which occurs when $\frac{\gamma_{2}(t)}{1+\gamma _{I2}} < \gamma_{\rm th}$, and $\frac{\gamma_{1}(t)}{1+\gamma _{I1}} < \gamma_{\rm th}$. The probability of the silent event can be calculated by $$\begin{aligned}
\label{Out_U5}
& {\rm Pr}\big\{q_1^1(t)=q_2^1(t)=0\big\} = \frac{1}{\Omega_0^2}\int\limits_{0}^{\gamma_{\rm th} ({1+\gamma _{I2}})} {\int\limits_{0}^{\gamma_{\rm th} ({1+\gamma _{I1}})} e^{-(\frac{\gamma_{1}}{\Omega_0})} e^{-(\frac{\gamma_{2}}{\Omega_0})} d{\gamma_{1}} d{\gamma_{2}}} .\end{aligned}$$ The expression in (\[Out\_U5\]) for $P \to \infty$, and consequently $\gamma_{\rm th} \to 0$, converges to $$\begin{aligned}
\label{Out_U6}
{\rm Pr}\big\{q_1^1(t)=q_2^1(t)=0\big\}
\to \frac{\gamma_{\rm th}^2 }{\Omega_0^2} {\hat \Omega_{IS}} \; \textrm{as} \;P \to \infty,\end{aligned}$$ where ${\hat \Omega_{IS}}$ is given in (\[Otage\_S\_final\]).
Now, by adding (\[Out\_U4\]), (\[Out\_U44\]), and (\[Out\_U6\]), we obtain the asymptotic outage probability as in (\[eq\_out1asl\]). This completes the proof.
[^1]: Content presented in this article is subjected to Australian Provisional Patent Application 2019903224, filing date 2/Sep/2019, see http://pericles.ipaustralia.gov.au/ols/auspat/applicationDetails.do?applicationNo=2019903224
[^2]: This work has been published in part at IEEE ICC 2018 [@8403595].
[^3]: M. M. Razlighi N. Zlatanov are with the Department of Electrical and Computer Systems Engineering, Monash University, Melbourne, VIC 3800, Australia (e-mails: [email protected] and [email protected].)
P. Popovski is with the Department of Electronic Systems, Aalborg University, Denmark (e-mail: [email protected]).
[^4]: ICI emerges when BSs and users in neighbouring cells transmit and receive on the same frequency band.
[^5]: Time slot is a time interval that is equal or smaller than the duration of the coherence interval. Moreover, we assume that a time slot is long enough such that a capacity achieving codeword can be transmitted during one time slot.
[^6]: The subscripts 1 and 2 are used to symbolize the BS-U1 and the BS-U2 channels, respectively.
[^7]: For D-TFDD Type 3, $\gamma_{I1}(t)=\gamma_{I2}(t)$.
[^8]: For D-TFDD Type 2, $\sigma_1^2=\sigma_2^2$.
[^9]: Note that the defined throughput region is not the capacity region.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Abelian Lagrangians containing $\lambda\varphi^{4}$-type vertices are regularized by means of a suitable point-splitting scheme combined with generalized gauge transformations. The calculation is developped in details for a general Lagrangian whose fields (gauge and matter ones) satisfy usual conditions. We illustrate our results by considering some special cases, such as the $(\overline{\psi}\psi)^{2}$ and the Avdeev-Chizhov models. Possible application of our results to the Abelian Higgs model, whenever spontaneous symmetry breaking is considered, is also discussed. We also pay attention to a number of features of the point-split action such as the regularity and non-locality of its new “interacting terms”.'
author:
- |
Winder A. Moura-Melo$^{\mbox{a,b}}$ [^1] .1cm and J.A. Helayël-Neto$^{\mbox{b,c}}$ [^2]\
\
$^{\mbox{a}}$Deptartamento de Ciências Exatas, Universidade Federal de Lavras\
Caixa Postal 37, 37200-000, Lavras, MG, Brasil\
$^{\mbox{b}}$Centro Brasileiro de Pesquisas Físicas\
Rua Xavier Sigaud 150 - Urca, 22290-180 - Rio de Janeiro, RJ - Brasil.\
$^{\mbox{c}}$ Grupo de Física Teórica, Universidade Católica de Petrópolis\
Av. Barão do Amazonas 124, 25685-070, Petrópolis, RJ, Brasil.
title: '[Generalized Gauge Transformations and Regularized $\lambda\varphi^{4}$-type Abelian Vertices]{}'
---
Introduction
============
In the quantum field-theoretical framework of modern Physics, products of fields at the same space and/or time point are not well-defined since these fields are taken as operator-valued distributions. As a consequence of such ill-defined products, we are led to divergent results when calculating relevant quantities, such as, physical masses and coupling constants. Physically speaking, such divergences arise because we describe elementary particles as if they were point-like entities and, consequently, carrying infinite density of mass, charge, and so forth.\
\
Even though there are several regularization methods to deal with such problems, those based on point-splitting may offer some advantages respect to others when performed in a suitable way. Essentially, the procedure works by taking the field products initially at the same point, and later on at different points, by [*splitting*]{} them. The result is such that the new Lagrangian contains only regularized interaction terms. Already in 1934, Dirac employed such an idea in order to split products of quantities at the same points which appeared in density matrices of electronic and positronic physical distributions (see Ref.[@Dirac], for details) .\
\
More recently, several results have been obtained by means of this method for both Abelian (QED) and non-Abelian (Standard Model) cases. For example, the values of some important physical parameters such as the top quark and Higgs scalar masses, have been got free from divergences and were shown to be in good agreement with other schemes. This procedure was also shown to respect the gauge invariance of the theories (for details see the papers listed in Refs. [@OW; @OW2]).\
\
Nevertheless, these works did not pay enough attention to the explicit construction and form of the new [*point-split*]{} gauge transformations. Such an issue was the subject of a more recent paper, Ref.[@GNW], where the Abelian infinitesimal form of these new transformations (the so-called [*generalized gauge transformations*]{}, denoted by ggt’s) was proven to exist to all orders in the gauge coupling constant. The explicit forms of such ggt’s as well as of a generalized QED-Lagrangian were presented up to fourth order. This new Lagrangian, obtained from the original QED action, was shown to be regularized, i.e., its interaction terms (including some new ones which appear from the splitting) presented no product of fields at the same point. On the other hand, those new terms also displayed non-locality property. As expected, as we set the point-splitting parameter to zero, we recover the original results.\
\
Although the ggt’s have been built up for QED, we do not see any restriction in applying them to other Abelian theories containing usual vector gauge fields coupled to matter fields in a suitable way. Therefore, we intend here to apply the scheme discussed above to a quite general Lagrangian which contains, among others, a $\lambda\fy^4$-type vertex, in order to obtain its [*generalized*]{}, say, point-split version. This new Lagrangian will be explicitly constructed up to the second order in the gauge coupling constant. In addition, the present work sets out not only to show the good applicability of this alternative regularization procedure to a quite interesting class of Abelian interaction vertex, but also to motivate further investigation towards its non-Abelian version, which includes among others, the so-celebrated Higgs mechanism of the Salam-Weinberg electroweak theory.\
\
Here, it is worthy noticing that the gauge transformation parameter may explicitly appear in point-split actions, depending on the number of matter fields involved in the interaction vertex. Actually, while for 3-vertices (two matter plus one gauge field) the gauge parameter is generally absent from the point-split action, in matter 4-vertices its presence turns out to be, as far as we have understood, a natural ingredient to preserve gauge invariance under generalized gauge transformations to a given order in the gauge coupling constant. This comes from the fact that, in the framework of ggt’s the gauge invariance, in a generalized sense, has to be constructed and checked order by order in a gauge coupling constant expansion. Such an issue will become clearer throughout this work.\
\
Our paper follows the outline below. In Section 2 the Lagrangian which will be worked out is presented as well as a survey of the point-splitting scheme combined with generalized gauge transformations. Then, we apply such a procedure to our Lagrangian and step-by-step we worked out its generalized (split) expression up to the second order in the gauge coupling constant. Section 3 is devoted to applications of the results obtained in the previous section to some specific cases, say, $(\overline{\psi}\psi)^2$ and a modified version of the Avdeev-Chizhov models. We close our paper by pointing out some Concluding Remarks. Among others interests, we pay attention to the applicability of our results to the Abelian Higgs model whenever spontaneous symmetry breaking is concerned.
The Lagrangian and the regularization procedure
===============================================
We shall start this section by considering the following Lagrangian (which has the form of the massive scalar Electrodynamics with self-interaction term, or the Abelian Higgs model -provided that $m^{2}< 0$):[^3] (x)=-14 F\_F\^ +(D\_)\^(D\^) - -()\^2 \[1\] , with $D_{\mu} =
\partial_{\mu} +ieA_{\mu}$ and $F_{\mu\nu}=\partial_{\mu} A_{\nu}
-\partial_{\nu} A_{\mu}$. Clearly, the matter fields are considered to be complex[^4] and their products are taken at the same space-time point, say, $x$. This Lagrangian is invariant under the usual local gauge transformations: A\_(x)=-\_(x); (x)= +ie(x)(x) ; (x)=-ie(x)(x) \[2\]. Now, in order to obtain a point-split version of the Lagrangian above, i.e., a form free from same point product of fields in interaction terms, we begin by writing the generalized version of the gauge transformations, ggt’s (denoted by $\delta_{g}$) up to $e^2$ (see Ref.[@GNW] for further details): & & A\_(x)= -\_ (x)=A\_(x) \[gtA\],\
& & (x)=+ie (1)(2) +12 (ie)\^[2]{} \[(1) +(3)\] (1,3)(4) +[O]{}(e\^[3]{}) \[gtfi\],\
& & (x)= -ie (-1)(-2) +12 (ie)\^[2]{} \[(-1) +(-3)\] (-1,-3)(-4)+[O]{}(e\^[3]{}) \[gtfid\], where we have defined: & & (n)=(xna);(n)=(xna);(n)=(xna);\[def1\]\
& & (m,n)=\_[b0\^[+]{}]{} \^[xna-b]{}\_[xma +b]{} A\_()d\^ \[def2\]. the point-splitting being implemented by the (constant) 4-vector $a_\mu\equiv a$.\
\
From the definition above, we realize the first price to be paid in order to avoid product of fields at the same point: the non-locality of the new model. We shall come back to this point later.\
\
These ggt’s can be shown to satisfy the [*generalized Abelian condition*]{} up to $e^{2}$, i.e., the commutator of two distinct ggt’s (each of them with its respective parameter $a_1$ and $a_2$) vanishes up to this order: $$[\delta_{g1},\delta_{g2}]\fy(x)={\cal O}(e^{3}); \qquad
[\delta_{g1},\delta_{g2}]\fyd(x)={\cal O}(e^{3}).$$ is worthy noticing that, as the parameter $a$ is set to zero, all the results above recover the usual ones (hereafter, by consistency, the same should happen to all point-split results). Furthermore, we should stress that the point-splitting acts only in transformations which present same point product, which is the case for $\delta\fy$ and $\delta\fyd$, but not for $\delta
A_{\mu}$.[^5]\
\
Now, we discuss the invariance of the ordinary Lagrangian, eq. (\[1\]), under the ggt’s above (more precisely, up to order $e^{2}$). The kinetic gauge term is clearly invariant since $\gt A_{\mu}=\delta A_{\mu}$. The mass term for matter fields can be shown to be invariant in its action form, $\int
m^2 \fyd\fy d^{4}x$, with suitable change of variables within the integration (see Ref. [@GNW] for more details). Contrary, the other terms are not invariant and must have their points split up. We choose to do the point-splitting (P.S) in the following way (like as in (\[def1\]), $A_{\mu}(\pm n)$ stands for $A_{\mu}(x \pm n)$): & (D\_(x))\^(D\^(x)) &(D\_)\^(D\^)\_[P.S]{}=\
& & =, \[8\]\
& ((x)(x))\^[2]{}& ()\^[2]{}\_[P.S]{}= (-1)(1)(-2)(2) \[psfi4\]. And the split Lagrangian takes the form: \^[(0)]{}\_[P.S]{}= -14 F\_(x)F\^(x) - (x)(x) +(D\_)\^(D\^)\_[P.S]{} -()\^[2]{}\_[P.S]{} \[L0ps\] . Here, it is worthy noticing that, while the kinetic matter term, $\partial_{\mu}\fyd(x)\partial^{\mu}\fy(x)$, involves a product at the same point, it does not need to be split because the action of the ggt’s on it will produce regularized terms. Now, taking $\gt$ of such split terms up to order $e$, we get: ((D\_)\^(D\^)\_[P.S]{})= (ie)+[O]{} (e\^[2]{}) . which is, at first glance, non-vanishing. But, if we take its action form, we can perform a change of variables to show that the integrals exactly cancel each other. In other words, the r.h.s. of the previous expression gives rise to a vanishing term in the full split action.\
Next, for the self-interaction term, we get: & (()\^[2]{}\_[P.S]{})= & ie(-2)(2)+\
& & +ie (-1)(1)+[O]{}(e\^[2]{}) . Contrary to the previous one, the term above seems to be intrinsically non-vanishing; in fact, we did not see any way to set it to zero, neither by a suitable change of variables nor by partial integration. Therefore, we must search for a new term, $ \Omega^{(1)}_{P.S}$, such that $(\fyd\fy)^{2}_{P.S}+
\Omega^{(1)}_{P.S}$ be invariant under $\gt$ at least up to order $e$. This term exists and can be explicitly written as: \^[(1)]{}\_[P.S]{}=-ie ( {-2,2}(-2)(2) +{-3,3}(-1)(1)) \[omega1\] , with the definition: {-n,+n}= \_[b0\^+]{}\^[x+na-b]{}\_[x-na+b]{}dy\^ \_, \[chaven\] where $(-\infty,y)$ stands for $\int^{y}_{-\infty}\,
A^{\nu}(\eta)d\eta_{\nu}$.\
Therefore, the split Lagrangian, whose action is invariant under $\gt$ up to first order, ${\cal L}^{(1)}_{P.S}$, is the sum of ${\cal L}^{(0)}_{P.S}$ and $ -\frac{\lambda}{4}\Omega^{(1)}_{P.S}$ (eqs. (\[L0ps\]) and (\[omega1\])).\
\
It is precisely in this sense that gauge invariance has to be taken in the framework of generalized gauge transformations, ggt’s. Actually, since the ggt’s themselves take the form of an infinite series in the gauge coupling constant, then it is expected that the split (and regularized) action also presents a similar form, with its “[*generalized gauge invariance*]{} being constructed and checked order by order.\
\
Now, calculating $\gt{\cal L}^{(1)}_{P.S}$ at order $e^2$, we get (after suitable change of variables in the action form of the terms): ((D\_)\^(D\^)\_[P.S]{})\_[e\^2]{}= (ie)\^[2]{}(x) (4) , (with $U\stackrel{\leftrightarrow}{\partial} V = U\partial
V-(\partial V)U$). Again, we cannot set this term to zero. Instead, according to $\Omega^{(1)}_{P.S}$, we must search for a new term, $ \Sigma^{(2)}_{P.S}$, such that $(D_{\mu}\fy)^{\dagger}(D^{\mu}\fy)_{P.S} + \Sigma^{(2)}_{P.S}$ be invariant under $\gt$ at least up to order $e^2$. Such term can be found and its simplest form is: $$\Sigma^{(2)}_{P.S}=-(ie)^2 \,\Sigma_{\mu}\, \left[\fyd(x)
\stackrel{\leftrightarrow}{\partial^{\mu}} \fy(4)\right],$$ with $\Sigma_{\mu}$ being a function of $\Lambda$ and $A_{\mu}$. In fact, $\Sigma_{\mu}$ must be an object such that $\gt\Sigma_{\mu}= \Lambda(1)
A_{\mu}(3)-\Lambda(3)A_{\mu}(1)$. It is easy to check that the following expression satisfies such a requiriment: \^[(2)]{}\_[P.S]{}=-(ie)\^2{\[A\_(1)+A\_(3)\](1,3) +(\[1\]+\[3\]) \_[x+a]{}\^[x+3a]{} F\_()d\^}(x) (4) \[sigma2\] , with $[\pm n]= \frac12 \left[ (-\infty,\pm n) + (\infty,\pm n)\right]$. In addition, we may see that as $a\to 0$ then $ \Sigma^{(2)}_{P.S}$ vanishes. \[The expression inside $\{\,\}$ was already obtained in Ref.[@GNW] for the interacting vertex of QED; for details, see eq. (24) in that paper\][^6].\
\
Now, for the self-interaction sector, we get: & & (()\^[2]{}\_[P.S]{}+ \^[(1)]{}\_[P.S]{}) \_[e\^2]{}= 12 (ie)\^[2]{}{2{-2,2} .+\
& & .+2{-3,3} +.\
& &.-\[(-2)+(-4)\] (-2,-4)+.\
& & . +\[(2)+(4)\](2,4).+\
& & . -2(4) (-,2).+\
& & . -2(-4) (-,-2).+\
& & . +2(-3)(3)(-2)(2) +.\
& & . +2(-4)(4)(-1)(1)}.\[eq16\] The non-vanishing of this term is evident. Searching for a new term, $ \Omega^{(2)}_{P.S}$, such that $(\fyd\fy)^{2}_{P.S} + \Omega^{(1)}_{P.S}+
\Omega^{(2)}_{P.S}$ be invariant under $\gt$ at least up to order $e^2$, is more difficult than for the former ones, $ \Omega^{(1)}_{P.S}$ and $\Sigma^{(2)}_{P.S}$. Such difficulties arise from its rather complicated structure. Fortunately, an explicit expression may be indeed found. For that, we notice that the six last terms above have similar structure, say, $\Lambda(\pm n) \, (\pm m,\pm p) \,\fyd\fy\fyd\fy$-type factors. Actually, for such terms, the simplest $ \Omega^{(2)}_{P.S}$-type ‘[*counter-term*]{}’ has the general form:$$\frac12 (ie)^{2}\left(\frac12 \frac{\Lambda(\pm n)}{\Lambda(\pm p)
-\Lambda(\pm m)}\, (\pm m,\pm p)^{2}\right).\\$$ By remembering the definitions of the above quantities, it is easy to see that such a expression vanishes as $a\to 0$.\
On the other hand, for the first two terms in eq. (\[eq16\]), those proportional to $\{-n,+n\}$, the task of finding $ \Omega^{(2)}_{P.S}$-type [*counter-terms*]{} appear to be very easy if we take into account that:$$\gt\{-n,+n\}\, \arrowvert_{e^0}=
\Lambda(-n)\fyd(-n-1)\fy(n-1) -\Lambda(n)\fyd(-n+1)\fy(n+1).$$In fact, as it can be readily checked, those first two terms have the following $ \Omega^{(2)}_{P.S}$-type [*counter-term*]{}:[^7]$$\frac12 (ie)^2 \left( -2 \,\{-2,+2\}\,
\{-3,+3\}\right).\\$$ Therefore, the full $ \Omega^{(2)}_{P.S}$-term takes over the form: & \^[(2)]{}\_[P.S]{}= &-12 (ie)\^2 \[omega2\]. Finally, the ${\cal L}^{(2)}_{P.S}$ Lagrangian, whose action is invariant under $\gt$ up to order $e^2$, may be written as: \^[(2)]{}\_[P.S]{}=[L]{}\^[(0)]{}\_[P.S]{}+ \^[(2)]{}\_[P.S]{} - (\^[(1)]{}\_[P.S]{}+ \^[(2)]{}\_[P.S]{}) \[L2ps\], with the expressions for the above terms being given by (\[L0ps\]), (\[omega1\]), (\[sigma2\]), and (\[omega2\]).\
\
The form of eq. (\[L2ps\]) deserves further remarks. The explicit presence of the gauge parameter, $\Lambda$, in eq. (\[omega2\]), may seem to be spurious, since it is well-known that (usual) gauge invariance is explicitly broken by the presence of the gauge parameter in the action. Actually, such a symmetry is broken whenever it is taken in the usual sense, but by reassessing the meaning of gauge invariance in the context of ggt’s, then the scenario may be changed. As we have already pointed out, gauge invariance has in this framework to be constructed and checked by means of an order by order (in the gauge coupling constant) algorithm.\
\
Therefore, when finding out a split action, this has to be also done order by order and its ‘[*generalized gauge invariance*]{}’ must be verified according such a “perturbative” procedure. Hence, if we wish to verify whether a given split action is invariant under ggt’s, say, up to 2nd order, for concreteness, then we must check: first, if as the splitting parameter vanishes, $a\to0$, the split action restores the original one; and, if the split action is actually invariant under ggt’s up to 2nd order, then $\gt S^{(2)}_{P.S.}={\cal O}(e^3)$.\
\
In our present case, eq.(\[L2ps\]), both requirements are verified, even though the gauge parameter is explicitly present in its expression. Whether other good split actions without explicit presence of such a parameter may be found for $\lambda\fy^4$-type vertices is not so clear to us. Indeed, in the case of 3-leg vertices, like as $\fyd A_\mu\gamma^\mu\fy$, we have found a $\Lambda$-explicitly dependent action which was shown to satisfy both of the requiriments above (see footnote that follows eq.(\[sigma2\])).\
Applications to some self-interacting models
============================================
Here, in order to illustrate the applicablity of our results, we shall deal with some $\lambda\varphi^4$-type models. Whenever necessary, we shall pay attention to specific points which were not still presented.\
\
i) [**The $(\overline{\psi}\psi)^2$ model**]{}\
The model which will be worked out is described by the following Lagrangian:[^8] \_(x)=-14 F\^F\_ +( iD\_\^ -m\_f)-g()\^2 \[Lpsi\] , with $D_{\mu}$ and $F_{\mu\nu}$ previously defined.\
\
Here, due to the anti-commutative character of fermionic fields, we must pay special attention in changing the order of such fields. Moreover, the kinetic term is slightly different from that for scalar field and it must be taken apart. Fortunately, such a term was already studied in Ref.[@GNW] and, if we perform the following splitting: &(x)iD\_\^(x) &(iD\_\^)\_[P.S]{}=\
& & (x)i\_\^(x) -e(x-a) A\_(x)\^(x+a) , one can readily show that $\int {d^{4}x}\,(\overline{\psi}\,iD_{\mu}
\gamma^{\mu}\psi)_{P.S}$ is invariant up to order $e$. At second order, such a variation does not vanish, but it is exactly canceled by the following term (quite similar to eq.(\[sigma2\]) above; see also eq. (24) in Ref.[@GNW]): \^[(2)]{}\_[,P.S]{}= - (-2) \^ (2) {\[A\_(-1)+A\_(1)\](-1,+1)+(\[-1\]+\[1\])\^[x+a]{}\_[x-a]{}d \^F\_()} \[sigma2psi\] . Now, the $(\overline{\psi}\psi)^2$-term is split in the same way as $(\fyd\fy)^2$: $$(\overline{\psi}(x)\psi(x))^2\stackrel{P.S}{\longrightarrow}
\,(\overline{\psi}\psi)^2_{P.S}= \overline{\psi}(-1)\psi(1)\overline{\psi}
(-2)\psi(2).$$\
So, ${\cal L}^{(0)}_{P.S}$ for $\psi$-like fields reads: \^[(0)]{}\_[,P.S]{}=-14 F\^(x)F\_(x) -m\_f (x)(x) +i(D\_\^)\_[P.S]{} -g()\^2\_[P.S]{} \[L0psi\] . To get ${\cal L}^{(2)}_{\psi,\,P.S}$, we may use the $\Omega^{(1)}_{P.S}$ and $\Omega^{(2)}_{P.S}$ obtained in the previous section with suitable change of $\fy$ by $\psi$ and $\fyd$ by $\overline{\psi}$. Indeed, as we kept the original order of those matter fields in the previous results, we may write: \^[(1)]{}\_[,P.S]{}=\^[(1)]{}\_[P.S]{}\_[,]{}\^[(2)]{}\_[,P.S]{}= \^[(2)]{}\_[P.S]{}\_[,]{}\[omegaspsi\] . Finally, we get: \^[(2)]{}\_[,P.S]{}=[L]{}\^[(0)]{}\_[,P.S]{} +\^[(2)]{}\_[,P.S]{} +g(\^[(1)]{}\_[,P.S]{}+\^[(2)]{} \_[,P.S]{})\[L2psi\].\
ii)[**The ‘Avdeev-Chizhov’ model**]{}\
Some years ago, Avdeev and Chizhov[@AC] proposed an Abelian model which includes antisymmetric rank-2 real tensors that describe matter, rather than gauge degrees of freedom. They are coupled to a usual vector gauge field as well as to fermions. The model was shown to reveal interesting properties: for instance, these new matter fields were shown to play an important role in connection with extended electroweak models in order to explain some observed decays like $\pi^-\to e^- +\overline{\nu}+\gamma$ and $K^+\to \pi^0 +e^+ +\nu$ [@Chi], and a classical analysis of its dynamics has shown that some longitudinal excitations may carry “physical degrees of freedom ” (see Ref.[@AC2], for further details). In addition, some works have been devoted to the study of its supersymmetric generalization [@NogCia], as well as its connection with non-linear sigma models [@NPFH].\
\
Starting off from these interesting features, it was shown that the coupling between tensorial and fermionic fields generates anomalies in the quantized version of the model and could also spoil its renormalizability [@LRS].The removal of the fermions has the additional usefulness of allowing us to write the new Lagrangian in a shorter form by means of complex field tensors, $\fy_{\mu\nu}$ and $\fyd_{\mu\nu}$ [@LRS2]. Thus, the [*modified*]{} Avdeev-Chizhov model reads: \_[AC]{}(x)= -14 F\_F\^ +(D\_\^) (D\^\_)\^ -\_ \^\_\^ \[Lac\], with $D_\mu$ and $F_{\mu\nu}$ already defined. Once $\fy_{\mu\nu}$ is taken to satisfy a complex self-dual relation: $$\fy_{\mu\nu}(x)=+i\tilde{\fy}_{\mu\nu}(x)
\, , \qquad \tilde{\fy}_{\mu\nu}=\frac12 \epsilon_{\mu\nu\alpha\beta}\fy^
{\alpha\beta},$$then it can be split into two (real tensors) parts:$$\fy_{\mu\nu}
(x)=T_{\mu\nu}(x) +i \tilde{T}_{\mu\nu}(x)\quad\mbox{and}\quad \fyd_{\mu\nu}(x)=
T_{\mu\nu}(x) -i \tilde{T}_{\mu\nu}(x),$$where $T_{\mu\nu}$ and $\tilde{T}_{\mu\nu}$ are real and antisymmetric fields. Actually, regarded as matter fields, it is known that they describe spin$\_$0 excitations (see, for example, Refs. [@AC2; @nogue]).\
\
Now, performing similar splittings in ${\cal L}_{AC}(x)$, as we have done in former cases, we get, after some calculation, ${\cal
L}^{(2)}_{P.S}(\fy^{\mu\nu})$: [^9] \^[(2)]{}\_[P.S]{}(\^)=[L]{}\^[(0)]{}\_[P.S]{}(\^) +\^[(2)]{}\_[P.S]{}(\^) - (\^[(1)]{}\_[P.S]{} (\^)+ \^[(2)]{}\_[P.S]{}(\^)) where the terms above have the following expressions: \^[(0)]{}\_[P.S]{}(\^)=-14 F\_F\^ +(D\_\^)\_[P.S]{}(D\^\_)\^\_[P.S]{} -(\_\^\_ \^)\_[P.S]{} (with the splittings previously performed); &\^[(2)]{}\_[P.S]{}(\^)=-(ie)\^2 &A\^(1)-12(\^ (3)-\^(1)) \]\
& & (its slight difference with respect to $\Sigma^{(2)}_{P.S}$, eq. (\[sigma2\]), is due to the tensor indices); \^[(1)]{}\_[P.S]{}(\^)=(-ie)( {-2,+2}\^\_\_(-2)\^(2) +{-3,+3}\^\_\_(-1)\^(1)), and $\Omega^{(2)}_{P.S}(\fy^{\mu\nu})$ which is easily obtained from $\Omega^{(2)}_{P.S}$ by making the interchanges: & & {-2,+2}{-3,+3}{-2,+2}\^\_{-3,+3}\^\_\
& & \_\^\_ \^, where we have defined $\{-n,+n\}_\mu\mbox{}^\nu$ in the same way as $\{-n,+n\}$, with $\fyd\fy$ changed by $\fyd_{\mu\alpha}\fy^{\alpha\nu}$ in its definition, eq. (\[chaven\]).\
\
Usually, the 2-form field, $\varphi_{\mu\nu}$, is treated as a gauge potential (the so-called Cremmer-Scherk-Kalb-Ramond field [@CS; @KR]), and in (3+1) dimensions it describes a massless scalar excitation. However, when mixed to the Maxwell field, $A_\mu$, by means of a topological mass term that linkes two Abelian factors, it yields a massive spin$\_$1 excitation in the spectrum. On the other hand, we should stress that its coupling to charged matter may be realized only non-minimally, i.e., by means of its (3-form) field-strength [@nossos].\
\
Furthermore, from the point of view of the point-splitting procedure, since its usual gauge transformation, $\delta\varphi_{\mu\nu}(x)=\partial_\mu\xi_\nu(x)
-\partial_\nu\xi_\mu(x),$ does not involve products of quantities at the same space-time point, such a tranformation does not undergo any change in going to the generalized case, say, $\delta_g\varphi_{\mu\nu}(x)=\delta\varphi
_{\mu\nu}(x)$. Therefore, by viewing $\varphi_{\mu\nu}$ as a gauge potential, its ggt’s take easier expressions than when it is treated as a matter field (a general fact, at least in the Abelian framework, in dealing with such a procedure). This readily implies remarkable simplifications whenever working with $\varphi_{\mu\nu}$ as a gauge potential.\
Concluding Remarks
==================
The point-splitting procedure combined with generalized gauge transformations has yielded regularized Lagrangians which contains $\lambda\fy^4$-type interaction. The result is such that the generalized Lagrangians have their interacting terms defined at different space-time points. Nevertheless, this property introduces non-locality at the level of the regularized theory.\
\
In general, non-local theories cannot be quantized with the usual methods and the interpretation of their results are not quite obvious. Moreover, we know that non-locality can lead to troubles as long as the causality of the theory is concerned. However, these problems arise only for the regularized theory, in much the same way as ghosts are present and unitarity is temporarily lost for regularized theories before the regularization parameter is removed.\
\
Nevertheless, Osland and Wu [@OW] obtained some standard results in QED starting by a split Lagrangian (with regularity and non-locality properties)[^10]. Their method works for the calculation of the quantities with a dependence in the splitting parameter, which is set to zero at the end of calculations in order to get the standard results.\
\
What we may learn from these calculations is that, when point-splitting is combined with generalized gauge transformations in order to obtain regularized (Abelian) Lagrangians, the task becomes more difficult with the increasing of the number of matter fields at the same vertex; in general, the complications which arise from the presence of extra (Abelian) gauge fields are minor ones. So, the calculations involving $\lambda\fy^4$-type vertices are harder to be performed than for ‘lower vertices’, $\fy A_\mu\fy$, $\fy A_{\mu}A^{\mu}\fy$, and so forth. In addition, higher-order terms in the coupling constant are, in general, more complicated to be handled than lower ones.\
\
Our present study is a good example of such complications and the reason why they arise. For instance, in dealing with the 2nd order calculations of the split version of the self-interacting vertex we have seen that a [*counter-term*]{} for that term involved the explicit presence of the gauge parameter, scenario which is expected to become even more intricate at higher orders.\
\
Another point that should be stressed is that this procedure is independent of the space-time dimension, and so, of the canonical dimension of the fields (matter or gauge ones)[^11]. Hence, the expressions for our $\Sigma$ and $\Omega$ terms remain valid in other dimensions. Therefore, our present results could be equally well applied to four-matter (scalar, spinorial, and rank-2 tensorial) vertices in lower or higher space-time dimensions. However, special attention should be paid if dimensional reduction and/or compactification, spontaneous symmetry breaking, or other mechanisms are involved. As we shall discuss below for the case of the Higgs model, the present procedure is suitably applied only at the stage in which the true physical excitations are taken into account.\
\
On the other hand, if we are dealing with a renormalizable theory (scalar, for simplicity) in $(2+1)$ dimensions, an extra $f\fy^6$-term is allowed. In this case, our results could be applied to the model, including the $\lambda\fy^4$-term, but the extra term should be worked out apart. As we have already said, the task of working out the split version of a matter vertex tends to become more difficult as the number of matter fields increases. Thus, we expect even more work in dealing with $\fy^6$-like matter vertices than we had in the present case. Still concerning possible relevant applications in this space-time, we may think of applying the present procedure for studying some Abelian (and non-Abelian in a further stage, too) models connected with the Chern-Simons term. For instance, we may study some points concerning the radiation produced by accelerated point-like charges[@inprogress], an issue which still demands several answers (see Ref.[@teseprd], for more details).\
\
Another relevant question that we may rise up here is the issue of the point-splitting in connection with the Abelian Higgs mechanism. The spontaneous symmetry breaking, as realized by a charged scalar, obliges a shift of the Higgs field around its vacuum expectation value and induces the appearance, among others, of trilinear matter couplings not present in the original action. Here, we may wonder whether the splitting should be performed before or after the breaking. Indeed, although our results are directly applied to the unbroken phase, it does not seem to be the best choice. Actually, we claim that the most suitable way to implement point-splitting is after the breaking takes place, for in the broken regime all possible vertices show up and we can really control the theory we are dealing with, since only at this stage we are quantizing the truly physical excitations. In this case, while our results are applicable to some terms of the Lagrangian written around the true ground state, such as the kinetic and quartic ones, the trilinear vertex, in turn, should be worked out apart (expected to be of easier manipulation than the present one).\
\
We also hope that the present paper could help us whenever dealing with the non-Abelian case. In this scenario novel features will arise mainly because $\gt A^{a}_{\mu}$ will take more complicated (and lengthier) forms, and they will imply in new ggt’s for the matter fields which, in turn, will also take lengthier expressions than those for the Abelian case.\
\
Furthermore, in view of the special role that supersymmetry and supersymmetric gauge theories play in the programme of building up fundamental interaction models, it would be advisable to extend the point-splitting method to treat supersymmetric theories in superspace. Point-split super-actions both in the space-time and Grassmann coordinates may be an interesting issue since now gauge invariance and supersymmetry must be simultaneously checked and many features of the method must be revealed: the advantage of implementing the point- -splitting procedure in superspace is that supersymmetry is manifest and one needs not checking Ward identities (as it would be the case in a component-field approach) to undertake that supersymmetry is kept upon point-splitting.\
\
we claim that some questions concerning this issue should eventually become clearer. For example, how could Feynman rules for such a kind of Lagrangian be formulated? Or still, as we may see, there are some new ‘ interaction terms’ in the generalized Lagrangian. Could these new terms have some physical interpretation and/or relevance?\
\
**Acknowledgments**
The authors are grateful to Dr. A.L.M.A. Nogueira for useful discussions concerning the Avdeev-Chizhov model. WAMM is grateful to CNPq and FAPEMIG for the financial support. JAHN thanks CNPq for partial financial support.\
[99]{} P.A.M. Dirac, Proc. Cambridge Phil. Soc (1934)150; P. Osland and T.T. Wu, Zeit. Physik (1992)569; P. Osland and T.T. Wu, Zeit. Physik (1992)585, 593; Phys. Lett. (1992)315; R. Gastmans, C. Newton and T.T. Wu, Phys. Lett. (1994)84; R. Gastmans, C. Newton and T.T. Wu, Phys. Rev. (1996)5302; R. Gastmans and T.T. Wu, Phys. Rev. (1998)1203; L.V. Avdeev and M.V. Chizhov, Phys. Lett. (1994)212; M.V. Chizhov, Mod. Phys. Lett. (1993)2752; L.V. Avdeev and M.V. Chizhov, hep-th/9407067; A.L.M.A. Nogueira, M.Sc. Thesis (CBPF, 1996) \[unpublished\]; V. Lemes, A.L.M.A. Nogueira and J.A. Helayël-Neto, Int. J. Mod. Phys. (1998)3145; M. Negrão, A. Penna-Firme and J.A. Helayël-Neto, hep-th/9808174; V. Lemes, R. Renan and S.P. Sorella, Phys. Lett. (1995)158; V. Lemes, R. Renan and S.P. Sorella, Phys. Lett. (1995)37; A.L.M.A. Nogueira’s M.Sc. Thesis in Ref. [@NogCia]. W.A. Moura-Melo and J.A. Helayël-Neto, in Proc. XVII Brazilian National Meeting Particles and Fields (Soc. Bras. Física, 1996) 643; C. Cremmer and J. Scherk, Nucl. Phys. (1974)117; M. Kalb and P. Ramond, Phys. Rev. (1974)2273; W.A. Moura-Melo, N. Panza and J.A. Helayël-Neto, Int. J. Mod. Phys. (1999)3949; [*ibid*]{} in Proc. XVIII Brazilian National Meeting Particles and Fields (1997)146; W. A. Moura-Melo, M.Sc. Thesis (CBPF, 1997) \[unpublished\]. W.A. Moura-Melo and J.A. Helayël-Neto, work in progress. W. A. Moura-Melo and J.A. Helayël-Neto, Phys. Rev. (2001)065013; W.A. Moura-Melo, Ph.D. Thesis, CBPF, 2001 \[unpublished\].
[^1]: Email: [email protected], [email protected]
[^2]: Email: [email protected].
[^3]: We shall use Minkowski metric $diag(\eta_{\mu\nu})=(+,-,-,-)$ and greek letters running 0,1,2,3.
[^4]: For further applications to fermionic fields, the Hermitian conjugation must be changed to Dirac conjugation. On the other hand, if the matter fields are rank-2 tensors, then additional attention must be paid to their indices. See Section III for details. In addition, in dealing with the actual Abelian Higgs model a important question that now takes place is whether the present scheme is more suitable applied before or after the spontaneous symmetry breaking be performed. Such a point will be discussed later (see Concluding Remarks, for more details).
[^5]: In the Abelian case, $\gt A_{\mu}=\delta A_{\mu}$ holds, but in the non-Abelian scenario, where the ordinary gauge transformation for $A^{a}_{\mu}$ involves products at the same point, the point-splitting will act on it, and its non-Abelian ggt’s will be different from the usual one.Indeed, such ggt’s were already worked out for $SU(2)$ [@GNW2], and more recently for $SU(N)$ [@GWu].
[^6]: An alternative, but apparently non-equivalent, form for $\Sigma^{(2)}_{P.S}$ was obtained in Ref.[@cax96] and reads:$$-(ie)^{2}\left\{[1]A_{\mu}(3)
-[3]A_{\mu}(1)
-\frac12 \left( \frac{[1]^2}{\Lambda(1)}\partial_{\mu}\Lambda(3) -
\frac{[3]^2}{\Lambda(3)}\partial_{\mu}\Lambda(1) \right)\right\}
\fyd(x)\stackrel{\leftrightarrow}{\partial^{\mu}} \fy(4).$$Despite the explicit presence of the gauge parameter in the expression above, it can be verified that it vanishes as $a\to0$ and leads us to a split action invariant under ggt’s up to the 2nd order.
[^7]: In fact, if $\fy$ (and $\fyd$) are fermionic fields, then $\fy$ (or $\fyd$) has anticommutative property, but the bilinear $\fyd\fy$ has commutative behavior. Therefore, even for fermionic fields, we can change the order of $\{-2,+2\}$ by $\{-3,+3\}$ and vice-versa, without any extra minus sign.
[^8]: It is worthy noticing the (power-counting) non-renormalizability of this self-interaction vertex: $[g]=[mass]^{-1}$ in (3+1) dimensions.
[^9]: We have already found a similar expression for this model in Ref.[@cax96]. There, a slightly modified splitting was employed, as well as a lengthier form for $\Sigma^{(2)}_{P.S}(\fy^{\mu\nu})$.
[^10]: It is worthy noticing that their Lagrangian (eq. (2.7) in Ref.[@OW]) is different from the ‘[*correct*]{}’ generalized QED-Lagrangian, up to fourth order (eq. (24) in Ref.[@GNW]). Such a difference may be explained by noticing that, in Ref. [@OW], the generalized gauge covariance is not taken in its precise meaning.
[^11]: In fact, the ggt’s depend on the splitting parameter, the constant vector $a_\mu$, and on the Abelian (or non-Abelian) character of the gauge fields.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The ground state of the $S=1$ antiferromagnetic Heisenberg chain belongs to the Haldane phase – a well known example of symmetry-protected topological phase. A staggered field applied to the $S=1$ antiferromagnetic chain breaks all the symmetries that protect the Haldane phase as a topological phase, reducing it to a trivial phase. That is, the Haldane phase is then connected adiabatically to an antiferromagnetic product state. Nevertheless, as long as the symmetry under site-centered inversion combined with a spin rotation is preserved, the phase is still distinct from another trivial phase. We demonstrate the existence of such distinct symmetry-protected *trivial* phases using a field-theoretical approach and numerical calculations. Furthermore, a general proof and a non-local order parameter are given in terms of an matrix-product state formulation.'
author:
- Yohei Fuji
- Frank Pollmann
- Masaki Oshikawa
title: |
Distinct Trivial Phases Protected by a Point-Group Symmetry\
in Quantum Spin Chains
---
*Introduction.—*While symmetry broken phases can be completely classified using the Landau theory, there still exists no exhaustive understanding of topological quantum phases. Topological quantum phases are gapped phases of matter that are distinct from trivially disordered states but cannot be characterized by any local order parameter. Over the past few years, new theoretical frameworks have been developed to understand and classify many different topological phases. For example, topological phases of noninteracting fermions are now completely classified using K-theory [@Kitaev09; @Wen12]. More generally, two gapped ground states belong to the same phase if and only if they can adiabatically connected with respect to local Hamiltonians [@Verstraete05; @XChen10; @Schuch11]. Even when different states are connected by a general adiabatic process, it is possible that they can no longer be adiabatically connected if we impose symmetries on the Hamiltonian. These are either states with spontaneous symmetry breaking or belong to the class of symmetry-protected topological (SPT) phase [@ZCGu09; @Pollmann10; @XChen10; @XChen11a; @XChen11b; @Schuch11; @XChen13]. Examples of SPT phases include topological insulators [@Kane05a], which are protected by time reversal symmetry, and the Haldane phase [@Haldane83a; @*Haldane83b; @Affleck87a; @*Affleck88] in one dimension, which is protected by either one of the time reversal, bond-centered inversion, or the dihedral group of the spin rotations [@Pollmann10]. In these examples, SPT phases are nontrivial in the sense that they cannot be adiabatically connected to a trivial product state, once an appropriate symmetry is imposed. They also support gapless edge states and/or nontrivial degenerate structures in the entanglement spectrum.
While the notion of the SPT phases is now established and widely recognized, in this work we demonstrate that site-centered inversion symmetry allows to distinguish different trivial one-dimensional phases. That is, there are multiple “symmetry-protected *trivial*” (SPt) phases, i.e., symmetric phases connected adiabatically to product states, which are still distinct in the presence of the imposed symmetry. We note that, the word “symmetry-protected trivial phase” is sometimes used in place of the standard terminology of SPT (symmetry-protected topological) phase, because the entanglement in such a phase is short-ranged and is removable in an adiabatic process if the symmetry is disregarded. In contrast, in what is called an SPt phase in this work, the entanglement can be completely removed adiabatically, even in the presence of the imposed symmetry, to reduce the state to a product state. However, it is still distinct from another trivial phase. The SPt phases introduced here represent a new class of 1D quantum phases that are protected by a point-group symmetry but not captured by the cohomology classification. While each of these phases is trivial by itself, the quantum phase transition between them is experimentally detectable (e.g., by a divergence of some susceptibilities). Moreover, we derive non-local order parameters that could be used to characterize each phase. While the concept of SPt phases is very general, we illustrate it for clarity with a simple model of spin-1 chain in the following.
*The model.—*In order to make the discussion concrete, let us first consider the following simple model of $S=1$ chain: $$\begin{aligned}
{\mathcal{H}}& = \sum_i \left[
\vec{S}_i \cdot \vec{S}_{i+1} +
D_z (S^z_i)^2 - h_z (-1)^i S^z_i \right] .
\label{eq.Ham}\end{aligned}$$ The first term is the standard Heisenberg model with antiferromagnetic exchange interactions which stabilize the celebrated Haldane gap [@Haldane83a; @*Haldane83b]. The $D_z$ term is the uniaxial single-ion anisotropy, which is commonly present in magnetic ions with $S=1$ such as Ni$^{2+}$. The model with uniaxial anisotropy $D_z$ and $h_z =0$ is well understood [@denNijs89; @Tasaki91; @WChen03]. For small $D_z \ge 0$, the system is in the Haldane phase and undergoes a quantum phase transition into the “large-$D$” phase at $D_z\approx1$. Both phases are gapped and have the full symmetry of the Hamiltonian. The Haldane phase is a well-known example of SPT phases [@Pollmann10; @Pollmann12a] as discussed in the introduction. The large-$D$ phase is a trivial phase which is adiabatically connected to the product state $\left| D \right\rangle = \left| \cdots 0000 \cdots \right\rangle$, where $|0\rangle$ represents the local spin state with $S^z=0$. This state is the exact ground state of the Hamiltonian in the limit $D_z\rightarrow\infty$. The $h_z$ term represents a staggered field, which occurs in many quasi-one-dimensional materials, including Haldane gap systems, with an alternating crystal structure under an applied (uniform) field. For simplicity, we only include the staggered field term without the uniform one.
In the limit $h_z \to \infty$, the spins are fully polarized along the staggered field, and the ground state is reduced to another trivial product state $\left| N \right\rangle = \left| \cdots +-+- \cdots \right\rangle$, where $+$ and $-$ represent the local spin states with $S^z=+1$ and $S^z=-1$, respectively. It was recognized earlier that there is no phase transition for $0 < h_z < \infty$ (for $D_z =0$) [@Tsukano98a]. That is, the Haldane phase is adiabatically connected to the Néel state $|N\rangle$ with imposed antiferromagnetic (AF) order. In the SPT framework, this is naturally understood; since the staggered field breaks all the symmetries that protect the Haldane phase as an SPT phase, it reduces the Haldane phase to a trivial phase which also includes the Néel state $|N\rangle$.
Now let us discuss the model with both $D_z$ and $h_z$. In fact, this model has been studied in Ref. [@Tsukano98b] by a field theory and numerical methods where a quantum phase transition between the large-$D$ phase and the imposed AF phase was found. This is rather surprising, since both phases are trivial and are adiabatically connected to product states $|D\rangle$ and $|N\rangle$ that have the full symmetry of the Hamiltonian. It is perhaps even more surprising in the light of the recent concept of the SPT phases, where the existing classification scheme [@XChen11a; @XChen11b; @Schuch11] would not distinguish them. While the nature of the phase transition were studied in Ref. [@Tsukano98b], why (and when) these two trivial phases are distinguished was not completely clarified. In the remainder of this paper, we demonstrate that this is an example of distinct SPt phases and identify the symmetry which protects them.
*Bosonization.—* The standard bosonization procedure of $S=1$ chains starts from two coupled $S=1/2$ chains, and the low-energy effective field theory for ${\mathcal{H}}$ is given by the Hamiltonian [@Schulz86; @Berg08; @Berg11] $$\begin{aligned}
H_\textrm{eff} &= \frac{v}{2\pi} \int dx \; \left[ K (\partial_x
\theta)^2 +\frac{1}{K} (\partial_x \phi)^2 \right] \notag \\ & + {g_\textrm{eff}}\int dx \cos{(2\phi)}, \label{eq:EffHam}\end{aligned}$$ where $\phi$ and $\theta$ are dual field of each other, satisfying $[\phi(x), \theta(x')] = i(\pi/2)[\textrm{sgn} (x-x')+1]$. The Hamiltonian represents the so-called sine-Gordon field theory, which is ubiquitous in many problems in $1+1$ dimensions. Its properties essentially depend on the coupling constant $K$. When $K>2$, the coupling ${g_\textrm{eff}}$ is irrelevant under the renormalization group (RG), and the system is renormalized in the low-energy limit into the free boson theory with ${g_\textrm{eff}}=0$, which is nothing but a gapless Tomonaga-Luttinger liquid (TLL). On the other hand, if $K<2$ and ${g_\textrm{eff}}$ is non-vanishing, the interaction is RG-relevant, and the system acquires an excitation gap. In the absence of the staggered field, the Haldane and the large-$D$ phases correspond to ${g_\textrm{eff}}>0$ and ${g_\textrm{eff}}<0$ respectively, both with $K<2$. It is easy to see that, within the effective Hamiltonian , the two phases with ${g_\textrm{eff}}>0$ and ${g_\textrm{eff}}<0$ are always separated by the critical point ${g_\textrm{eff}}=0$. This actually comes from the fact that $\cos(2\phi)$ is the only interaction compatible with the symmetry and the compactification, $\phi \sim \phi +\pi$ and $\theta \sim \theta+2\pi$, up to subleading terms $\cos(2n\phi)$ $(n \geq 2)$. In general, the effective theory can also have the $\sin{(2\phi)}$ term, which can be combined with the $\cos{(2\phi)}$ term as $\cos{(2 \phi + \alpha)}$ with a phase shift $\alpha$. It is clear that, by changing $\alpha$ from $0$ to $\pi$, the two phases with ${g_\textrm{eff}}>0$ and ${g_\textrm{eff}}<0$ are adiabatically connected without closing the gap [@Berg08]. Thus, for the two phases to be distinct, $\sin{(2\phi)}$ has to be forbidden by some symmetry.
Symmetry operation Symbol Transformation for spins Transformation for fields $(\phi, \theta)$
------------------------------- ------------------- ----------------------------------------------------------------- --------------------------------------------------------------------------------
Bond-centered inversion ${\mathcal{I}}_b$ $\vec{S}_i {\rightarrow}\vec{S}_{1-i}$ $\phi (x) {\rightarrow}-\phi (-x)$, $\theta (x) {\rightarrow}\theta (-x) +\pi$
Site-centered inversion ${\mathcal{I}}_s$ $\vec{S}_i {\rightarrow}\vec{S}_{-i}$ $\phi (x) {\rightarrow}-\phi (-x) +\pi$, $\theta (x) {\rightarrow}\theta (-x)$
$\pi$ rotation about $z$ axis ${\mathcal{R}}_z$ $S^{x,y}_i {\rightarrow}-S^{x,y}_i$, $S^z_i {\rightarrow}S^z_i$ $\phi {\rightarrow}\phi$, $\theta {\rightarrow}\theta +\pi$
In fact, in the framework of bosonization, this is how a symmetry protects the Haldane phase as an SPT phase which is distinct from the trivial large-$D$ phase. Any of the three symmetries, which are known to protect the Haldane phase, forbids the $\sin{(2\phi)}$ interaction [@Fuji14]. Here, for brevity, among these three symmetries, we only show the representation of the bond-centered inversion ${\mathcal{I}}_b$ in terms of the bosonic field $\phi$, $\theta$ in Table \[table:SymOp\]. The action of ${\mathcal{I}}_b$, $\phi(x) \to - \phi(-x)$, forbids $\sin{(2 \phi)}$, which leads to the distinction of the two phases with ${g_\textrm{eff}}>0$ and ${g_\textrm{eff}}<0$. On the other hand, Table \[table:SymOp\] shows that the site-centered inversion ${\mathcal{I}}_s$ also has the same action on $\phi$, forbidding $\sin{(2\phi)}$. However, ${\mathcal{I}}_s$ by itself is not sufficient to keep the distinction between the two phases; they are adiabatically connected without gap closing, because of the vertex operators $e^{\pm i \theta}$ allowed in the absence of the U(1) symmetry of spin rotation about $z$ axis. According to Table \[table:SymOp\], the combined operation ${\mathcal{I}}' = {\mathcal{I}}_s \times {\mathcal{R}}_z$ (${\mathcal{R}}_z$ is the global $\pi$-rotation about $z$ axis), as well as ${\mathcal{I}}_b$, forbids $e^{\pm i \theta}$. They still allow the next leading ones $e^{\pm 2 i \theta}$. Nevertheless, they just replace the direct transition between the ${g_\textrm{eff}}>0$ and ${g_\textrm{eff}}<0$ phases by an intermediate phase with spontaneous $\mathbb{Z}_2$-symmetry breaking. Thus ${\mathcal{I}}'$ alone should still maintain the distinction between two SPt phases.
*Numerical results.—*The above bosonization analysis suggests that we can introduce microscopic models with less symmetries than Eq. , but with the symmetry under ${\mathcal{I}}'$, to maintain the two distinct phases. As an example, we consider the following Hamiltonian: $$\begin{aligned}
{\mathcal{H}}' = {\mathcal{H}}+ \sum_i d_x \left( S^y_i S^z_{i+1} - S^z_i S^y_{i+1} \right)
\label{eq:Spin1Model}\end{aligned}$$ The new term $d_x$ represents the (uniform) Dzyaloshinskii-Moriya (DM) interaction with the DM vector parallel to $x$ axis. This term breaks not only the U(1) spin-rotational symmetry about $z$ axis, but also both ${\mathcal{I}}_s$ and ${\mathcal{R}}_z$ as individual symmetries. However, ${\mathcal{H}}'$ with a non-vanishing $d_x$ still preserves the symmetry ${\mathcal{I}}'$ under the composite operation.
![(Color online) Correlation lengths calculated for the spin-1 chain are plotted against $D_z$. The parameters are varied as (a) $h_z=0.1$, $d_x=0$, and (b) $h_z=d_x=0.1$. Each color and symbol denotes the different number of kept states $\chi$ from $50$ to $200$. Insets show the nonlocal order parameters $\mathcal{O}_s(L)/\operatorname{Tr}\Lambda^4$ for $L=100$ and $200$ (see text).[]{data-label="fig:CorrLength"}](CorrLength_v2.pdf){width="45.00000%"}
We numerically study Hamiltonian using infinite density-matrix renormalization group (iDMRG) [@White92; @McCulloch08; @Kjall13]. The correlation lengths as functions of $D_z$ are plotted in Fig. \[fig:CorrLength\] for different numbers $\chi$ of kept states and parameters of the model. A divergent correlation length with increasing $\chi$ indicates a critical point. For $h_z=0.1$ and $d_x=0$, we find that the Haldane phase and the Néel state ($D_z \rightarrow -\infty $) are adiabatically connected since all of the three symmetries protecting the Haldane phase are broken. However, as found in Refs. [@Tsukano98b; @XDeng13], the transition at $D_z \sim 1$ still exists (see Fig. \[fig:CorrLength\] (a)). This indicates that there is a phase transition between two trivial phases connected to $| N \rangle$ and $| D \rangle$. To confirm that this transition is protected by ${\mathcal{I}}'$ alone, we further introduce $d_x$ in Fig. \[fig:CorrLength\] (b). A single transition in Fig. \[fig:CorrLength\] (a) is now split into two transitions, but the two phases are still separated by (two) transitions and thus are distinct. In the intermediate phase, an AF order along $x$ axis occurs and thus ${\mathcal{I}}'$ is spontaneously broken. Further details about this calculation are shown in .
Once an explicit dimerization in introduced, e.g., by adding a term $\delta \sum_i (-1)^i \vec{S}_i \cdot \vec{S}_{i+1}$ with $\delta \neq 0$, ${\mathcal{I}}'$ is broken without affecting any other symmetries in ${\mathcal{H}}'$, and there is only one trivial phase. Numerically, we observe that the correlation length remains finite for all values of $D_z$ when $d_x=0$ . In fact, this can also be shown analytically by considering the limit of $\delta=1$ with $d_x =0$. Here, the entire chain is decomposed into isolated dimers. In particular, at $D_z = h_z =0$, the ground state is simply given by a product of spin singlet states on each dimer. It can be shown, by solving the two-spin problem explicitly , that this dimerized state is connected adiabatically to both $D_z \to \infty$ and $h_z \to \infty$ limits. Thus, the two trivial product states $|D\rangle$ and $|N\rangle$ can be adiabatically connected through the dimerized limit, and belong to a single phase, in the presence of $\delta$. This fact rules out the possibility that the two trivial phases are distinct under the two-site translation invariance and some on-site symmetry, as indicated in Refs. [@XChen11a; @XChen11b].
*Matrix-product state formulation.—*Matrix-product states (MPS) can represent gapped ground states of local Hamiltonians in one dimension faithfully. Thus, the classification of gapped phases in one dimension, including the SPt phase proposed in the present work, can be proven rigorously within the MPS formalism. Let us begin with the general MPS [@Vidal03; @Orus08], without assuming any translation invariance: $$\begin{aligned}
\left| \psi \right\rangle &=& \sum_{\{m_n\}} \cdots \Gamma^{[n-1]}_{m_{n-1}}
\Lambda^{[n-\frac{1}{2}]} \Gamma^{[n]}_{m_n} \Lambda^{[n+\frac{1}{2}]}
\Gamma^{[n+1]}_{m_{n+1}} \cdots \nonumber \\ && \times \left| \cdots
m_{n-1} m_n m_{n+1} \cdots \right\rangle,\end{aligned}$$ where $\Lambda^{[a]}$ is a $\chi_a \times \chi_a$ positive diagonal matrix, $\Gamma^{[n]}$ is a $\chi_{n-1/2} \times \chi_{n+1/2}$ matrix, and $m_n$ represents the physical degrees of freedom on site $n$. An MPS representation is not unique for a given state but we can always choose the canonical MPS [@PerezGarcia07] satisfying $\operatorname{Tr}{\left[\left(\Lambda^{[a]}\right)^2\right]} =1$, and $${\mathcal{E}}^{[n]}(\mathbb{I}_{\chi_{n+1/2}}) = \mathbb{I}_{\chi_{n-1/2}},
\; \;
\bar{{\mathcal{E}}}^{[n]} (\mathbb{I}_{\chi_{n-1/2}}) =
\mathbb{I}_{\chi_{n+1/2}},
\label{eq.canonical}$$ where $\mathbb{I}_{\chi}$ is the $\chi \times \chi$ identity matrix, and ${\mathcal{E}}^{[n]}$ and $\bar{{\mathcal{E}}}^{[n]}$ are completely positive maps defined by $$\begin{split}
{\mathcal{E}}^{[n]}(X) & \equiv \sum_m \Gamma_m^{[n]} \Lambda^{[n+\frac{1}{2}]} X
\Lambda^{[n+\frac{1}{2}]} \left( \Gamma_m^{[n]} \right)^\dagger, \\
\bar{{\mathcal{E}}}^{[n]} (Y) & \equiv \sum_m \left( \Gamma_m^{[n]} \right)^\dagger
\Lambda^{[n-\frac{1}{2}]} Y \Lambda^{[n-\frac{1}{2}]} \Gamma_m^{[n]}.
\end{split}$$ By introducing the metric $|X|^2 \equiv \textrm{Tr}[ X (\Lambda^{[a]})^2 X^\dagger]$ in the vector space of $\chi_a \times \chi_a$ matrices, we can introduce a singular value decomposition of ${\mathcal{E}}^{[n]}$ and $\bar{{\mathcal{E}}}^{[n]}$. The canonical condition Eq. means that the identity matrices are left/right “eigenvectors” of ${\mathcal{E}}^{[n]}$ and $\bar{{\mathcal{E}}}^{[n]}$ belonging to the largest singular value $1$. In the following we assume that the MPS is pure, that is the largest singular value $1$ is nondegenerate . In order to consider the symmetry ${\mathcal{I}}'$, we define $n \in \mathbb{Z}$ so that ${\mathcal{I}}_s$ can be identified with $n \to -n$ with the inversion center at site $n=0$. Following Refs. [@PerezGarcia08; @Pollmann10], if $\left| \psi \right>$ is invariant under the combined symmetry ${\mathcal{I}}'$, it satisfies $$\begin{aligned}
\sum_{m'} u_{mm'} \left( \Gamma_{m'}^{[n]} \right)^T =
e^{i\theta_{{\mathcal{I}}'}^{[n]}} \left( U_{{\mathcal{I}}'}^{[-n-\frac{1}{2}]}
\right)^\dagger \Gamma_m^{[-n]} U_{{\mathcal{I}}'}^{[-n+\frac{1}{2}]},
\label{eq.fe_calIp}\end{aligned}$$ where $u_{mm'}$ is the representation of ${\mathcal{R}}_z$ acting on the physical Hilbert space of each site, $\theta_{{\mathcal{I}}'}^{[n]}$ is a phase, and $U_{{\mathcal{I}}'}^{[a]}$ is a $\chi_a \times \chi_a$ unitary matrix commuting with $\Lambda^{[a]}$. ${{\mathcal{I}'}}$ also implies that $\Lambda^{[a]} = \Lambda^{[-a]}$ and $\chi_a = \chi_{-a}$. Using the above relation twice, we obtain $$\begin{aligned}
\bar{{\mathcal{E}}}^{[n]} \left( A^{[n-\frac{1}{2}]} \right)
& = e^{- i (\theta^{[n]}_{{\mathcal{I}'}}+ \theta^{[-n]}_{{\mathcal{I}'}})}
A^{[n+\frac{1}{2}]},
\label{eq.cmpEA}\end{aligned}$$ where $ A^{[a]} \equiv \left( U^{[-a]}_{{\mathcal{I}'}}\right)^T \left( U^{[a]}_{{\mathcal{I}'}}\right)^\dagger$. Since $|A^{[a]}|^2=1$, Eq. implies $A^{[n\pm 1/2]}$ are left/right eigenvectors of $\bar{{\mathcal{E}}}^{[n]}$ belonging to the singular value $1$. The assumption of the pure MPS, namely non-degeneracy of the singular value $1$ of $\bar{{\mathcal{E}}}^{[n]}(X)$, implies $A^{[a]} = e^{ i \phi^{[a]}_{{\mathcal{I}'}}} \mathbb{I}_{\chi_a}$, where $\phi^{[a]}_{{\mathcal{I}'}}+ \phi^{[-a]}_{{\mathcal{I}'}}= 0\mod{2\pi}$. Combining these with Eq. and the canonical condition, we obtain $\theta^{[n]}_{{\mathcal{I}'}}+ \theta^{[-n]}_{{\mathcal{I}'}}= \phi^{[n+1/2]}_{{\mathcal{I}'}}- \phi^{[n-1/2]}_{{\mathcal{I}'}}\mod{2\pi}$. In particular, for $n=0$, we find $$\begin{aligned}
2 (\theta^{[0]}_{{\mathcal{I}'}}- \phi^{[1/2]}_{{\mathcal{I}'}}) = 0 \mod{2\pi}.\end{aligned}$$ As a consequence, $\theta^{[0]}_{{\mathcal{I}'}}-\phi^{[1/2]}_{{\mathcal{I}'}}$ is quantized to either $0$ or $\pi$; it cannot change unless the system undergoes a quantum phase transition. This implies that, in the presence of the ${{\mathcal{I}'}}$ symmetry, there are two distinct phases corresponding to $\theta^{[0]}_{{\mathcal{I}'}}-\phi^{[1/2]}_{{\mathcal{I}'}}=0$ and $\pi$. Let us now consider the limits of the trivial product states $|D\rangle$ and $|N\rangle$. Here, all the matrices $\Gamma^{[n]}$, $\Lambda^{[n]}$, and $U^{[a]}_{{\mathcal{I}'}}$ are reduced to scalars ($1 \times 1$ matrices) and thus commute with each other. Then the fundamental relation for $n=0$ reads $\theta^{[0]}_{{\mathcal{I}'}}-\phi^{[1/2]}_{{\mathcal{I}'}}= 0$ for $|D\rangle$ and $\theta^{[0]}_{{\mathcal{I}'}}-\phi^{[1/2]}_{{\mathcal{I}'}}= \pi$ for $|N\rangle$. This establishes that, under the ${\mathcal{I}}'$ symmetry, the two product states $|D\rangle$ and $|N\rangle$ indeed belong to distinct phases, which are always separated by a quantum phase transition.
As in the case of SPT phases, no local order parameter can distinguish SPt phases. However, using the MPS framework, we can directly derive *non-local* order parameters [@Pollmann12b] which are sensitive to the phase factor $\theta^{[0]}_{{\mathcal{I}'}}-\phi^{[1/2]}_{{\mathcal{I}'}}$. In particular, we can define an operator $\mathcal{I}'_s(2L+1)$ which inverts a block of $2L+1$ consecutive sites. For $L$ much larger than the correlation length, we find that $$\begin{aligned}
\mathcal{O}_s(L) = \langle\psi|\mathcal{I}'_s(2L+1)|\psi\rangle \approx \operatorname{Tr}\Lambda^4 e^{i(\theta^{[0]}_{{\mathcal{I}'}}-\phi^{[1/2]}_{{\mathcal{I}'}})}. \end{aligned}$$ The insets in Fig. \[fig:CorrLength\] show that the different SPt phases are indeed distinguished by $\mathcal{O}_s(L)/\operatorname{Tr}\Lambda^4= \pm 1$ while $\mathcal{O}_s(L)=0$ when ${\mathcal{I}}'$ is broken.
*Conclusion and discussion.—* We demonstrated that there exists two distinct SPt phases in the presence of the symmetry under the site-centered inversion combined with a spin rotation. We showed the existence of such phases by field-theoretical arguments based on bosonization and presented a general proof based on the MPS formalism. While it is known that distinct trivial phases can exist in translation-invariant systems [@XChen11a; @XChen11b], it is surprising that only point-group symmetries can stabilize distinct trivial phases in simple 1D systems. Our finding implies that more studies are needed for complete classification of quantum phases in one dimension under symmetries. We derived non-local order parameter that could be measured in optical lattice realizations [@Endres11]. Moreover, even without any measurement of the non-local order parameter, the quantum phase transition separating the distinct SPt phases can be detected in standard experimental measurements, such as a divergence of the low-temperature specific heat when the gap is closing. This, in fact, is more in line with the operational definition of the SPt phases.
The notion of the SPt phases is not restricted to one-dimensional systems. In fact, what appear as examples of SPt phases in two dimensions were discussed in Ref. [@HYao07; @HYao10]. The lack of universal theoretical description of quantum many-body systems in higher dimensions makes a systematic analysis of SPt phases more difficult than in one dimension. Nevertheless, it would be certainly an interesting direction for the future.
#### Acknowledgment—
YF thanks P. Lecheminant, S. Nishimoto, and K. Totsuka for fruitful discussions. YF was supported in part by the Program for Leading Graduate Schools, MEXT, Japan. This work was supported in part by JSPS Grant-in-Aid for Scientific Research (KAKENHI) No. 25400392 and JSPS Strategic International Networks Program No. R2604 “TopoNet.” Numerical calculations were performed on supercomputers at the Institute for Solid State Physics, University of Tokyo.
[40]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1063/1.3149495) [****, ()](\doibase 10.1103/PhysRevB.85.085103) [****, ()](\doibase
10.1103/PhysRevLett.94.140601) [****, ()](\doibase 10.1103/PhysRevB.82.155138) [****, ()](\doibase 10.1103/PhysRevB.84.165139) [****, ()](\doibase 10.1103/PhysRevB.80.155131) [****, ()](\doibase 10.1103/PhysRevB.81.064439) [****, ()](\doibase 10.1103/PhysRevB.83.035107) [****, ()](\doibase 10.1103/PhysRevB.84.235128) [****, ()](\doibase 10.1103/PhysRevB.87.155114) [****, ()](\doibase 10.1103/PhysRevLett.95.146802) [****, ()](\doibase 10.1016/0375-9601(83)90631-X) [****, ()](\doibase 10.1103/PhysRevLett.50.1153) [****, ()](\doibase 10.1103/PhysRevLett.59.799) [****, ()](http://projecteuclid.org/euclid.cmp/1104161001) [****, ()](\doibase 10.1103/PhysRevB.40.4709) [****, ()](\doibase 10.1103/PhysRevLett.66.798) [****, ()](\doibase 10.1103/PhysRevB.67.104401) [****, ()](\doibase 10.1103/PhysRevB.85.075125) [****, ()](\doibase 10.1143/JPSJ.67.302) [****, ()](\doibase 10.1103/PhysRevB.57.R8087) [****, ()](\doibase 10.1103/PhysRevB.34.6372) [****, ()](\doibase 10.1103/PhysRevB.77.245119) [****, ()](\doibase 10.1103/PhysRevLett.106.110405) [](http://arxiv.org/abs/1410.4211) [****, ()](\doibase 10.1103/PhysRevLett.69.2863) [](http://arxiv.org/abs/0804.2509) [****, ()](\doibase 10.1103/PhysRevB.87.235106) [****, ()](\doibase
10.1103/PhysRevB.87.195101) [****, ()](\doibase 10.1103/PhysRevLett.91.147902) [****, ()](\doibase 10.1103/PhysRevB.78.155117) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.100.167202) [****, ()](\doibase 10.1103/PhysRevB.86.125441) [****, ()](\doibase 10.1126/science.1209284) [****, ()](\doibase 10.1103/PhysRevB.76.161104) [****, ()](\doibase 10.1103/PhysRevLett.105.166402) [****, ()](http://stacks.iop.org/1742-5468/2004/i=06/a=P06002) [****, ()](\doibase
10.1103/PhysRevB.86.195122)
Details for numerical calculation {#details-for-numerical-calculation .unnumbered}
=================================
In this supplemental material, we present several numerical details about the spin-1 chain, $$\begin{aligned}
H = \sum_i \left[ \left( 1+\delta (-1)^i \right) \vec{S}_i \cdot \vec{S}_{i+1} + D_z (S^z_i)^2 +h_z (-1)^i S^z_i + d_x (S^y_i S^z_{i+1} -S^z_i S^y_{i+1}) \right]. \end{aligned}$$
Central charge, magnetization, and entanglement spectrum
--------------------------------------------------------
For $h_z=0.1$ and $d_x=\delta=0$, from the divergent behavior of the correlation length in Fig.1(a) of the main text, we expect a Gaussian transition at $D_z \sim 1$. To confirm this, we calculate the von Neumann entanglement entropy $S$ for a bipartition of the system into two half chains as a function of the correlation length $\xi$. From conformal field theory, the entanglement entropy is known to scale as [@Calabrese04S; @Kjall13S] $$\begin{aligned}
S = \frac{c}{6} \log (\xi /a) + c',\end{aligned}$$ where $c$ is the central charge, $a$ is the lattice spacing (we set $a=1$), and $c'$ is a nonuniversal constant. As shown in Fig. \[fig:EEandMx\] (a), the entanglement entropy $S$ at $D_z=1.00$ is well fitted by a linear function of $\log (\xi)$, and the central charge is estimated as $c \approx 1.02$. This is close to the expected value $c=1$ at the Gaussian transition.
{width="80.00000%"}
For $h_z=d_x=0.1$ and $\delta=0$, a Gaussian transition at $d_x=0$ splits into two Ising transitions, and we have an intermediate phase between these transitions. In fact, between the two peaks in Fig.1(b) of the main text, the $\mathbb{Z}_2$ spin reversal symmetry in $x$ axis ($S^x_i \rightarrow -S^x_i$) is spontaneously broken. Then the staggered magnetizations along $x$ axis, $M_x$, take finite expectation values in the intermediate phase, as shown in Fig. \[fig:EEandMx\] (b). Such an intermediate Néel phase between two distinct gapped symmetric phases is also observed in a two-leg spin-$1/2$ ladder [@ZXLiu12S] when the $U(1)$ symmetry is explicitly broken.
To see a clear signature of the topologically trivial phases, we check the degeneracy in the entanglement spectra. For $h_z=d_x=\delta=0$, the Haldane phase is protected by time reversal, bond-centered inversion, and dihedral group of the spin rotations, as observed in Ref. [@Pollmann10S]. From Fig. \[fig:EntSpec\] (a), the whole entanglement spectrum is two-fold degenerate in a region $-0.3 \lesssim D_z \lesssim 1$. On the other hand, as shown in Fig. \[fig:EntSpec\] (b), once we introduce a finite staggered magnetic field $h_z$, this two-fold degeneracy is lifted, and the Haldane phase is merged with the imposed AF ordered phase which is topologically trivial.
{width="90.00000%"}
Adiabatic connection between |D> and |N>
----------------------------------------------
A nonzero dimerization $\delta$ breaks the site-centered inversion symmetry but still preserves the two-site translational invariance. In the main text, we discuss an adiabatic continuity between the two trivial states $| D \rangle$ and $| N \rangle$ in the absence of the combined symmetry, $\mathcal{I}' = \mathcal{I}_s \times \mathcal{R}_z$. We first show the correlation length as a function of $D_z$ for $h_z=\delta=0.1$ and $d_x=0$ in Fig. \[fig:CL\_Dimer\]. Compared with Fig.1(a) in the main text, the correlation length around $D_z \simeq 1$ becomes shorter and exhibits a saturating behavior by increasing $\chi$. This indicates the absence of the phase transition due to the dimerization which breaks the combined symmetry $\mathcal{I}'$.
![Correlation length as a function of $D_z$ for the spin-1 chain with $h_z=\delta=0.1$ and $d_x=0$. Each color and symbol denotes the different number of $\chi$ varied from $50$ to $200$.[]{data-label="fig:CL_Dimer"}](CL_Dimer.pdf){width="50.00000%"}
![Energy spectra of the two-site Hamiltonian . $D_z$ are varied and $h_z=0$ on the left panel, while $h_z$ are varied and $D_z=0$ on the right panel.[]{data-label="fig:Energy_JD"}](Energy_JD.pdf){width="70.00000%"}
In fact, this is easily and rigorously seen from the perfectly dimerized limit $\delta=1$. In this limit, the Hamiltonian with $d_x=0$ is reduced to the sum of independent two-site Hamiltonians, $$\begin{aligned}
\label{eq:2body}
H_\textrm{two-site} = 2\vec{S}_1 \cdot \vec{S}_2 + D_z[(S^z_1)^2+(S^z_2)^2] + h_z(S^z_1-S^z_2). \end{aligned}$$ Again we can obtain the two trivial states $\left| D \right>$ and $\left| N \right>$ in the limits $D_z \rightarrow \infty$ and $h_z \rightarrow \infty$, respectively. Therefore, the continuity between $\left| D \right>$ and $\left| N \right>$ is confirmed by finding a path on which no level crossing occurs in the lowest energy spectrum between these limits. In Fig. \[fig:Energy\_JD\], we plot energy spectra of the two-site Hamiltonian. The singlet state at $D_z=h_z=0$ is adiabatically connected to both the states $\left| 00 \right>$ and $\left| -+ \right>$. From our numerical data we thus conclude that, by breaking the site-centered inversion symmetry, the two states $\left| D \right>$ and $\left| N \right>$ are adiabatically connected and thus no longer distinguished.
[4]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](http://stacks.iop.org/1742-5468/2004/i=06/a=P06002) [****, ()](\doibase 10.1103/PhysRevB.87.235106) [****, ()](\doibase
10.1103/PhysRevB.86.195122) [****, ()](\doibase 10.1103/PhysRevB.81.064439)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Cosmological simulations have shown that dark matter haloes are connected to each other by large-scale filamentary structures. Cold gas flowing within this “cosmic web” is believed to be an important source of fuel for star formation at high redshift. However, the presence of such filamentary gas has never been observationally confirmed despite the fact that its covering fraction within massive haloes at high redshift is predicted to be significant ($\sim$ 25%), (Dekel et al. 2009). In this work, we investigate in detail whether such cold gas is detectable using low-ionisation metal absorption lines, such as CII $\lambda$1334 as this technique has a proven observational record for detecting gaseous structures. Using a large statistical sample of galaxies from the Mare Nostrum N-body+AMR cosmological simulation, we find that the typical covering fraction of the dense, cold gas in 10$^{12}$ M$_{\odot}$ haloes at z $\sim$ 2.5 is lower than expected ($\sim$ 5%). In addition, the absorption signal by the interstellar medium of the galaxy itself turns out to be so deep and so broad in velocity space that it completely drowns that of the filamentary gas. A detectable signal might be obtained from a cold filament exactly aligned with the line of sight, but this configuration is so unlikely that it would require surveying an overwhelmingly large number of candidate galaxies to tease it out. Finally, the predicted metallicity of the cold gas in filaments is extremely low ($\leq$ 10$^{-3}$ Z$_\odot$). Should this result persist when higher resolution runs are performed, it would significantly increase the difficulty of detecting filamentary gas inflows using metal lines. However, even if we assume that filaments are enriched to Z$_\odot$, the absorption signal that we compute is still weak. We are therefore led to conclude that it is extremely difficult to observationally prove or disprove the presence of cold filaments as the favorite accretion mode of galaxies using low-ionisation metal absorption lines. The Ly$\alpha$ emission route looks more promising but due to the resonant nature of the line, radiative transfer simulations are required to fully characterize the observed signal.'
author:
- |
\
$^1$ Department of Physics, Denys Wilkinson Building, Keble Road, Oxford, OX1 3RH, United Kingdom\
$^2$ Centre de Recherche Astrophysique de Lyon, UMR 5574, 9 Avenue Charles Andre, F69561 Saint Genis Laval, France\
$^3$ Institut d�astrophysique de Paris & UPMC (UMR 7095), 98, bis boulevard Arago , 75 014, Paris, France\
title: 'Are cold flows detectable with metal absorption lines?'
---
galaxies: formation – galaxies: high-redshift – galaxies: intergalactic medium – cosmology: theory
=-0.6in =0.2in
Introduction
============
How galaxies get their gas is a long-standing issue. For decades, the standard theoretical picture of galaxy formation has stipulated that all gas accreted into dark matter haloes is shock heated before it radiatively cools and settles into a galactic disk [@silk77; @rees77 although @binney77 first suggested this need not be the case]. This picture has recently been revisited both by analytic studies [@birnboim03; @dekel06] and hydrodynamical simulations (both in 1D: @birnboim03 and in 3D within an explicit cosmological context, @keres05 [@keres09; @ocvirk08]; hereafter OPT08). These studies have established that in haloes below a critical mass shocks are unstable and cannot propagate outwards, so that cold diffuse gas and/or cold filaments can penetrate deep into the halo without experiencing shock-heating. In contrast, at the other end of the mass spectrum, very massive haloes easily sustain a virial shock that is stable against gas cooling so that diffuse/filamentary gas is shock heated to the virial temperature of the halo as it enters. Finally, at intermediate halo masses, either gravitationally shock-heated hot gas and/or a hot galactic wind coexists with cold inflowing filaments (OPT08): some dense cold filaments are stable against the pressure force exerted by the hot gaseous material. In other words, the vast majority of galaxy-size host halos, especially at high redshift ($z>2$), are predicted to be threaded by cold gas filaments in a $\Lambda$CDM model of structure formation. Therefore, the question which naturally arises is whether or not the existence of these cold gas filaments can be observationally confirmed, and by which technique.
Since the advent of high-resolution spectroscopy has enabled astronomers to study the kinematics of the intergalactic medium at high redshifts in a wealth of detail [e.g. @pettini02; @adelberger03; @shapley03], and given the fact that cold gas is thought to flow into massive haloes ($10^{12} {\mbox{${\rm M}_{\odot}$}}$) at $z=2.5$ along filaments with velocities of $\gtrsim 200\,{\rm km s^{-1}}$ [@dekel09], it is sensible to think that the spectra of Lyman-break galaxies [@steidel96] might reveal these filaments as redshifted absorption features. Interestingly, a recent study reported that few Lyman-break galaxies (LBGs) show redshifted metal absorption lines, suggesting that inflowing gas in haloes with $4\times10^{11}<M_{\rm vir}<10^{12}M_{\odot}$ is rare [@steidel10]. Taken at face value, this appears to contradict the theoretical prediction that cold filaments are prominent in the vast majority of high-z halos.
However, there exist a variety of reasons as to why these filaments should be very difficult to detect. The first of these is the covering fraction of the filaments. This is estimated in [@dekel09] to be around ($\sim 25\%$) for four massive halos with $M_{\rm vir}\sim10^{12}M_{\odot}$ in the simulation, counting only relatively dense ($N_{{\rm H}}>10^{20}{\rm cm}^{-2}$) and cold ($T<10^{5}K$) gas within a radial distance $20<r<100$ kpc from the central galaxy hosted by these halos. Whereas $25\%$ is indeed a non-negligible covering fraction, its exact dependence on redshift and halo mass remains to be determined. For instance, @faucher10a also measured the covering fraction for a Milky way-type progenitor LBG in a smoothed particle hydrodynamic simulation but found a significantly smaller value ($\sim 2\%$). Secondly, low-ionisation lines can be produced not only by cold filamentary gas but also by a galaxy’s interstellar medium, so that when using a single galaxy to probe the circumgalactic medium, distinguishing absorption produced by filaments from that produced by the ISM is key to prove/disprove the presence of cold filaments. Thirdly, there is the geometry of the accretion: in order to produce a strong absorption signal, a filament needs to be well aligned with the line of sight to maximise its column density. Finally, there is the issue of metallicity: a metal-poor filament will be transparent to metal line observations.
The aim of this letter is to quantify the aforementioned effects to better assess the detectability of the absorption signal produced by cold filaments. We show that the actual probability of detecting such flows with metal lines is much smaller than the high covering fraction derived in @dekel09 would suggest due to i) the low density and ii) low metallicity of the filaments [*compared with the densities and metallicites of the interstellar medium of the host galaxy.* ]{}
Simulations
===========
To investigate the statistical properties of cold filaments, we analyze the simulation, carried out with the [WMAP1]{} cosmology [@spergel03] using the adaptive mesh refinement code [ramses]{} [@teyssier02]. Details of the simulation can be found in OPT08, @dekel09, @devriendt10, so we just briefly describe the simulation setup and modelling strategy here.
The simulation follows the evolution of a periodic cosmological volume of comoving side length $L=50\,{h^{-1}}$Mpc, and contains $1024^3$ dark matter particles with $m_{\rm dm}=1.4\times10^7 {\rm M}_{\odot}$. Dark matter halos are identified using the [AdaptaHop]{} algorithm [@aubert04; @tweed09], resulting in 3419 halos with $M_{\rm vir} \ge 10^{11} {\rm M}_\odot$ at $z=3.8$ and 6456 haloes at $z=1.5$. The spatial resolution of the simulation is kept fixed at around $1\,h^{-1}$kpc physical over the entire redshift range.
Gas in the simulation can radiatively cool by atomic processes down to $10^4 K$ [@sutherland93]. A fraction of the cold and dense gas ($n_{\rm H}>0.1 {\rm cm}^{-3}$) turns into stars following the Schmidt-Kennicutt law [@kennicutt98], and massive stars explode as type II supernovae, redistributing energy and metals into the interstellar/intergalactic medium. The Sedov Blast wave solution is adopted for supernova explosions [@dubois08], and reionization is implemented by instantaneously turning on a uniform UV background at z=8.5 [@haardt96].
Results
=======
Covering Fraction of Dense Gas
------------------------------
OPT08 showed that the transition mass () separating cold from hot dominated accretion increases with increasing redshift. According to these authors, by $z\sim2.5$, the cold streams feeding massive haloes (${\mbox{$M_{\rm 200}$}}\gtrsim 3\times10^{11}{\mbox{${\rm M}_{\odot}$}}$) begin to disappear, but based on measurements of a handful of haloes in the simulation at $z\sim2.5$, @dekel09 find that some massive haloes with ${\mbox{$M_{\rm 200}$}}\sim10^{12}{\mbox{${\rm M}_{\odot}$}}$ still show high covering fractions ($\lesssim 25\%$) of dense ($N_{\rm H}>10^{20} {\rm cm}^{-2}$) and cold ($\rm T<10^5 K$) inflowing gas within $20 < r < 100$ kpc. In order to obtain statistically significant results on the covering fraction in the simulation, we analyse [*all*]{} the halos in the simulation and show the results in Fig. \[fig:covfrac\]. Using the complete sample, we find that the average covering fraction of haloes with ${\mbox{$M_{\rm 200}$}}\sim10^{12}{\mbox{${\rm M}_{\odot}$}}$ at $z=2.5$ is $\sim 5\%$, about a factor 5 less than the covering fraction reported in @dekel09 for their sub-sample of ${\mbox{$M_{\rm 200}$}}\sim10^{12}{\mbox{${\rm M}_{\odot}$}}$ halos at $z\sim2.5$. We note that this lower value is more consistent with the recent findings of @faucher10a. However, whilst low covering fractions ($\sim 5\%$) are computed for most halos at redshifts $z\lesssim 3$, those of higher redshift ($z\sim 3.8$) halos are much larger ($\sim 25\%$), as can be seen in Fig. \[fig:covfrac\]. This strong redshift evolution in the covering fraction of cold filaments between $z\sim3.8$ and $z\sim2.5$ reflects the aforementioned rapid transition from cold to hot dominated accretion. We note that these results are consistent with the OPT08 value of $\sim 10^{13}{\mbox{${\rm M}_{\odot}$}}$ for at $z=4$.
An interesting feature present in Fig. \[fig:covfrac\] is that the covering fraction is higher in more massive haloes at a given redshift. This seems to contradict previous findings (OPT08) that it is the small haloes which are mainly fed by cold mode accretion. It should be noted, however, that the covering fraction shown in Fig. \[fig:covfrac\] does not account for the accretion of cold, [*diffuse*]{} gas (cold gas with lower column densities) which is only present in halos with masses incapable of sustaining a virial shock at all. Indeed, these small haloes are usually located within (or around) filaments whose density is low, whereas the filaments around more massive haloes tend to be denser. Furthermore, in massive haloes, satellite galaxies contribute more importantly to the covering fraction, as do extended, warped, galactic disks and dense gas bridges which result from tidal interactions between galaxies. These latter effects partly explain the trend, but the primary driver of the covering fraction increase with halo mass is the density of the accreted gas, which is higher in more massive haloes. Fig. \[fig:covfrac\] (dotted lines) substantiates this claim by showing how the covering fractions drop when the very dense gas ($\rm N_H >
10^{21} cm^{-2}$) which belongs to the ISM of satellite galaxies is excluded from the measurement.
![The covering fraction of cold ($T<10^5 K$) and dense ($N_{\rm H} > 10^{20}$) gas within $20 < r < 100$ kpc physical as a function of the virial mass of halos (). Different colours indicate the covering fractions at different redshifts. Solid lines include the contribution from the interstellar medium of satellite galaxies to the covering fraction. To exclude this latter contribution, we also plot the covering fraction with upper density cut ($10^{20}<N_{\rm H}<10^{21}$ cm$^{-2}$, dotted lines). Error bars correspond to the interquartile range ($25\le f \le 75\%$). For a given halo mass, halos at higher redshift show larger covering fractions. This can be understood in terms of the rapid development of a virialized hot medium between $z=3.8$ and $z=2.5$. []{data-label="fig:covfrac"}](fig1.ps){width="8.8cm"}
CII Absorption
--------------
![An example of the contribution from filamentary gas to the CII $\lambda1334$ absorption. Upper panels show the hydrogen column density and mass-weighted metallicity distributions for a $10^{12}{\mbox{${\rm M}_{\odot}$}}$ halo at $z\sim3.8$. A cold filament is indicated on the left panel, which is receding from the observer and therefore should produce a redshifted absorption line. A white square denotes the central $\sim$10 kpc$^2$ region over which the absorption spectrum is obtained. Bottom panel shows an absorption profile of this galaxy (thick grey solid line). We also compute the absorption without the central disc by neglecting the opacity of central cells ($r\leq0.1r_{200}$) (green solid line) and absorption without the outer gas ($r>0.1r_{200}$) (blue dashed line). Also included is the absorption produced when the outer gas ($r > 0.1 r_{200}$) is assumed to have solar metallicity in the absence of the central disc (red dotted line). It can be seen that the contribution from the filament to the low-ionisation metal line is negligible compared to that of the ISM of the galaxy in every case. An uncorrected spectrum is included to show the effect of the correction applied to the velocity distribution of gas (thin solid grey line).[]{data-label="fig:example"}](fig2_1.ps "fig:"){width="8cm"} ![An example of the contribution from filamentary gas to the CII $\lambda1334$ absorption. Upper panels show the hydrogen column density and mass-weighted metallicity distributions for a $10^{12}{\mbox{${\rm M}_{\odot}$}}$ halo at $z\sim3.8$. A cold filament is indicated on the left panel, which is receding from the observer and therefore should produce a redshifted absorption line. A white square denotes the central $\sim$10 kpc$^2$ region over which the absorption spectrum is obtained. Bottom panel shows an absorption profile of this galaxy (thick grey solid line). We also compute the absorption without the central disc by neglecting the opacity of central cells ($r\leq0.1r_{200}$) (green solid line) and absorption without the outer gas ($r>0.1r_{200}$) (blue dashed line). Also included is the absorption produced when the outer gas ($r > 0.1 r_{200}$) is assumed to have solar metallicity in the absence of the central disc (red dotted line). It can be seen that the contribution from the filament to the low-ionisation metal line is negligible compared to that of the ISM of the galaxy in every case. An uncorrected spectrum is included to show the effect of the correction applied to the velocity distribution of gas (thin solid grey line).[]{data-label="fig:example"}](fig2_2.ps "fig:"){width="8.5cm"}
In order to more carefully investigate the possibility of detecting cold filaments using metal absorption lines, we compute the strength of the ${\rm CII} ~\lambda 1334$ absorption. Our choice of line is dictated not only by the temperature of the filamentary gas which is not high enough to significantly produce more highly ionized metallic elements such as CIV, but also because it is empirically known to yield the strongest absorption feature [@steidel10] We do not attempt to model absorption by CII accurately which would require detailed radiative transfer, but instead derive an upper limit by making several extreme assumptions. Firstly, we assume that all the carbon present in filaments is eligible for the ${\rm C II}~ \lambda 1334$ transition. Secondly, we use the solar abundance ratio ($[C/Z]_{\odot}\simeq 0.178$, @asplund09) to obtain the carbon column density for a given metallicity in the simulation. The optical depth of a grid cell is computed as $\tau = \sigma_{\rm C II} n_{\rm C II} \Delta l$, where $n_{\rm CII}$ is the carbon number density, $\Delta l$ is the size of the grid cell, and $\sigma_{\rm C II}$ is the cross-section for the line transition, which is calculated as $\sigma_{\rm C II}=(3\pi \sigma_{\rm T}/8)^{1/2} f \lambda_0\simeq 1.5 \times 10^{-18} {\rm cm}^2$. Here $\sigma_{\rm T}$ is the Thomson cross-section, $\lambda_0$ is the rest-frame wavelength of the transition, and $f$ is the corresponding oscillator strength. We then correct the optical depth by assuming that each grid cell has a Gaussian velocity distribution with a dispersion ($\sigma_{\rm 1D}/2$), which is obtained from the line-of-sight velocity dispersion ($\sigma_{\rm los}$) computed using the closest 27 neighbouring cells. We have tested the validity of the correction by using a high resolution in the series (12 pc resolution, Devriendt et al. [*in prep.*]{}), and found that this procedure yields a accurate approximation of the maximum absorption strength and FWHM of the absorption profile that one would derive by using a much higher number of resolution elements. Finally stars are assumed to dominate the UV emission and we use the @maraston05 spectral energy distributions to derive the continuum flux around $\lambda\sim1334\AA$, which depends on their mass, age, and metallicity. The emission from each star particle in the galaxy is then used to estimate an observed flux, as attenuated by intervening gas present along the line of sight. Note that only the flux emitted by the central 10 kpc$^2$ of a galaxy is used to compute the absorption line so as to mimic the observational resolution of FWHM$\simeq 0.40''$ for galaxies at $z\sim2$ [@steidel10].
Fig. \[fig:example\] shows the HI column density map, projected metallicity distribution, and the corresponding absorption profile for a galaxy residing in a ${\mbox{$M_{\rm 200}$}}\sim10^{12}{\mbox{${\rm M}_{\odot}$}}$ halo at $z\sim3.8$. A dense filament is clearly seen not only in the H column density map, but also in the metallicity map, with metallicity ($\sim 10^{-4}-10^{-3} Z_{\odot}$). The filamentary gas is receding from the observer; hence if detectable it should produce a redshifted absorption line. However, it turns out that the absorption signal is dominated by the ISM of the galaxy lighting up the filament. When the absorption due to the ISM is arbitrarily removed by neglecting the opacity from the gas inside the central gas disc ($r<0.1 r_{200}$), the absorption feature vanishes (green line). Even when the gas outside $r > 0.1 r_{200}$ is assumed to have solar metallicity, the absorption strength is still much smaller (red dotted line) than the absorption produced by the galaxy’s ISM (grey and blue dashed lines). This strongly suggests that the primary reason why it is so difficult to detect the cold filament is that the density of the filamentary gas is much lower than that of the galaxy’s own ISM. As a result, and since the two absorption lines are not well enough separated in velocity space, the filament absorption signal is completely swamped by the high level of ISM absorption in the red wing of the line.
In order to see if we can bring out the filamentary signal by stacking absorption profiles, we analyse the absorption spectra of 132 and 386 massive galaxies ($M_{200}\ge 10^{12}{\mbox{${\rm M}_{\odot}$}}$) at $z=3.8$ and $z=2.5$ respectively along 6 projections ($+x, -x, +y, -y, +z, -z$) and find that the optical depth for the CII $\lambda 1334$ transition by the filaments is fundamentally small regardless of redshift. Fig. \[fig:stat\] shows that the stacked (mean) absorption strength is unaffected by the presence of filaments. To check whether the signal from filaments could potentially become noticeable if the ISM was less metal-enriched, we also examined the case where only a tiny fraction (1%) of the carbon in the ISM is eligible for the CII transition (blue lines in Fig. \[fig:stat\]), but the difference between absorption profiles resulting from all the gas along the line of sight versus the case where we exclude the gas in the central region ($r<0.1 r_{200}$) is still minute. Therefore, we conclude that the presence of cold filaments is very difficult to confirm with low-ionisation metal absorption lines.
The optical depth of filaments may be under-estimated due to the finite resolution of the simulation. For example, the density of the filamentary structure at $z=7$ in the simulation ($0.005 \lesssim n_{\rm H} \lesssim 0.1$) is more than an order of magnitude lower than that of the simulation ($0.1 \lesssim n_{\rm H} \lesssim 1$) where the filaments are fully resolved (Powell et al. [*in prep.*]{}). However, this increase does not suffice to produce a strong absorption signal. Moreover, we find that the effect of increasing resolution affects the ISM density more considerably (it becomes more than four orders of magnitude higher in the than in the simulation). As a consequence the ISM causes a stronger absorption signal, which more than compensates the increased contribution from the cold filament.
![Stacked absorption profiles of massive galaxies ($M_{200}\ge 10^{12}{\mbox{${\rm M}_{\odot}$}}$) at $z=3.8$ (left) and $z=2.5$ (right). We take 6 projections ($+x, -x, +y, -y, +z, -z$) for 132 galaxies at $z=3.8$ and 386 galaxies at $z=1.5$. The stacked spectra are obtained by taking the mean of the normalised intensity. We also show the absorption spectra expected when only a fraction (1%) of the ISM carbon is assumed to produce the absorption (blue lines, see the text). The contribution from the outer regions (i.e. filament if any) is negligible in the absorption profile. []{data-label="fig:stat"}](fig3.ps){width="8.5cm"}
It should also be noted that finite resolution affects the metallicity of the filamentary gas, so that in the real Universe its metallicity may be higher than the values ($\sim 10^{-4}-10^{-3} Z_{\odot}$) we report for the simulation. Indeed, our simulation cannot resolve all the small galaxies that might pollute the pristine filamentary gas. Besides, it is well known that the energy from supernovae will artificially be dissipated when simulations are run with an insufficient level of resolution. Indeed, in such a case, supernovae only explode in dense grid cells, resulting in substantial radiative losses and negligible momentum transfer to the surrounding gas. Under these circumstances, metals cannot disperse properly. For comparison, the ultra-high resolution simulation indicates that the metallicity of the filamentary gas around a $\sim 10^{9}{\mbox{${\rm M}_{\odot}$}}$ galaxy can reach values up to $10^{-2} Z_{\odot}$ already at $z \simeq 7$. Yet, such an increase in metallicity would still be inconsequential for the absorption spectra. Furthermore, we believe that the metallicity in the filaments is not likely to rise appreciably beyond these values, because most of the supernova ejecta escapes in a direction perpendicular to that of the elongated filament, which makes it difficult to efficiently enrich the filamentary gas with metals (Geen et al 2010, [*in prep.*]{}). However, for the sake of completeness, we present in Fig. \[fig:example\] the case of a halo for which the gas outside $r > 0.1 r_{200}$ is arbitrarily assumed to have solar metallicity. We find that the absorption strength of the filament (red dotted line) is still much smaller than the absorption produced by the galaxy’s ISM.
Conclusions and Discussion
==========================
Cosmological simulations predict that high-z galaxies grow by acquiring gas from cold streams [@dekel09], but no observational confirmation has been obtained yet. Based on a statistical sample from the simulation, we argue that low-ionisation metal absorption features, such as CII $\lambda1334$, arising from intervening cold filaments are extremely hard to distinguish from absorption by the ISM of high-z star-forming galaxies. This is primarily because the optical depth for the low-ionisation transition from cold filamentary gas is minuscule, compared with that of the ISM of the host galaxy. Moreover, the filamentary absorption is not redshifted enough with respect to the ISM absorption, so that the residual ISM absorption in the red wing of the line is still prominent. This small optical depth of filaments mainly finds its source in the intrinsically low densities and metallicities of the cold gas when compared to those of the galaxy’s ISM. Another factor is the geometry of the flow which lowers the probability of detecting filaments as their column density will rarely be maximised by being aligned with the line of sight.
As an alternative to using a single galaxy, one could probe circumgalactic regions by using a paired background galaxy. This method alleviates the importance of ISM absorption since when probing circumgalactic regions in this way, the absorption by the ISM of the foreground galaxy will occur at a different spatial position from that produced by filaments. Unfortunately, the rare occurrence of suitable foreground-background galaxy pairs makes it difficult to probe more than one line of sight per foreground galaxy. As a result, high resolution individual spectra are hard to obtain and one has to resort to stacking the spectra of multiple galaxies [@steidel10]. Contrary to what might be expected, stacking will wash out the cold filament absorption signal since absorption by inflowing gas does not neatly separate from that caused by outflows as was the case when probing the circumgalactic region using single galaxies. Indeed, cold filament absorption against the background galaxy light will not only be redshifted, but also blueshifted as one expects that on average as many cold streams will be detected moving towards as away from the observer. Absorption by outflows will also suffer the same fate. This will make it all the more difficult to argue whether the observed absorption features are driven by infalling gas or by outflows from high-z star-forming galaxies. Moreoever, the metal column density of cold filaments is minuscule, as already mentioned. The hydrogen column densities of the cold filaments are distributed around $10^{20} {\rm cm^{-2}}$ at these redshifts, yielding corresponding carbon column densities around $10^{16} {\rm cm^{-2}}$ if an average $Z=0.001 Z_{\odot}$ is used. Even if all carbon atoms are assumed to be eligible for the CII transition, the optical depth is only around $10^{-2}$ in the line. On the other hand, outflows are expected to be very metal rich, so that even though higher transitions like CIV are expected to dominate the absorption signal, the amount of CII absorption from these outflows might still swamp that produced by the cold filaments.
Finally, we have assessed possible numerical resolution issues on column density and metallicity of the filaments using the very high resolution numerical simulation suite and found that our conclusions remain by and large unchanged.
Based on these considerations, we conclude that the presence of the cold filament is difficult to disprove/prove with low-ionisation metal line absorption. Instead, the Lyman $\alpha$ emission route seems more promising to detect cold filaments, but the line profile is more sensitive to the kinematics of the intervening gas [@verhamme06; @verhamme08]. Therefore this will require full blown radiative transfer calculations [e.g. @faucher10b].
Acknowledgements {#acknowledgements .unnumbered}
================
The simulation was run at the Barcelona Supercomputing Center and the simulations at the CINES and the STFC HPC Facility DiRAC (Oxford node). CP acknowledges support from a Leverhulme visiting professorship at the Astrophysics department of the University of Oxford, TK from a Clarendon DPhil studentship. JD’s research is supported by the Oxford Martin School, as is AS’s along with Beecroft and STFC. We also acknowledge support from the Franco-Korean PHC STAR program.
[99]{} Adelberger K. L., Steidel C. C., Shapley A. E., Pettini M., 2003, ApJ, 584, 45 Aubert D., Pichon C., Colombi S., 2004, MNRAS, 352, 376 Asplund M., Grevesse N., Sauval A. J., Scott P., 2009, ARA&A, 47, 481 Binney J., 1977, ApJ, 215, 483 Birnboim Y., Dekel A., 2003, MNRAS, 345, 349 Dekel A., Birnboim Y., 2006, MNRAS, 368, 2 Dekel A., et al., 2009, Nature, 457, 451 Devriendt J., et al. 2010, MNRAS, 403, L84 Dubois Y., Teyssier R., 2008, A&A, 477, 79 Faucher-Giguere C. ., & Kere[š]{} D. 2010, arXiv:1011.1693 Faucher-Gigu[è]{}re, C.-A., Kere[š]{}, D., Dijkstra, M., Hernquist, L., & Zaldarriaga, M. 2010, ApJ, 725, 633 Haardt F., Madau P., 1996, ApJ, 461, 20 Kennicutt R. C., 1998, ApJ, 498, 181 Kere[š]{} D., Katz N., Weinberg D. H., Dav[é]{} R., 2005, MNRAS, 363, 2 Kere[š]{} D., Katz N., Fardal M., Dav[é]{} R., & Weinberg D. H. 2009, MNRAS, 395, 160 Maraston C., 2005, MNRAS, 362, 799 Ocvirk P., Pichon C., Teyssier R., 2008, MNRAS, 390, 1326 Pettini M., Rix S. A., Steidel C. C., Adelberger K. L., Hunt M. P., Shapley A. E., 2002, ApJ, 569, 742 Rees M. J., Ostriker J. P., 1977, MNRAS, 179, 541 Shapley A. E., Steidel C. C., Pettini M., Adelberger K. L., 2003, ApJ, 588, 65 Silk J., 1977, ApJ, 211, 638 Spergel, D. N., et al. 2003, ApJS, 148, 175 Steidel C. C., Giavalisco M., Pettini M., Dickinson M., & Adelberger K. L. 1996, ApJ, 462, L17 Steidel C. C., et al., 2010, ApJ, 717, 289 Sutherland R. S., Dopita M. A., 1993, ApJS, 88, 253 Teyssier R., 2002, A&A, 385, 337 Tweed D., Devriendt J., Blaizot J., Colombi S., Slyz A., 2009, A&A, 506, 647 Verhamme A., Schaerer D., & Maselli A. 2006, A&A, 460, 397 Verhamme A., Schaerer D., Atek H., & Tapken C. 2008, A&A, 491, 89
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present results from a study of the coherence properties of a system involving three discrete states coupled to each other by two-photon processes via a common continuum. This tripod linkage is an extension of the standard laser-induced continuum structure (LICS) which involves two discrete states and two lasers. We show that in the tripod scheme, there exist two population trapping conditions; in some cases these conditions are easier to satisfy than the single trapping condition in two-state LICS. Depending on the pulse timing, various effects can be observed. We derive some basic properties of the tripod scheme, such as the solution for coincident pulses, the behaviour of the system in the adiabatic limit for delayed pulses, the conditions for no ionization and for maximal ionization, and the optimal conditions for population transfer between the discrete states via the continuum. In the case when one of the discrete states is strongly coupled to the continuum, the population dynamics reduces to a standard two-state LICS problem (involving the other two states) with modified parameters; this provides the opportunity to customize the parameters of a given two-state LICS system.'
address: |
$^1$Fachbereich Physik der Universität, 67653 Kaiserslautern, Germany\
$^2$Helsinki Institute of Physics, PL 9, 00014 Helsingin yliopisto, Finland
author:
- |
R. G. Unanyan$^1\thanks{Permanent address: Institute for Physical
Research, Armenian National Academy of Sciences,
378410 Ashtarak-2, Armenia}$, N. V. Vitanov$^2$, B. W. Shore$^1\thanks{Permanent address: Lawrence Livermore National
Laboratory, Livermore, CA 94550, USA}$, and K. Bergmann$^1$
title: Coherent properties of a tripod system coupled via a continuum
---
[H]{} § &=&
0[\_0]{}
Introduction {#Sec-intro}
============
Coherent interaction between discrete quantum states via a continuum is an intriguing process. Although the continuum is traditionally seen as an incoherent medium, (partial) transfer of coherence can nevertheless occur through a continuum. In particular, much theoretical and experimental attention has been devoted to laser-induced continuum structure (LICS) [@Fano61; @Knight84; @Knight90; @Pavlov81; @Heller81; @Dai87; @Hutchinson88; @Shao91; @Cavalieri91; @Cavalieri93; @Faucher93a; @Faucher93b; @Faucher94; @Cavalieri95; @Eramo97; @Cavalieri98; @Halfmann98; @Yatsenko99; @Kylstra98], where the interaction between a discrete state $\psi_2$ and a structureless, flat continuum creates a structure in the continuum which affects significantly the interaction of another discrete state $\psi_1$ with this continuum. For example, the ionization probability for state $\psi_1$, when plotted as a function of the frequency of the ionizing laser, exhibits the so-called Fano profile [@Fano61]. The physical nature of LICS is closely related to autoionizing states [@Fano61; @Lambropoulos81; @Nakajima93; @Nakajima94b; @Karapanagioti95a; @Karapanagioti95b; @Nakajima96; @Chen99].
It has been suggested by Carroll and Hioe a few years ago [@Carroll92; @Carroll93] that a continuum can serve as an intermediary for population transfer between two discrete states in an atom or a molecule by using a sequence of two counterintuitively ordered delayed laser pulses. This scheme is an interesting variation of the process of stimulated Raman adiabatic passage (STIRAP) [@Gaubatz88; @Kuklinski89; @Gaubatz90; @Bergmann98] (and references therein) where a discrete intermediate state is used. The Carroll-Hioe analytic model, which involves an unbound quasicontinuum of equidistant states, suggests that complete population transfer is possible, the ionization being suppressed. Later, Nakajima [*et al*]{} [@Nakajima94a] demonstrated that this result derives from the very stringent restrictions of the model which are unlikely to be met in a realistic physical system with a real continuum, in particular with a non-zero Fano parameter $q$. It has subsequently been recognized that although complete population transfer is unrealistic, significant partial transfer may still be feasible [@Carroll95; @Carroll96; @Yatsenko97; @Vitanov97; @Paspalakis97; @Paspalakis98b]. It has been shown that, at least in principle, the detrimental effect of the nonzero Fano parameter and the Stark shifts can be overcome by using the Stark shifts induced by a third (nonionizing) laser [@Yatsenko97] or by using appropriately chirped laser pulses [@Vitanov97; @Paspalakis97]. It has been concluded [@Yatsenko97; @Vitanov97] that the main difficulty in achieving efficient population transfer is related to the incoherent ionization channels, of which at least one is always present and leads to inevitable irreversible population losses. It has been suggested [@Carroll96; @Yatsenko97] that these losses can be reduced (although not eliminated) by choosing an appropriate region in the continuum where the ionization probability is minimal. Later, it has been shown that the incoherent ionization can be suppressed very effectively by using a Fano-like resonance induced by an additional laser from a third state $\psi_3$, resulting in a considerable increase in the transfer efficiency [@Unanyan98a].
In the present paper, we investigate the coherence properties of a scheme comprising [*three*]{} discrete states coupled via a common continuum. This tripod linkage can be viewed as an extension of the standard LICS, involving two discrete states and two lasers, with the inclusion of an extra state by using a third laser. Such a scheme can also appear in standard two-state LICS when the two lasers are tuned near an autoionizing state; the latter is strongly coupled to the continuum by configuration interaction. The present scheme can also be viewed as a variation of the tripod scheme comprising three discrete states coupled via a (common) fourth discrete state [@Unanyan98b; @Theuer99]. In contrast to the three-state scheme in [@Unanyan98a], in which the additional laser used to suppress incoherent ionization is tuned in the continuum much above the region where the main lasers are tuned (thus reducing the coupled three-state dynamics to a pair of two-state LICS systems), here the additional laser is tuned in the same region as the two main lasers, which means that we have to deal with generally irreducible three-state dynamics. Some properties of this tripod scheme have been studied in [@Paspalakis98c] in the particular case when the Fano parameters are equal and the additional state is a strongly coupled autoionizing state. In the present paper we establish the basic properties of this system in the general case of arbitrary Fano parameters and arbitrary strong ionization rates. We derive the population trapping conditions, which are now two, in contrast to the single trapping condition in two-state LICS. Furthermore, we obtain the solution for coincident pulses and the behaviour of the system in the adiabatic limit for delayed pulses, including the optimal conditions for population transfer between the discrete states via the continuum.
This paper is organized as follows. In Sec. \[Sec-tripod\], we introduce the tripod-continuum system, present the basic equations and definitions, and derive the trapping conditions. In Sec. \[Sec-coincident\], we consider the case when all laser fields have the same time dependence. In Sec. \[Sec-delayed\], we examine the case of delayed laser pulses with a special attention to population transfer in the near-adiabatic regime. In Sec. \[Sec-AdbElimination\], we explore the case when the third state $\psi_3$ is strongly coupled to the continuum and eliminate it adiabatically to simplify the dynamics and gain insight of the tripod-continuum interaction. Finally, in Sec. \[Sec-conclusion\] we summarize the conclusions.
Tripod-continuum system {#Sec-tripod}
=======================
The system {#Sec-system}
----------
We shall ignore any continuum-continuum transitions, such as above threshold ionization (ATI) [@Agostini83], which become important only for very high laser intensity. We also neglect spontaneous emission from the bound states, which is justified when these states are ground or metastable or when the interaction time is short compared to the atomic relaxation times. Finally, we ignore incoherent ionization channels [@Yatsenko97; @Vitanov97; @Unanyan98a], i.e., we assume that each laser drives only one transition between a bound state and the continuum.
The total wave function can be written as a linear superposition of the three discrete states and the continuum. We then substitute this expansion into the time-dependent Schrödinger equation and eliminate the continuum using the rotating-wave and Markov approximations [@Knight90]. The time evolutions of the probability amplitudes of the three bound states obey the equation ($\hbar=1$) \[SEq\] id[dt]{}(t) = (t) (t), where $\C(t)=[C_1(t),C_2(t),C_3(t)]^T$. The time-dependent Hamiltonian describing the system separates into real and imaginary parts \[Hamiltonian\] \[H\] = + i,\
\[A\] = -12 -2\_1 & &\
& -2\_2 &\
& & 0 ,\
\[B\] = -12 \_1 & &\
& \_2 &\
& & \_3 , where \_1(t) = \_1 + §\_1(t) - §\_3(t),\
\_2(t) = \_2 + §\_2(t) - §\_3(t). Here $\d_k$ ($k=1,2$) is the two-photon laser detuning between state $\psi_k$ and state $\psi_3$. The quantity $\G_k(t)$ is the ionization rate of $\psi_k$ ($k=1,2,3$), which is proportional to the generally time-dependent (e.g., pulse-shaped) intensity of the corresponding laser. $\S_k(t)$ is the total laser-induced dynamic Stark shift for state $\psi_k$ ($k=1,2,3$), which is a sum of the Stark shifts, induced by each laser and proportional to the corresponding laser intensity. As evident from Eq. (\[A\]) and as shown in Fig. \[Fig-system\], we have chosen the Stark-shifted RWA energy of state $\psi_3$ as the zero energy level. The dimensionless constants $\q$, $\qa$, and $\qb$ are the Fano asymmetry parameters [@Fano61; @Knight84; @Knight90; @Burke65], which characterize the transitions between the corresponding pairs of states via the continuum and depend on the atomic structure. With the exception of the Fano parameters, all other variables involved in Eqs. (\[Hamiltonian\]) can be controlled externally by the laser fields.
We shall assume that the system is initially in state $\psi_1$, \[initial\] C\_1(-) = 1, C\_2(-) = C\_3(-) = 0, and the quantities of interest are the populations of the discrete states at $t\rightarrow +\infty$, $P_k = |C_k(+\infty)|^2$ ($k=1,2,3$), and the ionization probability $P_i = 1 - P_1 - P_2 - P_3$. Because we choose the initial conditions (\[initial\]) and we intend to explore how the additional state $\psi_3$ affects the interaction between states $\psi_1$ and $\psi_2$, we shall refer to $\G_1(t)$, $\G_2(t)$, and $\G_3(t)$ as ionization rates induced by the pump, Stokes, and control lasers, respectively.
Eigenvalues and trapping conditions {#Sec-eigenvalues}
-----------------------------------
It has been shown in [@Gazazyan87] that if the matrices $\A(t)$ and $\B(t)$ commute, \[normal\] (t)(t) = (t)(t), then the eigenvalues of $\H(t)$ read as \[Lk\] \_k(t) = \_k\^A(t) + i\_k\^B(t), (k=1,2,3), where $\lambda_k^A(t)$ and $\lambda_k^B(t)$ are eigenvalues of $\A(t)$ and $\B(t)$, respectively. The importance of relation (\[Lk\]) derives from the fact that the eigenvalues of $\B(t)$ are given by \[Lb\] \_1\^B(t) = \_2\^B(t) = 0, \_3\^B(t) = -12 (t), where \[Gamma\] (t) = \_1(t) + \_2(t) + \_3(t), i.e., $\B(t)$ has two zero eigenvalues which correspond to [*nondecaying*]{} eigenstates of $\H(t)$. The fulfillment of relation (\[normal\]) requires that \[Trap\] \_1(t) = 12+ 12(-)\_2(t),\
\_2(t) = 12+ 12(-)\_1(t). Equations (\[Trap\]) will be referred to as [*the population trapping conditions*]{}. Hence there are two such conditions imposed on the interaction parameters, rather than just one as in two-state LICS. It is easily verified that for $\G_3=0$, Eqs. (\[Trap\]) reduce to the well known trapping condition in LICS [@Knight90], \[TrapLICS\] \_1(t)-\_2(t) = 12 .
Given Eqs. (\[Trap\]), the eigenvalues of $\A(t)$ are \[La\] \_1\^A= a+ ,\
\_2\^A= a- ,\
\_3\^A= -12 \[\_1+ \_2\], with a = 14 \[(\_3-\_1) + (\_3-\_2) + (\_1+\_2)\],\
b = 14 \_3 \[(-)\_1 + (-)\_2 - \_3\] , where all quantities but the $q$’s are time dependent. (For typographic simplicity, here and subsequently we often omit the explicit time argument).
Eigenstates and adiabatic basis
-------------------------------
Important information of the interaction dynamics is contained in the instantaneous eigenstates of $\H(t)$ — the [*adiabatic states*]{}. They are derived readily when the trapping conditions (\[Trap\]) are fulfilled, which we shall assume.
Because of the degeneracy of the two zero eigenvalues of $\B\t$, there is an ambiguity in the corresponding two eigenstates of $\B\t$ since any linear combination of them would be a zero-eigenvalue eigenstate of $\B\t$ too. This implies, in particular, that the zero-eigenvalue eigenstates of $\B\t$ are not necessarily eigenstates of $\A\t$. Any eigenstate of $\A\t$, however, is an eigenstate of $\B\t$, and hence of $\H\t$ too. The common time-dependent eigenstates of $\A\t$, $\B\t$, and $\H\t$ are given by \[phi1\] \_1 = -\
--\
,\
\[phi2\] \_2 = +\
-+\
-,\
\[phi3\] \_3 =\
\
, where the time-dependent angles $\theta$, $\phi$, and $\chi$ are defined by \[angles\] \[theta\] = ,\
\[phi\] = ,\
\[chi\] 2=\
&+&. The use of adiabatic states is appropriate in two cases – in the near-adiabatic regime and for coincident pulses – because then the couplings between the adiabatic states vanish and it is possible to derive analytic estimates for the population dynamics. We shall do this in Secs. \[Sec-coincident\] and \[Sec-delayed\].
The basis of $\stateO_1(t)$, $\stateO_2(t)$, and $\AState_3(t)$
---------------------------------------------------------------
In some cases it is convenient to employ an alternative time-dependent basis composed of states $\stateO_1(t)$, $\stateO_2(t)$, and $\AState_3(t)$, where \[states’\] \_1 =\
-\
0 , \_2 =\
\
-, and $\AState_3(t)$ is the adiabatic state (\[phi3\]). Obviously, the adiabatic states $\AState_1(t)$ and $\AState_2(t)$ are linear superpositions of states $\stateO_1(t)$ and $\stateO_2(t)$, \_1 = \_1 - \_2 ,\
\_2 = \_1 + \_2 . Like states $\AState_1(t)$ and $\AState_2(t)$, states $\stateO_1(t)$ and $\stateO_2(t)$ do not decay; the only decaying state in the ($\stateO_1, \stateO_2, \AState_3$)-basis is $\AState_3(t)$. States $\stateO_1\t$ and $\stateO_2\t$ are (zero-eigenvalue) eigenstates of $\B\t$, but not generally of $\A\t$ and $\H\t$. It can easily be shown that they become eigenstates of $\A\t$ and $\H\t$ only when $\qa=\qb$.
The transformation from the bare-state basis (\[SEq\]) to the ($\stateO_1, \stateO_2, \AState_3$)-basis, $\C(t) = \R(t) \C^\prime(t)$, is carried out by the time-dependent rotation matrix \[R\] = & &\
-& &\
0 & -& . The Schrödinger equation in the new basis reads \[SEq’\] i \^(t) = \^(t) \^(t), with $\C^\prime(t)=[C^\prime_1(t),C^\prime_2(t),C_3(t)]^T$ and (an overdot meaning a time derivative) \[H’\] \^= \^[-1]{}- i\^[-1]{}\
= & - i& -i\
+ i& & i\
i& -i& -12i, where $\G$ is given by Eq. (\[Gamma\]) and \[parameters\] = \[ \_3(\_1+\_2) + (\_1+\_2)\^2\
&& -(\_1+\_2)(\_1+\_2) \],\
= ,\
= -12 (\_1 + \_2),\
= . Note that $\cot 2\chi = (\hbb-\haa)/2\hab$.
Coincident pulses {#Sec-coincident}
=================
The case of equal Fano parameters {#Sec-Equal}
---------------------------------
The above theory allows to derive analytic formulae for the bound-state populations and the ionization probability in the case when all ionization rates have the same time dependence, \[coincident\] \_k(t) = \_k f(t),(k=1,2,3). Then the mixing angles $\theta$, $\phi$, and $\chi$ are constant and the nonadiabatic couplings (which are proportional to derivatives of these angles) vanish identically. The solution can be found by transformation to the adiabatic basis, where the Hamiltonian is diagonal. Let us also assume for simplicity that all Fano parameters are equal, $\q = \qa = \qb \equiv q$. If the population is initially in state $\psi_1$, the populations of the bound states and the ionization after the interaction are easily found to be \[Pcoincident\] \[P1\] P\_1 = \[(\_2+\_3)\^2 + \_1\^2 e\^[-]{}\
&+& 2\_1(\_2+\_3) e\^[-/2]{}12q\],\
\[P2\] P\_2 = (1 + e\^[-]{} - 2 e\^[-/2]{}12q),\
\[P3\] P\_3 = (1 + e\^[-]{} - 2 e\^[-/2]{}12q),\
\[Pi\] P\_i = (1 - e\^[-]{}), where $\g = \g_1 + \g_2 + \g_3$ and \[area\] = \_[-]{}\^ (t) dt. The results are similar when the system is initially in state $\psi_2$ or $\psi_3$; then the initial-state population is given by Eq. (\[P1\]), the populations of the other two states by Eqs. (\[P2\]) and (\[P3\]), and the ionization by Eq. (\[Pi\]). Obviously, a similar population trapping phenomenon as for two-state LICS takes place, limiting the maximum possible ionization probability to $\frac13$ for $\G_1 = \G_2 = \G_3$ (compared to $\frac12$ for two-state LICS).
In Fig. \[Fig-coincident\], the populations (\[Pcoincident\]) are plotted against the pulse area $\area$ for the case of equal ionization rates. As the pulse area increases, the populations tend to their adiabatic limits $P_1 \rightarrow \frac49$, $P_2 = P_3 \rightarrow \frac19$, $P_i \rightarrow \frac13$.
The general case {#Sec-general}
----------------
In the general case of unequal Fano parameters one can still find an analytic solution by an appropriate change of the independent variable (time) and transformation to the adiabatic basis where all nonadiabatic couplings vanish, but the resulting formulae are too cumbersome to be presented here. The qualitative behaviour of the populations (\[Pcoincident\]) remains essentially the same. A simple estimate exists for the maximum possible ionization probability (achieved in the limit of strong ionization rates), which is equal to the initial population of the only decaying adiabatic state $\AState_3\t$ \[Eq. (\[phi3\])\], P\_[i,]{} = \^2\^2= . Hence, the stronger the Stokes and control pulses $\G_2$ and $\G_3$, the smaller the ionization.
Delayed pulses {#Sec-delayed}
==============
Minimal and maximal ionization
------------------------------
### No ionization {#Sec-NoIonization}
It is easily seen from Eq. (\[phi3\]) that when $\theta(-\infty)=0$ and/or $\phi(-\infty)=\frac 12\pi $, the only decaying adiabatic state $\AState_3(t)$ is not populated initially. As Eqs. (\[theta\]) and (\[phi\]) show, this happens when \[CIorder\] \_[t-]{} = 0. In the adiabatic limit state $\AState_3(t)$ remains unpopulated and hence, the ionization probability is zero throughout the interaction, $P_i(t) = 0$. In other words, in the adiabatic limit the ionization probability is zero when the pump pulse is delayed with respect to the Stokes pulse and/or the control pulse. The pulse ordering (\[CIorder\]) generalizes the counterintuitive pulse order in the two-state LICS and provides the most appropriate conditions for coherent processes via the continuum, such as population transfer between the bound states, which we shall discuss in Sec. \[Sec-Transfer\].
### Complete ionization {#Sec-TotalIonization}
As follows from Eq. (\[phi3\]), when $\theta(-\infty)=\frac12\pi$ and $\phi(-\infty)=0$, the decaying state $\AState_3(t)$ is the only adiabatic state populated initially. According to Eqs. (\[theta\]) and (\[phi\]), this happens when the pump pulse $\G_1(t)$ arrives before both the control and Stokes pulses, i.e., \[Iorder\] \_[t-]{} = \_[t-]{} = 0. In the adiabatic regime no population is transferred to the other adiabatic states and the ionization probability is given by $P_i = 1 - |\AState_3(+\infty)|^2$. Since the decay rate of state $\AState_3(t)$ is $\frac12\G(t)$ \[see Eq. (\[Lb\])\], we find that P\_i = 1 - e\^[-]{}, where $\area$ is given by Eq. (\[area\]), i.e., $P_i$ can approach unity for strong ionization rates, even though the trapping conditions (\[Trap\]) are satisfied. The pulse order (\[Iorder\]) generalizes the intuitive pulse order in the two-state LICS.
Population transfer via continuum {#Sec-Transfer}
---------------------------------
### Adiabatic limit {#Sec-AdbLimit}
An intriguing process based on LICS is population transfer between two bound states via a common continuum, which has received considerable attention recently [@Carroll92; @Carroll93; @Nakajima94a; @Carroll95; @Carroll96; @Yatsenko97; @Vitanov97; @Paspalakis97; @Paspalakis98b; @Unanyan98a]. We will show that the tripod system enables the same process, providing at the same time a greater flexibility.
Let us consider the pulse timing when the control pulse $\G_3(t)$ arrives first and disappears last, i.e., \[order\] \_[t]{} = \_[t]{} = 0. As we have shown above (Sec. \[Sec-NoIonization\]), the ionization probability in this case is zero, $P_i(t)=0$, because the only decaying adiabatic state $\AState_3(t)$ is not populated initially. Hence, the population is distributed amongst the bound states throughout the interaction. Suppose also that the Stokes pulse precedes the pump pulse (counterintuitive order), i.e., \_[t-]{} = 0, \_[t+]{} = 0. It follows from Eqs. (\[angles\]) that (-) = 0, (+) = 12,\
(-) = 12,(+) = 12. The initial and final values of $\chi$, however, depend on the Fano parameters. $\bullet$ For $\qa=\qb\neq \q$, we have $\chi(\pm\infty)=0$. Hence, && \_1(-) = \_1,\_1(+) = -\_2,\
&& \_2(-) = \_2,\_2(+) = \_1,\
&& \_3(-) = \_3, \_3(+) = \_3. Thus in the adiabatic limit, the population is transferred from state $\psi_1$ to state $\psi_2$ via the adiabatic state $\AState_1(t)$.
$\bullet$ For $\qa\neq \qb$, we have $\chi(-\infty)=\frac12\pi$ and $\chi(+\infty)=0$. Hence, && \_1() = -\_2,\
&& \_2() = \_1,\
&& \_3() = \_3. Thus in the adiabatic limit, the population returns to the initial state $\psi_1$, staying all the time in the adiabatic state $\AState_2(t)$.
$\bullet$ For $\qa = \qb = \q$, we have $\haa(t)-\hbb(t)=0$ and $\hab(t)=0$ in Eq. (\[H’\]). Hence, states $\stateO_1(t)$ and $\stateO_2(t)$ are degenerate and the coupling between them is given by $\dot\theta(t) \sin\phi(t)$. For the pulse ordering (\[order\]), states $\stateO_1(t)$ and $\stateO_2(t)$ have the following asymptotic behaviour \[see Eqs. (\[states’\])\]: \[limits\] \_1(-) = \_1, \_1(+) = -\_2,\
\_2(-) = \_2, \_2(+) = \_1. Hence, in the adiabatic limit, the bare-state populations are P\_1 && \^2 \_[-]{}\^ (t) (t) dt,\
P\_2 && \^2 \_[-]{}\^ (t) (t) dt,\
P\_3 && 0. The populations of states $\psi_1$ and $\psi_2$ depend only on the angles $\theta(t)$ and $\phi(t)$, which in turn depend on the time delay $\tau$ between $\G_1(t)$ and $\G_2(t)$. This dependence provides the possibility to control the created coherent superposition of $\psi_1$ and $\psi_2$ through the pulse delay. This property of the tripod-continuum system is similar to the one for a discrete tripod system coupled via a discrete state [@Unanyan98b], rather than a continuum, which has been demonstrated experimentally recently [@Theuer99].
### Optimal conditions for population transfer
Although in the general case of $\qa\neq \qb$ the population returns to the initial state $\psi_1$ in the adiabatic limit, it is still possible to transfer population to state $\psi_2$ for certain ranges of interaction parameters. These ranges are most easily determined in the ($\stateO_1,\stateO_2,\AState_3$)-basis which is more convenient than the adiabatic basis. As is evident from the asymptotic limits (\[limits\]) of $\stateO_1(t)$ and $\stateO_2(t)$, only state $\stateO_1(t)$ is populated initially, and if the atom stays in $\stateO_1(t)$ at all times, the desired population transfer from $\psi_1$ to $\psi_2$ will occur. In order to achieve this, transitions from $\stateO_1\t$ to both states $\stateO_2(t)$ and $\AState_3(t)$ must be suppressed. This restriction determines the ranges of interaction parameters for which significant population transfer from $\psi_1$ to $\psi_2$ is possible.
State $\stateO_1(t)$ is coupled to the decaying state $\AState_3(t)$ with a coupling proportional to $\dot\theta(t)$. Hence, the detrimental transitions from $\stateO_1(t)$ to $\AState_3(t)$ can be avoided if the interaction is sufficiently adiabatic, which requires that \[condition12\] |(t)(t)| .
On the other hand, the interaction should not be too adiabatic, because then, as we have shown in Sec. \[Sec-AdbLimit\], the population returns to state $\psi_1$. This conclusion is confirmed when examining the nature of the interaction between states $\stateO_1(t)$ and $\stateO_2(t)$ \[see Eq. (\[H’\])\]. Indeed, the effective detuning in this subsystem $\hbb(t)-\haa(t)$ has different signs at $t\rightarrow\pm\infty$, which means that there is a level-crossing transition and hence, complete population transfer between states $\stateO_1(t)$ and $\stateO_2(t)$ occurs in the adiabatic limit. According to Eqs. (\[limits\]), such a complete transfer means complete population return to $\psi_1$ in the bare-state basis. Obviously, only in the case of $\qa=\qb$, the coupling $\hab(t)$ vanishes identically and $\stateO_1(t)$ and $\stateO_2(t)$ are only coupled by a weak nonadiabatic coupling, which vanishes in the adiabatic limit. However, the case $\qa=\qb$ is exceptional and it is difficult to find atomic states which satisfy this condition. For $\qa \neq \qb$, there is a residual coupling $\hab(t)$ between $\stateO_1(t)$ and $\stateO_2(t)$ which remains nonzero in the adiabatic limit and causes transitions between these states. Refering to the Landau-Zener formula [@Zener32], we conclude that in order to avoid population transfer from $\stateO_1(t)$ to $\stateO_2(t)$, the relation \[condition13\] \[(t\_0)\]\^2 12 | (t\_0) - (t\_0) | must be fulfilled, where $t_0$ is the crossing point: $\haa(t_0) = \hbb(t_0)$. It is possible to refine condition (\[condition13\]) by including effects of asymmetry [@Vitanov96] and nonlinearity [@Vitanov99b] at the crossing and finite transition times [@Vitanov99a].
Conditions (\[condition12\]) and (\[condition13\]) provide the restrictions on the interaction parameters needed for significant population transfer from $\psi_1$ to $\psi_2$.
### Numerical examples
In our numerical simulations we have used Gaussian pulse shapes for $\G_1(t)$ and $\G_2(t)$ and constant $\G_3$, \[shapes\] \_1(t) = \_1 e\^[-(t-)\^2/T\^2]{},\
\_2(t) = \_2 e\^[-(t+)\^2/T\^2]{},\
\_3(t) = \_3, where $2\tau$ is the delay between the pump and Stokes pulses and $T$ is their width.
It is possible to simplify conditions (\[condition12\]) and (\[condition13\]) when the control pulse is much stronger than the pump and Stokes pulses, $\G_3 \gg \G_1,\G_2$. Then for $\g_1=\g_2$ the crossing point is given by $t_0\approx 0$ and conditions (\[condition12\]) and (\[condition13\]) become \[range\] \_3 T . Hence appreciable population transfer from $\psi_1$ to $\psi_2$ is only possible if the difference $|\qa - \qb|$ is sufficiently small.
In Fig. \[Fig-optimal\], the populations of the discrete states and the ionization probability are plotted against the pulse width $T$ of the pump and Stokes pulses. The detunings $\D_1(t)$ and $\D_2(t)$ are chosen to satisfy the trapping conditions (\[Trap\]) at any time; as noted in the introduction, this can be achieved, at least in principle, by using the Stark shifts induced by an additional (nonionizing) laser [@Yatsenko97] or by using appropriately chirped laser pulses [@Vitanov97; @Paspalakis97]. In this case, the Stark shifts $\S_k(t)$ $(k=1,2,3)$ are unimportant because they enter Eq. (\[SEq\]) through $\D_1(t)$ and $\D_2(t)$ only \[which are given the values prescribed by Eqs. (\[Trap\])\], and are therefore set equal to zero. The figure shows that a reasonably high efficiency of population transfer to state $\psi_2$ can be achieved in a certain range of $T$; this range is predicted correctly by condition (\[range\]), which in this case reads as $0.2 \ll \g_3 T \ll 8$. For small $T$, the interaction is nonadiabatic and the population is distributed mainly between the initial state (due to a transition from $\stateO_1$ to $\stateO_2$) and the continuum (due to a transition from $\stateO_1$ to $\AState_3$). As $T$ increases, the interaction becomes increasingly adiabatic and the ionization probability $P_i$ is reduced, as well as the initial-state population $P_1$. For large $T$ the interaction becomes almost completely adiabatic and the population returns to the initial state because of the level crossing transition from $\stateO_1$ to $\stateO_2$.
As Eqs. (\[Trap\]) show, for large and constant $\G_3$, the trapping conditions are satisfied approximately at $\d_1 \approx \frac12 \qa\G_3$ and $\d_2 \approx \frac12\qb\G_3$. The implication is that in this case it may be easier to satisfy the two (constant) trapping conditions for the tripod system than the single (time-dependent) trapping condition (\[TrapLICS\]) for the two-state LICS. In Fig. \[Fig-3D\], the population of state $\psi_2$ is plotted against the sum and the difference of the detunings $\d_1$ and $\d_2$ for three values of the constant ionization rate $\g_3$. For $\g_3=0$ when state $\psi_3$ is uncoupled, $P_2$ depends only on the two-photon detuning $\d_1-\d_2$ between states $\psi_1$ and $\psi_2$, as expected. The figure shows that for $\g_3=\g_0$ and $\g_3=4\g_0$, there is a region in the ($\d_1,\d_2$)-plane, where $P_2$ achieves higher values than for $\g_3=0$ (and also regions where $P_2$ achieves lower values). The maximum transfer efficiency is approximately 0.30 for $\g_3=0$, 0.76 for $\g_3=\g_0$, and 0.95 for $\g_3=4\g_0$. This shows that, indeed, in a certain detuning range it is easier to satisfy the two trapping conditions (\[Trap\]) for the tripod system than the single trapping condition for the two-state LICS. In this example, the Stark shifts $\S_k(t)$ $(k=1,2,3)$ of the three bound states were neglected and set to zero. Their inclusion would not introduce any qualitative change \[because the trapping conditions (\[Trap\]) are not satisfied anyway, even at the maxima\], but could only modify slightly the values of $P_2$.
Effective two-state LICS system {#Sec-AdbElimination}
===============================
Finally, we discuss the case when $\G_3$ is large compared to $\G_1$ and $\G_2$. For example, such a situation arises when state $\psi_3$ is an autoionizing state whose coupling to the continuum (by configuration interaction) is usually much stronger than laser ionization rates. Then we can eliminate state $\psi_3$ adiabatically by setting $dC_3/dt = 0$ in Eq. (\[SEq\]), determining $C_3$ in terms of $C_1$ and $C_2$ from the resulting algebraic equation, and replacing $C_3$ in the other two equations. We also make a (population preserving) phase transformation that shifts the zero energy level to coincide with the modified energy of state $\psi_1$. We thus reduce the initial three-state problem to an effective two-state one, involving states $\psi_1$ and $\psi_2$ only, i \^[ae]{}(t) = (t) \^[ae]{}(t), where $\C^{\rm ae}(t) = [\Cae_1(t),\Cae_2(t)]^T$ and = 12 -i\_1 & -(+ i)\
-(+ i) & 2- i\_2 , with \_1(t) = \_1(t) \^2,\
\_2(t) = \_2(t) \^2,\
= ,\
(t) = \_2 - \_1 + \_2(t) - \_1(t)\
&+& \_2(t) - \_1(t) . Hence we obtain a standard two-state LICS problem with modified ionization rates $\Gae_1\t$ and $\Gae_2\t$, Fano parameter $\qae$ and detuning $\Dae\t$. Hence the presence of a third state, strongly coupled to the continuum, modifies the properties of the two-state problem involving states $\psi_1$ and $\psi_2$. It may happen that the modified parameters, and in particular $\qae$, have more suitable values for observing and investigating LICS and related phenomena, such as population transfer. In particular, if the Fano parameters $\qa$, $\qb$, and $\q$ are large, the effective Fano parameter $\qae$ will be small, which can facilitate the observation of LICS [@Halfmann98; @Yatsenko99].
It is possible to obtain further insight of the tripod-continuum system by adiabatic elimination of the only decaying adiabatic state $\AState_3(t)$ both in the adiabatic basis and in the ($\stateO_1,\stateO_2,\AState_3$)-basis.
Summary and Conclusions {#Sec-conclusion}
=======================
In the present paper we have investigated the coherence properties of a system involving three discrete states coupled to each other by two-photon processes via a common continuum. In this tripod scheme, there exist two population trapping conditions, rather than one as in standard LICS. In some cases, e.g., for strong and constant control pulse, it may be easier to satisfy these conditions than the single trapping condition in standard LICS. Depending on the pulse timing, various effects can be observed. We have derived some basic properties of the tripod scheme, such as the solution for coincident pulses (sharing the same time dependence), the behaviour of the system in the adiabatic limit for delayed pulses, the conditions for no ionization and for maximal ionization, and the optimal conditions for population transfer between the discrete states via the continuum. In the case of a strongly coupled state, by adiabatically eliminating this state, we have found that the tripod scheme reduces to an effective standard two-state LICS system with modified Fano parameter and ionization rates; such modification may provide better conditions for observing and investigating LICS and related phenomena.
Acknowledgements {#acknowledgements .unnumbered}
----------------
RGU thanks the Alexander von Humboldt Foundation for a Fellowship. The work of NVV is supported by the Academy of Finland. BWS thanks the Alexander von Humboldt Foundation for a Research Award; his work is supported in part under the auspices of the U.S. Department of Energy at Lawrence Livermore National Laboratory under contract W-7405-Eng-48.
, Phys. Rev. [**124**]{}, 1866 (1961).
, Comments At. Mol. Phys. [**15**]{}, 193 (1984).
, Phys. Rep. [**190**]{}, 1 (1990).
, Phys. Lett. [**89A**]{}, 441-443 (1981).
, Phys. Lett. [**82A**]{}, 4-6 (1981).
, Phys. Rev. A [**36**]{}, 5205 (1987).
, Phys. Rev. Lett. [**60**]{}, 105-107 (1988).
, Phys. Rev. Lett. [**67**]{}, 3669-3672 (1991).
, Phys. Rev. Lett. [**67**]{}, 3673-3676 (1991).
, Phys. Rev. A [**47**]{}, 4219-4226 (1993).
, Phys. Rev. Lett. [**70**]{}, 3004-3007 (1993).
, J. Phys. B [**26**]{}, L309 (1993).
, Phys. Rev. A [**50**]{}, 641-648 (1994).
, Phys. Rev. A [**51**]{}, 2974 (1995).
, J. Phys. B [**30**]{}, 3789-3796 (1997).
, Phys. Rev. A [**57**]{}, 2915-2919 (1998).
, Phys. Rev. A [**58**]{}, R46-R49 (1998).
, Phys. Rev. A [**59**]{}, 2926-2947 (1999).
, J. Phys. B: At. Mol. Opt. Phys. [**31**]{}, L719-L728 (1998).
, Phys. Rev. A [**24**]{}, 379 (1981).
, Phys. Rev. Lett. [**70**]{}, 1081-1084 (1993).
, Phys. Rev. A [**50**]{}, 595-610 (1994).
, Phys. Rev. Lett. [**74**]{}, 2431-2434 (1995).
á[nchez, and E. Cormier]{}, Phys. Rev. A [**53**]{}, 2587-2597 (1995).
, Z. Phys. D [**36**]{}, 17 (1996).
, Phys. Rev. A [**58**]{}, 1267-1274 (1998).
, Phys. Rev. Lett. [**68**]{}, 3523-3526 (1992).
, Phys. Rev. A [**47**]{}, 571-580 (1993).
U. Gaubatz, P. Rudecki, M. Becker, S. Schiemann, M. Külz, K. Bergmann, Chem. Phys. Lett. [**149**]{}, 463, (1988)
J. R. Kuklinski, U. Gaubatz, F. T. Hioe, and K. Bergmann, Phys. Rev. A [**40**]{}, 6741 (1989).
U. Gaubatz, P. Rudecki, S. Schiemann, and K. Bergmann, J. Chem. Phys. [**92**]{}, 5363 (1990).
K. Bergmann, H. Theuer, and B. W. Shore, Rev. Mod. Phys. [**70**]{}, 1003 (1998).
, Phys. Rev. A [**50**]{}, R913-R916 (1994).
, Phys. Lett. A [**199**]{}, 145 (1995).
, Phys. Rev. A [**54**]{}, 5147-5151 (1996).
, Opt. Commun. [**135**]{}, 406-412 (1997).
, Phys. Rev. A [**56**]{}, 741-747 (1997).
, Opt. Commun. [**142**]{}, 34-40 (1997).
, J. Phys. B: At. Mol. Opt. Phys. [**31**]{}, 775-794 (1998).
, Phys. Rev. A [**57**]{}, 462-466 (1998).
R. G. Unanyan, M. Fleischhauer, B. W. Shore, and K. Bergmann, Opt. Commun. [**155**]{}, 144 (1998).
H. Theuer, R. G. Unanyan, C. Habscheid, K. Klein, and K. Bergmann Opt. Express [**4**]{}, 77-83 (1999).
, J. Phys. B: At. Mol. Opt. Phys. [**31**]{}, 2753-2767 (1998).
, J. Phys. B [**18**]{}, L281 (1983)
, Adv. Phys. [**56**]{}, 521 (1965).
, Sov. Phys. JETP [**66**]{}, 909 (1987).
C. Zener, Proc. Roy. Soc. Lond. A [**137**]{}, 696 (1932).
N. V. Vitanov and B. M. Garraway, Phys. Rev. A [**53**]{}, 4288-4304 (1996); erratum Phys. Rev. A [**54**]{}, 5458 (1996).
N. V. Vitanov and K.-A. Suominen, Phys. Rev. A [**59**]{}, 4580-4588 (1999).
N. V. Vitanov, Phys. Rev. A [**59**]{}, 988-994 (1999).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present 86.8 GHz SiO J=2-1 observations of the molecular outflow from the protostellar source NGC 1333 IRAS 2A. Silicon monoxide is a sensitive tracer of shocks in molecular gas, and the emission clearly traces terminal bowshocks at the ends of the outflow lobes, as well as emission knots and a source coincident with the protostar. Our three-field mosaic from the BIMA interferometer covers the entire outflow length, including both lobes. The most prominent emission feature, the eastern bowshock, shows SiO at the leading edge, clearly displaced from the maximum seen in CO. This is in contrast to recent observations by Chandler & Richer (2001) of the HH 211 jet, where SiO traced primarily internal working surfaces and was less extended than the CO. Our results are consistent with prompt entrainment at the leading edge of the IRAS 2A bowshocks.'
author:
- 'F.P. [^1]'
- 'H.-R.'
- 'L.M.'
- 'R.L.'
title: |
SiO Observations of the NGC 1333\
IRAS 2A Protostellar Jet
---
Summary
=======
The class 0 source IRAS 2A drives an unusual, highly collimated jet-like outflow first studied in CO (Sandell [*et al*]{}. 1994). Such a young, symmetric, and focused outflow provides an excellent opportunity to observe the expected bow shock of a jet interacting with the molecular ambient gas. We used the BIMA interferometer to map this outflow in SiO in order to trace shocks within the outflow. Our initial three-field, C-array (10”$\times 7''$ beam) mosaic covers the positions of IRAS 2A and the east and west bow shocks. In addition to this outflow of roughly E-W orientation, a larger, less collimated bipolar outflow originates from nearly the same position (Liseau [*et al.*]{} 1988), and is possibly associated with the nearby source IRAS 2C (Knee & Sandell 2000; Sandell & Knee 2001). Because the position of IRAS 2C agrees poorly with the symmetry axes of the flows seen in CO, the possibility exists that the two perpendicular outflows seen in CO might actually represent a single, much wider, limb-brightened bipolar outflow (see Figure 1).
![Mosaic of the IRAS 2A region. Thin contors show 12CO J=2-1 taken from the BIMA observations of Engargiola & Plambeck 1999. The greyscale indicates blue-shifted CO. Thick contors show SiO J=2-1 emission. Crosses of decreasing size give the positions of IRAS 2A, 2B, and 2C.[]{data-label="fig:map"}](wilkin_fig1.eps){width="30pc"}
Our recent observations with BIMA in the D-array (13.5” beam) detect SiO in the E-W outflow but not the N-S, strongly suggesting that these are two separate outflows. Nonetheless, it still appears likely that the driving source of the N-S flow remains to be discovered, and further searches for binarity of IRAS 2A are suggested. In Figures 2 and 3 we present the higher resolution wideband and channel maps of SiO in the E-W outflow. The much brighter, redshifted, eastern bow shock shows a wider range of velocities, and is associated with a dense clump seen in CS by Sandell [*et al.*]{}, and whose V-shaped geometry resolved by Blake (1996) suggests a bow shock driven into a steep density gradient. Presumably the eastern bow shock is much brighter because the driving jet is interacting with this dense clump. The actual emission peaks in both the east and west lobes are unresolved in our maps, and higher resolution observations are needed to confirm the expected bow shock geometry at the ends of the outflow lobes.
![3-Field mosaic of IRAS 2A outflow in SiO j=2-1. In this map, positions are relative to the bright, eastern emission peak. The submm point source locations are shown in Figures 1 and 3.[]{data-label="fig:map"}](wilkin_fig2.eps){width="30pc"}
![Channel maps at $5\,{\rm km}/{\rm s}$ intervals. The large, medium and small crosses indicate the positions of IRAS 2A, 2B and 2C, respectively.[]{data-label="fig:channels"}](wilkin_fig3.eps){width="30pc"}
[99]{} Blake, G. A. 1996, IAU Symp. 178: Molecules in Astrophysics: Probes & Processes, 178, 31 Chandler, C.J. & Richer, J.S. 2001, ApJ 555, 139 Engargiola, G. & Plambeck, R. L. 1999, The Physics and Chemistry of the Interstellar Medium, Eds.: V. Ossenkopf, J. Stutzki, and G. Winnewisser Liseau, R., Sandell, G., Knee, L.B.G. 1988, A&A 192, 153 Sandell, G., Knee, L.B.G., Aspin, C., Robson, I.E., & Russell, A.P.G. 1994, 1994, A&A 285, L1 Sandell, G., Knee, L.B.G. 2001, ApJ 546, L49
[^1]: Partially supported by NSF International Researchers Fellows Program
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the design and performance of the LIGO Input Optics subsystem as implemented for the sixth science run of the LIGO interferometers. The Initial LIGO Input Optics experienced thermal side effects when operating with 7 W input power. We designed, built, and implemented improved versions of the Input Optics for Enhanced LIGO, an incremental upgrade to the Initial LIGO interferometers, designed to run with 30 W input power. At four times the power of Initial LIGO, the Enhanced LIGO Input Optics demonstrated improved performance including better optical isolation, less thermal drift, minimal thermal lensing and higher optical efficiency. The success of the Input Optics design fosters confidence for its ability to perform well in Advanced LIGO.'
author:
- 'Katherine L. Dooley'
- 'Muzammil A. Arain'
- David Feldbaum
- 'Valery V. Frolov'
- Matthew Heintze
- Daniel Hoak
- 'Efim A. Khazanov'
- Antonio Lucianetti
- 'Rodica M. Martin'
- Guido Mueller
- Oleg Palashov
- Volker Quetschke
- 'David H. Reitze'
- 'R. L. Savage'
- 'D. B. Tanner'
- 'Luke F. Williams'
- Wan Wu
title: Characterization of thermal effects in the Enhanced LIGO Input Optics
---
Introduction
============
The field of ground-based gravitational-wave (GW) physics is rapidly approaching a state with a high likelihood of detecting GWs for the first time in the latter half of this decade. Such a detection will not only validate part of Einstein’s general theory of relativity, but initiate an era of astrophysical observation of the universe through GWs. Gravitational waves are dynamical strains in space-time, $h =
\Delta L/L$, that travel at the speed of light and are generated by non-axisymmetric acceleration of mass. A first detection is expected to witness an event such as a binary black hole/neutron star merger [@Abadie2010Predictions].
The typical detector configuration used by current generation gravitational-wave observatories is a power-recycled Fabry-Perot Michelson laser interferometer featuring suspended test masses in vacuum as depicted in Figure \[fig:IFOschematic\]. A diode-pumped, power amplified, and intensity and frequency stabilized Nd:YAG laser emits light at $\lambda = 1064$ nm. The laser is directed to a Michelson interferometer whose two arm lengths are set to maintain destructive interference of the recombined light at the anti-symmetric (AS) port. An appropriately polarized gravitational wave will differentially change the arm lengths, producing signal at the AS port proportional to the GW strain and the input power. The Fabry-Perot cavities in the Michelson arms and a power recycling mirror (RM) at the symmetric port are two modifications to the Michelson interferometer that increase the laser power in the arms and therefore improve the detector’s sensitivity to GWs.
![(Color online) Optical layout of a Fabry-Perot Michelson laser interferometer, showing primary components. The four test masses, beam splitter and power recycling mirror are physically located in an ultrahigh vacuum system and are seismically isolated. A photodiode at the anti-symmetric port detects differential arm length changes.[]{data-label="fig:IFOschematic"}](figures/IFOsimple_thesis.pdf){width="1.0\columnwidth"}
A network of first generation kilometer scale laser interferometer gravitational-wave detectors completed an integrated 2-year data collection run in 2007, called Science Run 5 (S5). The instruments were: the American Laser Interferometer Gravitational-wave Observatories (LIGO)[@Abbott2009LIGO], one in Livingston, LA with 4 km long arms and two in Hanford, WA with 4 km and 2 km long arms; the 3 km French-Italian detector VIRGO[@Acernese2008Virgo] in Cascina, Italy; and the 600 m German-British detector GEO[@Luck2006Status] located near Hannover, Germany. Multiple separated detectors increase detection confidence through signal coincidence and improve source localization via waveform reconstruction.
The first generation of LIGO, now known as Initial LIGO, achieved its design goal of sensitivity to GWs in the 40–7000 Hz band, including a record strain sensitivity of $2\times10^{-23}/\sqrt{\mathrm{Hz}}$ at 155 Hz. However, only nearby sources produce enough GW strain to appear above the noise level of Initial LIGO and no gravitational wave has yet been found in the S5 data. A second generation of LIGO detectors, Advanced LIGO, has been designed to be at least an order of magnitude more sensitive at several hundred Hz and above and to give an impressive increase in bandwidth down to 10 Hz. Advanced LIGO is expected to open the field of GW astronomy through the detection of many events per year [@Abadie2010Predictions]. To test some of Advanced LIGO’s new technologies and to increase the chances of detection through a more sensitive data taking run, an incremental upgrade to the detectors was carried out after S5 [@Adhikari2006Enhanced]. This project, Enhanced LIGO, culminated with the S6 science run from July 2009 to October 2010. Currently, construction of Advanced LIGO is underway. Simultaneously, VIRGO and GEO are both undergoing their own upgrades [@Acernese2008Virgo; @Luck2010Upgrade].
The baseline Advanced LIGO design [@AdvLigoSysDesign] improves upon Initial LIGO by incorporating improved seismic isolation [@Robertson2004Seismic], the addition of a signal recycling mirror at the output port [@Meers1988Recycling], homodyne readout, and an increase in available laser power from 8 W to 180 W. The substantial increase in laser power improves the shot-noise-limited sensitivity, but introduces a multitude of thermally induced side effects that must be addressed for proper operation.
Enhanced LIGO tested portions of the Advanced LIGO designs so that unforeseen difficulties could be addressed and so that a more sensitive data taking run could take place. An output mode cleaner was designed, built and installed, and DC readout of the GW signal was implemented [@Fricke2011DC]. An Advanced LIGO active seismic isolation table was also built, installed, and tested [@KisselThesis Chapter 5]. In addition, the 10 W Initial LIGO laser was replaced with a 35 W laser [@Frede2007Fundamental]. Accompanying the increase in laser power, the test mass Thermal Compensation System [@Willems2009Thermal], the Alignment Sensing and Control [@DooleyAngular], and the Input Optics were modified.
This paper reports on the design and performance of the LIGO Input Optics (IO) subsystem in Enhanced LIGO, focusing specifically on its operational capabilities as the laser power is increased to 30 W. Substantial improvements in the IO power handling capabilities with respect to Initial LIGO performance are seen. The paper is organized as follows. First, in Section \[sec:role\] we define the role of the IO subsystem and detail the function of each of the major IO subcomponents. Then, in Section \[sec:problems\] we describe thermal effects which impact the operation of the IO and summarize the problems experienced with the IO in Initial LIGO. In Section \[sec:design\] we present the IO design for Advanced LIGO in detail and describe how it addresses these problems. Section \[sec:performance\] presents the performance of the prototype Advanced LIGO IO design as tested during Enhanced LIGO. Finally, we extrapolate from these experiences in Section \[sec:aLIGO\] to discuss the expected IO performance in Advanced LIGO. The paper concludes with a summary in Section \[sec:summary\].
Function of the Input Optics {#sec:role}
============================
{width="100.00000%"}
The Input Optics is one of the primary subsystems of the LIGO interferometers. Its purpose is to deliver an aligned, spatially pure, mode-matched beam with phase-modulation sidebands to the power-recycled Fabry-Perot Michelson interferometer. The IO also prevents reflected or backscattered light from reaching the laser and distributes the reflected field from the interferometer (designated the *reflected port*) to photodiodes for sensing and controlling the length and alignment of the interferometer. In addition, the IO provides an intermediate level of frequency stabilization and must have high overall optical efficiency. It must perform these functions without limiting the strain sensitivity of the LIGO interferometer. Finally, it must operate robustly and continuously over years of operation. The conceptual design is found in Ref. [@Camp1996InputOutput].
As shown in Fig. \[fig:IOblock\], the IO subsystem consists of four principle components located between the pre-stabilized laser and the power recycling mirror:
- electro-optic modulator (EOM)
- mode cleaner cavity (MC)
- Faraday isolator (FI)
- mode-matching telescope (MMT)
Each element is a common building block of many optical experiments and not unique to LIGO. However, their roles specific to the successful operation of interferometry for gravitational-wave detection are of interest and demand further attention. Here, we briefly review the purpose of each of the IO components; further details about the design requirements are in Ref. [@Camp1997Input].
Electro-optic modulator
-----------------------
The Length Sensing and Control (LSC) and Angular Sensing and Control (ASC) subsystems require phase modulation of the laser light at RF frequencies. This modulation is produced by an EOM, generating sidebands of the laser light which act as references against which interferometer length and angle changes are measured [@Fritschel2001Readout]. The sideband light must be either resonant only in the recycling cavity or not resonant in the interferometer at all. The sidebands must be offset from the carrier by integer multiples of the MC free spectral range to pass through the MC.
Mode cleaner
------------
Stably aligned cavities, limited non-mode-matched (junk) light, and a frequency and amplitude stabilized laser are key features of any ultra sensitive laser interferometer. The MC, at the heart of the IO, plays a major role.
A three-mirror triangular ring cavity, the MC suppresses laser output not in the fundamental TEM$_{00}$ mode, serving two major purposes. It enables the robustness of the ASC because higher order modes would otherwise contaminate the angular sensing signals of the interferometer. Also, all non-TEM$_{00}$ light on the length sensing photodiodes, including those used for the GW readout, contributes shot noise but not signal and therefore diminishes the signal to noise ratio. The MC is thus largely responsible for achieving an aligned, minimally shot-noise-limited interferometer.
The MC also plays an active role in laser frequency stabilization [@Fritschel2001Readout], which is necessary for ensuring that the signal at the anti-symmetric port is due to arm length fluctuations rather than laser frequency fluctuations. In addition, the MC passively suppresses beam jitter at frequencies above 10 Hz.
Faraday isolator
----------------
Faraday isolators are four-port optical devices which utilize the Faraday effect to allow for non-reciprocal polarization switching of laser beams. Any backscatter or reflected light from the interferometer (due to impedance mismatch, mode mismatch, non-resonant sidebands, or signal) needs to be diverted to protect the laser from back propagating light, which can introduce amplitude and phase noise. This diversion of the reflected light is also necessary for extracting length and angular information about the interferometer’s cavities. The FI fulfils both needs.
Mode-matching telescope
-----------------------
The lowest-order MC and arm cavity spatial eigenmodes need to be matched for maximal power buildup in the interferometer. The mode-matching telescope is a set of three suspended concave mirrors between the MC and interferometer that expand the beam from a radius of 1.6 mm at the MC waist to a radius of 33 mm at the arm cavity waist. The MMT should play a passive role by delivering properly shaped light to the interferometer without introducing beam jitter or any significant aberration that can reduce mode coupling.
Thermal problems in Initial LIGO {#sec:problems}
================================
The Initial LIGO interferometers were equipped with a 10 W laser, yet operated with only 7 W input power due to power-related problems with other subsystems. The EOM was located in the 10 W beam and the other components experienced anywhere up to 7 W power. The 7 W operational limit was not due to the failure of the IO; however, many aspects of the IO performance did degrade with power.
One of the primary problems of the Initial LIGO IO [@Adhikari1998Input] was thermal deflection of the back propagating beam due to thermally-induced refractive index gradients in the FI. A significant beam drift between the interferometer’s locked and unlocked states led to clipping of the reflected beam on the photodiodes used for length and alignment control (see Fig. \[fig:IOschematic\]). Our measurements determined a deflection of approximately 100 [rad]{}/W in the FI. This problem was mitigated at the time by the design and implementation of an active beam steering servo on the beam coming from the isolator.
There were also known limits to the power the IO could sustain. Thermal lensing in the FI optics began to alter significantly the beam mode at powers greater than 10 W, leading to a several percent reduction in mode matching to the interferometer [@UFLIGOGroup2006Upgrading]. Additionally, absorptive FI elements would create thermal birefringence, degrading the optical efficiency and isolation ratio with power [@Khazanov1999Investigation]. The Initial LIGO New Focus EOMs had an operational power limit of around 10 W. There was a high risk of damage to the crystals under the stress of the 0.4 mm radius beam. Also, anisotropic thermal lensing with focal lengths as severe as 3.3 m at 10 W made the EOMs unsuitable for much higher power. Finally, the MC mirrors exhibited high absorption (as much as 24 ppm per mirror)–enough that thermal lensing of the MC optics at Enhanced LIGO powers would induce higher order modal frequency degeneracy and result in a power-dependent mode mismatch into the interferometer [@Bullington2008Modal; @Arain2007Note]. In fact, as input power increased from 1 W to 7 W the mode matching decreased from 90% to 83%.
In addition to the thermal limitations of the Initial LIGO IO, optical efficiency in delivering light from the laser into the interferometer was not optimal. Of the light entering the IO chain, only 60% remained by the time it reached the power recycling mirror. Moreover, because at best only 90% of the light at the recycling mirror was coupled into the arm cavity mode, room was left for improvement in the implementation of the MMT.
{width="100.00000%"}
Enhanced LIGO Input Optics Design {#sec:design}
=================================
The Enhanced LIGO IO design addressed the thermal effects that compromised the performance of the Initial LIGO IO, and accommodated up to four times the power of Initial LIGO. Also, the design was a prototype for handling the 180 W laser planned for Advanced LIGO. Because the adverse thermal properties of the Initial LIGO IO (beam drift, birefringence, and lensing) are all attributable primarily to absorption of laser light by the optical elements, the primary design consideration was finding optics with lower absorption [@UFLIGOGroup2006Upgrading]. Both the EOM and the FI were replaced for Enhanced LIGO. Only minor changes were made to the MC and MMT. A detailed layout of the Enhanced LIGO IO is shown in Figure \[fig:IOschematic\].
Electro-optic modulator design
------------------------------
We replaced the commercially-made New Focus 4003 resonant phase modulator of Initial LIGO with an in-house EOM design and construction. Both a new crystal choice and architectural design change allow for superior performance.
The Enhanced LIGO EOM design uses a crystal of rubidium titanyl phosphate (RTP), which has at most 1/10 the absorption coefficient at 1064 nm of the lithium niobate (LiNbO$_3$) crystal from Initial LIGO. At 200 W the RTP should produce a thermal lens of 200 m and higher order mode content of less than 1%, compared to the 3.3 m lens the LiNbO$_3$ produces at 10 W. The RTP has a minimal risk of damage, because it has both twice the damage threshold of LiNbO$_3$ and is subjected to a beam twice the size of that in Initial LIGO. RTP and LiNbO$_3$ have similar electro-optic coefficients. Also, RTP’s $dn/dT$ anisotropy is 50% smaller. Table \[tab:EOMcrystals\] compares the properties of most interest of the two crystals.
units LiNbO$_3$ RTP
--------------------------------------- ------------- ----------- ---------
damage threshold MW/cm$^2$ 280 $>600$
absorption coeff. at 1064 nm ppm/cm $< 5000$ $< 500$
electro-optic coeff. ($n_z^3 r_{33}$) pm/V 306 239
$dn_y/dT$ 10$^{-6}$/K 5.4 2.79
$dn_z/dT$ 10$^{-6}$/K 37.9 9.24
\[tab:EOMcrystals\]
We procured the RTP crystals from Raicol and packaged them into specially-designed, custom-built modulators. The crystal dimensions are $4 \times 4 \times 40$ mm and their faces are wedged by $2.85^\circ$ and anti-reflection (AR) coated. The wedge serves to separate the polarizations and prevents an etalon effect, resulting in a suppression of amplitude modulation. Only one crystal is used in the EOM in order to reduce the number of surface reflections. Three separate pairs of electrodes, each with its own resonant LC circuit, are placed across the crystal in series, producing the three required sets of RF sidebands: 24.5 MHz, 33.3 MHZ and 61.2 MHz. A diagram is shown in Fig. \[fig:EOM\]. Reference [@Quetschke2008ElectroOptic] contains further details about the modulator architecture.
![(Color online) Electro-optic modulator design. (a) The single RTP crystal is sandwiched between three sets of electrodes that apply three different modulation frequencies. The wedged ends of the crystal separate the polarizations of the light. The p-polarized light is used in the interferometer. (b) A schematic for each of the three impedance matching circuits of the EOM. For the three sets of electrodes, each of which creates its own $C_{crystal}$, a capacitor is placed parallel to the LC circuit formed by the crystal and a hand-wound inductor. The circuits provide 50 $\Omega$ input impedance on resonance and are housed in a separate box from the crystal.[]{data-label="fig:EOM"}](figures/EOMthesis.pdf "fig:") ![(Color online) Electro-optic modulator design. (a) The single RTP crystal is sandwiched between three sets of electrodes that apply three different modulation frequencies. The wedged ends of the crystal separate the polarizations of the light. The p-polarized light is used in the interferometer. (b) A schematic for each of the three impedance matching circuits of the EOM. For the three sets of electrodes, each of which creates its own $C_{crystal}$, a capacitor is placed parallel to the LC circuit formed by the crystal and a hand-wound inductor. The circuits provide 50 $\Omega$ input impedance on resonance and are housed in a separate box from the crystal.[]{data-label="fig:EOM"}](figures/EOMcircuit_thesis.pdf "fig:")
Mode cleaner design
-------------------
The MC is a suspended 12.2 m long triangular ring cavity with finesse $\mathcal{F}$=1280 and free spectral range of 12.243 MHz. The three mirror architecture was selected over the standard two mirror linear filter cavity because it acts as a polarization filter and because it eliminates direct path back propagation to the laser [@Raab1992Estimation]. A pick-off of the reflected beam is naturally facilitated for use in generating control signals. A potential downside to the three mirror design is the introduction of astigmatism, but this effect is negligible due to the small opening angle of the MC.
The MC has a round-trip length of 24.5 m. The beam waist has a radius of 1.63 mm and is located between the two 45$^\circ$ flat mirrors, MC1 and MC3. See Figure \[fig:IOschematic\]. A concave third mirror, MC2, 18.15 m in radius of curvature, forms the far point of the mode cleaner’s isosceles triangle shape. The power stored in the MC is 408 times the amount coupled in, equivalent to about 2.7 kW in Initial LIGO and at most 11 kW for Enhanced LIGO. The peak irradiances are 32 kW/cm$^2$ and 132 kW/cm$^2$ for Initial LIGO and Enhanced LIGO, respectively.
The MC mirrors are 75 mm in diameter and 25 mm thick. The substrate material is fused silica and the mirror coating is made of alternating layers of silica and tantala. In order to reduce the absorption of light in these materials and therefore improve the transmission and modal quality of the beam in the MC, we removed particulate by drag wiping the surface of the mirrors with methanol and optical tissues. The MC was otherwise identical to that in Initial LIGO.
Faraday isolator design
-----------------------
The Enhanced LIGO FI design required not only the use of low absorption optics, but additional design choices to mitigate any residual thermal lensing and birefringence. In addition, trade-offs between optical efficiency in the forward direction, optical isolation in the backwards direction, and feasibility of physical access of the return beam for signal use were considered. The result is that the Enhanced LIGO FI needed a completely new architecture and new optics compared to both the Initial LIGO FI and commercially available isolators.
Figure \[fig:FI\] shows a photograph and a schematic of the Enhanced LIGO FI. It begins and ends with low absorption calcite wedge polarizers (CWP). Between the CWPs is a thin film polarizer (TFP), a deuterated potassium dihydrogen phosphate (DKDP) element, a half-wave plate (HWP), and a Faraday rotator. The rotator is made of two low absorption terbium gallium garnet (TGG) crystals sandwiching a quartz rotator (QR) inside a 7-disc magnet with a maximum field strength of 1.16 T. The forward propagating beam upon passing through the TGG, QR, TGG, and HWP elements is rotated by $+22.5^\circ -
67.5^\circ + 22.5^\circ + 22.5^\circ = 0^\circ$. In the reverse direction, the rotation through HWP, TGG, QR, TGG is $-22.5^\circ +
22.5^\circ + 67.5^\circ + 22.5^\circ = 90^\circ$. The TGG crystals are non-reciprocal devices while the QR and HWP are reciprocal.
![(Color online) Faraday isolator photograph and schematic. The FI preserves the polarization of the light in the forward-going direction and rotates it by 90 degrees in the reverse direction. Light from the MC enters from the left and exits at the right towards the interferometer. It is ideally p-polarized, but any s-polarization contamination is promptly diverted $\sim 10$ mrad by the CWP and then reflected by the TFP and dumped. The p-polarized reflected beam from the interferometer enters from the right and is rotated to s-polarized light which is picked-off by the TFP and sent to the Interferometer Sensing and Control (ISC) table. Any imperfections in the Faraday rotation of the interferometer return beam results in p-polarized light traveling backwards along the original input path.[]{data-label="fig:FI"}](figures/FI_cropped2.jpg "fig:"){width="0.9\columnwidth"} ![(Color online) Faraday isolator photograph and schematic. The FI preserves the polarization of the light in the forward-going direction and rotates it by 90 degrees in the reverse direction. Light from the MC enters from the left and exits at the right towards the interferometer. It is ideally p-polarized, but any s-polarization contamination is promptly diverted $\sim 10$ mrad by the CWP and then reflected by the TFP and dumped. The p-polarized reflected beam from the interferometer enters from the right and is rotated to s-polarized light which is picked-off by the TFP and sent to the Interferometer Sensing and Control (ISC) table. Any imperfections in the Faraday rotation of the interferometer return beam results in p-polarized light traveling backwards along the original input path.[]{data-label="fig:FI"}](figures/FI_thesis.pdf "fig:")
### Thermal birefringence
Thermal birefringence is addressed in the Faraday rotator by the use of the two TGG crystals and one quartz rotator rather than the typical single TGG [@Khazanov2000Suppression]. In this configuration, any thermal polarization distortions that the beam experiences while passing through the first TGG rotator will be mostly undone upon passing through the second. The multiple elements in the magnet required a larger magnetic field than in Initial LIGO. The 7-disc magnet is 130 mm in diameter and 132 mm long and placed in housing 155 mm in diameter and 161 mm long. The TGG diameter is 20 mm.
### Thermal lensing
Thermal lensing in the FI is addressed by including DKDP, a negative $dn/dT$ material, in the beam path. Absorption of light in the DKDP results in a de-focusing of the beam, which partially compensates for the thermal focusing induced by absorption in the TGGs [@Mueller2002Method; @Khazanov2004Compensation]. The optical path length (thickness) of the DKDP is chosen to slightly over-compensate the positive thermal lens induced in the TGG crystals, anticipating other positive thermal lenses in the system.
### Polarizers
The polarizers used (two CWPs and one TFP) each offer advantages and disadvantages related to optical efficiency in the forward-propagating direction, optical isolation in the reflected direction, and thermal beam drift. The CWPs have very high extinction ratios ($>10^5$) and high transmission ($>$ 99%) contributing to good optical efficiency and isolation performance. However, the angle separating the exiting orthogonal polarizations of light is very small, on the order of 10 mrad. This small angle requires the light to travel relatively large distances before we can pick off the beams needed for interferometer sensing and control. In addition, thermally induced index of refraction gradients due to the 4.95$^{\circ}$ wedge angle of the CWPs result in thermal drift. However, the CWPs for the Enhanced LIGO FI have a measured low absorption of 0.0013 cm$^{-1}$ with an expected thermal lens of 60 m at 30 W and drift of less than 1.3 [rad]{}/W [@UFLIGOGroup2006Upgrading].
The advantages of the thin film polarizer over the calcite wedge polarizer are that it exhibits negligible thermal drift when compared with CWPs and it operates at the Brewster angle of 55$^\circ$, thus diverting the return beam in an easily accessible way. However, the TFP has a lower transmission than the CWP, about 96%, and an extinction ratio of only 10$^3$.
Thus, the combination of CWPs and a TFP combines the best of each to provide a high extinction ratio (from the CWPs) and ease of reflected beam extraction (from the TFP). The downsides that remain when using both polarizers are that there is still some thermal drift from the CWPs. Also the transmission is reduced due to the TFP and to the fact that there are 16 surfaces from which light can scatter.
### Heat conduction {#sec:heatconduction}
Faraday isolators operating in a vacuum environment suffer from increased heating with respect to those operating in air. Convective cooling at the faces of the optical components is no longer an effective heat removal channel, so proper heat sinking is essential to minimize thermal lensing and depolarization. It has been shown that Faraday isolators carefully aligned in air can experience a dramatic reduction in isolation ratio ($>$ 10-15 dB) when placed in vacuum [@TheVIRGOCollaboration2008Invacuum]. The dominant cause is the coupling of the photoelastic effect to the temperature gradient induced by laser beam absorption. Also of importance is the temperature dependence of the Verdet constant–different spatial parts of the beam experience different polarization rotations in the presence of a temperature gradient [@Barnes1992Variation].
To improve heat conduction away from the Faraday rotator optical components, we designed a housing for the TGG and quartz crystals that provided improved heat sinking to the Faraday rotator. We wrapped the TGGs with indium foil that made improved contact with the housing and we cushioned the DKDP and the HWP with indium wire in their aluminum holders. This has the additional effect of avoiding the development of thermal stresses in the crystals, an especially important consideration for the very fragile DKDP.
Mode-matching telescope design
------------------------------
The mode matching into the interferometer (at Livingston) was measured to be at best 90% in Initial LIGO. Because of the stringent requirements placed on the LIGO vacuum system to reduce phase noise through scattering by residual gas, standard opto-mechanical translators are not permitted in the vacuum; it is therefore not possible to physically move the mode matching telescope mirrors while operating the interferometer. Through a combination of needing to move the MMTs in order to fit the new FI on the in-vacuum optics table and additional measurements and models to determine how to improve the coupling, a new set of MMT positions was chosen for Enhanced LIGO. Fundamental design considerations are discussed in Ref. [@Delker1997Design].
Performance of the Enhanced LIGO Input Optics {#sec:performance}
=============================================
The most convincing figure of merit for the IO performance is that the Enhanced LIGO interferometers achieved low-noise operation with 20 W input power without thermal issues from the IO. Additionally, the IO were operated successfully up to the available 30 W of power. (Instabilities with other interferometer subsystems limited the Enhanced LIGO science run operation to 20 W.)
We present in this section detailed measurements of the IO performance during Enhanced LIGO. Specific measurements and results presented in figures and the text come from Livingston; performance at Hanford was similar and is included in tables summarizing the results.
Optical efficiency
------------------
The optical efficiency of the Enhanced LIGO IO from EOM to recycling mirror was 75%, a marked improvement over the approximate 60% that was measured for Initial LIGO. A substantial part of the improvement came from the discovery and subsequent correction of a 6.5% loss at the second of the in-vacuum steering mirrors directing light into the MC (refer to Fig. \[fig:IOschematic\]). A 45$^\circ$ reflecting mirror had been used for a beam with an 8$^\circ$ angle of incidence. Losses attributable to the MC and FI are described in the following sections. A summary of the IO power budget is found in Table \[tab:pwrbudget\].
Livingston Hanford
--------------------------- ------------ -----------
MC visibility 92% 97%
MC transmission 88% 90%
Composite MC transmission 81% (72%) 87%
FI transmission 93% (86%) 94% (86%)
- TFP loss 4.0% 2.7%
IO efficiency (PSL to RM) 75% (60%) 82%
\[tab:pwrbudget\]
### Mode cleaner losses
The MC was the greatest single source of power loss in both Initial and Enhanced LIGO. The MC visibility, $$V = \frac{P_{\mathrm{in}} - P_{\mathrm{refl}}}{P_{\mathrm{in}}},
\label{eq:vis}$$ where $P_{\mathrm{in}}$ is the power injected into the MC and $P_{\mathrm{refl}}$ the power reflected, was 92%. Visibility reduction is the result of higher order mode content of $P_{\mathrm{in}}$ and mode mismatch into the MC. The visibility was constant within 0.04% up to 30 W input power at both sites, providing a positive indication that thermal aberrations in the MC and upstream were negligible.
88% of the light coupled into the MC was transmitted. 2.6% of these losses were caused by poor AR coatings on the second surfaces of the $45^\circ$ MC mirrors. The measured surface microroughness of $\sigma_{rms}< 0.4$ nm [@1998Component] caused scatter losses of $[4 \pi \sigma_{rms}/\lambda]^2 < 22$ ppm per mirror inside the MC, or a total of 2.7% losses in transmission.
Another source of MC losses is via absorption of heat by particulates residing on the mirror’s surface. We measured the absorption with a technique that makes use of the frequency shift of the thermally driven drumhead eigenfrequencies of the mirror substrate [@Punturo2007Mirror]. The frequency shift directly correlates with the MC absorption via the substrate’s change in Young’s modulus with temperature, $dY/dT$. A finite element model (COMSOL [@COMSOL]) was used to compute the expected frequency shift from a temperature change of the substrate resulting from the mirror coating absorption. The measured eigenfrequencies for each mirror at room temperature are 28164 Hz, 28209 Hz, and 28237 Hz, respectively.
We cycled the power into the MC between 0.9 W and 5.1 W at 3 hour intervals, allowing enough time for a thermal characteristic time constant to be reached. At the same time, we recorded the frequencies of the high Q drumhead mode peaks as found in the mode cleaner frequency error signal, heterodyned down by 28 kHz. See Figure \[fig:MCabsorption\]. Correcting for ambient temperature fluctuations, we find a frequency shift of 0.043, 0.043, and 0.072 Hz/W. As a result of drag-wiping the mirrors, the absorption decreased for all but one mirror, as shown for both Hanford and Livingston in Table \[tab:MCabsorption2\].
![(Color online) Data from the MC absorption measurement. Power into the MC was cycled between 0.9 W and 5.1 W at 3 hour intervals (bottom frame) and the change in frequency of the drumhead mode of each mirror was recorded (top frame). The ambient temperature (middle frame) was also recorded in order to correct for its effects.[]{data-label="fig:MCabsorption"}](figures/MCdrumhead_fixed.pdf){width="1.0\columnwidth"}
mirror Livingston Hanford
-------- -------------------- -----------------
MC1 2.1 ppm (18.7 ppm) 5.8 (6.1 ppm)
MC2 2.0 ppm (5.5 ppm) 7.6 (23.9 ppm)
MC3 3.4 ppm (12.8 ppm) 15.6 (12.5 ppm)
: Absorption values for the Livingston and Hanford mode cleaner mirrors before (in parentheses) and after drag wiping. The precision is $\pm 10\%$.
\[tab:MCabsorption2\]
### Faraday isolator losses
The FI was the second greatest source of power loss with its transmission of 93%. This was an improvement over the 86% transmission of the Initial LIGO FI. The most lossy element in the FI is the thin film polarizer, accounting for 4% of total losses. The integrated losses from AR coatings and absorption in the TGGs, CWPs, HWP, and DKDP account for the remaining 3% of missing power.
Faraday isolation ratio
-----------------------
The isolation ratio is defined as the ratio of power incident on the FI in the reverse direction (the light reflected from the interferometer) to the power transmitted in the reverse direction and is often quoted in decibels: isolation ratio = $10
\log_{10}(P_{\mathrm{in-reverse}}/P_{\mathrm{out-reverse}})$. We measured the isolation ratio of the FI as a function of input power both in air prior to installation and *in situ* during Enhanced LIGO operation.
To measure the in-vacuum isolation ratio, we misaligned the interferometer arms so that the input beam would be promptly reflected off of the $97\%$ reflective recycling mirror. This also has the consequence that the FI is subjected to twice the input power. Our isolation monitor was a pick-off of the backwards transmitted beam taken immediately after transmission through the FI that we sent out of a vacuum chamber viewport. Refer to the “isolation check beam” in Fig. \[fig:IOschematic\]. The in air measurement was done similarly, except in an optics lab with a reflecting mirror placed directly after the FI.
![Faraday isolator isolation ratio as measured in air prior to installation and *in situ* in vacuum. The isolation worsens by a factor of 6 upon placement of the FI in vacuum. The linear fits to the data show a constant in-air isolation ratio and an in-vacuum isolation ratio degradation of 0.02 dB/W.[]{data-label="fig:IR"}](figures/FaradayIR.pdf){width="1.0\columnwidth"}
Figure \[fig:IR\] shows our isolation ratio data. Most notably, we observe an isolation decrease of a factor of six upon placing the FI in vacuum, a result consistent with that reported by Ref. [@TheVIRGOCollaboration2008Invacuum]. In air the isolation ratio is a constant 34.46 $\pm$ 0.04 dB from low power up to 47 W, and in vacuum the isolation ratio is 26.5 dB at low power. The underlying cause is the absence of cooling by air convection. If we attribute the loss to the TGGs, then based on the change in TGG polarization rotation angle necessary to produce the measured isolation drop of 8 dB and the temperature dependence of the TGG’s Verdet constant, we can put an upper limit of 11 K on the crystal temperature rise from air to vacuum. Furthermore, a degradation of 0.02 dB/W is measured in vacuum.
Thermal steering
----------------
We measured the *in situ* thermal angular drift of both the beam transmitted through the MC and of the reflected beam from the FI with up to 25 W input power. Just as for the isolation ratio measurement, we misaligned the interferometer arms so that the input beam would be promptly reflected off of the recycling mirror. The Faraday rotator was thus subjected to up to 50 W total and the MC to 25 W.
Pitch and yaw motion of the MC transmitted and interferometer reflected beams were recorded using the quadrant photodiode (QPD) on the IO table and the RF alignment detectors on the Interferometer Sensing and Control table (see Fig. \[fig:IOschematic\]). There are no lenses between the MC waist and its measurement QPD, so only the path length between the two were needed to calibrate in radians the pitch and yaw signals on the QPD. The interferometer reflected beam, however, passes through several lenses. Thus, ray transfer matrices and the two alignment detectors were necessary to determine the Faraday drift calibration.
![(Color online) Mode cleaner and Faraday isolator thermal drift data. (a) Angular motion of the beam at the MC waist and FI rotator as the input power is stepped. The beam is double-passed through the Faraday isolator, so it experiences twice the input power. (b) Average beam angle per power level in the MC and FI. Linear fits to the data are also shown. The slopes for MC yaw, MC pitch, FI yaw, and FI pitch, respectively, are 0.0047, 0.44, 1.8, and 3.2 [rad]{}/W.[]{data-label="fig:drift"}](figures/forthesis_refldriftx10.pdf "fig:"){width="1.0\columnwidth"} ![(Color online) Mode cleaner and Faraday isolator thermal drift data. (a) Angular motion of the beam at the MC waist and FI rotator as the input power is stepped. The beam is double-passed through the Faraday isolator, so it experiences twice the input power. (b) Average beam angle per power level in the MC and FI. Linear fits to the data are also shown. The slopes for MC yaw, MC pitch, FI yaw, and FI pitch, respectively, are 0.0047, 0.44, 1.8, and 3.2 [rad]{}/W.[]{data-label="fig:drift"}](figures/alldrift.pdf "fig:"){width="1.0\columnwidth"}
Figure \[fig:drift\] shows the calibrated beam steering data. The angle of the beam out of the MC does not change measurably as a function of input power in yaw (4.7 nrad/W) and changes by only 440 nrad/W in pitch. For the FI, we record a beam drift originating at the center of the Faraday rotator of 1.8 [rad]{}/W in yaw and 3.2 [rad]{}/W in pitch. Therefore, when ramping the input power up to 30 W during a full interferometer lock, the upper limit on the drift experienced by the reflected beam is about 100 [rad]{}. This is a thirty-fold reduction with respect to the Initial LIGO FI and represents a fifth of the beam’s divergence angle, $\theta_{div}$ = 490 [rad]{}.
Thermal lensing
---------------
We measured the profiles of both the beam transmitted through the mode cleaner and the reflected beam picked off by the FI at low ($\sim$ 1 W) and high ($\sim$ 25 W) input powers to assess the degree of thermal lensing induced in the MC and FI. Again, we misaligned the interferometer arms so that the input beam would be promptly reflected off the recycling mirror. We picked off a fraction of the reflected beam on the Interferometer Sensing and Control table and of the mode cleaner transmitted beam on the IO table (refer to Fig. \[fig:IOschematic\]), placed lenses in each of their paths, and measured the beam diameters at several locations on either side of the waists created by the lenses. A change in the beam waist size or position as a function of laser power indicates the presence of a thermal lens.
![(Color online) Profile at high and low powers of a pick-off of the beam transmitted through the MC. The precision of the beam profiler is $\pm 5\%$. Within the error of the measurement, there are no obvious degradations.[]{data-label="fig:MC_lensing"}](figures/MCTrans_datafit.pdf){width="1.0\columnwidth"}
![(Color online) Faraday isolator thermal lensing data. With 25 W into the Faraday isolator (corresponding to 50 W in double pass), the beam has a steeper divergence than a pure TEM$_{00}$ beam, indicating the presence of higher order modes. Errors are $\pm
5.0\%$ for each data point.[]{data-label="fig:FI_lensing"}](figures/REFL_datafit.pdf){width="1.0\columnwidth"}
As seen in Fig. \[fig:MC\_lensing\] and \[fig:FI\_lensing\], the waists of the two sets of data are collocated: no thermal lens is measured. For the FI, the divergence of the low and high power beams differs, indicating that the beam quality degrades with power. The $M^2$ factor at 1 W is 1.04 indicating the beam is nearly perfectly a TEM$_{00}$ mode. At 25 W, $M^2$ increases to 1.19, corresponding to increased higher-order-mode content. The percentage of power in higher-order modes depends strongly on the mode order and relative phases of the modes, and thus cannot be determined from this measurement [@Kwee2007Laser].
The results for the MC are consistent with no thermal lensing. The high and low power beam profiles are within each other’s error bars and well below our requirements.
We also measured the thermal lensing of the EOM prior to its installation in Enhanced LIGO by comparing beam profiles of a 160 W beam with and without the EOM in its path. The data for both cross-sections of the beam is presented in Fig. \[fig:EOMlensing\]. We observe no significant thermal lensing in the y-direction and a small effect in the x-direction. An upper limit for the thermal lens in the x-direction can be calculated to be greater than 4 m, which is 10 times larger than the Rayleigh range of the spatial mode. The mode matching degradation is therefore less than 1%. Although a direct test for Advanced LIGO because of the power used, this measurement also serves to demonstrate the effectiveness of the EOM design for Enhanced LIGO powers.
![(Color online) EOM thermal lensing data. The x- and y-direction beam profiles with 160 W through the EOM (closed circles and squares) place a lower limit of 4 m on the induced thermal lens when compared to the beam profiles without the EOM (open circles and squares).[]{data-label="fig:EOMlensing"}](figures/EOMlensing.pdf){width="1.0\columnwidth"}
Mode-matching
-------------
We measured the total interferometer visibility (refer to Eq. \[eq:vis\]) as an indirect way of determining the carrier mode-matching to the interferometer. In this case, $P_{\mathrm{in}}$ is the power in the reflected beam when the interferometer cavities are unlocked and $P_{\mathrm{refl}}$ is the power in the reflected beam when all of the interferometer cavities are on resonance.
The primary mechanisms that serve to reduce the interferometer visibility from unity are: carrier mode-matching, carrier impedance matching, and sideband light. We measured the impedance matching at LLO to be $>$ 99.5%; impedance matching therefore makes a negligible contribution to the power in the reflected beam. We also measured that due to the sidebands, the carrier makes up 86% of the power in the reflected beam with the interferometer unlocked and 78% with the interferometer locked; to compensate, we reduce the total $P_{\mathrm{refl}}/P_{\mathrm{in}}$ ratio by 10%. With the interferometer unlocked, there is also a 2.7% correction for the transmission of the RM.
Initially, anywhere between 10% and 17% of the light was rejected by the interferometer due to poor, power-dependent mode matching. After translating the mode-matching telescope mirrors during a vacuum chamber incursion and upgrading the other IO components, the mode mismatch we measured was 8% and independent of input power. The MMT thus succeeds in coupling 92% of the light into the interferometer at all times, marking both an improvement in MMT mirror placement and success in eliminating measurable thermal issues.
Implications for Advanced LIGO {#sec:aLIGO}
==============================
As with other Advanced LIGO interferometer components, Enhanced LIGO served as a technology demonstrator for the Advanced LIGO Input Optics, albeit at lower laser powers than will be used there. The performance of the Enhanced LIGO IO components at 30 W of input power allows us to infer their performance in Advanced LIGO. The requirements for the Advanced LIGO IO demand are for similar performance to Enhanced LIGO, but with almost 8 times the laser power.
The Enhanced LIGO EOM showed no thermal lensing, degraded transmission, nor damage in over 17,000 hours of sustained operation at 30 W of laser power. Measurements of the thermal lensing in RTP at powers up to 160 W show a relative power loss of $< 0.4\%$, indicating that thermal lensing should be negligible in Advanced LIGO. Peak irradiances in the EOM will be approximately four times that of Enhanced LIGO (a 45% larger beam diameter will somewhat offset the increased power). Testing of RTP at 10 times the expected Advanced LIGO irradiance over 100 hours show no signs of damage or degraded transmission.
The MC showed no measurable change in operational state as a function of input power. This bodes well for the Advanced LIGO mode cleaner. Compared with the Enhanced LIGO MC, the Advanced LIGO MC is designed with a lower finesse (520) than Initial LIGO (1280). For 150 W input power, the Advanced LIGO MC will operate with 3 times greater stored power than Initial LIGO. The corresponding peak irradiance is 400 kW/m$^2$, well below the continuous-wave coating damage threshold. Absorption in the Advanced LIGO MC mirror optical coatings has been measured at 0.5 ppm, roughly four times less than the best mirror coating absorption in Enhanced LIGO, so the expected thermal loading due to coating absorption should be reduced in Advanced LIGO. The larger Advanced LIGO MC mirror substrates and higher input powers result in a significantly higher contribution to bulk absorption, roughly 20 times Enhanced LIGO, however the expected thermal lensing leads to small change ($< 0.5 \%$) in the output mode [@Arain2007Note].
The Enhanced LIGO data obtained from the FI allows us to make several predictions about how it will perform in Advanced LIGO. The measured isolation ratio decrease of 0.02 dB/W will result in a loss of 3 dB for a 150 W power level expected for Advanced LIGO relative to its cold state. However, the Advanced LIGO FI will employ an *in situ* adjustable half wave plate which will allow for a partial restoration of the isolation ratio. In addition, a new FI scheme to better compensate for thermal depolarization and thus yield higher isolation ratios will be implemented [@Snetkov2011Compensation]. The maximum thermally induced angular steering expected is 480 [rad ]{}(using a drift rate of 3.2 [rad]{}/W), approximately equal to the beam divergence angle. This has some implications for the Advanced LIGO length and alignment sensing and control system, as the reflected FI beam is used as a sensing beam. Operation of Advanced LIGO at high powers will likely require the use of a beam stabilization servo to lock the position of the reflected beam on the sensing photodiodes. Although no measurable thermal lensing was observed (no change in the beam waist size or position), the measured presence of higher order modes in the FI at high powers is suggestive of imperfect thermal lens compensation by the DKDP. This fault potentially can be reduced by a careful selection of the thickness of the DKDP to better match the absorbed power in the TGG crystals.
Summary {#sec:summary}
=======
In summary, we have presented a comprehensive investigation of the Enhanced LIGO IO, including the function, design, and performance of the IO. Several improvements to the design and implementation of the Enhanced LIGO IO over the Initial LIGO IO have lead to improved optical efficiency and coupling to the main interferometer through a substantial reduction in thermo-optical effects in the major IO optical components, including the electro-optic modulators, mode cleaner, and Faraday isolator. The IO performance in Enhanced LIGO enables us to infer its performance in Advanced LIGO, and indicates that high power interferometry will be possible without severe thermal effects.
The authors thank R. Adhikari for his wisdom and guidance, B. Bland for providing lessons to K. Dooley and D. Hoak on how to handle the small optics suspensions, K. Kawabe and N. Smith-Lefebvre for their support at LHO, T. Fricke for engaging in helpful discussions, and V. Zelenogorsky and D. Zheleznov for their assistance in preparing for the Enhanced LIGO IO installation. Additionally, the authors thank the LIGO Scientific Collaboration for access to the data. This work was supported by the National Science Foundation through grants PHY-0855313 and PHY-0555453. LIGO was constructed by the California Institute of Technology and Massachusetts Institute of Technology with funding from the National Science Foundation and operates under cooperative agreement PHY-0757058. This paper has LIGO Document Number LIGO-P1100056.
[10]{}
Abadie, J., et al., “[Predictions for the rates of compact binary coalescences observable by ground-based gravitational-wave detectors]{},” Classical and Quantum Gravity **27**, 173001+ (2010).
Abbott, B. P., et al., “[LIGO: the Laser Interferometer Gravitational-Wave Observatory]{},” Reports on Progress in Physics **72**, 076901+ (2009).
Acernese, F., et al., “[The Virgo 3 km interferometer for gravitational wave detection]{},” Journal of Optics A: Pure and Applied Optics **10**, 064009+ (2008).
Lück, H., et al., “[Status of the GEO600 detector]{},” Classical and Quantum Gravity **23**, S71–S78 (2006).
Adhikari, R., P. Fritschel, and S. Waldman, “[Enhanced LIGO]{},” Tech. Rep. T060156, LIGO Laboratory (2006).
Lück, H., et al., “[The upgrade of GEO 600]{},” Journal of Physics: Conference Series **228**, 012012+ (2010).
, “[Advanced LIGO Systems Design]{},” Tech. Rep. T010075, LIGO Laboratory (2009).
Robertson, N. A., et al., “[Seismic isolation and suspension systems for Advanced LIGO]{},” in “Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series,” , vol. 5500 of *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*, Hough, J., and G. H. Sanders, eds. (2004), vol. 5500 of *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*, pp. 81–91.
Meers, B. J., “[Recycling in laser-interferometric gravitational-wave detectors]{},” Physical Review D **38**, 2317–2326 (1988).
Fricke, T., et al., “[DC Readout Experiment in Enhanced LIGO]{},” Classical and Quantum Gravity, accepted for publication (2011).
Kissel, J. S., “Calibrating and improving the sensitivity of the [LIGO]{} detectors,” Ph.D. thesis, Louisiana State University (2010).
Frede, M., B. Schulz, R. Wilhelm, P. Kwee, F. Seifert, B. Willke, and D. Kracht, “[Fundamental mode, single-frequency laser amplifier for gravitational wave detectors]{},” Opt. Express **15**, 459–465 (2007).
Willems, P., A. Brooks, M. Mageswaran, V. Sannibale, C. Vorvick, D. Atkinson, R. Amin, and C. Adams, “[Thermal Compensation in Enhanced LIGO]{},” (2009).
Dooley, K., et al., “[Angular Sensing and Control of the Enhanced LIGO Interferometers]{},” in preparation .
Camp, J., D. Reitze, and D. Tanner, “[Input/Output Optics Conceptual Design]{},” Tech. Rep. T960170, LIGO Laboratory (1996).
Camp, J., D. Reitze, and D. Tanner, “[Input Optics Design Requirements Document]{},” Tech. Rep. T960093, LIGO Laboratory (1997).
Fritschel, P., R. Bork, G. González, N. Mavalvala, D. Ouimette, H. Rong, D. Sigg, and M. Zucker, “[Readout and Control of a Power-Recycled Interferometric Gravitational-Wave Antenna]{},” Appl. Opt. **40**, 4988–4998 (2001).
Adhikari, R., A. Bengston, Y. Buchler, T. Delker, D. Reitze, Q.-z. Shu, D. Tanner, and S. Yoshida, “[Input Optics Final Design]{},” Tech. Rep. T980009, LIGO Laboratory (1998).
, and [IAP Group]{}, “[Upgrading the Input Optics for High Power Operation]{},” Tech. Rep. E060003, LIGO Laboratory (2006).
Khazanov, E. A., O. V. Kulagin, S. Yoshida, D. B. Tanner, and D. H. Reitze, “[Investigation of self-induced depolarization of laser radiation in terbium gallium garnet]{},” IEEE Journal of Quantum Electronics **35**, 1116–1122 (1999).
Bullington, A. L., B. T. Lantz, M. M. Fejer, and R. L. Byer, “[Modal frequency degeneracy in thermally loaded optical resonators]{},” Appl. Opt. **47**, 2840–2851 (2008).
Arain, M., “[A Note on Substrate Thermal Lensing in Mode Cleaner]{},” Tech. Rep. T070095, LIGO Laboratory (2007).
Quetschke, V., “[Electro-Optic Modulators and Modulation for Enhanced LIGO and Beyond]{},” Coherent Optical Technologies and Applications pp. CMC1+ (2008).
Raab, F., and S. Whitcomb, “[Estimation of Special Optical Properties of a Triangular Ring Cavity]{},” Tech. Rep. T920004, LIGO Laboratory (1992).
Khazanov, E., N. Andreev, A. Babin, A. Kiselev, O. Palashov, and D. H. Reitze, “[Suppression of self-induced depolarization of high-power laser radiation in glass-based Faraday isolators]{},” J. Opt. Soc. Am. B **17**, 99–102 (2000).
Mueller, G., R. S. Amin, D. Guagliardo, D. McFeron, R. Lundock, D. H. Reitze, and D. B. Tanner, “[Method for compensation of thermally induced modal distortions in the input optical components of gravitational wave interferometers]{},” Classical and Quantum Gravity **19**, 1793+ (2002).
Khazanov, E., et al., “[Compensation of thermally induced modal distortions in Faraday isolators]{},” IEEE Journal of Quantum Electronics **40**, 1500–1510 (2004).
, “[In-vacuum optical isolation changes by heating in a Faraday isolator]{},” Appl. Opt. **47**, 5853–5861 (2008).
Barnes, N. P., and L. B. Petway, “[Variation of the Verdet constant with temperature of terbium gallium garnet]{},” J. Opt. Soc. Am. B **9**, 1912–1915 (1992).
Delker, T., R. Adhikari, S. Yoshida, and D. Reitze, “[Design Considerations for LIGO Mode-Matching Telescopes]{},” Tech. Rep. T970143, LIGO Laboratory (1997).
“[Component Specification: Substrate, Mode Cleaner Flat Mirror]{},” Tech. Rep. E970148, LIGO Laboratory (1998).
Punturo, M., “[The mirror resonant modes method for measuring the optical absorption]{},” Tech. Rep. VIR-001A-07, VIRGO (2007).
“[COMSOL]{},” http://www.comsol.com.
Kwee, P., F. Seifert, B. Willke, and K. Danzmann, “[Laser beam quality and pointing measurement with an optical resonator.]{}” The Review of scientific instruments **78** (2007).
Snetkov, I., I. Mukhin, O. Palashov, and E. Khazanov, “[Compensation of thermally induced depolarization in Faraday isolators for high average power lasers]{},” Opt. Express **19**, 6366–6376 (2011).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We observe that at the initial stage of a high-current discharge, a low-divergence short x-ray pulse ($\approx\thinspace$$0.5^{\circ}$, $<\thinspace$$2$ ns, $h\nu$$\thinspace>\thinspace$$500$ eV) with the energy of $\sim$$21\enspace\upmu$J is formed over a ferrite surface, which propagates parallel to the surface in the anode direction. The high directionality of the radiation points to its coherent nature. We propose that the radiation is due to the short-lived magnetization of the ferrite surface excited by a high-power electromagnetic pulse. The radiation is coherent due to the equivalent excitation conditions for all emitters. The excitation pulse and the radiation it generates move at the same speed ($\sim$$c$). Thereby, the emitted waves propagating parallel to the ferrite surface are phase-matched, providing the high radiant intensity of the radiation.'
author:
- 'Ivan N. Tilikin'
- 'Nikolai V. Pestovskii'
- 'Sergey A. Pikuz'
- 'Tatiana A. Shelkovenko'
- 'Sergey N. Tskhai'
- 'Sergey Yu. Savinov'
title: 'Coherent x-ray radiation induced by high-current breakdown on a ferrite surface'
---
In the studies of VUV plasma radiation in a high-current discharge on a ferrite surface \[1,2\], a short ($\tau~<~2$ ns) directional pulse of fairly hard coherent radiation ($h\nu >500$ eV) was detected at the initial stage of the discharge. The radiant intensity of this radiation was significantly (by more than an order of magnitude) higher than that of the radiation from the ferrite surface, measured perpendicular to the surface in [@1; @2]. In this paper, we report the first results on the study of this radiation and provide their possible interpretation. The experiments were performed on a BIN generator with an output current amplitude of up to 270 kA and a rise time of 80 ns \[3\]. The impedance of the generator forming line was $\approx 1~\Omega$, and the voltage at the generator output reached 240 kV, with the charging voltage of the forming line being of about 350 kV. The generator load was a rectangular ferrite ((Ni-Zn) $Fe_{2}O_{4}$) prism of grade M1000NN with transverse dimensions 10 $\times$ 20 mm$^{2}$. The prism was mounted perpendicular to the diode axis, see Fig. 1a. By changing the electrode length on the cathode side, the length of the working part of the ferrite prism was varied from 1.5 cm to 7 cm. The current flow path on the ferrite surface was set by a pattern drawn with a graphite pencil. This path was generally the same during consecutive discharges in the experiment \[1,2\]. The generator load was unmatched, and its impedance varied greatly during the pulse. The pressure in the discharge chamber did not exceed $10^{-4}$ Torr.
The radiation from the discharge was studied using calibrated diamond photoconductive detectors (PCDs) with flat spectral response $C=5 \cdot 10^{-4}$ A/W in the energy range from 10 eV to 4 keV. In the high-energy region, the sensitivity smoothly decreases in accordance with the absorption of carbon \[4\]. The transverse size of the detector crystals was $\approx$3 mm, whereas in the detection direction it was 0.5 mm, which provided sufficient sensitivity of the detector up to energies of 10 keV. Detector response time was less than 1 ns. The total time resolution of the recording channel was 2 ns, with the bandwidth of the Tektronix TDS 3104B oscilloscope and the cable lines taken into account. The detectors were typically placed at a distance of 20 cm from the end face of the ferrite prism at various angles with respect to the discharge direction (x axis in Fig. 1a). In each experiment, we used two detectors with gray filters made from a nickel mesh having 2000 lines per inch (78.74 lines per mm) with a transmission of 36%.
![(a) Experimental layout. (b) Angular intensity distribution for the radiation from ferrite. []{data-label=""}](FIG1.pdf){width="80mm"}
The arrangement of the detectors is shown in Fig. 1a. The angular range of radiation detection in the ferrite surface plane xy was $-8^{\circ} < \theta < +22^{\circ}$ (azimuthal angle), and in the xz plane, orthogonal to the ferrite surface, it was $-1^{\circ} < \alpha < +5^{\circ}$ (polar angle). To estimate the width of energy distribution over the cross section of the generated beam, we used uncoated Fuji TR photographic plates sensitive to both x-ray and UV radiation; the plates were placed at a distance of 26 cm from the end face of the ferrite prism. The time dependence of the load current was calculated by numerical integration of the signal from the Rogowski coil (with a bandwidth of $>$500 MHz). The time dependences of the discharge current and the radiation intensities recorded along and perpendicular to the ferrite surface (PCDs in positions 2 and 7) are shown in Fig. 2. The experimental conditions were the following: the length of the working part of the ferrite prism was $l = 6.5$ cm; one detector was placed parallel to the working surface of the prism at a distance $L = 20$ cm from its end face, and the other was placed perpendicular to the working surface at the same distance.
{width="80mm"} {width="60mm"}
It is seen in Fig. 2a that at the initial (prebreakdown) stage of the discharge, when there is almost no discharge current yet, a short ($\tau < 2$ ns) radiation pulse is observed along the discharge axis, with the intensity of this pulse exceeding the intensities of radiation detected in the same time points in the perpendicular direction by an order of magnitude. Note that the actual signal duration is probably shorter since the measured value coincides with the time resolution of the recording channel. The spectral composition of the radiation was estimated as follows: two PCDs were mounted side by side in the xy plane along the discharge axis ($\alpha =0^{\circ}$, $\theta = \pm 2^{\circ}$). One PCD was used with a mesh, as in the previous experiments, and, in front of the other, filters (Al, thickness $d = 50$ $\upmu$m; Be, $d = 10$ $\upmu$m; polypropylene, $d = 4$ $\upmu$m) were installed. The experimental results are given in Table \[tab:table1\].
[lcdr]{} & &\
none & & 100%\
50 $\upmu$m Al & $>5$ keV & 9 %\
4 $\upmu$m PP & $150-280$ eV and $>700$ eV & 93 %\
10 $\upmu$m Be & $>800$ eV & 42 %\
It is seen from the table that over 40% of the radiation falls within the x-ray range ($>$800 eV), and most of the UV radiation has energy above 150 eV. Note that the radiation from the discharge, measured in the direction perpendicular to the ferrite surface, is in the 10–800 eV range \[1,2\]. We studied the angular distribution of radiation intensities. The detectors were placed at a distance $L = 15$ cm from the prism end face. The ferrite prism had a length $l=2$cm. The measurement results are presented in Fig. 1b. It can be seen that the radiation is concentrated in a region with angular sizes of $\sim4^{\circ}$ and $\sim5^{\circ}$ ($\pm 2.5^{\circ}$) in the planes perpendicular and parallel to the working surface of the ferrite prism. Regarding the fact that in our measurement geometry, the angular resolution is $\sim2^{\circ}$, the presented results should be thought of as an evaluation.
Independent estimation of the angular distribution of the investigated radiation was done using the energy distribution over the cross section of the generated beam. Fuji TR photographic plates were used, placed at a distance $L = 26$ cm from the end face of the ferrite prism, with the length of its working part $l = 4.5$ cm ($R = l + L = 30.5$ cm). The result is shown in Fig. 2b. The darkening of the photographic plate corresponds to the time-integrated radiation energy emitted in the given direction, recorded with a 15 $\upmu$m Al filter having two transmission bands: 1.3–1.6 keV and above 3.4 keV. As already indicated, during a series of pulses with a characteristic transverse size of the discharge region of $\sim$200 $\upmu$m, the discharge on the ferrite surface always follows the path, originally set by a graphite pattern \[1,2\].
It is seen from Fig. 1b that the cross section of the beam is characterized by dimensions $s\times~h\approx~1$ cm $\times~0.2$ cm, with a size $s$ close to the transverse size of the ferrite prism; the investigated radiation is likely to be generated over its entire surface. The observed halo may be associated with inhomogeneities on the ferrite surface at different scales, particularly, comparable to the wavelength of the generated radiation. In addition, as noted in the first works considering the discharge on the ferrite surface, the discharge channel modifies the surface, creating a track with an amorphous structure and increased conductivity \[5, 6\]. Radiation scattering on this structure may produce a halo seen in a photographic plate. Taking into account the experimental geometry, the angular divergence of the observed radiation is $\sim0.5^{\circ}$.
The energy characteristics of the radiation were studied. The total energy registered by the detector is determined as follows
$$\varepsilon = \frac{1}{CR} \int_{-\infty}^{\infty} V(t)dt.
\label{eq1}$$
![Dependences of the total radiation energy on the length of the discharge gap obtained along to the ferrite surface (radiation detectors in positions 2 and 7 in Fig. 1a.). Inset shows the corresponding dependence obtained in the perpendicular direction \[1,2\]](FIG3.pdf){width="80mm"}
Here, $R$ = 75 $\Omega$ is the detector load resistance, $V(t)$ is the instantaneous value of the signal, measured in Volts, at time $t$, $C$ is the detector sensitivity. The duration of the radiation pulse may turn out to be significantly less than the time resolution of the recording channel ($\sim$2 ns), and the pulse duration and shape cannot be obtained by electrical measurements only. Nevertheless, since the signal spectral width (see Table \[tab:table1\]) is scarcely beyond the region of the detector spectral sensitivity (10 eV – 4 keV), relation (1) can be used to estimate the total energy incident on the detector. The value of $\int_{-\infty}^{\infty} V(t)dt$ was determined as the area under the curve $V(t)$ (see Fig. 2a). During measurements, the length of the ferrite working part was varied from 1.5 cm to 7 cm by changing the lenght of the negative electrode. In several series of experiments, the length of the working part was first increased (2, 3, 5, 6, and 7 cm), and next decreased (6.5, 5.5, 4.5, 3.5, 2.5, and 1 cm). At each length in each series, 2 or 3 consecutive shots were made, with the total length of the ferrite sample being constant. For estimates, the beam cross section was considered a rectangle with dimensions $s = 1$ cm (width) and $h = 0.2$ cm (height). The value of h is smaller than the transverse dimensions of the detector crystals (detector active area is of circle shape with the radius $r_{D}\approx3$ mm); consequently, the total energy transferred by the radiation pulse is
$$E = \varepsilon \frac{d}{r_{D}}.$$
Squares in Fig. 3 show the measured dependence of the total energy of the directional radiation on the ferrite prism length. It follows from the figure that the radiation energy nonlinearly increases with this length, and the growth rate increases as well. The solid line in the figure corresponds to the approximation of the experimental data by the quadratic dependence of energy on the length of the ferrite working part ($E\sim l^{2}$). The maximum radiation energy at a discharge gap length of 7 cm was $\sim$$21\thinspace\upmu$J. For comparison, the inset of Fig. 3 presents a similar dependence for the total radiation energy obtained perpendicular to the ferrite surface \[1,2\]. It is seen that the dependence has a fundamentally different character: with increasing the ferrite prism length, the total radiated energy increases, but the growth rate decreases significantly.
What is the physical nature of the observed radiation? In view of the sharp asymmetry in the angular distribution of the radiation intensity in the absence of focusing and limiting devices, one can conclude that the radiation is coherent, and the asymmetry in the spatial distribution is due to interference phenomena. We encounter a similar phenomenon when Cherenkov radiation arises, that is electromagnetic radiation of optically transparent media occurs, caused by a charged particle moving in a medium at a speed exceeding the speed of light in this medium \[7–9\]. The Cherenkov radiation condition can be derived considering the interference phenomena and Huygens-Fresnel principle.
For a particle moving with velocity v in an isotropic medium, the wavefronts of these elementary waves are spheres, and their radii increase with speed $u = c/n$, where c is the speed of light in vacuum, $n$ is the refractive index of the medium. According to the Huygens-Fresnel principle, the elementary waves are mutually canceled except for their common envelope. If $v < u$, there is no envelope at all because all the spheres are nested, and there are no intersection points for the wavefronts belonging to different spheres. Consequently, the Cherenkov radiation does not occur if the charged particle moves with the speed $v < u$. If the particle moves with a speed higher than the speed of waves in the medium ($v > u$), the wavefronts of elementary waves intersect, and their common envelope is a cone with the vertex in the instantaneous position of the charged particle. In this case, the elementary waves are enhanced due to interference, and the light generated by the medium at each unit length of the particle trajectory propagates along the generatrix of the cone (Cherenkov cone), with its axis coinciding with the direction of the charged particle motion, and the angle at the vertex being 2$\beta$, where $\beta$ satisfies the well-known relation:
$$\cos \beta = \frac{u}{v} = \frac{c}{nv}.$$
{width="80mm"}
Thus, sharp asymmetry appears in the angular distribution of the radiation intensity in the complete absence of focusing and limiting devices. What happens in our case? We assume that at the prebreakdown stage of the discharge, when a high voltage is applied to the cathode, a longitudinal electric field appears in the interelectrode gap (a bias current arises), which causes the magnetic field in the perpendicular direction. Thus, a high-power magnetic field pulse passes through the discharge gap, traveling over ferrite at a speed of $\sim c$. This pulse induces short-lived magnetization of the ferrite surface. Accordingly, each unit area of the ferrite surface becomes a temporary coherent source of interfering elementary electromagnetic waves, forming the resultant radiation.
The coherence of the radiation is due to the equivalent excitation conditions for all emitters. The excitation pulse and the radiation it generates move in the same direction at the same speed. The envelope of the wavefronts of the elementary waves exists only in a small region near the ferrite surface where the phase matching of the radiation from elementary sources is ensured. As it was mentioned above, the elementary waves are mutually canceled except for their common envelope. Thereby, the emitted radiation has low angular divergence and propagates parallel to the ferrite surface towards the anode. Fig. 4 shows the diagram explaining the formation of the radiation region as a result of coherent addition of elementary electromagnetic waves.
In conclusion, we emphasize the important feature and novelty of the considered phenomenon: no optically transparent medium with a refractive index $n$ and no charge moving at a speed $v > c/n$ are involved. The studied radiation is formed as the electromagnetic excitation pulse passes over the surface of the ferrite prism. This pulse and the radiation it generates move in the same direction at the same speed, and the radiation region is formed as a result of coherent addition of elementary electromagnetic waves. Consequently, the radiation with high radiant intensity is emitted.
This study was supported by the Russian Science Foundation, project no. 19-79-30086.
[9]{} I.N. Tilikin, S.N. Tskhai, T.A. Shelkovenko, S.Yu. Savinov, and S.A. Pikuz, Plasma Phys. Rep. [**44**]{}, 600 (2018).
I.N. Tilikin, S.N. Tzhai, T.A. Shelkovenko, S.Yu. Savinov, S.A. Pikuz, and A.R. Mingaleev, IEEE Trans. on Plasma Sci. [**46**]{}, 3982 (2018).
S.A. Pikuz, T.A. Shelkovenko, and D.A. Hammer, Plasma Phys. Rep. [**41**]{}, 291 (2015).
R.B. Spielman, L.E. Ruggles, R.E. Pepping, S.P. Breeze, J.S. McGurn, and K.W. Struve, Rev. Sci. Instrum [**68**]{}, 782 (1997).
K. Watanabe, S. Kashiwabara, and R. Fujimoto, Appl. Phys. Lett. [**50**]{}, 629 (1987).
S. Kashiwabara, K. Watanabe, and R. Fujimoto, Journ. of Appl. Phys. [**62**]{}, 787 (1987).
P.A. Cerenkov, Comptes Rendus Acad. Sci. URSS [**8**]{} (2), 451 (1934) \[P.A. Cerenkov, Doklady Akademii Nauk SSSR [**2**]{} (8), 451 (1934) (In Russian)\].
P.A. Cerenkov, Phys. Rev. [**52**]{}, 378 (1937).
I.M. Frank and I.E. Tamm, Comptes Rendus Acad. Sci. URSS [**14**]{}, 109 (1937) \[I.M. Frank and I.E. Tamm, Doklady Akademii Nauk SSSR [**14**]{}, 107 (1937) (In Russian)\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Cuprate films offer a unique opportunity to observe vortex tunneling effects, due to their unusually low superfluid density and short coherence length. Here, we measure the magnetoresistance (*MR*) due to vortex motion of a long meander line of a superconducting film made of underdoped $YBa_2Cu_3O_{7-\delta}$. At low temperatures (*T*), the *MR* shows a significant deviation from Arrhenius activation. The data is consistent with two dimensional Variable Range Hopping (VRH) of single vortices, i.e. $MR\propto exp[-(T_0/T)^{1/3}]$. The VRH temperature scale $T_0$ depends on the vortex tunneling rates between pinning sites. We discuss its magnitude with respect to estimated parameters of the meander thin film.'
author:
- 'G. Koren, Y. Mor, A. Auerbach and E. Polturak'
bibliography:
- 'AndDepBib.bib'
- 'apssamp.bib'
title: 'Quantum vortex tunneling in $YBa_2Cu_3O_{7-\delta}$ thin films'
---
Very soon after the discovery of the high temperature superconductors, the resistive transition was observed to broaden under magnetic fields, instead of shifting to lower temperatures [@Palstra]. This was explained as due to thermally activated vortex motion at high temperatures just below $T_c$, which gave rise to an induced voltage across the superconductor and thereby to a flux flow resistance $R_{ff}$. At any given magnetic field in this regime, this resistance has the form of an Arrhenius law $R_{ff} \propto exp[-(U_0/k_BT)]$, where $U_0$ is the activation energy which was discussed by several authors . Generally, flux flow and flux creep are possible at high temperatures where the pinning is relatively weak compared to the thermal energy. At low temperatures, where the pinning is stronger and thermal activation is much weaker, the dominant mechanism for flux motion is via quantum tunneling. Measurements of magnetic relaxation and transport by Stein *et al.* [@Stein] have shown a signature of quantum flux creep in $Y_{1-x}Pr_xBa_2Cu_3O_{7-\delta}$ crystals. This behavior, where the creep is temperature independent, was observed in a very limited temperature range of about 2-4 K. Vortex tunneling in a 2D superconductor at temperatures much lower than the transition temperature $T_c$ was discussed theoretically by Fisher, Tokuyasu and Young (FTY) [@Fisher], and more recently by Auerbach, Arovas and Gosh (AAG) [@AAG]. FTY studied quantum vortex tunneling in 2D films at T close to zero near the superconductor to insulator glass transition. They found that this tunneling occurs via variable range hoping (VRH), but instead of the usual 1/3 power law in the exponent, at high fields they predicted a different behavior where the exponent ranges between 2/3 and 4/5. This resulted from taking into account vortex-vortex interactions. At low fields ($H<<H_{c2}$), AAG calculated the tunneling rate of a single vortex between two pinning sites and the resulting flux tunneling resistivity. They found that this resistivity depends on temperature as the well known VRH in 2D, namely $\rho\propto
exp[-(T_0/T)^{1/3}]$. In the present study we set up an experiment to test this AAG prediction in a thin film of $YBa_2Cu_3O_{7-\delta}$ (YBCO) patterned into a very long and narrow meander line. We note that in a short microbridge of YBCO of typically a few hundred $\mu m$ length, and under a magnetic field of several Tesla, the voltage induced by moving vortices becomes immeasurably small, and in practical terms, a critical current develops already at about 10-20 K below $T_c$. This prohibits measurements of $R_{ff}$ at low bias and low temperatures. In contrast, in a much longer meander line, the total DC component of the voltage generated by moving vortices is large, and a de-facto resistive state persists down to very low temperatures, thus enabling the vortex tunneling study.\
A question still arises as for why quantum vortex tunneling has not been unequivocally identified in low $T_c$ superconductors? We believe that this is due to the fact that conventional BCS superconductors are generally three dimensional and characterized by a large coherence length and high superfluid density, all of which greatly inhibit vortex tunneling. In contrast, underdoped cuprate films, such as investigated here, offer a unique system of low superfluid density and quasi 2D superconductivity with very short coherence lengths. According to a recent theoretical work [@AAG], in such two dimensional “bosonic” superconductors, vortex tunneling can be manifested in terms of deviation from Arrhenius thermal activation of the magnetoresistance. Therefore, observing the $(T_0/T)^{1/3}$ exponent would be an exciting signature of the bosonic character of cuprate superconductivity as well as a confirmation of short length-scale vortex tunneling effects. Concerning previous experimental results, we note that the phenomenon of vortex quantum creep has been investigated by both magnetization relaxation and transport in three dimensional cuprate crystals [@Stein]. Unfortunately, although a temperature independent relaxation time and voltage drops have been observed at low temperatures, these experiments did not necessarily probe the low current regime where tunneling (rather than above barrier classical diffusion) is expected to dominate. The transport results explicitly show non-saturation at low currents which indicates that no vortex tunneling occurred in the low current, Ohmic regime. In our experiments, we study thin films with low current bias. This ensures that deviation from Arrhenius activation which we find is indeed related to vortex tunneling and not to classical above-barrier motion.\
In the present study, we used epitaxial *c-axis* oriented films of YBCO, with the magnetic field applied parallel to this axis and perpendicular to the substrate wafer. The films were prepared by 355 nm laser ablation deposition on (100) $SrTiO_3$ wafers of 10$\times10\, mm^2$ area. We investigated films of 50 and 100 nm thickness. In order to reduce their contact resistance, some films were coated with a 20 nm thick gold layer deposited at a high temperature (780 $^\circ$C) in oxygen ambient. This gold overlayer consists of loosely connected ball-like grains as seen by atomic force microscope images. Its contribution to the resistance versus temperature of the bilayer was non-metallic and very small. This was verified in a control experiment in which another bilayer of 20 nm gold on 100 nm YBCO was prepared with the YBCO doped to have a very low $T_c$. Thus, its resistance could be measured down to very low temperatures. This bilayer was measured as deposited, without patterning. After measuring the resistance, the Au layer was removed using Ar ion milling, and the resistance was measured again. A simple calculation of two parallel resistors shows that the resistivity of the gold layer below 10 K was above 5 $m\Omega\,cm$. This value is much higher than the resistivity of the 60 K YBCO phase which is about 0.1 $m\Omega\,cm$ at 100 K. Considering also the 1:5 thickness ratio between the gold and YBCO layers used, we find that the presence of the gold layer changed the resistance of the meander line by about 0.25$\%$, which is negligible. Nevertheless, the aim of obtaining a low contact resistance with the YBCO film where the contact is made normal to the surface was certainly achieved. We note that the *MR* results measured with and without the gold layer were basically the same. However, those obtained with the gold overcoating were less noisy, and therefore we chose to present them here.\
The YBCO films and Au/YBCO bilayers were annealed *in-situ* in a controlled oxygen ambient to have a transition temperature of 50-60 K, in order to avoid apparent critical currents at low temperatures. We note that the presence of a critical current interferes with the resistance studies at low bias where the current should be proportional to the voltage. The films were patterned by deep UV lithography using a PMMA resist and Ar ion beam milling, into a 4 m long meander line. This meander line had 14 $\mu m$ lines-width with 4 $\mu m$ lines-spacing on $8\times
9\,mm^2$ area of the wafer, while the remaining area was used for the four $1.5\times 1.5\,mm^2$ contacts. In the patterning process, we limited the development time of the photo-resist in order to keep it continuous, but this left several shorts, effectively shortening the meander line. In addition, we found that in the patterning process a few defects were formed in the long meander line which we bridged with small silver paste dots of $\sim 0.5\,mm$ diameter. The gold overcoating layer was also very helpful in improving the contacts to these bridging dots. This bridging procedure and the shorts mentioned before led to a reduced effective length of our meander lines of about 1 m. This length was estimated from the measured resistance, the resistivity and the cross section area of the bare YBCO meander lines without the gold coating. For the transport measurements we used the standard four contacts technique in a dc mode, and cooling was done in a liquid helium cryostat with a base temperature of 2 K.\
 Resistance versus temperature of the YBCO meander line (of $\sim$1m length and $100\,nm \times 14\,\mu m$ cross section area) under various cooling conditions with and without a magnetic field. The resistance of the meander line over a wider temperature range under 2 T field cooling is shown in the inset.](Fig1.eps){height="7cm" width="9cm"}
 The resistance data of Fig. 1 plotted on a log scale versus inverse temperature. The standard vortex activation regions where $R\propto exp(-U_0/kT)$ are shown for fields of 2 T and 6 T.](Fig2.eps){height="7cm" width="9cm"}
The resistance versus temperature results of the meander line are shown in Fig. 1 for zero field cooling (ZFC) and under magnetic fields of 1, 2, 4, and 6 T applied normal to the wafer. One can see the typical behavior characteristic of underdoped YBCO in the 2-300 K temperature range in the inset to this figure. In the main panel, the superconducting transition temperature range is shown in more detail. The well known thermally activated broadening of the transition under increasing magnetic field is clearly observed, and in addition a significant resistive tail is seen already at zero field. To see the thermal activation functional dependence, and the data at low temperatures, the data of Fig. 1 was reploted in Fig. 2 on a log scale versus inverse temperature. One can see that the regimes where thermal activation holds are small, about 5 K at 2 T and 7 K at 6 T. The onset of the superconducting transition under ZFC is $T_c^{onset}=58\,K$, while that of the 6 T field cooled curve is shifted down to about 40 K. This is qualitatively similar to the results of Palstra *et al.*, but with about twice the measured shift of the 90 K phase of YBCO single crystals [@Palstra1990]. Also seen in Fig. 2 is a knee in the ZFC curve at 42-34 K just below the main transition. This is a clear signature of weak links, which is very similar to the results on grain boundary junctions [@Gross]. These weak links can be due to defects in the STO substrate which are copied into the epitaxial film, presence of grains of the minority ($\sim$3%) *a-axis* oriented phase, lithography induced defects and so on. It is therefore clear that any possible vortex tunneling unmasked by other effects could be observed only at temperatures below about 20-30 K.\
 The magnetoresistance $MR=R(H)-R(0)$ derived from the data of Fig. 1, plotted on a log scale as a function of $T^{-1/3}$. The straight lines are linear fits to the data in the relevant low temperature regimes. In the inset the slopes of these straight lines (on a ln scale) $T_0^{1/3}$ are plotted versus field, together with a fit to $4.1/B^{1/8}$. ](Fig3.eps){height="7cm" width="9cm"}
 The data of Fig. 3 plotted versus inverse temperature to compare possible standard vortex activation to vortex tunneling by VRH as in Fig. 3. The straight lines are guides to the eye. ](Fig4.eps){height="7cm" width="9cm"}
Next we study the temperature dependence of the magnetoresistance of the meander line. We use the usual definition of $MR=R(H)-R(0)$ where $R(H)$ and $R(0)$ are the resistances under a magnetic field H and at zero field, respectively. Using the magnetoresistance rather than the resistance under field itself has the advantage of eliminating all extrinsic effects such as those contributed by weak links and series resistors of the repairing silver paste dots. We can safely assume that these contributions to the magnetoresistance are small since the relevant defects occupy a very small fraction of the area of the meander line, and their resistance has a much weaker dependence on the perpendicular magnetic field. In Fig. 3 the $MR$ is plotted versus $T^{-1/3}$ where a linear dependence should indicate VRH in the YBCO planes, while in Fig. 4 the $MR$ is plotted versus $1/T$ to test for a possible Arrhenius activation process. As one can see a linear behavior is found in both figures at low temperatures, but the VRH range of Fig. 3 extends over 2-10 K which is about three times as large as the activation range of 2-5 K as seen in Fig. 4. Although this temperature range is relatively small, the fact that the $log(MR) \propto T^{-1/3}$ behavior exists in a significantly larger temperature range suggests that we actually observe VRH as predicted by AAG [@AAG]. Their expression for the magneto-resistivity in a 2D superconductor is:
$$MR(B,T)=\big{(}\frac{h}{2e}\big{)}^2\gamma_0[n_v(B)]exp[-(T_0/T)^{1/3})]$$
where $\gamma_0$ is the vortex conductivity which depends on the vortex density $n_v(B)$, and $T_0$ is given by:
$$T_0=K\delta \overline{V}\big{(}\frac{\pi n_s}{n_{pin}}\big{)}^2$$
where $K$ is a dimensionless factor of order unity, $\delta \overline{V}$ is the average variation of the pinning energy, $n_s$ is the pairs density and $n_{pin}$ is the pinning sites density. It should be noted that this theory strictly applies only to the 2D case, while the actual film has a quasi 3D nature with pancake vortices which have very weak Josephson coupling in the c direction. The vortices in the film, form straight lines which minimize the total pinning potentials of the layers. The minimal instanton action (wavefunction overlap) for the tunneling events involves motion of an individual vortex which has a pinning site in its close neighborhood. After the tunneling event, the vortex lines up to its new position via classical (above barrier) relaxation. This satisfies the ingredients of the variable range hopping model, where the mean tunneling distance in a thin film is of the order of $l_{2D}/\sqrt{ N_{layers}}$, where $l_{2D}$ is the mean distance between uncorrelated pinning sites in each 2D layer, and $N_{layers}$ is the number of layers in the film. Therefore, for the quasi 3D case of a thin film, Eq. (2) has to be modified by replacing $n_{pin}$ by $n_{pin}\times N_{layers}$ which yields:
$$T_0(film)=K\delta \overline{V}\big{(}\frac{\pi
n_s}{n_{pin}N_{layers}}\big{)}^2.$$
One can see that for any given field the results of Fig. 3 at low temperatures are in good agreement with Eq. (1). Although the FTY theory [@Fisher] is applicable only at high fields, we also tested their prediction of $\rho \propto exp[-(T_0/T)^{2/3})]$, but this dependence fit our $MR$ data only in the narrow regime of 2-5 K. In the inset of Fig. 3, the $T_0^{1/3}$ coefficient of Eq. (1) is plotted versus field. We see that this coefficient is not exactly a constant as assumed for low fields by the theory, but has a small decreasing contribution of about 30% with increasing magnetic field. This behavior can be explained by the effect of decreasing barrier height for vortex tunneling when the vortex-vortex interaction is increased with increasing field.\
Next we make a consistency check of our data at H=2 T with Eq. (3). From the data of Fig. 2 just below $T_c$ we find that the pinning energy of our films is $U_0=V\approx 550\,K$. For $\delta
\overline{V}$ of about 10% of $V$ one gets $\delta
\overline{V}\approx 55\,K$. The measured $T_0$ from the inset of Fig. 3 at 2 T is $3.8^3\approx 55\,K$. Thus, for our data Eq. (3) implies that $\pi n_s \approx n_{pin}N_{layers}$. For pairs, $n_s$ equals half the doping $p$ per copper in the $CuO_2$ plane. In our 60 K YBCO phase, $p\approx 0.12$ [@Segawa] which yields $n_s\approx 0.06$ per copper. Thus $n_{pin}\approx \pi 0.06/170$ per copper in a single $CuO_2$ plane, where $N_{layers}\approx
170$ is the number of $CuO_2$ planes in the 100 nm thick YBCO film. This implies that the average distance between pinning sites in a single $CuO_2$ plane is $l_{2D}\approx 3.9/\sqrt{\pi
0.06/170}\approx 117\,\rm\AA$ which is very reasonable. (the 3.9$\rm \,\AA$ here is the in-plane lattice constant of YBCO).\
 Magnetoresistance as a function of magnetic field at a few temperatures. The straight lines are just connecting the data points, while the parabolic curve is a quadratic fit to the data at 10 K for fields of 1-6 T.](Fig5.eps){height="7cm" width="9cm"}
In Fig. 5 we plot the measured magnetoresistance $MR$ versus magnetic field for a few temperatures. One can see that at the low temperatures of 2.5 and 5 K the $MR$ is almost linear in field. In contrast, at higher temperatures the dependence on the field is more complex, as seen from the data at 7.5 and 10 K. The linear $MR$ dependence on field at a constant temperature can be attributed to a constant terminal vortex velocity in the YBCO film. This would give rise to an induced voltage across the meander line which increases linearly with field. The behavior of the $MR$ at higher temperatures and higher fields can originate in the weak links which are distributed along the meander line. Under these conditions, the weak links become normal, and this gives rise to an $MR$ which reflects the distribution of the strength of the weak links which can be quite complex.\
Finally, we point out that the main difference between the present study and the work on the vortex glass dynamics [@Fisher], is that here we probed the low field (low vortex density) regime. The single vortex tunneling theory [@AAG] predicts a magnetoresistance linear in H as observed in Fig. 5, and ignores vortex interaction effects which are of higher order in the magnetic field. Experimentally, this is justified since at the highest field used of 6 T, the distance between vortices is of about 20 nm, while the distance between pinning sites in the whole film is only $117/\sqrt{170}\approx 0.9$ nm. Therefore, collective tunneling effects might well be relevant only at high fields and temperatures. Since our experiment probes the low field regime, (much lower than $H_{c2}$), we expect single vortex variable range hopping to dominate our results.\
In conclusion, vortex variable range hopping in two dimensions is consistent with our experimental magnetoresistance results in a long meander line of underdoped YBCO thin films at low temperatures.\
[*Acknowledgments:*]{} AA acknowledges useful discussions with Steve Kivelson and Bert Halperin. This research was supported in part by the Israel Science Foundation (grant \# 1564/04), the US-Israel Binational Science Foundation, the Heinrich Hertz Minerva Center for HTSC, the Karl Stoll Chair in advanced materials, and by the Fund for the Promotion of Research at the Technion.\
[99]{} \[Bib\]
T. T. M. Palstra, B. Batlogg, L. F. Schneemeyer and J. V. Waszczak, Phys. Rev. Lett. **61**, 1662 (1988). P. W. Anderson and Y. B. Kim, Rev. Mod. Phys. **36**, 39 (1964). Y. Yeshurun and A. P. Malozemoff, Phys. Rev. Lett. **60**, 2202 (1988). M. Tinkham, Phys. Rev. Lett. 61, 58 (1988). T. Stein, G. A. Levin, C. C. Almasan, D. A.Gajewski and M.B.Maple, Phys. Rev. Lett. **82**, 2955 (1999). M. P. A. Fisher, T. A. Tokuyasu and A. P. Young, Phys. Rev. Lett. **66**, 2931 (1991). A. Auerbach, D. P. Arovas and S. Ghosh, Phys. Rev. B **74**, 064511 (2006). T. T. M. Palstra, B. Batlogg, R. B. Van Dover L. F. Schneemeyer and J. V. Waszczak, Phys. Rev. B **41**, 6621 (1990). R. Gross, P. Chaudhari, D. Dimos, A. Gupta and G. Koren, Phys. Rev. Lett. **64**, 228 (1990). Y. Ando and K. Segawa, Phys. Rev. Lett. **88**, 167005 (2002).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we consider a simple cooperative network consisting of a source, a destination and a cluster of decode-and-forward half-duplex relays. At each time-slot, the source and (possibly) one of the relays transmit a packet to another relay and the destination, respectively, resulting in inter-relay interference (IRI). In this work, with the aid of buffers at the relays, we mitigate the detrimental effect of IRI through interference cancellation. More specifically, we propose the min-power scheme that minimizes the total energy expenditure per time slot under an IRI cancellation scheme. Apart from minimizing the energy expenditure, the min-power selection scheme, also provides better throughput and lower outage probability than existing works in the literature. It is the first time that interference cancellation is combined with buffer-aided relays and power adaptation to mitigate the IRI and minimize the energy expenditure. The new relay selection policy is analyzed in terms of outage probability and diversity, by modeling the evolution of the relay buffers as a Markov Chain (MC). We construct the state transition matrix of the MC, and hence obtain the steady state with which we can characterize the outage probability. The proposed scheme outperforms relevant state-of-the-art relay selection schemes in terms of throughput, diversity and energy efficiency, as demonstrated via examples.'
author:
- |
Nikolaos Nomikos, Themistoklis Charalambous, Ioannis Krikidis, \
Dimitrios Skoutas, Demosthenes Vouyioukas, \
and Mikael Johansson, [^1] [^2] [^3]
title: 'A Buffer-aided Successive Opportunistic Relay Selection Scheme with Power Adaptation and Inter-Relay Interference Cancellation for Cooperative Diversity Systems'
---
Introduction
============
Cooperative relaying is an efficient technique to combat fading and path-loss effects in wireless systems. It enables multiple nodes to create virtual multiple-input multiple-output (MIMO) configurations in order to provide spatial transmit and/or receive diversity to single-antenna destinations [@LAN]. Traditional cooperative systems are characterized by the half-duplex constraint that relay nodes cannot receive and transmit data simultaneously, a fact that results to bandwidth loss. In order to overcome this bandwidth limitation, several techniques have been proposed in the literature [@DIN]. Among them, the successive relaying scheme in [@FAN] incorporates multiple relay nodes and proposes a transmission overlap (source-relay, relay-destination) in order to mimic an ideal full-duplex transmission. However, this scheme refers to scenarios with a long distance between the relays and thus inter-relay interference is not considered. An extension of this work is discussed in [@chao_ISIT], where the authors consider that IRI is strong (co-located or clustered relays) and can always be decoded at the affected nodes; this decoded IRI is exploited in a superposition coding scheme that significantly improves the diversity-multiplexing tradeoff (DMT) performance of the system. On the other hand, relay selection has been introduced as a promising practical solution that achieves full diversity by keeping the implementation complexity low. In earlier works in which relays were assumed to lack data buffers, relay selection was mainly based on the max-min criterion and its variations (see, for example, [@BLE; @KAR; @KRI_minmax; @BLE1] and references therein). As a result, based on either proactive or reactive criteria, the relay that received the source signal is the same with the one that is subsequently forwarding the signal towards the destination.
In the adoption of buffer-aided relays, this coupling is broken as different relays could be selected for transmission and reception. Buffering at the relay nodes is a promising solution for cooperative networks and motivates the investigation of new protocols and transmission schemes. Ikhlef *et al.* [@IKH2] propose a novel criterion based on the $\max-\max$ Relay Selection (MMRS), in which the relay with the best Source-Relay (SR) link is selected for reception and the relay with the best Relay-Destination (RD) link is selected for transmission. The MMRS implementation with buffers of finite size is subsequently also discussed. Another work that adopts MMRS is [@IKH3] which aims to recover the half-duplex loss by adopting successive transmissions. As the proposed topology aims to mimic full-duplex relaying, different relays are selected in the same time slot. However, relays are considered isolated and the effect of inter-relay interference is ignored, while relay buffers are never considered to be full or empty. Also, outage performance is examined for a fixed rate and the achieved capacity for adaptive rate. Krikidis *et al.* [@KRICHAR] proposed the $\maxlink$ protocol, which allows all the SR and RD links to enter the competition for the best link through which a signal will be transmitted, thus providing additional freedom in the scheduled transmissions at each time slot. Adaptive link selection, but for a single relay setup, is proposed by Zlatanov *et al.* in [@ZLAadapt]. The performance potentials of the buffer-aided relaying concept are further investigated in [@ZLA2], where the authors show that half-duplex relaying with storage capabilities outperforms ideal full-duplex relaying. Although most of the studies refer to one-way relaying, the work in [@LIU] studies the employment of buffers in a two-way relay network and investigates an opportunistic scheduling that maximizes the system sum-rate while ensuring buffer stability.
In the majority of the aforementioned works the main target is outage probability reduction or throughput improvement. Another characteristic that is of utmost importance in wireless cooperative networks is energy-efficiency. The proposed algorithm in [@NOMspringer] selects the end-to-end paths based on their achievable capacities and uses hierarchical and adaptive modulation. In this way the power of the source and the transmitting relay are adjusted according to the level required for successful reception. In various works (see, e.g., [@FENG; @ZHOU; @CHEN; @SHENG] and references therein) relays are assumed to be battery operated. Due to this practical assumption, the proposed algorithms select the best relays according to two factors. First, the selected relay will spend the minimum amount of energy that is required to surpass the outage threshold of the network. Second, by considering the remaining energy of each node, relays that were selected often will be protected in order to extend their operation and increase the network lifetime. A variation of this energy-efficiency metric is provided in [@AMIN], where circuit-energy consumption is considered and dynamic time allocation is implemented. Also in [@MOUS], by modeling the residual energy of each relay as an energy state, the transition probabilities among states are calculated and an expression of the average network lifetime is obtained in a variable gain amplify-and-forward relay network. Optimal energy allocation techniques are presented in [@IKKI], where the effects of co-channel interference at the relays and the destination are studied. Closed-form expression for the outage probability is derived and in addition, under different global and individual energy constraints, resource allocation algorithms are provided in a multi-relay network. Finally, in [@CHIEH], a multi-user network is investigated aiming to improve the energy efficiency per user, so that relays and subcarriers are selected and allocated accordingly for each user.
This work proposes a buffer-aided successive opportunistic relaying protocol with IRI cancellation. Through the successive nature of this protocol, we aim to recover the half-duplex constraint of cooperative relaying. In contrast to other works in the literature, we consider the arising IRI and with the aid of buffers we mitigate the detrimental effect of IRI through interference cancellation, if feasible. More specifically, we propose the $\minpow$ relay selection policy that acts in conjunction with interference cancellation and adjusts the power levels required accordingly to support the end-to-end communication. The $\minpow$ relay selection policy in terms of outage probability and diversity, is analyzed by modeling the evolution of the relay buffers as a MC. The construction of the state transition matrix and the related steady state of the MC are studied; then, the derivation of the outage probability is presented. The contribution of this work is twofold.
- Buffer-aided relays and interference cancellation are combined for the first time, thus decoupling the necessity of the receiving relay to transmit in the next time slot, even if the channel is in outage. Hence, an extra degree of freedom is obtained for choosing which relay to transmit, so that IRI cancellation is also achieved, if it is feasible. Also, the current channel state is the deciding factor in choosing a relay rather than the prediction of the next channel state.
- Power adjustment is also included in our scheme. In this way, the total energy expenditure in the network is minimized, as well as the inter-relay interference, thus reducing the outage probability of the network.
The $\minpow$ relay selection policy outperforms the current state-of-the-art schemes, with which it is compared. The outage, throughput and energy efficiency performance metrics considered are uniformly improved.
The structure of this paper is as follows. In Section \[sec:model\], we present the system model. Section \[policies\] reviews the most relevant relay selection schemes and Section \[sec:minpow\] introduces the $\minpow$ relay selection policy proposed herein. Then, a model of this communication scheme and an outage probability analysis is performed in Section \[sec:outage\], while illustrative examples and numerical results are provided in Sections \[sec:examples\] and \[sec:numerical\], respectively. Finally, conclusions are discussed in Section \[sec:conclusions\].
System model {#sec:model}
============
We assume a simple cooperative network consisting of one source $S$, one destination $D$ and a cluster $\mathcal{C}$ with $K$ Decode-and-Forward (DF) relays $R_k \in \mathcal{C}$ ($1 \leq k\leq K$). All nodes are characterized by the half-duplex constraint and therefore they cannot transmit and receive simultaneously. A direct link between the source and the destination does not exist and communication can be established only via relays [@BLE]. Each relay $R_k$ holds a buffer (data queue) $Q_k$ of capacity $L$ (number of data elements) where it can store source data that has been decoded at the relay and can be forwarded to the destination. The parameter $l_k \in \mathbb{Z}_{+}$, $l_k \in[0, L]$ denotes the number of data elements that are stored in buffer $Q_k$; at the beginning, each relay buffer is empty (i.e., $l_k =0$ for all $k$) We denote by $\mathcal{T}$ all the relays for which their buffer is not empty, i.e., $\mathcal{T} = \{R_k: l_k>0 \}$, $\mathcal{T} \subseteq \mathcal{C}$.
Time is considered to be slotted and at each time-slot the source $S$ and (possibly) one of the relays $R_k$ transmit with power $P_S$ and $P_{R_k}$, respectively. The source node is assumed to be saturated (it has always data to transmit) and the information rate is equal to $r_0$ bits per channel use (BPCU). The retransmission process is based on an Acknowledgment/Negative-Acknowledgment (ACK/NACK) mechanism, in which short-length error-free packets are broadcasted by the receivers (either a relay $R_k$ or the destination $D$) over a separate narrow-band channel in order to inform the network of that packet’s reception status.
All wireless links exhibit fading and Additive White Gaussian Noise (AWGN). The fading is assumed to be stationary, with frequency non-selective Rayleigh block fading. This means that the fading coefficients $h_{ij}$ (for the $i\rightarrow j$ link) remain constant during one slot, but change independently from one slot to another according to a circularly symmetric complex Gaussian distribution with zero mean and unit variance. The channel gains are $g_{ij} = |h_{ij}|^2$ and exponentially distributed, taking values in the range $(0,1]$. The power level chosen by the transmitter $i$ is denoted by $P_{i}$. $n_{i}$ denotes the variance of thermal noise at the receiver $i$, which is assumed to be AWGN.
Since we implement successive relaying, we (may) have concurrent transmissions by the source and one relay taking place at the same time slot. This results in IRI and the source has to consider the interference power that the candidate relay for reception will receive by the transmitting relay.
![The system model: Source $S$ communicates with Destination $D$ via a cluster of relays $R_k \in \mathcal{C}$, $k\in[1,K]$.[]{data-label="overflow"}](Figures/relay_net){width="0.45\columnwidth"}
It is worth noting that our focus is to investigate the performance of a new buffer-aided successive opportunistic relay selection scheme under a global Channel State Information (CSI) assumption and hence, the implementation issues are beyond the scope of this work. Note, however, that conventional centralized/distributed half-duplex relay selection approaches can be applied for the implementation of the proposed scheme (e.g., [@BLE1; @NOS; @SHAH]).
The interference power at the $k$-[th]{} receiver, $I_{k}$, in its general case, includes the interference from all the transmitters (belonging in a set $\mathcal{S}$) in the network apart from the intended transmitter $i$ and the thermal noise, and is given by $$\begin{aligned}
\label{interference}
I_{k}=\sum_{j\neq i, j \in \mathcal{S}}{g_{jk}P_{j} + n_{k}}.\end{aligned}$$ Therefore, the Signal-to-Interference-and-Noise Ratio (SINR) at the receiver $k$ is given by $$\begin{aligned}
\label{SINR1}
\Gamma_{k}=\frac{g_{ik}P_{i}}{\sum_{j\neq i, j \in \mathcal{S}}{g_{jk}P_{j} + n_{k}}}\ .\end{aligned}$$ The Quality of Service (QoS) is measured in terms of SINR. Hence, independently of nodal distribution and traffic pattern, a transmission from transmitter to its corresponding receiver is successful (error-free) if the SINR of the receiver is greater or equal to the *capture ratio* $\gamma_0$. Therefore, we require that $$\begin{aligned}
\label{SINR_geq}
\frac{g_{ik}P_{i}}{\sum_{j\neq i, j \in \mathcal{S}}{g_{jk}P_{j} + n_{k}}} \geq \gamma_0.\end{aligned}$$ The value of $\gamma_0$ depends on the modulation and coding characteristics of the radio. Due to battery limitations, we assume that each transmitting node $i$ (source and relays) has a maximum power $P_{i}^{\max}$.
Relay selection policies {#policies}
========================
In this Section, we briefly review the main buffer-based relay selection policies used in the literature.
$\max-\min$ relay selection
---------------------------
The $\max-\min$ relay selection policy [@BLE], [@KAR] is considered as the optimal selection scheme for conventional DF relay setups without buffers and is used as our baseline relay selection strategy. The selected relay is the one that provides the “best” (strongest) end-to-end path between the source and the destination. The $\max-\min$ selection policy ensures a full diversity (equal to the number of the relays (e.g. $K$)) and refers to a standard cooperative protocol where communication is performed in two time slots: a) in the first time slot, the source transmits b) in the second time slot, the selected relay forwards the data received by the source towards the destination. The $\max-\min$ relay selection policy can be written as $$\begin{aligned}
R^*=\arg\max_{R_k\in \mathcal{C}} \min \{ g_{S,R_k} , g_{R_k,D} \}.\end{aligned}$$ where $R^*$ denotes the selected relay. Although the $\max-\min$ selection scheme achieves a diversity equal to $K$, it does not take into account the fact that relays are equipped with buffers and their potential benefits. The selected relay forwards the received data (in case of a successful decoding) immediately in the next time slot to the destination and therefore the ability of the relays to store at least a limited number of data packets is not elaborated.
$\max-\min$ relay selection in successive relaying
--------------------------------------------------
The $\max-\min$ relay selection policy implemented in a successive relaying network takes a different form if the IRI can be canceled at the relays. In[@NOM], a reactive relay selection policy is proposed. More specifically, instead of considering only the [*SR*]{} and [*RD*]{} channel gains, the feasibility of IC is also examined. In this way, with a very simple IC condition the following two cases for relay selection are given.
- If candidate relay $R_k$ can perform IC then it may be selected to receive from the source based on the following value: $$\begin{aligned}
R^*=\arg\max_{R_k\in \mathcal{C}} \min \{ g_{S,R_k} , g_{R_k,D} \}.\end{aligned}$$
- On the other hand, if $R_k$ cannot perform IC then it can be selected after competing with the rest of the relays as below: $$\begin{aligned}
R^*=\arg\max_{R_k\in \mathcal{C}} \min \left\{ \frac{g_{S,R_k}} {g_{R_t,R_k}} , g_{R_k,D} \right\}.\end{aligned}$$
It is obvious that by having two simultaneous transmissions by the source and the transmitting relay reduces the diversity of the network as $R^*$ can not participate in the selection process due to the half-duplex constraint. On the other hand, if IRI is effectively mitigated, a full-duplex behavior can be achieved and the half-duplex loss is leveraged.
$\max-\max$ relay selection (bound)
-----------------------------------
The $\max-\max$ relay selection [@IKH2] is the first policy that exploits buffering capability at the relay nodes and is used as a reference selection scheme. Given that the relay nodes are equipped with buffers and thus can store the data received from the source, the $\max-\max$ policy splits the relay selection decision in two parts and selects the relay with the best source-relay link for reception and the relay with the best relay-destination link for transmission. The $\max-\max$ selection policy respects the conventional two-slot cooperative transmission where the first slot is dedicated for the source transmission and the second slot for the relaying transmission, but the relay node may not be the same for both phases of the protocol. The $\max-\max$ relay selection policy can be written as $$\begin{aligned}
&R_r^*=\arg\max_{R_k\in \mathcal{C}}\{ g_{S,R_k} \}, \\
&R_t^*=\arg\max_{R_k\in \mathcal{C}}\{ g_{R_k,D} \},\end{aligned}$$ where $R_r^*$ and $R_t^*$ denote the relay selected for the first phase and the second phase of the cooperative protocol, respectively. It has been proven that the $\max-\max$ relay selection policy also ensures full diversity equal to the number of the relays and it provides a significant coding gain in comparison to the conventional $\max-\min$ selection scheme. However, it is worth noting the the above selection strategy assumes that no relay’s buffer can be empty or full at any time and thus all relays have always the option of receiving or transmitting [@IKH2 Sec. III. C]. Therefore, the $\max-\max$ relay selection considered provides the optimal performance that can be achieved by such a scheme, thus yielding the lowest outage bound.
$\maxlink$ relay selection
--------------------------
The previous relay selection schemes are associated with a two-slot cooperative protocol where the schedule for the source and relay transmission is fixed a priori. Here, this limitation is relaxed and each slot is allowed to be allocated dynamically to the source or a relay transmission, according to the instantaneous quality of the links and the status of the relays’ buffers. More specifically, the proposed $\maxlink$ relay selection scheme exploits fully the flexibility offered by the buffers at the relay nodes and at each time selects the strongest link for transmission (source or relay transmission) among the available links. A source-relay link is considered to be available when the corresponding relay node is not full and therefore can receive data from the source, while a relay-destination link is considered to be available when the relay node is not empty and thus can transmit source’s data towards the destination. The proposed scheme compares the quality of the available links and adjust the relay selection decision and the time slot allocation to the strongest link. If a source-relay link is the strongest link, the source transmits and the corresponding relay is selected for reception; on the other hand, if a relay-destination link is the strongest link, the corresponding relay is selected for transmission. The $\maxlink$ relay selection policy can be analytically expressed as follows: $$\begin{aligned}
R^*=\arg\max_{R_k\in \mathcal{C}}\Bigg\{ \bigcup_{R_k\in \mathcal{C}: \Psi(Q_k)\neq L}\{g_{S,R_k} \} \bigcup_{R_k\in \mathcal{C}: \Psi(Q_k)\neq 0}\{g_{R_k,D} \} \Bigg\},\end{aligned}$$ where $R^*$ denotes the selected relay (either for transmission or reception) and the function $0\leq\Psi(Q_k)\leq L$, gives the number of data elements that are stored in buffer $Q_k$.
$\minpow$ relay selection policy {#sec:minpow}
================================
In this section, we present a novel relay selection policy called $\minpow$. This relay selection scheme is associated with a one-slot cooperative protocol (similar to the $\maxlink$ relay selection policy [@KRICHAR] and the one in [@IKH3] where however, IRI and power minimization are not considered), rather than two-slot protocols (as in [@IKH2]). At each time slot, the source $S$ transmits data to a selected relay (denoted by $R_r \in \mathcal{K}$) with a non-full buffer, and another relay with a non-empty buffer (denoted by $R_t\in\mathcal{T}$, $R_t \neq R_r$) transmits data to the destination $D$.
A packet is successfully transmitted from the transmitting relay $R_t$ to the destination $D$ if the Signal-to-Noise Ratio (SNR), denoted by $\text{SNR}_{R_t D}$, is greater than or equal to the capture ratio $\gamma_0$, i.e., $$\begin{aligned}
\label{eq:SNR_RD}
\frac{g_{R_t D}P_{R_t}}{n_D} \geq \gamma_0 \; , \quad R_t\in\mathcal{T}, R_t \neq R_r \; .\end{aligned}$$ A packet is successfully transmitted from source $S$ to the receiving relay $R_r$, if the Signal-to-Interference-and-Noise Ratio (SINR) at the receiving relay, denoted by $\text{SINR}_{S R_r}$ is greater than or equal to $\gamma_0$, i.e., $$\begin{aligned}
\label{eq:SNR_SR}
\frac{g_{S R_r}P_{S}}{g_{R_t R_r} P_{R_t} \mathbb{I}(R_t R_r) + n_{R_r}} \geq \gamma_0 \; , \quad R_r\in\mathcal{C}, R_r \neq R_t \; ,\end{aligned}$$ where $ \mathbb{I}(R_t R_r)$ is a factor indicating whether interference cancellation is satisfied and it is described by $$\begin{aligned}
\label{eq:index}
\mathbb{I}(R_t R_r) =
\begin{cases}
0 , & \text{if } \displaystyle \frac{g_{R_t R_r} P_{R_t}}{g_{S R_r}P_{S} + n_{R_r}} \geq \gamma_0 \; , \\
1 , &\text{otherwise.}
\end{cases}\end{aligned}$$
The following proposition states that if the maximum powers $P_{S}^{\max}$ and $P_{R_t}^{\max}$ are large enough (thus not imposing any limitations/constraints), for each pair of relays $R_r$ and $R_t$, then we can always find power levels such that interference cancellation conditions are satisfied.
\[prop:1\] Let $P_{S}^{\max}=\infty$ and $P_{R_t}^{\max}=\infty$. For each pair of relays $R_r$ and $R_t$, there exist $P_{S}$ and $P_{R_t}$ such that $\mathbb{I}(R_t R_r) =0$, $\text{SNR}_{R_t D} \geq \gamma_0$ and $\text{SINR}_{S R_r} \geq \gamma_0$. The minimum power levels $P_{S}^*$ and $P_{R_t}^*$ are achieved when $\text{SNR}_{R_t D} = \text{SINR}_{S R_r} = \gamma_0$, and are given by
\[eq:P\_optimal\] $$\begin{aligned}
P_{S}^* &= \frac{ \gamma_0 n_{R_r}}{g_{S R_r}} \; , \label{eq:P_S_optimal} \\
P_{R_t}^* &=\max \left\{ \frac{\gamma_0 n_D}{g_{R_t D}} , \frac{n_{R_r}\gamma_0 (\gamma_0 + 1)}{g_{R_t R_r}} \right\} \; \label{eq:P_Rt_optimal}.\end{aligned}$$
For IC to take place, according to , we have $$\begin{aligned}
\label{eq:index2}
{g_{R_t R_r} P_{R_t}} \geq \gamma_0({g_{S R_r}P_{S} + n_{R_r}})\end{aligned}$$ Given that $P_{R_t}$ is chosen such that is fulfilled, then becomes $$\frac{g_{S R_r}P_{S}}{n_{R_r}} \geq \gamma_0\; ,$$ and since $P_S$ decreases monotonically with $P_{R_t}$ (see equation for $\mathbb{I}(R_t R_r) =0$), while it requires a minimum value such that equation is fulfilled, the minimum power of $S$ is given with equality; i.e., $$\begin{aligned}
P_{S}^* = \frac{ \gamma_0 n_{R_r}}{g_{S R_r}} \; .\end{aligned}$$ Substituting into we have ${g_{R_t R_r} P_{R_t}} \geq n_{R_r}\gamma_0 (\gamma_0 + 1)$. Hence, the minimum $P_{R_t}$ is given by $$\begin{aligned}
P_{R_t}^* =\max \left\{ \frac{\gamma_0 n_D}{g_{R_t D}} , \frac{n_{R_r}\gamma_0 (\gamma_0 + 1)}{g_{R_t R_r}} \right\} \; .\end{aligned}$$
Proposition \[prop:1\] provides the minimum power levels of $S$ and $R_t$, provided that their maximum power levels do not impose any constrain and hence the IC conditions are satisfied. In the next proposition, we find the conditions under which IC cannot take place and we find the optimal power levels of $S$ and $R_t$.
\[prop:2\] For each pair of relays $R_r$ and $R_t$, interference cancellation is feasible if and only if $$\begin{aligned}
\label{eq:Pmax_condition}
P_{R_t}^{\max} \geq \gamma_0 \left[ \gamma_0 \left(\frac{\gamma_0 n_D}{g_{R_t D}} + \frac{n_{R_r}}{g_{R_t R_r}} \right) + \frac{n_{R_r}}{g_{R_t R_r}} \right] \; .\end{aligned}$$ When interference cancellation is infeasible, the signal from $R_t$ can be decoded successfully at $D$ if and only if $$\begin{aligned}
\frac{\gamma_0 n_D}{g_{R_t D}}&\leq P_{R_t}^{\max} \; .\label{subeq:ineq1} \end{aligned}$$ In addition, the signal from $S$ can be decoded successfully at $R_r$ if and only if $$\begin{aligned}
\gamma_0 \left(\frac{g_{R_t R_r}}{g_{S R_r}} P_{R_t} + \frac{n_{R_r}}{g_{S R_r}}\right) &\leq P_{S}^{\max} \; . \label{subeq:ineq2}\end{aligned}$$ When , and hold, the minimum power levels $P_{S}^\dagger $ and $P_{R_t}^\dagger $ are achieved when $\text{SNR}_{R_t D} = \text{SINR}_{S R_r} = \gamma_0$, and are given by
\[eq:P\_optimal2\] $$\begin{aligned}
P_{S}^\dagger &= \gamma_0 \left(\frac{g_{R_t R_r}}{g_{S R_r}} P_{R_t} + \frac{n_{R_r}}{g_{S R_r}}\right) \; , \label{eq:P_S_optimal2} \\
P_{R_t}^\dagger &=\frac{\gamma_0 n_D}{g_{R_t D}} \; \label{eq:P_Rt_optimal2}.\end{aligned}$$
Interference cancellation cannot take place when the maximum power of $R_t$ is not high enough, such that its signal can be decoded by $R_r$, i.e., $$\begin{aligned}
\label{eq:P_R_t_max}
\frac{g_{R_t R_r} P_{R_t}^{\max}}{g_{S R_r}P_{S} + n_{R_r}} < \gamma_0 \; .\end{aligned}$$ but it should definitely be high enough to be decoded at the destination, i.e., $$\begin{aligned}
\frac{\gamma_0 n_D}{g_{R_t D}}\leq P_{R_t} &\leq P_{R_t}^{\max} \label{subeq:ineq11}.\end{aligned}$$ Then, $P_S$ should be high enough, so that the transmitting signal can be successfully decoded at $R_r$. Hence, $$\begin{aligned}
\gamma_0 \left(\frac{g_{R_t R_r}}{g_{S R_r}} P_{R_t} + \frac{n_{R_r}}{g_{S R_r}}\right) \leq P_{S} &\leq P_{S}^{\max} \label{subeq:ineq21}\end{aligned}$$ Substituting into we have $$\begin{aligned}
\label{eq:condition}
\gamma_0 \left(\frac{g_{R_t R_r}}{g_{S R_r}} \frac{\gamma_0 n_D}{g_{R_t D}} + \frac{n_{R_r}}{g_{S R_r}}\right) \leq P_{S} &\leq P_{S}^{\max} \; .\end{aligned}$$ Substituting into we obtain that $$\begin{aligned}
P_{R_t}^{\max} < \gamma_0 \left[ \gamma_0 \left(\frac{\gamma_0 n_D}{g_{R_t D}} + \frac{n_{R_r}}{g_{R_t R_r}} \right) + \frac{n_{R_r}}{g_{R_t R_r}} \right] \; .\end{aligned}$$
Making use of Propositions \[prop:1\] and \[prop:2\], we will now describe the $\minpow$ relay selection algorithm. The algorithm is implemented in the following steps:\
1) First, for each possible pair of relays, we carry an *IC feasibility check*, i.e., we check through if interference cancellation is feasible.\
2) If IC is *feasible*, then
- assuming IC took place, $P_{R_t}^*$ and $P_{S}^*$ are as given in and , respectively. Note that since the feasibility criterion is satisfied, then $P_{R_t}^* \leq P_{R_t}^{\max}$. Nevertheless, we need to check separately whether $P_{S}^* \leq P_{S}^{\max}$.
- assuming IC did not take place, $P_{R_t}^\dagger $ and $P_{S}^\dagger$ are as given in and , respectively. $P_{R_t}^\dagger \leq P_{R_t}^{\max}$ is satisfied by the feasibility check, but we need again to check separately whether $P_{S}^\dagger \leq P_{S}^{\max}$.
- The minimum energy expenditure at a specific time slot for each pair is the minimum sum of the powers for the two cases, i.e., $\min \left\{ P_{S}^*+ P_{R_t}^*, P_{S}^\dagger + P_{R_t}^\dagger \right\}$. Note that if IC can take place, this might require $P_{R_t}^* > P_{R_t}^\dagger$, such that $P_{S}^*+ P_{R_t}^* > P_{S}^\dagger + P_{R_t}^\dagger$.
3\) If IC is *infeasible*, then we use case (ii) only from step 2).\
4) We compare the minimum energy expenditure for all possible relay pairs and we choose the minimum among them.
Note that in the worst case scenario (in which all the queues are neither empty nor full), there will be $K\times (K-1)$ combinations. Hence, the worst case complexity of the problem is $\mathcal{O}(K^2)$.
Model and outage probability analysis {#sec:outage}
=====================================
In this section, the outage probability behavior of the $\minpow$ relay selection scheme follows the theoretical framework of [@KRICHAR], which is also a relay network with finite buffers. The main differentiation compared to [@KRICHAR] is that we have additional ways of transmission through successive relaying; in other words, we have additional links from a certain buffer state to others. This will be more clearly reflected in the illustrative example.
Construction of the state transition matrix of the MC {#sec:outageA}
-----------------------------------------------------
We first formulate the state transition matrix of the Markov Chain (MC), denoted as $\mathbf{A}$, $\mathbf{A}\in \mathbb{R}^{{\left( {L + 1} \right)^K} \times {\left( {L + 1} \right)^K}}$. More specifically, ${{\bf{A}}_{i,j}} = {P}\left( {{s_i} \to {s_j}} \right) = {P}\left( {{X_{t + 1}} = {s_j}|{X_t} = {s_i}} \right)$ are the transition probabilities to move from a state $s_i$ to a state $s_j$. The transition probability depends on the number of relays that are available for cooperation.
As we consider finite buffers, relays that have full buffers cannot compete in the selection of the best relay that will receive the source’s signal. Also, relays with empty buffers are not able to transmit and as a result they are excluded from the best transmitting relay selection. Moreover, when there is no possibility of transmitting successively through two selected relays our system reduces to the $\maxlink$ relay selection scheme. The number of links that are available in this mode is reduced if the relays have full or empty buffers.
Let $N_i$ denote all the possible out-going links from state $s_i$. Among them, $n_i$ links are considered through non-successive (also referred to as *single-link*[^4]) transmissions, while $N_i-n_i$ through successive transmissions. These transmissions endure for one transmission phase. Since the source and one relay will transmit simultaneously, inter-relay interference will arise and the transmission power in each hop will be adjusted, according to Section \[sec:minpow\], depending on whether or not interference cancellation took place. In what follows, we denote with superscript $ns$ the non-successive (single-link) transmissions and with $s$ the successive ones. Since we have i.i.d fading channels, single link transitions can occur with equal probability. On the other hand, due to the inter-relay interference, the successive transitions will have equal probabilities in the high SNR regime as interference cancellation is always feasible there. Thus, the probability of a single link transition from state $s_i$ to a state $s_j$ is equal to: $$\label{eq:single_tr}
p^{\rm ns}_{ij} \buildrel \Delta \over = \frac{1}{n_i}\left[ {1 - {{\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{P}} \right)} \right)}^{{D_{{\rm ns}}}}}} \right] \;.$$ On the other hand, single link outage will occur when $$\label{eq:single_out}
{\bar p}^{\rm{ns}}_{ij} \buildrel \Delta \over = 1 - \sum\limits_{k = 1}^{{D_{\rm{ns}}}} {p^{\rm{ns}}_{ij} = \frac{n_i}{N_i}{{\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{P}} \right)} \right)}^{{D_{\rm{ns}}}}}} \;,$$ where $P$ is the transmission power, $r_0$ is the end-to-end spectral efficiency and ${D_{ns}}$ is the total number of links that the $K$ relays offer, i.e., ${D_{ns}} = \sum_{k = 1}^K {\Phi \left( {{Q_k}} \right)}$. More specifically, each relay will offer $$\label{eq:link_num}
\Phi \left( {{Q_k}} \right) = \left\{ \begin{array}{l l }
2 & {\rm{ \hspace{1 mm} if \hspace{1 mm} }} 0 < \Psi \left( {{Q_k}} \right) < L;\\
1 & {\rm{ \hspace{1 mm} elsewhere}}{\rm{.}}
\end{array} \right.$$
Furthermore, a successive transition, will be successful with a probability equal to $$\label{eq:succ_tr}
p^{\rm s} _{ij} \buildrel \Delta \over = \frac{1}{N_i - n_i}\left[ {1 - {{\left[ {\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{{{P_{s}^*}}}} \right)} \right)\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{P_{R_t}^*}} \right)} \right)} \right]}^{{D_{s}}}}} \right] \; ,$$ where $P_{s}^*$ is defined by if IC took place or if IC was not feasible. Similarly, $P_{R_t}^*$ is defined by or . By $D_{s}$ we denote the total number of possible end-to-end successive links as all the possible relay pairs are examined. $D_{s}$ is defined as $$\label{eq:succ_links}
{D_{s}} = \sum\limits_{k \ne l}^K {\Phi \left( {{Q_k}{Q_l}} \right)} \; ,$$ where $\Phi \left( {{Q_k}{Q_l}} \right)$ is equal to $$\begin{aligned}
\label{eq:succ_phi}
\Phi \left( {{Q_k}{Q_l}} \right) &= \prod\limits_{k = 1}^K {\Psi \left( {{Q_k}} \right)} \prod\limits_{l = 1}^K {\Psi \left( {{Q_l}} \right)}= \left\{ \begin{array}{l l}
1 &{\rm{ \hspace{1 mm} if \hspace{1 mm} }}0 < \Psi \left( {{Q_k}} \right) \le L \cap 0 \le \Psi \left( {{Q_l}} \right) < L \; ,\\
0 &{\rm{ \hspace{1 mm} elsewhere} \;.}
\end{array} \right.\end{aligned}$$ Moreover, an outage in the successive transitions will occur when $$\label{eq:succ_out}
{\bar p}^{\rm s}_{ij} \buildrel \Delta \over = 1 - \sum\limits_{k = 1}^{D_{\rm s}} p^{\rm s}_{ij} = \frac{{N_i - n_i}}{N_i}{{\left[ {\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{{{P_{S}^*}}}} \right)} \right)\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{P_{R_t}^*}} \right)} \right)} \right]}^{D_{\rm s}}}$$
Steady state distribution and outage probability {#steady_st}
------------------------------------------------
Since we have defined the entries of the transition matrix $\mathbf{A}$, the next step is to find the steady state distribution of the MC. In this way, the relationship among the different ways of leaving and reaching specific buffer states will be defined.
\[prop:3\] The state transition matrix $\mathbf{A}$ is Stochastic, Indecomposable[^5] and Aperiodic[^6] (SIA).
In order to prove that $\mathbf{A}$ is SIA, we have to show that it is (i) row stochastic, (ii) indecomposable, and (iii) aperiodic.\
(i) *Row Stochasticity*: For any MC the transition from state $s_i$ to a state $s_j$ for all possible states $s_i$ sums up to $1$, i.e., $$\label{eq:row_stoch}
\sum\limits_{i = 1}^{{{\left( {L + 1} \right)}^K}} {{{\bf{A}}_{i,j}}} = 1.$$ In the MC that models the buffer states of the relays, if there exists a transition from state $s_i$ to $s_j$ then there exists a transition from $s_j$ to $s_i$. This fact applies to our scheme where transitions due to successive relaying, offer additional links from one state to another and vice versa. However, since the states are not symmetric and the number of links to other states is not the same, the transition probabilities are not the same. Thus, the transition matrix $\mathbf{A}$ is not symmetric. As a result the $\mathbf{A}$ is row stochastic, but not necessarily doubly stochastic.\
(ii) *Indecomposability*: Due to the structure of the problem all the possible states of the considered MC can communicate and hence its state space is a single communicating class; in other words, it is possible to get to any state from any state. Hence, the MC is indecomposable.\
(iii) *Aperiodicity*: Aperiodicity of a MC is easily established due to the fact that the diagonal entries that correspond to original (non-virtual) nodes are nonzero. All links receive nonnegative weights and the diagonals (outage probabilities) are strictly positive. Also, the probability of being at any state after $M$ and $M +1$ transitions is greater than zero; hence, all states are aperiodic and therefore, the MC is aperiodic.
\[lemma:1\] The stationary distribution of the row stochastic matrix [**A**]{.nodecor} of the MC that models the buffer states is given by $\bm{\pi}=(\mathbf{A}-\mathbf{I}+\mathbf{B})^{-1}\bf{b}$, where $\bm{\pi}$ is the stationary distribution, $\bm{b}=(1~1~\ldots~1)^T$ and $\mathbf{B}_{i,j}=1, ~\forall i,j$.
The proof is similar to that of [@KRICHAR Lemma 1, Lemma 2]. It is included in Appendix \[sec:proofs\] for completeness.
Derivation of the outage probability
------------------------------------
Using the steady state of the MC and the fact that an outage event occurs when there is no change in the buffer status, first for successive relaying and then for single link transmissions, the outage probability of the system can be expressed as [@KRICHAR] $$\label{eq:p_out}
{P_{out}} = \sum\limits_{i = 1}^{{{\left( {L + 1} \right)}^k}} {{{\bm{\pi }}_i}{{\bar p}_{ij}} = {\rm{diag}}\left( {\bf{A}} \right){\bm{\pi }}}$$ By constructing the state transition matrix **A** that captures in its diagonal the probabilities where no change in buffer states happened, and the corresponding steady state probabilities, we can easily compute the outage probability of the system.
Illustrative Examples {#sec:examples}
=====================
In the previous section, we have described the theoretical framework for the computation of the outage probability. In the following, we will present two illustrative examples that showcase the behavior of our approach for different parameters. The first example consists of two relays $(K=2)$ with finite buffer size equal to two $(L=2)$, while the second one examines the case of infinite buffers $(L\rightarrow\infty)$ at the relays.
Illustrative example of K=2 relays and L=2 buffer size
------------------------------------------------------
In what follows, we present an illustrative example that showcases the behavior of our approach for different parameters. Since we have a scheme that employs successive transmissions the simplest case is when two relays are available. Assuming that each relay has a buffer size equal to two, we show its state transition diagram in Fig. \[overflow\], with the nine possible states for the buffers of the two relays.
![State diagram of the Markov chain representing the states of the buffers and the transitions between them for a case with $K=2
$ and $L=2$. Compared to the $\maxlink$ scheme in [@KRICHAR], the $\minpow$ model includes extra transition states due to the successive nature of the protocol.[]{data-label="overflow"}](Figures/2x2_v2){width="0.65\columnwidth"}
The steady state of the system for different values of SNR can be found by using the method described in \[steady\_st\] and the corresponding state transition matrix **A** is as follows. $$\begin{aligned}
{\small
\mathbf{A} = \left(\begin{array}{ccccccccc}
\bar{p}_{1} & p_{12} & p_{13} & {0} & {0} & {0} & {0} & {0} & {0} \\
p_{21} & \bar{p}_{2} & p_{23} & p_{24} & p_{25} & {0} & {0} & {0} & {0} \\
p_{31} & p_{32} & \bar{p}_{3} & {0} & p_{35} & p_{36} & {0} & {0} & {0} \\
0 & p_{42} & {0} & \bar{p}_{4} & p_{45} & {0} & p_{47} & {0} & {0} \\
0 & p_{52} & p_{53} & p_{54} & \bar{p}_{5} & p_{56} & p_{57} & p_{58} & {0} \\
0 & {0} & p_{63} & {0} & p_{65} & \bar{p}_{6} & {0} & p_{66} & {0} \\
0 & {0} & {0} & p_{74} & p_{75} & {0} & \bar{p}_{7} & p_{78} & p_{79} \\
0 & {0} & {0} & {0} & p_{85} & p_{86} & p_{87} & \bar{p}_{8} & p_{89} \\
0 & {0} & {0} & {0} & {0} & {0} & p_{97} & p_{98} & \bar{p}_{9}
\end{array} \right) \; .}\end{aligned}$$
Buffer State $\bm{\pi}$ ($P=0$dB) $\bm{\pi}$ ($P=10$dB) $\bm{\pi}$ ($P=20$dB) $\bm{\pi}$ ($P=30$dB)
-------------- ---------------------- ----------------------- ----------------------- -----------------------
$S_1$ 0.0413774 0.0005918 0.0000055 0
$S_2$ 0.1264042 0.1021872 0.1001714 0.05115971
$S_3$ 0.1263465 0.1021954 0.1001701 0.05115971
$S_4$ 0.1145455 0.1481283 0.1498316 0.19884035
$S_5$ 0.1826525 0.2937944 0.2996428 0.39768046
$S_6$ 0.1145455 0.1481283 0.1498316 0.19884035
$S_7$ 0.1263465 0.1021954 0.1001701 0.05115971
$S_8$ 0.1264042 0.1021872 0.1001714 0.05115971
$S_9$ 0.0413774 0.0005918 0.0000055 0
: Buffer states for $K=2$ relays and $L=2$ buffer size
\[K2L2\_pi\]
In Table \[K2L2\_pi\] we observe that when SNR increases, the steady state distribution decreases in buffer states $S_1$ and $S_2$, and eventually it becomes practically zero. That means that for high SNR, the probability of outage is zero. On the other hand, when SNR increases, the steady state distribution in state $S_5$ increases, which dominates the states for large SNRs.
In the numerical results (Section \[sec:numerical\]), we evaluate the theoretical framework for the case of $K=2$, $L=2$ in order to examine its behavior compared to obtained numerical results from simulations.
Example of $L\rightarrow\infty$ buffer size and $P\rightarrow\infty$ transmission power {#sec:examples2}
---------------------------------------------------------------------------------------
The asymptotic analysis yields interesting results due to the nature of the problem. In [@KRICHAR] it was shown that as the buffer size approaches infinity, the states where no full or empty relays exist are dominant[^7].
\[remark:1\] In the proposed scheme, we consider $K(K-1)$ possible relay pairs for transmission and reception in a successive way. When we are in a dominant state, each pair provides one possible transition to another state. Thus, the successive transmissions contribute $K(K-1)$ transitions. Moreover, we have the possibility of single link transmissions. A single link transmission can be either a transmission in a source-relay link or in relay-destination link. As a result, in a dominant state each relay offers 2 possible transitions and in total $K$ relays offer $2K$ transitions. In conclusion, in a dominant state the whole set of transitions is equal to $K(K-1)+2K=K(K+1)$.
From Remark \[remark:1\], the total number of transitions, denoted as $N_{i}$ in Section \[sec:outageA\], is equal to $K(K+1)$ when we are in a dominant state. Among them, single link transitions, denoted as $n_{i}$ in Section \[sec:outageA\], are equal to $2K$ so ${D_{{\rm{ns}}}} = 2K$. At the same time, successive transitions, denoted as $N_{i}-n_{i}$ in Section V.A, are equal to $K(K-1)$ possible ways of leaving the state; thus, ${D_{{\rm{s}}}} = K(K-1)$. Therefore, based on equation , the outage probability becomes $$\begin{aligned}
\label{eq:Pout}
{\bar p}^{\rm ns} = \frac{2K}{K\left( {K + 1} \right)}{\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{P}} \right)} \right)^{2K}} \; .\end{aligned}$$
\[lemma:3\] The diversity order[^8] of $\minpow$ is equal to $2K$.
The outage probability of $\minpow$ is bounded by the outage probability of single link transmission, i.e., by considering each possible outage case we have that the outage probability $\mathbb{P}_{\rm out}$, is given by $\mathbb{P}_{\rm out}=\mathbb{P}({\bar p}^{\rm ns}=1|{\bar p}^{\rm s}=1)+\mathbb{P}({\bar p}^{\rm s}=1|{\bar p}^{\rm ns}=1)=\mathbb{P}({\bar p}^{\rm ns}=1|{\bar p}^{\rm s}=1)$, since the probability of having an outage when searching for a successive transmission is larger due to the requirement of simultaneously not having two links in outage. On the contrary, in single link, the same links are examined and if one of them fulfills the threshold for error-free transmission then the outage event is avoided.
The diversity order is derived using its definition and equation , i.e., $$\begin{aligned}
d &= -\lim_{P \rightarrow \infty} \frac{\log \mathbb{P}_{\rm out}(P)}{\log P}
=-\lim_{P \rightarrow \infty} \frac{\log \frac{2K}{K\left( {K + 1} \right)}{\left( {1 - \exp \left( { - \frac{{{2^{2{r_0}}} - 1}}{P}} \right)} \right)^{2K}}}{\log P}\\
&\stackrel{(a)}{\approx}-\lim_{P \rightarrow \infty} \frac{\log \frac{2K}{K\left( {K + 1} \right)}{\left(\frac{2^{2{r_0}} - 1}{P} \right)^{2K}}}{\log P}=-\lim_{P \rightarrow \infty} \frac{\log \frac {2K}{K\left(K + 1\right)}+\log\left( \frac{2^{2 r_0} - 1}{P} \right)^{2K}}{\log P} \\
&=-\lim_{P \rightarrow \infty} \frac{2K\log\left( \frac{2^{2 r_0} - 1}{P} \right)}{\log P}
=-2K\lim_{P \rightarrow \infty} \frac{\log \left(2^{2{r_0}} - 1\right)}{\log P} + 2K\lim_{P \rightarrow \infty} \frac{\log P}{\log P}
=2K \; .\end{aligned}$$ Note that approximation $(a)$ emerges from the fact that for $x \rightarrow 0$, then $1+e^{-x} \approx x$.
For non-infinite buffer sizes possible gains in outage probability, compared to [@KRICHAR], derive from better interconnection between the buffer states. In this way, the buffer states where relays are full or empty are often avoided and the states where increased diversity is offered, are more usual.
Numerical Results {#sec:numerical}
=================
In line with the previous discussion, we have developed a simulation setup for the $\minpow$ scheme to evaluate its performance with a spectral efficient target $r_0=1$ bps/Hz, in terms of: $1)$outage probability, $2)$average throughput and $3)$power reduction. Min-power is compared to best-relay selection (BRS) [@BLE], successive opportunistic relaying (SOR) [@NOM], hybrid relay selection ($\max-\max$) [@IKH2] and $\maxlink$ selection [@KRICHAR]. In addition, we provide a selection bound corresponding to the case presented in [@IKH3] where inter-relay interference is ignored and additionally all links are always available for selection i.e. buffers are never full or empty. Also, the selection bound scheme is coupled with single link transmissions when successive transmissions fail, in order to provide a fair comparison with $\minpow$.
Outage Probability
------------------
Fig. \[pout\_K2L2\], illustrates the outage probability results. Each scheme employs $K=2$ relays with buffer size $L=2$. The selection policy that offers the worst performance is SOR. The lack of buffers prohibits the combination with a more robust scheme such as $\maxlink$ and IRI degrades the outage performance. Max-min shows better behavior as IRI is not present. Furthermore, $\max-\max$ offers about 1.5 dB improvement due to the use of buffers, over BRS. Even better results are achieved by $\maxlink$ as outage performance is improved by almost 4 dB due to the flexibility in the link selection. As $\maxlink$ is a part of $\minpow$ when successive transmissions are not possible, we observe similar results between these two schemes. For $K=2$ and $L=2$, $\minpow$ exhibits a 0.5 dB gain for high SNR. The increased interconnection between buffer states guarantees that states $S_{1}$ (00) and $S_{9}$ (22) offering the least diversity, are more often avoided compared to $\maxlink$. Also, the theoretical curve of the outage probability matches the simulation results validating the analysis in Section IV.
![Outage probability for increasing transmit SNR for $K=L=2$. As $\maxlink$ is a part of $\minpow$ when successive transmissions are not possible, they exhibit similar behavior. Still, $\minpow$ exhibits a 0.5 dB gain for high SNR due to increased interconnection between buffer states, thus avoiding $S_{1}$ (00) and $S_{9}$ (22) offering the least diversity. Also, the theoretical curve of the outage probability matches the simulation results.[]{data-label="pout_K2L2"}](Figures/Outage/OutageVSSchemes/OutageProbabilityK2L2){width="0.5\columnwidth"}
In Fig. \[pout\_K2Lv\], we depict outage probability performance for increasing transmit SNR and varying $L$. As $L$ increases, the curves become steeper thus indicating the increase in diversity as more links are available for selection. For $L=100$ and $L=\infty$ the outage curves almost match but still retain a gap from the selection bound case. This is logical as inter-relay interference degrades the performance of $\minpow$, while for the selection bound we assumed that the relays are isolated and inter-relay interference is negligible. Still, the $\minpow$ curve follows closely the selection bound especially for high SNR as very strong interference increases the probability of interference cancellation.
![Outage probability for increasing transmit SNR for $K=2$ and varying $L$. For increased $L$, the curves become steeper thus indicating increased diversity as more links are available for selection. For $L = 100$ and $L=\infty$ the outage curves almost match but still retain a gap from the selection bound, as IRI degrades the performance of $\minpow$.[]{data-label="pout_K2Lv"}](Figures/Outage/OutageVSL/OutageK2Lvarious){width="0.5\columnwidth"}
Next, Fig. \[pout\_KvL4\], shows the results for varying $K$ while $L=4$. As we saw in the asymptotic analysis example, diversity increases with a rate twice the number of $K$. Here we have a fixed finite buffer size but still each relay addition obviously improves the achieved diversity of the network. This derives from the fact that the possibility of inter-relay interference cancellation increases with an order equal to $K(K-1)$ as we have more relay pairs to select from. It is interesting to note, whenever $K$ increases to three from two the improvement is bigger than the gain introduced by adding one more relay when $K=3$ as $r_0$ is fulfilled even for low $K$ values.
![Outage probability for increasing transmit SNR for varying $K$ and $L=4$. Each relay addition improves the diversity of the network as we have more relay pairs to select from, thus increasing the chances for IRI cancellation. The improvement when $K$ increases to three from two is larger than the gain introduced by adding one more relay when $K=3$ as $r_0$ is fulfilled even for low $K$ values.[]{data-label="pout_KvL4"}](Figures/Outage/OutageVSK/OutageKvariousL4){width="0.5\columnwidth"}
Average Throughput
------------------
For the second set of comparisons among relay selection schemes, we present in Fig. \[thr\_K2L2\] the average throughput performance, measured in bps/Hz. First, we see that the compared selection policies are divided in two groups. The first one consists of the half-duplex schemes, namely BRS, $\max-\max$ and $\maxlink$. Due to the constant transmission rate, equal to 1 bps/Hz, these schemes can achieve a maximum average throughput of 0.5 bps/Hz. In line with the outage probability performance, $\maxlink$ outperforms BRS and $\max-\max$ and reaches the upper bound nearly 2.5 dB prior to the others. In the second group we have SOR and $\minpow$. These schemes aim to lift the half-duplex constraint and increase the average throughput through successive transmissions. It is observed that the schemes of the second group reach the upper-limit more slowly than those of the first group. This is explained by the fact that in the low SNR regime, single-link transmissions are often performed, thus reducing the throughput by one-half in these cases. Min-power achieves the best performance reaching 1 bps/Hz for high SNR. SOR however, does not reach the upper bound even for high SNR as IRI causes many outages.
![Average throughput for increasing transmit SNR for $K=L=2$. It is observed that $\minpow$ reaches its upper-limit more slowly than the half-duplex schemes. In the low SNR regime, single-link transmissions are often performed, thus reducing the throughput by one-half in these cases.[]{data-label="thr_K2L2"}](Figures/Throughput/ThroughputVSSchemes/Throughput_K2L2only){width="0.5\columnwidth"}
Fig. \[thr\_K2Lv\] illustrates average throughput for $\minpow$ as $L$ and transmission power increase. From the results we see that $\minpow$ for buffer-size above eight, follows the selection bound and their performance gap becomes negligible at about 8 dB. It is important to note that when the SNR is low, interference cancellation does not take place often and the proposed scheme chooses half-duplex transmissions instead of successive ones. This explains the gap between the selection bound and the cases of $L=100, \infty$ in the low SNR regime. Note that in this comparison we have $K=2$ relays and there is no flexibility in pair selection when a successive transmission is performed.
![Average throughput for increasing transmit SNR for $K=2$ and varying $L$. For buffer-size above eight, $\minpow$, follows the selection bound and their gap becomes negligible at about 8 dB. When the SNR is low, IC does not take place often and the proposed scheme chooses half-duplex transmissions instead of successive ones, thus explaining the gap between the selection bound and the cases of $L=100, \infty$[]{data-label="thr_K2Lv"}](Figures/Throughput/ThroughputVSL/ThroughputK2Lvarious){width="0.5\columnwidth"}
The third parameter that we examine is the number of relays in the cluster. Fig. \[thr\_KvL4\] shows the gain in average throughput as both $K$ and the transmission power increase. From the analysis of the $\minpow$ relay selection scheme, for each transmission we check $K(K-1)$ pairs to see whether or not interference cancellation can be performed. As we add more relays, the number of possible relay pairs increases from 2 in the case of $K=2$ to 6 in the case of $K=3$, while for $K=4$ we have 12 possible pairs. As a result, even for low SNR, successive transmissions are more possible for increasing $K$ and throughput tends to reach its maximum value more rapidly.
![Average throughput for increasing transmit SNR for varying $K$ and $L=4$. As more relays are added, the possible relay pairs increase from 2 in the case of $K=2$ to 6 in the case of $K=3$, while for $K=4$ we have 12 possible pairs. So, even for low SNR, successive transmissions are more possible and throughput reaches its maximum value more rapidly.[]{data-label="thr_KvL4"}](Figures/Throughput/ThroughputVSK/ThroughputKvariousL4){width="0.5\columnwidth"}
Power Reduction
---------------
The third metric that we study is the power reduction achieved by $\minpow$. The value on y-axis refers to differential power gain i.e. the power in dB that is saved when the proposed scheme is employed. The reference scheme that is used to calculate the power gain is a buffer-aided relay selection scheme that does not employ power minimization in the selection process and transmits each time with the maximum available power. For example, if the fixed power scheme uses 6 dB to perform a transmission and $\minpow$ reduces this requirement to 2 dB we keep the difference of 4 dB and we calculate the average after the end of all transmissions for this step of maximum power value. As stated in Section \[sec:minpow\], when $\minpow$ relay selection is used, a search for the pair of relays that require the minimum sum of transmission powers starts for a given SNR threshold. Similarly, for the case of single link transmission, the link requiring the minimum power to achieve a successful transmission is selected. We note that when a successive transmission is feasible both the source and a relay transmit, resulting in increased power reduction margin as the most appropriate relay pair is selected. On the contrary, when a single-link transmission occurs, we calculate power reduction by comparing the power used by the transmitting node in the fixed power scheme to the power used by the transmitting node in the selected single-link.
The first parameter that influences power reduction performance is $L$. Fig. \[pow\_K2Lv\], contains the curves for $K=2$ and various buffer sizes. Here the limiting factor is inter-relay interference. We see that for increased values of $L$ greater than 8, differences in power are minor but still, the selection bound is not met. This comes as a consequence of the difficulty to cancel IRI since only two relays are employed in the transmission.
![Power reduction for increasing transmit SNR for varying $L$ and $K=2$. For increased values of $L$ greater than 8, power differences are minor, however the selection bound is not met. As only two relays are employed in the transmission IRI cancellation is difficult.[]{data-label="pow_K2Lv"}](Figures/Power/PowerVSL){width="0.5\columnwidth"}
In order to examine the effect of additional relays, we present corresponding results in Fig. \[pow\_KvL8\]. We see that adding more relays to the cluster, the relay selection alternatives increase, thus leading to improved power minimization. As fixed transmission rate is adopted, a saturation is observed since differences in the required power for successful transmissions are very close after $K=3$ relays.
![Power reduction for increasing transmit SNR for varying $K$ and $L=8$. Since fixed transmission rate is adopted, a saturation is observed as differences in the required power for transmission are very close after $K=3$ relays.[]{data-label="pow_KvL8"}](Figures/Power/PowerVSK){width="0.5\columnwidth"}
Conclusions {#sec:conclusions}
===========
In this paper we proposed an opportunistic relaying protocol that minimizes the total energy expenditure per time slot under an IRI cancellation scheme. By making use of power adaptation and buffer-aided relays, the detrimental effect of IRI through interference cancellation is mitigated. It is the first time that interference cancellation is combined with buffer aided relays to mitigate IRI and on top this is combined with power adaptation for minimizing the energy expenditure; the examples demonstrate the improvements achieved. For the evaluation of the performance of the $\minpow$ relay selection policy, we performed comparisons with other schemes and the results showed that by combining successive transmission and a buffer-aided half-duplex protocol, gains were achieved in both outage and throughput performance. Moreover, we studied the effect of changing the buffer sizes and relay numbers on the outage, throughput and power reduction metrics. When the number of buffers and relays becomes large enough, the gain obtained by our proposed scheme vanishes. This observation suggests the need for additional approaches that will offer new characteristics to $\minpow$, thus improving its potential. A future direction includes the study of network coding schemes for further throughput improvement in this network setup and exploitation of the inter-relay interference by using superposition coding schemes.
Proofs {#sec:proofs}
======
Proof of Lemma \[lemma:1\] {#appendix1}
--------------------------
Since the MC considered is SIA, then there exists a unique solution $\bm{\pi}=(\bm{\pi}_1 ~ \bm{\pi}_2 ~\ldots ~ \bm{\pi}_{(L+1)^K} )^T$ such that $$\begin{aligned}
\mathbf{A}\bm{\pi}&=\bm{\pi}, \label{Api} \\
\sum_{i=1}^{(L+1)^K} \bm{\pi}_i &=1. \label{Bb}\end{aligned}$$ The solution can be easily found by solving the set of linear equations. In this case, is written as $$\begin{aligned}
\label{B1}
\mathbf{B}\bm{\pi}=\mathbf{b},\end{aligned}$$ where $\bm{b}=(1~1~\ldots~1)^T$ and $\mathbf{B}_{i,j}=1, ~\forall i,j$. Then, by adding and , we obtain that $$\begin{aligned}
\label{eq:p1}
\mathbf{A}\bm{\pi}-\bm{\pi}+\mathbf{B}\bm{\pi}=\mathbf{b} \Rightarrow (\mathbf{A}-\mathbf{I}+\mathbf{B})\bm{\pi}=\bf{b}\end{aligned}$$ Now, we have to show that the inverse of $(\mathbf{A}-\mathbf{I}+\mathbf{B})$ exists, and hence $\bm{\pi}=(\mathbf{A}-\mathbf{I}+\mathbf{B})^{-1}\bf{b}$. Let $\bf{y}$ be a row vector such that $$\begin{aligned}
\label{eq:p2}
\bf{y}(\mathbf{A}-\mathbf{I}+\mathbf{B})=0.\end{aligned}$$ It is sufficient to show that $\bf{y}=\bf{0}$. Multiplying by $\bm{\pi}$, we have $$\begin{aligned}
\bf{y}(\mathbf{A}-\mathbf{I}+\mathbf{B})\bm{\pi}=\bf{y}(\mathbf{A}\bm{\pi}-\bm{\pi}+\mathbf{B}\bm{\pi}) =\bf{y}\mathbf{b}=0.\end{aligned}$$ Since $\mathbf{b}$ has strictly positive entries and $\bf{y}\mathbf{b}=0$, we conclude that $\bf{y}=\bf{0}$. We have shown that there exists a unique solution and that it is given by $\bm{\pi}=(\mathbf{A}-\mathbf{I}+\mathbf{B})^{-1}\bf{b}$.
[1]{}
J. N. Laneman, D. N. C. Tse, and G. W. Wornell, “Cooperative diversity in wireless networks: Efficient protocols and outage behavior”, *IEEE Trans. Inform. Theory*, vol. 50, pp. 3062–3080, Dec. 2004.
Z. Ding, I. Krikidis, B. Rong, J. S. Thompson, C. Wang, and S. Yang, “On combating the half-duplex constraint in modern cooperative networks: protocols and techniques,” *IEEE Wireless Commun.*, vol. 19, pp. 20–27, Dec. 2012.
Y. Fan, C. Wang, J. S. Thompson, and H. V. Poor, “Recovering multiplexing loss through successive relaying using repetition coding,” *IEEE Trans. Wireless Commun.*, vol. 6, pp. 4484–4493, Dec. 2007.
C. Wang, Y. Fan, I. Krikidis, J. S. Thompson, and H. V. Poor, “Superposition-coded concurrent decode-and-forward relaying,” in *Proc. IEEE Int. Symp. Inf. Theory*, Toronto, Canada, July 2008, pp. 2390–2394.
A. Bletsas, A. Khisti, D. Reed, and A. Lippman, “A simple cooperative diversity method based on network path selection,” [*IEEE J. Select. Areas Commun.*]{}, vol. 24, pp. 659-672, March 2006.
D. S. Michalopoulos and G. K. Karagiannidis, “Performance analysis of single relay selection in Rayleigh fading,” [*IEEE Trans. Wireless Commun.*]{}, vol. 7, pp. 3718–3724, Oct. 2008.
I. Krikidis, J. S. Thompson, Steve McLaughlin and Norbert Goertz, “Max-min relay selection for legacy amplify-and-forward systems with intereference," [*IEEE Trans Wireless Commun.*]{}, vol. 8, pp. 3016-3027, June 2009.
A. Bletsas, H. Shin, and M. Z. Win, “Cooperative communications with outage-optimal opportunistic relaying,” [*IEEE Trans. Wireless Commun.*]{}, vol. 6, pp. 3450–3460, Sept. 2007.
A. Ikhlef, D. S. Michalopoulos, and R. Schober, “Max-max relay selection for relays with buffers," [*IEEE Trans. Wireless Commun.*]{}, vol. 11, pp. 1124–1135, March 2012.
A. Ikhlef, K. Junsu, and R. Schober, “Mimicking full-duplex relaying using half-duplex relays with buffers," [*IEEE Trans. Vehicular Tech.*]{}, vol. 61, pp. 3025–3037, Sept. 2012.
I. Krikidis, T. Charalambous, and J. S. Thompson, “Buffer-aided relay selection for cooperative diversity systems without delay constraints," [*IEEE Trans. Wireless Commun.*]{}, vol. 11, pp. 1957–1967, May 2012.
N. Zlatanov, R. Schober, and P. Popovski, “Buffer-aided relaying with adaptive link selection," [*IEEE J. Select. Areas in Commun.*]{}, vol. 31, pp. 1–13, Aug. 2012.
N. Zlatanov and R. Schober, “Buffer-aided half-duplex relaying can outperform ideal full-duplex relaying,” *IEEE Commun. Lett.*, accepted for publication, 2013.
H. Liu, P. Popovski, E. de Carvalho, and Y. Zhao, “Sum-rate optimization in a two-way relay network with buffering,” *IEEE Commun. Lett.*, vol. 17, pp. 95-98, Jan. 2013.
N. Nomikos, D. N. Skoutas, D. Vouyioukas, C. Verikoukis, and C. Skianis, “Capacity maximization through energy-aware multi-mode relaying", in [*Springer Wireless Pers. Commun.*]{} DOI 10.1007/s11277-012-0899-5, 2012.
Z. Zhou, S. Zhou, J.-H. Cui, S. Cui, “Energy-efficient cooperative communication based on power control and selective single-relay in wireless sensor networks," in [*IEEE Trans. Wireless Commun.*]{}, vol. 7, pp. 3066–3078, Aug 2008.
F. Ke, S. Feng, and H. Zhuang, “Relay selection and power allocation for cooperative network based on energy pricing," [*IEEE Commun. Letters*]{}, vol. 14, pp.396-398, May 2010.
W. J. Huang, Y. W. Peter Hong, and C. C. Jay Kuo, “Lifetime maximization for amplify-and-forward cooperative networks,” *IEEE Trans. Wireless Commun.*, vol. 7, pp. 1800–1805, May 2008.
Z. Sheng, B. J. Ko, and K. K. Leung, “Power efficient decode-and-forward cooperative relaying," [*IEEE Wireless Commun. Letters*]{}, vol. 1, pp. 444-447, Oct 2012.
O. Amin, and L. Lampe, “Opportunistic energy efficient cooperative communication," [*IEEE Wireless Commun. Letters*]{}, vol. 1, pp. 412–415, Oct 2012.
S. Mousavifar and C. Leung,“Lifetime analysis of a two-hop amplify-and-forward opportunistic wireless relay network," *IEEE Trans. Wireless Commun.*, accepted for publication, 2013.
S. Ikki and P. S. Aissa, “Regenerative cooperative diversity networks with co-channel interference: Performance analysis and optimal energy allocation," [*IEEE Trans. Vehicular Tech.*]{}, accepted for publication, 2012.
Y. H. Chieh, and H. Ching-Yao, “Energy efficient subcarrier-power allocation and relay selection scheme for OFDMA-based cooperative relay networks,” in [*Proc. IEEE Int. Conf. Commun.*]{}, Kyoto, Japan, June 2011, pp. 1–6.
N. Nomikos and D. Vouyioukas, “A successive opportunistic relaying protocol with inter-relay interference mitigation,” [*Proc. IEEE Wireless Commun. Mob. Comp. Conf*]{}, Limassol, Cyprus, Aug. 2012, pp. 1–5.
A. Nosratinia and T.E. Hunter, “Grouping and partner selection in cooperative wireless networks,” [*IEEE J. Select. Areas in Commun.*]{}, vol. 25, no. 2, pp. 369–378, Feb. 2007.
V. Shah, N. B. Mehta, and R. Yim, “Splitting algorithms for fast relay selection: Generalizations, analysis, and a unified view,” [*IEEE Trans. Wireless Commun.*]{}, vol. 9, pp. 1525–1535, Apr. 2010.
[^1]: N. Nomikos, D. Vouyioukas and D. Skoutas are with the Department of Information and Communications Systems Engineering, University of the Aegean,Karlovassi 83200, Samos, Greece (E-mail: [{nnomikos,dvouyiou,d.skoutas}@aegean.gr]{}).
[^2]: T. Charalambous and M. Johansson are with the Automatic Control Lab, Electrical Engineering Department and ACCESS Linnaeus Center, Royal Institute of Technology (KTH), Stockholm, Sweden (Emails: [{themisc,mikaelj}@kth.se]{}).
[^3]: I. Krikidis is with the ECE Department, University of Cyprus, Nicosia 1678 (E-mail: [[email protected]]{}).
[^4]: By single-link, we define the transmissions that take place either between the source and a receiving relay or those between a transmitting relay and the destination.
[^5]: A stochastic matrix $P\in \mathbb{R}^{m\times m}$ is said to be decomposable if there exists a nonempty proper subset $\mathcal{S} \subset \{v_1,v_2, \ldots, v_m \}$ such that $p_{ji}=p_{ij}=0$ whenever $v_i\in \mathcal{S}$ and $v_j \notin \mathcal{S}$. $P$ is indecomposable if it is not decomposable.
[^6]: In a finite state Markov Chain, a state $i$ is aperiodic if there exists $k$ such that for all $k' \geq k$, the probability of being at state $i$ after $k'$ steps is greater than zero; otherwise, the state is said to be periodic. A stochastic matrix is aperiodic if every state of the Markov chain it describes is aperiodic.
[^7]: By dominant we mean the states for which the probability of being at that state, when steady state is reached, is greater than zero.
[^8]: The diversity order (or diversity gain), denoted herein by $d$, is the gain in spatial diversity, used to improve the reliability of a link and it is defined as follows: $d = -\lim_{\textrm{SNR} \rightarrow \infty} \frac{\log \mathbb{P}_{\rm out}(\textrm{SNR})}{\log \textrm{SNR}}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
A central question in derandomization is whether randomized logspace ($\RL$) equals deterministic logspace (${\ComplexityFont{L}}$). To show that $\RL={\ComplexityFont{L}}$, it suffices to construct explicit pseudorandom generators (PRGs) that fool polynomial-size read-once (oblivious) branching programs (roBPs). Starting with the work of Nisan [@n92], pseudorandom generators with seed-length $O(\log^2 n)$ were constructed (see also [@inw94; @gr14]). Unfortunately, improving on this seed-length in general has proven challenging and seems to require new ideas.
A recent line of inquiry (e.g., [@bv10; @gmrtv12; @imz12; @rsv13; @svw14; @hlv17; @lv17; @chrt17]) has suggested focusing on a particular limitation of the existing PRGs ([@n92; @inw94; @gr14]), which is that they only fool roBPs when the variables are read in a particular *known* order, such as $x_1<\cdots<x_n$. In comparison, existentially one can obtain logarithmic seed-length for fooling the set of polynomial-size roBPs that read the variables under any fixed *unknown* permutation $x_{\pi(1)}<\cdots<x_{\pi(n)}$. While recent works have established novel PRGs in this setting for subclasses of roBPs, there were no known $n^{o(1)}$ seed-length explicit PRGs for general polynomial-size roBPs in this setting.
In this work, we follow the “bounded independence plus noise” paradigm of Haramaty, Lee and Viola [@hlv17; @lv17], and give an improved analysis in the general roBP unknown-order setting. With this analysis we obtain an explicit PRG with seed-length $O(\log^3 n)$ for polynomial-size roBPs reading their bits in an unknown order. Plugging in a recent Fourier tail bound of Chattopadhyay, Hatami, Reingold, and Tal [@chrt17], we can obtain a ${\widetilde{O}}(\log^2 n)$ seed-length when the roBP is of constant width.
author:
- 'Michael A. Forbes [^1]'
- 'Zander Kelley [^2]'
bibliography:
- 'prg-hlv.bib'
date: 'August 19, 2018'
title: 'Pseudorandom Generators for Read-Once Branching Programs, in any Order'
---
Introduction
============
A central goal in complexity theory is to understand the power of randomness in computation, in particular the $\mathsf{P}$ vs $\BPP$ problem. A particularly natural method of showing $\mathsf{P}=\BPP$ is to construct an explicit $\varepsilon$-error pseudorandom generator (PRG) with sufficiently small seed-length $\ell$, ideally logarithmic. That is, a function $G:{\{0,1\}}^\ell\to{\{0,1\}}^n$ such that for any sufficiently efficiently computable $f$, $$\left|\underset{y \in {\{0,1\}}^\ell}{\mathbb{E}}f(G(y)) - \underset{x \in {\{0,1\}}^n}{\mathbb{E}}f(x)\right|
\le\varepsilon
\;.$$ Given such a PRG, one can then replace the randomness of a $\BPP$ algorithm with the pseudorandom output and then enumerate over all such seeds to obtain a deterministic algorithm by majority vote (if $\varepsilon$ is a sufficiently small constant). After decades of work, the hardness-vs-randomness paradigm (see for example Vadhan [@v12]) shows that the construction of pseudorandom generators fooling general polynomial-size circuits is intimately tied to the quest for circuit lower bounds, which remain out of reach. As such, a long line of work has sought to derandomize subclasses of $\BPP$. A particularly fruitful model to study has been randomized logspace ($\RL$), as not only do PRGs for $\RL$ have natural applications, but they can also be unconditionally constructed, for example as done in the seminal work of Nisan [@n92].
In particular, Nisan [@n92] constructed a PRG fooling the non-uniform version of $\RL$, that is, the class of polynomial-size read-once (oblivious) branching programs (roBPs). A read-once branching program can be thought of as a finite automaton that takes in binary input strings $x$ of some fixed size $n$. Additionally, the transition function of the automaton is allowed to depend on the position $i$ of each bit $x_i$. We say that the branching program has width $w$ if the each layer of time the finite automaton has $w$ states. Visually, branching programs can be represented as a layered acyclic digraph with $n+1$ layers, each containing $w$ nodes; the transition function is then represented by assigning two outgoing edges at each interior node into the next layer. The existence of a logspace-computable PRG $G : {\{0,1\}}^{\ell(n)} \rightarrow {\{0,1\}}^n$ for branching programs of width $w = n^{O(1)}$ is sufficient to show that $\BPL \subseteq \DSPACE(\ell(n))$.
Nisan [@n92] gave a construction of a PRG with seed-length $\ell = O(\log^2n)$ for polynomial-width roBPs. Since then, there have been various constructions ([@inw94; @gr14]) recovering the same seed-length using different techniques, but there has been little quantitative progress towards the desired seed-length $\ell = O(\log n)$.[^3] In fact, it remains open even to achieve a seed-length of $\ell = O(\log^2 n / \log \log n)$, even for constant-width branching programs.
The constructions of Nisan and ([@inw94; @gr14]) all employ a common high-level approach which can be summarized by the following “communication” argument. The first half of a branching program can communicate with the second half only via the state reached in the middle layer. Since there are only $w$ states in this layer, the second half of the program should “learn” roughly only $\log w$ bits of information about the input bits fed to the first half. Because of this, it is safe to reuse all but roughly $\log w$ of the bits of entropy invested to generate the first half of the input string to generate the second half. This argument is then applied recursively to the left and right subprograms.
There is some feeling that this particular recursive paradigm will not yield generators with seed-lengths better than $O(\log^2 n)$ ([@bv10; @rsv13; @svw14]), and that new, more flexible techniques are required to make progress. A crucial feature of this paradigm is that the PRG knows the order in which its pseudorandom output will be read. In fact, it is known that Nisan’s generator fails to generate pseudorandom strings that fool branching programs if they read the bits of the string in a different order than anticipated ([@t09]). The search for a different paradigm motivates the following challenge: construct a PRG that fools branching programs which may read their input in any order. To formalize this, we define the notion of an unknown-order roBP: a function $g : {\{0,1\}}^n \rightarrow {\{0,1\}}$ of the form $g(x) = f( x_{\pi(1)}, x_{\pi(2)}, \ldots, x_{\pi(n)} )$, where $\pi$ is a permutation (independent of $x$) and $f$ is a roBP.
Bogdanov, Papakonstantinou, and Wan [@bpw11] constructed a PRG with seed-length $(1 - \Omega(1)) \cdot n$ for unknown-order roBP of width $w = n^{O(1)}$. Their primary motivation for doing so was to derive the first generator with nontrivial seed-length that fools read-once formulas. Read-once formulas can be simulated by small-width read-once branching programs for *some* order $\pi$, and hence existing generators for *known*-order roBPs ([@n92; @inw94; @gr14]) would not suffice. Impagliazzo, Meka, and Zuckerman [@imz12] achieved a generator with seed length $\ell = (nw)^{1/2 + o(1)}$ for unknown-order roBP of width $w$.[^4]
Our Work
========
Here, we give the first PRG with poly-logarithmic seed-length for $\poly(n)$-width unknown-order roBPs.
\[main1\] There exists an explicit $\nicefrac{1}{\poly(n)}$-error pseudorandom generator $G : {\{0,1\}}^{O( \log^3 n)} \rightarrow {\{0,1\}}^n$ for the class of functions computable by a $\poly(n)$-width read-once (oblivious) branching program in some variable order.
As a corollary, we also derive the first PRG with poly-logarithmic seed length for read-once formulas (see [@bpw11] for the reduction).
There exists an explicit pseudorandom generator $G : {\{0,1\}}^{O( \log^3 n)} \rightarrow {\{0,1\}}^n$ for read-once formulas with constant fan-in.
Our Techniques
--------------
We now briefly describe our proof technique at a high-level, with a more technical discussion given in . The main motivation comes from the “bounded independence plus noise” paradigm introduced by Haramaty, Lee, and Viola ([@hlv17; @lv17]). There, they study the addition (modulo 2) of a low-wise independent distribution with a pseudorandom noise distribution. The intuition is that to fool a function $f$, it suffices to create a distribution to dampen all non-constant Fourier coefficients. For low-degree Fourier coefficients, this can be achieved by a low-wise independent distribution. In the other extreme, high-degree Fourier coefficients are dampened by coordinate-wise independent noise. The addition of these two distributions can then inherit the best of both distributions and fool the desired function $f$.
However, the above outline has two challenges. First, the noise distribution (picking each coordinate independently amongst $\{0,1\}$) requires too large a seed-length. To address this, the work of Haramaty, Lee, and Viola ([@hlv17; @lv17]) proposed to use a pseudorandom noise distribution where a pseudorandom set of coordinates are first chosen, and then the elements within those coordinates are then substituted with *truly* random values. While this proposal as stated still requires a large seed-length, the key observation is that the number of truly random bits has shrunk from $n$ originally to $\approx \nicefrac{n}{2}$ (for if you choose a (pseudorandom) subset of $\{1,\ldots,n\}$ it has size $\approx \nicefrac{n}{2}$). Thus, one can hope to then recursively apply the construction in $\approx \log n$ rounds until no random bits are further required.
The second, more serious, challenge is to show that a single step of “bounded independence plus (pseudorandom) noise” actually fools the target function $f$, and doing so is the main contribution of this work (). The difficulty in addressing this is that there are too many high-degree Fourier coefficients, so that while each can be individually fooled by the construction, we cannot apply a union bound while maintaining a small seed-length. Indeed, nothing so far in this discussion has used anything about the structure of the function $f$, which clearly must be used to obtain a small seed-length.
To meet this challenge, we avoid a naive union bound by instead grouping high-degree Fourier coefficients into a small number of groups, each of which can be dampened at once. Specifically, we group high-degree Fourier coefficients into $n$ sets, where the $i$-th set contains those coefficients that “become” high-degree upon reading the $i$-th variable ().[^5] On an intuitive level, one can then appeal to the “bounded communication” aspect of a roBP to argue that in the $i$-th grouping those variables read after the $i$-th variable can be essentially ignored. We are then left with Fourier coefficients that are of *medium* degree (for if they were very-high degree they would have been put in the $j$-th group for some $j<i$). The number of such medium-degree Fourier coefficients is not too large (because the degree is not too large), and yet each such coefficient is dampened by the noise distribution (because the degree is not too small). This then allows us to apply a union bound to obtain that we have dampened all the Fourier coefficients in the $i$-th grouping, and by applying this for all $i$ we obtain the result.
The Constant-Width Case
-----------------------
Although we give the first PRG with poly-logarithmic seed-length for the general case of poly-width unknown-order roBPs, such a seed-length has been qualitatively achieved in the constant-width case as a result of a recent line of work. Reingold, Steinke, and Vadhan [@rsv13] gave a PRG with seed-length $O(\log^2 n)$ for unknown-order *permutation* branching programs of constant width. Steinke, Vadhan, and Wan [@svw14] gave a PRG with seed-length $\widetilde{O}(\log^3 n)$ for unknown-order width-3 branching programs. Chattopadhyay, Hatami, Reingold, and Tal [@chrt17] gave a PRG with seed-length $\widetilde{O}(\log^{w+1} n)$ for unknown-order branching programs of constant-width $w$. Central to each of these results is a bound on a certain key quantity: the level-$k$ Fourier mass of a branching program (see ). In each work, a bound on this quantity is established for the class of branching programs under consideration, and then this bound is used to deduce the result.
Although we also employ a Fourier analytic approach, a major contrast between our techniques and this line of work is that in general we have no need of any nontrivial bound on the Fourier mass of branching programs. However, we can still make use of one to replace otherwise naive bounds on Fourier mass in our argument. By incorporating the level-$k$ Fourier mass bound for constant-width branching programs derived in [@chrt17] into our approach, we get the following improvement on in the constant-width case.
\[main2\] There exists an explicit $\nicefrac{1}{\poly(n)}$-error pseudorandom generator $G : {\{0,1\}}^{{\widetilde{O}}( \log^2 n)} \rightarrow {\{0,1\}}^n$ for the class of functions computable by a $O(1)$-width read-once (oblivious) branching program in some variable order.
Thus in the constant-width case, we nearly recover the $O( \log^2 n)$ seed-length of Nisan’s generator for the more challenging model of unknown-order branching programs.
Preliminaries {#prelim}
=============
Here we describe a convenient algebraic encoding of a branching program as a product of one-bit matrix-valued functions. Recall that a branching program of width $w$ is a $w$-state finite automaton where the transition map is allowed to depend on the number of bits read so far. Let us encode the $w$ states as the set of standard basis vectors in ${\mathbb{R}}^{w}$. Then, the transition map corresponding to the $i$-th input bit $x_i$ can be encoded by a pair of transition matrices $A_{i,0}, A_{i,1} \in {\mathbb{R}}^{w \times w}$, defined so that $A_{i,x_i}$ applied to the current state produces the appropriate successor state. Define the one-bit matrix-valued functions $F_i(x_i) = A_{i,x_i}$. With this notation in place, the value of a branching program $f$ on an input $x \in {\{0,1\}}^n$ is given by the $(1,1)$ entry of the product $$F(x) {:=}F_1(x_1) F_2(x_2) \cdots F_n(x_n)
\;.$$ This entry indicates whether the string $x$ defines a path through the program that takes the start state to the accepting state.
Let $U$ denote the uniform distribution over ${\{0,1\}}^n$ and let $X$ be an arbitrary distribution over ${\{0,1\}}^n$. To show that a branching program $f$ is $\varepsilon$-fooled by $X$, it suffices to bound the error ${\left| {\mathbb{E}}_X f(X) - {\mathbb{E}}_U f(U) \right|} = {\left| {\mathbb{E}}_X F(X)_{1,1} - {\mathbb{E}}_U F(U)_{1,1} \right|}$ by $\varepsilon$. However, it will be more convenient to simply bound the Frobenius norm of the entire error matrix $${\mathbb{E}}_X F(X) - {\mathbb{E}}_U F(U)
\;,$$ by $\varepsilon$. Recall that the Frobenius norm of a matrix is defined by $${\|M \|} {:=}\sqrt{\operatorname{tr}{\left( M^\top M \right)}} = \sqrt{ \sum_{i,j} M_{i,j}^2 }
\;,$$ so clearly such a bound is also sufficient. In this paper, ${\|\cdot \|}$ will always denote the Frobenius norm.
Fourier Analysis
----------------
For every vector $\alpha \in {\mathbb{F}}_2^n$, define the associated Fourier character ${\chi_{\alpha}}: {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}$ via $${\chi_{\alpha}}(x) = (-1)^{{\left\langle\alpha ,x \right\rangle}} .$$ We say that ${\chi_{\alpha}}$ is a degree-$k$ Fourier coefficient if ${\left|\alpha \right|} = k$, where ${\left|\alpha \right|}$ denotes the hamming weight of $\alpha$. Any function $F : {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}^{w \times w}$ can be expanded in the basis of Fourier characters, with coefficients from the ring ${\mathbb{R}}^{w \times w}$. The Fourier expansion of $F$ is $$F(x) = \sum_{\alpha \in {\mathbb{F}}_2^n} {\widehat}{F}_{\alpha} {\chi_{\alpha}}(x),$$ where the Fourier coefficients ${\widehat}{F}_{\alpha}$ are given by $${\widehat}{F}_\alpha {:=}\underset{x \in {\mathbb{F}}_2^n}{\mathbb{E}}F(x) {\chi_{\alpha}}(x).$$ This identity can easily be checked with the aid of a few useful properties of Fourier characters; namely $${\chi_{\alpha}}(x) {\chi_{\beta}}(x) = \chi_{\alpha + \beta}(x)$$ and $$\underset{x \in {\mathbb{F}}_2^n}{\mathbb{E}}{\chi_{\alpha}}(x) =
\begin{cases}
1 &\text{ if } \alpha = 0 \\
0 &\text{ otherwise}.
\end{cases}$$
Aside from this, we will only need a few simple facts about Fourier analysis. Firstly, we note that the expectation of a function $F$ under the uniform distribution is conveniently encoded by its $0$-th degree Fourier coefficient: $${\mathbb{E}}_U F(U) = \sum_{\alpha \in {\mathbb{F}}_2^n} {\widehat}{F}_{\alpha} {\mathbb{E}}_U {\chi_{\alpha}}(U) = {\widehat}{F}_0 .$$ Next, we will need Parseval’s identity to get a bound on the sum of squares of Fourier coefficients.
$$\sum_{\alpha \in {\mathbb{F}}_2^n} {\|{\widehat}{F}_\alpha \|}^2 = \underset{x \in {\mathbb{F}}_2^n}{\mathbb{E}}{\| F(x) \|}^2
\;.$$
Let ${\left\langle\cdot ,\cdot \right\rangle}$ denote the Frobenius matrix inner product, which is defined by $${\left\langleM ,N \right\rangle} = \operatorname{tr}\! {\left(M^\top N \right)} = \sum_{i,j} M_{i,j} N_{i,j} .$$ First note that $$\begin{aligned}
{\|F(x) \|}^2
&= {\left\langleF(x) , F(x) \right\rangle} \\
&= {\left\langle \sum_{\alpha \in {\mathbb{F}}_2^n} {\widehat}{F}_{\alpha} {\chi_{\alpha}}(x) , \sum_{\beta \in {\mathbb{F}}_2^n} {\widehat}{F}_{\beta} {\chi_{\beta}}(x) \right\rangle} \\
&= \sum_{\alpha \in {\mathbb{F}}_2^n} \sum_{\beta \in {\mathbb{F}}_2^n} {\left\langle {\widehat}{F}_{\alpha} , {\widehat}{F}_{\beta} \right\rangle} \chi_{\alpha + \beta}(x)
\;.
\end{aligned}$$ Thus $$\begin{aligned}
{\mathbb{E}}_{x \in {\mathbb{F}}_2^n} {\| F(x) \|}^2
&=\sum_{\alpha \in {\mathbb{F}}_2^n} \sum_{\beta \in {\mathbb{F}}_2^n} {\left\langle {\widehat}{F}_{\alpha} , {\widehat}{F}_{\beta} \right\rangle} \underset{x \in {\mathbb{F}}_2^n}{\mathbb{E}}\chi_{\alpha + \beta}(x) \\
&=\sum_{\alpha \in {\mathbb{F}}_2^n} {\left\langle {\widehat}{F}_{\alpha} , {\widehat}{F}_{\alpha} \right\rangle} \\
&=\sum_{\alpha \in {\mathbb{F}}_2^n} {\| {\widehat}{F}_{\alpha} \|}^2
\;.
\qedhere
\end{aligned}$$
We remark that if $F$ is a branching program, then upon any input $x$, $F(x)$ is equal to some transition matrix that has exactly $w$ entries of value $1$ and its remaining entries are all zeros. Thus for branching programs we have ${\|F(x) \|}^2 = w$ for any $x$, and the above identity gives $\sum_{\alpha \in {\mathbb{F}}_2^n} {\|{\widehat}{F}_{\alpha} \|}^2 = w$.
Finally, we define $\mathcal{L}_k(F)$, the level-$k$ Fourier mass of a function, as in [@rsv13]: $$\mathcal{L}_k(F) {:=}\sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ |\alpha| = k}} {\| {\widehat}{F}_{\alpha} \|}
\;.$$ Recalling that ${\| {\widehat}{F}_{\alpha} \|} = {\| {\mathbb{E}}_x F(x) {\chi_{\alpha}}(x) \|} \leq {\mathbb{E}}_x {\| F(x) \|} = w^{1/2}$, note that we have the trivial bound $$\mathcal{L}_k(F) \leq \binom{n}{k} w^{1/2}
\;,$$ for branching programs.[^6]
Chattopadhyay, Hatami, Reingold, and Tal [@chrt17] derive the following bound on the level-$k$ Fourier mass of a branching program which significantly improves upon the trivial bound in the case of small $w$.
Suppose $F : {\mathbb{F}}^2_n \rightarrow {\mathbb{R}}^{w \times w}$ encodes a branching program of width $w$. Then $$\mathcal{L}_k(F) \leq O( \log n)^{wk}
\;.
\qedhere$$
We also define the level-$k$ Fourier complexity of width-$w$ branching programs, $$\mathcal{L}(n,w;k) {:=}\max_{F} \sum_{i=1}^k \mathcal{L}_i(F)
\;,$$ where the maximum is taken over all functions $F : {\mathbb{F}}_2^n \to {\mathbb{R}}^{w \times w}$ that encode a branching program.
Pseudorandom Primitives
-----------------------
In this section, we collect some basic pseudorandom distributions over ${\mathbb{F}}_2^n$ that will serve as building blocks for our generator. It is important that the defining property of each of these distributions remains unaffected by a re-ordering of bit positions. The fact that we build our generator out of permutation-invariant components is what allows it to fool branching programs that read their input in any order.
First we introduce $\delta$-biased distributions, which fool linear functions over ${\mathbb{F}}_2^n$, i.e. Fourier characters.
Let $D$ be a distribution over ${\mathbb{F}}_2^n$. We say $D$ is $\delta$-biased if, for every nonzero $\alpha \in {\mathbb{N}}_2^n$, we have $${\left| {\mathbb{E}}_D {\chi_{\alpha}}(D) \right|} \leq \delta
\;.
\qedhere$$
It is possible to sample from a $\delta$-biased distribution using $O(\log n + \log\nicefrac{1}{\delta})$ random bits ([@nn93; @aghp92]).
Next we have $k$-wise independent and $\gamma$-almost $k$-wise independent distributions, which look locally uniform and thus fool functions that only depend on a few bits.
Let $D$ be a distribution over ${\mathbb{F}}_2^n$. We say $D$ is $k$-wise independent if, for every $f : {\mathbb{F}}_2 \rightarrow [-1, 1]$ that depends on at most $k$ bits, we have $${\mathbb{E}}_D f(D) = {\mathbb{E}}_U f(U).$$ If $D$ merely satisfies $${\left| {\mathbb{E}}_D f(D) - {\mathbb{E}}_U f(U) \right|} \leq \gamma$$ for every such $f$, we say that $D$ is $\gamma$-almost $k$-wise independent.
It is possible to sample from a $k$-wise independent distribution using $O(k \cdot \log n)$ random bits ([@v12]) and from a $\gamma$-almost $k$-wise independent distribution using $O(k + \log \log n + \log 1/\gamma)$ random bits ([@nn93; @aghp92]). We remark that Fourier characters ${\chi_{\alpha}}(x)$ only depend on $|\alpha|$ bits of $x$, so these distributions also fool low-degree Fourier characters.
The Generator
=============
We adopt our construction from the “bounded independence plus noise” framework developed by Haramaty, Lee, and Viola in [@hlv17; @lv17]. In fact this framework is essentially equivalent to the “mild pseudorandom restriction” framework developed by Gopalan, Meka, Reingold, Trevisan and Vadhan [@gmrtv12] and subsequently employed by various authors (e.g., [@rsv13; @svw14; @chrt17]), but we find the bounded independence plus noise perspective more convenient to work with.
We actually give two slightly different constructions. The first construction only uses $k$-wise independence as a core pseudorandom primitive (along with appropriate recursion), and suffices for proving . The second construction replaces the use of *exact* $k$-wise independence with the use small-bias spaces and *almost* $k$-wise independence. This allows the error analysis to be more general (but also slightly more involved due to additional parameters). In particular, this second construction also suffices for handing general roBPs (), but now the added generality can be tuned to achieve a better seed-length for constant-width roBPS, as required to prove .
We proceed to give the first construction.
Let $D_1, D_2, \ldots, D_r$ denote $r$ independent copies of a $2k$-wise independent distribution and let $T_1, T_2, \ldots, T_r$ denote $r$ independent copies of a $k$-wise independent distribution over ${\mathbb{F}}_2^n$. We then define the distribution $G_r$ recursively as follows. Let $G_0$ be equal to the all-ones string in ${\mathbb{F}}_2^n$, and set $$G_{i+1} {:=}D_i + T_i \land G_i,$$ where $\land$ denotes bitwise AND and $+$ denotes addition over ${\mathbb{F}}_2^n$ (i.e. bitwise XOR).
\[mainlem1\] Suppose $F : {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}^{w \times w}$ encodes a branching program. Then $G_r$, with parameters $k = \lceil 5 \lg n + 2 \lg w \rceil$ and $r = \lceil 2 \lg n + \frac{1}{2} \lg w \rceil$, fools $F$ with error $$\varepsilon = {\| {\mathbb{E}}_{G_r} F(G_r) - {\mathbb{E}}_U F(U) \|} \leq O{\left(\tfrac{1}{n} \right)} .$$
This proves , since in the case of width $w = n^{O(1)}$ the price of sampling such a distribution is $O(r \cdot k \cdot \log n) = O(\log^3 n)$ random bits.
We now define the second construction, $G_r^*$. This time, let $D_1, D_2, \ldots, D_r$ denote $r$ independent $\delta$-biased distributions, and let $T_1, T_2, \ldots, T_r$ denote $r$ independent $\gamma$-almost $k$-wise independent distributions. We set $G_0^*$ equal to a $320k$-wise independent distribution, and again we define $$G_{i+1}^* {:=}D_i + T_i \land G_i^*.$$
\[mainlem2\] Suppose $F : {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}^{w \times w}$ encodes a branching program. Then $G_{r}^*$, with depth $r = \lceil \lg n \rceil$, fools $F$ with error $$\varepsilon = {\| {\mathbb{E}}_{G_r^*} F(G_r^*) - {\mathbb{E}}_U F(U) \|} \leq O{\left( \left( \sqrt{\delta} \mathcal{L}(n,w;k) + \left(\tfrac{1}{2}\right)^{k/2} + \sqrt{\gamma} + \gamma 4^k \right) \cdot nwr \right)} .$$
Since $G_r^*$ can be sampled at the cost of $$O \big( \left(\log n + \log 1/\delta + k + \log1/\gamma \right) \cdot r + k\cdot \log n \big)$$ random bits, it suffices to set
- $r = \lceil \lg n \rceil$
- $k = \lceil 3 \lg(nw/\varepsilon) \rceil$
- $\gamma = (nw / \varepsilon)^{-9}$
- $\delta = (nw \mathcal{L}(n,w;k) / \varepsilon)^{-3}$
to get a generator with seed length $$\ell = O \big( \left(\log(nw/\epsilon) + \mathcal{L}(n,w;k) \right) \cdot \log n \big)$$ that $O(\varepsilon)$-fools branching programs $F$ of width $w$. From this we derive the following two corollaries by invoking either the trivial bound or the bound from [@chrt17] on the level-$k$ Fourier mass of width-$w$ branching programs. The second of these corollaries proves .
There exists an explicit PRG with seed length $$\ell = O \big( \log(nw/\varepsilon) \cdot \log^2 n \big)$$ that $\varepsilon$-fools unknown-order branching programs of width $w$.
There exists an explicit PRG with seed length $$\ell = O \big( w \cdot \log(nw/\varepsilon) \cdot \log n \cdot \log \log n \big)$$ that $\varepsilon$-fools unknown-order branching programs of width $w$.
Proof Strategy {#strategy}
==============
We show that $G_r$ successfully fools branching programs with the following inductive analysis. By adding and subtracting the term $F(D_i + T_i \land U)$, we have
$$\begin{aligned}
{\|{\mathbb{E}}_{G_i} F(G_i) - {\mathbb{E}}_U F(U) \|}
&= {\|{\mathbb{E}}_{D_i} {\mathbb{E}}_{T_i} {\mathbb{E}}_{G_{i-1}} F(D_i + T_i \land G_{i-1}) - {\mathbb{E}}_U F(U) \|} \\\
&\leq {\|{\mathbb{E}}_{D_i} {\mathbb{E}}_{T_i} {\mathbb{E}}_{U} F(D_i + T_i \land U) - {\mathbb{E}}_U F(U) \|} \\
& \; + {\mathbb{E}}_{D_i} {\mathbb{E}}_{T_i} {\|{\mathbb{E}}_{G_{i-1}} F(D_i + T_i \land G_{i-1}) - {\mathbb{E}}_{U} F(D_i + T_i \land U) \|}.\end{aligned}$$
Since, for any fixed vectors $d,t \in {\mathbb{F}}_2^n$, the function $F'(x) : = F(d + t \land x)$ is again some branching program, we argue that $F'$ is fooled inductively. The bulk of our proof is then spent arguing that ${\mathbb{E}}_{D,T,U} F(D + T \land U) \approx {\mathbb{E}}_U F(U)$.
The starting point of this argument is the observation that the “noise-like” distribution $T \land U$ successfully fools any function that is divisible by a high-degree Fourier character. Specifically, suppose that $T$ is $k$-wise independent and that $\alpha \in {\mathbb{F}}_2^n$ has hamming weight $|\alpha| \geq k$. Let $g : {\mathbb{F}}_2^n \rightarrow [-1, 1]$ be an arbitrary function such that $g$ and ${\chi_{\alpha}}$ depend on disjoint sets of input bits. Then $T \land U$ fools the function $f {:=}{\chi_{\alpha}}\cdot g$ with error $\varepsilon = \nicefrac{1}{2^k}$: $${\mathbb{E}}_U f(U) = {\left( {\mathbb{E}}_U {\chi_{\alpha}}(U) \right)} {\left( {\mathbb{E}}_U g(U) \right)} = 0
\;,$$ and $$\begin{aligned}
|{\mathbb{E}}_{T,U} f(T \land U)|
&=\left| {\mathbb{E}}_T \big( {\mathbb{E}}_U {\chi_{\alpha}}(T \land U) \big) \big({\mathbb{E}}_U g(U) \big) \right|\\
&\leq {\mathbb{E}}_T| {\mathbb{E}}_U {\chi_{\alpha}}(T \land U) | | {\mathbb{E}}_U g(U) |\\
&\leq {\mathbb{E}}_T |{\mathbb{E}}_U {\chi_{\alpha}}(T \land U) | \\
&={\mathbb{E}}_T {{\mathbbold{1}}\! \! {\left( \alpha \land T = 0 \right)}} \\
&\leq \frac{1}{2^k}
\;.\end{aligned}$$
Overall, we wish to enact the following plan. If $F$ encodes our branching program, we wish to use Fourier analysis to rewrite $F$ as a sum of simpler terms, and then use linearity of expectation together with a triangle inequality to argue that $D+T \land U$ fools each term separately.[^7] Recall the product structure of $F$, $$F(x) = F_1(x_1) F_2(x_2) \cdots F_n(x_n),$$ where $F_i(x_i) = A_{i, x_i}$. We can imagine taking the Fourier expansions of these one-bit factors: $$F_i(x_i) = \tfrac{1}{2} (A_{i,0} + A_{i,1}) + \tfrac{1}{2} (A_{i,0} - A_{i,1}) (-1)^{x_i} {=:}B_{i,0} + B_{i,1} (-1)^{x_i}
\;,$$ so that $F$ has the form $$F(x) = \prod_{i=1}^n \big( B_{i,0} + B_{i,1} (-1)^{x_i} \big)
\;.$$ If we expand this product completely, we recover the Fourier expansion of $F$. Certainly the terms in this sum are simple enough, but the problem is that there are too many of them; we cannot afford a $2^n$-fold triangle inequality. Instead, we expand this product more “slowly”: only until we see that some term has collected a degree-$k$ Fourier character as a factor. By the above observation, this term can be killed immediately, and we go on expanding the remaining terms. At the end, we are left with only low-degree Fourier coefficients, which can all be fooled by the distribution $D$. By doing this carefully, we get by with only $n$ applications of the triangle inequality.
Proof of
=========
Suppose $$F(x) = F_1 (x_1) F_2(x_2) \cdots F_n(x_n)$$ encodes a branching program. We define the subprograms of $F$, $$F^{\leq i}(x_1,\ldots,x_i){:=}F_1(x_1) F_2(x_2) \cdots F_i(x_i)$$ and $$F^{> i}(x_{i+1},\ldots,x_n) {:=}F_{i+1}(x_{i+1}) F_{i+2}(x_{i+2}) \cdots F_{n}(x_n)
\;,$$ so that $F = F^{\leq i} \cdot F^{> i}$ for any $i$ (we define $F^{> n}$ as the empty product, so that it is an identity matrix $I$). With this notation in place, we can re-express $F$ in the following convenient form.
\[prop1\] Suppose $F : {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}^{w \times w}$ encodes a branching program. Then for any $k\ge 1$, $F$ can be written as $$F = {\widehat}{F}_{0} + L + \sum_{i = 1}^n H_i \cdot F^{> i},$$ where $$L = \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\widehat}{F}_\alpha \chi_\alpha,$$ and $$H_i = \sum_{\substack{\alpha \in {\mathbb{F}}_2^i \\ |\alpha| = k \\ \alpha_i = 1}} {\widehat}{F}^{\leq i}_\alpha \chi_\alpha.$$
We verify that the expression has the same Fourier expansion as $F$. $$\begin{aligned}
\sum_{i=1}^n H_i \cdot F^{>i}
&= \sum_{i=1}^n
\Bigg( \sum_{\substack{\alpha \in {\mathbb{F}}_2^i \\ |\alpha| = k \\ \alpha_i = 1}} {\widehat}{F}^{\leq i}_\alpha \chi_\alpha \Bigg)
\Bigg( \sum_{\beta \in {\mathbb{F}}_2^{n-i}} {\widehat}{F}^{> i}_\beta \chi_\beta\Bigg) \\
&= \sum_{i=1}^n
\sum_{\substack{\alpha \in {\mathbb{F}}_2^i \\ |\alpha| = k \\ \alpha_i = 1}}
\sum_{\beta \in {\mathbb{F}}_2^{n-i}}
{\widehat}{F}_{\alpha \beta} \chi_{\alpha \beta} \\
&= \sum_{i=1}^n
\sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ \alpha_1 + \alpha_2 + \cdots + \alpha_i = k \\ \alpha_i = 1}}
{\widehat}{F}_{\alpha} \chi_{\alpha}(x) \\
&= \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ |\alpha| \geq k}}
{\widehat}{F}_{\alpha} \chi_{\alpha}
\;.
\qedhere
\end{aligned}$$
Now we derive a useful expectation bound for functions whose Fourier expansions are only supported at degree $k$.
\[prooflem1\] Let $$H(x) = \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ |\alpha| = k}} {\widehat}{H}_{\alpha} {\chi_{\alpha}}(x)$$ be some function whose Fourier expansion is supported only at degree $k$. Let $D$, $T$, and $U$ denote respectively a $2k$-wise independent, a $k$-wise independent, and a uniform distribution over ${\mathbb{F}}_2^n$. Then we have $${\mathbb{E}}_{D,T} \| {\mathbb{E}}_U H(D + T \land U) \|^2
\leq \frac{1}{2^k}\sum_{|\alpha| = k} \|{\widehat}{H}_\alpha \|^2
\;.$$
Firstly, note that $$\begin{aligned}
{\mathbb{E}}_U H(D + T \land U)
&={\mathbb{E}}_U \sum_{|\alpha| = k} {\widehat}{H}_\alpha \cdot \chi_\alpha(D + T \land U) \\
&={\mathbb{E}}_U \sum_{|\alpha| = k} {\widehat}{H}_\alpha \cdot \chi_\alpha(D) \cdot \chi_\alpha(T \land U) \\
&=\sum_{|\alpha| = k} {\widehat}{H}_\alpha \cdot \chi_\alpha(D) \cdot {\mathbb{E}}_U \chi_\alpha(T \land U) \\
&=\sum_{|\alpha| = k} {\widehat}{H}_\alpha \cdot \chi_\alpha(D) \cdot {{\mathbbold{1}}\! \! {\left(\alpha \land T = 0 \right)}}
\;.
\end{aligned}$$
Letting ${\left\langle\cdot ,\cdot \right\rangle}$ denote the Frobenius matrix inner product, we have $$\begin{aligned}
\| {\mathbb{E}}_U H(D + T \land U) \|^2
&= {\left\langle {\mathbb{E}}_U H(D + T\land U) , {\mathbb{E}}_U H(D + T\land U) \right\rangle}\\
&= {\left\langle \sum_{|\alpha| = k} {\widehat}{H}_\alpha \chi_\alpha(D) {{\mathbbold{1}}\! \! {\left(\alpha \land T = 0 \right)}} , \sum_{|\beta| = k} {\widehat}{H}_\beta \chi_\beta(D) {{\mathbbold{1}}\! \! {\left(\beta \land T = 0 \right)}} \right\rangle}\\
&= \sum_{|\alpha| = k} \sum_{|\beta| = k} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\beta \right\rangle} \cdot \chi_{\alpha + \beta}(D) \cdot {{\mathbbold{1}}\! \! {\left((\alpha \lor \beta) \land T = 0 \right)}}
\end{aligned}$$ and $$\begin{aligned}
{\mathbb{E}}_{D} \| {\mathbb{E}}_U H(D + T \land U) \|^2
&= {\mathbb{E}}_D \sum_{|\alpha| = k} \sum_{|\beta| = k} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\beta \right\rangle} \chi_{\alpha + \beta}(D) {{\mathbbold{1}}\! \! {\left((\alpha \lor \beta) \land T = 0 \right)}} \\
&= \sum_{|\alpha| = k} \sum_{|\beta| = k} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\beta \right\rangle} \Big( {\mathbb{E}}_D \chi_{\alpha + \beta}(D) \Big) {{\mathbbold{1}}\! \! {\left( (\alpha \lor \beta) \land T = 0 \right)}} \\
&= \sum_{|\alpha| = k} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\alpha \right\rangle} {{\mathbbold{1}}\! \! {\left(\alpha \land T = 0 \right)}} \\
&= \sum_{|\alpha| = k} {\|{\widehat}{H}_\alpha \|}^2 {{\mathbbold{1}}\! \! {\left(\alpha \land T = 0 \right)}}
\;.
\end{aligned}$$ Finally, $$\begin{aligned}
{\mathbb{E}}_{D,T} \| {\mathbb{E}}_U H(D + T \land U) \|^2
&= {\mathbb{E}}_T {\mathbb{E}}_D \| {\mathbb{E}}_U H(D + T \land U) \|^2 \\
&= {\mathbb{E}}_T \sum_{|\alpha| = k} {\|{\widehat}{H}_\alpha \|}^2 {{\mathbbold{1}}\! \! {\left(\alpha \land T = 0 \right)}}\\
&= \sum_{|\alpha| = k} {\|{\widehat}{H}_\alpha \|}^2 \cdot {\mathbb{E}}_T {{\mathbbold{1}}\! \! {\left(\alpha \land T = 0 \right)}} \\
&=\sum_{|\alpha| = k} {\|{\widehat}{H}_\alpha \|}^2 \cdot \left( \tfrac{1}{2} \right)^k
\;.
\qedhere
\end{aligned}$$
We now have the tools in place to prove our main technical lemma.
\[prooflem2\] Suppose $F : {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}^{w \times w}$ encodes a branching program. Let $D$, $T$, and $U$ denote respectively a $2k$-wise independent, a $k$-wise independent, and a uniform distribution over ${\mathbb{F}}_2^n$. Then $D + T \land U$ fools $F$ with error $$\varepsilon = {\|{\mathbb{E}}_{D,T,U}F(D + T \land U) - {\mathbb{E}}_U F(U) \|} \leq \frac{n w}{2^{k/2}} .$$
To analyze $\| {\mathbb{E}}_{D,T,U} F(D + T \land U) - {\mathbb{E}}_{U} F(U) \|$, we use the preceding expansion of $F$ together with linearity of expectation and the triangle inequality. Recalling that ${\mathbb{E}}_U F(U) = {\widehat}{F}_0$, this gives $$\begin{aligned}
\| {\mathbb{E}}_{D,T,U} F(D + T \land U) - {\mathbb{E}}_{U} F(U) \| &\leq \\
\| {\mathbb{E}}_{D,T,U} L(D + T \land U) \|& + \sum_{i=1}^n \| {\mathbb{E}}_{D,T,U} H_i(D + T \land U) F^{> i}(D + T \land U) \|
\;.
\end{aligned}$$ The low-degree term is dealt with easily by $D$: $$\begin{aligned}
{\mathbb{E}}_{D,T,U} L(D + T \land U)
&={\mathbb{E}}_D {\mathbb{E}}_T {\mathbb{E}}_U \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\widehat}{F}_\alpha \chi_\alpha(D + T \land U) \\
&={\mathbb{E}}_D {\mathbb{E}}_T {\mathbb{E}}_U \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\widehat}{F}_\alpha {\chi_{\alpha}}(D) {\chi_{\alpha}}(T \land U) \\
&={\mathbb{E}}_T {\mathbb{E}}_U \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\widehat}{F}_\alpha \cdot \Big( {\mathbb{E}}_D {\chi_{\alpha}}(D) \Big) \cdot {\chi_{\alpha}}(T \land U) \\
&=0
\;.
\end{aligned}$$ Now for each $i$ we have $$\begin{aligned}
\| {\mathbb{E}}_{D,T,U} H_i(D + T \land U) F^{> i}(D + T \land U) \|
&=\left\| {\mathbb{E}}_{D,T} \big({\mathbb{E}}_U H_i(D + T \land U) \big) \big({\mathbb{E}}_U F^{> i} (D + T \land U) \big) \right\|\\
&\le{\mathbb{E}}_{D,T} \| {\mathbb{E}}_U H_i(D + T \land U) \| \|{\mathbb{E}}_U F^{> i} (D + T \land U) \| \\
&\le{\mathbb{E}}_{D,T} \| {\mathbb{E}}_U H_i(D + T \land U) \| w^{1/2} \\
&\le\Big( {\mathbb{E}}_{D,T} \|{\mathbb{E}}_U H_i(D + T \land U) \|^2 \Big)^{1/2} w^{1/2} \\
&\le\left( \tfrac{1}{2} \right)^{k/2} \Big( \sum_{\alpha \in {\mathbb{F}}_2^n} {\|{\widehat}{F}^{\leq i}_\alpha \|}^2 \Big)^{1/2} w^{1/2}\\
&=\left( \tfrac{1}{2} \right)^{k/2} w,
\end{aligned}$$ where we get the final equality by applying the Parseval identity to $F^{\leq i}$.
Proof of
---------
Recall the induction framework outlined in :
$$\begin{aligned}
{\|{\mathbb{E}}_{G_i} F(G_i) - {\mathbb{E}}_U F(U) \|}
&= {\|{\mathbb{E}}_{D_i} {\mathbb{E}}_{T_i} {\mathbb{E}}_{G_{i-1}} F(D_i + T_i \land G_{i-1}) - {\mathbb{E}}_U F(U) \|} \\
&\leq {\|{\mathbb{E}}_{D_i} {\mathbb{E}}_{T_i} {\mathbb{E}}_{U} F(D_i + T_i \land U) - {\mathbb{E}}_U F(U) \|} \\
& \; + {\mathbb{E}}_{D_i} {\mathbb{E}}_{T_i} {\|{\mathbb{E}}_{G_{i-1}} F(D_i + T_i \land G_{i-1}) - {\mathbb{E}}_{U} F(D_i + T_i \land U) \|}.
\end{aligned}$$
We have seen how to carry out the inductive step; it remains to establish the base case. To do this, we wish to think of $F(G_r)$ as a function of $G_0$ only, with $D_1, D_2, \ldots, D_r$ and $T_1, T_2, \ldots, T_r$ fixed.
Specifically, we do the following. Define the strings $g_0 {:=}x$ and $g_{i} {:=}D_i + T_i \land g_{i-1}$, and define the function $f(x) {:=}F(g_r)$. Note that with this setup we have $F(G_r) = f(G_0)$, and so
$$\begin{aligned}
{\|{\mathbb{E}}_{G_r} F(G_r) - {\mathbb{E}}_U F(U) \|}
&= {\|{\mathbb{E}}_{D_r} {\mathbb{E}}_{T_r} {\mathbb{E}}_{G_{r-1}} F(D_r + T_r \land G_{r-1}) - {\mathbb{E}}_U F(U) \|} \\
&\leq \frac{nw}{2^{k/2}} + {\mathbb{E}}_{D_r} {\mathbb{E}}_{T_r} {\|{\mathbb{E}}_{G_{r-1}} F(D_r + T_r \land G_{r-1}) - {\mathbb{E}}_{U} F(D_r + T_r \land U) \|} \\
&\leq \frac{nw}{2^{k/2}} + \frac{nw}{2^{k/2}} + \cdots + \frac{nw}{2^{k/2}}+ \underset{\substack{D_1, D_2, \ldots, D_r \\ T_1, T_2, \ldots, T_r}} {\mathbb{E}}{\| {\mathbb{E}}_{G_0} f(G_0) - {\mathbb{E}}_U f(U) \|}.
\end{aligned}$$
Now we must show that the function $f$ is fooled by $G_0$ for most values of $D_i$ and $T_i$. Luckily, for $r$ large enough, $f$ is often a constant function and therefore fooled by any distribution.
In particular, let $T_i[j]$ denote the $j$-th bit of $T_i$ and define the indicator random variables $$Y_j = \bigwedge\limits_{i=1}^r T_i[j].$$ Note that $f(x)$ depends on the $j$-th bit of $x$ only if $Y_j = 1$, and so $f(x)$ is constant if $\sum_{j=1}^n Y_j = 0$. Also note that ${{\Pr}\! {\left(Y_j = 1 \right)}} = 2^{-r}$. By applying a Markov inequality, we have
$$\begin{aligned}
{\mathbb{E}}_{\substack{D_1, D_2, \ldots, D_r \\ T_1, T_2, \ldots, T_r}} {\| {\mathbb{E}}_{G_0} f(G_0) - {\mathbb{E}}_U f(U) \|}
&\leq {{\Pr}\! {\left(\sum_{j=1}^n Y_j \geq 1 \right)}} \underset{\substack{D_1, D_2, \ldots, D_r \\ T_1, T_2, \ldots, T_r}} \max {\| {\mathbb{E}}_{G_0} f(G_0) - {\mathbb{E}}_U f(U) \|} \\
&\leq {\mathbb{E}}\left[\sum_{j=1}^n Y_j \right] \cdot 2 w^{1/2} \\
&= \frac{2 n w^{1/2}}{2^r}
\;.
\end{aligned}$$
If we set $k \geq 5 \lg n + 2 \lg w$ and $r \geq 2 \lg n + \frac{1}{2} \lg w$, $G_r$ fools $F$ with error $$\varepsilon = {\|{\mathbb{E}}_{G_r} F(G_r) - {\mathbb{E}}_U F(U) \|} \leq r \cdot \frac{nw}{2^{k/2}} + \frac{2 n w^{1/2}}{2^r} \leq \frac{3}{n}
\;.
\qedhere$$
Proof of
=========
The proof of follows the same story as the previous section with details differing in two places. First, we derive an analogue of which is slightly messier due to our now weaker pseudorandom primitives. Secondly, in order to get the best possible seed-length in the small error regime, this time we analyze the base case with a bit more care. In particular, we upgrade the Markov argument to a Chernoff bound for $\gamma$-almost $k$-wise independent variables.
Let $$H(x) = \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ |\alpha| = k}} {\widehat}{H}_{\alpha} {\chi_{\alpha}}(x)$$ be some function whose Fourier expansion is supported only at degree $k$. Let $D$, $T$, and $U$ denote respectively a $\delta$-biased, a $\gamma$-almost $k$-wise independent, and a uniform distribution over ${\mathbb{F}}_2^n$. Then we have $${\mathbb{E}}_{D,T} \| {\mathbb{E}}_U H(D + T \land U) \|^2
\leq \Big( 2^{-k} + \gamma \Big) \left( \delta \cdot \Big( \sum_{|\alpha| = k} \|{\widehat}{H}_\alpha \| \Big)^2
+\sum_{|\alpha| = k} \|{\widehat}{H}_\alpha \|^2 \right)
\;.$$
As before, we have $$\| {\mathbb{E}}_U H(D + T \land U) \|^2
= \sum_{|\alpha| = k} \sum_{|\beta| = k} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\beta \right\rangle} \chi_{\alpha + \beta}(D) {{\mathbbold{1}}\! \! {\left( (\alpha \lor \beta) \land T= 0 \right)}}
\;.$$ We analyze the terms with $\alpha = \beta$ and $\alpha \neq \beta$ separately. For the cross terms we have $$\begin{aligned}
{\mathbb{E}}_D
&{\mathbb{E}}_T \sum_{|\alpha| = k} \sum_{\beta \neq \alpha}{\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\beta \right\rangle} \chi_{\alpha + \beta}(D) {{\mathbbold{1}}\! \! {\left((\alpha \lor \beta) \land T = 0 \right)}} \\
&= \sum_{|\alpha| = k}\sum_{\beta \neq \alpha} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\beta \right\rangle} \Big( {\mathbb{E}}_D \chi_{\alpha + \beta}(D) \Big) \Big( {\mathbb{E}}_T {{\mathbbold{1}}\! \! {\left((\alpha \lor \beta) \land T = 0 \right)}} \Big) \\
&\leq \sum_{|\alpha| = k}\sum_{\beta \neq \alpha} {\| {\widehat}{H}_\alpha \|} {\| {\widehat}{H}_\beta \|} \cdot \delta \cdot \Big( 2^{-k} + \gamma \Big) \\
&\leq \delta \cdot \Big( 2^{-k} + \gamma \Big) \Big( \sum_{|\alpha| = k} {\| {\widehat}{H}_\alpha \|} \Big)^2
\;.
\end{aligned}$$
For the like terms we have $$\begin{aligned}
{\mathbb{E}}_T
&{\mathbb{E}}_D \sum_{|\alpha| = k} {\left\langle {\widehat}{H}_\alpha , {\widehat}{H}_\alpha \right\rangle} \chi_0(D) {{\mathbbold{1}}\! \! {\left(T \land \alpha = 0 \right)}} \\
&= \sum_{|\alpha| = k} {\|{\widehat}{H}_\alpha \|}^2 \cdot {\mathbb{E}}_T {{\mathbbold{1}}\! \! {\left(T \land \alpha = 0 \right)}} \\
&\leq\sum_{|\alpha| = k} {\|{\widehat}{H}_\alpha \|}^2 \cdot \Big(2^{-k} + \gamma \Big)
\;.
\qedhere
\end{aligned}$$
We now derive the following analogue of .
Suppose $F : {\mathbb{F}}_2^n \rightarrow {\mathbb{R}}^{w \times w}$ encodes a branching program. Let $D$, $T$, and $U$ denote respectively a $\delta$-biased, a $\gamma$-almost $k$-wise independent, and a uniform distribution over ${\mathbb{F}}_2^n$, and let Then $D + T \land U$ fools $F$ with error $$\varepsilon = {\|{\mathbb{E}}_{D,T,U}F(D + T \land U) - {\mathbb{E}}_U F(U) \|} \leq \left( \sqrt{\delta} \mathcal{L}(n,w;k) + \left(\tfrac{1}{2}\right)^{k/2} + \sqrt{\gamma} \right) \cdot nw.$$
Again we use to split $F$ into high and low degree components:
$$\begin{aligned}
\| {\mathbb{E}}_{D,T,U} F(D + T \land U) - {\mathbb{E}}_{U} F(U) \| &\\
\leq \| {\mathbb{E}}_{D,T,U} L(D + T \land U) \|& + \sum_{i=1}^n \| {\mathbb{E}}_{D,T,U} H_i(D + T \land U) F^{> i}(D + T \land U) \|.
\end{aligned}$$
For the low-degree component we have
$$\begin{aligned}
{\| {\mathbb{E}}_{D,T,U} F(D + T \land U) \|}
&= {\| {\mathbb{E}}_D {\mathbb{E}}_T {\mathbb{E}}_U \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\widehat}{F}_\alpha \chi_\alpha(D + T \land U) \|} \\
&= {\| {\mathbb{E}}_T {\mathbb{E}}_U \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\widehat}{F}_\alpha \Big( {\mathbb{E}}_D {\chi_{\alpha}}(D) \Big) {\chi_{\alpha}}(T \land U) \|}\\
&\leq \sum_{\substack{\alpha \in {\mathbb{F}}_2^n \\ 0 < |\alpha| < k}} {\| {\widehat}{F}_\alpha \|} \cdot \delta \\
&= \delta \sum_{i=1}^{k-1} \mathcal{L}_i(F).
\end{aligned}$$
Now we proceed as before.
$$\begin{aligned}
\| {\mathbb{E}}_{D,T,U} F(D + T \land U) - &{\mathbb{E}}_{U} F(U) \| \\
&\le\delta \sum_{i=1}^{k-1} \mathcal{L}_i(F) + \sum_{i=1}^n \| {\mathbb{E}}_{D,T,U} H_i(D + T \land U) F^{> i}(D + T \land U) \| \\
&\le\delta \sum_{i=1}^{k-1} \mathcal{L}_i(F) + \sum_{i=1}^n \Big( {\mathbb{E}}_{D,T} \|{\mathbb{E}}_U H_i(D + T \land U) \|^2 \Big)^{1/2} w^{1/2} \\
&\le\delta \sum_{i=1}^{k-1} \mathcal{L}_i(F) + \sum_{i=1}^n \Big( \delta \mathcal{L}_k(F^{\leq i})^2 + \left(\tfrac{1}{2}\right)^k + \gamma \Big)^{1/2} w \\
&\le\left( \sqrt{\delta} \mathcal{L}(n,w;k) + \left(\tfrac{1}{2}\right)^{k/2} + \sqrt{\gamma} \right) \cdot nw
\;.
\qedhere
\end{aligned}$$
Proof of
---------
We proceed as in the proof of , except this time we derive a sharper bound on the quantity $$\underset{\substack{D_1, D_2, \ldots, D_r \\ T_1, T_2, \ldots, T_r}} {\mathbb{E}}{\| {\mathbb{E}}_{G_0^*} f(G_0^*) - {\mathbb{E}}_U f(U) \|}.$$ Again, define the random variable $$Y = T_1 \land T_2 \land \cdots \land T_r.$$ Recall that $G_0^*$ is a $320 k$-wise independent distribution, so if $|Y| \leq 320 k$ then $${\| {\mathbb{E}}_{G_0^*} f(G_0^*) - {\mathbb{E}}_U f(U) \|} = 0.$$ Therefore
$$\begin{aligned}
\underset{\substack{D_1, D_2, \ldots, D_r \\ T_1, T_2, \ldots, T_r}} {\mathbb{E}}{\| {\mathbb{E}}_{G_0^*} f(G_0^*) - {\mathbb{E}}_U f(U) \|}
&\leq {{\Pr}\! {\left(|Y| \geq 320 k \right)}} \underset{\substack{D_1, D_2, \ldots, D_r \\ T_1, T_2, \ldots, T_r}} \max {\| {\mathbb{E}}_{G_0^*} f(G_0^*) - {\mathbb{E}}_U f(U) \|} \\
&\leq {{\Pr}\! {\left(|Y| \geq 320 k \right)}} \cdot 2 w^{1/2}.
\end{aligned}$$
We appeal to the following extension of the Chernoff bound for $k$-wise independent variables.
[[(see [@svw14], Lemma A.1)]{.nodecor} ]{} Suppose $X_1, X_2, \ldots X_t$ are $\gamma$-almost $k$-wise independent variables with $X_i \in \{0,1\}$. Then $${{\Pr}\! {\left(\frac{1}{t}\sum_{i=1}^t X_i \geq \frac{1}{2} + \frac{a}{2} \right)}} \leq \left(\frac{40k}{a^2 t} \right)^{\lfloor k/2 \rfloor} + 2 \gamma \left( \frac{2}{a}\right)^k.$$
As a result, if $T$ is a $\gamma$-almost $k$-wise independent distribution and $\alpha$ is any fixed bitmask with hamming weight $|\alpha| \geq 320k$, we have $${{\Pr}\! {\left(|\alpha \land T| \geq \tfrac{3}{4} |\alpha| \right)}} \leq \left( \tfrac{1}{2} \right)^{\lfloor k/2 \rfloor} + 2 \gamma \cdot 4^k.$$ Noting that $ \left(\frac{3}{4}\right)^r n \leq 1$, a simple union bound argument shows that $${{\Pr}\! {\left(|T_1 \land T_2 \land \cdots \land T_r| \geq 320 k \right)}} \leq r \cdot \left( \left( \tfrac{1}{2} \right)^{\lfloor k/2 \rfloor} + 2 \gamma \cdot 4^k \right) .$$
To conclude, we have $$\begin{aligned}
{\|{\mathbb{E}}_{G_r^*} F(G_r^*) - {\mathbb{E}}_U F(U) \|}
&\leq r \cdot \left( \sqrt{\delta} \mathcal{L}(n,w;k) + \left(\tfrac{1}{2}\right)^{k/2} + \sqrt{\gamma} \right) \cdot nw
+ r \cdot \left( \left( \tfrac{1}{2} \right)^{\lfloor k/2 \rfloor} + 2 \gamma \cdot 4^k \right) \cdot 2w^{1/2} \\
&\leq O{\left( \left( \sqrt{\delta} \mathcal{L}(n,w;k) + \left(\tfrac{1}{2}\right)^{k/2} + \sqrt{\gamma} + \gamma 4^k \right) \cdot nwr \right)}
\;.
\qedhere
\end{aligned}$$
[^1]: Email: `[email protected]`. Department of Computer Science, University of Illinois at Urbana-Champaign. Supported by NSF grant CCF-1755921.
[^2]: Email: `[email protected]`. Department of Computer Science, University of Illinois at Urbana-Champaign. Supported by NSF grant CCF-1755921.
[^3]: By this we mean quantitative progress in the constant-error regime. Recently in [@bcg17], Braverman, Cohen, and Garg give a hitting set (a “one-sided” PRG) with a better seed-length in the small-error regime than Nisan’s generator.
[^4]: In fact, their generator fools the more general model of branching programs that may read the input bits any number of times and in any adaptive order.
[^5]: Note that this grouping depends on the order of the variables. However, this grouping only occurs in the *analysis*, and that the construction itself is oblivious to the variable order.
[^6]: Actually, one can easily derive the slightly better bound $\mathcal{L}_k(F) \leq \binom{n}{k}^{1/2} w^{1/2}$ by first applying Cauchy-Schwarz followed by Parseval’s identity.
[^7]: This approach is inspired by the similar arguments of Haramaty, Lee, and Viola employed in [@hlv17]. However, the generators they produce with this idea have seed length $\geq n^{1/2}$, while we achieve generators with poly-logarithmic seed-length.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss transport equations resulting from relativistic diffusions in the proper time. We show that a solution of the transport equation can be obtained from the solution of the diffusion equation by means of an integration over the proper time. We study the stochastic processes solving the relativistic diffusion equation and the relativistic transport equation. We show that the relativistic transport equation for massive particles in the light-cone coordinates and for the massless particles in spatial momentum coordinates are related to the (generalized) Bessel diffusion which has an analytic solution. The solution describes a particle moving in a fixed direction whose frequency distribution is the Bessel process. An approach to an equilibrium in a moving frame is discussed. We formulate the equilibrating diffusion and transport processes in a Lorentz covariant way.'
author:
- |
Z. Haba\
Institute of Theoretical Physics, University of Wroclaw,\
50-204 Wroclaw, Plac Maxa Borna 9, Poland\
PACS:02.50.Ey,05.10.Gg,25.75.-q
title: Relativistic diffusive transport
---
Introduction
============
There are many versions of the relativistic diffusion and of the relativistic transport (see [@hanggirev][@debbaschrev][@chev] for reviews). However, if we assume that the diffusing particle does not change its mass, the diffusion evolves in the proper time and that it is generated by the second order Lorentz invariant differential operator, then its definition is unique as shown by Schay [@schay] and Dudley [@dudley]. The theory can be generalized to describe a particle diffusing on a manifold (pseudoriemannian background metric) [@lejan][@franchi]. We have extended these models to a diffusion of a particle with spin [@habajpa]. We have shown [@habagen] that in the models of the relativistic diffusion the energy and the angular momentum grow exponentially fast in the proper time. Such a growth is rather unphysical in a steady world around us. In refs.[@habapre][@habagen] we have studied a modification of the model by an addition of some drag terms (friction) resulting from the requirement that a definite equilibrium state is achieved at a large time. In [@habakomp] we have shown that a relativistic diffusion equilibrating to the Bose-Einstein equilibrium is a linearized version of the Kompaneets equation [@komp] applied for a long time to a description of photons diffusing in an electron gas [@rybicki]. There are phenomena in astrophysics [@rybicki] [@peebles][@dodelson] as well as in particle physics [@ion][@svet] which can be described by relativistic kinetic and diffusion equations. The general theory of relativistic kinetic equations has been developed in [@stewart][@israel]. Various forms of relativistic kinetic equations can be derived from relativistic quantum field theory [@zachar][@elze] [@kelly]. These kinetic equations are applied to the quark-gluon plasma [@svet] and to the heavy ion collisions [@ion] often in the diffusion approximation.
In our earlier papers [@habapre][@habajpa] we have concentrated on the diffusion in the proper time. The proper time is a reasonable mathematical tool (see [@landau] for its use in relativistic dynamics) to keep explicit Lorentz invariance. However, we should be able to describe the relativistic phenomena also in the coordinate time (we call it the laboratory time). In relativistic dynamics there is an equation relating the proper time with the coordinate time and we can exchange one into the other. The probability density as a function of the proper time trajectories should be independent of the proper time. This requirement is equivalent to an evolution equation of the probability density in the coordinate time (the kinetic equation). It is natural to interpret an analogous requirement (the diffusion probability distribution defined on the Minkowski space-time is independent of the proper time) as the kinetic equation (the transport equation) in the stochastic case.
We can obtain the transport equation by a random time change as well. This is a stochastic analogue of the change of the proper time (associated with a moving particle) into a laboratory time. In this paper the random time change (which is a standard notion in the theory of diffusion processes [@skorohod]) is formulated as a calculational tool useful for analytic as well as numerical calculations. The method involves an integration over the proper time analogous to Feynman’s proper time formulation of quantum mechanics [@feynman].
In the case of particles of moderate velocity (non-zero mass) the meaning of the proper time diffusion can be elaborated by means of an expansion of momenta in powers of $(mc)^{-1}$. In this paper we study in detail the ultrarelativistic case of massless particles (then the proper time is a formal affine parameter on the trajectory) and concentrated beams of massive particles when only light-cone momenta $p_{+}$ are relevant. We perform probabilistic calculations of the integral over the proper time by an application of some results of Yor [@yor][@yor2]. It comes out that the solution of the transport equation obtained as an integral over the proper time from the solution of the relativistic diffusion is expressed by Bessel functions. This is not accidental as the transport equation coincides with the Bessel diffusion. We compare our probabilistic proper time method of the solution of the transport equation with the theory of Bessel diffusions.
In general, the relativistic diffusion will have no limit for a large proper time (no equilibrium). For the same reason the solution of the transport equation will have no limit as the laboratory time tends to infinity. The diffusion equation must be supplemented by some drag terms describing a friction if the diffusion is to achieve an equilibrium. The equilibrium distribution depends on the velocity of the frame where it is observed. So, the equilibrium is a covariant but not an invariant notion. In this paper we formulate the dependence of the relativistic dynamics on the equilibration reference frame in a covariant way.
The plan of the paper is the following. In sec.2 we review the mathematical scheme of the relativistic diffusion theory in the proper time. In this section we also formulate and prove the basic theorem which allows to obtain a solution of the transport equation if the diffusion in the proper time is known. In sec.3 we discuss various coordinates which will be applied in subsequent sections. In sec.4 transport equations for relativistic diffusions with drifts leading to an equilibrium are discussed. In sec.5 the relativistic diffusion process as a function of the Brownian motion is obtained in an explicit form. In sec.6 a diffusion of massless particles is discussed. Sec.7 is the main part of this paper. We study a relation between the relativistic diffusion in the proper time and the transport equation in the ultrarelativistic case of massless particles and the light-cone variables for massive particles. We solve the transport equation by an application of the theorem of sec.2. These equations are solved also by another method. The result is that the diffusing particle has the momentum whose probability distribution is the Bessel process. In sec. 8 we discuss the frame dependence of the equilibrating diffusion process. We formulate the theory in terms of relativistic invariant variables. The results are summarized in the last section of this paper.
The general scheme of the relativistic diffusion and relativistic transport
===========================================================================
Following [@schay][@dudley][@habapre] we consider an evolution in the proper time of points on the phase space $(x,p)$ $$\frac{dx^{\mu}}{d\tau}=\frac{1}{m}p^{\mu},$$ $$\frac{dp^{\mu}}{d\tau}=F^{\mu}$$ preserving the mass-shell ${\cal H}_{+}$ $$\eta^{\mu\nu}p_{\nu}p_{\mu}=p^{\mu}p_{\mu}=p_{0}^{2}-{\bf
p}^{2}=m^{2}c^{2}.$$ In eq.(2) $F$ can be a random force leading to a diffusive behaviour of a test particle described by coordinates $(x,p)$. From eq.(1) and (3) (for $m>0$) it follows that $\tau$ really has the meaning of the proper time. Then, the evolution of a scalar function of the trajectory is determined by the equation $$\partial_{\tau}\phi={\cal G}\phi=(p^{\mu}\partial_{\mu}^{x}+{\cal
A})\phi,$$ where (in the case of a diffusion) ${\cal A}$ is an $O(3,1)$ invariant second order differential operator; differentiation over space-time coordinates has an index $x$ whereas differentiation without an index concerns momenta. The probability density $\Phi$ evolves according to an adjoint equation $$\partial_{\tau}\Phi_{\tau}={\cal
G}^{*}\Phi_{\tau}=(-p^{\mu}\partial_{\mu}^{x}+{\cal
A}^{*})\Phi_{\tau}$$ resulting from ($x\in R^{4}$, ${\bf p}\in R^{3}$; fourvectors have Greek indices and threevectors Latin indices) $$\int dxd{\bf p}\phi_{\tau}(x,{\bf p})\Phi(x,{\bf p})= \int dxd{\bf
p} \phi(x,{\bf p})\Phi_{\tau}(x,{\bf p}).$$ ${\cal
A}$ is the second order differential operator such that ${\cal
A}1=0$ and the quadratic term is negatively definite. Then, there exists a diffusion process $(x_{\tau},p_{\tau})$ starting from $(x,p)$ at $\tau=0$ such that $$\phi_{\tau}=E[\phi(x_{\tau},p_{\tau})]\equiv{\cal K}_{\tau}\phi,$$ where $${\cal K}_{\tau}(x,{\bf p};x^{\prime},{\bf p}^{\prime})=E[\delta(
x_{\tau}(x,p)-x^{\prime})\delta({\bf p}_{\tau}(x,{\bf p})-{\bf
p}^{\prime})].$$ From eq.(6) $$\Phi_{\tau}={\cal K}_{\tau}^{*}\Phi$$ where $${\cal K}_{\tau}^{*}(x,{\bf p};x^{\prime},{\bf
p}^{\prime})=E[\delta(
x_{\tau}(x^{\prime},p^{\prime})-x)\delta({\bf
p}_{\tau}(x^{\prime},{\bf p}^{\prime})-{\bf p})].$$ The probability distribution $\Phi$ is independent of $\tau$ if it satisfies the transport equation $${\cal G}^{*}\Phi=0.$$
We define an equilibrium distribution $\Phi_{E}$ as an $x^{0}$ independent solution of the transport equation. Let $$\Phi=\Psi\Phi_{E}.$$ Then, $\Psi$ satisfies a diffusion equation $$\partial_{\tau}\Psi=\hat{{\cal G}}\Psi,$$where $$\hat{{\cal G}}=-p^{\mu}\partial_{\mu}^{x}+\hat{{\cal A}}$$and $\hat{{\cal A}}=\Phi_{E}^{-1}{\cal A}\Phi_{E}$ is a second order differential operator closely related to ${\cal A}$. The solution of eq.(13) can be expressed by a stochastic process $$\Psi_{\tau}=E[\Psi(\hat{x}_{\tau},\hat{p}_{\tau})].$$
Then, if $\Phi$ is a solution of the transport equation (11) then $\Psi$ is a solution of a diffusion equation ( we set $t=\frac{x^{0}}{c}\geq 0$) $$\partial_{t}\Psi=\frac{c}{p_{0}}{\bf p}\nabla_{\bf x} \Psi+\frac{c}{p_{0}}\hat{{\cal
A}}\Psi.$$ A solution of the diffusion equation (16) can again be expressed by a diffusion process $(\tilde{{\bf x}}_{t},\tilde{{\bf p}_{t}})$ $$\Psi_{t}=E[\Psi(\tilde{{\bf x}}_{t},\tilde{{\bf p}_{t}})].$$ We can establish a relation between the solutions of the diffusion equations (13) and (16)
[**Theorem**]{}
Let $(\hat{x}_{\tau},\hat{p}_{\tau})$ be the diffusion process (15) solving the proper time diffusion equation (13) then the solution of the transport equation (16) with the initial condition $\Psi({\bf x},{\bf p})$ reads $$\Psi_{t}({\bf x},{\bf p})=\frac{1}{mc}\int_{0}^{\infty}d\tau
E\Big[\Psi(\hat{{\bf x}}_{\tau},\hat{{\bf p}}_{\tau})
\hat{p}_{0}(\tau)\delta\Big(t-\frac{1}{mc}\int_{0}^{\tau}ds\hat{p}_{0}(s)\Big)\Big]$$ where $$\hat{p}_{0}(s)=\sqrt{m^{2}c^{2}+\hat{{\bf
p}}_{s}^{2}}.$$
[**Proof**]{}: Let us first check the initial condition. For small $t$ the proper time $\tau$ is also small and $\int_{0}^{\tau}p_{0}(s)ds\simeq p_{0}\tau$, then $x_{\tau}\simeq
x$ and $p_{\tau}\simeq p$. An integration over $\tau$ gives $\Psi({\bf x},{\bf p})$ when $t\rightarrow 0$. Next, for a derivation of the diffusion equation (16) we apply the formula $\delta(f(\tau))=\vert f^{\prime}\vert^{-1}\delta(\tau-\tau(t))$ in order to express the time $\tau$ in the diffusion process $({\bf x}_{\tau},{\bf p}_{\tau})$ by $t$. Then, $\vert
f^{\prime}\vert^{-1}=p_{0}(\tau)^{-1}$ cancels the $p_{0}$ term in eq.(18). There remains a function $\Psi$ of the process $({\bf
x}_{\tau},{\bf p}_{\tau})$ at the time $\tau(t)$ in eq.(18).This is the random time change well-known from the theory of diffusion processes. It is proved in [@skorohod] that after the random change of time eq.(16) is satisfied (we shall still discuss the random time change in sec.7,eqs.(76)-(79)).
We shall apply the formula (18) for an explicit calculation of expectation values. Although in general the formula (18) may be difficult to use for an analytic treatment it still can be very useful for numerical calculations.
We can express the solution of the transport equation (16) in terms of a kernel $\hat{{\cal K}}_{t}$ $$\Psi_{t}=\tilde{{\cal K}}_{t}\Psi.$$ Then, from eqs.(17)-(18) the kernel can be expressed in the form $$\begin{array}{l}
\tilde{{\cal K}}_{t}({\bf x},{\bf p};{\bf x}^{\prime},{\bf
p}^{\prime})\cr =\frac{1}{mc}\int_{0}^{\infty}d\tau
E[\delta(\hat{{\bf x}}_{\tau}({\bf x},{\bf p})-{\bf
x}^{\prime})\delta(\hat{{\bf p}}_{\tau}({\bf p})-{\bf
p}^{\prime}))
p_{0}(\tau)\delta(t-\frac{1}{mc}\int_{0}^{\tau}ds\hat{p}_{0}(s))]
\cr=E[\delta(\tilde{{\bf x}}_{t}({\bf x},{\bf p})-{\bf
x}^{\prime})\delta(\tilde{{\bf p}}_{t}({\bf p})-{\bf
p}^{\prime})].\end{array}$$
In this section we have applied a notation suggesting the choice of spatial momenta ${\bf p}$ as coordinates on the mass-shell (3). We shall apply some other coordinates. The definition of the adjoint in eq.(6) is with respect to the Lebesgue measure which is not relativistic invariant. A density factor ensuring the invariance is contained in $\Phi$. In the next sections we shall apply light-cone coordinates. Then, the evolution in $x_{0}$ is replaced by an evolution in $x_{-}=x^{0}-x^{3}$ but the integration in eq.(6) is still with respect to a Lebesgue measure.
Various coordinate systems
==========================
The relativistic diffusion is generated by the Laplace-Beltrami operator on the mass-shell $p^{2}=m^{2}c^{2}$. If we choose the spatial momenta ${\bf p}$ as coordinates on the mass-shell then the diffusion generator reads $$\begin{array}{l}
2\gamma^{-2}{\cal A}=\triangle_{H}=
(\delta_{jk}+m^{-2}c^{-2}p_{j}p_{k})\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial
p_{k}}+3m^{-2}c^{-2}p_{k}\frac{\partial}{\partial p_{k}}.
\end{array}$$ We shall also use $\kappa^{2}=m^{-2}c^{-2}\gamma^{2}$ as the diffusion constant.
The spatial momenta are convenient for a physical interpretation. However, for a derivation of explicit solutions some other coordinates are more useful. Let us consider the Poincare coordinates $(q_{1},q_{2},q_{3})$ on ${\cal H}_{+}$ which are related to momenta $p$ as follows $$p_{3}+p_{0}=\frac{mc}{q_{3}},$$$$p_{3}-p_{0}=-\frac{mc}{q_{3}}(q_{1}^{2}+q_{2}^{2}+q_{3}^{2}),$$$$p_{1}=\frac{mc q_{1}}{q_{3}},$$$$p_{2}=\frac{mc q_{2}}{q_{3}},$$ where $q_{3}\geq 0$. Then, the metric is $$ds^{2}=(mc)^{2}q_{3}^{-2}
(dq_{1}^{2}+dq_{2}^{2}+dq_{3}^{2})$$ and $$(mc)^{2}\triangle_{H}=q_{3}^{2}(\partial_{1}^{2}+\partial_{2}^{2}+\partial_{3}^{2})
-q_{3}\partial_{3}.$$
The Poincare coordinates are closely related to the light-cone coordinates $(p_{+},p_{a})$ (where $a=1,2$) $$p_{+}=p_{0}+p_{3}.$$ We have $q_{a}=p_{a}p_{+}^{-1}$ and $q_{3}=mcp_{+}^{-1}$. Then $$\begin{array}{l}
\triangle_{H}=\partial_{1}^{2}+\partial_{2}^{2}+(mc)^{-2}p_{+}^{2}\partial_{+}^{2}
+(mc)^{-2}p_{1}^{2}\partial_{1}^{2}\cr+
(mc)^{-2}p_{2}^{2}\partial_{2}^{2}+2(mc)^{-2}p_{+}p_{a}\partial_{a}\partial_{+}
+ 2(mc)^{-2}p_{1}p_{2}\partial_{1}\partial_{2} \cr
+(mc)^{-2}3p_{+}\partial_{+}+(mc)^{-2}3p_{a}\partial_{a}\end{array}$$ where $\partial_{a}=\frac{\partial}{\partial p_{a}}$. The generator ${\cal G}$ in the light-cone coordinates is $${\cal
G}=p_{-}\partial^{x}_{+}+p_{+}\partial^{x}_{-}-p_{a}\partial^{x}_{a}+\frac{\gamma^{2}}{2}\triangle_{H}.$$ The time evolution is in $x_{-}$ instead of $x_{0}$. In the formulae of sec.2 we replace $p_{0}$ by $p_{+}$ and $x_{0}$ by $x_{-}$. The integration measure in eq.(6) is $dx_{1}dx_{2}dx_{+}dx_{-}dp_{1}dp_{2}dp_{+}$.
Transport equations with friction
=================================
In order to achieve an equilibrium at a large time we add a friction term $K$ to the diffusion generator (21) $$K=K_{j}\frac{\partial}{\partial p_{j}}.$$ The transport equation (11) with friction reads $$\begin{array}{l}
\frac{1}{m}\eta^{\mu\nu} p_{\nu}\frac{\partial }{\partial
x^{\mu}}\Phi\cr=
\frac{\kappa^{2}m^{2}c^{2}}{2}\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial
p_{k}}(\delta_{jk}+m^{-2}c^{-2}p_{j}p_{k})\Phi-\frac{3\kappa^{2}}{2}\frac{\partial}{\partial
p_{k}}p_{k}\Phi-\frac{\partial}{\partial p_{k}}K_{k}\Phi.
\end{array}$$ Eq.(30) determines the drift $K$ if $\Phi_{E}$ is space-time independent $$\begin{array}{l}
K_{k}=\kappa^{2}m^{2}c^{2}\Phi_{E}^{-1}\Big(\frac{1}{2}\frac{\partial}{\partial
p_{j}}(\delta_{jk}+m^{-2}c^{-2}p_{j}p_{k})-\frac{3}{2}(mc)^{-2}p_{k}\Big)\Phi_{E}.\end{array}$$ Let us consider a class of solutions $\Phi_{E}$ depending solely on $p_{0}$. We write $\Phi_{E}$ in the form $$\Phi_{E}=p_{0}^{-1}\exp(f(\beta cp_{0})).$$ The $p_{0}^{-1}$ factor ensures the Lorentz invariance of the measure $d{\bf p}p_{0}^{-1}$. From eq.(31) we obtain $$K_{k}=\frac{\kappa^{2}}{2}p_{k}\beta cp_{0}f^{\prime}(\beta c
p_{0}).$$
We choose $\Phi$ in the form $$\Phi=p_{0}^{-1}\exp( f(c\beta p_{0}))\Psi$$
Then, the transport equation (16) reads $$\partial_{t}\Psi=\frac{c}{p_{0}}({\bf p}\nabla_{{\bf x}}+\hat{{\cal A}})\Psi$$ where $$\hat{{\cal A}}=K_{j}\partial^{j}+{\cal A}$$ and ${\cal A}$ is defined in eq.(21).
Explicitly $$\begin{array}{l}
\frac{\partial}{\partial
x^{0}}\Psi=\frac{1}{p_{0}}p_{j}\frac{\partial}{\partial x_{j}}\Psi
\cr
+\frac{\kappa^{2}m^{3}c^{2}}{2p_{0}}(\delta_{jk}+m^{-2}c^{-2}p_{j}p_{k})\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial
p_{k}}\Psi+\frac{m\kappa^{2}}{p_{0}}(\frac{3}{2}+\frac{1}{2}
c\beta f^{\prime}(\beta cp_{0})p_{0})p_{k}\frac{\partial}{\partial
p_{k}}\Psi.
\end{array}$$(the formulae (35)-(37) are not obvious but come out from calculations). The transport equation (without friction) in light-cone coordinates is $$p_{-}\partial^{x}_{+}+p_{+}\partial^{x}_{-}\Psi-p_{a}\partial_{a}\Psi-\frac{\gamma^{2}}{2}\triangle_{H}^{*}\Psi=0$$ where $x_{+}=x^{0}+x^{3}$ and $x_{-}=x^{0}-x^{3}$. The adjoint of the operator (27) in eq.(38) is with respect to the Lebesgue measure $dp_{+}dp_{1}dp_{2}$.
We look for an equilibrium solution of the transport equation in the form (it is not normalizable with respect to the integration over the transverse momenta $p_{a}$) $$\Phi_{E}=p_{+}^{-1}\exp(f(\beta c p_{+})).$$ We shall still discuss equilibrium distributions of the form (39) in sec.8. Here, we only point out that the light-cone coordinates are appropriate for particle beams moving in the direction of the third axis such that the remaining momenta $p_{-}$ and $p_{a}$ are small. An analogue of eq.(31) gives $$\begin{array}{l}
K=K_{-}\partial_{+}+K_{a}\partial_{a}=\frac{1}{2}\beta c
\kappa^{2}p_{+}^{2}f^{\prime}\partial_{+}
+\kappa^{2}p_{+}\partial_{+}-\frac{3}{2}\kappa^{2}p_{a}\partial_{a}
\end{array}$$ for the friction leading to the equilibrium (39). A particular (non-normalizable but relativistic invariant) distribution corresponds to $f=0$. We note that the change of coordinates $(p_{1},p_{2},p_{3})\rightarrow (p_{1},p_{2},p_{+})$ transforms the relativistic invariant measure $dp_{1}dp_{2}dp_{3}p_{0}^{-1}$ into $$dp_{1}dp_{2}dp_{+}p_{+}^{-1}.$$ With the friction (40) the diffusion generator (4) reads
$$\begin{array}{l}
2\gamma^{-2}{\cal
A}=\partial_{1}^{2}+\partial_{2}^{2}+(mc)^{-2}p_{+}^{2}\partial_{+}^{2}
+(mc)^{-2}p_{1}^{2}\partial_{1}^{2}\cr+
(mc)^{-2}p_{2}^{2}\partial_{2}^{2}+2(mc)^{-2}p_{+}p_{a}\partial_{a}\partial_{+}
+ 2(mc)^{-2}p_{1}p_{2}\partial_{1}\partial_{2} \cr
+5(mc)^{-2}p_{+}\partial_{+} + \beta c
\frac{1}{m^{2}c^{2}}p_{+}^{2}f^{\prime}\partial_{+}
\end{array}$$
and the diffusion operator $\hat{{\cal A}}$ of eqs.(13)-(14) and eq.(36) in the evolution equation for the proper time probability distribution $$\partial_{\tau}\Psi=(-p_{-}\partial^{x}_{+}-p_{+}\partial^{x}_{-}\Psi+p_{a}\partial_{a}\Psi
+\hat{{\cal A}})\Psi$$is $$\begin{array}{l}
2\gamma^{-2}\hat{{\cal
A}}=\partial_{1}^{2}+\partial_{2}^{2}+(mc)^{-2}p_{+}^{2}\partial_{+}^{2}
+(mc)^{-2}p_{1}^{2}\partial_{1}^{2}\cr+
(mc)^{-2}p_{2}^{2}\partial_{2}^{2}+2(mc)^{-2}p_{+}p_{a}\partial_{a}\partial_{+}
+ 2(mc)^{-2}p_{1}p_{2}\partial_{1}\partial_{2} \cr
+(mc)^{-2}p_{+}\partial_{+}+(mc)^{-2}2p_{a}\partial_{a} + \beta c
\frac{1}{m^{2}c^{2}}p_{+}^{2}f^{\prime}\partial_{+}+ 2\beta c
\frac{1}{m^{2}c^{2}}p_{+}p_{a}f^{\prime}\partial_{a}.
\end{array}$$
Stochastic equations
====================
In the Poincare coordinates the diffusion process is a solution of the linear stochastic differential equations ($\gamma^{2}=m^{2}c^{2}\kappa^{2}$) $$dq_{a}=\kappa q_{3}db_{a},$$ $a=1,2$, $$dq_{3}=-\frac{\kappa^{2}}{2}q_{3}d\tau+\kappa
q_{3}db_{3}=-\kappa^{2}q_{3}d\tau+\kappa q_{3}\circ db_{3},$$ where Stratonovitch differentials are denoted by a circle and the Ito stochastic differentials without the circle (the notation is the same as in [@ikeda]). The Brownian motion appearing on the rhs of eqs.(43) is defined as the Gaussian process with the covariance $$E[b_{a}(\tau)b_{c}(s)]=\delta_{ac}min(\tau,s).$$ The solution of eq.(43) is $$q_{3}(\tau)=\exp(-\kappa^{2}\tau+\kappa b_{3}(\tau))q_{3}$$ and $$q_{a}(\tau)=q_{a}+\kappa\int_{0}^{\tau}q_{3}(s)db_{a}(s).$$
In the light-cone coordinates $$p_{\pm}=p_{0}\pm p_{3}$$ we have $$dp_{+}=\kappa^{2}p_{+}d\tau+\kappa p_{+}\circ
db_{+}=\frac{3\kappa^{2}}{2}p_{+}d\tau+\kappa p_{+} db_{+},$$ $$dp_{a}=\kappa^{2}p_{a}d\tau-\kappa p_{a} db_{+}+\gamma db_{a}$$ where $a=1,2$.
The solution of eqs.(49) is $$\begin{array}{l}
p_{a}(\tau)=\exp(\kappa^{2}\tau-\kappa
b_{+}(\tau))p_{a}+\gamma\int\exp(\kappa^{2}s-\kappa
b_{+}(s))db_{a}(s)\equiv
e_{a}+\delta_{a}.\end{array}$$ The similarity between the solutions in light-cone coordinates and the Poincare coordinates is not accidental. It follows from the relation (23) between these coordinates.
A relativistic transport equation can be important in applications to high-energy plasma. In an electromagnetic field there is an additional drift term in the diffusion equation (4) $$K=\frac{e}{mc}F_{j\nu}p^{\nu}\partial^{j},$$ where $F_{\mu\nu}$ denotes the electromagnetic field tensor. The stochastic equations are linear and can be solved if the only components of $F$ are $F_{12}=B$ and $F_{30}=E$. In such a case the stochastic equations (48)-(49) read $$dp_{1}=\kappa^{2}p_{1}d\tau+\alpha Bp_{2}d\tau-\kappa p_{1}
db_{+}+\gamma db_{1},$$ $$dp_{2}=\kappa^{2}p_{2}d\tau-\alpha Bp_{1}d\tau-\kappa p_{2}
db_{+}+\gamma db_{2},$$ $$dp_{+}=\kappa^{2}p_{+}d\tau+\alpha Ep_{+}d\tau +\kappa p_{+}\circ
db_{+},$$ where $$\alpha=\frac{e}{mc}.$$ Eqs.(52)-(54) should be supplemented by equations determining the coordinates $dx_{a}=m^{-1}p_{a}d\tau$, $dx_{+}=m^{-1}p_{-}d\tau$ and $$dx_{-}=m^{-1}p_{+}d\tau.$$
A particular solution $\Phi_{E}$ of the transport equation in an electric field (51) ($F_{30}=E$, the remaining $F_{\mu\nu}=0$ in eq.(51)) $$-p_{+}\partial^{x}_{-}\Phi-p_{-}\partial^{x}_{+}\Phi+p_{a}\partial_{a}^{x}\Phi
+{\cal A}^{*}\Phi-\alpha E\partial_{+}p_{+}\Phi=0,$$ where ${\cal A}^{*} $ is the adjoint of ${\cal A}$ (eq.(41) with $f=0$), is
$$\Phi_{E}=p_{+}^{-1}.$$
Let $\Phi=p_{+}^{-1}\Psi$ then the proper time diffusion equation for $\Psi$ takes the form$$\partial_{\tau}\Psi=
(-p_{+}\partial^{x}_{-}-p_{-}\partial^{x}_{+}+p_{a}\partial_{a}^{x}
+\hat{{\cal A}}+\alpha E p_{+}\partial_{+})\Psi.$$ The transport equation reads $$\begin{array}{l}
\partial^{x}_{-}\Psi=-p_{+}^{-1}p_{-}\partial^{x}_{+}\Psi+p_{+}^{-1}p_{a}\partial_{a}^{x}\Psi
+p_{+}^{-1}\hat{{\cal A}}\Psi+\alpha
E\partial_{+}\Psi,\end{array}$$ where $\hat{{\cal A}}$ is defined in eq.(42) with $f=0$. In the light-cone coordinates the stochastic equation for the process $\hat{p}_{+}$ of eq.(15) in an electric field $E$ (but without friction) is $$d\hat{p}_{+}=\frac{1}{2}\kappa^{2}\hat{p}_{+}d\tau+\alpha E
\hat{p}_{+}d\tau +\kappa \hat{p}_{+}db_{+}.$$ It has the solution $$\hat{p}_{+}(\tau)=\exp(\alpha
E\tau+\kappa b_{+}(\tau)).$$ If we have a friction leading to the Jüttner equilibrium distribution (without an electromagnetic field and $f^{\prime}=-1$) then the stochastic equation for $\hat{p}_{+}$ is $$d\hat{p}_{+}=\frac{1}{2}\kappa^{2}\hat{p}_{+}d\tau-\frac{1}{2}\beta
c\kappa^{2}\hat{p}_{+}^{2}d\tau +\kappa \hat{p}_{+}db_{+}.$$ We discuss its solutions in sec.7.
Massless particles
==================
There is a substantial simplification of stochastic equations if $m=0$. In order to obtain the limit of zero mass we let $\delta_{jk}\rightarrow 0$ in eq.(21). The diffusion generator in the limit $m\rightarrow 0$ reads $$\begin{array}{l}
\triangle_{H}= p_{j}p_{k}\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial
p_{k}}+3p_{k}\frac{\partial}{\partial p_{k}}.
\end{array}$$
The transport equation for a diffusion with a friction is (when $m=0$ we use only the diffusion constant $\kappa^{2}$) $$\begin{array}{l}
\eta^{\mu\nu}p_{\nu}\frac{\partial }{\partial x^{\mu}}\Phi=
\frac{\kappa^{2}}{2}\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial
p_{k}}p_{j}p_{k}\Phi-\frac{3\kappa^{2}}{2}\frac{\partial}{\partial
p_{k}}p_{k}\Phi-\frac{\partial}{\partial
p_{k}}K_{k}\Phi.\end{array}$$ If $\Phi_{E}$ depends only on momenta then from eq.(61) $$\begin{array}{l}
K_{k}=\kappa^{2}\Phi_{E}^{-1}\Big(\frac{1}{2}\frac{\partial}{\partial
p_{j}}p_{j}p_{k}-\frac{3}{2}p_{k}\Big)\Phi_{E}.\end{array}$$ Then, eq.(13) reads (in the massless case the proper time is just an affine time parameter without any physical meaning) $$\begin{array}{l}
\partial_{\tau}\Psi_{\tau}=
\frac{1}{2}\kappa^{2}p_{j}p_{k}\partial^{j}\partial^{k}\Psi_{\tau}
+2\kappa^{2}p_{j}\partial^{j}\Psi_{\tau}+\frac{1}{2}\kappa^{2}p_{j}p_{k}(\partial^{j}
\ln\Phi_{E})
\partial^{k}\Psi_{\tau}-p_{\mu}\partial_{x}^{\mu}\Psi_{\tau}.
\end{array}$$ The transport equation follows from eqs.(62)-(63). In an explicit form $$\begin{array}{l}
\frac{\partial}{\partial
x^{0}}\Psi=\frac{1}{p_{0}}p_{j}\frac{\partial}{\partial x_{j}}\Psi
+\frac{\kappa^{2}}{2p_{0}}p_{j}p_{k}\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial p_{k}}\Psi\cr
+2\kappa^{2}p_{j}p_{0}^{-1}\partial^{j}\Psi
+\frac{1}{2}\kappa^{2}p_{j}p_{k}p_{0}^{-1}(\partial^{j}
\ln\Phi_{E})
\partial^{k}\Psi.
\end{array}$$ We discuss in more detail the Jüttner equilibrium distribution [@juttner] $$\Phi_{E}=p_{0}^{-1}\exp(-c\beta p_{0}).$$Then, from eq.(62) we obtain $$K_{k}=-\frac{c\kappa^{2}}{2}\beta p_{k}p_{0}.$$ The stochastic equation for the diffusion (15) in the proper time with the friction (66) reads $$d\hat{p}_{j}=
\frac{3\kappa^{2}}{2}\hat{p}_{j}d\tau
-\frac{c\beta\kappa^{2}}{2}\hat{p}_{j}\vert\hat{{\bf p}}\vert d\tau+\kappa
\hat{p}_{j}db.$$ From eq.(67) we obtain the stochastic equation for $\hat{p}_{0}=\vert \hat{ {\bf p}}\vert$ $$d\hat{p}_{0}=\frac{3\kappa^{2}}{2}\hat{p}_{0}d\tau-\frac{c\kappa^{2}}{2}\beta
\hat{p}_{0}^{2}d\tau +\kappa \hat{p}_{0}db.$$
Solution of the transport equation
===================================
In this section we wish to apply the Theorem of sec.2 in order to derive a solution of the transport equation. Let us begin with the spatial momenta as coordinates on the mass-shell ${\cal H}_{+}$. The stochastic equation for the transport diffusion $(\tilde{{\bf
x}}_{t},\tilde{{\bf p}}_{t})$ (17) of a massive particle solving eq.(37) in ${\bf p}$ coordinates reads $$\begin{array}{l}
d\tilde{p}_{j}=\frac{\tilde{p}_{j}\kappa^{2}m}{2} c^{2}\beta
f^{\prime}dt+\frac{3\kappa^{2}mc}{2\tilde{p}_{0}}\tilde{p}_{j}dt
+mc\sqrt{mc}\tilde{p}_{0}^{-\frac{1}{2}}\kappa db_{j}\cr + \sqrt{mc}
\kappa
(\tilde{p}_{0}^{\frac{1}{2}}-mc\tilde{p}_{0}^{-\frac{1}{2}})\tilde{{\bf
p}}^{-2}\tilde{p}_{j}\tilde{{\bf p}}d{\bf b}\cr \equiv
\frac{\tilde{p}_{j}\kappa^{2}m}{2} c^{2}\beta
f^{\prime}dt+\frac{3\kappa^{2}mc}{2\tilde{p}_{0}}\tilde{p}_{j}dt
+\sqrt{c} \kappa{p}_{0}^{-\frac{1}{2}}e_{j}^{n} db_{n},
\end{array}$$ $$d\tilde{x}^{j}=-c\tilde{p}_{j}\tilde{p}_{0}^{-1}dt,$$ here we defined the square root of the metric (the second order terms in eq.(21)) $$e^{n}_{j}e^{n}_{k}=\delta^{jk}+(mc)^{-2}p_{j}p_{k}.$$ The process (69)-(70) gives the solution of the transport equation (37) in the form $$\Psi_{t}({\bf x},{\bf p})=E[\Psi(\tilde{{\bf x}}_{t}({\bf x},{\bf p}),\tilde{{\bf
p}}_{t}({\bf p}))].$$ Unfortunately, the non-linear equations (69)-(70) are difficult to solve explicitly. We could also consider the proper time equations in these coordinates discussed extensively in [@habapre]. Then, we could apply the Theorem of sec.2. However, such a method would be fruitful only in numerical calculations. Computing the expectation values (18) can be more efficient then solving eqs.(69)-(70) directly.
The light-cone coordinates are more useful for analytic solutions.The formula (18) for the solution of the transport equation in these coordinates takes the form $$\begin{array}{l}
\Psi(x_{-};x_{+},x_{a},p_{+},p_{a})\cr=\frac{1}{m}\int_{0}^{\infty}d\tau
E\Big[\Psi(\hat{x}_{a}(\tau),\hat{x}_{+}(\tau),\hat{p}_{+}(\tau),\hat{p}_{a}(\tau))
\delta\Big(x_{-}-\frac{1}{m}\int_{0}^{\tau}ds\hat{p}_{+}(s)ds\Big)\hat{p}_{+}(\tau)\Big],\end{array}$$ where $\hat{p} $ is the stochastic process (15) generated by $\hat{{\cal A}}$ of eq.(42). If there is no friction then the stochastic equations of the diffusion (42) are linear and can be solved explicitly (see eqs.(50) and (58)). Then, the expectation value of a function $h$ of $\hat{p}_{a}$ has a representation in terms of its Fourier transform $$\begin{array}{l}E[h(\hat{p}_{1}(\tau),\hat{p}_{2}(\tau))]=\int
dk_{1}dk_{2}\tilde{h}(k_{1},k_{2})
E[\exp(ik_{a}\hat{p}_{a}(\tau))]\cr =\int
dk_{1}dk_{2}\tilde{h}(k_{1},k_{2})
E\Big[\exp(ik_{a}e_{a}(\tau))\exp\Big(-\frac{\gamma^{2}}{2}k_{a}k_{a}
\int \exp(2\gamma^{2}s+2\gamma b_{+}(s))ds\Big)\Big]\end{array}$$ ($e_{a}$ is defined in eq.(50)).The joint distribution law of $b,\int \exp b,\int\exp 2b$ can be explicitly calculated [@alili]. Hence, a solution of the transport equation (including the electromagnetic field (51)) can be obtained in a form of an integral kernel.
We consider here a simplified version of the Theorem of sec.2 when the initial condition $\Psi$ depends solely on $p_{+}$. Let us note that $p_{+}$ enters the generator of the diffusion with the factor $(mc)^{-1}$ (or with $\kappa=(mc)^{-1}\gamma$). Hence, it is irrelevant for moderate velocities. Our restriction to probabilities depending only on $p_{+}$ in fact applies to the ultrarelativistic limit when the transverse momenta can be neglected. In such a case the transport equation with the Jüttner friction (39) ($f^{\prime}=-1$) but without the electric field reads $$\begin{array}{l}
\partial_{-}\Psi=\frac{\kappa^{2}}{2}p_{+}\partial_{+}^{2}\Psi
+\frac{\kappa^{2}}{2}\partial_{+}\Psi -\frac{1}{2}c\beta\kappa^{2}
p_{+}\partial_{+}\Psi.
\end{array}$$ The transport equation with the electric field $E$ and without friction ($f=0$ in eq.(39)) defined in eq.(56) can be written explicitly as
$$\begin{array}{l}
\partial_{-}\Psi=\frac{\kappa^{2}}{2}p_{+}\partial_{+}^{2}\Psi
+(\frac{\kappa^{2}}{2}+\alpha E)\partial_{+}\Psi.
\end{array}$$
Eq.(71) gives the solution of eqs. (72)-(73) in the form $$\begin{array}{l}\Psi(x_{-},p_{+})\cr
=m^{-1}\int_{0}^{\infty}d\tau E\Big[\Psi(p_{+}(\tau,p_{+}))
\delta\Big(x_{-}-\frac{1}{m}\int_{0}^{\tau}ds p_{+}(s,p_{+})\Big)
p_{+}(\tau,p_{+})\Big].\end{array}$$ We can solve the diffusion equations (72)-(73) by means of the stochastic process (17). For eq.(72) the stochastic equation is $$d\tilde{p}_{+}=\frac{\kappa^{2}}{2}dx_{-}-\frac{1}{2}\beta c
\kappa^{2}\tilde{p}_{+}dx_{-}+\kappa\sqrt{\tilde{p}_{+}}db$$ whereas the stochastic process solving eq.(73) satisfies the equation $$d\tilde{p}_{+}=(\frac{\kappa^{2}}{2}+\alpha
E)dx_{-}+\kappa\sqrt{\tilde{p}_{+}}db.$$ We shall discuss solutions of eqs.(72) and (73) together with the transport equation for massless particles because all these equations are related to the Bessel diffusion.
Before solving the equations let us explain the text-book method [@skorohod] of the random change of time in eq.(67) (then we show that the Theorem of sec.2 is its efficient realization). Let us treat the formula (1) for $x^{0}(\tau)$ as a definition of the proper time $\tau$ $$\tau=\int_{0}^{x^{0}}\vert \hat{{\bf p}}_{s}\vert^{-1}ds$$ (here $\hat{{\bf p}}_{s}$ is the solution of the proper time diffusion (67); $x^{0}(\tau)$ can be defined implicitly by (77)). We can see from eq.(77) that $\tau $ depends only on events earlier than $x^{0}$. As a consequence $\hat{p}_{j}(x^{0})=p_{j}(\tau(x^{0}))$ is again a Markov process. Then, differentiating the momenta and coordinates according to the rules of the Ito calculus [@ikeda] we obtain the following Langevin equations (for mathematical details of a random change of time see [@skorohod][@ikeda]; a random time change from $x^{0}$ to $\tau$ is discussed in [@dun]) $$\begin{array}{l}
dp_{j}(x^{0})=
\frac{3\kappa^{2}}{2}\vert{\bf p}\vert^{-1}p_{j}dx^{0}
-\frac{c\beta\kappa^{2}}{2}p_{j} dx^{0}+
\kappa p_{j}\vert{\bf p}\vert^{-\frac{1}{2}}db,
\end{array}$$ $$dx^{j}=-p_{j}\vert {\bf p}(x^{0})\vert^{-1}dx^{0}.$$ Let $\Psi({\bf x},{\bf p})$ be an arbitrary function of ${\bf x}$ and ${\bf p}$ and $(x(x^{0},x,{\bf p}),{\bf p}(x^{0},{\bf p}))$ the solution of eqs.(78)-(79) with the initial condition $({\bf x},{\bf
p})$ then $$\Psi_{t}({\bf x},{\bf p})=E\Big[\Psi\Big({\bf
x}(x^{0},{\bf x},{\bf p}),{\bf p}(x^{0},{\bf p})\Big)\Big]$$ is the solution of the transport equation $$\begin{array}{l}
\frac{\partial}{\partial
x^{0}}\Psi=\frac{1}{p_{0}}p_{j}\frac{\partial}{\partial x_{j}}\Psi
+\frac{\kappa^{2}}{2p_{0}}p_{j}p_{k}\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial p_{k}}\Psi-\frac{\kappa^{2}}{2}\beta
c p_{k}\frac{\partial}{\partial
p_{k}}\Psi+\frac{3\kappa^{2}}{2p_{0}} p_{k}\frac{\partial}{\partial
p_{k}}\Psi
\end{array}$$ with the initial condition $\Psi({\bf
x},{\bf p})$. We can prove that eq.(81) is satisfied by differentiation of eq.(80) and application of the rules of the Ito calculus [@ikeda] (or the well-known relation between Langevin equation and the diffusion equation).
We return to the diffusion in the proper time in order to perform the average over the random time $\tau(x^{0})$ as prescribed by eq.(18) (this formula is treating the random change of time as a computational tool). From eqs.(67)-(68) it follows that $$\hat{{\bf p}}_{\tau}\vert \hat{{\bf
p}}_{\tau}\vert^{-1}\equiv {\bf
n}=(\cos\phi\sin\theta,\sin\phi\sin\theta,\cos\theta)=const$$ is time independent. Hence, according to eq.(18) the solution of the transport equation (64) takes the form $$\Psi_{t}({\bf x},\vert {\bf p}\vert,{\bf
n})=\frac{1}{mc}\int_{0}^{\infty}d\tau E\Big[\Psi({\bf
x}(\tau,{\bf x}),\vert \hat{{\bf p}}_{\tau}\vert,{\bf
n})\delta\Big(t-\frac{1}{mc} \int_{0}^{\tau}\vert \hat{{\bf
p}}_{s}\vert ds\Big)\vert \hat{{\bf p}}_{\tau}\vert\Big],$$ where $${\bf x}(\tau,{\bf x})={\bf x}-\int_{0}^{\tau}ds\vert \hat{{\bf
p}}_{s}\vert {\bf n}.$$
In order to calculate the expectation value (83) we apply the results of Yor [@yor][@yor2] who obtained (using the Feynman-Kac formula for an exponential potential) the joint distribution of $$B^{(\mu)}(\tau)=b_{\tau}+\mu \tau$$ and $$A^{(\mu)}(\tau)=\int_{0}^{\tau}ds\exp\Big(2 B^{(\mu)}(s)\Big).$$ We need some rescalings (using $\alpha b(s)=b(\alpha^{2} s)$) in order to bring our formulae to the Yor’s form. We write $$\exp(\kappa^{2}s+\kappa
b(s))=\exp\left(2\left(2\kappa^{2}\frac{s}{4}+b\left(\kappa^{2}\frac{s}{4}\right)\right)\right).$$ It follows that $\mu=2$ in Yor’s formula.
After such an rescaling an expectation value of a function of $B_{\tau}=\kappa^{2}\tau +\kappa b(\tau)$ and $A(\tau)=\int_{0}^{\tau}ds\exp
2B_{s}$ is [@yor][@yor2]$$\begin{array}{l}
E\Big[F\Big(B(\tau),A(\tau)\Big)\Big]=\exp(-\frac{\tau\kappa^{2}}{2})\int
F(v,\frac{4}{\kappa^{2}}\vert{\bf p}\vert y)\exp(2v) \cr
\exp\Big(-\frac{1}{2y}(1+\exp(2v))\Big)\theta(y^{-1}\exp(v),\frac{\tau\kappa^{2}}{4})y^{-1}dydv
\end{array}$$ where ($v\in R$,$y\in R_{+}$) $$\begin{array}{l}
\theta(r,\tau)=r(2\pi\tau)^{-\frac{1}{2}}\exp(\frac{\pi^{2}}{2\tau})\int_{0}^{\infty}
\exp(-\frac{\xi^{2}}{2\tau}-r\cosh(\xi))\sinh(\xi)\sin(\frac{\pi\xi}{\tau})d\xi.
\end{array}$$ The $\tau$-integral in the Theorem can be calculated by an application of another Yor’s formula $$\int_{0}^{\infty}d\tau\exp(-2\nu\tau)\theta(r,\tau)=I_{\sqrt{2\nu}}(r),$$ where $I_{\alpha}$ is the modified Bessel function of order $\alpha$ [@grad]. Hence, $$\begin{array}{l}\int_{0}^{\infty}d\tau E[F(B(\tau),A(\tau))]=
\frac{4}{\kappa^{2}}\int F(v,\frac{4}{\kappa^{2}}\vert{\bf p}\vert
y)\exp(2v)\cr
\exp\Big(-\frac{1}{2y}(1+\exp(2v))\Big)I_{2}(y^{-1}\exp(v))y^{-1}dydv
\end{array}$$ We can apply the formula (91) only to the expectation values (18) of diffusions without friction (because only in this case the integrand (18) depends solely on $A$ and $B$) corresponding to the Jüttner distribution with $f=0$ (this is the Lorentz invariant but non-normalizable distribution $dxd{\bf p}\Phi_{E}=dxd{\bf
p}p_{0}^{-1}$). Then, with the proper rescalings we have (we omit the index $\mu=2$ in $A$ and $B$) $${\bf p}_{\tau}={\bf n}\vert{\bf p}\vert\exp
\Big(2B(\frac{\tau\kappa^{2}}{4})\Big),$$ $${\bf x}_{\tau}={\bf x}-\frac{4}{\kappa^{2}}{\bf n}\vert {\bf
p}\vert A(\frac{\tau\kappa^{2}}{4}).$$ We can insert these formulae in eqs.(18) and (83). After the change of variables $$v=\frac{1}{2}\ln(\frac{r}{\vert{\bf p}\vert})$$ we obtain from the Theorem (in eq.(18) the integrals over $\delta$-function can be performed directly and only the $v$-integral remains) $$\begin{array}{l}
\Psi_{t}({\bf x},{\bf p})=t^{\prime
-1}\int_{0}^{\infty}dr\frac{r}{\vert {\bf p}\vert}
\exp\Big(-\frac{1}{2t^{\prime}}(\vert {\bf p}\vert +r)\Big) \cr
I_{2}(t^{\prime -1}\sqrt{\vert{\bf p}\vert r} ) \Psi({\bf x}-{\bf
n}ct,{\bf n}r),
\end{array}$$ where $t^{\prime}=\frac{c\kappa^{2}}{4}t$.
Calculations in the light-cone coordinates for a massive particle diffusing without a friction ($f=0$ in eq.(39)) are similar. We obtain from eq.(58) that $\mu=0$ for $E=0$.Then, from eqs.(88)-(90) (we change variables $r=p_{+}\exp(2v)$) $$\begin{array}{l}
\Psi_{t_{-}}(p_{+})=p_{+}\int_{0}^{\infty}d\tau
E\Big[\Psi\Big(p_{+}\exp(2B)\Big)
\exp(2B)\delta(t-\frac{4}{m\kappa^{2}}p_{+}A)\Big]\cr=p_{+}
\int\exp(2v)\Psi(\exp(2v)p_{+})I_{0}(\frac{\exp (v)}{y})
\delta(t-\frac{4}{m\kappa^{2}}p_{+}y)\cr
\exp(-\frac{1}{2y}(1+\exp(2v)))dvy^{-1}dy
=t_{-}^{-1}\int dr
\Psi(r)\exp(-\frac{p_{+}+r}{2t_{-}})I_{0}(\frac{1}{t_{-}}\sqrt{rp_{+}}),
\end{array}$$ where $t_{-}=\frac{1}{4}\kappa^{2}x_{-}$.
We can generalize this result to a particle in a constant electric field. Then, from eq.(58) after a rescaling similar to eq.(87) we obtain $$\mu=\frac{2\alpha E}{\kappa^{2}}.$$ Hence, in the transition function we shall have $I_{\mu}(\frac{1}{t_{-}}\sqrt{p_{+}r})$. The solution of eq.(73) obtained from eq.(74) is $$\begin{array}{l}
\Psi_{t_{-}}(p_{+})=t_{-}^{-1}\int dr(\frac{r}{p_{+}})^{\mu}
\Psi(r)\exp(-\frac{p_{+}+r}{2t_{-}})I_{\mu}(\frac{1}{t_{-}}\sqrt{rp_{+}}).
\end{array}$$
We shall check the results (95)-(98) of the probabilistic averaging by solving the transport equations (73) and (81) directly. Let $\rho=\vert {\bf p}\vert $ then from eq.(78) $\rho$ satisfies the stochastic equation $$d\rho=\frac{3\kappa^{2}c}{2}dt-
\frac{c^{2}\beta\kappa^{2}}{2}\rho dt+\sqrt{c}\kappa \rho^{\frac{1}{2}}db$$ whereas $$dx^{j}=-c\vert {\bf p}\vert^{-1}p_{j}dt=-cn_{j}dt.$$ Eq.(100) has the solution $${\bf x}(t)={\bf x}-{\bf
n}ct.$$
The transport equation (64) for massless particles with the Jüttner friction can be expressed in the form $$\partial_{t}\Psi=\frac{c\kappa^{2}}{2} r\partial_{r}^{2}\Psi+(\frac{\sigma}{4}-\frac{\nu}{2}
r)\partial_{r}\Psi -{\bf n}\nabla_{{\bf x}}\Psi$$ where $\sigma=6\kappa^{2}c $ and $\nu=\beta c^{2}\kappa^{2}$. The transport equation (102) is related to the (generalized) Bessel diffusion $$\partial_{t}\chi=\frac{\kappa^{2}c}{2}\partial_{\rho}^{2}\chi
+ \frac{\sigma -\kappa^{2} c}{2\rho}\partial_{\rho}\chi-\nu
\rho\partial_{\rho}\chi.$$ The diffusion process $\rho_{t}$ solving the diffusion equation (103) (i.e., $\chi_{t}=E[\chi(\rho_{t})]$) is the solution of the stochastic equation $$d\rho=\frac{\sigma -\kappa^{2}c}{2\rho}dt-\nu \rho dt+\kappa db.$$ Let $$r=\frac{1}{4}\rho^{2}$$ then $$dr=\frac{\sigma}{4}dt -\frac{\nu}{2}rdt+\kappa \sqrt{c}\sqrt{r}db.$$ If $\nu=0$ then the Kolmogorov transition function ( the probability density to go from $x$ to $y$ in time $t$) for the Bessel process is (ref.[@ikeda],sec.IV.8; see also [@yorpit][@watanabe]) $$P(t,x,y)=\frac{1}{t^{\prime}}(xy)^{1-\frac{\tilde{\sigma}}{2}}y^{\tilde{\sigma}-1}I_{\frac{\tilde{\sigma}}{2}-1}
(\frac{xy}{t^{\prime}})$$ where $\tilde{\sigma}=\frac{\sigma}{\kappa^{2}c}$ and $t^{\prime}=\frac{t\kappa^{2}c}{4}$. Eq.(107) coming from ref.[@ikeda] coincides with our results (95)-(96).
We can calculate the Laplace transform of the process $p_{+}(t)$ applying the results (91),(96) and (98) of the Theorem of sec.2 to $\Psi=\exp(-\lambda r)$ and using the integrals 6.614 from ref. [@grad] $$\begin{array}{l}
E[\exp(-\lambda p_{+}(t_{-},p_{+}))]=t_{-}^{-1}\int dr \exp(-\lambda
r)\exp(-\frac{p_{+}+r}{2t_{-}})I_{0}(\frac{1}{t_{-}}\sqrt{rp_{+}})
\cr=(1+\lambda t_{-})^{-1}\exp\Big(-\lambda p_{+}(1+\lambda
t_{-})^{-1}\Big). \end{array}$$ The diffusion process (73) in an electric field $E$ has the Laplace transform $$(1+\lambda t_{-})^{-1-2\alpha E\kappa^{-2}}\exp\Big(-\lambda
p_{+}(1+\lambda t_{-})^{-1}\Big).$$The Laplace transform (108)-(109) has been calculated earlier in [@ikeda] (sec.IV.8).
The solution of the transport equation with friction cannot be calculated with Yor’s formula (88). The diffusion process (see eqs.(117)-(118) below) depends on the variables $A$ and $B$ but in eq.(18) we still need its integral over time. Nevertheless, we can solve the diffusion equations (72) and (75) directly applying their relation to the (generalized) Bessel diffusion (104) with $\nu>0$. We have calculated the transition function of the process $\rho_{t}$ with $\nu\neq 0$ in [@habarep]$$\begin{array}{l}
P_{t}(x,y)=\exp(t\nu(\alpha+1))(\sinh \nu t)^{-1}
y^{\alpha+1}x^{-\alpha}I_{\alpha}(\nu xy(\sinh \nu t)^{-1})
\cr\exp(-\frac{\nu}{2}(\coth (\nu t)+1)y^{2}-\frac{\nu}{2}(\coth (\nu
t)-1))x^{2}),
\end{array}$$ where $$\alpha=(\kappa^{2}c)^{-1}\frac{\sigma}{2}-1$$ ($\alpha=2$ in the model of the diffusion of massless particles). The transition function (108) describes the imaginary time evolution in quantum mechanics with the potential $V(x)=\frac{\nu}{2}x^{2}+gx^{-2}$ (then $\alpha=\frac{1}{2}\sqrt{1+8g}$,[@habarep]).
The Laplace transform of the transition function of the process $r_{t}$ (with $\nu>0$) starting from $k$ which is the square of the generalized Bessel process can be calculated using the transition function (110) (it has been has been derived earlier in [@ikeda]). We obtain $$\begin{array}{l} E[\exp(-\lambda r_{t})]=\int
dyP_{t}(k,y)\exp(-\lambda
y^{2})=\Big(-\lambda\kappa^{2}\nu^{-1}(\exp(-\frac{\nu}{2}
t)-1)+1\Big)^{-\frac{\sigma}{2c\kappa^{2}}}\cr
\exp\Big(-\Big(-\lambda\kappa^{2}\nu^{-1}(\exp(-\frac{\nu}{2}
t)-1)+1\Big)^{-1}\lambda\exp(-\frac{\nu}{2} t)k\Big).\end{array}$$ For the massless particles $\sigma=6c\kappa^{2}$ , $\nu=\kappa^{2}c\beta$; for the process $p_{+}$ with the Jüttner friction ($f^{\prime}=-1$) we have $\sigma=2c\kappa^{2}$ and $\nu=\kappa^{2}c\beta$ . The expansion of eq.(111) for the diffusion of massless particles $$E[\exp(-\lambda r_{t})]-\frac{1}{2}(c\beta)^{3} \int
dyy^{2}\exp(-c\beta y-\lambda y)
=\sum_{n=1}^{\infty}c_{n}(k)\exp(-n\frac{\nu}{2} t)$$ shows that $d\mu=\frac{1}{2} (c\beta)^{3}y^{2}\exp(-c\beta y)$ is the invariant measure for the process $\rho_{t}$ and that the convergence to the equilibrium is exponential with the speed $\frac{2}{\nu}$ proportional to the temperature.
Applying the transition function (110) we can calculate the solution of the transport equation (81) as an expectation value over the stochastic process $$\Psi_{t}({\bf x},\vert{\bf p}\vert,{\bf n})=E[\Psi({\bf x}-c{\bf
n}t,\rho_{t},{\bf n})] =\int_{0}^{\infty}dyP_{t}(\sqrt{\vert{\bf
p}\vert},\sqrt{y})\psi({\bf x}-c{\bf n}t,y,{\bf n})$$
From eq.(113) it follows that if $\Psi$ as the function of ${\bf x}$ has no limit at infinity then the limit $t\rightarrow$ does not exist (this is a consequence of the wave propagation). If the limit $t\rightarrow \infty$ in eq.(113) exists then it follows that $\Psi_{t_{-}}\rightarrow 1$ and $\Phi_{t_{-}}\rightarrow \Phi_{E}$ (up to a normalization constant). Moreover, for the momentum probability distribution of an “observable” $\phi$ in a “state” $\Phi$ we obtain
$$\begin{array}{l}
lim_{t\rightarrow \infty}\int d{\bf x}d{\bf p}\phi({\bf
p})\Phi_{t}({\bf x},{\bf p}) \cr=\Big(\int d{\bf p}\Phi_{E}({\bf
p})\phi({\bf p})\Big)\frac{1}{8\pi}(\beta c)^{3} \int d{\bf p}d{\bf
x}\exp(-c\beta\vert{\bf p}\vert)\Psi({\bf x},{\bf p})
\end{array}$$
As an example of the solution (113) we could consider the plane wave with the initial condition $\Psi({\bf x},\vert{\bf p}\vert,{\bf
n})=\exp(i\vert{\bf p}\vert {\bf nx})$. Then, from eq.(111) $$\begin{array}{l}
\Psi_{t}({\bf x},\vert{\bf p}\vert,{\bf
n})=\Big(\frac{i}{c\beta}({\bf
nx}-ct)(\exp(-\frac{1}{2}\kappa^{2}c\beta t)-1)+1\Big)^{-3}\cr
\exp\Big(i\Big(\frac{i}{c\beta}({\bf
nx}-ct)(\exp(-\frac{1}{2}\kappa^{2}c\beta
t)-1)+1\Big)^{-1}\exp(-\frac{1}{2}\kappa^{2}c\beta t)\vert{\bf
p}\vert({\bf nx}-ct)\Big)\end{array}$$ As a result we obtain again a wave (note that $\vert{\bf p}\vert({\bf x}-{\bf
n}ct){\bf n}=p_{\mu}x^{\mu}$ in the relativistic notation) moving in the direction ${\bf n}$ with a decreasing amplitude and exponentially decreasing wave vector.
We can solve the stochastic equations (68) and (75) and calculate expectation values of functions of the processes with a friction. The solution of eq.(68) reads $$\vert{\bf p}_{\tau}\vert=\vert {\bf
p}\vert\exp(\kappa^{2}\tau+\kappa
b_{\tau})\Big(1+\frac{\kappa^{2}}{2}\beta\vert{\bf p}\vert
\int_{0}^{\tau}ds\exp(\kappa^{2}s+\kappa b_{s})\Big)^{-1}$$ The stochastic equation (75) for $p_{+}$ can also be solved with the result $$p_{+}(\tau)= p_{+}\exp(\kappa
b_{\tau})\Big(1+\frac{\kappa^{2}}{2}\beta p_{+}
\int_{0}^{\tau}ds\exp(\kappa b_{s})\Big)^{-1}$$ Unfortunately, with the Yor’s formula we are unable to calculate the expectation value (18). However, we can obtain expectation values $E[F({\bf p}_{\tau})]$ of any function of the process. This may be sufficient for a calculation of some mean values with the probability distribution $\Phi$ as a solution of the transport equation (11), for example applying eq.(18)$$\int dxd{\bf p}\phi({\bf p})\Phi(x,{\bf p})= \int d{\bf
p}\phi({\bf p})\Phi_{E}({\bf p})\int_{0}^{\infty}d\tau E[\int
d{\bf x}\Psi({\bf x},\hat{{\bf p}}_{\tau}({\bf
p}))\hat{p}_{0}(\tau)].$$ The rhs is of the Yor’s form (88) also for the processes (116)-(117) with a friction.
Transport equations in a moving frame
=====================================
We consider a covariant form of the equilibrium distribution $$\Phi_{E}=p_{0}^{-1}\exp(f(c\beta^{\mu}p_{\mu})).$$ The four-vector $\beta^{\mu}$ can be related to the velocity of the frame of reference in such a way that in the rest frame $\beta=(\frac{1}{kT},0,0,0)$, where $T$ is the temperature and $k$ is the Boltzmann constant (see [@cmb][@patria][@kujaw][@weldon][@svet2] for a discussion of such equilibrium distributions).Then, the probability distribution $d\sigma(x,p,\beta)=dxd{\bf p}\Phi_{E}$ transforms in a covariant way under Lorentz transformations $\Lambda$ $$d\sigma(\Lambda x,\Lambda p,\Lambda \beta)=d\sigma(x,p,\beta).$$ An expectation value in a “state” $\Phi$ satisfying the transport equation (11) is $$\begin{array}{l}
\int d{\bf p}d{\bf x}\phi({\bf x},{\bf p})\Phi_{t}({\bf x},{\bf
p})=\int d\sigma(x,p,\beta)\phi({\bf x},{\bf p})E[\Psi(\hat{{\bf
x}}_{\tau}({\bf x},{\bf p}),\hat{{\bf p}}_{\tau}({\bf p}))].
\end{array}$$ We applied the formula (18) in order to express the time evolution of the solution $\Phi_{t}$ of the transport equation by the proper time. In eq.(119) $\tau(t)$ is expressed by the coordinate time from the equation $t=(mc)^{-1}\int_{0}^{\tau}\hat{p}_{0}ds$ .
If $m>0$ then the drift (31) reads $$K_{j}=-\frac{m^{2}c^{3}\kappa^{2}}{2}\beta_{j}f^{\prime}+\frac{\kappa^{2}c}{2}\beta^{\mu}p_{\mu}p_{j}f^{\prime}.$$ In the massless case $$K_{j}=\frac{c\kappa^{2}}{2}\beta^{\mu}p_{\mu}p_{j}f^{\prime}.$$
The covariant form of the transport equation takes the form $$\begin{array}{l}
\frac{\partial}{\partial
t}\Psi=\frac{c}{p_{0}}p_{j}\frac{\partial}{\partial x_{j}}\Psi
+\frac{\kappa^{2}m^{3}c^{3}}{2p_{0}}(\delta_{jk}+m^{-2}c^{-2}p_{j}p_{k})\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial
p_{k}}\Psi\cr+\frac{m\kappa^{2}c}{p_{0}}(\frac{3}{2}+\frac{1}{2}
c\beta^{\mu}p_{\mu}f^{\prime} )p_{k}\frac{\partial}{\partial
p_{k}}\Psi+\frac{m^{3}c^{4}\kappa^{2}}{2p_{0}}\beta_{k}\frac{\partial}{\partial
p_{k}}\Psi.
\end{array}$$ In the massless case $$\begin{array}{l}
\frac{\partial}{\partial
t}\Psi=\frac{c}{p_{0}}p_{j}\frac{\partial}{\partial x_{j}}\Psi
+\frac{c\kappa^{2}}{2p_{0}}p_{j}p_{k}\frac{\partial}{\partial
p_{j}}\frac{\partial}{\partial p_{k}}\Psi+\frac{\kappa^{2}}{2p_{0}}
c^{2} \beta^{\mu}p_{\mu}p_{k}f^{\prime} \frac{\partial}{\partial
p_{k}}\Psi+\frac{c\kappa^{2}}{2p_{0}} p_{k}\frac{\partial}{\partial
p_{k}}\Psi.
\end{array}$$ In the proper time formalism the diffusion process solving the diffusion equation (15) satisfies the equation $$d\hat{p}_{j}=\frac{3\kappa^{2}}{2}\hat{p}_{j}d\tau-\frac{m^{2}c^{3}\kappa^{2}}{2}\beta_{j}f^{\prime}d\tau
+\frac{\kappa^{2}}{2}\beta^{\mu}\hat{p}_{\mu}\hat{p}_{j}f^{\prime}
d\tau +\kappa e^{j}_{n}db^{n}.$$ where $e^{j}_{n}$ (the square root of the metric in the second order differential operator (21), see [@habapre]) has been defined in eq.(69). The diffusion process $\hat{{\bf p}}_{\tau}$ (124) is the same as the process ${\bf p}_{\tau}$ generated by ${\cal A}$ (eq.(7)) because ${\hat{\cal A}}=\Phi_{E}^{-1}{\cal A}\Phi_{E}={\cal
A}$. We have obtained this equality by direct calculations but there should be a deeper reason for it.
We can derive from eq.(124) the formula for $dp_{0}$. Let us define $$\pi=\beta^{\mu}p_{\mu}$$ and $$X=\beta^{\mu}x_{\mu}.$$ Then, a direct calculation (using the Ito stochastic calculus [@ikeda]) leads to the proper time stochastic equations
$$\begin{array}{l}
d\pi=\frac{3\kappa^{2}}{2}\pi d\tau
+\frac{c\kappa^{2}}{2}\pi^{2}f^{\prime}d\tau\cr
-\frac{m^{2}c^{3}\kappa^{2}}{2}\beta^{\mu}\beta_{\mu}f^{\prime}d\tau
+\frac{\kappa}{p_{0}}(\beta_{0}p_{j}-\beta_{j}p_{0})e^{j}_{n}db^{n}
\end{array}$$
$$\frac{dX}{d\tau}=\pi.$$
In the massless case eq.(127) reads $$\begin{array}{l}
d\pi=\frac{3\kappa^{2}}{2}\pi d\tau
+\frac{c\kappa^{2}}{2}\pi^{2}f^{\prime}d\tau +\kappa\pi db.
\end{array}$$
Using the variables $(X,\pi)$ we can write a proper time diffusion equation for an evolution of the probability distribution (13) $\Psi(X,\pi)$ assuming that it depends solely on the variables $(X,\pi)$ $$\begin{array}{l}
\partial_{\tau}\Psi=\pi\partial_{X}\Psi+(\frac{3\kappa^{2}}{2}\pi
+\frac{c\kappa^{2}}{2}\pi^{2}f^{\prime}
-\frac{m^{2}c^{3}\kappa^{2}}{2}\beta^{\mu}\beta_{\mu})\partial_{\pi}\Psi
\cr
+\frac{\kappa^{2}}{2}(\pi^{2}-m^{2}c^{2}\beta^{\mu}\beta_{\mu})\partial_{\pi}^{2}\Psi
\end{array}$$ In the massless case this diffusion equation reads $$\begin{array}{l}
\partial_{\tau}\Psi=\pi\partial_{X}\Psi+(\frac{3\kappa^{2}}{2}\pi
+\frac{c\kappa^{2}}{2}\pi^{2}f^{\prime})\partial_{\pi}\Psi
+\frac{\kappa^{2}}{2}\pi^{2}\partial_{\pi}^{2}\Psi
\end{array}$$ The transport equation for massless particles is the same as the one for the Bessel diffusion $$\partial_{X}\Psi=\frac{\kappa^{2}}{2}\pi\partial^{2}_{\pi}\Psi
+(\frac{3\kappa^{2}}{2} +\frac{c\kappa^{2}}{2}\pi
f^{\prime})\partial_{\pi}\Psi.$$ In order to preserve an interpretation of $\beta$ as the temperature the four-vector $\beta$ should be time-like. We can see from eq.(130) that the model simplifies substantially (the transport equation reduces to the Bessel diffusion) if we let $\beta^{\mu}\beta_{\mu}\rightarrow 0$. In a formal limit $\beta=(\beta_{-},0,0,0)$ (when $\beta$ is on the light cone) and with $\partial_{\tau}\Psi=0$ we obtain the model (72) corresponding to the light cone coordinates with the exception of the factor $\frac{3\kappa^{2}}{2}\partial_{\pi}$ in eq.(130) which is replaced by $\frac{\kappa^{2}}{2}\partial_{+}$ in eq.(72). The incorrect factor in the formal limit comes from the incorrect probability measure in this limit which in the light-cone coordinates should be $dp_{1}dp_{2}dp_{+}p_{+}^{-1}\exp f$ and not $d{\bf
p}p_{0}^{-1}\exp f$. The factor $p_{+}^{-1}$ is responsible for the difference between the term $\frac{3\kappa^{2}}{2}\partial_{\pi}$ in eq.(130) and $\frac{\kappa^{2}}{2}\partial_{+}$ in eq.(72).
Relativistic invariance means that a Lorentz transformation of the processes $X$ and $\pi$ can be shifted into a transformation of the frame of reference, i.e., for a function $F$
$$E[F(\Lambda x(\tau,\beta,\Lambda x, \Lambda p),\Lambda
p(\tau,\beta,\Lambda p))]= E[F(x(\tau,\Lambda^{-1}\beta, x,
p),p(\tau,\Lambda^{-1}\beta, p))]$$
Discussion
==========
In our earlier papers we have developed a formalism of relativistic diffusions evolving in the proper time. Such a formalism is useful because the explicit relativistic invariance is a strong guiding principle when building relativistic diffusion equations. In this paper we have shown that a solution of the stochastic equations evolving in the proper time allows a construction of the solution of the transport equation (in the laboratory time) by means of an integration over the proper time. We have shown how the method works in some models which can be solved exactly. These examples are not typical for the proper time equations. For massless particles the proper time is just an affine parameter on the trajectory without a physical meaning. A function depending solely on $p_{+}$ can describe a massive particle whose motion is restricted to a straight line (in the direction of the third axis). The results concerning the transport equations described by the Bessel diffusion of momenta could find applications in astrophysics, in plasma physics and in heavy ion collisions. The momentum $\vert{\bf p}\vert$ evolves in time as the Bessel process which has some distinguished features among all diffusion processes [@yorpit][@watanabe]. In particular, the (generalized) Bessel diffusion has the exceptional property that an exponential initial distribution remains exponential for any time.
In general, the method of proper time could be applied for approximate calculations and computer simulations. If $m>0$ then an expansion parameter $(mc)^{-1} $ appears in all our equations. We could use it in the Theorem (eq.(18)) in order to calculate the integral over $\tau$ with an expanded argument of the $\delta $-function ($\hat{{\bf p}}_{s}={\bf b}_{s}$ in the lowest order of the expansion in $(mc)^{-1}$)$$\delta\Big(t-\tau-\frac{1}{2m^{2}c^{2}}\int_{0}^{\tau}ds {\bf
b}_{s}^{2}\Big).$$ We can now calculate the expectation value (18) in a way similar to the one in this paper because the probability distribution of the integral $\int_{0}^{\tau}ds {\bf b}_{s}^{2}$ is known.
The original relativistic diffusion in the proper time of Schay [@schay] and Dudley [@dudley] is explicitly Lorentz invariant because its proper time evolution is generated by a Lorentz invariant differential operator. In applications we need diffusions which have a limit for a large time. The necessary drag terms which force the process to an equilibrium cannot be Lorentz invariant. The reason is that the notion of the equilibrium is itself frame dependent. We described this frame dependence in the last section. The stochastic process transforms in a covariant way with respect to the Lorentz transformation if together with the process we transform also the frame.
[99]{}J. Dunkel and P. Hänggi,Phys.Rep.[**471**]{},1(2009)
C. Chevalier and F. Debbasch, AIP Conf.Proc.[**913**]{},42(2007) C. Chevalier and F. Debbasch,
Journ.Math.Phys.[**49**]{},043303(2008)
G.Schay,PhD thesis,Princeton University,1961 R.Dudley, Arkiv for Matematik,[**6**]{},241(1965) J. Franchi and Y. Le Jan, Commun.Pure Appl.Math.[**60**]{},187(2007) J. Franchi, Commun.Math.Phys.[**290**]{},523(2009) Z. Haba,Journ.Phys.[**A42**]{},445401(2009) Z. Haba, arXiv:0909.2880 Z. Haba, Phys.Rev.[**E79**]{},021128(2009)
Z.Haba,Mod.Phys.Lett.A, arXiv:0910.2253 A.S. Kompaneets, JETP,[**47**]{},1939(1956)(in Russian)
G.B. Rybicki and A.P. Lightman,
Radiative Processes in Astrophysics,Wiley-VCH,1979 P.J.E. Peebles, Physical Cosmology, Princeton University Press,1971 S. Dodelson, Modern Cosmology, Academic Press, New York,2003 R.Rapp and H. van Hees,Int.Journ.Mod.Phys.[**E**]{}(2009) arXiv:0903.1096 B.Svetitsky, Phys.Rev.[**D37**]{},2484(1988) J.M. Stewart, Non-equilibrium Relativistic Kinetic Theory,Lect.Notes in Physics,Vol.10,Springer,1971 W. Israel, Journ.Math.Phys.[**4**]{},1163(1963)
P. Carruthers and F. Zachariasen,
Phys.Rev.[**D13**]{},950(1976) H.-Th. Elze and U.Heinz, Phys.Rep.[**183**]{},81(1989) P.F. Kelly, Q. Liu, G.Lucchesi and C. Manuel,
Phys.Rev.Lett.[**72**]{},3461(1994),Phys.Rev.[**D50**]{},4209(1994)
L.D. Landau and E.M. Lifshits,
Field Theory, Pergamon Press,Oxford,1981 I.I. Gikhman and A.V. Skorohod,
Stochastic Differential Equations,Springer,Berlin,1972 R.P. Feynman, Phys.Rev.[**80**]{},440(1950),[**84**]{},108(1951)
H. Matsumoto and M. Yor, Probability Surveys,[**2**]{},312(2005) M.Yor,Adv.Appl.Prob.[**24**]{},509(1992) J. Dunkel, P. Hänggi and S. Weber, Phys.Rev.[**E 79**]{},010101(R)(2009) N. Ikeda and S. Watanabe, Stochastic Differential Equations and Diffusion Processes,North Holland,1981 L. Alili, H. Matsumoto and T. Shiraishi, in Lecture Notes in Math., Vol.1755,Springer,2001
J.Pitman and M.Yor, Z.Wahr.verw.Geb.[**59**]{},425(1982)
T. Shiga and S. Watanabe, Z.Wahr.verw.Geb.[**27**]{},37(1973)
I.S.Gradshteyn , I.M. Ryzhik, Tables of Integrals,Series and Products,
Nauka, Moscow,1971(in Russian)
F. Jüttner, Ann.Phys.(Leipzig)[**34**]{},856(1911)
Z. Haba, Reports Math.Phys.[**18**]{},257(1980)
P.J.E. Peebles and D.T. Wilkinson, Phys.Rev.[**174**]{},2168(1968)
G.R. Henry, R.B. Feduniak, J.E. Silver and M.A. Peterson,
Phys.Rev.[**176**]{},1451(1968) R.K. Patria,Proc.Phys.Soc.[**88**]{},791(1966)
J.H. Eberly and A.Kujawski, Phys.Rev.[**155**]{},10(1967)
H.A. Weldon, Phys.Rev.[**D26**]{},1394(1982)
T.Matsui,B. Svetitsky and L.D.McLerran,Phys.Rev.[**D34**]{},783(1986)
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.